International Phenomenological Society

Foundationalism and the Infinite Regress of Reasons

Metaepistemology and Skepticism by Richard Fumerton
Review by: Peter Klein
Philosophy and Phenomenological Research, Vol. 58, No. 4 (Dec., 1998), pp. 919-925
Published by: International Phenomenological Society

Richard Fumerton's admirable defense of foundationalism depends upon some arguments against
the possibility of the infinite regress of justificatory reasons. One reason for preferring
foundationalism is, he claims, that infinitism (the view that justificatory reasons are infinite)
should be rejected.1 I will try to show that the arguments against infinitism do not succeed. Since
a key ingredient in his rejection of infinitism is his account of inferential justification, I will offer
some critical comments about that along the way.

There is one important claim on which Fumerton and I agree that should be mentioned before I
begin to discuss his rejection of infinitism. It is central to that discussion:

(B) S's belief that P should be explicated in terms of S's having a disposition to endorse P under
some set of appropriate circumstances. (p. 58)2

Now, what is the essential claim of infinitism? This: The reasons that justify a belief are
members of a chain (perhaps branching) that is infinitely long and non-repeating.

Fumerton's reasons for rejecting infinitism can be understood only within the context of his
account of inferential justification. Here is what he says:

To be justified in believing one proposition P on the basis of another proposition E, one must be
(1) justified in believing that E and (2) justified in believing that E makes probable P. (p. 36)

Given this account of inferential justification, Fumerton presents three reasons for rejecting
infinitism:3

Al. Given clause (1), "[f]inite minds cannot complete an infinitely long chain of reasoning, so, if
all justification were inferential we would have no justification for believing anything." (p. 57)
[emphasis added]

A2. Given clause (2), not only would the chain be infinitely long, the "infinite regresses are
mushrooming out in an infinite number of directions. If finite minds should worry about the
possibility of completing one infinitely long chain of reasoning, they should be downright
depressed about the possibility of completing an infinite number of infinitely long chains of
reasoning." (p. 57) [emphasis added]

A3. Finally, "it is terribly difficult to even imagine how one might continue to appeal to still
more and more beliefs in justifying one's belief that one is in pain now." (p. 59)
There is a common thread of an argument in Al and A2 that implicitly appeals to a principle that
we can call the Completion Requirement: In order for a belief to be justified for someone, that
person must have actually completed the process of reasoning to the belief. Of course, the
infinitist cannot agree to that because to do so would be tantamount to rejecting infinitism. More
importantly, the infinitist should not agree because the Completion Requirement demands more
than what is required to have a justified belief.

To see that, apply the Completion Requirement to a foundationalist conception of justification

coupled with the dispositional account of belief mentioned in (B) above. The result would be that
most, if not all, of our beliefs are not justified. For example, consider what I might believe as I
drive down a road in Montana: I believe that I am in a car, that the steering wheel is working,
that the day is bright and sunny, that I am seeing a mountain, that its name is "Holland Peak,"
that if it were raining, I might not be able to see Holland Peak, that Holland Peak is the highest
mountain I can see at the moment, that the car in front of me belongs to an old friend, etc. I also
believe that 2+2=4 and that pears don't typically grow on orange trees. I also believe that the sun
is a great distance from the earth and that yesterday it rained. There are thousands and thousands
and thousands (if not infinitely many) such beliefs. (Remember that to believe P is to have a
disposition to endorse P in the appropriate circumstances.)I have not carried out the process of
reasoning to any of those beliefs from some foundational beliefs (even if there were foundational
beliefs). In fact, I couldn't have explicitly entertained any significant number of the propositions
believed. There are just too many. But that is not sufficient to show that they are not justified
unless we are willing to embrace skepticism for this reason.

But there is something deeply correct about Fumerton's insistence that S's belief is not justified
merely because there is a justification available to S. In discussing the requirements for a belief's
being justified, Fumerton draws an important distinction between S's merely having a
justification for P and S's belief that P being justified. He claims, correctly I believe, the former
is necessary but not sufficient for the latter.

Fumerton says that in order for P to be justified for S, the belief that P must be appropriately
causally related to the justification. In particular, he says this:

For a belief to be actually justified, we might again want to require more than that the believer
have justification for holding the belief. We might want to require in addition that there be some
legitimate chain of reasoning leading from non-inferentially justified beliefs to the conclusion in
question, the existence of which is causally sustaining the belief. (p. 106)

Earlier he had stated the requirement more forcefully:

The expression "S has a justification for believing P" will be used in such a way that it implies
nothing about the causal role played by that justification in sustaining the belief. The expression
"S's belief that P is justified" will be taken to imply both that S has justification and that S's
justification is playing the appropriate causal role in sustaining the belief. (p. 92)

I think that an infinitist can and should grant the distinction between S's merely having a
justification for the belief that P and the belief that P being justified for S. Suppose that S
believes that P because of wishful thinking, not because of the justification S has. The belief
would not be justified. But note that one can accept the requirement(s) mentioned in the two
passages just cited without accepting the Completion Requirement. What is required in those
passages is that there be a legitimate chain of reasoning which is "causally sustaining" the belief
that P. It is not required that S actually have carried out that process by rehearsing the
justification out loud or sotto voce. That's a good thing, too. For, as mentioned above, the
Completion Requirement would rule out most, if not all, of my beliefs as justified - even on
foundationalist grounds.

Let us grant that, given our current understanding of a belief's being justified, it is required that it
have an appropriate causal history as well as an appropriately structured set of supporting
reasons. Infinitists, foundationalists and coherentists can agree to that. Infinitists differ with
foundationalists (and coherentists) about what is necessary in order for S to have a justification
for a belief. And there might be, as Fumerton claims, some appropriate causal ancestry of a
belief required if it is to be justified. But unless Fumerton can produce an argument to show that
either we never do (or never can) have an infinite, non-repeating set of beliefs that serve as
reasons for and causes of the succeeding beliefs in the chain of reasons, the causal requirement
will not aid in the rejection of infinitism.4

But let us grant, for the sake of argument, that somehow it could be shown - either through
philosophic argument, or perhaps even by cognitive science, that our beliefs do not (or can not)
have the requisite causal history as required by infinitism (or foundationalism or coherentism, for
that matter). What would be the consequences to infinitism (foundationalism or coherentism)? I
think that is very far from clear-cut. The infinitist is claiming that a normatively acceptable set of
reasons must be infinitely long and non-repeating if we are to avoid the pitfalls of
foundationalism( arbitrariness) and coherentism "begging the question). If infinitism correctly
specifies our current concept about what is required for a belief to have the appropriate
normative pedigree and if it were to turn out that beliefs don't (or can't) have the requisite causal
structure, then we have at least three choices: (1) We can revise our concept of the normative
structure of good reasoning or (2) we can adopt a form of Pyrrhonism (withholding assent to any
proposition requiring a justification) or (3) we can accept an antinomy. But it would not follow
that the normative constraints were incorrectly described - unless, perhaps, epistemic oughts
imply epistemic cans.5

Let us turn to A2. This looks to be a stronger objection since it claims that the infinite chains
seem to be "mushrooming." Since this objection depends upon the Completion Requirement, it
already can be set aside. Nevertheless, clause (2) in the requirements for inferential justification -
that S must be justified in believing that E makes probable P - is interesting in itself and deserves
comment. Bluntly, that requirement strikes me as just too strong.

First, it requires that some propositions of the form "E makes probable P" be non-inferentially
justified if we are to have any justified beliefs (see p. 189). Whatever one means by one
proposition's making another one probable, I don't share Fumerton's optimism that such
propositions can be non-inferentially justified. Indeed, it seems to me that those particular
propositions are always open to this question: What about E makes it the case that whenever E, P
becomes probable? Something's appearing to be a car, for example, might make it probable that
it is a car. But isn't that true, if it is true, because there is something true in general about how
things appear and how they are? And wouldn't that general truth be a reason for thinking that the
particular instantiation of it is true? So, I don't think that those propositions are non-inferentially
justified. Indeed, they seem to be just the types of propositions that call out for inferential
justification.

Second, it seems to me that to be justified in believing one proposition, P, on the basis of another
proposition, E, all that is required, ceteris paribus, is that (1) one be justified in believing that E
and (2)' that it be true that E makes probable P. It is not required that one be justified in believing
(2)'. That requirement would force anyone having a justified belief to be an epistemologist, and a
darn good one at that - since the beliefs about what makes propositions probable have to be
justified. That's just too much to ask!

But what is Fumerton's defense of clause (2)? I think the best defense occurs in his diagnosis of
the astrologer case. Here, in part, is what he says:

...if I am talking to an astrologer who infers from the present alignment of planets that there will
be prosperity this year, I am perfectly entitled to challenge the reasonability of this conclusion by
challenging the reasonability of the astrologer's thinking that there is some connection between
the two states of affairs.... We will reject our astrologer's beliefs as unjustified for the reason that
the astrologer has no reason to believe that there is a probabilistic connection between
astrological evidence and astrological predictions. (pp. 86-87)

To the contrary, I think it is just as plausible to suggest that we will reject the astrologer's beliefs
as unjustified for the reason that there just is no probabilistic connection between astrological
evidence and astrological predictions. If there were such a connection, then, ceteris paribus, I
think we would hold that the predictions were justified but that the astrologer wasn't justified in
believing that they were justified.

Fumerton does have one other defense of clause (2). It is connected with the way in which he
envisions the nature of skeptical arguments. Here is what he says, in part:

If one examines classic arguments for strong local skepticism, one can discover, I think, a
recurring pattern. First, the skeptic indicates the class of propositions under skeptical attack.
Then the skeptic attempts to exhaustively characterize the most plausible candidate for
something that could conceivably justify, or make rational, belief in this kind of proposition.
Next the skeptic attempts to drive a logical wedge between the available justification and the
proposition it is supposed to justify.... The strong skeptic goes on to argue that the logical gap
cannot be bridged using any legitimate nondeductive reasoning. (p. 31)

He then discusses two examples: sensations/physical world; memory/past events. Presumably, a

conjunction of various propositions about one's particular sensations is "E" and a proposition
about the physical world is "P." He writes that clause (2) results from "reflection on the above
examples of skeptical reasoning." (p. 36) In particular, he says:
Thus the argument for strong skepticism with respect to the physical world relies on clause 2 by
insisting that a belief in physical objects inferred from what we know about the character of our
sensations is rational only if we have some reason to suppose that there is a connection between
the occurrence of certain sensations and the existence of certain objects. (p. 36)

Implicitly this argument for clause (2) is this: Since skeptical arguments appeal to clause (2), if
one is to give skepticism a fair hearing, one must accept it.6 Of course, there is the contrary
intuition that defenders of common sense will have if Fumerton is right that skepticism depends
upon clause (2). What better reason could one have for rejecting it since it is responsible for
making skepticism seem plausible!

I will not press that response here, for I, too, believe in giving skepticism a fair hearing. What I
do not see, though, is that the case has been made for claiming that skepticism depends
essentially upon clause (2). It seems just as plausible to hold that skeptics claim that it is not
rational to believe in physical objects on the basis of sensations because the "logical gap" cannot
be bridged. But that can be understood as a challenge to the truth of the claim that beliefs about
sensations make beliefs about physical objects probable. Thus it seems to me that it is just as
plausible to hold that what is required to give skepticism a fair hearing is not clause (2), but
clause (2)': E makes probable P.

Let me now turn very briefly to A3. The claim is that it is difficult to imagine how one might
continue to appeal to still more and more beliefs in justifying one's belief that one is in pain now.
I don't think it is difficult at all. After someone says that they believe that they are in pain (and if
any one ever says such things, it must be in a philosophical context), we can perfectly well ask
(remember, we are in a philosophical context): What is it about such beliefs that gives you such
confidence in their truth? What makes such beliefs (almost) always true?

This response to A3 is a key to appreciating the initial plausibility of infinitism. Suppose that
some proposition, say F, is offered as a putative foundational one. Echoing Pyrrhonism, the
infinitist will claim: Either F has some characteristic which makes it such that it is (highly likely
to be) true or it doesn't. If it does, then the possession of that characteristic can be and should be
offered as a reason for thinking that it is true - and the regress continues. If it doesn't, then there
is nothing that distinguishes it from non-foundational propositions and it becomes arbitrary to
treat it as foundational.

A careful look at that argument for infinitism would take us too far afield. My point here has
been that stronger arguments against infinitism are needed before it is rejected.7

1 The term "infinitism" is not original with me. Paul Moser speaks of "epistemic infinitism" in
his paper "A Defense of Epistemic Intuitionism" in Metaphilosophy, vol. 15, no. 3/4 (July/Oct.,
1984) pp. 196-209. Also, John Post in The Faces of Existence, Ithaca: Cornell University Press,
1987, refers to a position similar to the one under discussion here as the "infinitist's claim", p. 91.

2 All page references in the text are to Richard Fumerton's book, Metaepistemology and
Skepticism, Boston and London: Rowman & Littlefield Publishers, 1995.
3 In fairness, it should be pointed out that these arguments are not put forth by Fumerton with
equal enthusiasm. He says of A3 that it is "hardly an argument" but that it is "another reason" for
rejecting what I have called infinitism. (See p. 59.)

4 It is important to note that infinitism can appeal to the fact that we seem to have orders of
dispositions - dispositions to have other dispositions. So not all of the members of the infinite set
of beliefs need be formed dispositions. It would be sufficient for us to have the disposition to
form a belief (a disposition) under the appropriate circumstances if the second (or even higher)
order dispositions could play the appropriate causal role in sustaining the justified belief.

5 It seems highly dubious that epistemic oughts imply epistemic cans. Would it not be possible
for it to be true that the rules of inference that are most truth conducive are such that we are not
"wired" to employ them? If so, there is a perfectly good sense in which we ought to reason in
some way that we can't.

6 Fumerton also argues that the skeptic employs clause (1) of the definition of inferential
justification. I ignored that claim since I am not objecting to clause (1).

7 There is one other objection to infinitism that I did not find in Fumerton's book but which
seems to jump from the pages. (I might have missed it.) It was suggested to me by one of my
graduate students - Troy Cross. Consider the expression "e makes probable p". If the relevant
notion of "making probable" is some sort of statistical one then the longer the chains, the less
likely the conclusion. If the chains were infinitely long, the conclusion would be so improbable
that it would not be reasonable to accept it.

There are three ways around that worry. The first is that there is an infinite number of probability
gradations available for the justifying propositions given any required probability level of the
putatively justified proposition. The second is that it is the proposition, itself, that is located in
the chain rather than a proposition with a probability assigned. The third is to simply reject the
reading of "makes probable" in frequency terms and treat "p is probable" as roughly synonymous
with "p is epistemically acceptable and can, ceteris paribus, be used to make other propositions
acceptable."
Replies to My Three Critics, Fumerton
Metaepistemology and Skepticism
Philosophy and Phenomenological Research, Vol. 58, No. 4 (Dec., 1998), pp. 927-937
International Phenomenological Society

Cohen and Klein on the Principle of Inferential Justification:

I claim that (PIJ) is analytic. Both Cohen and Peter Klein claim that the requirement that we be
aware of probabilistic connections in order to have inferential justification is too strong. Surely,
Cohen would argue, the principle is no more plausible than is the falsity of skepticism and if
forced to choose between (PIJ) and skepticism, we should jettison (PIJ). Klein also argues that
the second clause of (PIJ) is far too strong a requirement for justified belief. But I argue in the
book that there are many contexts in which we do seem to unequivocally insist on awareness of
probabilistic connections before we allow for reasonable belief based on inference. Were a
subject unaware of a probabilistic connection between furrowed brows and criminality, litmus
paper's turning red in a solution and the solution being acidic, or the character of tea leaves and
predictions of the future, we would never allow the subject to justifiably believe propositions
about criminality, acidity, and the future, based on, respectively, propositions describing
furrowed brows, red litmus paper, and tea leaves. How can we in good conscience jettison the
principle of inferential justification just because the skeptic is successfully using our own
standards of rationality to embarrass us?

One might respond that the "inferences" described above are not fundamental inferences
reflecting epistemic principles, at all.3 We may casually speak of inferences to the conclusion
that a solution is acidic from a premise describing the color of litmus paper it contains, but, one
might insist, the "argument" in question is highly enthymematic. We only reach the conclusion
against a complex set of background information, including the information that usually or
always acidic solutions turn litmus paper red. The information about the correlation between red
litmus paper and acid solutions functions as a premise in the fleshed out characterization of the
reasoning we employ. And even hard core externalists will allow that if one employs a premise
to reach a conclusion, the premise must be justified if the conclusion is to be justified based on
that premise. To argue for inferential internalism, one must be sure one is focussing on genuine,
fundamental inferences, and when one does it may not be nearly so obvious that we do insist on
anything like awareness of probabilistic connections.

At this point, however, the inferential externalist owes us some kind of account of how one
would recognize a fundamental epistemic principle. One can trot out examples, of course,
beginning with everyone's favorites, the rules of deductive logic. But deductive reasoning by
itself is never going to save us from skepticism with respect to the vast majority of what we
believe. How shall we recognize fundamental non-deductive principles of reasoning? The
difficulty is that for most externalists and naturalists, epistemic principles, principles licensing an
inference from beliefs of one sort to beliefs of another are almost certainly going to be contingent
empirical truths. The whole point of a reliabilist's understanding of justified belief is that we are
supposed to secure a connection between justified belief and true belief. But it will always be a
contingent feature of the world that relative to an environment X, true beliefs of kind A, taken as
input and processed by the relevant software of the brain, result in true beliefs of kind B. The
connection between input beliefs and output beliefs sanctioned by an inferential rule is
essentially a nomological connection. As such it looks suspiciously like the connection between
litmus paper's turning red in a solution and the solution's being acidic. As a matter of contingent
fact, the occurrence of the one state of affairs might typically signal the occurrence of the other.
But this would be only a causal connection which one might suppose we need to discover a
posteriori before we would be entitled to employ the litmus test in arriving at conclusions. By
the same token, probabilistic connections as understood by externalists and naturalists are
contingent nomological connections which it seems we should also need to discover a posteriori
before we could be viewed as rational in implicitly following the relevant rule sanctioning the
inference in question. Again, how are we supposed to tell when a contingently reliable inference
is genuinely fundamental in a way which obviates any need for awareness of the reliability of the
inference?

Notice that the inferential internalist is likely to have a principled answer to the question of how
to distinguish fundamental (or genuine) epistemic principles from derivative (or spurious)
epistemic principles. The inferential internalist is likely to suppose that fundamental epistemic
principles, deductive and non-deductive, are necessary truths knowable a priori. At least such a
view is the inferential internalist's only hope of ending a potentially vicious regress when it
comes to inferential justification. There must be some probabilistic connections, according to the
inferential internalist, which one knows without inference. Otherwise every inferentially justified
belief would lead one to a regress of inferences designed to justify acceptance of inferential
connections.

Setting aside the above problem, however, let's look at an obvious candidate for a fundamental
inference, a deductive inference. In the case of someone S who infers P from E where E logically
entails P, is the inferential internalist right in maintaining that in order for S to justifiably believe
P on the basis of E, S must be aware (have at least a justified belief) that E entails P (that the
inference in question is legitimate)? Frankly, the answer still seems to me obviously yes. We can
easily imagine someone who is caused to believe P as a result of believing E where E does in
fact entail P but where the entailment is far too complicated for S to understand. Unless S "sees"
that P follows from E, would we allow that the inference in question could generate justified
belief?

Now to be sure a great deal hinges on what is involved in someone's believing one proposition
on the basis of another. Certainly, if S believes P after he believes E because he read in the tea
leaves that if E is true P is true, then we don't really have a case of S's inferring P from E at all. S
is inferring P via a chain of inferences - one from propositions describing tea leaves to a
proposition asserting the conditional, if E then P; the other, inferring P from E and if E then P.
The inferential externalist can treat this situation just as he treated the astrological inference or
the litmus paper inference. The inferential internalist, the argument goes, has misidentified the
inference in question, and in so doing has misidentified the source of the intuition that a genuine
inference from E to P (where E entails P) requires awareness of the connection between E and P.
But even if the inferential internalist must be careful in describing the relevant hypothetical
situation, it still seems possible to describe a case in which S is caused by his justified belief in E
alone to believe P, where E does entail P, but where S is completely unaware of the entailment.
And I would still suggest that in any philosophically interesting sense of inferential justification,
S would still not be justified in believing P on the basis of E.

The inferential externalist has one last move. He can argue that believing P on the basis of E
requires more than simply being caused to believe P by a belief that E. The hypothetical situation
described above might still fail to constitute a situation in which S genuinely infers P from E
where E entails P but S has no justification for believing P on the basis of E. There is, I think,
some plausibility to the claim that inference involves more than mere causal connection between
beliefs. But the source of this intuition will, I think, provide little solace for the inferential
externalist. In fact, I suspect that we may not concede genuine inference unless there is veridical
or nonveridical "perception" of a connection between the proposition from which P is inferred
and P. But this perception will be just what the inferential internalist claims is the justification
for believing in a connection that is necessary for inferentially justified belief.4

Peter Klein and "Infinitism":

In defending a version of foundationalism in the book I discuss regress arguments for the view.
In his thought-provoking comments, Klein argues that there is no vicious regress requiring one to
accept the existence of noninferentially justified beliefs. One can plausibly and consistently
claim that every justified belief owes its justification to the having of other justified beliefs, and
one can do so without embracing the kind of circular reasoning inherent in coherence theories of
justification. Specifically, there is no reason to deny that one can have an infinite number of
beliefs such that the having of each is causally sustained by the having of other justified beliefs.
If my belief that P is to be justified, it may be true that it must be causally sustained by an infinite
conjunction of beliefs, but there is no reason to suppose that one cannot have an infinite number
of such beliefs and no reason to suppose that they cannot stand in the appropriate causal
relationship to my belief that P. Furthermore, even if one were to discover a priori or empirical
reasons to suppose that our belief system doesn't have the appropriate structure to yield justified
belief according to "infinitism", why should we reject it is an account of justified belief? Perhaps
the correct account of the structure of justification will lead inexorably to skepticism.

If we remind ourselves that beliefs need not be occurrent, it should be obvious that Klein is right
in suggesting that we not only can, but do have an infinite number of justified beliefs. I
suggested in the book that most often we can trace the infinite number of justified beliefs we
have to our belief in general propositions which entail infinitely many other propositions where
we have dispositional awareness of the entailment. I justifiably believe that all triangles have
three sides and also believe that if I were to draw a triangle on a wall it would have three sides, if
I were to draw a triangle on paper it would have three sides, if I were to draw..., and so on for the
indefinitely many ways I can imagine drawing a triangle.

I'm also inclined to agree with Klein that we shouldn't reject his account of justification even if it
should lead to skepticism. As an aside on the subject of skepticism, however, I do think he is
wrong to suggest that classical skeptical arguments should not be construed as relying on clause
(2) of (PIJ) - that one could as easily construe the skeptic as simply arguing with respect to some
problematic inference from E to P, that E does not make probable P. The classic skeptic often
wants to base a positive assertion that we lack justification for believing P on the basis of E on
the claim that we simply have no reason to believe or deny that E makes probable P.5 The lack of
reason to believe that there is a probabilistic connection between E and P is sufficient to deny
someone an inferentially justified belief in P on the basis of E.

Klein notes that I am somewhat puzzled by the idea that one could construe my belief that I'm in
pain as being justified by a structure of justified beliefs infinitely complex. The idea that one
would move to some more general belief about propositions describing beliefs about pains and
pains in order to find the source of my justification for believing that I'm in pain (when I am)
seems to me about as plausible as the radical behaviorist committed to the view that the way one
discovers one's pain is to notice one's grimaces and screams. Surely the knowledge that we are in
pain when we are in pain is far more secure than knowledge of more abstract and esoteric
philosophical claims about the character of beliefs directed at mental states.

But I have not yet touched on my main concern with Klein's infinitism. In referring to my
discussion of regress arguments for foundationalism, Klein quotes passages from Chapter 2 of
the book. In chapter 3, however, I stress what I call a conceptual regress argument for
foundationalism. The gist of the argument is that we need a concept of noninferential
justification with which to define in a non-circular way justified belief. In discussing infinitism,
Klein doesn't tell us what makes a belief justified. He does tell us that the justification of one
belief always involves the having of other (perhaps only dispositional) beliefs, but unless those
other beliefs are themselves justified beliefs the account is unsatisfactory. I presume the infinitist
doesn't want to allow that my belief that P can be justified solely by virtue of the fact that I can
produce an argument for P, and an argument for the soundness of the argument for P, and an
argument for the soundness of the argument for the soundness of P, and so on ad infinitum.
Suppose I believe P and support my belief with the argument that there is a God1 who says that
P is true and whatever God1 says is true. My reason for thinking that this is a sound argument is
that there is a God2 who says that the God1 argument is sound and whatever God2 says is true.
My reason for thinking that the God2 argument is sound is that there is a God3 who says..., and
off we go. I take it that Klein would agree that though I'm having a great deal of fun producing
God arguments after God arguments, none of them result in any justified beliefs because I don't
have any justification for believing the premises of any of the arguments. But if the infinitist
understands inferential justification in such a way that the inference from E to P yields justified
belief in P only if the belief that E is justified, the infinitist's story doesn't shed any light on the
nature of justification. The mere existence of an infinite number of justified beliefs doesn't avoid
the problem of conceptual regress.

It's true that we must not fear a vicious regress every time the truth conditions for a proposition
involve classes with an infinite number of members. Zero Mostel's musical injunction that
everyone ought to have a maid doesn't violate the principle that ought implies can for we could
live in a universe containing an infinite number of maids. The fact that P's truth entails that it is
true that P is true and true that it is true that P is true and so on, ad infinitum is metaphysically
unproblematic. But some regresses are vicious. Consider an analogy I offer in the book. Suppose
a philosopher tells us that nothing is intrinsically good - that all good things are made good by
producing something else that it is good. Furthermore, we are assured, we mustn't worry about a
regress here, as there is no reason to suppose that there are only a finite number of good things.
With an infinite number of good things, there will also be a another good thing to make good
everything that is good. Would the existence of an infinite number of good things obviate the
need to recognize a distinction between intrinsic and instrumental goodness? No it wouldn't. We
need the concept of intrinsic goodness in order to understand the concept of instrumental
goodness. We could form no idea of instrumental goodness if the only way to understand
something's being good was in terms of its leading to something else that is good. In our search
for the conceptual source of goodness we would always be led to an idea of goodness that was
itself defined in terms of an undefined concept of goodness.

In the same way, the infinitist never gives us a non-circular analysis of justification. As an
analysis of justification it resembles a recursive definition without a base clause. Classical
foundationalists and some externalists (process reliabilists, for example) actually agree on this
one point. Unless one can define a concept of justified belief that does not invoke the concept of
justified belief, one will have no understanding of the way in which justified "output" depends on
justified "input."

3 In this discussion I am greatly indebted to Mike Huemer who in correspondence raised a

number of thoughtful objections to (PIJ) and helped me clarify my thought on the subject.

4 I leave open the possibility, here, that a non-veridical perception can yield justified belief in a
probabilistic or logical connection (just as hallucinatory experience can yield justified belief
about the external world). Again, I'm indebted to Mike Huemer at Rutgers for many e-mail
"conversations" which helped me to clarify my thoughts about this matter.

5 This is particularly true when the skeptic is understanding the probabilistic connection as a
causal connection or a probabilistic connection understood in terms of empirical frequencies.
Skepticism Avoided
Patrick Hawley

1.2 The Agrippan argument

One venerable argument for skepticism about justification goes like this. Suppose that one of my
beliefs is justified. Since my belief is justified, it is justified by something. Now, if what justifies
my belief is not itself justified, then my belief is not justified. So, what justifies my belief is
justified. But if what justifies my belief is justified, then it is justified by something. And so on.
This process either moves in a circle, or goes on infinitely. Since neither a circle nor an infinite
chain suffices to justify my belief, my belief is not justified. Thus we have a reductio of the claim
that my belief is justified. This argument can, at any time, be given by anyone, about any belief.
So, no belief can ever be justified.1

I suspect that few philosophers are driven to skepticism about justification by the Agrippan
argument on its own, for there are many ways to try to avoid the conclusion. According to
coherentism, a circular chain of justification can be legitimate. A foundationalist thinks that each
chain eventually ends. Infinitism says that an infinite chain of justification can be acceptable.
These views about the structure of justification appear in various forms and combinations.

Although few philosophers are swayed by the Agrippan argument, many philosophers rely on
parts of the Agrippan argument. The best argument for coherentism relies on the Agrippan
argument (minus the premise that circular chains of justification are always illegitimate). The
best argument for foundationalism relies on the Agrippan premise that neither circular nor
infinite chains suffice for justification. Thus, plenty of non-skeptics should be interested in the
Agrippan argument. An examination of the Agrippan argument might shed light on discussions
about the structure of justification…

1.3 The Agrippan premises

The Agrippan argument, as just presented, is a reductio of the hypothesis that some belief is
justified, relying on the following three premises:2

Premise I: For any A, if A is justified, then A is justified by something.

Premise II: For any A and B, if A is justified, and A is justified by B, then B is justified.
Premise III: Neither circular nor infinite chains suffice to confer justification…

1.3.2 Premise III Premise III: Neither circular nor infinite chains suffice to confer justification.
Two analogies will reveal the plausibility of Premise III. Fully defending Premise III would
require arguing against various forms of coherentism. Although important, this is not my present
task.

First Analogy
Think of a belief's being justified by something as like a book's being physically supported by
something. For example, my copy of The Great Chain of Being is supported by a shelf.
Moreover, my book is part of a structure of support: my book is supported by the shelf; the shelf
is supported by the wall; the wall is supported by the foundation of the building; and so on.

Now, plausibly, if a book is supported, then it is part of an acceptable support structure. And,
clearly, some structures of support are not acceptable. Self-supporting structures, for example,
are not acceptable: the book cannot be physically supported by itself. In addition - if the support
relation is transitive - no circular structures are acceptable. You see the point: if, as seems
plausible, "A is justified by B" describes a relation between A and B that is anti-symmetric and
transitive, then A's being part of a circular chain does not suffice to make A justified.3

We can rule out infinite structures of physical support with empirical premises. Infinite structures
of physical support require an infinite amount of matter. If there is a finite amount of matter in
the universe, then infinite structures of physical support are impossible. Similarly we may look
for empirical premises to rule out infinite chains of justification. Since humans are finite, they
cannot understand infinite chains of justification. So infinite chains do not suffice to make
human beliefs justified.4

An analogy is not an argument. Still, if being justified by something is relevantly like being
physically supported by something, then we have reason to accept Premise III.

Second Analogy
Think of a belief's being justified as like a person's catching a cold. A healthy person can be
infected by coming into contact with a person who is already infected. Consider, for simplicity, a
cold caused by a virus that cannot survive outside a human body. And suppose that a person can
be infected with this virus only once in his lifetime.

Now if Arnold was infected by Ziggy, and Ziggy was infected by Leonard, then it does not
follow that Arnold was infected by Leonard. At least, it is not obvious that the relation of being
infected by is transitive. Still epidemiological circles are ruled out because we are assuming that
you can only be infected by this virus once in your lifetime. So it cannot be that you infected
yourself; and it cannot be that Arnold was infected by Ziggy, and Ziggy was infected by Arnold.
Moreover, since there have only been finitely many people existing in the world, a chain of
infection cannot trace back through an infinite number of different people.

According to this analogy, A gets justified by being "infected" by some B that is already
"infected". If A is justified by B, then A is justified only if B is already justified. A can't "infect"
itself with justification, nor can A get "infected" by being part of a circular chain. If there are
only finitely many justifiable things then we can rule out infinite chains too.

If being justified is relevantly like catching a cold, then it seems that we have some reason to
accept Premise III.

So, it appears that Premise III, as well as Premise I and Premise II, are plausible…

1 Of ancient lineage, such skeptical reasoning appears in Sextus Empiricus, Outlines of

Pyrrhonism and Against the Mathematicians. Diogenes Laertius (IX 88) attributes it to one
Agrippa, a philosopher otherwise unmentioned in ancient manuscripts. J. Barnes discusses the
Agrippan reasoning helpfully in The Toils of Scepticism.

2 In presenting the premises, I make no mention of belief. In Premises I and II, the quantifiers
could range over beliefs, collections of beliefs, or something else. For present purposes, assume
that the quantifiers range over anything epistemically relevant. (Although I am discussing the
argument as if the conclusion is that no belief can be justified, it should be noted that similar
reasoning may be used to argue that no action, nor person, nor anything else can be justified.)

3 J. Barnes makes this and other helpful points in The Toils of Skepticism.

4 A thought like this goes back to at least Aristotle, Posterior Analytics 1.3. Much more ought to
be said to turn this thought into an argument. Recently, Peter Klein has defended the viability of
infinite chains. See [38].

Philosophy and Phenomenological Research Vol. LXVI, No. 3, May 2003

When Infinite Regresses Are Not Vicious*
PETER KLEIN
Rutgers University
When Infinite Regresses are Not Vicious

Infinite regresses are notoriously suspect. Indeed, even the label “infinite regress” as opposed to
“infinite series” suggests that there might be problems lurking. But setting that verbal uneasiness
aside, there remain many reasons offered for distrusting regresses. I want to focus on one of them
here, namely what Carl Gillett calls “The Structural Objection” to my proposal that in order for a
person, S, to be justified in believing a proposition, p, there must be an infinite set of
propositions available to S that can be arranged in a nonrepeating series such that the first
member, r1 is a reason for p, and the second member, r2 is a reason for r1, and r3 is a reason for
r2, etc., and no r1 repeats in the series.1 I argued that this partial account of justification, what I
call “infinitism,” offers a better solution than either foundationalism or coherentism to the
“regress of reasons” problem that is posed in The Outlines of Pyrrhonism by Sextus Empiricus.2
The reasons I gave were based upon epistemic considerations, and Gillett doesn’t challenge
those. Rather, the Structural Objection is based upon “metaphysical worries.” I think this type of
worry is related to some other objections to infinitism developed by Ernest Sosa and Richard
Fumerton. In this paper I will briefly address those as well.

I will argue for two main points. First, the regress imbedded in infinitism need not be subject to
the Structural Objection; and second, the Structural Objection does not pose a real problem for
any regress. I will not be arguing for the correctness of my proposal directly. That is, as will
become apparent soon, my proposal rests on two principles of reasoning which together entail
infinitism and I will not present my arguments for those principles here.3 The purpose of this
paper is to show that the objections raised by Gillett do not pose a grave problem for infinitism.

So, what is my proposal and what is the Structural Objection? Gillett cites the two epistemic
principles (referred to above) which, if correct entail infinitism. The principles are:

PAA: For all x, if a person S, has a justification for x, then there is some reason, r,1 available to
S for x; and there is some reason, r2, available to S for r1, etc. (HKIRR, p. 299)

PAC: For all x, if a person, S, has a justification for x, then for all y, if y is in the evidential
ancestry of x for S, then x is not in the evidential ancestry of y for S. (HKIRR, p. 298)

Gillett argues that these two principles are specific instances of two, more general principles
which together lead to a vicious “In Virtue of’ regress, or more simply, what he calls an “IV
Regress.” Here are those general principles:

(IVA) For all entities, an entity x has the property H only in virtue of, amongst other possible
necessary conditions, some entity y having the property H; and y is H only in virtue of some
entity z being H, etc.

(IVB) For all entities, if entity x has the property H only in virtue of, amongst other possible
necessary conditions, some entity y having the property H, then y does not have H in virtue of x
being H.
Now, the alleged difficulty faced by all IV Regresses is this:

The question consequently arises how it could ever come to pass that any member of the chain
has the property H? For it appears that whatever entity, or structure of entities, is added to the
chain of prior entities, consistent with the governing principles (IVA) and (IVB), this addition
will not be sufficient for the dependent property H to feed back to any member of the regress.
Adding more entities one by one, or even as structured groups, will not suffice, for once again
these additional entities will only have H in virtue of some still further entity that is H. Thus, the
objection ultimately concludes, there is no entity, or structure of entities, that can be added to an
1V Regress, consistent with its governing principles, that will suffice for any of its dependent
properties to feed back to any members of the chain. Let us call this the ‘Structural Objection,’
since it is the argument that the structure of an IV Regress means that it cannot produce any of its
dependent properties. (p. 713)4

Gillett is relying on our understanding of what “in virtue of’ means in this context and, as
Aristotle remarked, that phrase is multiply ambiguous. Aristotle did point out that it is used in the
same senses as “cause” (1022a14). That isn’t much help, given the many senses of “cause,” but it
does seem that in this context “in virtue of’ is, at base, probably a phrase meant to be roughly
synonymous with “because of’ and I take the dependence here to at least involve explanatory
dependence. True, Gillett calls this worry a “metaphysical worry” and perhaps such a worry is
not strictly equivalent to an explanatory worry. But it does seem that whatever could be meant by
a “metaphysical worry,” if it is true (as I think would be intended here) that if x is metaphysically
dependent on y, then y provides an explanation (at least in part) of x, then showing that the
regresses envisioned by infinitism do not involve an explanatory regress would be sufficient to
show that they do not involve a metaphysical regress. Thus, I will take “x has H depends upon y
having H’ to be true only if y having H is a necessary part of an explanation of x having H.

The immediate question, then, is this: Are PAC and PAA instances of (IVA) and (IVB)? My
answer is “No.”

Imagine a game we can call “jigscotch” (a combination of a jigsaw puzzle with hopscotch).
There is a pile of cardboard pieces about the size of hopscotch squares that look like jigsaw
pieces, and there is already one such piece lying separately on the ground. You are allowed to
take one piece from the pile and fit it into a piece that is already lying on the ground. You can
then get another piece from the pile and fit it into the second piece. You win only if there is a pile
of such pieces that would never run out (maybe one would be added to the pile every time you
took one away) and you are able to locate and fit an appropriate piece from the pile into a piece
that is already located on the ground.5 Of course, you can never use all of the pieces - indeed, you
can win without actually using any piece. What is required is that the pieces are there for you to
use and you have the ability to use them. Thus, this principle (partially) describes the game:

G1: S wins the game of jigscotch starting with p, only if there is some piece, p1, usable by S that
fits p, and there is some piece, p2, usable by S that fits p1, etc.
There is an infinite regress here similar to the one that Gillett thinks is an IV Regress imbedded
in infinitism, namely: If S wins starting with p, then S wins starting with p1; if S wins starting
with p1, then S wins starting with p2, etc. The analogous regress entailed by infinitism is: If S is
justified in believing p, then S is justified in believing r1; if S is justified in believing r1, then S is
justified in believing r2, etc. But G1 is not an instance of (IVA) because, although S wins starting
with p only if S wins starting with p1, etc., S doesn’t win starting with p in virtue of winning
starting with p1. S wins in virtue of, amongst other things, the conditions specified in G1
obtaining. That they obtain explains, or partially explains, S’s winning.6 I think the parallel with
S’s being justified is clear.

Nevertheless, the drawback with using analogies in order to appeal to intuitions is that if the
analogs are very close, as I intend these to be, intuitions are likely to be identical in the two
cases.7 Indeed, someone might think that being justified in believing p consists (in part) in being
justified in believing r1, and being justified in believing r1, consists in (in part) in being justified
in believing r2, just as winning in jigscotch starting with p consists (in part) in winning starting
with p1, and winning starting with p1 consists (in part) in winning starting with p2, etc. And it
might be thought that this is enough to make both of these regresses examples of IV Regresses.

To see that this is not sufficient, consider a somewhat less analogous case. Consider a line AB
and some subsegment of it, says. Now, s is a subsegment of AB only if there is another
subsegment of s, say s1, that is not identical to s (or AB), and there is some subsegment, s2. etc.
In addition, any subsegment consists (in part) of its own subsegments, but it is not a subsegment
in virtue of its having subsegments. Rather, each is a subsegment in virtue of being a segment
between the endpoints of the given segment that is not equivalent to the given segment. That
explains why it is a subsegment. My point is that .necessary conditions, even those that entail the
existence of a constituent, are not necessarily part of explanatory or in-virtue-of conditions. In
other words, “A holds only if B holds” can be true without “A holds in virtue of B holding”
being true, even when B is a constituent of A.

Return to infinitism. What makes the proposition r1 a reason for p for S is not that there is
another proposition, r2, that is a reason for r1, for S. Rather, as Gillett notes, my suggestion is
that what makes x a reason for y for S is that x satisfies some person-neutral features (what I
called “objective availability”) and some person-centered features (what I called “subjective
availability”). Infinitism is not committed to any particular view about those conditions of
availability. For example, it could be that x is objectively available as a reason for y just in case x
is true and such that the objective probability of y given x is sufficiently high; or it could be that
x is objectively available as a reason for y just in case the relevant epistemic community would
(in the appropriate circumstances) accept the argument “x, therefore y.” There are many other
possible theories of what makes a proposition an objectively available reason.

A proposition, x, is subjectively available to S as a reason for y just in case x is “properly hooked

up” to some of S’s current mental states (including dispositional ones). Again, there are many
ways to specify the appropriate range of S’s mental states and many ways to characterize what it
means to be properly hooked up to them. Infinitism per se need not take a stand on those issues.
What is crucial to note, here, is that a given proposition, x, is not an available reason in virtue of
another proposition, y, being an available reason. Each is an available reason because the
objective and subjective conditions of availability are fulfilled. So, being an available reason is
not a troublesome dependent property.

But Gillett does not claim that it is. What he says is that being a non-arbitrary reason is the
troublesome dependent property, and I think he means by “non-arbitrary reason” exactly what I
meant by “justified.” It is true, as mentioned above, that if some proposition p is justified (or is a
non-arbitrary reason) for S, there is another proposition, r1, that is justified (or is a non-arbitrary
reason) for S, and another, r2, that is justified (or a non-arbitrary reason) for S, etc. It is not
necessary for S to complete (or even begin!) the process of giving reasons for p in order for p to
be justified for S - just as S did not have to complete (or even begin) the process of piecing
together the jigscotch pieces to win the game. Being justified or a non-arbitrary reason is not a
troublesome dependent property because a proposition being justified (or being a non-arbitrary
reason) does not arise in virtue of another proposition being justified (or being a non-arbitrary
reason) - a proposition is justified for S in virtue of being a member of a set of propositions each
member having the required properties specified in PAA and PAC. The joint satisfaction of PAA
and PAC partially explains why a proposition, p, is justified.8 Thus, I do not think that the
infinite regress envisioned by infinitism is an instance of an IV Regress.

In order to make that more clear, I would like to address two related criticisms of infinitism
developed by Ernest Sosa and Richard Fumerton. Sosa says:

...the main reason for accepting formal foundationalism [as against infinitism] ... is the very
plausible idea that epistemic justification is subject to the supervenience that characterizes
nonnative and evaluative properties generally.... Epistemic justification is supervenient. The
justification of a belief supervenes on such properties of it as its content and its basis (if any) in
perception, memory or inference.9

Fumerton says:

The mere existence of an infinite number of justified beliefs doesn’t avoid the problem of a
conceptual regress.... the infinitist never gives us a non-circular analysis of justification. As an
analysis of justification it resembles a recursive definition without a base clause.... Unless one
can define a concept of justified belief that does not invoke the concept of justified belief, one
will have no understanding of the way in which justified “output” depends on justified “input.”10

Briefly put, these criticisms are:

Sosa: Infinitism is incompatible with normative properties supervening on non-normative

properties; and that is a drawback of infinitism.11

Fumerton: Infinitism involves a conceptual opacity, since “justification” is defined in such a way
that it is used, and never eliminated, in the definiens as well as the definiendum.

Both criticisms focus on supposed weaknesses in the explanatory power of infinitism. That is the
similarity between these criticisms and the one proposed by Gillett. But they, too, fail to
appreciate the explanatory resources available to infinitism. It is true that: 1) a proposition is
justified only if it is a member of an infinite series of non-repeating propositions related to each
other by the property of “being a subjectively and objectively available reason for” and 2) each
member of the series is such that it is justified only if some other member is justified. But 1) and
2) do not make the concept of justification conceptually opaque or preclude epistemic
normativity from supervening on the non-normative because the concept of justification (or
being a non-arbitrary reason) is characterized by appealing to the concept of being an available
reason, and that, in turn, can be characterized in ways that neither makes use of “being an
available reason” nor any other normative notion (as illustrated in the options mentioned above
for defining objective and subjective availability). But let me hasten to add that such
supervenience is not required by infinitism because it is compatible with the view that “being an
objectively available reason” is a basic normative concept which is “he floating” (or that “being
an objectively available reason” is defined in terms of a free floating normative property) and in
no way supervenes on the nonnormative.12 The points here are that (1) infinitism is sufficiently
flexible to allow room for the type of supervenience that Sosa thinks is a desideratum of a good
theory of justification’13 and (2) pace Fumerton, it does not lead to conceptual opacity since
justification need not be understood, defined or explained in terms of justification.

Before turning to my second main point, namely that IV Regresses are not necessarily vicious
anyway, it must be noted that Gillett has (correctly) anticipated that someone defending
infinitism could attempt to respond to the Structural Objection in just the way that I have -
namely, that the series need not be composed of members that pass some dependent property
from one to the other. He calls such a view “Modest” infinitism and claims that it is not an
epistemic view that is a “substantive option” to coherentism and foundationalism. The reason he
gives is that the “. . .Modest picture holds simply that a[n] infinite regress accompanies whatever
conditions are involved in the metaphysical ‘guts’ of justification.” (p. 715) He says that if the
Modest view were correct, “one is left wondering why an unending array of non-arbitrary
reasons is necessary for justification...” Let me respond briefly by, first, making a minor
correction in the characterization of PAA and, second, by distinguishing two general types of
theories of justification and locating my brand of infinitism within those types of theories.

As Gillett correctly pointed out, I claimed that the best reason for adopting infinitism was that it
was entailed by two seemingly plausible epistemic principles, PAA and PAC. Gillett thinks of
those principles as instantiations of two more general metaphysical principles which do involve
appealing to dependent properties that arise only in virtue (at least in part) of some other object
possessing the property. Recall that PAC and PAA were stated as follows:

PAC: For all x, if a person, S, has a justification of x, then for all y, if y is in the evidential
ancestry of x for S, then x is not in the evidential ancestry of y for S.

PAA: For all x, if a person, S, has a justification for x, then there is some reason, r1, available to
S for x; and there is some reason, r2, available to S for r1, etc.

But strictly speaking, at least on one reading, neither PAA conjoined with PAC nor the two
parallel general metaphysical principles formulated by Gillett respectively entail the denials of
coherentism and foundationalism or the parallel metaphysical views that would be incompatible
with an infinite series of objects with dependent properties unless one takes the “etc” in PAA to
include the claim that “there is no last member in the series.” That is what I had intended; but it
was not presented clearly.

Let me try to make clear why that clause is needed. Imagine a view that we could call
“Infinitistic Foundationalism,” namely that there are an infinite number of non-repeating
propositions between any given proposition and a foundational one. We could imagine a view,
Infinitistic Coherentism, that holds that the set of coherent propositions is infinite. Both PAA and
the first general metaphysical principle need to be stated more carefully so as to preclude those
readings. Here is the revised PAA:

PAA*: For all x, if a person, S, has a justification for x, then there is some reason, r1, available to
S for x; and there is some reason, r2, available to S for r1, etc, and there is no last reason in the
series.

That is what I had intended to express because (I think) it is part of the picture of reasoning that
the Pyrrhonians had in mind that leads to the “regress of reasons” problem. (I leave it to the
reader to revise the parallel general metaphysical principle.) Together with the original PAC,
PAA* entails what I meant by “infinitism.” So defined, substantive infinitism is incompatible
with substantive coherentism and substantive foundationalism.14 Substantive coherentism would
deny PAC and substantive foundationalism would deny PAA*.

All three views (infinitism, foundationalism and coherentism) could come in two basic varieties:
1) a holistic or emergent variety and 2) a dependent-property variety. The holistic variety thinks
of justification as emerging in virtue of a set of propositions having certain features. For
example, a holistic foundationalist would maintain that a proposition is justified for S in virtue
of there being a set of propositions correctly characterized by PAC and by a principle
incompatible with PAA*, namely something like this:

F*: The set of propositions (whether infinite or finite) is such that each member is the last
member of a series of propositions or properly related to a last member.15

A holistic coherentist would hold PAA* and a principle incompatible with PAC, namely
something like this:

C*: There is a set of propositions (whether infinite or finite) available to S such that some (or
perhaps all) of them are in each other’s evidential ancestry.

On such holistic views, propositions are justified for S in virtue of being members of sets of
propositions with the features specified by the particular holistic account of justification, i.e.,
either a holistic coherence, foundational, or infinitist account. Justification emerges whenever the
set of propositions has the salient features, but it is not a dependent property that is passed from
one member of the set to another. Most coherence theories are holistic. My brand of infinitism is
a holistic one.

Each type of epistemic theory could come in a dependent-property variety by holding that
justification is a property attached to a proposition that can be transferred from one proposition
to another. Justification would be a property similar to (the medieval notion of) momentum that
gets passed from one object to another. On such a substantive coherence theory, the property
would be transferred from one proposition to another and eventually back to the first one.
(Analogously, basketball players standing in a circle pass the ball from one to another.) On a
substantive foundationalist theory, justification arises spontaneously, so to speak, in the basic
propositions and then gets transferred to other propositions.16 I think it is fair to say that most
foundationalist theories are dependent property views. Dependent-property infinitism would
appeal to the type of dependent property that Gillett thinks would lead to the Structural
Objection.17

I think we are now in a position to see how to respond to the questions posed by Gillett, namely,
i) in what way is Modest infinitism a substantive alternative to coherentism and foundationalism
and ii) why would infinitism hold that “an unending array of non-arbitrary reasons is necessary
for justification?’ In the holistic varieties of epistemic theories of justification, the determinant
(to use the useful terminology suggested by Gillett) of some proposition being justified is that it
is a member of a set of propositions with some feature(s) specified by the epistemic theory. It is
in virtue of being a member of such a set that any given proposition is justified. In the
dependent-property views, propositions have the dependent property in virtue of some other
proposition having it - with the caveat that in foundationalist theories some propositions (i.e.,
basic ones) have the property in virtue of some other feature (for example, the content of the
proposition or the causal history of the belief that has the proposition as its content or some
feature of the conversational context). Thus, i) each one of the six sub-theories is a substantive
alternative to the other five theories because the six theories are contraries and ii) a reason for
thinking that an unending array of non-arbitrary reasons is necessary for justification is that
PAA* and PAC are true. Of course, I have not argued for either PAC or PAA* here. My point is
merely that Modest infinitism - what I prefer to call “holistic infinitism”4s incompatible with the
other five sub-theories and that PAA* and PAC (as I have characterized their constituent terms),
if true, provide a good (indeed entailing) reason for being a holistic infinitist.

Now to my final point, namely, that IV Regresses are not vicious, per se. In the terminology just
employed, my claim is that dependent-property infinitism should not to be discarded simply
because it appeals to an unending array of propositions each having an epistemic property in
virtue of some other proposition, not already employed, having that property. As I’ve argued
above, I do not think dependent-property infinitism is the only form of infinitism or even best
form of infinitism, but the question here is whether dependent-property infinitism should be
rejected because IV regresses arc vicious, per se.

Consider a classic example of an IV Regress.’18 Suppose we were to assume (contrary to fact)

that the “natural state” of physical objects (or some subset of them - what Aristotle called
“sublunary bodies”) is to be at rest. One possible view is that if there is such an object, x, that has
motion, then there is another such object, y (not identical to x), that has motion, and then another,
z, that has motion and is not identical to y or x, and there is no last member of the series. That is,
it could be held that there is an infinite IV Regress because the motion of one physical object
depends upon the motion of the another, that depends upon the motion of another ... without end.
In this case, let us grant that the motion of one object is cited in the explanation of the motion of
another . . . without end.
Now what exactly is the basis for infiniphobia? Why is such a regress thought to be vicious? The
regress has this form (where the variable ranges over a specified type of object): x has H in
virtue (at least in part) of y having H, y has H in virtue (at least in part) of z having H, and there
is no last member of the series. It would be vicious if we ran out of objects of the appropriate
type! But an IV Regress allows for the possibility that the number of such objects is infinite. The
IV Regress concerning the motion of sublunary objects is designed to say something true and
important about the motion of such objects - namely that the motion of each one depends (at
least in part) upon the motion of another that has not already appeared in the series. (Never mind
that this is false. The issue is whether it could be true and, if true, whether it would be
important.) Now, I agree with Gillett that the “structure of the IV Regress means that it cannot
produce any of its objects” if that means that the existence of the regress does not explain why
there is any sublunary body that is moving. But the regress was not designed to say everything
important about motion. In particular, it was not designed to say why there is some motion of
sublunary bodies rather than none at all. Put another way, Gillett is correct that “the question
consequently arises how it could ever come to pass that any [emphasis added] member of the
chain has the property H?’ (p. 713) But the IV Regress is not designed to answer that question. It
is designed to answer the quite different question “How does it come to pass that each member
of the chain - taken individually - has property H?’ The answer is that there is always another (as
yet unmentioned) member of the chain in virtue of which each member has H. What explains the
existence of such a chain of entities with dependent properties? This is the old question: “Why is
there some motion of sublunary bodies rather than none at all?’ Or more generally, “Why does
anything have H?’ The N Regress is completely silent on that. But that doesn’t make an IV
Regress vicious. It merely makes it less than a full account of everything that is interesting about
H-ness.

*I want to thank Anne Ashbaugh, Brian Mclaughlin, Ernest Sosa, and especially Carl Gillett for
their help with this paper. Needless to say, at least one of those people will not agree with
everything I have to say here!

1 Carl Gillett, “Infinitism Redux? A Response to Klein.” Philosophical and Phenomenological

Research, 66.3.709-717. The papers of mine that Gillett is discussing are: “Foundationalism and
the Infinite Regress of Reasons,” Philosophy and Phenomenological Research, 58.4, 1998, 919-
925; “Human Knowledge and the Infinite Regress of Reasons,” [HKIRR] Philosophical
Perspectives, 13, J. Tomberlin (ed.), 1999, 297-325. Page references in the text are to Gillett’s
article and to my papers.

2 Sextus Empiricus wrote:

The Mode based upon regress ad infinitum is that whereby we assert that the thing adduced as a
proof of the matter proposed needs a further proof, and this again another, and so on ad
infinitum, so that the consequence is suspension [of assent], as we possess no starting-point for
our argument ... We have the Mode based upon hypothesis when the Dogmatists, being forced to
recede ad infinitum, take as their starting-point something which they do not establish but claim
to assume as granted simply and without demonstration. The Mode of circular reasoning is the
form used when the proof itself which ought to establish the matter of inquiry requires
confirmation derived from the matter; in this case, being unable to assume either in order to
establish the other, we suspend judgement about both. (Sextus Empiricus, Outlines of
Pyrrhonism, I, 166-169.)

3 A more recent defense of PAC and PAA (in slightly revised forms) appears in “Infinitism is
the Best Solution to the Regress Problem,” Contemporary Debates in Philosophy, ed. Matthias
Steup, (Blackwell). spring 2003, forthcoming.

4 1 changed the quotation slightly by making the names of the general metaphysical principles
conform to the names used in this paper.

5 There are some other necessary conditions for winning, but they aren’t relevant here; just as
there are some necessary conditions for being justified not captured by PAC and PAA.

6 Note that you do not have to win starting with p1, (temporally) before you win starting with p.

7 Here are the analogs:

Cardboard pieces: propositions

Winning: being justified in believing
The pieces fitting together: objective availability of reasons
S’s being able to fit the pieces together: subjective availability of reasons

This might seem to be a game that is very easy to win since you can actually do nothing once the
game starts, and win. But remember that I am trying to develop an analogy with being justified
in believing that p. I take it that to be justified in believing p is equivalent to being entitled to
believe that p. S might be so entitled even if S does not believe that p or has yet to actually form
the beliefs that provide the reasons for p. The relevant necessary conditions for S’s being entitled
to believe that p are that there are an infinite number reasons which are objectively and
subjectively available to S. Winning requires that there be an infinite number of pieces that fit
together and that S have the ability to fit them together. So, it is not so easy to win after all. In
addition, S’s winning and determining whether S won are two different acts. The latter is harder
yet. (That is one reason why justification doesn’t iterate necessarily.)

8 There are other necessary explanatory conditions, e.g., there being no non-overridden counter
evidence subjectively available to S. Presumably that is a feature of foundationalism and
coherentism as well. Thus, it is not relevant to the discussion here.

9 Ernest Sosa, “The Raft and the Pyramid,” Epistemology (Oxford: Blackwell Publishers, 2000).
p. 142. Sosa does not use the expression “infinitism,” but I think that it is clear from the context
that the contrasting theory to formal foundationalism is the view that 1 call ‘infinitism.’

10 Richard Fumerton, “A Reply to my Critics,“ Philosophy and Phenomenological Research,

58.4, December, 1998, 936-937.
11 Alvin Goldman also argues that satisfying the supervenience condition is desideratum of an
acceptable account of justification. See his “What is Justified Belief?” in Justification and
Knowledge, ed. G.S. Pappas (Dordrecht: D. Reidel. 1970), 1-23.

12 I say that because I am not convinced that the normative does, indeed, supervene on the non-
normative, or even that this is the best way to cut nature or non-nature at its joints. But
considering that issue would take us too far afield.

13 In the “Raft and the Pyramid,” Sosa distinguishes between “formal foundationalism” and
various forms of “substantive foundationalism.” In the terminology I am employing, the former
holds that propositions are justified in virtue of some non-normative property which they all
possess. Holistic or warrant-emergent forms of coherentism would qualify as formal
foundationalism because such views hold that propositions are warranted (at least prima facie
justified) just in case they are members of a set of coherent propositions with certain features. I
think infinitism, then, would be a form of formal foundationalism. Whether that makes the
category of “formal foundationalism” more or less interesting, I leave to the reader. On the one
hand, it seems more interesting since all of the competing substantive accounts (foundationalism,
coherentism, infinitism) turn out to be forms of one view - formal foundationalism. On the other
hand, the category seems less interesting since it fails to distinguish between interesting
differences between the substantive views.

14 See fn. 13.

15 Recall that PAA* and PAC are meant to characterize (significant) necessary but not sufficient
conditions for justification. F* and C* are also meant to state (significant) necessary conditions.

16 I do not mean to imply that a foundationalist would have no explanation of the way in which
warrant or justification would arise in basic propositions. They would have some story to tell
which accounts for the fact that some propositions have (at least) prima facie warrant that does
not depend upon the warrant of any other proposition. It take it that is what is meant by “self-
justified” propositions.

17 The distinction between the two types of coherentist theories has been well recognized but I
do not think it has been noted that there are two forms of foundationalism and infinitism. For
further discussion of these two forms of coherentism, see Laurence BonJour in “Can Empirical
Knowledge Have a Foundation?” American Philosophical Quarterly, 15.1 (1978). 1-13.

18 This is an example of an IV Regress that Gillett cites. (See his fn. 5.)
Philosophy and Phenomenological Research Vol. LXVI, No. 3, May 2003
Infinitism Redux? A Response
CARL GILLETI‘
Illinois Wesleyan University

Foundationalist, Coherentist. Skeptic etc.. have all been united to Klein in one respect - all accept
epistemic justification cannot result from an unending, and non-repeating. chain of reasons. Peter
Klein has recently challenged this minimal consensus with a defense of what he calls
“Intinitism” - the position that justification can result from such a regress. Klein provides
surprisingly convincing responses to most of the common objections to Infinitism, but I will
argue that he fails to address a venerable metaphysical concern about a certain type of regress.
My conclusion will be that until Klein answers these metaphysical worries he will not have
restored Infinitism as a viable option in epistemology.

Infinite regresses have played important roles in many philosophical debates, but they have
recently been most prominent in epistemology. A wide array of positions, Foundationalist,
Coherentist, Skeptical etc., have all been adopted in response to a looming regress of epistemic
reasons and these positions are united in one respect - they all accept that epistemic justification
cannot result from an unending, and non-repeating, chain of reasons. In his commentary in this
journal on Richard Fumerton’s (1995) and in a longer paper elsewhere, Peter Klein has recently
challenged this minimal consensus with a defense of what he calls “Infinitism” - the position that
justification can result from such a regress.1

Klein provides surprisingly convincing responses to what he identifies as the toughest objections
to Infinitism. But my goal will be to argue that Klein fails to address a venerable metaphysical
concern, dating back at least to Aquinas, about a certain type of regress. The standard response to
such problems is to abandon one of the theses that generate the relevant regress, but I will argue
that the very nature of Infinitism precludes this type of reaction. My conclusion will be that until
Klein successfully offers an answer to these metaphysical worries about regresses, then he will
not have restored Infinitism as a viable option in epistemology.

The promise of Klein’s Infinitism is that, unlike other positions in epistemology, it would allow
us to hold two plausible principles and still be justified in our beliefs. The first of these
principles, where ‘x’, ‘y’ etc. are all beliefs, is what Klein calls the “Principle of Avoiding
Circularity” (PAC):

(PAC) For all x, if a person, s, has a justification for x, then for all y, if y is in the evidential
ancestry of x for s, then x is not in the evidential ancestry of y for s. (p. 298)

By “evidential ancestry” Klein means that “if r is a reason for p. and q is a reason for r, then r is
in the evidential ancestry of p, and q is in the evidential ancestry of p and r” (p. 298). PAC is a
highly intuitive prohibition against circular reasoning and Klein’s other principle is equally
attractive. This is what he calls the “Principle of Avoiding Arbitrariness” (PAA):

(PAA) For all x. if a person, s, has a justification for x, then there is some reason, r1, available to
s for x; and there is some reason, r2, available to s for r1; etc. (p. 299)
Again, PAA makes a familiar and appealing demand, namely that our reasons cannot be arbitrary
and must always themselves be supported by non-arbitrary reasons.

PAA and PAC are both highly attractive principles and together they entail that justified belief
depends upon the existence of an unending and nonrepeating regress of non-arbitrary reasons.
For PAA implies there must always be some further non-arbitrary reason in order for prior
reasons to be non-arbitrary and PAC entails this chain will be non-repeating. Furthermore, given
the truth of PAA and PAC, it appears that there can only be justified beliefs if such a regress of
reasons can result in any of its members being non-arbitrary reasons. Klein accepts these
implications, for he summarizes the attractions of his account as follows:

It is the straightforward appeal of these principles [PAC and PAA] that is the best reason for
thinking that if any beliefs are justified, the structure of reasons must be infinite and
nonrepeating. (p. 299)

As we shall shortly see, Klein further contends against the Skeptic that for all we know such an
infinite chain may produce justified beliefs.

Klein is only committed to the existence of an unending and non-repeating regress of non-
arbitrary reasons being a necessary, and not sufficient, condition for justified belief, since he
contends there are other necessary conditions (p. 318, n. 8). Yet in order to make even this more
limited claim plausible Klein addresses four objections that have convinced most philosophers
otherwise. I lack the space to examine all of Klein’s subtle and careful responses, but I encourage
the reader to examine his papers directly since he does a surprisingly successful job of answering
many common worries. My focus here, however, must be limited to Klein’s interpretation of the
fourth objection to Infinitism that he considers-what he calls “The Specter of Skepticism” and
tells us “is the most difficult objection because it is the most difficult to fully understand” (p.
312).2

As his name for it implies, Klein takes this concern about the unending regress of reasons to be a
distinctly epistemic problem suggesting that “the objection rests upon a Cartesian-like view that
the whole point of reasoning is to ‘settle’ an issue” (p. 312). Klein argues that this view results in
a ultimately skeptical demand for “lifetime” or “final” guarantees and, specifically, final
guarantees both that there actually exists an unending array of reasons for our beliefs and that
humans have the cognitive capacities to comprehend these reasons. With regard to such
demands, Klein explicitly sides with Richard Foley (1990) in arguing that they should not be
legislated against and are the understandable result of the epistemic drive to take our own
methods of inquiry as objects of study. Nonetheless, Klein also follows Foley in arguing that we
should reject these demands for “final” guarantees and he consequently argues that:

... we do have limited guarantees. And for all I know, there might be an infinite number of such
limited guarantees. Thus, although no a priori argument is available whose conclusion is that
there is an infinite regress ... of reasons, as we have seen there is also no such argument for the
claim that there is no such set of reasons available. (p. 316)
Klein therefore concludes that so long as there might be such an unending and non-repeating
chain of reasons, and we lack evidence that there is not, then the truth of Infinitism is compatible
with our beliefs being justified, contrary to skeptical intuitions and received opinion in
philosophy.

Klein admits that this final worry is the hardest to understand and I therefore wish to explore an
interpretation of this concern about regresses. I will argue that these worries are metaphysical
and completely general in nature, and arise where instances of two general principles hold. The
first principle can be defined as follows (I use ‘entity’ widely to refer to individuals, states,
properties, events, and processes):

(I) For all entities, an entity x has a property H only in virtue of, amongst other possible
necessary conditions, some entity y having the property H, and y is H only in virtue of some
entity z being H; etc.

This principle simply implies that it is a necessary condition of any entity having the property H
that some other entity has H, and so on. The second principle is:

(II) For all entities, if entity x has the property H only in virtue of, amongst other possible
necessary conditions, some entity y having the property H, then y does not have H in virtue of x
being H.

Principle (II) entails that there are no circles, or mutual dependencies, in any chain of in virtue of
relations that result in an entity’s having the property H. It should be obvious that (I) and (II) are
general analogues of Klein’s more specific principles PAA and PAC, and we will return to this
point shortly. More importantly, however, it should also be clear that principles (I) and (II)
generate a regress of entities bearing ‘in virtue of relations to each other. For (I) demands that for
any entity to have the property H there must be some distinct entity that is H, and this further
entity has H only in virtue of some further entity being H, and so on. And principle (n) prevents
the same entity from appearing twice in the resulting chain. To ease my discussion, I will call the
latter an ‘IV Regress’, to mark the role of ‘in virtue of relations’ in it, and I shall refer to a
property such as H as a ‘dependent property’ of the regress.’

At least since Aquinas there have been general metaphysical concerns about whether a regress
like an IV Regress can actually result in any of its members having its dependent property.4
These concerns may be put as follows. Let us call an arbitrary member of the regress ‘so’ and
consider what effect its addition would have to the regress. Regardless of the nature of this
entity, and its place in the chain, all the members, s-sn+1, prior in the regress to s, only
instantiate H in virtue of sn being H. But sn will only have property H in virtue of some still
further entity being H. And, of course, this will be the situation regardless of how many more
entities are added to the regress. The question consequently arises how it could ever come to pass
that any member of the chain has the property H? For it appears that whatever entity, or structure
of entities, is added to the chain of prior entities, consistent with the governing principles (I) and
(II), this addition will not be sufficient for the dependent property H to feed back to any member
of the regress. Adding more entities one by one, or even as structured groups, will still not
suffice, for once again these additional entities will only have H in virtue of some still further
entity that is H. Thus, the objection ultimately concludes, there is no entity, or structure of
entities, that can be added to an IV Regress, consistent with its governing principles, that will
suffice for any of its dependent properties to feed back to any of the members of its chain. Let us
call this the ‘Structural Objection’, since it argues that the structure of an IV Regress means that
it cannot produce any of its dependent properties.5

We should note that, as well as being completely general, the Structural Objection is also
ontological in nature. That is, the Objection does not depend upon a distinctly epistemic demand,
such as that for explanation, sufficient reasons or non-arbitrary reasons, since the Objection is
solely based upon the metaphysics of IV Regresses. It is apparently true that in Aquinas’ most
famous application of something like the Structural Objection the driving concern is a demand
for explanations or sufficient reasons (Rowe (1975) and (1997)). But this is merely an artifact of
the particular theses Aquinas implicitly used to generate an IV Regress, specifically some
version of the ‘Principle of Sufficient Reason’. However, IV Regresses may result from
instances of principles (I) and (11) which are non-epistemic in character, since the relata of ‘in
virtue of‘ relations take many forms.6

The Structural Objection is therefore rather different than the skeptical concern about infinite
regresses raised by Klein. The latter is a distinctively epistemic worry, whereas the Structural
Objection is a general metaphysical concern about the potential of IV Regresses to produce their
dependent properties. And the arguments about guarantees that Klein adduces to turn the
Skeptic’s epistemological objection do not suffice to allay the ontological concerns about IV
Regresses.7 The driving force behind this Objection is not a demand for “final” or “life-time’’
guarantees that there is an infinite chain of non-arbitrary reasons, or even that humans have the
cognitive capacities to comprehend such a chain of reasons. Rather, the Structural Objection is
based upon a general a priori argument that no IV Regress can produce its dependent property;
and hence that PAC and PAA cannot produce an IV Regress that results in its dependent
property of being a non-arbitrary reason.8

The standard response to IV Regresses, and the arguments based upon them, is to reject one of
the theses that generate the regress. Thus Aquinas argues, for example, that one IV Regress
shows that we should reject the thesis that all beings are contingent, whilst his opponents often
conclude that this regress shows we should reject the Principle of Sufficient Reason. Similarly, in
the epistemological debate, in response to the IV Regress of reasons we have seen result from
principles like PAA and PAC, Foundationalists argue we must abandon principles like PAA,
Coherentists cast off PAC, and Skeptics conclude we can never have justified beliefs. What
unifies these reactions is that they accept the Structural Objection that IV Regresses cannot result
in their dependent property and hence abandon one, or more, of the theses that produce the
relevant regress. Given their position, Klein and other Infinitists cannot take such an approach
and must face the Structural Objection square on.

However, it is often objected that an Infinitist like Klein is not, or at least need not, be committed
to an IV Regress of reasons. Perhaps the Infinitist may only be committed to the claim that it is
necessary for justified belief that there must be some infinite regress of beliefs, all of which are
non-arbitrary reasons. But where the beliefs are not non-arbitrary reasons only in virtue of some
further belief being a non-arbitrary reason. Thus the objection is that a regress is necessary for
justified belief under Infinitism, but it not the type of viciously problematic IV Regress
underlying the Structural Objection.

We can better understand this objection and the problems it faces using a couple of crude
technical terms. Let us therefore take B to be a ‘marker’ of A just in case it is true that “A only if
B”, but not true that “A in virtue of B’; and refer to B as ,a ‘determinant’ of A just in case it is
true that “A only if B” and it is true that “A in virtue of B”. Using these crude notions we can
apparently distinguish different kinds of infinitism and articulate the idea behind the objection.
For example, we may define ‘Modest’ Infinitism as the position that an infinite regress of
reasons is a marker of justified belief.9 Whilst reserving the name ‘Bold’ Infinitism for a position
that claims that an infinite regress of reasons is a determinant of justified belief. The objector’s
point is apparently that the Modest position really is a species of Infinitism and clearly is not
committed to an IV Regress.10

In response we must be careful to mark the different features of the Modest and Bold positions.
The Bold view is what I have been refening to as ‘Infinitism’ and it is a position whose truth is
incompatible with the truth of either Foundationalism or Coherentism. As a result, the Bold
position does appear to be a substantive epistemological position, though of course it faces the
Structural Objection. But what of the Modest variant? We should note that by avoiding
commitment to an IV Regress the Modest picture holds simply that a infinite regress
accompanies whatever conditions are involved in the metaphysical ‘guts’ of justification. The
latter conditions being correctly described by either Foundationalism or Coherentism is thus
compatible with the truth of the Modest view. Given its compatibility with such positions, it
appears questionable whether the Modest view can restore Infinitism as a substantive option in
epistemology and whether it is correctly labeled ‘Infinitism’.

Building upon this latter point, we can use what I will call the ‘Infinitist’s Dilemma’ to further
sharpen our view of the difficulties facing Infinitists. The Dilemma proceeds as follows: When
PAA and PAC are true either a belief, ‘x’, is a non-arbitrary reason only in virtue of some other
belief, ‘y’, being a non-arbitrary reason (where y cannot be in the evidential ancestry of x), or it
is not. If the former is true, then being a non-arbitrary reason is a dependent property of an IV
Regress and one faces the Structural Objection. On the other hand, if there can be a belief that is
a non-arbitrary reason, but not in virtue of some further belief being a non-arbitrary reason, then
one is left without any reason why an unending array of non-arbitrary reasons is necessary for
justification and one plausibly abandons Infinitism. (And if one gives such a reason, then one
will plausibly have argued that it is in virtue of another belief y being a non-arbitrary reason that
x is a non-arbitrary reason. Hence one would be forced back to the first horn of the dilemma).
Thus, the Infinitist’s Dilemma apparently shows that either one is an Infinitist and must face the
Structural Objection, or one plausibly abandons Infinitism.11 Consequently, Klein and other
Infinitists cannot avoid commitment to IV Regresses that result in their dependent properties, on
pain of ceasing to be Infinitists.

To conclude, Peter Klein has achieved a great deal by showing that Infinitism can answer many
of the concerns commonly leveled against it. But we have nonetheless found that a general
metaphysical concern about IV regresses still dogs even Klein’s subtle Infinitism. These
concerns do not derive from skeptical, or other epistemic, grounds, but from venerable
metaphysical arguments that IV Regresses cannot produce their dependent properties, whether
the property of being a non-arbitrary reason or some other feature. The Structural Objection thus
provides an a priori argument that there cannot be the unending, non-repeating chain of non-
arbitrary reasons necessary for the truth of Infinitism. As we have seen, unlike other participants
in the epistemological debate, Infinitists such as Klein cannot avoid the Structural Objection and
Infinitism must remain unrestored until an answer is successfully provided to it.12

1 See Klein (1998) and (1999). I focus on the latter paper given its more detailed account and all
page references are to this paper unless otherwise noted. The term “Infinitism” is attributed by
Klein to Moser (1984).

2 Let me briefly summarize the other objections Klein addresses, as well as his responses, to
reassure readers that he does not answer the problems I will later outline. (I refer the reader to
Klein (1999) for the full details of his responses). The first objection is what Klein dubs “The
Infinite Mind Objection” which is expressed in a variety of ways, but presses the point that a
finite human mind cannot encompass, in whatever manner, the infinite array of reasons
Infinitism implies are necessary for justification. Klein distinguishes between ‘occurrent’ and
‘dispositional’ beliefs and argues that Infinitism implies only that humans must have the
potential for an infinite array of dispositional beliefs. Furthermore, he argues that when we are
discussing dispositional beliefs it is not implausible that we might have an infinite number of
such beliefs (pp. 306-310). The next concern that Klein addresses is Aristotle’s argument that not
all beliefs can be inferential which seems to imply that, contrary to Infinitism, there is an end to
the chains of inference connecting our beliefs. Here Klein wields the distinction between the
genesis and the justification of a belief, agreeing with Aristotle that our beliefs are not generated
by an endless chain of inference, but pressing the Infinitist point that our beliefs may nonetheless
be justified by such an endless chain (pp. 310-11). Third, Klein considers the recent reductios of
Infinitism, offered by Oakley (1976), and Post (1980) and (1987). that seek to show that if
Infinitism is true, then any contingent proposition is justified. Klein answers these arguments by
arguing that Oakley uses an unsound principle, at odds with the truth of PAC, and that Post relies
upon an implausible construal of Infinitism, specifically taking it to be the claim that an
unending regress of reasons is by itself sufficient for justification (pp. 311-12).

3 My interpretation of the problems surrounding vicious regresses is superficially similar to

Clarke (1988), but ultimately diverges from his account. Clarke’s view is that reductio
arguments based upon vicious regresses ultimately depend upon premises that some property is
both “conditionally” and “unconditionally” instantiated by individuals. Consequently, someone
such as Klein who argues that justification is always only “provisional”, i.e. “conditional” to use
Clarke’s term, can slip the vicious regress objection under this interpretation of its nature. As
will become clear below, such claims of “provisionality” do not ameliorate the problems I
outline. (See Black (1988) for a nice abstract formulation of the structure of vicious regress
arguments).

4 Aquinas (1964).

5 In effect, the Structural Objection illuminates that what Aristotle termed an “actual infinity”
must exist in order for Infinitism to be true and all the problems surrounding the actual infinite
consequently dog Infinitism. (Sosa (1980) has defended the notion of the “actually infinite” in
defending lnfinitist positions from criticism, but Moser (1985) provides plausible arguments that
Sosa’s notion does not succeed with this task).

6 For example, consider these two non-epistemic principles:

(I*) An entity x has a property of being in motion only in virtue of some entity y having the
property of being in motion; and y is in motion only in virtue of some entity z being in motion;
etc.

And the further principle:

(II*) For any entity x. if x has the property of being in motion only in virtue of some entity y
having the property of being in motion. then y is not in motion in virtue of x being in motion.

Together (I*) and (II*) generate an IV Regress whose dependent property is being in motion,
even though these principles are non-epistemic in character.

7 In the course of his discussion of such skeptical concerns Klein also argues (p. 314) that we
should reject what he calls the “Completion Requirement” of Fumerton (1995). And he also
responds to Sosa (1980)’s claim that Infinitism cannot adequately account for the differences
between having a justification and being justified (pp. 314-315). I contend that his responses in
neither case suffice to provide an answer to the Structural Objection.

8 We should note that specific application of the concerns driving the Structural Objection have
often been used in attacking Infinitism. For example, Moser (1985) provides a recent critique of
Infinitism of this kind and his objection is in effect an application of the Structural Objection to
the case of epistemic justification.

Fumerton (1998) responds to Klein’s (1998) defense of Infinitism on rather different grounds by
objecting that an Infinitist cannot provide a non-circular definition of justification. However,
Klein’s (1999) makes it plausible that this objection may be met, since it outlines an account of
justification which does not itself rely upon this notion and instead uses the idea of “subjectively
and objectively available reasons”.

9 Note that the Modest view basically entails that PAC and PAA are incorrectly interpreted as
instances of principles (I) and (11).

10 Thanks to Peter Klein for helping me see this objection.

11 And note that the Dilemma can be applied if the Infinitist moves her focus from the property
of being a non-arbitrary reason to some other property, for example being what Klein calls a
“subjectively and objectively available” reason. For it appears that the Infinitist must still be
committed to some infinite regress of beliefs where the dependent property is now being a
“subjectively and objectively available” reason. Furthermore, the beliefs in this regress must be
connected by “in virtue of’ relations, for otherwise we have no reason why such an unending
chain of beliefs is necessary for justification. As a result. one either ceases to be an Infinitist, or
one is again committed to an IV Regress of beliefs.

12 A version of this paper was read at the 2001 Pacific APA conference in San Francisco.
Thanks to the audience there for comments and to Seth Crook, Paul Lodge and, especially, Ben
Haines and Peter Klein for discussion of the issues of this paper.
John Greco
Ernest Sosa and His Critics

Supervenience and Normative Epistemology

The issues above are brought into clear focus by another important theme in Sosa’s work: that of
the supervenience of the evaluative. In general, Sosa thinks, we should accept the thesis that the
evaluative supervenes on the non-evaluative. In other words, we should accept the idea that a
thing has its evaluative properties in virtue of its non-evaluative properties.

For example, suppose we think that a particular car is a good one. Surely this must be in virtue of
other properties that the car has, for example the mileage that it gets, its ability to accelerate, its
look, etc. To deny this would be to accept that two cars could be alike in all their non-evaluative
properties (both intrinsic and relational) and yet differ in their evaluative ones. But this seems
absurd. The same reasoning holds for epistemically evaluative properties.

Suppose S and Twin-S live lives indistinguishable physically or psychologically,

indistinguishable both intrinsically and contextually, on Earth and Twin-Earth respectively.
Surely there can then be no belief of S epistemically justified without a matching belief held by
Twin-S with equal epistemic justification. Epistemic justification must accordingly supervene
upon or derive from physical or psychological properties of the subject of belief, properties
either intrinsic or contextual. (KP, 110)

An important aim of epistemology, Sosa reasons, is to specify the non-evaluative basis of

supervenience, thus allowing a special sort of insight into the nature of justification and
knowledge. In this respect, coherentists and foundationalists share a common goal: to specify
such a basis in relatively simple and complete terms.

It is from this perspective that the argumentative account of justification seems clearly hopeless.
According to (AJ), for one to justify a belief is for one “correctly and seriously to use
considerations or reasons in its favor.” But again, how are we to understand “correctly”? The
most obvious way is in terms of some epistemically evaluative property. Alternatively,
“considerations or reasons” will have to be understood that way. And therefore the
argumentative account fails to get beyond the epistemically evaluative. What is required for that,
as we saw above, is that something else be considered more ultimate.

Also from this perspective, certain arguments against foundationalism can be seen in a new light.
For example, “doxastic ascent” arguments charge that there can be no property F in virtue of
which belief B is foundationally justified, unless one is justified in believing that B has F. But
then B is not foundational at all, since its justification depends on B′: the belief that B has F. But
this line of reasoning, Sosa argues, would implicate all of substantive epistemology with
foundationalism, coherentism included. For anyone who accepts it would have to accept the
following as well:

that a belief B is justified in virtue of membership in a coherent system only if one is justified in
believing that it has such membership.
And more generally,

that a belief B is justified in virtue of any property X only if one is justified in believing that B
has X.

Clearly, such commitments entail an infinite regress of justified beliefs. More importantly from
the present perspective, however, they are inconsistent with the supervenience of the
epistemically evaluative. No matter what we specify as a non-epistemic source of justification,
coherence included, such commitments require that something else is needed; viz., another
justified belief. This “would then preclude the possibility of supervenience, since it would entail
that the source of justification always includes an epistemic component” (KP, 183)…
Richard Fumerton
Ernest Sosa and His Critics

Sosa’s Conception of Epistemic Ascent to Reflective Knowledge

In a number of places, most recently in “Two False Dichotomies,” Sosa suggests an answer to
this question of what is necessary to achieve reflective knowledge. He begins by emphasizing the
following principle of epistemic ascent:

(KA) If one really knows that P and one considers whether one does, then one must be justified
in thinking that one does.

Notice that Sosa does not assert that knowledge implies knowing that one knows, or even that
knowledge implies having the capacity to know that one knows. Nor is it clear whether he would
accept an analogous principle of ascent for justification:

(JA) If one really believes P justifiably then if one considers whether one does then one must be
justified in thinking that one believes P justifiably.

This latter issue is important if we fear regress from Sosa’s ascent principle. After all, while one
might think it initially plausible to suppose that someone’s knowing P requires that person to
believe justifiably that he knows that P if he considers the question, do we also want to insist that
if the person were to consider the question of whether he justifiably believes that he knows that P
he would find himself justifiably believing that he justifiably believes that he knows that P, and
so on ad infinitum? In no time at all the higher-level beliefs will presumably get too complicated
for any normal epistemic agent to keep things straight.

I want to be clear about the nature of the regress I fear from JA. I’m not suggesting that a
principle of ascent need involve one in conceptual regress. Just because one thinks that my
justifiably believing P would require me to justifiably believe that I am justified in believing P if
I consider the question, it doesn’t follow that that justified metabelief need be constitutive of my
justifiably believing P. In other words, it needn’t be an analytic truth that if I justifiably believe P
then upon consideration I would justifiably believe that I have such justification. But even if the
principle were not analytic, a principle of ascent concerning justification might require
something of which finite epistemic agents are clearly incapable. Now it may be that Sosa would
reject JA and for that reason deny that he faces any problematic regresses, but I’m not sure why
JA has any less initial plausibility than KA, particularly if we can make a distinction between
animal-level justification and reflective justification analogous to Sosa’s distinction between
animal-level knowledge and reflective knowledge…
Peter Klein
Ernest Sosa and His Critics

Ernest Sosa has developed a comprehensive theory of propositional knowledge, which he labels
“virtue perspectivism.”1 It integrates many of the salutary aspects of foundationalism and
coherentism, internalism and externalism, evidentialism and reliabilism. In this chapter I want to
focus on his treatment of philosophical skepticism and offer some friendly criticisms. I say
“friendly” for two reasons. He is a friend of mine and I share many of his basic commitments in
epistemology.

In order to understand Sosa’s response to philosophical skepticism it is crucial to see that Sosa
adopts a Cartesian distinction between two forms of knowledge – cognitio and scientia. Cognitio
is true, justified belief that is appropriately caused or subjunctively related to the object of the
belief. Tempting though it would be, I will not comment on Sosa’s account of cognitio beyond
what is required to discuss his response to skepticism.2 Scientia is a reflective form of knowledge
that results from “ascending” to a position from which we can assess whether we satisfy the
conditions required by cognitio. Bluntly and somewhat contentiously put, this form of
knowledge is reserved for epistemologists (professional or amateur).3

Here is a crucial passage from Descartes’s replies to the objections to the Meditations cited, by
Sosa, in which Descartes draws the distinction between the two grades of knowledge:

That an atheist can clearly know that the three angles of a triangle are equal to two right angles, I
do not deny: I merely say that this knowledge (cognitio) of his is not true science (scientia),
because no knowledge which can be rendered doubtful should, it seems, be called science. Since
he is supposed to be an atheist, he cannot be certain that he is not deceived even in those things
that seem most evident to him, as has been sufficiently shown; and although this doubt may
never occur to him, nevertheless it can occur to him, if he examines the question, or it may be
suggested by someone else, and he will never be safe from it, unless he first acknowledges God.
(HRPP, 236)

I say that this is a crucial passage for two reasons. In addition to displaying the two types of
knowledge taken over from Descartes by Sosa, it illustrates the deep structural similarities
between Sosa’s view, virtue perspectivism, and Descartes’s overall approach. As Sosa writes:

In structure, virtue perspectivism is thus Cartesian, though in content it is not. Radical

rationalism admits only (rational) intuition and deduction (along with memory) as its faculties of
choice (or anyhow of top choice) and wishes to validate all knowledge in terms of these
faculties; thus the Cartesian grand project. Virtue perspectivism admits also perception and
introspection, along with intuition and deduction, as well as inductive and abductive reasoning.
(RK, 423)

Virtue perspectivism is structurally Cartesian because it seeks to show that we have cognitio, not
through rational theology as Descartes attempted, but through other means that will be discussed
below. Thus, developing a response to philosophical skepticism is an integral part of virtue
perspectivism.
Sosa takes philosophical skepticism to be the view that there is “no way to attain full
philosophical understanding of our knowledge. A fully general theory of knowledge is
impossible” (PSEC, 93). Thus, the skepticism that Sosa has in mind is a form of Academic
Skepticism rather than a form of Pyrrhonian Skepticism that would refrain from assenting to
either the view that we can attain a full philosophical understanding of our knowledge or that we
cannot attain such an understanding. In this chapter I want to consider Sosa’s reasons for
rejecting Academic Skepticism because they strike me as less than compelling. I don’t think the
reasons for Academic Skepticism are compelling either (and have argued for that elsewhere4 ).
Thus, I side with the Pyrrhonians on this issue. But I won’t defend Pyrrhonism here – at least not
directly. The issue here is whether Sosa has found a good response to Academic Skepticism.

As I see it, Sosa’s response to Academic Skepticism has two main tenets:

(T1) There is no good purely internalist response to philosophical (i.e., Academic) skepticism,
where “internalism” is understood as the view that appropriate reasons, and only such reasons,
can convert a true belief into reflective knowledge.

(T2) There is a viable externalist response where “externalism” is the denial of internalism.5

I think neither of these tenets has been established.

Consideration of (T1)

The argument for (T1) is that in order to develop an adequate response to Academic Skepticism
one would need a “legitimating account of absolutely all one’s own knowledge” and such an
account is impossible, from a purely internalist perspective, because such an

account admits only justification provided by inference or argument and, since it rules out
circular or endlessly regressive inferences, such an account must stop with premises that it
supposes or “presupposes” that one is justified in accepting, without explaining how one is
justified in accepting them in turn. (PSEC, 96)

In other words, pure internalist responses fall victim to the Pyrrhonian trilemma.6

Thus, by a “legitimating account” Sosa means an argument that is not circular and not “endlessly
regressive” and whose premises are not arbitrarily appealed to and whose conclusion is that the
person giving the argument has methods of arriving at beliefs which when properly deployed are
at least likely to produce knowledge. Sosa thinks this kind of argument can’t be given, but he
also thinks that once the reason for that is recognized, it is no longer bothersome that there is no
such internalist response available.

Sosa claims that an internalist argument which concludes that our methods of arriving at beliefs
are reliable will have to employ that proposition as a premise (overtly or covertly) and will then,
of necessity, beg the question. He likens the situation to the one in which we desire to locate a
saint who blesses all and only those who don’t bless themselves. (Shades of the barber who
shaves all who don’t shave themselves!) As he says, in the search for such a saint, we may “turn
up likely prospects each of whom eventually is seen to fall short, until someone ... reflects that
there could not possibly be such a saint, and this for evident, logical reasons” (PSEC, 109).

Thus, even though he claims that we could not produce such an argument without begging the
question (having eliminated justificational surds and endless regresses as appropriate ways of
responding to the skeptic), once we recognize that what is being requested by the skeptic is
logically impossible to produce, we can “go ahead and ‘beg’ the question against such a sceptic
(though ‘begging the question’ and ‘arguing circularly’ may now be misnomers for what we do,
since it is surely no fallacy, not if it constitutes correct and legitimate intellectual procedure)”
(PSEC, 111).

I want to ask three questions:

Is Sosa correct that the kind of anti-skeptical argument required, though circular, is not viciously
so?

Is Sosa correct that every attempt at providing a legitimating argument for the claim that we have
knowledge will either beg the question or lead us to a vicious infinite regress or be based upon
some arbitrary supposition?

Is Sosa correct that every legitimating response to academic skepticism must be in the form of an
argument powerful enough to justify us in believing that we have reflective knowledge?

It will probably not be surprising that I think the answer to each question is a qualified “no.” Let
us take them one at a time.

(a) What makes an argument circular is not easy to pin down and I have no well worked out
general account to offer here. Nevertheless, I do think we can say that the fallacy of circular
reasoning occurs when the conclusion of an argument is warranted only on the basis of the
transfer of warrant from the premises and the conclusion is used (either overtly or covertly) as
one of the premises. A fallacy has occurred because in order for the argument to transmit some
warrant to the conclusion, there must be some warrant already present in each of the premises
but there is no such prior warrant for the proposition that doubles as the conclusion. Note: This
does not rule out the possibility that there might be some form of legitimate reasoning both from
and to propositions that are mutually probability enhancing if those propositions have some
warrant from sources other than that provided by the reasoning in question. Such reasoning,
though circular in some sense, might augment the initial warrant of the propositions.7

So, there might be some room for “self-supporting arguments” (PSEC, 111) that are not
viciously circular. But I take it that (i) the kind of reasoning employed by the antiskeptic
envisioned by Sosa is designed to produce some warrant for the proposition, our methods of
arriving at beliefs are reliable when properly employed, and further, (ii) Sosa thinks that the
anti-skeptic internalist has no means of arriving at that conclusion without employing it as a
premise. That is why he holds it to be logically impossible for such an anti-skeptic to avoid
begging the question. But, if such an anti-skeptic has no warrant for that proposition independent
of the reasoning offered and that proposition is employed as a premise, then the reasoning is
fallacious and, pace Sosa, the anti-skeptic ought not to be allowed to “go ahead and beg the
question” against the skeptic because this form of arguing is surely fallacious since it does not
constitute “correct or legitimate intellectual procedure.”

Now, to (b). I think that there are many arguments that conclude with the appropriate general
claim about the reliability of belief processes that are not instances of circular reasoning.

Consider this argument:

(1) My methods of arriving at beliefs are M-type methods.

(2) M-type methods are reliable.
(3) My methods of arriving at beliefs are reliable.

Does that argument commit a fallacy of any sort? No. No premise employs the conclusion. And I
can see no reason why an argument for either premise must employ the conclusion in one of its
premises, etc.

Consider this argument:

(1) If some condition C holds, then my methods of arriving at beliefs using M-type methods is
reliable.
(2) Condition C holds.
(3) My methods of arriving at beliefs using M-type methods are reliable.

Does that argument commit a fallacy of any sort? No. No premise employs the conclusion. And I
can see no reason why an argument for either premise must employ the premise in one of its
premises, etc.

So, why think that for “evident, logical reasons” every argument to the conclusion that my belief
acquisition methods are reliable must be circular? I think the answer is that any sufficiently
careful and self-conscious person who believes that his/her use of M-type belief acquisition
methods are reliable will employ those very methods in acquiring the belief in the conclusion.
But that is not a fallacy of circular reasoning. It is merely making one’s practices consistent with
one’s beliefs. If I thought that my reliable belief acquisition methods were M-type methods, it
would be imprudent of me, to say the least, to use some other method of acquiring beliefs – no
matter what the beliefs were. Indeed, I would legitimately be accused of an inconsistency
between my practices and my beliefs.

Now, perhaps this is what Sosa meant when he said that it might be a “misnomer” to call the
argument he had in mind an instance of circular reasoning because there is no fallacy involved. If
this is what he meant, we are in complete agreement here. An argument to the conclusion that
my belief acquisition methods are reliable need not be circular.

But avoiding that horn of the Pyrrhonian trilemma only throws us back on the other two horns.
Must internalists embrace arbitrary foundations or fall victim to a vicious regress of reasons? I
agree with Sosa that foundationalism is not a viable theory for explaining what provides a belief
with the suitable epistemic quality to pass muster as knowledge, and I have argued for that
elsewhere.8 I realize that to reject foundationalism without giving an argument here might be
considered brash, to say the least. But I hope it is understandable in the context of this
discussion. So, that leaves us with the regress.

I have argued elsewhere that some infinite regresses of reasons are not vicious.9 Here, I merely
wish to consider Sosa’s reasons for thinking that all such regresses are vicious. That is important
in the context of this discussion, since if he has not provided us with a good enough reason for
thinking that all such regresses are vicious, then there remains the possibility of a purely
internalist response to Academic Skepticism. Here is what he says:

the main reason for accepting formal foundationalism [as against infinitism] ... is the very
plausible idea that epistemic justification is subject to the supervenience that characterizes
normative and evaluative properties generally.... Epistemic justification is supervenient. The
justification of a belief supervenes on such properties of it as its content and its basis (if any) in
perception, memory or inference. (RP2, 142)

Now I am not certain that the normative does generally supervene on the non-normative or even
that this way of understanding properties cuts nature and non-nature correctly at the joints, but let
us assume that it is essential that a good account of justification must not preclude such
supervenience.10 The question before us then is this: Why think that the supervenience of the
normative on the non-normative is precluded if a proposition is justified if and only if there is an
infinite set of non-repeating propositions each of which serves as a reason for an earlier one in
the chain?11 Probably because it is thought that since the justificatory status of one proposition
obtains if and only if there is another proposition with the same justificatory status, ad infinitum,
the justificatory status of any proposition depends upon the justificatory status of another in a
sense of “depends upon” that rules out the possibility that justification supervenes on some non-
normative properties.

But that is an unwarranted inference. To see that, suppose that the following (contrary to fact)
were true: A child is a genius if and only if its parents are geniuses. If that were true, would it
follow that every account of what it is to be a genius must depend upon appealing to geniusness?
No. We could say that a person being a genius depends upon their scoring above a certain level
on some designated IQ test. That is compatible with the infinite regress of genius transmission.
The same holds here. A proposition, p, might be justified if and only if there is another
proposition, q, that is justified and transmits to p whatever non-normative features in virtue of
which it is justified. In other words, whatever non-normative properties on which justification
supervenes can be carried along by one justified proposition to another justified proposition.
Thus, I think this argument by Sosa gives us no good reason for thinking that every infinite
regress of reasons is vicious.

In “Philosophical Scepticism and Epistemic Circularity,” he hints at another reason for thinking
that such regresses are vicious. There he says when considering the resources available to an
internalist for responding to skepticism:
the justification of any given belief requires appeal to other beliefs that constitute one’s reasons
for holding the given belief. Of course, when one combines this with rejection of circularity, the
case for scepticism is very strong, assuming that for limited humans an infinite regress of reasons
or justifications is out of the question. (PSEC, 100)

This paragraph seems to suggest that the demand that there be an infinite number of distinct
reasons available for our beliefs cannot be satisfied because we are “limited humans.” But I think
this misconstrues what an infinitist must require. The infinitist need not require that in order for
us to have a good enough reason for believing a given proposition, we have traversed an infinite
path of reasons – that would, indeed, be impossible for a mortal being, who, like us, can only
entertain a finite number of propositions at any given time. That requirement is tantamount to
insisting that we complete the process of justifying a proposition in order for us to have good
enough reasons for believing it. And the infinitist would certainly eschew that requirement.
Rather, the infinitist could require only that there be an infinite set of distinct reasons available
for justifying the proposition and that we actually have some good reasons for the proposition
that are not ultimately overridden by our currently held reasons against it. The latter is sufficient
to make it more reasonable to believe than withhold. Assenting to it – that is, thinking that we
know the proposition to be true – might require having completed the process of justifying the
proposition, and that is why a Pyrrhonian infinitist would not assent to any proposition that
requires a justification. Tentative, provisionally justified belief – yes; assent – no.

In partial summary: I think Sosa’s reasons for holding that the regress would be vicious are
inadequate for motivating the conclusion that there is no viable internalist response.12 Such a
response can hold both that a proposition is justified if and only if there are the appropriate
reasons for it while also maintaining that epistemic properties supervene on non-epistemic ones.
Further, our “finite minds” pose no obstacle to requiring that the set of such reasons is infinite
and non-repeating.

But that being said, is it even required that, in order for the internalist to have a “legitimating
response” to the Academic Skeptic, she have some sort of argument that concludes with the
claim that we are capable of cognitio? That brings us to question (c).

(c) If “legitimating response” is being used as term-of-art and is defined in such a way that only
the possession of such an argument would qualify, then the claim would be true by decree and
lose its interest. I think, however, there is a deep difference here between what a Cartesian and
Sosa would find to be a legitimating response to Academic Skepticism and what a Pyrrhonian
and I would find to be a legitimate response. The Academic Skeptic is claiming that we do not
have cognitio. The Cartesian-like philosopher is claiming that we do.

The Pyrrhonian alternative is to withhold assent to the proposition that we have knowledge and,
consequently, to its denial. The Pyrrhonians would try to foster the habit of withholding assent to
the proposition that we have reflective knowledge by one of two methods. The first would be to
provide some general reasons for thinking that all stated arguments, and a fortiori, all stated
arguments for Academic Skepticism or for its denial either beg the question or rest on arbitrary
assumptions.13 Thus, even if there were a non-circular, non-vicious infinite series of reasons for
the claim that we have knowledge, we could never state that argument and so we ought to
withhold assent to the proposition that we have reflective knowledge. The second method would
be to carefully consider both the arguments for Academic Skepticism and those for its denial in
an effort to show that they are equally good or equally bad. In my defense of infinitism
(mentioned earlier) I do argue for the general claim. But, once again, this is not the place to
rehearse that general defense. However, I would like to take up the second alternative. So let us
begin with the specific argument for Academic Skepticism that Sosa seems to think is the most
powerful.

Here is his account of that argument:

A version of the [skeptic’s] argument may be laid out as follows:

1 All your sensory experience and information at t is compatible with your dreaming at t.
2 So you need some test which “indicates” that you are not then dreaming, such that you know
both (i) it is satisfied and (ii) if it is satisfied then you are not dreaming.
3 But how could such a test ever be available to you, if it is a condition of your knowing
(perceptually) anything beyond your experience that you know yourself not to be dreaming? ...

This nowhere assumes that in order to know that p one must be absolutely and justifiably certain
that p. We are not setting the standards too high and then complaining that we can’t possibly
measure up. Our argument is designed to show rather that, when compared with the possibility
that we are just dreaming, our thought that we really see is based upon no good reason whatever.
A sufficient reason must enable us to rule out alternatives that clearly would preclude our
knowing, as above [emphasis added]. The dismissal of the skeptic as setting the bar too high is
superficial because it overlooks this argument. (BIF, 142)

There is much in this diagnosis of the argument for Academic Skepticism with which I agree.
For example, the argument does not appeal to a “too high standards” requirement but it does
appeal to a version of the “eliminate all contraries” principle.14 The question before us is this:
Does the Academic Skeptic have the right to insist that I must eliminate all incompatible
propositions to p before we can attain scientia that p? I think Sosa and I agree that he/she does
not. (See BIF, 141) But, then, there is a legitimating internalist response to Academic Skepticism
available: namely, the Academic Skeptic has not given us a good reason for thinking that we do
not have cognitio.

That response could begin by pointing out that a consequence of the “eliminate all contraries”
principle is that in order to know that p our reasons for p must entail p. And that strikes me as a
burden that no internalist need accept.

To see that this is a consequence of adopting that prerequisite, note that both (~p & q) as well as
(~p and ~q) are contraries of p. Thus, our evidence for p would have to include both ~(~p & q)
and ~(~p & ~q). But that set of propositions entails p. So, the requirement that we eliminate all
contraries of p before we have any form of knowledge (scientia or cognitio) amounts to requiring
that our evidence entail p.15 That is a requirement we need not accept especially if we are
coherentists who employ both abduction and induction. Thus, it strikes me that the argument for
Academic Skepticism which Sosa finds most plausible is based upon a view of the requirements
for having adequate reasons or, in his terms, for employing a virtuous process of reasoning, that
sets inappropriate conditions on having good reasons for our beliefs.16

But suppose that the requirement is weakened. Perhaps the only alternatives we have to eliminate
are those which are known to be incompatible with p? That won’t help because if we are
circumspect and informed about what is incompatible with our knowing that p, we would have to
eliminate both of the contraries of p mentioned above, since we know them to be incompatible
with p.

Suppose that we weaken the requirement even further and require that the alternatives that have
to be eliminated are known to be incompatible and are mentioned or thought of in the context?17
Even this weaker principle seems incorrect. Recall Dretske’s oft-cited Zebra-in-the-Zoo Case.18
We are standing before some zebra-like looking animals and someone says, “How do you know
that these animals are not very cleverly disguised Seventh-Day Adventists who dress up as
zebras in order to catch you off guard so that they can begin their spiel?” Or “How do you know
that these are not aliens from another galaxy who have disguised themselves as zebras in order to
observe large numbers of humans as they visit the zoo?” Pretty silly. Pretty unmotivated. Now, if
Seventh Day Adventists or aliens have resorted to such disguises, or perhaps, even if you just
thought they had, then such an alternative might need to be eliminated with epistemic “priority,”
as Sosa puts it (CEP). But in the absence of either of those obtaining, even this weakened
requirement seems too restrictive.

Perhaps it could be claimed that although it is not normally required that we eliminate all known
and raised alternatives, it is required by anyone seeking scientia? Doing philosophy brings with
it unusual epistemic burdens. Keith Lehrer writes as follows:

generally arguments about where the burden of proof lies are unproductive. It is more reasonable
to suppose that such questions are best left to courts of law where they have suitable application.
In philosophy [emphasis added] a different principle of agnoiology [the study of ignorance] is
appropriate, to wit, that no hypothesis should be rejected as unjustified without argument against
it. Consequently, if the sceptic puts forth a hypothesis inconsistent with the hypothesis of
common sense, then there is no burden of proof on either side.19

The point here could be put in our present terminology as follows: When we are seeking scientia,
we put on the cloak of a philosopher and, then, it becomes appropriate to require that we
eliminate all known and raised alternatives. Recall Descartes’s method of “pretending” that all
his former opinions were false in order to arrive at what, if anything, is certain. A question, then,
becomes this: Is that an appropriate requirement when we are doing philosophy? Of course, we
could stipulate that it is appropriate. But then the results of employing such a requirement could
be ignored by anyone not agreeing to that precondition. I think the more interesting questions are
these: Why should anyone think that the mere raising of the skeptical scenario imposes any
obligation upon us to eliminate it in order to have the highest form of human knowledge? Where
does that obligation come from? Not our ordinary practices. From where, then?

Presumably the obligation originates from some reflections about our philosophical practices,
and in particular, from examining the Cartesian-like goal mentioned above of settling, in so far
as possible, whether we have cognitio. Interestingly enough, that is a goal shared by the
Academic Skeptics as shown by their adherence to some version of the elimination principle. (I
might note parenthetically that this is why the Pyrrhonians considered both to be dogmatists.)
But if the practice of philosophy includes Pyrrhonism, namely, continuing to inquire about
matters that were not self-evident (“skeptic” does come from the Greek “σκε´πτομαι” meaning to
inquire), then the obligation to eliminate all known and raised incompatible alternatives would
likely disappear. Why should we be obliged to eliminate alternative hypotheses about the scope
of our knowledge for which we have no reason whatsoever? Thus, I think there is a “legitimate
response” to Academic Skepticism that does not require us to have an argument powerful enough
to fully justify us in believing that we have cognitio-style knowledge. We can come to see that
the argument for Academic Skepticism imposes obligations on us that need not be accepted.

Now, this is not to gainsay the possibility that a legitimate question could arise about the general
reliability of our belief acquisition and sustaining methods.20 As I have argued above, I see no
reason why we can’t give non-circular arguments to the conclusion that those methods are
generally reliable. Those arguments would, of necessity, employ those methods. Further, those
arguments would initiate a process of reasoning that is unending and further inquiry would
always be in order for there would always be one more step to take. But that is a prospect that
can be gladly embraced once the Cartesian-like goals are put aside.

Consideration of (T2)

It is now time to turn directly to (T2), namely Sosa’s suggested response to the Academic
Skeptic. As I see it, the essential step involves the Principle of Ascent (PA – mentioned in “Two
False Dichotomies”). If that principle were true, it would allow us to ascend to a meta-level and
be justified, on coherentist grounds, in believing that we do indeed have cognitio. That is, we can
ascend and then assent. If correct, it would provide a basis for rejecting Academic Skepticism
and at the same time it would lead to the rejection of Pyrrhonism since it would provide the basis
for assenting to a hitherto non-evident proposition. Here is that principle:

(PA) [Ksp &CsKsp] → JsKsP

That is, if S knows that p and S considers whether he/she knows that p, then S is justified in
believing the proposition that he/she knows that p. I say it provides a basis for rejecting
Academic Skepticism because, if true, iterative reflective knowledge is made possible and even
necessary if we consider whether we do have first-order reflective knowledge. For if one knows
that p and considers whether one knows, and one is justified in (actually) believing that one
knows, then, according to Sosa, one knows that one has knowledge (barring defeaters, of
course). Thus, a limited version of the KK thesis holds.21

What makes Sosa think PA is true? Here is what he says in “Two False Dichotomies”:

Suppose, first, one consciously and occurrently believes that p, and, in that same specious
present, second, one consciously and occurrently considers whether one not only believes but
knows that p. Exactly three options open up: one might say either (a) “No, I don’t know that,” or
(b) “Who knows whether I know it or not; maybe I do, maybe I don’t,” or (c) “Yes, that is
something I do know.” One is better off, surely, if able to give the later answers; better off with
the second answer than the first, and better off yet with the third. Answer a, and even answer b,
would reveal a certain lack of integration in that stretch of consciousness; only answer c, of the
three, entirely avoids disharmony within that consciousness in that specious present. If one has to
give answer a, or even answer b, one thereby falls short, and one’s belief that p falls short. It is
not all that it could be.

The more general claim about any belief, p, in the above citation is particularized in the
following passage from “Philosophical Scepticism and Epistemic Circularity”:

But now suppose that by using way W of forming beliefs . . . we arrive at the conviction that W
is our way of forming beliefs. Now, so long as we do not go back on that conviction, does that
not restrict our coherent combinations of attitudes? Take: (e) B:[W is my overall way of forming
beliefs]. And compare (f) B:[W is reliable], (g) D: [W is reliable] and (h) Wh:[W is reliable]. It
is not evident that (e) & (f ) would be more satisfyingly coherent than either of (e) & (g) or (e) &
(h)? (PSEC, 107)

I cite these passages at length not only because they are crucial in the argument for PA and in
understanding Sosa’s argument that we can know that we have cognitio, whenever we reflect on
whether we do, but also because it underscores the deep difference between Sosa and me,
mentioned above, concerning our powers to justify beliefs. His model is the Cartesian; mine is
the Pyrrhonian.

The first thing to note about the argument for PA is that the conjunct, “Ksp,” plays no role. The
argument is that it is simply more coherent in the specious present to believe that my beliefs are
knowledge than to deny or withhold that they are knowledge. The quick response to this is that it
strikes me as at least bordering on what the Pyrrhonians would have thought was dogmatism.
From the mere fact that I believe something, is it more coherent to believe that I also know it? I
say that is the “quick” response. I would now like to consider PA a bit more carefully, but I
admit that my “long” response will merely be an elaboration of this initial reaction.

To begin that more careful response, let us grant the privileged access to our occurrent belief
states presupposed in these passages. The issues, though, are these: Is it more “satisfyingly
coherent” to believe that I know that p whenever I believe that I believe that p? Is that the only
alternative among the three possible epistemic attitudes (believe, deny, withhold) that “avoids
disharmony?” I think the answer to those questions is “It depends upon what else I believe in the
same specious present.”

Now exactly what falls under the scope of all the beliefs of mine in the specious present is
difficult to determine. Perhaps I can only have one (perhaps compound) occurrent belief at a
time. But in that case, the specious present might not contain both the belief that p and the belief
that I know that p unless one’s account of JKp was such that it contained Bp. That strikes me as
plausible but not obviously correct. The more important consideration is that if the specious
present didn’t include a significant portion of the beliefs relevant to p, it is difficult, if not
impossible, to conceive of “coherence” as doing much work in a theory of justification. With
what set of beliefs is a particular one, say p, supposed to cohere in order to be justified? It must
be more than p’s constituent beliefs. So, let us assume that in whatever way relevance is
determined, a significant portion of my most relevant beliefs about some subject matter, p, are to
be included in the scope of the beliefs to be considered when the question of what best coheres
with p arises.

To see why I think that the options for achieving coherence among my beliefs are not as Sosa
thinks are “surely” correct, let’s begin by considering my belief about the reliability of methods
W employed by someone else in arriving at their beliefs. Suppose that I’m at a carnival and, in
particular, at the booth where the carnival employee “guesses” a person’s weight. The rules of
the “game” are simple: Someone gives the carnie 25 cents and that someone gets a panda bear
IFF the carnie misses by more than 3 lbs of the person’s correct weight. I hear the carnie say to
someone, “You weigh 185.” Assuming he/she believes what they said, am I justified in believing
that the carnie knows that the person weighs 185 lbs?22 Yes, because I think that the carnie has
employed some special skill – a skill not had by just any person – in determining people’s
weights.

On the other hand, suppose that I overhear someone – let’s call that person “Anyperson” – say to
someone else, “You weigh 185 lbs.” Let us assume that I think Anyperson’s skills in determining
weights is typical – that is, Anyperson is at least not more likely to be right than wrong. In that
case, would I be justified in believing that Anyperson knows that the person weighs 185 lbs? It
strikes me that the answer to that question is that it would be preferable on coherence grounds to
at least withhold such a judgment if not to deny that it is justified. Most people are not very good
at telling another’s weight within 3 lbs.

Now, suppose I am Anyperson and believe that someone weighs 185 lbs. I believe of myself that
I don’t have any special weight-guessing skills. I’m just like anyone else. I ask myself: Do I
know that the person weighs 185 lbs.? I believe the person’s weight is 185 – he looks to be about
Bob’s weight and Bob told me a while ago that he weighed 185. It seems to me that in this case
the most “satisfyingly coherent” attitude to ascend to is at least withholding assent about whether
I have knowledge.

A person with Cartesian leanings might wish that humans, and in particular they themselves,
were disposed not to believe something until they also believed (in the same specious present)
that the belief is known. And a Cartesian might feel some “disharmony” here. But must we be
Cartesians? The question here is whether if I discover a belief of mine it must be more
“satisfyingly coherent” to believe that such a belief rises to the level of knowledge than to
withhold assent to the proposition that it is knowledge. Before answering that question, wouldn’t
I first want to believe that I had figured out what the conditions of knowledge were and then
wouldn’t I want to believe that my belief satisfied them? Suppose that I thought that the
conditions of knowledge required something akin to eliminating all known and raised
incompatible propositions. Would I think I had done that? In particular, would I think that I had
eliminated the evil genius hypothesis? I, for one, do not see how to do that.

Note, my argument here does not in any way depend upon accepting the elimination
requirement. Indeed, I don’t accept it. The point here is that whether it is more coherent to
believe about an occurrent belief of mine that it is knowledge rather than have either of the other
alternative epistemic attitudes depends upon what else I believe. In particular, it depends upon
what I believe my skills are in determining the truth in matters that are not already evident and it
depends upon whether I thought that I could then conform my beliefs, or the particular belief in
question, to the results of that determination. If I thought that my skills were no better (or not that
much better) than those of Anyperson and that my epistemic virtues were no better (or not that
much better) than those of Anyperson, then merely discovering that I have a belief would not
provide a basis for thinking that the belief is knowledge. Indeed, I see no “disharmony” at all in
the mental states of persons who withhold believing that their beliefs rise to the level of
knowledge. Pyrrhonism strikes me as a viable “harmonious” option. When they ascend, they
don’t assent.

Acknowledgments

I would like to thank Anne Ashbaugh, John Greco, Brian McLaughlin and especially Ernest Sosa
for their help with this paper. Needless to say some of those people don’t agree with all of the
claims that I make!

Notes

1 In this chapter I will be relying primarily on the following papers by Sosa: “Philosophical
Skepticism and Epistemic Circularity,” Proceedings of the Aristotelian Society 68 (1994), 263–
90 (cited as PSEC); “How to Resolve the Pyrrhonian Problematic: A Lesson from Descartes,”
Philosophical Studies 85 (1997): 229–49; “Reflective Knowledge in the Best Circles,” The
Journal of Philosophy 94 (1997): 410–30 (cited as RK); “The Raft and the Pyramid,” as
reprinted in Ernest Sosa and Jaegwon Kim, eds., Epistemology (Oxford: Blackwell, 2000) (cited
as RP2); “Two False Dichotomies: Internalism/Externalism and Foundationalism/ Coherentism,”
in Walter Sinnott-Armstrong, ed., Pyrrhonian Skepticism (Oxford: Oxford University Press,
2003); and “Circularity and Epistemic Priority,” in R. Schantz, ed., The Externalist Challenge:
New Studies in Cognition and Intentionality (de Gruyter, forthcoming) (cited as CEP).

2 I plan to discuss these issues in another paper.

3 I have a slight worry – that sometimes rises to the level of a nagging fear – that attributing the
highest form of human knowledge to those of us who are epistemologists might be a bit
jingoistic.

4 See my “Skepticism and Closure: Why the Evil Genius Argument Fails,” Philosophical Topics
23, 1 (Spring 1995): 213–36.

5 More specifically, Sosa’s form of externalism includes both the requirement that there is some
appropriate hook-up between our beliefs and the world so that it is not accidental that our beliefs,
however coherent, are true when justified, as well as a requirement that there be broad coherence
among our beliefs. Thus “externalism” can include some internalist features but it is
incompatible with pure internalism that holds that the epistemic quality of our beliefs depends
completely on the reasons we have for them.
6 See Sextus Empiricus, Outlines of Pyrrhonism, I, 166–9.

7 Sosa does give some examples of the kind of argument that might be circular in some sense but
not vicious. See CEP: “Is there no difference between having and lacking an ‘affirming’
perspective on oneself, on the surrounding world, and on the relations between the two, a
perspective on the basis of which one can coherently endorse one’s own beliefs? Is such a
coherent picture of things of no value simply because it cannot possibly be elaborated while
avoiding all circularity? This should seem the opposite of obvious to a philosopher, since it
discounts a main perennial objective of philosophical reflection: namely, the elaboration of such
a coherent, ‘affirming’ view of oneself and one’s place in the scheme of things. What is more,
the member beliefs of such a system are enhanced precisely through being part of the integrated
view and not just loose from other beliefs in a less integrated mind. So these member beliefs
attain a higher epistemic grade, a higher grade of knowledge, reflective knowledge, than they
would attain otherwise, other things being equal.”

8 I agree with his reasons and have offered some of my own in “Human Knowledge and the
Infinite Regress of Reasons,” in J. Tomberlin, ed., Philosophical Perspectives, vol. 13 (Malden,
MA/Oxford: Blackwell, 1999), pp. 297–325; “Why Not Infinitism?” in Richard Cobb-Stevens,
ed., Epistemology: Proceedings of the Twentieth World Congress in Philosophy (Bowling Green,
OH: Philosophy Documentation Center, 2000), vol. 5, 199–208; “The Failures of Dogmatism
and a New Pyrrhonism,” Acta Analytica 15, 24 (2000): 7–24; and in “Academic and Pyrrhonian
Skepticism,” to appear in Steven Luper, ed., The Skeptics: Contemporary Essays (Ashgate
Press).

9 See the papers cited in note 8, above.

10 Could it not be that some ancestor properties are essentially neither normative nor non-
normative and that both “emerge” under various circumstances? I am not here endorsing this
view. Rather, I merely wish to point to the importance of the assumption that all properties are
either normative, non-normative, or both.

11 Strictly speaking, the existence of such a chain is not sufficient for a proposition’s being
justified. There must also not be reasons subjectively available (i.e., either a reason which S
currently has or one that is appropriately epistemically hooked-up with S’s currently held
reasons) that would override a step in the chain of reasons. But that is not the admission of an
externalist feature into an internalist account as Sosa views internalism and that complication can
be ignored here.

12 I have discussed objections to infinitism, including the two that Sosa presents, more fully in
the papers cited in note 8, above, as well as in “When Infinite Regresses are Not Vicious” to be
given at the Pacific APA, March 2001, and published in Philosophical and Phenomenological
Research 66 (2003): 718–29.

13 The qualification, namely “stated” arguments, leaves open the possibility that there is an
infinite regress of reasons on either or both sides of the issue. Of course, if that were the case,
then it would not be settled that either Academic Skepticism or its denial is the correct view to
hold.

14 See Certainty, a Refutation of Skepticism (Minneapolis: University of Minnesota Press,

1981); and my “Skepticism and Closure: Why the Evil Genius Argument Fails.”

15 See Certainty, esp. pp. 100–4.

16 Here I would identify my views closely with those of Michael Williams in Unnatural Doubts
(Oxford: Basil Blackwell, 1991).

17 Some contextualists have held that merely raising the various alternatives is sufficient to
make them such that they ought to be eliminated. See David Lewis in “Elusive Knowledge,”
Australasian Journal of Philosophy 74 (1996): 549–67. It is also endorsed by Keith DeRose in
his “Solving the Skeptical Problem,” Philosophical Review 104 (1995): 1–52 and his
“Contextualism and Knowledge Attributions,” Philosophy and Phenomenological Research 52
(1992): 913–29.

18 The Zebra-and-the-Zoo case is due to Fred Dretske. See his “Epistemic Operators,” The
Journal of Philosophy 67, 24 (1970): 1015–16.

19 Keith Lehrer, “Why Not Scepticism?” in Pojman, ed., The Theory of Knowledge (Belmont,
CA: Wadsworth Publishing, 1993): 53. Lehrer no longer seems to hold this view, or at least, it is
highly qualified since he now believes that we can use “I see a zebra” in our response to the
skeptic who posits this objection, “You are asleep and dreaming that you see a zebra.” See
Lehrer, Theory of Knowledge, 2nd edn. (Boulder, CO: Westview Press, 2000): esp. 133–4 and
chapter 9. Thus, he is claiming that we can use the very proposition that is being challenged to
rebut the challenge. In that case, it does appear that we must believe that our current belief (or
“acceptance” as Lehrer would prefer) can be used to rebut a challenge to it and, hence, there is at
least an implicit implication that our current beliefs are epistemically privileged over a challenge
to them.

20 Perhaps we find ourselves regularly believing p and then not-p. Perhaps we share the worry
that Descartes raised, namely that “nevertheless in whatever way they suppose that I have arrived
at the state of being that I have reached . . . since to err and deceive oneself is a defect, it is clear
that the greater will be the probability of my being so imperfect as to deceive myself ever, as is
the Author to whom they assign my origin the less powerful” (Descartes’s “First Meditation” in
The Philosophical Works of Descartes, trans. Haldane and Ross (New York: Dover Publications,
1931): 147).

21 Note that PA does commit Sosa to an infinite regress of propositions of increasing complexity
for which S has a justification. He is not committed to S’s being able to believe them – since S
might not be able to consider propositions with many iterated knowledge predicates.
22 By “justified” here I mean to be speaking somewhat casually. That is, I do not mean to be
endorsing the view that it is settled that the carnie’s belief is justified (in the sense that there is a
completed justification available to him).
Ernest Sosa
Ernest Sosa and His Critics

What is required in order to know that p, let us suppose, is that one be able to answer potential
objections to the belief that p. And consider now a belief whose presence and truth are required
for such successful defense against a potential objection. Objections to this belief would
themselves seem objections to the original belief, as held then by that believer. But this threatens
a vicious regress. For the system must now contain not only true beliefs that defend against first-
level objections to the original belief, but also true beliefs that defend against (second-level)
potential objections against the first-level defensive beliefs. And so replies on. No belief will
hence amount to knowledge unless it is embedded in an infinitely complex actual corpus. And
this lies beyond human capabilities.

So my doubts about the proposal are (a) that it reduces to the trivial alternative that one knows
that p iff one knows that p; and (b) that if in some non-ad hoc way Habib and Lehrer are able to
rule out of consideration any such objection as “But you don’t know that!”, so as to avoid the
triviality fate, they still face a potentially vicious regress. For consider one’s defending beliefs
against objections to one’s belief that p, and call these “defenders.” The problem is that
objections to defenders would seem to count indirectly as objections to the original belief, so that
a full defense of the original belief against all potential objections to it would require a mind
housing an actual infinity of defending beliefs. None of our minds would seem spacious enough
for that…
Andrew D. Cling
The Trouble with Infinitism

ABSTRACT

One way to solve the epistemic regress problem would be to show that we can acquire
justification by means of an infinite regress. This is infinitism. This view has not been popular,
but Peter Klein has developed a sophisticated version of infinitism according to which all
justified beliefs depend upon an infinite regress of reasons. Klein’s argument for infinitism is
unpersuasive, but he successfully responds to the most compelling extant objections to the view.
A key component of his position is his claim that an infinite regress is necessary, but not
sufficient, for justified belief. This enables infinitism to avoid a number of otherwise compelling
objections. However, it commits infinitism to the existence of an additional feature of reasons
that is necessary and, together with the regress condition, sufficient for justified belief. The
trouble with infinitism is that any such condition could account for the connection between
justification and truth only by undermining the rationale for the regress condition itself.

1. INTRODUCTION

The epistemic regress problem is that the possibility of knowledge, or even justified belief,
seems incompatible with the requirement that knowledge and justified belief depend upon the
support of propositional evidence. For this condition can be satisfied only if either beliefs can
figure essentially in their own support or support requires an infinite regress of distinct beliefs.
So if justified beliefs cannot figure essentially in their own support and cannot depend upon an
infinite regress, then knowledge and justified belief are impossible.

One way to solve this problem would be to show that we can acquire justified belief by means of
an infinite regress. This position – infinitism – has not been taken seriously by many
philosophers and has been adopted by even fewer. Recently, however, an important
contemporary epistemologist, Peter Klein, has developed a sophisticated version of infinitism,
arguing that it provides a superior alternative to foundationalist, coherentist, reliabilist, and other
popular accounts of justified belief.1 To assess infinitism, I shall examine the strengths and
weaknesses of Klein’s position.

Klein gives an argument for infinitism and responds to objections that have been made to the
view. I shall show that although Klein’s argument for infinitism is unpersuasive, he is able to
avoid the most powerful extant objections to the view. A key element of Klein’s infinitism is his
claim that an infinite regress of reasons is necessary, but not sufficient, for justified belief. This
claim allows him to sidestep some of the otherwise compelling objections to infinitism.
However, it also commits him to the existence of an additional necessary condition of justified
belief. The trouble with infinitism is that any such feature must be positively relevant to the truth
of justified beliefs and this undermines the rationale for the regress condition itself.

2. INFINITISM
Infinitism is the view that there can be a justified belief the justification for which depends upon
an infinite sequence of distinct propositions each of which is supported by its successor. Peter
Klein defends a strong version of infinitism. He thinks that justified beliefs are possible and that
all justified beliefs require an infinite regress of reasons. Klein argues for his regress condition
using claims about the conditions on justification.

Klein supports his view that justified belief requires infinite justificatory regresses with two
principles, the “Principle of Avoiding Circularity”:

(PAC) For all x, if a person, S, has a justification for x, then for all y, if y is in the evidential
ancestry of x for S, then x is not in the evidential ancestry of y for S.2

and the “Principle of Avoiding Arbitrariness”:

(PAA) For all x, if a person, S, has a justification for x, then there is some reason, r1, available to
S for x; and there is some reason, r2, available to S for r1; etc.3

Taken together, (PAC) and (PAA) imply that justified belief in any proposition requires an
infinite regress of reasons. For (PAA), as Klein glosses it, implies that any person S can have
justified belief in a proposition P only if there is a proposition R1 such that R1 is a reason for P,
in some objective sense, and R1 is such that, given S’s doxastic predicament, S is able to “call
on” R1 as a reason for P.4 So (PAA) says that a person S can have justified belief in a proposition
P only if there is objective propositional evidence for P to which S has access, and that a
proposition can provide propositional evidence for another only if there is propositional evidence
for it, and so on. (PAC) says that a justifiably believed proposition cannot be in its own
evidential ancestry. So justified belief in any proposition requires an infinite regress of distinct
propositions. Specifically, the conjunction of (PAC) and (PAA) implies Klein’s regress
condition:

(K) For any person S and any proposition P1, S justifiably believes P1 only if P1 is the first
member in an infinite sequence of distinct propositions σ (P1, P2,...) such that every member Pn
of σ is such that its successor Pn+1 is available to S as a reason for Pn.

Klein must distinguish (K) from this trivial logical regress condition:

(L) For any person S and any proposition P1, S justifiably believes P1 only if P1 is the first
member in an infinite sequence of distinct propositions σ (P1, P2,...) such that every member Pn
of σ is such that its successor Pn+1 implies Pn.

If Pn+1 is available to S as a reason for Pn only if Pn+1 deductively or inductively implies Pn,
then (K) implies (L), but not vice versa. However, since any proposition is the first member of an
infinite sequence of distinct propositions such that every member is deductively or inductively
implied by its successor, any belief satisfies (L).5 So the infinitist needs additional conditions on
justified belief to distinguish regresses that are justification-affording – those upon which actual
justified beliefs depend – from regresses that are not justification-affording.
 3. (PAC) AND (PAA) ARE DOUBTFUL

Klein’s argument for the regress condition is doubtful. Klein supports (PAC) by claiming that it
expresses “an obvious presupposition of good reasoning”.6 Klein supports (PAA) by giving a
general argument against epistemologies that are incompatible with it: reliabilist theories,
traditional foundationalist and non-traditional coherence theories, and non-evidentialist
epistemologies according to which supporting propositions need not be supported.7 The linchpin
of Klein’s criticism of such theories is his view that any belief P that is justified either has a
property Ф that explains P’s being justified or it does not. If P has Ф, then P has propositional
support since Ф must be indicative of the truth and thus the fact that P has Ф is itself a reason to
believe that P is true. If, however, P does not have Ф, then it is arbitrary.8 Neither of these lines
of support is persuasive.

(PAC) is false, as I have argued elsewhere, for there are justification-affording arguments in
which propositions figure essentially in their own evidential ancestry.9 Here I shall simply note
that the epistemic defects of direct self-support – the kind of support at work if C supports C –
need not afflict extended chains of support in which propositions figure further “upstream” in
their own evidential ancestry, for example cases in which C supports P and P supports C. The
problem with direct self-support is neither logical nor factual. Lee was the victor at
Chancellorsville entails that Lee was the victor at Chancellorsville, and Lee was the victor at
Chancellorsville. Any circular argument will be doxastically and pragmatically ineffective since
one must already accept the conclusion in order to accept the premise. As a result, no one can
come to accept the conclusion, or be persuaded to accept the conclusion, of a circular argument
as a result of reasoning from the premise to the conclusion.10 Still, the doxastic and pragmatic
limitations of such arguments need not blunt their epistemic force since the fact that an argument
cannot persuade those who doubt its premise to accept its conclusion does not imply that the
premise does not provide evidence for the conclusion. The epistemic problem with immediate
self-support – P supports P – is that it cannot provide a person S with new evidence for P. Thus
persons cannot come to be justified in believing that P or enhance their justification for believing
that P by inferring P from itself. Since extended chains of support in which the terminal
conclusion C figures in its own evidential ancestry – for example, cases in which C supports P
and P supports C – do not have this limitation, it is not obvious that they cannot be probative. So
Klein is wrong to think that (PAC) is obviously true.

Klein can sidestep this criticism, however, for his argument for his regress condition does not
require (PAC). Klein need not claim that no proposition can be in its own evidential ancestry,
only that no belief can be such that its justification depends upon this. Specifically, Klein can use
this weaker principle:

(PAC)∗ For all x, there is no person, S, such that S has a justification for x and S has a
justification for x only if x is in the evidential ancestry of x for S.

(PAC)∗ leaves it open that beliefs can enhance their justification by being in their own evidential
ancestry but blocks justification when this is essential.
Circularity aside, Klein’s argument for (PAA) is unsound. He is right to insist that there be a
principled distinction between justified and unjustified beliefs that is relevant to truth. He is
wrong to suppose that this distinction requires that justified beliefs depend upon propositional
reasons that are available to the believer. Klein argues for (PAA) by rejecting traditional
foundationalist theories, non-traditional coherence theories, and theories according to which
justification can be grounded in unsupported beliefs. According to all such theories, justification
does not always require propositional support. Foundationalists hold that some beliefs can be
justified without the support of other beliefs – by non-propositional features of sensory
experience, for example. Non-traditional coherentists hold that a person’s beliefs can acquire
justification from the coherence of the person’s system of beliefs as a whole. Still other theories
– contextualism and pragmatism, for example – hold that justified beliefs can ultimately be
supported by beliefs that are unjustified in any of these other ways. Despite their other
differences, all such accounts imply that there is a feature Ф that explains justification, but that Ф
does not require propositional support for all justified beliefs.

Klein’s argument for (PAA) turns on his claim that any feature that could explain the fact that
S’s belief that P is justified would be relevant to truth only if P has Ф were itself a reason for S
to believe that P. Thus, Klein:

Pick your favorite accounts of the property, [Ф]. I think ... that the old Pyrrhonian question is
reasonable: Why is having [Ф] truth-conducive? Now, either there is an answer available to that
question or there isn’t. (... ) If there is an answer, then the regress continues – at least one more
step, and that is all that is needed here, because that shows that the offered reason that some
belief has [Ф] or some set of beliefs has [Ф] does not stop the regress. If there isn’t an answer,
the assertion is arbitrary.11

To support (PAA), this argument must run as follows: (i) Any belief P either has a property Ф
that is relevant to P’s being true or it does not. (ii) If P does not have Ф, then P is arbitrary, but
(iii) if P is arbitrary, then P is unjustified for S. (iv) If P does have Ф, then the proposition P has
Ф is a reason for believing that P is true, and (v) if P has Ф is a reason for believing that P is
true, then there is a reason available to S for P. Therefore, (vi) for any belief P that is justified for
S, there must be a reason available to P for S.

This argument is unsound, for (v) is false. Indeed, (v) is undermined by an essential feature of
Klein’s own account of justified belief. Klein holds that beliefs are justified only if they are
supported by available reasons and that:

[t]here are two conditions that must be satisfied in order for a reason to be available to S. It must
be both “objectively” and “subjectively” available.12

Klein leaves the precise character of objective availability open – I shall say more about this
below – but a crucial feature of objective availability is that neither the fact that P implies C nor
the fact that S takes P to be a reason for C is sufficient for P’s being objectively available to S as
a reason for C. On the other hand, P is subjectively available to S as a reason for C only if C is
“properly hooked up with S’s own beliefs ....”13 P is subjectively available to S as a reason for C
only if, in some way, S has access to P as a reason for C. Klein need not claim that the subjective
availability of the successor of each member of an infinite regress of reasons requires that we
have an infinite number of occurrent beliefs, or even an infinite number of settled dispositions to
form occurrent beliefs. It is enough, according to Klein, that we have the second-order
dispositions to form the dispositions to have an infinite number of occurrent beliefs in the
relevant circumstances.14 Having such second-order dispositions does not even require that we
already have the concepts necessary to have the requisite beliefs, we need only the ability to
“develop new concepts and ways of specifying them”.15

The problem with Klein’s argument for (PAA) can now be made clear. Even given Klein’s weak
conditions on subjective availability, the most that his argument could show is that if a person S
is justified in believing a proposition P, then there will be a proposition Q that is objectively
available to S as a reason for P. By Klein’s own account, however, it does not follow from this
that Q will be subjectively available to S as a reason for P.

For suppose that Klein is right to think that an acceptable account of justified belief, when
applied to a given belief P, can provide a person with evidence for the truth of P. Suppose, that is
to say, that one way in which S might acquire evidence for P would be for S to decide whether or
not, given an understanding of S’s own epistemic predicament and knowledge of an epistemic
principle to the effect that P’s having Ф is sufficient for P’s being justified for S, S is justified in
believing that P is true. Under these conditions, the fact that S’s belief that P has Ф would be an
objective reason for S to believe that P is true since P has Ф, by implying that S is justified in
believing that P, would itself be evidence that P is true. Now suppose that S’s belief that P has
Ф. It follows, given our assumptions, that P has Ф would be an objectively available reason for
P, since the fact that S’s belief that P has Ф indicates that P is true. It does not follow, however,
that P has Ф is a reason for P that is subjectively available to S, for S need not have the
dispositions necessary to form the concept of Ф or to take P has Ф to be a reason for P, but this
is what Klein’s argument requires. Klein is simply assuming that factors relevant to justification
must be subjectively available to persons with justified beliefs. The history of epistemology
shows that properties that are sufficient for justification are likely to be quite complex, subtle,
and subjectively available to comparatively few persons, even persons who are otherwise quite
adept at marshalling other kinds of propositional evidence for their beliefs. So even if one way to
acquire evidence for P were to decide whether P has a justification-explaining property Ф, this
does not imply that P’s having Ф would itself be a reason, in the required sense, for every person
to believe any proposition that has Ф. So the mere fact that an epistemological theory must draw
a principled distinction between justified and unjustified beliefs does not show that every
justified belief must be supported by a reason.

Klein might be confusing the question of whether some beliefs can be justified without
propositional support with the question of whether there are special propositions that cannot be
supported by objectively available reasons. The first question is whether states of affairs that do
not require that propositional evidence for P be available to S can ever be sufficient for justified
belief in P. The second question is whether there are special propositions for which there can be
no propositional evidence. Even if it were true that there are states of affairs not requiring S to
have propositional evidence for P that are sufficient for S’s being justified in believing that P, it
would not have to be true that P be such that there can be no propositional evidence for it. For
example, even if foundationalists are right to think that some beliefs can be justified without
propositional support, it does not follow that there can be no objective propositional reasons for
such “immediately justified” beliefs. Nor does it follow that persons for whom such beliefs are
immediately justified can have no propositional evidence for them. Foundationalism implies that
some beliefs can be justified independently of propositional support, not that there are
propositions for which propositional support is, in some absolute sense, unavailable. For all
Klein’s argument shows, we might have propositional evidence for beliefs that we would be
fully justified in holding without that evidence.

I conclude that Klein has given no good reason to accept (PAC) or (PAA), and therefore that his
infinitism is unsupported. He gives no argument for (PAC) but, contrary to his claim, (PAC) is
not an obvious requirement of correct reasoning. I have shown, however, that Klein can
reformulate his argument in a way that avoids this criticism. The major problem with Klein’s
argument is that although he does give an argument for (PAA), it is unsound.

4. SOME INCONCLUSIVE OBJECTIONS

Infinitism might still be true. That Klein has not supported infinitism does not show that an
infinite regress of reasons is not required for justified belief or that no justified beliefs depend
upon an infinite regress of reasons. What is the trouble with infinitism? This turns out to be
difficult to say, for a carefully formulated version of infinitism such as Klein’s is resistant to a
number of popular objections.

Some philosophers, following Aristotle, argue that the trouble with infinitism is doxastic. They
claim that infinite justificatory regresses prevent us from having justified beliefs since it is
impossible for a finite mind to make an infinite number of inferences, believe an infinite
proposition, or have an infinite number of dispositions to have specific occurrent beliefs.16

Klein’s infinitism, however, does not require that anyone with justified beliefs be able to make
an infinite number of inferences, have an infinite number of occurrent beliefs, believe an infinite
proposition, or even to have an infinite number of dispositions to have specific beliefs. Klein’s
infinitism requires only that one be able to form dispositions to have the relevant occurrent
beliefs under the appropriate circumstances, and this could be explained by a finite number of
second-order dispositions, that is, dispositions to form the dispositions to have occurrent beliefs
under the appropriate circumstances. For example, mastery of a finite number of simple
mathematical principles could give us the ability to form the dispositions required to believe any
of an infinite number of mathematical propositions.17 In effect, Klein thinks of a justified
occurrent belief as the tip of an iceberg, supported ultimately by dispositional beliefs beneath the
surface of occurrent awareness which are themselves supported by the limitless ocean of reasons
to which a finite number of second-order dispositions gives us access.18

James Cornman and John Post argue that the trouble with infinitism is logical. In their view,
infinite justificatory regresses are unacceptable because if an infinite regress of distinct
propositions such that each member is implied by its successor could produce justified belief in
any proposition, then every proposition would be justified.19
Ernest Sosa shows, however, that this objection ignores the distinction, available to infinitism,
between actual and merely potential infinite justificatory regresses:

What an actual regress has that a merely potential regress lacks is the property of containing only
justified beliefs as members. What they both share is the property of containing no member
without successors that would jointly justify it.20

The crucial question is not whether there can be a regress σ (P1, P2,...) such that every member
Pn would be justified for S if its successor Pn+1 were, but whether there is a regress “each
member of which is justified by its immediate successors working jointly, where every member
of the regress is in addition actually justified”.21 Even if every proposition stands at the head of a
potentially justification-affording regress, it does not follow that every proposition stands at the
head of an actually justification-affording regress. Otherwise put, a properly formulated version
of infinitism should claim only that an infinite regress of reasons is necessary, but not sufficient,
for justified belief.

An actual justification-affording regress would be an infinite sequence of distinct propositions σ

(P1, P2,...) such that some person S is justified in believing P1 in part because for every member
Pn of σ, Pn+1 supports Pn for S. In short, an actual justification-affording regress would be a
regress that is an essential component of a sufficient condition of a person’s justifiably believing
a given proposition. In Klein’s view, any justified belief requires an actual justification-affording
regress supporting that belief. It does not follow that any infinite sequence of propositions σ (P1,
P2,...) such that every member is implied by its successor results in a justified belief. In addition,
each proposition in the sequence Pn must also be such that its successor Pn+1 is an objective
reason for Pn that is available to S as a reason for Pn, and perhaps other conditions must be
satisfied.

Paul Moser uses the distinction between actual and potential justificatory regresses to raise a
conceptual objection to infinitism. Moser’s objection would be a serious threat to Klein’s
infinitism since it attacks the claim that infinite regresses are necessary for justification. Moser
begins by noting that an infinite sequence of distinct propositions such that every member is
implied by its successor is insufficient for justified belief since the most that would have to be
true given such a regress is that any proposition in the regress is justified if its predecessor is.22 It
follows that a given infinite sequence of distinct propositions σ could be actually justification-
affording for a person S only if S is actually justified in believing the members of σ. This
requires, according to Moser, that each member Pn of σ be supported by an “external” justified
belief – a belief that is not itself a member of σ – en.23 However, a merely potentially
justification-affording regress of reasons cannot suffice for the justification of en itself, and “it is
useless to assume that the external justification consists of an actual infinite justificatory regress,
for such an assumption will only raise the question at hand once again,” namely, what
distinguishes regresses that are, from regresses that are not, justification-affording?24 So an
infinite regress of reasons cannot be necessary for justified belief.

The problem with this objection, as Klein notes,25 is that Moser simply assumes that the
“external” factor needed to explain the justification of each member of σ must be some
additional justified belief. To distinguish actual from merely potential justification-affording
regresses, infinitists must say that justification-affording regresses satisfy additional necessary
conditions on justification, but we cannot simply assume, as Moser does, that these conditions
must include additional justified beliefs. Moser is on to something, however, for the trouble with
infinitism does concern the factor that distinguishes actual from merely potential justification-
affording regresses.

5. THE TROUBLE WITH INFINITISM

The trouble with infinitism is neither doxastic, logical, nor conceptual, but epistemic. The regress
condition itself cannot explain the connection between justification and truth, but any additional
feature that could explain this connection would undermine the rationale for the regress
condition itself.

An important feature of normative epistemic properties such as justification is that they involve a
positive, relevant connection between what is believed by a person at a time and truth. Justified
beliefs need not be true, but they must be likely to be true or reasonably believed to be true. Any
account of the conditions on justification, therefore, must explain why beliefs that satisfy those
conditions are likely to be true or are reasonably believed to be true. So infinitism must
distinguish infinite sequences of propositions that are justification-affording – those upon which
actual justified beliefs depend – from those that are not justification-affording, in a way that
explains the relevant connection between justification and truth.

Infinitism is similar to coherence theories of justification. Both imply that justification is, at least
in part, a function of a belief’s relationship to other propositions in a believer’s space of reasons.
Infinitism and coherentism differ about the logical topology of justification-affording reason
spaces, however. Coherentism implies that the space of justification-affording reasons can be
finite and that a proposition can figure, sooner or later, in its own justification, if only by
contributing to the overall coherence of the system of beliefs of which it is a member. Since
infinitism requires propositional reasons but rejects the idea that a proposition can figure in its
own justification, it implies that only infinite sequences of propositions can constitute
justification-affording reason spaces.

Infinitism and coherentism thus face this challenge: logical relations alone cannot distinguish a
set of propositions that is likely to be true or reasonably believed to be true from a set that is
arbitrary, ungrounded fantasy. An important challenge for any coherence theory is to explain
coherence in such a way that the propositions in an otherwise coherent novel would not be
justified for anyone who accepts just those propositions. A similar problem confronts infinitism.
The fact that a proposition P is implied by another, which is implied by another, and so on,
provides a person with no reason to believe that P is true, even if that person has the dispositions
necessary to take every proposition in the sequence to be supported by its successor. For P and
the statements in its ancestry might be pure fiction with no grounding at all in how things are, are
likely to be, or are reasonably believed to be.

We have seen that Klein sidesteps the criticisms of Cornman, Post, and Moser by stressing that
an infinite regress is necessary, but not sufficient, for justified belief. He thereby commits
himself to the existence of an additional feature that is necessary and, together with the infinite
regress requirement, sufficient for justified belief. This additional condition must (perhaps
together with other factors) distinguish justification-affording in- finite regresses from infinite
regresses that are not justification-affording. Furthermore, since justification must have an
appropriate connection to truth, this condition must explain how justification-affording infinite
regresses are relevantly connected to truth. Specifically, an infinitist theory of justification must
imply that for any justified belief P, there is a property such that P has Ф and (i) P’s having Ф
does not require that P be supported by a justification-affording infinite regress of reasons, (ii)
P’s having Ф is a necessary condition of P’s being justified for S, (iii) Ф is positively relevant to
P’s being true or to P’s being reasonably believed to be true, (iv) P’s having Ф is not sufficient
for P’s being justified for S, and (v) Ф can be analyzed without reference to the concept of
justified belief itself.

Condition (i) is required since if P’s having Ф requires a justification-affording infinite regress
of reasons, then Ф will presuppose, and not explain, the distinction between justification-
affording and non-justification-affording regresses. Condition (ii) is needed since Ф must
distinguish regresses that are from regresses that are not justification-affording, and condition
(iii) is necessary since Ф must explain the link between justification-affording regresses and
truth. Condition (iv) is required, since if Ф is sufficient for P’s being justified, then, given (i), the
regress condition would be false. Finally, condition (v) is required since if Ф cannot be analyzed
without reference to the concept of justified belief, then the infinitist account of justification will
be viciously circular. The trouble with infinitism is that any viable Ф undermines the regress
requirement itself.

An essential feature of justification-affording regresses, according to Klein, is that the successor

Pn+1 of each member Pn of the regress be “objectively available” to S as a reason for Pn. As we
have seen, objective availability is not guaranteed by the fact that P deductively or inductively
implies C or by the fact that S takes P to be a reason for C. In some way, P must actually be a
good reason for S to believe that C is true. For example, according to Klein, that all fish wear
army boots and anything wearing army boots has fins is not an objectively available reason to
believe that all fish have fins, but that dark clouds are gathering over the mountains and it is
mid-winter in Montana is an objectively available reason to believe that a snowstorm is likely.26
Since Klein thinks objective availability is a necessary condition of justification, we may begin
by asking whether it can provide the needed connection between justification and truth. More
generally, we may consider whether a proposition’s being supported by objectively available
evidence provides the feature Ф that is both necessary and, together with the regress requirement,
sufficient for an infinite regress’ being justification-affording according to infinitism.

Klein claims that infinitism is compatible with a wide range of accounts of objective availability:

There are many accounts of objective availability. Each specifies either some normative or non-
normative property or, perhaps, a mixed property that is sufficient to convert a belief into a
reason. For example, one could say that a belief, r, is objectively available to S as a reason for p
if (1) r has some sufficiently high probability and the conditional probability of p given r is
sufficiently high; or (2) an impartial, informed observer would accept r as a reason for p; or (3) r
would be accepted in the long run by an appropriately defined set of people; or (4) r is evident
for S and r makes p evident for S; or (5) r accords with S’s deepest epistemic commitments; or
(6) r meets the appropriate conversational presuppositions; or (7) an intellectually virtuous
person would advance r as a reason for p.27

Klein thinks that no matter which of these accounts we adopt, we might use it to distinguish
infinite sequences of propositions that are, from infinite sequences of propositions that are not,
justification-affording. Unfortunately, Klein is mistaken about this, for none of these accounts is
available to infinitism. To see this, we must examine each proposal in detail.

Together with the regress condition, account (1) implies that S is justified in believing a
proposition P1 only if there is an infinite sequence of propositions σ (P1, P2,...) such that for
every member Pn of σ, Pn+1 is probable, to a sufficient degree, and the conditional probability
of Pn given Pn+1 is sufficiently high. To say that a proposition is probable is to say that, in some
sense, it is likely to be true. The likelihood of a contingent proposition, however, is not an
intrinsic feature of the proposition itself, but is a matter of its relationship to other propositions
or, perhaps, to probability-conferring non-propositional states of affairs. Just which propositions
(or non-propositional states of affairs) confer probability on the members of a justification-
affording infinite regress?

Not just any set of propositions can do the job, for then every regress would be justification-
affording, since every proposition is maximally probable given itself and highly probable given
some other propositions. Nor can we say that the propositions in a justification-affording infinite
regress must be sufficiently probable given the set of true propositions, for this will imply that
impeccable reasoning from limited data cannot result in justified beliefs. For example, in light of
the true proposition that there are black and white swans it is improbable that all swans are
white, yet one could be justified in believing the latter proposition on the basis of a limited set of
true propositions, as many careful observers were before the discovery of black swans. Nor will
it do to say that the propositions in a justification-affording regress are probable in light of what
S believes, for then any regress stemming from a given belief will be justification-affording for S
provided only that S has the dispositions necessary to sustain it. Evidently, then, the only thing
that will do is to say that each of the propositions in a justification-affording regress must be
probable in light of what S is justified in believing or, at least, in light of the reasons available to
S, but this would make infinitism unacceptably circular or would undermine the regress
requirement, or both.

Together with the regress condition, Klein’s second sample account of the objective availability
of a reason implies that S is justified in believing a proposition P1 only if there is an infinite
sequence of propositions σ (P1, P2,...) such that for every member Pn of σ, an impartial,
informed observer would accept Pn+1 as a reason for Pn. The crucial expression here is
‘impartial, informed observer’. One possibility is that an informed observer is one who believes
all the relevant truths, in which case infinitism will imply that sound reasoning from limited data
cannot result in justified beliefs. No informed observer would regard the fact that all the swans S
has seen have been white as a reason to believe that all swans are white, since such an observer
would have the information that not all swans are white. Another possibility is that an impartial,
informed observer has the relevant justified beliefs, but this would make the account of
justification viciously circular. Yet another possibility is that an impartial, informed observer is
one who knows the relevant necessary and sufficient conditions for justified belief – hence
knows what is and what is not an objective reason – and thus is in a position to know whether or
not S satisfies them. This account is also unacceptably circular. So infinitism cannot use the
concept of an informed, impartial observer to draw the needed distinction between infinite
regresses that are and infinite regresses that are not justification-affording.

Klein’s third sample account of the objective availability of a reason says that a proposition R is
an objective reason for a proposition P just in case it “would be accepted in the long run by an
appropriately defined set of people”.28 Presumably Klein intends this to imply not just that the
appropriately defined set of people would accept R but also that they would regard R as a reason
for P. Together with (K), this implies that S is justified in believing a proposition P1 only if there
is an infinite sequence of propositions σ (P1, P2,...) such that for every member Pn of σ, Pn+1
would be accepted in the long run by an appropriately defined set of people and would be
regarded by them as a reason for Pn. The crucial expression in this account is ‘appropriately
defined set of people.’ Either the definition in question implies that such persons, in the long run,
have the relevant justified beliefs, the relevant true beliefs, or it implies neither of these things. If
the definition implies that such persons have the relevant justified beliefs, then the account will
be unacceptably circular. If the account implies that such persons have the relevant true beliefs,
then it will be too strong by implying that good reasoning from limited data will not result in
justified beliefs. If it avoids either of these implications, however, then it fails to provide an
appropriate connection to the truth, for a set of persons defined in such a way that they need not,
even in the long run, have the relevant true or even only the relevant justified beliefs cannot
explain the connection between justification and truth.

Klein’s fourth sample account of the objective availability of a reason, together with (K), implies
that S is justified in believing a proposition P1 only if there is an infinite sequence of
propositions σ (P1, P2,... such that for every member Pn of σ, Pn+1 is evident for S and Pn+1
makes Pn evident for S. Since ‘evident’ is a term of epistemic appraisal that applies to a person
S’s belief that P only if S has evidence that is positively relevant to P’s being reasonable for S,
either this account results in unacceptable circularity or else it is sufficient for justified belief, or
both.29

Klein’s fifth sample account says that a proposition R is a reason for P just in case it “accords
with S’s deepest epistemic commitments”.30 Presumably Klein intends this account to imply not
just that R accords with S’s deepest epistemic commitments, but that R’s being a reason for P
does so, as well. Together with (K), this implies that S is justified in believing a proposition P1
only if there is an infinite sequence of propositions σ (P1, P2,... such that for every member Pn
of σ, Pn+1 and that Pn+1 is a reason for Pn accord with S’s deepest epistemic commitments. One
problem with this account is that it does not provide an appropriate connection between
justification and truth, for propositions that are in accord with a person’s deepest epistemic
commitments might not be true or likely to be true. One might argue that by being in accord with
our deepest epistemic commitments such propositions are reasonably believed to be true, but this
just pushes the problem back: what makes our deepest commitments themselves reasonable, in
an epistemically relevant sense?31

Even if this problem could be overcome, this approach to the concept of an objective reason is in
tension with the infinitist’s regress requirement. For if being in accord with one’s deepest
commitments makes it reasonable to accept one proposition on the basis of another, then being in
accord with one’s deepest commitments ought to suffice for justification, whether or not one’s
dispositions can sustain an infinite regress. If P is in accord with my deepest commitments, then,
by this account, it is reasonable for me to believe P whether or not I can sustain an infinite
justificatory regress supporting P. A related problem is raised by reflective persons whose
deepest epistemic commitments imply that an infinite regress of propositions is not a necessary
condition of epistemic justification. If reasonability is determined by our deepest commitments,
then such persons can have reasonable beliefs without satisfying the regress condition, hence (K)
would be false.

Klein’s sixth sample account of the objective availability of a reason says that a proposition R is
an objective reason for P if it “meets the appropriate conversational presuppositions.”32 Given
this, Klein’s account would imply that S is justified in believing a proposition P1 only if there is
an infinite sequence of propositions σ (P1, P2,... such that for every member Pn of σ, Pn+1
meets the appropriate conversational presuppositions. Here the crucial expression is ‘appropriate
conversational presuppositions.’ Presumably, meeting conversational presuppositions is a matter
of conforming to social conventions about what is and what is not appropriate to offer as a
reason. However the details are specified, either meeting such conditions suffices for justified
belief, true belief, or neither. If meeting such conditions implies that the relevant beliefs are
justified, then the account is either unacceptably circular, or undermines the claim that an infinite
regress is necessary for justified belief, or both. If the account implies that beliefs satisfying the
appropriate conversational presuppositions are true, then it will be too strong, since it will imply
that false beliefs cannot play a role in inferential justification. If it avoids either of these
conditions, however, then it fails to provide an appropriate connection between justified beliefs
and truth, for beliefs that satisfy a given set of social conventions need not be an adequate
standard specifying the extent to which our beliefs are likely to be true. One might try to avoid
this problem by claiming that conformity to social conventions is sufficient for reasonable belief.
This just pushes the problem back: not just any conventions will be conducive to reasonable
belief, so what distinguishes those that are from the rest? Furthermore, infinitism runs into
difficulty if it turns out that, in some contexts, it is a conversational presupposition that infinite
regresses of reasons are not required for justified belief. Either such conventions do not override
the regress requirement – in which case we need a distinction between justification-affording and
non-justification-affording conventions – or they do, in which case (K) is false.

Together with (K), Klein’s final sample account of the objective availability of a reason implies
that S is justified in believing a proposition P1 only if there is an infinite sequence of
propositions σ (P1, P2,...) such that for every member Pn of σ, Pn+1 would be advanced by an
intellectually virtuous person as a reason for Pn. An intellectually virtuous person might be
either a person whose dispositions are optimally attuned to forming beliefs that are reasonable
under the conditions the person happens to be in, or one whose dispositions are optimally attuned
to forming beliefs that are true under those conditions.

However, we specify the relevant dispositions, an intellectually virtuous person must either be
one whose beliefs all satisfy the regress requirement or one whose beliefs do not all satisfy the
regress requirement. If all of an intellectually virtuous person’s beliefs must satisfy the regress
requirement, then Klein’s account of justified belief will be un-informatively circular, since the
account of the objective availability of a reason is supposed to explain the difference between
regresses that are justification-affording and those that are not. To tell which regresses are
justification-affording we would first need to identify an intellectually virtuous person, but to
identify an intellectually virtuous person we would first need to know which regresses are
justification-affording.

On the other hand, suppose that intellectually virtuous persons need not be such that all of their
beliefs satisfy the infinite regress requirement. Under this possibility we must consider both the
case in which an intellectually virtuous person is one whose dispositions are optimally attuned to
forming justified beliefs, and the case in which an intellectually virtuous person is one whose
dispositions are optimally attuned to forming true beliefs.

If intellectually virtuous persons are persons who are optimally disposed to have justified beliefs
– hence have only justified beliefs – but their beliefs need not satisfy the regress requirement,
then justified beliefs need not be supported by an infinite regress of reasons. So this account of
the objective availability of a reason is incompatible with the regress requirement.

This problem can be avoided if intellectually virtuous persons are persons who are maximally
disposed to have true beliefs, yet need not be such that all of their beliefs satisfy the regress
condition, for it will not follow from these assumptions that there can be justified beliefs that are
not supported by infinite regresses of reasons. In this case, however, either being optimally
disposed to having true beliefs guarantees that one’s beliefs are true or it does not. If so, then
reasoning from strong but incomplete data to justified false beliefs will be impossible.

So now suppose being optimally disposed to having true beliefs does not guarantee having true
beliefs (or justified beliefs, given the conception of intellectual virtue in question). On this
alternative, S will be justified in believing a proposition P1 only if there is an infinite sequence of
propositions σ (P1, P2,...) such that for every member Pn of σ, Pn+1 would be advanced by a
person optimally disposed to having true beliefs as a reason for Pn.

One objection to this position is that if being optimally disposed to having true beliefs does not
require satisfying the regress condition, then the regress condition cannot be a condition on
justification. For, it may be argued, the goal of inquiry is to have as many true and as few false
beliefs as possible on topics of interest to us, and justification is nothing more than what is
required in order for us to attain that end. So if we can be optimally disposed to forming true
beliefs without satisfying the regress condition, then the regress condition cannot be a condition
on justification.

This is tempting, but not decisive. For the infinitist may argue that our epistemic goal is not
simply to have true beliefs, but to have true beliefs together with an understanding of the
evidence we have for them. An understanding of the evidence for our beliefs is not simply a
means to the goal of inquiry, it is partly constitutive of that end. Given this, an infinitist may
plausibly add that, since having evidence requires an infinite regress, satisfaction of the regress
requirement itself is partly constitutive of the goal of inquiry.
So we need other reasons for thinking that the claim that Pn+1 is a reason for Pn just in case a
person optimally disposed to having true beliefs (who nevertheless has at least some false beliefs
and not all of whose beliefs satisfy the regress condition) would advance Pn+1 as a reason for Pn
is unavailable to the infinitist. This account of objective availability might provide us with a
practical standard to use in deciding whether R is a reason for C. When the epistemic chips are
down I might ask, “if I were an intellectually virtuous person would I use R as a reason for C?”
Our question, however, is whether infinitism can use this account as an explanation of what it is
for R to be objectively available as a reason for C. Consider, then, the deliberations of an
intellectually virtuous person Sv. In deciding whether or not R is a reason for C, Sv could ask
whether an intellectually virtuous person would take R to be a reason for C. Since Sv is an
intellectually virtuous person, the question reduces to whether Sv does take R to be a reason for
C. In light of what? Not in light of all relevant truths, since this will imply that reasoning from
limited data cannot result in justified false beliefs; in any case, we are supposing that not all of
Sv’s beliefs are true. Not in light of relevant justified beliefs, since this will make the infinitist
account of justification circular. The remaining option is that Sv takes R to be a reason for C in
light of Sv’s relevant beliefs, in particular, in light of Sv’s belief that R and that R supports C.
The belief that R supports C, in turn, can only be grounded on Sv’s epistemic commitments. So
objectively available reasons are those that intellectually virtuous persons take to be reasons in
light of their own epistemic commitments.

What reasons are there for supposing that an intellectually virtuous person’s commitments,
especially commitments about what supports what, are true or are likely to be true? In the case
under consideration – the case in which an intellectually virtuous person, though optimally
disposed to form true beliefs, has at least some false beliefs – there is no such reason. For it is
possible that persons optimally disposed to form true beliefs are nevertheless exceptionally poor
at forming true beliefs about what supports what: they manage to get a maximal number of true
beliefs, and perhaps even to have what they take to be reasons for those beliefs, but the chains of
support they are disposed to construct in supporting them are filled with undetected howlers. So
being a proposition that an intellectually virtuous person, in this sense, would take to be a reason
is not the same thing as being a good reason. So Klein’s final account of the objective
availability of a reason cannot distinguish justification-affording regresses from the rest.

So none of Klein’s sample accounts of the objective availability of reasons provides infinitism
with a distinguishing feature of justification-affording regresses that explains the truth-relevant
character of justification. However, this shows only that Klein’s infinitism is an incomplete
account of justification, not that an infinitist account cannot be complete. Why suppose that there
cannot be a property that completes the infinitist account of justification in the appropriate way?
To show this, I shall argue that any viable account of undermines the regress requirement.
Property , according to the infinitist, is necessary and, together with the regress condition,
sufficient for justified belief. In particular, if a belief that P has Ф, then P must be true, likely to
be true, or reasonably believed to be true. I shall show, by elimination, that if P has Ф, then
either there is a non-doxastic mental state that provides evidence for P or there is some non-
mental, external state of affairs that explains why P is true, likely to be true, or reasonably
believed to be true. In either case, the rationale for the regress condition collapses.
What can it be for a belief P to have Ф? Ф must either be such that a belief P’s having Ф
requires justified beliefs or does not require justified beliefs. If P’s having Ф requires justified
beliefs, then either Ф will be sufficient for justified belief or the infinitist account will be
unacceptably circular, or both. So P’s having Ф cannot require that S have any justified beliefs.

Since P’s having Ф cannot require that S have any justified beliefs, either P’s having Ф is a
matter of S’s having unjustified beliefs that have Ф, a matter of S’s having unjustified beliefs that
lack Ф, or P’s having Ф is a matter of there being some non-doxastic, possibly non-mental, state
of affairs relevant to the truth of P. But P’s having Ф cannot be a matter of having unjustified
beliefs with Ф. For one thing, this will be uninformatively circular, for what we need is an
account of Ф itself and of its relationship to truth. Worse, even if this problem could be solved,
the account would imply that justification requires reasons that are not supported by reasons
available to the believer. Since the unjustified beliefs in question have Ф, they are objectively
available reasons for the truth of P. Yet since these beliefs are unjustified, they must fail to
satisfy the regress condition. This implies that sooner or later the justification for P depends upon
a reason that is not itself supported by a reason. This would undermine (PAA).

Perhaps, then, P’s having Ф is a matter of S’s having unjustified beliefs without Ф. But
unjustified beliefs, per se, cannot explain why P is true, likely to be true, or reasonably believed
to be true. To explain the connection to truth, such beliefs will need some further property that is
positively relevant to truth, in which case the same problems arise about this new property. Since
P’s having Ф cannot be a matter of S’s having justified beliefs, unjustified beliefs with Ф, or
unjustified beliefs without Ф, P’s having Ф requires at least one non-doxastic state relevant to
the truth of P.

There are, at last, two possibilities. One is that the non-doxastic state required by Ф is a non-
doxastic psychological state, the content of which is relevant to the truth of P. The other
possibility is that Ф involves some objective – that is, non-psychological – state of affairs
relevant to the truth of P. However, if Ф involves a non-doxastic, psychological state of
awareness the content of which is relevant to the truth of P, this state will itself be a reason
available to S for thinking that P is true.33 But if such mental states – states without propositional
content – can be reasons for beliefs, then (PAA) is gratuitous.

On the other hand, if Ф involves some objective, non-psychological state of affairs, then whether
or not S is justified in believing a proposition P will not just depend upon whether there is a
proposition Q that provides evidence for P and is available to S as evidence for P, and so on, but
on whether some external factor Ф makes Q a justification-affording reason for P. If there is
such an objective condition on justification, then there can always be two internally
indistinguishable persons S and S∗ such that both satisfy (K) with respect to a belief P, while S is
but S∗ is not justified in believing P. We have seen that this state of affairs is incompatible with
Klein’s argument for (PAA) since that argument presupposes that factors relevant to justification
must be internal and available to the believer. We now see that such an external condition puts
(PAA) itself at risk since, in this case, it is external factor Ф itself that makes the non-arbitrary
difference between justified and unjustified beliefs. Since Ф does the epistemic work, it is
gratuitous to insist, as (PAA) does, that any belief with Ф can be justified for S only if it is also
supported by another belief available to S.
Why, then, does Klein insist on the availability of propositional support for every justified
belief? Klein claims that having reasons is distinctive of “adult human knowledge.”34 In this
connection, Klein shares these doubts of Ernest Sosa’s about reliabilism:

Admittedly, there is a sense in which even a supermarket door “knows” when someone
approaches, and in which a heating system “knows” when the temperature in a room rises above
a certain setting. Such is “servo-mechanic” knowledge. And there is an immense variety of
animal knowledge, instinctive or learned, which facilitates survival and flourishing in an
astonishingly rich diversity of modes and environments. Human knowledge is on a higher plane
of sophistication, however, precisely because of its enhanced coherence and comprehensiveness
and its capacity to satisfy self-reflective curiosity. Pure reliabilism is questionable as an adequate
epistemology for such knowledge.35

Fully human knowledge, according to Klein, requires us to have reasons for our beliefs, not
simply to have beliefs that are likely to be true because our servo-mechanisms are properly tuned
to our environment. If there were an external condition on justification, however, then two
persons could be just alike in all internal respects and still be such that one has, while the other
lacks, reasons and, consequently, justified beliefs. This means that having a reason for P would
be a matter of having access to a proposition R that, from S’s own point of view, just happens to
be a reason for P, whether or not S is, or even can be, aware that it is a reason for P. According
to such an external account, there is no sense in which a person’s having a reason for P need be
less mechanical or more self-reflectively satisfying than a supermarket door’s “knowing” that
someone is approaching. Whether or not I have a genuine reason for a belief is not a matter of
how I conduct my doxastic affairs with respect to that belief but is a matter of whether or not my
affairs are attuned to my environment in the right sort of way. So if Ф involves an external state
of affairs, the rationale for the regress condition is undermined.

6. CONCLUSION

Klein’s argument for infinitism is unsupported since his argument for (PAA) is unsound. Worse,
infinitism cannot be a complete account of epistemic justification. If being a member of an
infinite regress is necessary, but not sufficient, for justification, then there must be a factor Ф that
explains the truth-conducive character of justification and that is both necessary and, together
with the regress requirement, sufficient for justification. However, any viable candidate
undermines the rationale for the regress condition itself.

The results of my argument are more general. The argument given in the previous section applies
to any property Ф that can provide a principled distinction between justified and unjustified
beliefs. If there is a principled, truth-relevant distinction between justified and unjustified beliefs,
then either non-doxastic mental states can provide justification-affording evidence for at least
some of our beliefs or there is some external condition on justification, or both.

7. ACKNOWLEDGEMENTS I am grateful to Peter Klein and William Lycan for very helpful
comments on an earlier version of this essay. The remaining errors are entirely my own.
NOTES

1 Klein (1999).

2 Klein (1999, p. 298).

3 Klein (1999, p. 299).

4 Klein (1999, pp. 299–300).

5 Post (1980).

6 Klein (1999, p. 298).

7 According to Klein, reliabilist theories cannot account for the rational requirements of adult
human knowledge, traditional foundational and non-traditional coherence theories cannot
terminate the regress since it is always legitimate to inquire why foundational propositions or
coherent sets of beliefs are likely to be true, and non-evidentialist theories can offer no reason for
thinking that any set of beliefs we might start with is more likely to be true than any other set of
beliefs (1999, pp. 301–306).

8 Klein (1999, p. 304). One notable gap in Klein’s argument for (PAA) is his failure to consider
intuitionistic foundationalism. According to the intuitionistic foundationalist, we have non-
propositional evidence for our basic beliefs in the form of non-doxastic sensory states of
appearing. If such a view is true, then (PAA) is false, since not all justified beliefs need be
supported by propositional reasons yet basic beliefs are not arbitrary since they are supported by
non-propositional sensory states that function as non-propositional reasons. The most thorough
and compelling defense of this position is provided by Moser (1985).

9 Cling (forthcoming).

10 Alston (1989, pp. 11–12, 16–18).

11 Klein (1999, p. 303). This argument is similar to an argument of BonJour’s (1978, pp. 5–6).

12 Klein (1999, p. 299).

13 Klein (1999, p. 300).

14 Klein (1999, pp. 308–309).

15 Klein (1999, pp. 308–309).

16 For objections along these lines, see Aristotle (Book I, Chapter 3, 72b5–15) and Williams
(1981, pp. 85–88).
17 Klein (1999, pp. 308–309).

18 The iceberg metaphor is adapted from Sosa (1980, pp. 9–10).

19 Cornman (1977, pp. 290–291), Post (1980, pp. 32–37).

20 Sosa (1980, p. 12).

21 Sosa (1980, p. 13).

22 Moser (1985, pp. 107–116).

23 Moser (1985, pp. 114–115).

24 Moser (1985, p. 115).

25 Klein (1999, p. 324, note 51).

26 Klein (1999, p. 300).

27 Klein (1999, p. 299).

28 Klein (1999, p. 299).

29 Chisholm (1966, pp. 18–21) uses ‘evident’ in this way.

30 Klein (1999, p. 299).

31 This sort of account of reasonability is defended by Foley (1990).

32 Klein (1999, p. 299). 33 Moser provides a detailed defense of this possibility (1985, pp. 141–
187).

34 Klein (1999, p. 302).

35 Sosa (1991, p. 95) quoted by Klein (1999, p. 302).

VERITAS Porto Alegre v. 50 n. 4 Dezembro 2005 p. 109-128 6
THE PROBLEM OF THE CRITERION, KNOWING THAT ONE KNOWS AND INFINITISM
Tito Alencar Flores

ABSTRACT – This essay discusses the Problem of the Criterion, one of the most important
problems in Epistemology. The answer given to this problem will shape a fundamental aspect of
theories of knowledge. It also analyzes some of the consequences epistemological theories may
face in accepting metaepistemological requirements. These consequences are given special
consideration with regard to Infinitism, the epistemological theory according to which the
reasons that support our beliefs must be infinite and non-repeating. Finally, this essay claims not
only that those propositions that require metajustification make any theory, even Infinitism, more
plausible, but also that the absence of such requirements creates difficult problems, such as the
so-called “Problem of Easy Knowledge.”
KEY WORDS – Infinitism. Problem of the Criterion. Metajustification. Knowledge.
Metaknowledge. Skepticism.

The Problem of the Criterion (PC) leaves us grappling with a very complicated question: should
we think that knowing that one knows is a condition of knowing in the first place? This question
can be posed in different ways: must a theory state that to know that an evidence or reason is
adequate is part of what is necessary to be justified in believing some proposition on the basis of
that evidence or reason? Or, still, is it right to claim that we can come to know some proposition
p on the basis of having arrived at p through a reliable process of belief formation without
knowing that the process is in fact reliable?1

As Chisholm described it once, the PC consists of having to make a special kind of decision,
“how do we decide, in any particular case, whether we have a genuine item of knowledge?” 2 Of
course, those are epistemic decisions, since they are about deciding whether we know that we
know something. As a result, the PC must be defined as a metaepistemological problem. With
that in mind, for present purposes, it will prove useful to call the PC a “metaepistemological
problem” (departing somewhat from the conventional use of the adjective
“metaepistemological”).

There are two obvious ways in which we can deal with the PC. On the one hand, we might say
that the PC is only about second-order knowledge.3 In this sense, granting that we have first-
order knowledge, we should only try to find out if we know that we have this piece of first-order
knowledge. An eventual failure to know that we know would have no impact on our first-order
knowledge. In this case, the situation in which S knows that some proposition p is true, but fails
to know that he knows p is true, would be unproblematic.

On the other hand, we might say that the claim underlying the PC states that in order to have
first-order knowledge we must know that we know. Then, in order to know that some
proposition p is true, we will have to know that the reason on which our believing that p is based
is adequate, or that the process of belief formation of p is reliable. Otherwise, we would be
missing some fundamental point regarding the phenomenon of knowledge. To use an expression
from Sextus Empiricus, lacking this second-order knowledge will make us so vulnerable to
criticism that we would be discredited.
In one sense, then, we could claim that, since the PC has this metaepistemological nature, the
question is not about whether we know something – i.e. it is not about first-order knowledge –
but only about whether we know that we know something – i.e. it is only about second-order or
metaknowledge. Yet, under another interpretation, we could say that the metaepistemological
nature of the PC makes it a problem about whether we know something, because failing to know
that we know would have a negative impact on our first-order knowledge. In this last sense, the
PC would point to the existence of a natural and resilient bridge between knowing and knowing
that one knows.

According to the theories that deny that we have to know that we know in order to know in the
first place, the phenomenon of knowledge can be understood and completely captured in the
absence of second-order knowledge. Thus, a person S should be granted knowledge that p if, for
example, he believes a true proposition p and uses a reliable process of belief formation, without
having to know that the process is in fact reliable. But, according to the other way to deal with
the PC, if S has a true belief that p, and this belief is a result from what is, unbeknownst to S, a
reliable process of belief formation, S could be seen, from some point of view (naturally, not the
point of view of S himself), as avoiding an accidentally true belief. But if S did not know that the
process was reliable, S would be missing something epistemically decisive. In the absence of
second-order knowledge, what S has doesn’t amount to much at all.

Following Stewart Cohen’s suggestion, we may call the theories that claim that we can come to
know that p is true without knowing that we know p, “Basic knowledge structure theories”
(BKT). Also, we may call the first-order knowledge acquired without metaknowledge “Basic
knowledge” (BK). According to Cohen,

S has basic knowledge of P just in case S knows P prior to knowing that the cognitive source of
S’s knowing P is reliable. Our knowledge has basic knowledge structure just in case we have
basic knowledge and we come to know our faculties are reliable on the basis of our basic
knowledge.4

The BKT permit, then, that a knowledge source “can deliver knowledge prior to one’s knowing
that the source is reliable” 5 and “hold that reliability knowledge is based on basic knowledge”.6
Before proceeding, I want to clarify a couple of things about the vocabulary Cohen uses. First,
the idea of “knowledge source” can be seen as indicating two things: a knowledge source could
be a method of belief formation, as well as the reason that is supposed to justify some particular
belief. I think this is what Cohen means. Moreover, we must bear in mind that “reliability
knowledge” is equivalent to metaknowledge, in the sense that, when one gets knowledge of the
reliability of a knowledge source, one knows that one knows. In light of these last clarifications,
the point is this: if the BKT hold that metaknowledge can be based on basic knowledge there
would be a decisive reason to abandon those theories. The threat faced by BKT is what has been
called “bootstrapping”.7 I will not discuss this problem in great detail, but I do want to clarify
what “bootstrapping” means.

We can think of bootstrapping through the following example.8 Suppose that a person, Michael,
is driving his car from California to Alaska. Soon after he passes a gas station, he reads a road
sign that says that the next gas station will be 300 miles away. Then, he looks at the car’s gas
gauge and comes to believe (1) the tank is full. According to a BKT like Reliabilism, this belief
is an item of knowledge for Michael, provided that (a) it is true that the tank is full and (b) the
gauge is reliable – both of which we can suppose for the sake of the argument. Using the terms
just defined, Michael has BK that (1). So, according to Reliabilism, even if Michael does not
know that he knows the tank is full or that the gas gauge is reliable, his belief that the tank is full
is an item of knowledge for him.

A bit later, seeing that it is getting dark and cold out there, Michael checks the gas gauge again.
Now, he comes to believe (2) the gauge marks “full.” According to a BKT, (2) is also an item of
Michael’s knowledge, provided that it is true that the gauge marks “full” and that Michael’s
vision is reliable – which we can also suppose for the sake of the argument. Again, Michael
doesn’t know that he knows (2), but he knows the gauge marks “full.” Once again, Michael has
BK that (2).

At this moment, Michael can believe (3) the gauge marks “full” and the tank is full. From any
point of view, Michael is entitled to infer (3) from (1) and (2). Actually, since Michael comes to
believe (3) deducting it from (1) and (2), Michael knows (3), given that deduction is a reliable
process of belief formation. I will insist on one point, since it seems very important: Michael
doesn’t know that he in fact knows (1) or (2), but according to the BKT, unbeknownst to him, he
knows (1) and (2), and once he deduces (3) from (1) and (2), he knows (3) as well.

Here is where the Bootstrapping comes in: since Michael knows (3), nothing prevents him from
inferring from (3), (4) the gauge is working properly on this occasion. Since the example I am
offering is quite similar to Vogel’s example, and given how crucial this step is, it seems useful to
see exactly what he says.

So, for example, when the gauge reads ‘F’ she believes that, on this occasion, the tank is full.
She also believes that, on this occasion, the gauge reads ‘F’. Moreover, Roxanne combines these
beliefs; she believes: (20) On this occasion, the gauge reads ‘F’ and F. Certainly, the perceptual
process by which Roxanne forms her belief that the gauge reads ‘F’ is a reliable one. By
hypothesis, her belief that the tank is full is also reached by a reliable process. Hence, there
seems to be no good reason to deny that Roxanne’s belief in (20) is the result of a reliable
process, and the reliabilist will say that she knows (20). Now, it is a completely straightforward
logical consequence of (20) that: (21) On this occasion, the gauge is reading accurately. Assume
that Roxanne deduces (21) from (22). Deduction is certainly a reliable process, so there is no loss
of reliability at this step. Consequently, it seems that Roxanne must be credited with knowing
(21).9

Going back to our example, the decisive step is when Michael comes to know the gauge is
reading accurately on some occasion based on his knowledge of (3) (the gauge marks “full” and
the tank is full). Furthermore, if Michael repeats this same procedure and makes a few inductive
inferences, he can then come to know that the gauge is working properly.10 As we have seen, (3)
is based on Michael’s basic knowledge of (1) and (2). So – without forgetting the inductive step
– we can claim that reliability knowledge is being based on basic knowledge.
Now, what would be the problem if Michael comes to believe (4) the gauge is working properly
on some particular occasion? In a certain sense, there is no problem at all, as long as Michael
infers (4) from (3). However, the point is that with (4) Michael suddenly comes to know that he
knows the tank is full, since (a) knowing that the gauge is working properly is equivalent to
reliability knowledge11 and (b) reliability knowledge is equivalent to knowing that one knows.
The problem for BKT, then, is that the description of Michael’s acquisition of second-order
knowledge is in perfect accord with a BKT like Reliabilism, but this acquisition of second-order
knowledge cannot be considered legitimate. So, this way of acquiring metaknowledge may be
seen as a reductio of all BKT’s.

In our example, Michael is basing reliability knowledge on basic knowledge. Since it is not only
Reliabilism that holds that reliability knowledge is based on basic knowledge, the problem must
be in the feature of all BKT’s that is shared by Reliabilism. The problem is that we cannot come
to know that we know in this way, but this way of acquiring metaknowledge is approved by any
BKT. So, we have a problem with BKT in general.

We must note also that, as was pointed out by Cohen,12 not only Externalist theories, like
Reliabilism, are BKT, and thus face the Problem of Bootstrapping. On the contrary, as we will
see, many Internalist theories, like Infinitism, can be BKT’s as well.

In what follows, I shall assume that BKT’s cannot provide a comprehensive analysis of
knowledge or justification, since all BKT face the Bootstrapping Problem. In turn, the most
important point is that the Bootstrapping Problem derives from a misdirected description of the
condition for first-order knowledge or justification. Hence, we cannot separate epistemic levels,
and we have to analyze the phenomenon of knowledge in a way that combines knowing and
knowing that one knows.

Infinitism

Infinitism is preferable to both Coherentism and Foundationalism, mainly in light of the fact that
Infinitism is the only theory capable of avoiding dogmatism as well as circularity. I am not going
to defend Infinitism over Foundationalism and Coherentism, though. I agree with Klein that no
criticism against Infinitism does succeed.13

However, I want to take into account a particular feature of Klein’s Infinitism and try to suggest
that it must (and can) be changed in order for Infinitism to be transformed into a more
compelling theory. The aspect of Infinitism I will treat is directly related to that which is
responsible for making some theory a BKT.

According to Klein, a person S will have justification to believe some proposition p if S has
reasons available to him. This is what the Principle of Avoiding Arbitrariness states,

(PAA) For all x, if a person, S, has a justification for x, then there is some reason, r1, available to
S for x, and there is some reason, r2, available to S for r1, etc, and there is no last reason in the
series.14
The notion of availability has two different aspects, one subjective and another objective. In spite
of the relevance that subjective availability has for Infinitism, we can put this notion aside. What
is crucial for our purposes here is the idea according to which a reason must be objectively
available for S.

The conception that a reason must be objectively available to some person S shows that not just
any proposition will be a reason for S to believe some other proposition. It means that the reason
r must be adequate from some objective perspective. Furthermore, by granting that not any
proposition can be a proper justifier, Infinitism need not produce a particular account of
objective availability. As Klein says, “there are many accounts of objective availability. Each
specifies either some normative or non-normative properties or, perhaps, a mixed property that is
sufficient to convert a belief into a reason”.15 We could think of these accounts of objective
availability in this way,

One could say that a belief, r, is objectively available to S as a reason for p if (1) r has some
sufficiently high probability and the conditional probability of p given r is sufficiently high; or
(2) an impartial, informed observer would accept r as a reason for p; or (3) r would be accepted
in the long run by an appropriately defined set of people; or (4) r is evident for S and r makes p
evident for S; or (5) r accords with S’s deepest epistemic commitments; or (6) r meets the
appropriate conversational presuppositions; or (7) an intellectually virtuous person would
advance r as a reason for p.16

By itself, this aspect doesn’t subject Infinitism to the Bootstrapping Problem. As we know, the
requirement of objective availability is quite common in almost all theories of knowledge or
justification. The aspect I want to point out is the idea according to which S is not supposed to be
justified that r is objectively available to him in order for S to have justification to believe p on
the basis of r. Actually, for Klein, S need not have either (a) a belief that r is an objectively
available reason for believing p, or (b) have justification – without believing – that r is
objectively available to S. According to Klein, all that matters is the objective property that r
must have in order for S to have justification to believe p on the basis of r.

In a footnote to “Human knowledge and infinite regress of reasons,” Klein says,

Let me make the distinction between the three views of justification absolutely clear. The ‘thin’
view (the one I think is correct) holds that S has a justification for p on the basis of r entails that
(a) S believes r and (b) r is reason for p. It does not require that, in addition, either (1) S believes
that r is a reason for p or (2) S is justified in believing that r is a reason for p. The ‘moderately
thick view’ (the one I think is plausible) adds (1) to the thin view. The ‘extremely thick view’
(the one I think cannot be correct) adds (2) and presumably (1) as well to the thin view.17

Klein still adds, “I think it is the reason available to S for p that determines whether S has a
justification for p regardless of S’s beliefs about those reasons.” 18 This characteristic of
Infinitism, or of any kind of Internalism, is what makes it a BKT.

It is quite important to make some distinctions here. First of all, no plausible theory would deny
the possibility of acquiring second-order knowledge, or deny the possibility for some person S to
have justification that he has justification to believe some proposition p. The point is whether we
will consider the metaepistemic requirement a necessary condition to be justified in the first
place. Second, there is a difference between (a) requiring a belief according to which a reason r
is a good or adequate reason to believe p, and (b) requiring justification to believe that a reason r
is a good or adequate reason to believe p.

At the outset, it seems that what subjects a theory to the Bootstrapping Problem is not related to
believing itself. What matters is justification only. Hence, Infinitism couldn’t be made
invulnerable to the Bootstrapping Problem if it required that S must believe, or must have some
disposition to believe, that r is the reason for believing p. Therefore, even Klein’s “moderately
thick view” wouldn’t keep Infinitism from being a BKT. What may prevent a theory from being
a BKT is the need to name the epistemic efficiency of what justifies some belief.

We could say, then, that the BKT could be avoided either with (a) the requirement that S needs
to have justification for believing that r is objectively available to S, or (b) the requirement that S
needs a justified belief that r is objectively available to him. I think (a) is more adequate, because
it seems enough to keep a theory from being a BKT, although not as demanding as (b), since (a)
doesn’t require an act of believing.

Demanding the identification of the reason that is being used in believing p does not prevent the
Bootstrapping Problem, because the mere identification of the reason is not enough to avoid a
question that resides precisely in our capacity to identify the epistemic efficacy of our reasons.
Undoubtedly, identifying the reason used – which also goes for the method of belief formation –
is a necessary condition, since we couldn’t have justification, or a justified belief, that the reason
r is objectively available if we are not aware of which reason, or method, is being used to support
some particular belief p. What is required to avoid the Bootstrapping Problem is some
justification for a belief about the epistemic effectiveness of the reasons we use to hold some
particular belief p. Therefore, in order to avoid Bootstrapping, we would have to uphold the two
first conditions offered by Klein, namely, (1) S believes r and (2) r is a [objectively available]
reason for p, but add that S must have justification to believe that the reason r is objectively
available to him – provided this last condition implies that S is capable of identifying the reasons
he actually uses for believing p.

If the absence of a requirement according to which S would need to have justification to believe
that the reason r is objectively available to him in order to have justification to believe p on the
basis of r correctly depicts Klein’s proposal, then Infinitism should be characterized as a BKT.
Any BKT will deny that we have to have justification that our reason to believe p really does
justify p in order to have justification to believe p on the basis of that reason in the first place.
Infinitism, having the basic knowledge structure, allows that a reason r, being objectively
available, grants justification to S before S knows that r is objectively available to him. Then, S
could know p without knowing that he knows p. (Granted, S could come to know that he knows
p if S knows that r is an appropriate reason to believe p).

Infinitism allows for basic knowledge in the same way Reliabilism does. As we can see, in this
respect Infinitism is identical to Reliabilism. In fact, the only difference between them is that,
with regard to Infinitism, we will talk about the objective availability of some reason, and with
regard to Reliabilism, we will talk about the reliability of some belief formation process instead.
Nonetheless, this difference is just not of consequence. What is central to both these theories is
the absence of epistemic priority of second-order knowledge over first-order knowledge.
Consequently, Reliabilism as well as Infinitism allow for BK and then permit the acquisition of
metaknowledge on the basis of BK.

Basing knowledge of the reliability of a process on a belief that results from that same process
must be equivalent to using a belief p which is objectively justified by reason r to show that r is
an objectively available reason to believe p. In both cases, what we have is the same: basic
knowledge illegitimately generating metaknowledge. Again, what makes this possible is that it is
not necessary for S to have some justification to believe that r is an objectively available reason
to believe p in order to have justification to believe p on the basis of r in the first place.
Whenever we are dealing with the reliability of a belief formation process, we could bootstrap in
order to come to know that the process is reliable; whenever we talk about reasons, we could
bootstrap in order to come to know that we have objectively available reasons for our beliefs. In
both cases, the problem resides in the inadequacy of the constraints that are placed upon what is
a necessary condition for having justification for believing some proposition in the first place. In
both cases, concerning reliability and objective availability, S will be able to use his BK in order
to show that the reason he uses to justify his beliefs is in fact appropriate, exactly because it is
not required of S that he have independent justification to believe that his reason is adequate. In
other words, the problem is that just using a reliable method, or just having an objectively
available reason, is not enough to have justification for believing in the first place.

Suppose, for example, that a person, Pedro, believes (1) a snowstorm is likely in Montana, and
that the reason Pedro has to believe (1) is that (2) dark clouds are gathering over the mountains
and it is mid-winter in Montana.19 Let us suppose that (2) is objectively available for Pedro, but
that he doesn’t know that (2) is adequate to justify his belief in (1). Accordingly, Pedro may
know that a snowstorm is likely, provided that (a) (1) is true; (b) he uses (2) to base his belief in
(1) and (c) (2) is an objective available reason for believing (1). However, since he doesn’t know
(c), he doesn’t know that he knows (1). So, if Pedro believes (1), this belief will be, unbeknownst
to him, an item of knowledge – a BK – for him.

If the description above is compatible with Infinitism, Infinitism cannot prevent Pedro from
using his BK to come to know that he knows that a snowstorm is likely. This situation is
analogous to the gas gauge example: if Pedro believes (1), and his belief in (1) is supported by
(2), then Pedro knows (1). If Pedro doesn’t know that (2) is adequate to justify (1), Pedro has BK
of (1). Yet Pedro must be capable of identifying that (2) is being used to justify (1).20 So, once he
knows (1) and realizes that (1) is based on (2), Pedro can believe (3) a snowstorm is likely and
dark clouds are gathering over the mountains and it is mid-winter in Montana.

Now, nothing prevents Pedro from, on the basis of (3), coming to believe (4) that dark clouds are
gathering over the mountains and it is mid-winter in Montana is an appropriate reason to believe
that a snowstorm is likely. This is possible because Pedro, according to Infinitism, knows (1),
and, by introspection, knows that (1) is based on (2).
As was the case with the gas gauge example, I think it would be useful to see what exactly
Cohen writes about it,

In fact, the problem of bootstrapping generalizes to evidentialist theory as well. That is, it
generalizes to any evidentialist theory that allows for basic knowledge. […] Consider
evidentialist foundationalism: according to that view, I can know the table is red on the basis of
its looking red, even though I have no prior evidence that something’s looking red is a reliable
indication that it is red. But once I know the table is red, I can appeal to that fact in reasoning. A
little introspection will tell me that the table appears red. So now I know that the table looks red
and that it is red. So now I have some evidence that something’s looking red is a reliable
indication that it is red. And by taking a few more looks, I can acquire more evidence.21

It is important to be clear about this: a BKT, as Infinitism, must concede that Pedro knows two
things: (A) his belief that a snowstorm is likely is based on his belief that dark clouds are
gathering over the mountains and it is midwinter in Montana, and (B) a snowstorm is likely. In
other words, he knows that a snowstorm is likely and knows why he believes so, even though he
doesn’t know that he knows that a snowstorm is likely, i.e. even though he doesn’t know that (2)
is appropriate to justify (1). In this circumstance, since Pedro knows (A) and (B), nothing can
prevent him from believing that (4) dark clouds are gathering over the mountains and it is mid-
winter in Montana is an appropriate reason to believe that a snowstorm is likely. Now, (4) is
equivalent to knowing that one knows that a snowstorm is likely, given that knowing that the
reason to believe (1) is appropriate is equivalent to knowing that one knows (1).

In this case, Pedro suddenly knows that he knows that a snowstorm is likely, since he now knows
that his reason, namely (2), is adequate to justify his believing (1). This means that the absence
of a requirement according to which Pedro needs to have a reason to believe that (2) is adequate
for believing (1) in order to have justification for believing (1) in the first place makes it possible
for Pedro to know that he knows (1) in a way that cannot be legitimate.

What is crucial here is that given that Pedro may identify (2) as the reason to believe (1), even
without any independent reason to believe that (2) is appropriate to justify (1), if Pedro
introspects a little he may find an illegitimate way to believe that (2) is appropriate to justify (1),
and then, appears as if knowing that he knows (1).

What happens to Infinitism is the same that happens to Reliabilism: there is no sufficient
constraint about what is necessary to have justification to believe in the first place. This way to
get second-order knowledge or justification is illegitimate only because the way to get first-order
knowledge, even though it may not seem wrong from some perspective, cannot be right. The
problem with this second-order knowledge acquisition derives from too lax a set of conditions
for first-order knowledge.

I am quite sure that there are various ways to manifest dissatisfaction with the restrictions that
concern a requirement of metajustification. Take, for instance, Lehrer’s appraisal,

Examples of alleged knowledge in which a person does not know that the information he accepts
is correct may be of some philosophical interest but such knowledge falls outside the concern of
knowledge used in a way that is characteristically human in critical reasoning and the life of
reason.22

In some sense, the Bootstrapping Problem can be seen as demonstrative of what is wrong with
the sense of “knowledge” that results when one ignores whether the information accepted is
correct – or ignores that a reason is objectively available, or that a process is reliable. Therefore,
the Bootstrapping Problem just helps to establish a very welcome result for epistemology: no
theory of knowledge can claim that a person S knows p without knowing that he knows p.

Avoiding The Bootstrapping Problem

If the previous diagnosis is right, it is not too difficult to see what kind of change must be made
in order to eliminate the threat of the Bootstrapping Problem. The idea consists of reestablishing
the epistemic priority of second-order knowledge. If the root of the Bootstrapping Problem
grows from an unbalanced relation between first and second-order knowledge, all we have to do
is bring proper equilibrium to this relationship. The point, then, is just to realize that we cannot
separate knowing from knowing that one knows, and find a way to keep them adequately related.
However, to accept that we cannot separate epistemic levels will be much simpler than to explain
how exactly we must relate them.

Focusing only on what matters most to our present concerns, it must be made clear that to keep
the levels adequately related means that we will have to claim that in order to have justification
to believe some proposition p on the basis of some reason r, one has to have justification to
believe that r is a proper justifier for p. This means that we will have to have metajustification for
believing p. Or, in other words, if we have justification to believe some proposition p, because p
results from a reliable method of belief formation m, then it must be the case that we have
justification to believe that method m is reliable. This sort of metajustification requirement will
make any theory immune to the Bootstrapping Problem.

Before proceeding, let us further clarify the idea of “metajustification” or “metaepistemic

requirement.” As it will be understood here, a metajustification must be directed toward some
particular belief. Thus, in the same way we can ask for justification for the belief that p, we can
ask for metajustification for the belief that p. Justification for believing p involves the reasons we
have to believe that some proposition is true; metajustification concerns our reason to believe
that the reason we have for believing that p adequately justifies our belief that p. Let us use an
example to clarify how the metajustification requirement may work. Suppose that I believe that
people from Angola speak Portuguese, and suppose that my justification to believe that it is true
that people in Angola speak Portuguese consists of the following reason: Angola was colonized
by Portugal. The metajustification to believe that people from Angola speak Portuguese will
consist of my justification to believe that my reason, namely that Angola was colonized by
Portugal, really justifies my belief that people from Angola speak Portuguese. It could be
something like “colonies inherit the language of their colonizers.” My metajustification to
believe that people from Angola speak Portuguese consists of my justification to believe that my
reason, that Angola was colonized by Portugal, is effective in justifying my belief that in Angola
people speak Portuguese, and the proposition “colonies inherit the language of their colonizers”
plays this role.
As you may have noticed, there is a potential problem lurking around here. I will come back to it
later.

In principle, there are many different ways to describe what it is that could make any theory
bulletproof against Bootstrapping. My own suggestion is that, in conjunction with the objective
conditions for justification proposed by Klein, we will have to add something along the lines of
the second clause of Fumerton’s Principle of Inferential Justification (PIJ): to be justified in
believing one proposition P, on the basis of another proposition E, one must be (1) justified in
believing E and (2) justified in believing that E makes probable P.23

Discussing the PIJ will lead us to consider two important points that are involved in the
metajustification requirement and, as long as we make the necessary clarifications, we can deal
with those two points.

One important clarification has to do with the very terms in which Fumerton expresses the
second clause of the PIJ. I don’t think it is necessary – for our purposes here, at least – to
understand the second clause of the principle including the specific content Fumerton holds is the
right one to make r a justifier for p, namely the idea according to which r makes probable p. We
can just leave open the possibility that that which makes r a justifier for p need not be defined as
a probabilistic relation between r and p, and substitute “r makes probable p” for “r justifies p.” I
don’t think it would be wrong, in a preliminary analysis, to interpret the second clause of the PIJ
as, “in order to have justification for believing p on the basis of r, or method m, one must have
justification to believe that r justifies p, or that the method m is reliable.”

Furthermore, we should read the PIJ as stating that one has to have justification for believing, but
not necessarily a justified belief, that the reasons are adequate or the processes are reliable. Since
having justification is enough to avoid the Bootstrapping Problem, we don’t need a justified
belief to make any theory invulnerable to the problem.

We can now deal with the following points. The first point has to do with the idea that accepting
a principle like the PIJ, or any other principle that states the necessity of a metaepistemic
requirement, implies being committed to some form of “level confusion”.24

I am not quite sure about the best way of characterizing a level confusion. If to state some
relation between the levels, particularly the relation according to which we cannot be justified in
the first place without having justification for the belief that we have justification, implies
confusing epistemic levels, then the second clause of the PIJ – and, in fact, any metaepistemic
requirement – confuses epistemic levels.

However, one could hardly say that any such thing is, per se, a case of level confusion. Actually,
talking about justification without making some necessary distinctions is exactly what seems to
be at the core of level confusion. A case of level confusion would be the result of the absence of
a clear distinction between having justification to believe p and having metajustification to
believe p. The failure to keep this distinction clear seems to me a more correct way to
characterize level confusion.
The pivotal problem with the acceptance of the second clause of the PIJ, though, is that it would
generate an infinite metaregress: “[I]f finite minds should worry about the possibility of
completing one infinitely long chain of reasoning, they should be downright depressed about the
possibility of completing an infinite number of infinitely long chains of reasons”.25 The point,
then, is that the idea of completing an infinite number of infinitely long chains of reasoning
cannot be claimed as a necessary condition to be justified in believing any proposition, since it is
just impossible to complete any such infinite chain.

The criticism against Infinitism seems to take, then, two forms: one concerns the first-order
regress, and refers to the necessity of completing one infinite chain of reasoning, to which I will
refer here as one infinity. The other concerns the second-order regress, or metaregress, and refers
to completing not just one, but an infinite number of infinite chains of reasoning, to which I will
refer here as infinite infinities.

Now it is necessary to identify what exactly is the problem that could afflict Infinitism. Although
it may be a platitude to say that not every infinite series or regress is a vicious one, it seems
appropriate to insist on it. Just to remember a single example,

A well-known example of a ‘benign’ regress is the so-called ‘truth regress.’ This regress, which
can be derived simply from the Tarsky schema and plausible ancillary assumptions about what
are fit substitutions into the schema, demonstrates that there are an infinite number of truths
corresponding to each truth. For, as is commonly known, where p is some proposition, if p is
true, then ‘p is true’ is true, as is ‘‘p is true’ is true’ is true, and so on.26

So, I cannot agree with the claim that an infinite regress or series is vicious only because of the
fact that it is infinite. There must be something more than the mere property of being infinite that
makes a particular series, chain, or regress a vicious one. Hence, the mere existence of some
infinite series or regress cannot counts as a reductio of Infinitism. On the contrary, I will assume
that, in order to constitute a refutation, it must be shown why some particular infinite series,
chain, or regress is vicious.

However, the idea of completing an infinite chain of reasoning does point to what will make
Infinitism an implausible theory. Accordingly, Infinitism would seem to imply that a person
would have to have infinite beliefs in order to have a single justified belief. More precisely, one
would have to have infinite beliefs in order to have a single justified belief. In this idea,
apparently, would reside Infinitism’s most serious problem.

Now, as far as I can see, this supposed problem is not necessarily related to the metaregress. The
impossibility of completing an infinite number of infinite chains of reasoning is the very same
problem we would find if we considered only the first-order regress, i.e. considering only the
impossibility of completing one infinite chain of reasoning. To be clear, if the criticism against
Infinitism is grounded on the idea that no human being is capable of completing one infinite
chain of reasoning, then it is just unnecessary to add infinite infinities to the criticism, since just
one infinity is enough to make the point. In fact, there cannot exist any difference between the
impossibility of completion of one infinity and infinite infinities. Nothing can become more
impossible to complete than one infinity, and talking about infinite infinities will just repeat the
same point. So, the difference between first and second-order regress is irrelevant to this
criticism, since it does not depend in any way on the metaregress, i.e. it does not depend on
infinite infinities. Therefore, I do not agree that there is a difference regarding problems posed by
one infinity – the first-order regress – that according to Fumerton may makes us “worried,” and
problems posed by infinite infinities – the metaregress – that would make us not only worried,
but “downright depressed.” Actually, both present the same problem and, as far as I can see, if
we were supposed to complete one infinite chain of reasoning, we should be downright
depressed from the very beginning.

If this is right, there should be no problem on the second-level, as long as there is no problem on
the first-level. So, if Infinitism correctly deals with the first-order regress, as I think it does, we
have got a reason at least to suspect that Infinitism may provide a correct way to handle this
metaregress and, in fact, all kinds of regresses.

I will not focus on the infinitist response to the criticism according to which Infinitism implies
that we would have to complete an infinite chain of reasoning. Klein has already shown that line
of criticism misrepresents properly formulated infinitist claim.27 For present purposes, suffice it
to say that, while it is true that no one would be able to complete an infinite chain of reasoning,
Infinitism does not claim that this would have to be done. In fact, according to Infinitism, it is
not necessary to complete an infinite chain of reasoning in order to have a justified belief.

I just would like to explain a little further why the idea that there must exist an infinite number of
reasons available for S, if S has justification to believe some proposition p, does not imply that S
will be committed to providing reasons nonstop all the time.

Klein has claimed that sometimes we are just not involved with the process of justification, and
because of this, we are not supposed to provide reasons all the time. He must be right about this,
although I think the explanation of why we can stop providing reasons and still be considered
non-dogmatics is not yet fully developed.28 Of course, getting tired, needing to sleep or even
dieing, could explain why we sometimes stop providing reasons for our beliefs.29 However,
claiming this would not explain why we stop providing reasons in a rational or non-dogmatic
fashion. So, this sort of suggestion is sufficient to show why we sometimes stop providing
reasons, but it is not enough to explain why we can do so in a non-dogmatic way.

I think the best way to give the required explanation is by focusing on the fact that we can stop
providing reasons in a non-dogmatic fashion once we eliminate whatever it is that creates a need
to provide reasons in the first place. This means that the need to provide reasons – or to create
reasons, if it were the case – must respect some criterion. As far as I can see, this criterion seems
to be that of doubt, which in turn is generated by disagreements, dialectical or in soliloquy. I will
not try to offer here a complete explanation of this point, but it is imperative to understand the
stopping places of Infinitism in these terms: disagreements create doubts, and doubts generate
the need to provide or create reasons. Without disagreement, that is, without doubt, there is no
need to provide reasons, but this has nothing to do with the requirement according to which if S
has justification to believe some proposition p, then there must be an infinite number of reasons
available to him. This means that, when doubt is eliminated, we can rationally stop providing
reasons. On the other hand, if we still have doubts about something, we can stop providing
reasons at will, but in no way would this manner of stopping avoid dogmatism. It is not rational
to stop providing reasons when I am fatigued if a disagreement is taking place, because that
which generates the need to provide reasons still persists. On the other hand, I can find a rational
stopping place for providing reasons if I no longer find disagreements. Of course, this kind of
stopping place is provisional, since the absence of disagreement today does not guarantee an
agreement tomorrow. In other words, I have no assurances that a belief that is bedrock to me at
present will be a bedrock belief to me in the future. However, this provisional stopping place
does not seem either problematic or mysterious, but quite natural indeed. So, we should say that,
according to Infinitism, justification is not provisional, but the rational stopping places are.

In any case, the important point is that the same kind of response that will be given to the
problem of the impossibility of completing an infinite chain of reasoning when it arises in the
first-order regress can be given with respect to the second-order regress, namely, that we are not
supposed either to complete an infinite chain of reasoning, to be justified in believing that some
proposition is true, or to provide reasons non-stop all the time.

To my mind, the most interesting problem here is that Klein himself has said that the second
clause of the PIJ would lead to the rejection of Infinitism: “[I]t is easy to see that if this condition
[the second clause of the PIJ] were coupled with Infinitism, the consequence would be that any
person having a justified belief must have a belief that gets ‘so complex’ that no human could
ever have it”. 30 Then, according to Klein, the acceptance of the second clause of the PIJ would
lead to unbearably complex beliefs.

This is interesting for many reasons. I suspect it reveals a blurry relationship between Infinitism
and the metaepistemic requirement. While Klein bases the criticism against Foundationalism on
the need for metaepistemic justification, he refuses to accept the second clause of the PIJ that just
represents one way to pose the metaepistemic requirement.

The criticism presented by Klein against Foundationalism does require metajustification. When
Klein asks the foundationalist, “why do you think that the basic proposition b is likely to be
true?” and the foundationalist replies, “because b has some property f that makes b
autonomously warranted,” Klein will insist on asking if this property f is a reason to think that b
is likely to be true.31 The absence of an answer for that last question entitles Klein to charge this
type of Foundationalism with dogmatism. This means that Klein is requiring some
metajustification: he is asking why the foundationalist thinks that this property f is adequate to
make his belief b justified for him. Moreover, Klein seems to assume that it will be necessary to
have this sort of metajustification in order to be justified in believing any proposition in the first
place. If this is the case, then it highlights the fact that Infinitism presupposes a metaepistemic
requirement, or, in other words, that the absence of metajustification makes Foundationalism a
theory that is impossible to uphold in a non-dogmatic way.

This result is not surprising, since the criticism presented by Klein seems to be Pyrrhonian in
essence. In fact, we could say that Klein is confronting Foundationalism with the Problem of the
Criterion, because his criticism appears to be based on the impossibility of establishing a gap
between first and second-order knowledge or justification. While the Pyrrhonians are committed
to the view that it is impossible to know something without knowing that one knows, Klein is
saying to the Foundationalist that we cannot be rational without having a reason to believe that
what justifies our beliefs really does justify them.

The necessity of a metaepistemic requirement would be a very welcome desideratum of

Infinitism. However, Klein thinks that the second clause of the PIJ will lead to the rejection of
Infinitism by generating unbearably complex beliefs.

But the problem of overcomplexity may be just apparent, and could be resolved with the idea
that what is necessary is the existence of an infinite number of adequate reasons that may have
different functions, such as reasons to believe p and reasons to believe that the reasons to believe
p are adequate. This would not be a case of over-complexity, but only a case of availability of
reasons with different functions.

Let us use an example to see both how complex our beliefs could become and if we really can
have justification without metajustification. Suppose that someone believes that women are
inferior to men on the basis of the belief that no woman has ever been president of the United
States. What is claimed is that in order to have justification to believe that women are inferior to
men, this person would have to have justification for believing that that reason, that no woman
has ever been president of the US, really justifies the belief that women are inferior to men.

Let us look closely at this case. If S believes that women are inferior to men, S is supposed to
have a reason, or reasons, to believe so. It is quite easy to imagine a disagreement about whether
women are inferior to men. This situation creates a need to provide reasons, i.e. it inaugurates for
S the process of justification of his belief that women are inferior to men. Then, S could say that
he believes that women are inferior to men because no woman has ever been president of the US.
This answer, as I think is clear, will not be enough to settle the matter. What, then, is S supposed
to do after that?

First of all, S is supposed to have a reason to believe that it is true that no woman has been
president of the US. Let us grant that S does, in fact, have an appropriate reason to believe that
no woman has ever been president of the US. Even so, that is obviously not enough to settle the
question about women’s inferiority. Having a reason to believe that no woman has ever been
president of the US would at best (let us suppose) be considered necessary, but by no means
sufficient, for a justified belief that women are inferior to men. This is a very important point
because it shows that the first clause of the PIJ is not enough to capture what it is that amounts to
having justification. What seems to be the case is that S 126 is supposed to answer the following
question: why do you think the reason “no woman has ever been president of the US” is in any
way appropriate to justify the belief that women are inferior to men? Answering this question is
decisive, given that a different reasoning could go: I believe that women are superior to men, and
my reason to believe so is precisely that no woman has ever been president of the US. All that
matters, again, is answering why S thinks that the reason “no woman has ever been president of
the US” is appropriate to justify the belief that women are superior to men.

The required metajustification can be achieved by one’s providing a reason about the epistemic
effectiveness of “no woman has ever been president of the US” for justifying “women are
inferior to men.” Now, why do we have to think that this reason will be unbearably complex? I
don’t think it is, because all S needs is just another reason. This metareason (it will be a
metareason if we are considering the belief that women are inferior to men) could be something
like “the capacity of being a president determines gender superiority.” That would provide what
the situation calls for.

Of course, we don’t have to think that the regress will be stopped at this point. On the contrary,
there is no reason to think that it will. If there is some doubt about the effectiveness of the
metareason, S will have to provide more reasons, if S has justification for believing that women
are inferior to men. The next step would be to provide a reason why S thinks that being president
determines gender superiority. But I can see no over-complexity here, since all we need in every
step is just more simple reasons. What we have to understand is that the propositions that serve
as reasons may have different functions. Then, we will still deal with individual reasons, reasons
that play different roles, but not obviously complex enough as to become impossible for a person
to grasp.

As long as we are interested in pushing the process of justification further, we would find the
same scenario with the same requirements. All we have to require is an infinite number of
available reasons, and understand that those reasons may serve different functions: on the one
hand, to justify, and, on the other, to serve as metajustification. The proposition that serves as
metajustification for some belief will not be more complex than the proposition that serves as a
reason to believe that proposition.

This example, I hope, shows that we cannot separate epistemic levels. It shows that no one can
be justified in believing that women are inferior to men if one doesn’t have metajustification –
i.e. if one doesn’t have justification to believe that the fact that no woman has ever been
president of the US adequately justifies the belief that women are inferior to men. If we don’t
have justification to believe that our reasons are effective in justifying our beliefs, then we don’t
have justification at all. The opposite view give us an odd result: S is justified in believing that
women are inferior to men, but S doesn’t need justification to believe that the reason that
supports his belief is appropriate! The point is that justification must come along with
metajustification, given that it is not possible to be justified in believing in the absence of
metajustification.

We could think that Klein’s non-acceptance of the second clause of the PIJ has to do only with
its content, and not with the idea of a metaepistemic requirement in itself. I bet this is probably
true. However, the rejection of the metaepistemic requirement is not only explained by the
rejection of the second clause of the PIJ, but, maybe principally, by the very idea that it is not
necessary to have justification that the reason r is objectively available to S in order for S to have
justification to believe p on the basis of r. be that as it may, the point is not only that we may
require a metaepistemic constraint, but that we must do so if we want to avoid the Bootstrapping
Problem and if we want to use the sort of criticism Klein uses against Foundationalism.

In conclusion, an Infinitist should accept the following: if S has justification to believe p, then S
has metajustification to believe p, which means that if S has justification to believe p, then S has
an infinite number of available reasons to believe p, and this “infinite number of available
reasons” must include reasons with different functions, in particular, the reasons according to
which S’s reasons are effective – in other words, reasons according to which the reasons used by
S are objectively available.

This conclusion appears to be in accordance with something like the second clause of the PIJ –
which will avoid the Bootstrapping Problem – as well as in accordance with the criticism
presented by Klein against Foundationalism.

In any case, I will not push this point further right now, because a defense of the inclusion of a
specific metaepistemic requirement along with the other requirements of Infinitism will exceed
the scope of this essay. The fundamental point is that Infinitism is capable of adopting the
metaepistemic constraint, thus becoming a non-BKT view. This means that Infinitism is capable
of avoiding the Bootstrapping Problem.

Conclusion

If Bootstrapping really is a problem, it is the result of having forgotten, or inverted, epistemic

priority. The Bootstrapping Problem has been created by a theoretical intent that aims to provide
solutions for a certain predicament, but fails to capture a fundamental aspect of the phenomenon
of knowledge. Any resolution of the Bootstrapping Problem must reestablish the epistemic
priority of second-order knowledge or justification. This can be achieved by imposing some
constraints according to which we cannot know without properly knowing that we know.

This metajustification, or metaknowledge, requirement may leave us to face one skeptical threat,
the infinite regress of metajustification. I think in this case we only have two options: to think
that Skepticism is the right position, since we cannot attain appropriate justification for believing
or else to think that we can deal with higher-order regresses in pretty much the same way
Infinitism handles first-order regress. As long as we understand what exactly has to be infinite in
Infinitism, we will see a way to deal with both kinds of regresses. Infinitism seems to be the view
that can best accommodate the metaepistemic requirement that is necessary to avoid the
Bootstrapping Problem. However, that is not my conclusion for now. My conclusion here is that
the only way to avoid the Bootstrapping Problem is adopting a metaepistemic requirement
according to which we cannot separate epistemic levels.32

1 For ease of exposition, I will speak about knowledge of reliability as if it amounted to knowing
that one knows. I am assuming that to know that a process is reliable is equivalent to knowing
that one knows. In the same way, I will assume that knowing that an evidence or reason is
adequate is also equivalent to knowing that one knows.

2 Chisholm, Roderick, The Foundations of Knowing, (University of Minnesota Press,

Minneapolis, 1983), p.62.

3 I will take the expression “second-order knowledge” and “metaknowledge” as synonymous

with “knowing that one knows”.
4 Cohen, Stewart, “Why Basic Knowledge is Easy Knowledge”, Philosophy and
Phenomenological Research, 2005, p. 417.

5 Cohen, Stewart, “Basic Knowledge and the problem of Easy Knowledge”, Philosophy and
Phenomenological Research, 2002, p. 310.

6 Cohen, Stewart (2002, p. 311)

7 According to Cohen, the BKT will face what he called the “Problem of Easy Knowledge”, and
Bootstrapping would be one way to raise this problem. The other way would be the Closure
Principle. I will not consider the problem that supposedly arises from the Closure Principle. The
main reason for that is that the sort of knowledge easily acquired by Closure does not seem to be
second-order knowledge. This means that the resolution of that “version” of the Problem of Easy
Knowledge should be completely different from the Bootstrapping problem. By the way, I think
the analysis provided by Klein in “Closure Matters: academic skepticism and easy knowledge”
Philosophical Issues, 2004, is essentially correct. As far as I understand, the entire problem
hinges on deciding under which circumstances it would be right to reason from “the table next to
me seems red” to “the table next to me is red”, and not on the reasoning that involves the Closure
Principle, that would go from “the table next to me is red” to “the table next to me is not white
but illuminated by red lights”. Then, what the Problem of easy knowledge calls for is an analysis
of that first way of reasoning. As Klein suggested, much can be resolved with the idea according
to which the Closure Principle does not specify a priority among reasons and, then, the reason
available to believe that the table is red need not be the same reason to believe that the table is
not white but illuminated by red lights. Anyway, I will not take into account in what follows the
questions that involve the Closure Principle.

8 I will try offering an example similar to Jonathan Vogel’s in “Reliabilism Leveled”, Journal of
Philosophy, 2000.

9 Vogel, Jonathan, “Reliabilism Leveled”, The Journal of Philosophy, 2000, p. 615.

10 There is something tricky here. One may easily overlook the difference between “reading
accurately” and “working properly.” We should bear in mind that knowing that “the gauge is
reading accurately” is not really the same as knowing that “the gauge is working properly,” since
it is possible that the gauge reads accurately on some occasions even when it is broken—i.e. the
gauge can accidentally read accurately.

11 In fact, Michael would need more inferences in order to make an inductive inference that the
gauge is reliable on all occasions. However, those are ancillary inferences that will present no
problem to this kind of analysis. All he would need is to make more inferences of the same type:
“now, the gauge is working well,” and after a certain amount of inferences of this kind, he would
be able to conclude that the gauge is working properly all the time.

12 This point was also made by Bergman, Michael, “Externalism and Skepticism”, The
Philosophical Review, 2000.
13 For example Klein, Peter, “When Infinite Regresses Are Not Vicious”, Philosophy and
Phenomenological Research, 2003, p 728.

14 Klein, Peter (2003, p 728).

15 Klein, Peter, “Human Knowledge and the infinite regress of reasons”, Philosophical
Perspectives, 1999, p. 299.

16 Klein, Peter, (1999, p. 299)

17 Klein, Peter, (1999, p. 322)

18 Klein, Peter, (1999, p. 322)

19 This example was used by Klein to show what would be an objectively available reason in
“Human knowledge and the infinite regress of reasons”.

20 It would be very odd to require that Pedro must have justification for believing (2), and that
his believing (2) has to have a causal role in believing (1), if Pedro has a justified belief that (1)
without his being at least capable of identifying (2) as his reason to believe (1).

21 Cohen, Stewart, (2002, p.318)

22 Lehrer, Keith, Theory of Knowledge (Westview Press, Boulder, 2000). p. 41. My emphasis.

23 Fumerton, Richard, Metaepistemology and Skepticism (Rowman and Littlefield Publishers,

1995), p. 36. I think clause (1) is not problematic and it is contained in Klein’s conditions
discussed above.

24 Huemer, Mike, “Fumerton’s Principle of Inferential Justification”, Journal of Philosophical

Research, 2002.

25 Fumerton, Richard, (1995, p. 57)

26 Nolan, Daniel, “What’s Wrong With Infinite Regresses?”, Metaphilosophy, 2001, p. 524.

27 Klein, Peter, 2003.

28 I think Klein himself would agree with that.

29 Klein, Peter, “Human Knowledge and the Infinite Progress of Reasoning”. Forthcoming.

30 Klein, Peter, (1999, 309)

31 Klein, Peter, “What IS Wrong with Foundationalism Is That It Cannot Solve the Epistemic
Regress Problem”, Philosophy and Phenomenological Research, 2004.
32 This paper couldn’t have been written without the help of Laura Catelli. I am also indebted to
Claudio Almeida, for his many suggestions, and to Julio Burdzinski, for his help with the first
version of this paper.
SCOTT F AIKIN
WHO IS AFRAID OF EPISTEMOLOGY’S REGRESS PROBLEM?
Philosophical Studies (2005) 126:191–217

ABSTRACT. What follows is a taxonomy of arguments that regresses of inferential justification

are vicious. They fall out into four general classes: (A) conceptual arguments from
incompleteness, (B) conceptual arguments from arbitrariness, (C) ought-implies-can arguments
from human quantitative incapacities, and (D) ought-implies can arguments from human
qualitative incapacities. They fail with a developed theory of ‘‘infinitism’’ consistent with
valuational pluralism and modest epistemic foundationalism.

Insofar as we strive to be rational, we strive to believe on the basis of good reasons. For those
reasons to be good, they must not only support our first belief, but they themselves must also be
believed for good reasons. This is where we begin to see a disturbing pattern. If that first belief is
to be held on the basis of good reasons, it seems we are in need of a very long chain of reasons.
This is a rough and ready picture of the regress problem. It seems endemic to the project of
believing on the basis of reasons. And thereby, it seems endemic to the very project of being
rational.

This problem is old. Aristotle made it famous when he used it to show the necessity for first
principles (A. Post. 72b 6–15).1 Agrippa and Sextus Empiricus made it infamous when they used
it to show the inescapability of skepticism (DL II.88–90 and PH 1.164–177). And the problem is
not just old, it is deep. Children, when they understand the game of giving and asking for
reasons, see its protean ability to manifest itself anywhere, and they easily exploit it by
continuously asking ‘‘why?’’ And it is not just deep, it is obvious. In many cases, the
obviousness of the problem is fodder for anti-intellectualism and irrationalism. Some may look
cursorily at the project of giving reasons, see the problem, and abjure the project from the start.
Given the regress problem, some often say, so much the worse for that notion of rationality.
Better faith, or revelation, or hedonism, or carelessness and inattention… or externalism.

But for as old and deep and obvious as the problem is, only recently have we made progress
understanding how and why it really is a problem. The longest-standing and most intuitive
starting point with the regress problem is to just take the regress as absurd. Aristotle and Sextus
(and presumably, his Stoic interlocutors2) just took it for granted. Modern and contemporary
epistemologists that motivate tensions between foundationalism and coherentism also often
assume that giving reasons on to infinity is patently absurd. A few, though, have not run past the
issue and have bothered to give arguments for why there cannot be an infinite series of
inferentially justified beliefs. What follows is a taxonomy of those arguments and a series of
criticisms.

I will argue three theses: (1) Arguments against the possibility of epistemic infinitism3 come in
two basic forms: one I will call ‘‘conceptual arguments,’’ the other ‘‘ought-implies-can
arguments.’’ Conceptual arguments proceed either (A) from the incompleteness of the regress, or
(B) from some arbitrary conclusion justified by it. Ought-implies-can arguments proceed from
the fact of human finitude in (C) the quantity of beliefs, or (D) the quality of beliefs necessary to
amount of epistemic justification. (2) Present instantiations of these arguments are either
incompletely probative against or are irrelevant to a developed theory of epistemic infinitism
consistent with valuational pluralism and weak foundationalism. And, (3) epistemic infinitism
offers a positive theory of justification, not a negative theory, so it is not the cynical
epistemology we may suspect. Instead, it offers a safe haven for evidentialism as an ethics of
belief.

This is not the first account or defense of epistemic infinitism. It is often noted that Peirce (1965)
and Popper (1935) may plausibly be read as infinitists, and a number of epistemologists have
expressed a sympathy with the position, for example, Reichenbach (1952), Foley (1978), and
Hardwig (1988). Dale Jaquette (1996) and Peter Klein (1999, 2000, 2003) and are the only
philosophers explicitly articulating and defending infinitist epistemologies. In the case of
Jaquette, a good deal of the defense rests on rejecting the semantics of classical logic. I will show
that a more modest means of defending infinitism is possible. In the case of Klein, the defense
depends on the unexceptionablity and details of two requirements of justification, the Principle
of Avoiding Circularity and the Principle of Avoiding Arbitrariness (1999, pp. 298–299). I hope
to show here that a more robust defense of infinitism is possible, in that the argument can often
be taken into and against the premises driving the objections.

Epistemologies posited on the viciousness of an infinite regress of inferential justification I will

call ‘‘finitist.’’4 Arguments for finitism come in two families: conceptual arguments and ought-
implies-can arguments. Conceptual arguments start from the deep, and I think right, intuition that
epistemic justification should be truth-conducive. That is, that epistemic justification aims at
(and is a reliable marker of) truth. Then the argument turns to how infinite regresses are not
truth-conducive. There are two ways this can be done: either by rooting the problem in the
incompleteness of the regress (so it is not conducive of anything) or by rooting the problem in
the constitution of the regress itself (in that because it is infinite, the regress allows arbitrariness,
so it is just as likely to justify a false belief as it is a true one). Conceptual arguments, then,
appeal only to the incompatibility of the concept of epistemic justification and regresses.

Ought-implies-can arguments appeal to facts beyond the concept of epistemic justification,

namely, the fact of human finitude. Ought-implies-can arguments start from the premise that any
obligation, if it is truly an obligation, is one we can perform. If we cannot perform what is asked
of us, then that task is not truly obligatory. There are two ways to show we cannot perform the
task of completing the series of infinite inferential justification. On the one hand, there are
arguments that the quantity of beliefs (and inferences) necessary is beyond us (for various
reasons). This is the argument from quantitative incapacity. On the other hand, there are
arguments that the quality (or kind) of belief necessary to complete the regress appropriately is
one we simply cannot have. That is, because some belief in or about the series (and necessary for
the series to provide epistemic justification) will be so complex, we cannot have it. And thereby,
we cannot maintain the series in a way capable of amounting to epistemic justification. This is
the argument from qualitative incapacity. The complete taxonomy5 looks as follows:

Finitism Conceptual Ought-Implies-Can (epistemic justification is (X is an obligation for S only

if truth conducive) X can be executed by S) \ A. incompleteness B. arbitrariness C. quantitative
D. qualitative (not anything- (does not incapacities incapacities conducive) distinguish T and F)
[A.] Conceptual arguments from incompleteness exploit the thought behind inferential
justification that beliefs get their justificatory status from the beliefs that directly precede them
inferentially.6 As a consequence, if reasons go on to infinity, then as far as the series goes, there
will always be a further belief necessary for all the preceding beliefs to be justified. If there is no
end to the chain of beliefs, then there is no justification for that chain to inherit in the first place.
William Alston captures the argument as follows:

If there is a branch [of mediately justified beliefs] with no terminus, that means that no matter
how far we extend the branch the last element is still a belief that is mediately justified if at all.
Thus, as far as this structure goes, whenever we stop adding elements we have still not shown
that the relevant necessary condition for mediate justification of the original belief is satisfied.
Thus the structure does not exhibit the original belief as mediately justified. (1986a, p. 82)

Henry Johnstone captures the thought: ‘‘X infinitely postponed is not an X,’’ since the series of
postponements shortly becomes, ‘‘inane stammering’’ (1996, p. 96). The same kind of thought
can be captured with an analogy. Take the one R.J. Hankinson in his commentary on Sextus
uses:

Consider a train of infinite length, in which each carriage moves because the one in front of it
moves. Even supposing that fact is an adequate explanation for the movement of each carriage,
one is tempted to say, in the absence of a locomotive, that one still has no explanation for the
motion of the whole. And that metaphor might aptly be transferred to the case of justification in
general. (1995, p. 189)

These stories generally have two moves: First is the distinction between a mediate perspective
(that from the relation between one belief and another, one train-car and another) and an
external, global perspective (that on the whole chain of beliefs, the whole train of cars). From the
mediate perspective, another belief or car is always necessary to explain why one belief in
question is justified or how one car is moving. And from the global perspective, we imagine
these beliefs and cars linked together, stretching out to infinity.

And now here’s the second move. From the global perspective, Alston says we stop adding
elements, Johnstone looks at the short run, and Hankinson evokes locomotives. The problem
with the arguments is here.

Let us start with the analogy. We should grant that trains do not move without locomotives, or
something of that ilk. But this is where the analogy between causal explanation and epistemic
justification may prove too thin to be probative. First, analogies between causal explanations and
structures of justification can be elucidatory and powerful, but they are not independent
arguments. Not a deep objection, really, but one that should make us aware of how certain
pictures or metaphors can do our thinking for us. Unless we investigate and endorse the aptness
of the analogy, it is empty verbiage. Here is where a second problem arises. The locomotive is a
non-regressive explanation only for certain purposes. Causal explanations certainly abound for
how the locomotive can do all that work. (Perhaps the story would include an account of what’s
going on in the boiler, how the coal and water got there, how the coal was turned into coal, how
those carbon-based molecules got to a place where they could become coal, etc.) So causal
stories are not necessarily non-regressive, anyways.

The third problem with the analogy is that it is based on the thought that explanatory regresses
and epistemic regresses share the exact (or relevantly similar) structure.7 This is a serious
presumption. One significant dissimilarity between the two structures is a modal one: that causal
stories require the actuality of their objects for them to be true, but epistemic regresses do not
require actual beliefs or speech acts in all cases. I will argue later that epistemic regresses may be
potential infinities, but the important thing to note here is that if such a dissimilarity between the
two structures is true and significant, the analogy is weakened profoundly. In most cases, the
standing problem with explanatory regresses is that they require actual infinities. If that truly is
the problem with such regresses, then the force of the analogy between explanation and
justification is weakened.

Now, it’s here that perhaps the depth of the analogy is to be felt – in the presumption against
actual infinities. The most powerful objections against such sets or series depend on the principle
of parsimony – simpler theories are more likely true than complex ones. Given that regresses
have infinite components, it seems the cards are stacked against them. Now, if parsimony not
only extends to actual objects, but also to potential ones, we can see the driving thought behind
the objection to epistemic regresses: namely that even as potential objects, they are too complex
to be true… even potentially true. Simpler alternatives must take precedence.

But I do not think the use of parsimony is a good move here. First, parsimony is a criterion for
theory-choice, not a criterion for theory-elimination.8 At this stage, when just faced with the
infinite series, there is still only one option. The criterion may be a consideration at another stage
of inquiry (like when competing theories surface), but it cannot at this stage serve to prevent
infinitism from being a contending theory of justification.

Second, I think parsimony is an invalid criterion in matters of normative judgment. It may be

right about descriptive theories – we do not want or need the world to be populated by
unnecessary objects. But it is wrong in valuational contexts, in that any pluralist about value will
insist that the criterion requiring the fewest kinds of goods is wrong-headed. Else, we must tackle
the bugbear behind the request that we explain our love for our spouses in terms of kegs of beer.
Insofar as we haven’t foreclosed the possibility of epistemic pluralism, we see we should allow
Ockham’s razor to rest on the shelf.

Johnstone’s argument is that the series, as it progresses, shortly becomes inane stammering, since
it only repeats steps (1996, pp. 96–97). But why must inferential justification repeat steps? If
steps are repeated, then it seems the rule of circularity is being broken, which is irrelevant to the
regress. Unless Johnstone can show that repeating steps is intrinsic to the series, the argument is
a red herring. Moreover, why does the argument require the perspective of the short run?
Problems may take more work than we initially suppose. How long does it take to tell a complete
story?

Alston’s argument is that when we stop adding elements (or stop moving to the next belief in the
regress), we leave the justification incomplete. That’s right. But why does he stop? If, from the
external, global perspective, it is clear that there is an infinity of beliefs to move to next, then
there is no need to stop. Alston seems to just have assumed that we need to stop, but in having
done so, he has assumed finitism, and if that is the case, he begs the question against infinitism.9
If the task is left incomplete when you stop, then do not stop.

The strategy of these arguments is that theories that postulate an infinite number of things in a
domain known finite are known vicious (Nolan, 2001, p. 531). We know that we stop giving
reasons, so we must know that an infinite series of reasons must be vicious. First, the premise
begs the question. Second, the inference is invalid, since the simple descriptive fact that we stop
giving reasons does not imply that the reasons may themselves not continue. We may leave off
giving reasons for pragmatic purposes (e.g., we may have elicited assent or stimulated action),
and continuing with reason-giving would be excessive for those purposes. But to assume that the
reasons relevant to epistemological reflection end there is already to assume infinitism is false,
and perhaps to mis-conceive the very project of epistemology as a normative (as opposed to
descriptive) discipline.

[B.] Conceptual arguments from arbitrariness also exploit the requirement that epistemic
justification be truth-conducive.10 The tactic here is to manufacture a regress of inferences and
show that in one use, the regress can be used to justify some belief, then turn to show that in
another use, the same regress can justify that belief’s negation. As a consequence, with an
infinite series of inferential-justificatory moves, we can justify any belief and its negation, and
thereby, we show that infinite regresses are not truth conducive.

There is a simple version of the argument. I will call it ‘‘the simplification reductio’’ (SR). Take
some belief that p. Justification for p can come from the belief that (p and q) from simplification.
Justification for (p and q) can come from the belief that (p and q and r). And so on. The
simplification rule allows us to add beliefs on to infinity, and since we can add ‘‘not p’’ for ‘‘p’’
in the series without any trouble, we get the reductio (Oakley, 1976, pp. 227–228; Foley, 1978,
p. 313).

There is a more complex version of the argument. I will call it ‘‘the modus ponens reductio’’
(MPR). Take some belief that p. Justification for p can come from the belief that [q and (if q then
p)]. And justification for that belief can come from the belief that (r and {if r, then [q and (if q,
then p)]}). And so on with an infinitely iterated chain of modus ponens inferences. Given that we
can replace ‘‘p’’ with ‘‘not-p’’ without damaging the chain, we get the reductio (Pollock, 1974,
pp. 28–29; Cornman, 1977, p. 290; Post, 1980, pp. 32–35, 1987, pp. 88–91).

There are a few serious objections to the two reductios. The first is to SR’s epistemic
backwardness. SR requires that p is justified on the basis of (p and q). What’s epistemically
backwards about the requirement is that if (p and q) justifies p, then p is part of its own
justificatory ancestry. So p justifies itself. As a consequence, SR is not necessarily an argument
against infinite regresses, but rather against reflexivity in justification-relations. Given the
content of the beliefs in question, the simpler belief should be part of the justificatory ancestry of
the more complex belief, not the other way around. Were these empirical beliefs, the direction of
inference should run from instances of believing that p and believing that q to believing that (p
and q).11
John Post’s version of MPR is designed to avoid this problem for SR. Pollock and Cornman’s
versions are open to the backwardness objection, but can be salvaged easily by Post’s
requirements that: (i) each belief in the series properly entails its successor, (ii) successors are
not entitled by predecessors, and (iii) no belief is justified on the basis of any beliefs succeeding
it (1980, p. 33; 1987, p. 90). So, the backwardness argument cannot work against Post’s MPR
argument.

A second problem with the SR and MPR arguments is that deductive inference rules play the role
of inferential justification. The thought is intuitive enough: if deductive validity is truth-
preservative and if epistemic justification is truth-conducive, then deductive validity is also
preservative of epistemic justification. But that inference is wrong, because neither reductio
provides a series of inferences that necessarily amount to epistemic justification. With SR,
simplifying (p and q and r and s) to p and simplifying (not-p and q and r and s) to not-p are
formally equivalent, but the two moves may still differ in some non-formal but epistemically
relevant way.12

SR and MPR both require that justification be exclusively inferential. It is a common mistake
with critics to suppose that infinitism depends on the requirement that all epistemic justification
is inferential.13 Infinitism does require that justification must have an inferential component.
That’s right. But there is nothing prima facie incoherent with the thought that an infinite regress
of inferential justification could have other factors at play than inferential relations between
beliefs. Some beliefs are more intuitive than others, some beliefs are caused by occurrent
experiential states, and some beliefs have a kind of formal character to them that all stand as sui
generis evidence. I think that this insight that drives foundationalism can be incorporated and
appreciated in most meta-epistemic theories, and it certainly can work here – namely, what may
distinguish the inconsistent regresses of SR and MPR may not be any of the doxastic states of the
subjects, but the subjects’ non-doxastic states. The state of seeing something to be true, the state
of being appeared to in a certain way, the state of being unable to think otherwise. These states
have justificatory purport only in the context of inferentially rich support, but that condition does
not mitigate their own independent, non-inferential justification. In this respect, an infinitism can
be considered to be a form of weak foundationalism (or at least that weak foundationalism and
infinitism are not necessarily inconsistent meta-epistemic theories).

Klein’s criticism of both SR and MPR is that the justifying conditions must be available to the
knowing subject as reasons (1999, p. 312). Klein’s subjective availability allows for a non-
occurrent sense of belief, where if a subject believes p, that subject would affirm or endorse p
(including sotto voce endorsements) in some restricted circumstances. Klein’s example is that S
may believe that a snowstorm’s immanent without the belief that she’s in Montana in mid-winter
looking at storm clouds gathering (p. 300). But when asked why she believes the former, she
may give the latter as an answer. My difference here is not with Klein’s notion of belief, but with
what supports those beliefs and in what way they may be dispositionally available to a subject –
some beliefs are dispositionally available to us because things seem a certain way to us. The
subject in question, when producing the inferential story for her questioner presumably does
consult with how things seem to her and fits her story properly with what those non-doxastic
states non-inferentially support.
A sketch of non-inferential support is necessary here. Modest foundationalisms are usually built
around the slogan that beliefs based on (some acceptable set of) non-doxastic states are innocent
until proven guilty. I think this thought is right, but it needs qualification. The conditions for
such asymmetry between those beliefs and ones without non-doxastic support are supplied by a
subject’s justified doxastic states. Take cases like Mu¨ller–Lyer lines. Subjects would be weakly
justified in taking the lines to be unequal only in cases where they believe on the basis of their
being appeared to equal-line-ly and the tacit belief that this is provides a good reason to believe
that the lines are equal. What defeats this weak justification is the subject being told and shown
that the sharp lines on the ends of the parallel lines distort how the lines look. In this case, the
appearance has been shown to be an inaccurate condition for believing, and beliefs to that effect
or to the contrary create a doxastic web wherein beliefs caused by the appearances can have
justificatory weight.14

The important feature of such a modest foundationalism is that these beliefs with defeasible non-
doxastic support have probabilities less than one for the subjects that consider them. If the
subject considers the beliefs defeasible, the subject is not certain. A classic argument for strong
(as opposed to modest) foundationalism is that the assignment of probability between 0 and 1
requires a doxastic backdrop. For any subject, no proposition can have a probability in isolation
– all judgments of indeterminate or uncertain probability implicitly are consequences of
weighing some evidence. If moderate foundationalism is true, then these beliefs have the
probability they do because of their relation to other beliefs. For those beliefs to have
probabilities, they must stand in some relation to others, and so on. Eventually, they must ground
out in certain beliefs (i.e., beliefs with the probability of one). Lewis’s famous use of the
argument ended with the slogan, ‘‘If anything is to be probable, then something must be certain’’
(1946, p. 186).15 Moderate foundationalism, then, on the argument, cannot provide regress-
enders. However, the moderate foundationalist position need not be posited on finding regress-
enders, but it may rather be that of capturing the role that non-doxastic states play in a subject’s
formed and available justificatory support. Strong foundationalism follows only if the tacit
rejection of infinitism on the basis of the incompleteness argument here is right. But
incompleteness arguments are not right, as shown earlier. As a consequence, if modest
foundationalism, as a theory of defeasible non-doxastic justificatory support, is right, then
infinitism has the resources to solve the problem of arbitrary regresses. Insofar as modest
foundationalism is the thesis that a subject’s non-doxastic states can play a sui generis but
defeasible role in justifying a subject’s beliefs, then the infinitist has the means of holding
arbitrariness at bay.

So finitist arguments in SR and MPR are not sufficient, since they saddle infinitism with
unnecessary commitments and derive their reductios from those commitments. Given that
infinitism is capable of accommodating the good thoughts behind foundationalism, such
arguments have no weight.16 Ernest Sosa had remarked that SR and MPR fail to distinguish
between actual and potential regresses of justification (1980, pp. 12–13). My conclusion bears a
strong resemblance to his claim, but given that non-inferential features of the regresses
distinguish them, I have provided a criterion by which knowers may pick out the actual
instantiations.17
[C.] Ought-implies-can arguments from the quantity of beliefs necessary to run the regress
exploit a crucial feature about humans: we are finite beings with finite minds.18 From this
important feature about us, we find important conclusions follow. Here are two versions of the
argument.

The finite lives argument: To run the regress of justification, some S would need infinite time
(because the inferences take time to make). S won’t live much longer than 100 years … tops!
Ought implies can. Therefore, the regress is an absurd epistemic requirement.

The finite minds argument: To run the regress of justification, some S would need an infinite
number of beliefs (since those beliefs can’t just get recycled). S has only a limited number of
beliefs (even the most opinionated person does not have that many). Ought implies can.
Therefore, the regress is an absurd epistemic requirement.

The problem with both arguments is that they require that we must occurrently think our way
through the regresses. The finite lives argument, from this requirement, then runs that we do not
have time to have all those occurrent thoughts. The finite minds argument, from this
requirement, then runs that we do not have the capacity to have all those different thoughts. But
the requirement is unnecessary baggage for a theory of justification. There is a difference
between having the ability to produce reasons and actually producing them, and it seems
excessive for an epistemology to require of a knower that she occurrently think out all her
reasons for every belief she considers to be justified.19 It seems more likely that having those
reasons, no matter how many there are (ten, ten thousand, or an infinite number), seems enough.
Giving those reasons by thinking them occurrently or speaking them out loud is a separate issue.
So the finitude of our lives is not relevant consideration.

The finite minds argument depends on the thought that a finite mind cannot have infinite beliefs.
But that simply is not true, or at least is not obviously true. Take beliefs about the successor
relation in its instantiations in mathematics. If the successor relation is right about defining
numbers, I can believe, say, about counting numbers that 2 succeeds 1, 3 succeeds 2, 4 succeeds
3, and so on to infinity. If we had a tool for understanding our empirical beliefs in this way, we
can apply this point more broadly than to just mathematical beliefs.20 Perhaps this model may do
the trick: Take the simple belief that there is a football in a specific position on a field. If the
space between it and the goal line is infinitely divisible, we have a potentially infinite number of
beliefs as to how close the football is to the goal line. It is this close, but it is capable of being
closer than that, namely this close… John Williams’s objection to the successor relation
algorithm for belief-production is that though we may be able to continue adding, there are
numbers so large that they ‘‘defeat human understanding’’ (1981, pp. 85– 86; See also Audi,
1993, p. 127, 1998, p. 189). But since we do not need such numbers, but only need indexicals to
pick out the successors here, such a problem does not arise.21 So long as we can understand that
there is a difference between each iteration of the moves, nothing defeats human understanding
here.

[D.] The final ought-implies-can argument is that from the kind of beliefs we must have to
complete the regress. The argument proceeds from the thought that for us to be responsible
reasoners, we must not only have our reasons properly arranged but we must have beliefs about
that arrangement. So, let some S believe that p on the basis of S’s belief that q and q on the basis
of r. S must not only have beliefs that p, q, and r but also S must justifiably believe that q
justifies p, r justifies q, and that p is justified by r through q. It is in this last belief that the
objection finds its roots – because when the chain is of an infinite length, that belief about the
chain would be infinitely complex. And we just cannot have thoughts like that (Foley, 1978, p.
314).

On the one hand, it seems right to require that knowers not only have reasons for their beliefs in
question but also have justified beliefs about the arrangement of those reasons. To borrow a
distinction from earlier, knowers should not only have a mediate perspective on their reasons, but
also a global perspective on them. Reasoners must not only be capable of traversing the space of
reasons but also be capable of surveying it. It also seems right to say that we can not have beliefs
so complex as to take in the whole chain. I would not challenge that thought.

But on the other hand, the requirement, in the form that produces this problem for knowers
having to survey a regress, is too stringent. There are two ways to loosen up the requirement. The
first way to weaken the requirement is to allow groupings of beliefs to be seen as wholes. So
instead of requiring that the global perspective on chains of reasons take all the reasons as
particulars, we can simply allow that some reasons can be surveyed as groups. So, say, when I
reflect on my reasons for believing that the publication of Descartes’s Meditations was a
momentous event in the history of philosophy, I do not have to look at every one of my beliefs
that support this belief individually. I can take them in as my knowledge of epistemology, my
knowledge of French and European intellectual culture, my knowledge about the history of
philosophy, and so on. So the complexity of my beliefs about my reasons is mitigated. Will this
weakened requirement help with the problem of infinitely complex beliefs? I do not think so,
since if I have got an infinite series of beliefs and if I group those individual beliefs into groups
(even very large groups), I will still end up with an infinity of groups. Hence, it won’t solve the
problem of complexity.

The second way to weaken the requirement is to say that though reason-giving can never leave
off, it may be right to allow reason-surveying to leave off at a certain point (or that the survey to
be one where we only need to have a feel for where the reasons are going). I only need a strategy
of responding, one that does not have to be worked out down to any rigorous level of detail.
Here, Klein (1999, p. 309) has argued that requiring more than this not only requires that
knowers have detailed beliefs about their reasons but also developed epistemological theories
about how those beliefs hang together. And though this is a noble pursuit for us as knowers, it is
above and beyond the call of duty. It requires that knowers be epistemologists (and good ones, at
that), and that is simply not a requirement for being a knower. Foley, himself, concedes that this
requirement would entail that epistemologists with competing theories of justification may have
to hold each other unjustified in all of their beliefs (1978, p. 314). Surely, if Chisholm was right
about foundationalism, it would not follow that Sellars was never justified in believing that the
cat’s on the mat or that it’s time for lunch.

The difference of these requirements is one of degree. Surely if knowledge is the culmination of
our responsible believing, then we must have some sort of synoptic view of how we’ve come to
and sustain that knowledge. This is different from having an epistemology or a belief about every
supporting belief, but only by degree in that the project of epistemic reflection arises out of (and
is only an amplification of the considerations giving rise to) self-aware and responsible
believing. Insofar as this difference is heeded, the qualitative argument need not refute infinitism.

Up to now, I have attended to specific versions of the ought-implies-can arguments without

questioning the ought-can premise. I intend to show how it is the real problem behind these
arguments. There are four reasons to reject it. First, it is not right to have such a constraint on our
epistemologies. At bottom, the principle insulates us from seriously engaging skeptical and other
deep challenges, and I think this is disengenuine philosophical procedure.22 Skeptics not only say
intelligible things, but also things that are real challenges to the way we take ourselves in
position to the world. And they may be right. And insofar as that is right, we’re obliged to
answer.

Second, even if any of the arguments from our qualitative or quantitative incapacities are right, it
does not follow that infinite regresses are vicious. Instead, it may just be that our cognitive duties
do not extend to full-blown epistemic justification. As a consequence, justification does not have
to be otherwise, but rather we should pursue some more modest and attainable end.23

And third, I am unsure ought really implies can. I ought to pay both my gas bill and my
telephone bill, but when I cannot, it certainly does not follow that I do not still need to pay those
bills. I ought to refrain from stealing, and my kleptomania makes it impossible to resist, but it
does not follow that the rules should be suspended. Punishment and blame may be mitigated in
such circumstances, but our incapacities do not necessarily change our obligations. In the
cognitive realm, modus tollens is a valid inference rule, and it ought to constrain the way I think
in the relevant circumstances. But there is compelling psychological evidence that humans are
incapable of such constraint consistently – e.g., the selection tasks in Wason and Johnson-Laird
(1972). It certainly does not follow that such a rule of reason be abjured in the face of the fact
that our minds are recalcitrant to constraint. In fact, I would say that such incapacities of mind
should spur us to strive to be better than we are than make due with the cold comfort that
everybody makes the same mistakes. When we reason, we do so to do it right.

When we take up the project of knowing, we are engaged in an activity that makes certain claims
upon us. Analogously, cyclists should move in certain ways, teachers ought to explain things
clearly, parents should take care of their children. The fact that there may be incompetent
cyclists, negligent parents, and obscurantist teachers does not change those requirements one jot.
Knowing, too. When we pursue it, we do so to get it right. Given that the rules outlined for these
roles are in terms of successful performance, ought-implies-can is a premise that severs the
normative relation between the goals of these practices and the requirements placed upon the
practitioners.24

Perhaps something more probative is necessary here. Let me offer a supplemental argument in
order ensure that my reasons for rejecting ought-implies-can that do not run the risk of being
taken to be a mere hortatory reason, but one rooted in the principle itself. It is a reductio from
two premisses. (1) That determinism is a conceptual possibility. That, I think, is not in need of
argument here. An implication of determinism is that there are cases where some S does not do
something because S cannot. So it is possible that there are cases where if some S fails to do
something, it is because S cannot. (2) Here might be the contentious premise. I will call it
conceptual pessimism: the realm of relevant consideration for normative inquiry (be it ethics,
epistemology, or aesthetics), must extend only to worlds where failure to do what you ought is
real. That is, normative inquiry is relevant only to worlds where there are mistakes. Worlds
without mistakes, be they heaven, the kingdom of ends, or the null set, may be relevant to
normative inquiry negatively (in that they are regulative ideals), but they cannot be worlds that
give us discrete information about normative judgment. So, (2) reads: it is conceptually
necessary that some S doesn’t do something but S ought to. I will introduce the ought-can
premise (3) for reductio: it is conceptually necessary that if S ought to do something, S can do it.

It is easy to see that (1), (2), and (3) form an inconsistent set.25 But perhaps the move should be
to reject (2). Here are some further reasons not. First, it is a weak version of the original sin
commitment, so if you are a theist of certain stripes, you are already committed to it. Second, it is
a central commitment to any theory of education that requires that training must involve mistakes
and their correction. So, if there is education, there must be mistakes. Third, if inquiry is itself a
developmental or self-correctional notion, then it analytically entails a requirement that we learn
from our mistakes. As a consequence, it is conceptually necessary that for any world with
inquiry, there are mistakes. Fourth, for any valuational pluralism, any world relevant to our
deliberation about goods and the epistemology of their proper order is a world where deliberation
is needed. Worlds where deliberation does not exist may be conceivable to us, but they
themselves are not parts of our deliberations. In the same way that total depravity is logically
possible but irrelevant to the concept of moral or epistemic normativity (see Descartes on
madness in Meditation 1 and Kant on ad populum appeals for moral justification), total
perfection is irrelevant, too. It is something that may impel our judgments to have a certain
categorical form, but these worlds cannot have any purchase on the content of our judgments.
Fifth, conceptual pessimism is related to fallibalism. It seems to be a reasonable thought that any
knower with limited evidential resources will have the attitude that any one of its beliefs could be
false, and some knowers may be of the attitude that they are quite sure that at least one is false. If
we think that fallibalism of some strength is an appropriate (and perhaps prerequisite) attitude for
knowers, then conceptual pessimism is not so objectionable. If it is reasonable for all possible
knowers (relevantly similar to us) to believe of themselves that they have false beliefs, then it is
reasonable for us to believe that of them. Though this may not be as probative as I had like it to
be, I think this proof shows clearly that the conceptual possibility of determinism, conceptual
pessimism, and the ought-can principle are inconsistent. I do not think we can reject (1) as easily
as (2) or (3). I have marshaled a few reasons to keep (2), and that is the best I can do for my
purposes here.

So far, I have criticized the standing arguments against regresses of inferential justification. I
have not given an argument for their plausibility or necessity. To close, I will outline some
positive possibilities for such an infinitist theory of justification.

First, a developed infinitism is minimally mutilative of our intuitive (internalist) commitments

about justification. We retain the thoughts that inferential justification is serial, and that its
iterations are not reflexive or symmetric but are transitive. We retain the importance of inferential
justification while making the right kinds of concessions to the good thoughts driving
foundationalism – namely, that our non-doxastic states play a role in justification. In this respect,
infinitism is a catholic epistemology – one that is woven from the best thoughts driving other
meta-epistemologies. Ironically enough, infinitism may be a common ground for a plurality of
epistemic theories.

Second, a developed infinitism also provides a groundwork for a coherent fallibalist model for
inquiry. As inquirers traverse and map the terrain of reasons, they test those reasons not only for
proper support between claims, but also for their acceptability in terms of their non-doxastic
components. Because all beliefs are supported by these chains and given that these chains are
open to inquiry and testing, it follows that the justification for any belief is open to challenge and
revision.26 Fallibalism is not a matter of principle, but it is a rule of thumb (a rule of proper
method) that ensues from proper consideration of the structural aspects of justification. This is a
step toward coherent fallibalism, because its instantiations are about the justificatory status of
beliefs, not truth. As a consequence, the justification-fallibalist avoids one paradox that plagues
the truth-fallibalist – the paradox of the preface. The justification-fallibalist believes that any one
(and perhaps a good number) of his beliefs could be unjustified, but that belief does not
contradict his commitment to their truth. There is still a tension between the two commitments,
certainly, but that tension is precisely what drives inquiry, not what (if fallibalism is put simply
in terms of truth) marks something self-defeating about fallibalism.27

Third, and finally, if the denial of the ought-can principle is the heart of infinitism, then
infinitism is a unique theory of evidentialist cognitive striving. A standing problem for
evidentialism as an ethics of belief is that it founders on the regress problem. If the regress is
absurd, then evidentialists must embrace one of the classical alternatives – foundationalism or
coherentism. Coherentism runs afoul of the intuitive rule against circularity, and foundationalism
may seem inconsistent with evidentialism, since the essence of foundationalism that that some
beliefs are exempt from reason-backing. From the skeptic’s perspective, it always looks like
foot-stomping.28 Moreover, foundationalism has seemed to many philosophers a hopeless
philosophical task – that of not only finding sufficiently stable but also sufficiently robust
foundations for knowledge, where the two objectives seem positively at odds.29 Given this
argument from elimination, non-evidentialism.30 But since (modest-foundationalist) infinitism is
still a working option for evidentialist theories, and if the criticisms of the standing alternatives
are right, then infinitism is a natural home for evidentialism. But note that infinitist evidentialism
may require more of us than we may be able to provide, and though that often suffices as an
objection to evidentialism independently, given our arguments against ought-can, such
objections are facile. As a consequence, from the perspective of a developed infinitism,
evidentialism is a theory of cognitive striving that may be quixotic, but nevertheless captures our
real obligations as thinking, rational beings.31

NOTES

1 Aristotle rejects regresses across the board. Most are simply rejected on the basis of his
principle that ‘‘Nature flees from the infinite’’ (Generation of Animals 715b.16). Other cases of
Aristotle’s rejection of various sorts of regress occur in Nichomachean Ethics 1094a.18, De
Caelo 300b.1–2, Physics 256a.15–16, and Metaphysics 1074a.29. Plato, too, seems implicitly
committed to the viciousness of at least some regresses with the third man argument in
Parmenides 84c–85b.
2 For an account of the shared premisses of Stoic and Skeptic exchanges on logic see Frede
(1983) and Barnes (1990).

3 The term ‘‘infinitism’’ came into use with Moser (1984, 1985) and Post (1987) and has been
used by McGrew (1995), Klein (1999, 2000) and Fumerton (1995).

4 The first use of the term ‘‘finitism’’ in epistemology is Johnstone (1996), but it has regular use
in metaphysics and mathematics (e.g., for any theory posited on the denial of the axiom of
infinity.)

5 The divisions in this taxonomy have been widely (but sometimes tacitly) drawn in the literature
on the possibility of epistemic recursion. Oakley (1976), Foley (1978), Williams (1981), Harker
(1984), Klein (1999 and 2000) make many of the same distinctions between standing arguments,
but there has been no systematic view of the arguments, and they are generally handled
piecemeal. A synthetic or synoptic response to the arguments requires that they be handled as
classes. Much of the work parsing the arguments has been done notably by finitists offering new
arguments and working to show the relevant difference the new argument has from previous
ones. See Oakley (1976, pp. 226–227) for the first discussion of the difference between
‘‘arguments against held beliefs’’ and ‘‘arguments against infinite series,’’ which becomes here
the distinction between ought-implies-can arguments and conceptual arguments.

6 In the standing literature, this objection has also been termed ‘‘the no starting point objection’’
(Klein, 2000, p. 204) and ‘‘the structural objection’’ (Gillett, 2003).

7 Klein (2003, p. 720) has recently criticized these arguments on the basis of conflating
explanatory and logical relations.

8 Daniel Nolan (2001, pp. 533–534) describes the use of regress arguments as tools of theory-
elimination as a return of ‘‘Aristotelian prejudice.’’ The thought is simply that regresses are
absurd simply because of their lack of economy. But it is far from clear that quantificational
extravagance is independent evidence of impossibility.

9 See Harker (1984, p. 258) for a similar argument that such a move begs the question against
the possibility of a non-vicious regress.

10 In a similar spirit, Klein (2000) has called these ‘‘Anything Goes Arguments.’’ A fine early
version of the argument is Deutscher’s, where he proposes:

Could it be one vast delusion system? Is a man reasonable in holding one belief merely because
he holds another whose propositional content is suitably related to the first, even if he holds the
second on account of a third which is suitably related to the second, and so on? Might not a man
just dream up a system and be ingenious enough to always extend his story in logical fashion?
How can the mere continuous extension of a belief system guarantee the rationality of the
members of the system? (1973, p. 6)
I take the SR and MPR to be articulations of that ‘‘logical fashion’’ of extension. The
arbitrariness argument bears a strong family resemblance to arbitrariness arguments against
contextualism, as noted by Oakley (1976, pp. 226–227).

11 Foley himself anticipates and concedes this objection (1978, p. 313).

12 Barnes (1990) has proposed this difference as a possibility, but does not pursue it.

13 This point was first noted by Harker (1984, p. 263).

14 This thought is behind the Cartesian method to open the Meditations – in order for our non-
doxastic states to have proper justificatory power, we have to prepare an appropriate doxastic
context for them. It is also behind the theory-ladenness of perception literature: if a layperson
and a physicist look into a particle accelerator’s gas chamber, both would be justified in saying
they see a certain kind of squiggle, but only the physicist is justified in saying she sees a proton,
neutron… or whatever.

15 McGrew (1995, p. 64) has a recent version of Lewis’s argument formulated explicitly as a
response to the varieties of modest foundationalism. Reichenbach (1952) offered what could be
seen as an infinitist response to the argument.

16 A further way to mitigate this difficulty with justified contradictions is to introduce a

paraconsistent (or positive-plus) conception of negation (found in Priest and Routley 1989, pp.
162–168). Such a marriage between infinite series of inferences and explanations with
paraconsistent logics is suggested by Jacquette (1996, 112–114) – though not as a solution to this
problem. I will not pursue this thought here, since I think that the epistemic solution above is
sufficient. However, it is important to acknowledge that both SR and MPR require a classical
interpretation of negation for the reductio and that such conceptions have been challenged. If the
epistemic argument above does not work, the infinitist does have a further fallback position.

17 Paul Moser’s criticism of Sosa’s argument is that since there must be information external to
the regresses to distinguish them, infinitism must fail (1985, p. 69). But this is off the mark.
Infinitism does not have to require that justification be an exclusively inferential affair. A similar
confusion is behind Nathan’s argument that the regress, in the end, must be self-supporting
(1977, p. 124). That would be so if the series were entirely (in Nathan’s words) ‘‘autonomous’’
inferential relations. But they needn’t be. Non-doxastic support is possible within the series.

18 In the standing literature, Klein calls these arguments ‘‘Finite Minds Arguments’’ (2000, p.
203). Oakley had earlier parsed this qualitative argument off from the logical arguments as
‘‘against infinite series of held beliefs’’ (1976, p. 226). Foley’s qualitative argument explicitly
evokes the distinction between the form of these arguments from the conceptual arguments
(1977).

19 Aune (1972, p. 329), Deutchser (1973, p. 6), and Alston (1986a, p. 24) have formulated this
distinction in other relevant epistemological contexts. Klein (1999, p. 300, pp. 308–309) is clear
that the requirement of occurrent beliefs for the regress is excessive, and his account of
subjective belief-availability explicitly answers this requirement.

20 The successor relation may provide an algorithm for producing beliefs in a formal context,
but it is often worried that such an algorithm cannot serve that kind of role in a more robust
empirical context. My model is designed to show that the successor relation can still serve such a
role, so long as there are indexicals capable of distinguishing truth conditions. Klein suggests
that indexicals could play that role (1999, pp. 307–308). His example is that we allow that there
be an infinite set of a-shaped things, and that a subject could go through the set successively and
say, ‘‘This is a-shaped.’’ Another possible example could be that if space is infinitely extended,
let some subject successively believe that some object is to the left of where it was previously.
Or, if we allow an infinite stretch of time, let some subject think successively, ‘‘Now I’m older
than I was before… namely, this old.’’ The point of these examples (from infinite spatial,
temporal, and set extension) is that a subject need not have infinitely complex beliefs (or ones
that merely defeat human understanding) to traverse the infinite series intelligibly.

21 In this refinement of the qualitative argument, we can see that it leads us to a consideration of
the qualitative argument – namely that the Williams argument moves from the issue of having a
number of beliefs to having beliefs of a certain kind of complexity.

22 Klein’s defense of infinitism against the finite minds arguments (2000, p. 205) also reflects
this close connection between the ought-can premise and skepticism. Klein points out that, ‘‘The
claim that we do not have an infinite number of adequate, non-repeating reasons available for our
beliefs is compatible with the infinitist’s view that the appropriate normative requirement for our
beliefs is depicted by infinitism. That type of infinitist would be a skeptical infinitist with regard
to conditional knowledge.’’ Importantly, the failure of our beliefs to actually live up to the
infinitist requirement does not undo the requirement.

23 Foley makes this concession – that his ought-can argument does not refute the skeptic (or
infinitist), but shows that we should construct a more modest theory of knowledge. Such theories
are ones of ‘‘degenerate justification’’ (1978, p. 316). Margolis makes a similar concession, but
argues that pragmatism is the alternative (1977, p. 127).

24 Feldman (2000) has the most explicit defense of epistemic requirements as ‘‘role oughts,’’
where insofar as a subject is engaged in some goal-oriented practice, the subject ought to follow
the rules of the practice that tie to that goal (p. 278). Importantly, the examples in the analogy
above are his. For arguments against ought-can on similar grounds, see Feldman (1988, pp. 240–
242) for an argument from ‘‘contractual obligation’’, Wolterstorff for an argument from
‘‘paradigm-obligations’’ (1997, p. 233), and Sinnott-Armstrong (1984) that ought only
conversationally implies can (as opposed to analytically).

25 The proof should run in reductio fashion:

(1) ) (~D fi ~C) Derived from the conceptual possibility of determinism: if determinism is true,
then it follows that there are cases where if someone doesn’t do something, it’s because that
person cannot.
(2) u (O & ~D) Conceptual Pessimism
(3) u (O fi C) Ought Implies Can
(4) u (C fi D) (1 contraposed)
(5) ) (O fi D) (3,4 modal hypothetical syllogism)
(6) u~(~O v D) (2 DeMorgan’s)
(7) u~(Ofi D) (6 Df. Conditional)
(8) ~) (O fi D) (8 Df. u~)
(9) ) (O fi D) & ~) (O fi D) (5,8 conjunction)
(10) ~ u (O fi C) (4-9 Reductio). QED
The formal features of this proof are from Saka (2000, pp. 94–95). The strength of Saka’s
operators are those of epistemic possibility. My argument here is about the conceptual relations
between failures and obligations. As a consequence, though my argument bears a strong
resemblance to Saka’s formally (precisely because it was inspired by his), my premisses and
conclusion are of a different order. One consequence of this argument, if the strength of the
operators is at the level of conceptual possibility, is that there may be duties which are
conceptually impossible to actually perform. (Thanks to the reviewer for this point.) But such
possibilities I think abound. It may be conceptually impossible to refute the skeptic, but if we
claim to know (and if knowledge entails being justified and right), we ought to. Or take theists
with philosophical inclinations – they seem to me duty-bound, insofar as they’re committed to a
God that has the relevant omni-attributes to refute the argument from evil. They’ve still got the
requirement, given their commitments, regardless (for the sake of my example) of the argument
being impossible to refute.

26 Klein also outlines the consequences of infinitist fallibalism (1999, p. 313).

27 Though the move to justification can mitigate the problems with fallibalism’s paradox of the
preface, it does not solve the self-referential paradox, which is a more serious problem for the
fallibalist. Again, the problem may be solved with the marriage of infinitism and
paraconsistency, suggested by Jacquette (1996, pp. 113–114).

28 See Sextus Empiricus on this problem for the foundationalist (PH 1. pp. 164–177). The
problem has been given new life by BonJour’s meta-justification requirement for internal
justification (1985, pp. 30–32). A similar thought drives the Sellars–Davidson argument against
non-doxastic support – that such states cannot provide justification for propositions unless they
stand in some inferential (and hence, logical) relation to beliefs (Sellars, 1963; Davidson, 1983,
p. 428). If the states do not have propositional content, then they cannot. Some doxastic
background is necessary for epistemic support at all. For the modest foundationalist – infinitist,
the regress of meta-justification need be no more vicious than the series on the first-order
justification.

29 See Rescher (1974) for this argument.

30 See Adler (2002, pp. 173–179) for an overview of this argument. Adler, instead of embracing
non-evidentialism, introduces his theory of ‘‘tacit con- firmation’’, which bears a strong
resemblance to coherentism (with both ‘‘bottom-up’’ and ‘‘top-down’’ justification). The point
here is that we needn’t devise new alternatives or embrace non-evidentialist ethics of belief in
the face of the regress argument, but accept infinitism as a real home for evidentialism.
31 Thanks to Peter Klein, John Post, Jeffrey Tlumak, Lenn Goodman, Allen Coates, James
Bednar, Robert B. Talisse, Derek Turner, Toni Nicolletti, Jason Carroll and Brian Ribeiro for
comments on earlier drafts of this paper.
Richard Fumerton
University of Iowa
CLASSICAL FOUNDATIONALISM

It's always hard to tell if a philosophical view is on the rise. However, if one is a proponent of
that view it never hurts to claim that it is and with the conversion of one of the most prominent
coherence theorists to foundationalism, it's as good a time as any to predict a resurgence of the
view. When I was asked to participate in this conference, I suggested that it might be useful to
present a kind of overview of what I take to be the main reasons so many epistemologists
abandoned classical foundationalism. It is always fun to paint with a broad stroke. Due solely to
time constraints, of course, one gets to dismiss all sorts of views one doesn't like, fail to defend
crucial premises of one's own arguments and assure one's critics that there are replies to well-
known objections not discussed. While I am interested in defending classical foundationalism,
there is a sense in which I am skeptical about the project as I have just described it. There are all
kinds of radically different epistemological views associated with "classical" foundationalism.
Each foundationalist may well band with anti-foundationalists in endorsing arguments designed
to refute all but the one correct foundationalist view. Indeed my defense of foundationalism will
proceed largely by encouraging us to abandon certain well-known versions of the view. Before
I'm done I'll no doubt have antagonized as many foundationalists as anti-foundationalists.

Classical Foundationalism:

So what makes a version of foundationalism classical? It seems plausible to begin an answer to

this question by turning to contemporary internalism/externalism debates in epistemology. Many
paradigmatic externalist epistemologies do exemplify a foundationalist structure. Let's say that a
foundationalist is someone who claims that there are noninferentially justified beliefs and that all
justified beliefs owe their justification ultimately, in part, to the existence of noninferentially
justified beliefs. A belief is noninferentially justified if its justification is not constituted by the
having of other justified beliefs. Since we are in the land of epistemic warrant, I should hasten to
add that if you have Plantinga-style reasons for being unhappy with justification as a useful
epistemic concept, you can replace reference to justification with reference to knowledge or
warrant. Interestingly enough certain versions of reliabilism and tracking accounts of knowledge
are foundationalist views, as we just defined foundationalism. Crude versions of reliabilism will
make a distinction between beliefs whose justification depends on the justification of other
beliefs, and beliefs whose justification does not. On Goldman's original view, there are beliefs
that are justified in virtue of their being produced by belief-independent unconditionally reliable
processes. These beliefs in turn can be processed by conditionally reliable belief-dependent
processes to yield additional justified beliefs. On Nozick's tracking analysis of knowledge, some
beliefs track facts where the tracking mechanism does not involve intermediate beliefs which
track facts, and we can view these beliefs as noninferentially justified. As I shall argue later,
these versions of foundationalism might recognize the validity of what I call the conceptual
regress argument for foundationalism, but I do want to distinguish these views from what I have
been calling classical foundationalism. Certainly their proponents thought that they were offering
revolutionary alternatives to traditional (foundationalist) epistemology. Shall we define, then,
classical foundationalism as foundationalism committed to internalism?
I have argued elsewhere that there is no one straightforward way of understanding the
internalism/externalism debate in epistemology.1 One can distinguish at least the following
versions of internalism (and corresponding versions of externalism):

Internal State Internalism:

Internal state internalism is the view that S's being justified in believing P is identical with S's
being in some internal state--where internal state might be understood as a state of mind
consisting of the exemplification of non-relational properties. While some classical versions of
foundationalism, Descartes's for example, might be versions of internal state internalism, there
are other traditional foundationalisms, acquaintance theories, for example, which are not in any
straightforward way versions of the view. Acquaintance with a fact, object or property is a
relation and the state which is S's being acquainted with X may be a state that involves
something "outside" of S.2

Strong Access Internalism:

In the empiricist tradition being "in the mind" seems sometimes to have been thought of as being
the object of a certain kind of knowledge. There is, then, a natural way of moving from internal
state internalism to a view often conflated with it, a view we might call strong access internalism.
Strong access internalism is the view that some condition X constitutes S's justification for
believing P only if S has or could have access to the fact that X obtains and constitutes S's
justification for believing P. If one defines one's access internalism in terms of potential access
one can distinguish as many different versions of the view as one can distinguish species of
potentiality (and one can distinguish a great many species of potentiality).

There are two importantly different versions of strong (actual or potential) access internalism.
One is unintelligible; the other merely implausible. If one maintains that for any set of conditions
X that one proposes as constitutive of S's justification for believing P those conditions must
always be fortified with some other set of conditions describing S's access to X, then the view is
hopeless. Call the satisfaction of access conditions to X, A1. Will X together with A1 constitute
justification for S to believe P? Not given the view. Our strong access requirements require
access (call it A2) to the new proposed sufficient conditions for justification X and A1. But the
conjunction of X, A1, and A2 will not constitute S's justification for believing P as the view
requires us to add access to these conditions and so on ad infinitum.

To avoid this problem, the strong access internalist must distinguish carefully a view about what
is constitutive of justification from a view about what is necessary for justification. If the view is
to be intelligible the access internalist must argue that when some set of conditions X constitute
S's justification for believing P those conditions will be such that they entail that S has access to
them. The access however, need not be part of what constitutes the justification. An analogy
might be helpful. P cannot be true unless it is true that P is true--P's truth entails (in some sense
of entails) that it is true that P is true. But it would be a serious mistake to argue that P's being
true is constituted by its being true that P is true. The correct analysis of what it is for P to be true
should not make reference to metatruths about P's truth even if the correct analysis of P's being
true must reveal why P's being true entails that it is true that P is true.
Inferential Internalism:

Lying at the heart of many classical foundationalists concern with skepticism is a view I call
inferential internalism. It is characterized by a principle concerning inferential justification:
(PIJ) To be justified in believing one proposition P on the basis of another E one must be 1)
justified in believing E and 2) justified in believing that E makes probable P
It is the second clause of the principle of inferential justification that is rejected by all paradigm
externalists and that creates enormous difficulties for the philosopher determined to avoid
skepticism. Indeed, the aid and comfort it provides the skeptic is often taken by self-proclaimed
foundationlists as a reason to reject it. Note that inferential internalism does not entail strong
access internalism.

Nonnaturalistic Internalism:

It's my own view that paradigm internalists and externalists might usefully be understood as
disagreeing over the question of whether one can reduce fundamental epistemic properties to so-
called natural properties. Naturalistic epistemologists are convinced that we can ultimately
understand epistemic properties of doxastic states in terms of facts about the causal origin of
those states. While causal considerations will undoubtedly be invoked by many internalists in
characterizing what it is to base one belief on another, their account of the critical notion of
noninferential justification is likely to rely on a concept of justification that is not to be defined
in terms of contingent facts about how the belief was produced or what the effects of having such
beliefs will be.

So what kind of internalism characterizes classical foundationalism. It depends, of course, on

one's version of foundationalism. I am leery, however, of supposing that strong access
internalism lies at the heart of the foundationalism that so dominated the history of philosophy. I
have argued that one version of strong access internalism is unintelligible so you won't be
surprised by the fact that I'm not interested in defending classical foundationalists who are
committed to that view. In fact, It's difficult to find discussion by the British empiricists (surely
almost all paradigmatic foundationalists) of levels of knowledge, justification or rational belief.
Hume may have thought that if you know directly some truth then you know that you know it
and know that you know that you know it, but I don't know where he says it. Furthermore,
depending on one's construal of access or potential access, it may be possible for externalists to
mimic access requirements for knowledge and justification. If the requirement is only of
potential access and the potentiality is understood in terms of logical possibility, I actually can't
imagine a reliabilist, for example, who shouldn't allow that whenever one knows or has a
justified belief it is logically possible for one to discover that fact. It's unlikely that a reliabilist
would embrace requirements of potential access if the potentiality is understood nomologically,
but even if such requirements were accepted by the reliabilist, we wouldn't feel that the view is
getting any closer to classical foundationalism if the access in question is analyzed in terms of
reliably produced belief. In any event it seems doubtful that people have the causal capacities
required to have justified belief if we accept potential access internalism where the potentiality is
defined in terms of lawful possibility. The view requires that it is possible for me to have
infinitely many ever increasingly complex beliefs, and speaking for myself, I have difficulty
keeping things straight when I try to form the fourth- or fifth-level metabelief.

It is important to note that some very influential arguments against traditional foundationalism,
BonJour's for example, began by embracing some version of strong access internalism and by
inviting the foundationalist to defend some version of foundationalism within the constraints of
that view. I think BonJour's argument was successful against foundationalists who accepted
strong access internalism, but, if I'm right, his argument is successful against anyone who
accepts strong access internalism including coherence theorists. The moral, of course, is that one
should reject strong access internalism in formulating one's foundationalism.

We haven't yet made much progress. We've tentatively suggested that classical foundationalists
accept some version of internalism, the view that there are noninferentially justified beliefs, and
the view that all justified beliefs owe their justification in part ultimately to the existence of these
noninferentially justified beliefs. To make further progress, we need to start distinguishing
specific versions of traditional foundationalism. Before doing that, however, perhaps we should
briefly examine the historically most influential argument for foundationalism, the familiar
regress argument.

The Regress Argument for Foundationalism:

If the only way to be justified in believing one proposition were to justifiably infer it from
another, the argument goes, we would face a vicious regress of justification. Invoking the first
clause of (PIJ), the foundationalist argues that to be justified in believing P on the basis of E1
one would need to be justified in believing E1, but if the only way to justifiably believe E1 is to
infer it from something else E2 we would be then faced with the task of justifiably believing E2,
which we could only do by inferring it from something else we justifiably believe E3, and so on
ad infinitum. If one is an inferential internalist, the regresses looming are even more
intimidating. In addition to justifiably believing E1, we would need to justifiably believe that E1
makes likely P, which we would need to infer from some proposition F1, which we would need
to infer from some proposition F2, and so on ad infinitum. We would also need to justifiably
believe that F1 makes it likely that E1 makes likely P, which we would need to infer from some
G1, which we would need to infer from some G2.... Given inferential internalism there is not one
but an infinite number of infinite regresses that seem to threaten our ability to justifiably believe
anything, on the supposition that all justification is inferential.

Peter Klein has recently argued in an unpublished paper for a view he calls infinitism. He accepts
the first clause of (PIJ) and agrees that it entails that if there are no noninferentially justified
beliefs, the existence of a justified belief would entail the existence of an infinite number of
justified beliefs, but he goes on to argue that there is nothing vicious about the regress. We can
and do have an infinite number of justified dispositional beliefs, enough and of the right sort to
allow us to have justified beliefs even if all justification is inferential. He emphasizes that one
should not insist that a chain of reasoning actually be completed for a belief to be justified. It is
enough that one is able to justify each belief in an infinitely complex hierarchy of justification by
reference to some other (dispositional) belief in the structure (but not all at once so to speak).
Klein is right that we do have an infinite number of justified beliefs, but I think he misses the real
point of the regress argument for noninferentially justified beliefs. The viciousness of the regress
is, I believe, conceptual. If one tries to analyze or understand the idea of inferential justification,
one finds oneself inevitably invoking the concept of justification to explain it. To be justified in
believing P on the basis of E just is in part to be justified in believing E. (If the inferential
internalist is correct the justification for believing P on the basis of E is also constituted by the
existence of justification to believe that E makes likely P). But so far our analysis of justification
in terms solely of inferential justification is blatantly circular. We still need some understanding
of what it is for a belief that E to be justified. If one tries to understand having a justified belief
that E in terms of its being inferred from something else we justifiably believe, we are getting
nowhere in our attempt to understand justification. We need some understanding of justification
that does not invoke the concept of justification.

I find the following analogy instructive. Suppose a philosopher argues that there is no need for
the concept of intrinsic goodness, that we can understand all goodness as instrumental. To say
that X is good is always just to say that X produces some Y which is good. Furthermore, we are
assured, we mustn't worry about a regress here, as there is no reason to suppose that there are
only a finite number of good things. With an infinite number of good things, there will always be
another good thing to make good everything that is good. Would the existence of an infinite
number of good things obviate the need to recognize a distinction between intrinsic and
instrumental goodness? No it wouldn't. We need the concept of intrinsic goodness in order to
understand the concept of instrumental goodness. We could form no idea of instrumental
goodness if the only way to understand something's being good was in terms of its leading to
something else that is good. In our search for the conceptual source of goodness we would
always be led to an idea of goodness that was itself defined in terms of an undefined idea of
goodness.

In this same way, the infinitist never gives us a non-circular analysis of justification. As an
analysis of justification it resembles a recursive definition without a base clause. Unless one can
define a concept of justified belief that does not invoke the concept of justified belief, one will
have no understanding of the way in which justified "output" depends on justified "input".

To escape the conceptual regress argument foundationalism, one can try to define inferential
justification without invoking the concept of justification. Klein's infinitist might try to argue that
all we need in order to be inferentially justified in believing P is the ability to produce a non-
question begging valid argument for P, a non-question begging valid argument for the claim that
we have produced a sound argument for P, a non-question begging valid argument for the claim
that we have produced a sound argument for the conclusion that we have produced a sound
argument for P, and so on ad infinitum. But this can't be right. I believe P and offer as my
evidence for P that a God1 believes P and everything God1 believes is true. I'm also ready and
willing to argue for the soundness of this argument. There is a God2, I claim, who believes that
my God1 argument is sound and whatever God2 believes is true. You want an argument for the
soundness of that argument, I have many different infallible Gods in my ontology, each ready to
be the subject matter of premises I need to prop up an infinitely complex edifice of beliefs, where
the lowest level is supported by everything above. I'm confident Klein wouldn't want to allow
that one can be justified in believing P through having this system of beliefs, and the reason is
that he would never concede that any of the premises in the God arguments were justifiably
believed. Inferential Justification requires justified belief in the relevant premises if the belief in
those premises is to yield justified belief in conclusions based on them.

The only other alternative to the foundationalist attempt to end the regress of justification with a
concept of noninferentially justified belief is to define some generic concept of justification that
has nothing to do with the specific way in which a belief is justified. If, for example, one takes
justification to be a fundamentally normative concept and one supposes that one can reduce
propositions describing epistemically justified beliefs to propositions describing what one ought
to believe, then one might be able to avoid the need to define inferential justification in terms of
noninferential justification. But one can't define epistemic justification in terms of what one
ought to believe for reasons that Firth, Plantinga, and others have made clear. And I don't know
of any other plausible way to define a generic concept of justification. An externalist coherence
theory of justification would do the trick, but BonJour showed us convincingly that coherence to
which one has no access (i.e. coherence about which one has no justified belief) is neither here
nor there when it comes to producing justified belief. In short I don't think one can define
inferential justification except in terms of a kind of justification that is not inferential, and that's
why I think one must embrace foundationalism (although not necessarily internalist
foundationalism).

Traditional Foundationalism and Infallible Belief:

But if we reject contemporary externalist epistemologies, how are we to understand this

fundamental concept of noninferentially justified belief? One version of internal state internalism
seeks to discover foundational knowledge in some characteristic of a belief state. Because
Descartes is often thought of as the quintessential foundationalist, it is perhaps not surprising that
critics of foundationalism have often been preoccupied with the question of whether or not there
are infallible beliefs. It's certainly true that many foundationalists either explicitly or implicitly
seemed to endorse the view that we have found the foundations of knowledge and justified belief
when we have found beliefs which cannot be mistaken. When Price introduced the notion of
sense-data, knowledge of which would be included in the foundations of empirical knowledge,
he contrasted sense data and their nonrelational properties with other sorts of things about which
one could be mistaken, implying that the way to find the correct foundations of knowledge is to
scrape away from one's beliefs all that could be false. Following Lehrer (1974) we might
formulate the following definition of infallible belief:

Ia. S's belief that P at t is infallible if S's believing P at t entails that P is true.

Let us construe the entailment broadly so that P may be said to entail Q if P formally,
analytically or synthetically entails Q.3

Although this is a perfectly intelligible definition of infallibility, it is far from clear that it has
much relevance to an attempt to understand the epistemic concept of noninferential justification.
The problems are familiar and I won't belabor them here. Every necessary truth is entailed by
every proposition, and thus if I happen to believe a necessary truth, P, the fact that I believe P
will entail that P is true. Thus, by the above definition my belief that P will be infallible
whenever P is a necessary truth even if I believe P solely on a whim. Surely this concept of
infallibility has precious little to do with whether or not my belief is justified.

To deal with this problem one could tinker with the definition of infallible belief, but once one
sees that mere entailment between the having of a belief and the truth of what is believed does
not provide justification, perhaps we should look elsewhere for the foundations of knowledge
and justified belief. By doing so we may be able to avoid a host of standard objections to
foundationalism. Although one can give a few examples of contingent propositions whose truth
is trivially entailed by their being believed, critics of infallible-belief foundationalism have
argued that there are simply not enough infallible beliefs upon which to support the edifice of
justified beliefs we are trying to erect. Thus we remember philosophers employing thought
experiments in an attempt to convince Cartesian foundationalists that their favorite candidates for
infallible beliefs were not infallible at all. Armstrong (1963) imagined us in a future
characterized by a utopian neurophysiology. We are wired to complex machines that inform us
that, despite the fact that we believe we are in pain, there is simply no indication of the sort of
neural activity associated with pain. Might it not be reasonable in such a situation to conclude
that we simply have a false belief that we are in pain? Lehrer (1974) argued that one can
genuinely (as opposed to merely verbally) confuse pains with itches and for that reason arrive at
a false belief that one is in pain.

Whatever the force of these specific thought experiments, there is a very general argument
designed to establish that the foundationalist's favorite candidates for noninferentially justified
empirical beliefs are not infallible. It is a Humean sort of argument that proceeds from the simple
observation that in the vast majority of cases, the belief that P is one state of affairs and P's being
the case is a different state of affairs. If these really are two distinct facts, then why couldn't one
have the one without the other?4 Although it does not add much to the logical force of the
argument, one can again employ our hunches about how the brain might work to rhetorically
bolster the argument. Consider again a standard candidate for an infallible empirical belief, my
belief that I am in pain now, for example. It is surely possible that the region of the brain
causally responsible for producing the belief that I am in pain is entirely different from the region
of the brain causally responsible for producing the pain. There may be a causal connection
between the occurrence of the "pain" brain event and the occurrence of the "belief" brain event,
or vice versa, but even if the causal connection holds it will be a contingent fact that it does. It
hardly seems that the neurophysiologist could discover these (or any other) causal connections
purely a priori. But if the brain state responsible for my belief that I am in pain is wholly
different from the brain state responsible for the pain, and if the connections between them are
merely nomological, then it is in principle possible to produce the one without the other. The
belief will not entail the truth of what is believed.

Infallible Justification:

The foregoing argument has a great deal of plausibility, I think, and in any event it has always
seemed strange to me to search for foundations in mere belief. What justifies me in believing that
I am in pain? The mere fact that I believe that I am in pain? Well what is it about this belief that
makes it so different from other beliefs? Why does my belief that I am in pain constitute a kind
of justification but my belief that there are ghosts does not constitute a kind of justification? The
appeal to belief as a justifier borders on a non sequitur if one is genuinely attempting to find a
useful characterization of a special kind of epistemic relation one can bear to truth that obviates
the need for inference.

As BonJour pointed out, this same lack of a genuine response seems to characterize those
foundationalists who seek to identify the source of noninferential justification with the fact that
makes the noninferentially justified belief true.5 When asked what justifies one in believing that
one is in pain, this foundationalist identifies the pain itself. But what is it about the pain that
makes it a justifier? When you believe that I am in pain, my pain doesn't justify you in believing
that I am in pain (according to most foundationalists), so there must be something different about
my relationship to my pain that enters into the account of what constitutes the justification. It is
the fact that I have a kind of access to my pain that you don't have that makes my belief
noninferentially justified while you must rely on inference. One still needs an account of what
this relation is, but before we consider such an account it is worth noting that we could have
defined the concept of infallibility in a way that makes it potentially more useful in developing a
foundationalist theory of justification. The relevant question is not whether my belief entails the
truth of what is believed. It is, rather, whether my justification entails the truth of what is
believed:

Ib. S's belief that P at t is infallible if S's justification for

believing P at t relevantly entails the truth of P.

It is necessary to qualify the entailment as relevant to circumvent the problems already discussed
in connection with Ia. Whenever I have any justification at all for believing a proposition that
turns out to be necessarily true, that justification will entail the necessary truth. But we do not
want just any sort of justification to yield infallibly justified belief even if the object of that belief
is a necessary truth. What is the difference between relevant and irrelevant entailment? This
question is notoriously difficult to answer, but intuitively it should have something to do with the
fact that would make true the proposition entailed and the fact that would make true the
proposition that entails it. More specifically, we could say that P relevantly entails Q only if the
fact that would make P true is at least a constituent of the fact that would make Q true. This
suggestion can be considered at best only preliminary, since we will obviously need a more
detailed account of facts and their constituents. That I have grey hair entails that someone has
grey hair, but is my having grey hair a constituent of the fact that is someone's having grey hair?
There is certainly a sense in which it is something one can point to in answer to the question
"What makes it true that someone has grey hair?" One cannot appropriately point to my having
grey hair as something that makes it true that two plus two equals four.

Acquaintance and Noninferential Justification:

I have suggested that neither a belief nor the fact that makes true what is believed is by itself a
plausible justification at all, let alone the kind of justification that might end a regress of
justification. Rather, we must stand in some sort of special relation to the truth of what is
believed, or more precisely, we must stand in some sort of special relation to the fact that makes
true what we believe. I have argued elsewhere that the most fundamental concept required to
make sense of traditional foundationalism is the concept of acquaintance. In order to explain my
acquaintance theory of noninferential justification, however, I must digress and sketch a highly
controversial theory of truth and intentionality.

I take the primary bearers of truth value to be thoughts (which I also refer to as propositions).
The secondary bearers of truth value are the linguistic items that express them. Thoughts I take
to be nonrelational properties of a mind or self. True thoughts correspond to or "picture" facts.
False thoughts fail to correspond. A fact is a nonlinguistic complex that consists of an entity's or
entities' exemplifying properties. The world contained facts long before it contained minds and
thoughts. In one perfectly clear sense, however, the world contained no truths before there were
conscious beings, for without conscious beings there would be no bearers of truth value. There
were facts that would have made true the relevant thoughts had they existed, and by employing
counterfactuals we can make good sense of such commonplace assertions as that it was true
hundreds of millions of years ago that there were no conscious beings.

On my view every intentional state is a thought. Believing that there are ghosts and fearing that
their are ghosts are species of the same thought that there are ghosts. We can represent true and
false belief respectively as follows:

S believes truly that P = Df 'P'*s and 'P' C P

S believes falsely that P = Df 'P'*s and it is not the case that
there exists some fact x such that
'P' C x

where s stands for S, 'P' stands for the thought that P, * indicates that the thought is a belief, C
stands for correspondence, P refers to the fact that P, and x is a variable.

This correspondence theory of truth avoids the need for such ontological nightmares as
nonexistent states of affairs to serve as the "objects" of false beliefs, and it preserves a much
more natural way of understanding the referents of sentences, analogous to the referents of
names and definite descriptions. Unlike Frege, we have no need for such mysteries as The True
and The False to serve as the referents of true and false sentences, respectively. Rather, we adopt
the more straightforward view that just as the successful use of a name refers to an individual, so
the successful — that is true — attempt to refer to the world with a descriptive sentence succeeds
in picking out a fact. Some names, such as "Pegasus," do not succeed in referring to any
individual, and some sentences, like "Dogs have eight legs," do not refer to any fact. Direct
theories of reference aside, having a referent is not necessary for having meaning, and the
thoughts that false sentences express give those sentences meaning despite the fact that they fail
to refer.

Acquaintance is not another intentional state to be construed as a nonrelational property of the

mind. Acquaintance is a sui generis relation that holds between a self and a thing, property, or
fact. To be acquainted with a fact is not by itself to have any kind of propositional knowledge or
justified belief, and for that reason I would prefer not to use the old terminology of knowledge by
acquaintance. One can be acquainted with a property or fact without even possessing the
conceptual resources to represent that fact in thought, and certainly without possessing the
ability to linguistically express that fact. But if this is true, what has acquaintance got to do with
epistemology?

In one of the most influential arguments against foundationalism, Sellars argued that the idea of
the given in traditional epistemology contains irreconcilable tensions. On the one hand, to ensure
that something's being given does not involve any other beliefs, proponents of the view want it to
be untainted by the application of concepts. The kinds of data that are given to us are also
presumably given in sense experience to all sorts of other creatures. On the other hand, the whole
doctrine of the given is designed to end the regress of justification, to give us secure foundations
for the rest of what we justifiably infer from the given. But to make sense of making inferences
from the given, the given would have to be propositional. Minimally, the given must have a truth
value. But the kind of thing that has a truth value involves the application of concepts or thought,
a capacity not possessed by at least lower-order animals.6

The solution to the dilemma presented by Sellars and others is to reemphasize that acquaintance
is not by itself an epistemic relation. Acquaintance is a relation that other animals probably bear
to properties and even facts, but it also probably does not give these animals any kind of
justification for believing anything, precisely because these other animals probably do not have
beliefs. Without thought there is no truth, and without a bearer of truth value there is nothing to
be justified or unjustified. But how does acquaintance give us noninferential justification? My
suggestion is that one has a noninferentially justified belief that P when one has the thought that
P and one is acquainted with the fact that P, the thought that P, and the relation of
correspondence holding between the thought that P and the fact that P. No single act of
acquaintance yields knowledge or justified belief, but when one has the relevant thought, the
three acts together constitute noninferential justification. When everything that is constitutive of
a thought's being true is immediately before consciousness, there is nothing more that one could
want or need to justify a belief. The state that constitutes noninferential justification is a state that
contains as constituents both the bearer of truth-value and the truth-maker.

The reader might well complain that if mere acquaintance with a fact does not constitute an
epistemic property, surely one cannot conjure up an epistemic property by multiplying acts of
acquaintance. But if this is intended to be a formal objection to the view I presented, it involves
committing the fallacy of division. Because none of the components of a complex state of affairs
constitutes the exemplification of an epistemic property, it does not follow that the complex does
not constitute the exemplification of such a property. Classical acquaintance theorists like
Russell appropriately emphasized the role of acquaintance with particulars, properties, and even
facts in grounding justification. But a fact is not a truth, and what one needs to end a regress of
justification is a direct confrontation with truth. To secure that confrontation, one needs to be
directly aware of not just a truth-maker (a fact to which a truth corresponds) but also a truth-
bearer (a thought) and the correspondence that holds between them.

Because the relations of acquaintance and correspondence that the above account appeals to are
sui generis, there is precious little one can say by way of trying to explain the concept to one
who claims not to understand it. Because acquaintance is not like any other relation, there is no
useful genus under which to subsume it. One can give examples of facts with which one is
acquainted and in this way present a kind of "ostensive" definition of acquaintance, but
philosophers who think the concept is gibberish are unlikely to find themselves acquainted with
their being acquainted with various facts. When one is acquainted with a fact, the fact is there
before consciousness. Nothing stands "between" the self and the fact. But these are metaphors
and in the end are as likely to be misleading as helpful. Correspondence, too, is sometimes
thought of as a picturing relation, but the picturing metaphor is largely responsible for the
caricature of the view one so often encounters in the cruder theories of "ideas" as pale copies of
reality. It is tempting to at least mention the metaphor of a Kodak print and the scene it depicts as
a way of explaining the relation that a true thought bears to the fact with which it corresponds,
but most thoughts are not "pictures" and the relation of correspondence has nothing to do with
any kind of similarity that holds between the thought and the fact it represents. Correspondence
is not like anything else; it cannot be informatively subsumed under a genus, and it cannot be
analyzed into any less problematic concepts.

Is acquaintance a source of infallible justification? The answer is in one sense straightforward. If

my being acquainted with the fact that P is part of what justifies me in believing P and if
acquaintance is a genuine relation that requires the existence of its relata, then when I am
acquainted with the fact that P, P is true. The fact I am acquainted with is the very fact that
makes P true. The very source of justification includes that which makes true the belief. In a way
it is this idea that makes an acquaintance foundation theory so attractive. I have no need to turn
to other beliefs to justify my belief that I am in pain because the very fact that makes the belief
true is unproblematically before consciousness, as is the correspondence that holds between my
thought and the fact. Again, everything one could possibly want or need by way of justification
is there in consciousness.

Notice that the infallibility of the justification provided by acquaintance is due to the presence of
the fact itself as a constituent of the justifier. It is interesting to note that in this respect there are
remarkable similarities between this classic version of foundationalism and at least some
paradigmatic externalist views. On certain causal theories of direct knowledge, for example, my
belief that P is justified by its being caused in the appropriate way by the fact that P, the very fact
that makes my belief true. If a causal relationship between the fact that P and my belief that P
were a kind of justification, then that justification too would be infallible. Its existence would,
trivially, entail the truth of what I believe. From the fact that a certain justification is infallible, it
does not follow that one could not mistakenly believe that one has an infallibly justified belief.
Certainly the causal theory I have just sketched would have no difficulty imagining a person who
mistakenly concluded that his belief that P was caused by the fact that P, and if the causal theory
were correct, that person could mistakenly infer that the justification in support of his belief
entailed the truth of what he believed. Similarly, I think that it is in principle possible for a
person to mistakenly conclude that he is acquainted with something actually known only through
inference. One might trust a philosopher with a mistaken epistemology, for example, and falsely,
perhaps even justifiably, believe that one is acquainted with a fact when one is not. Although this
complicates matters considerably, I also argue that it may be possible on an acquaintance theory
to have noninferential justification that does not entail the truth of what is believed. Specifically,
I have argued that one might be acquainted with a fact very similar to the fact that makes P true,
and such acquaintance might give one a justified but false belief that P. It should be clear that
this admission is perfectly compatible with the rather trivial claim that when one's justification
for believing P consists in part in being acquainted with the fact that P, that justification is
infallible in that it entails the truth of P.

If I am asked what reason I have for thinking that there is such a relation as acquaintance, I will,
of course, give the unhelpful answer that I am acquainted with such a relation. The answer is
question-begging if it is designed to convince someone that there is such a relation, but if the
view is true it would be unreasonable to expect its proponent to give any other answer. I can also
raise dialectical considerations and object to alternatives. One of the dialectical advantages of the
above view is that it can easily respond to some of the classic arguments against the existence of
noninferentially justified belief.

One of the most discussed arguments against foundationalism again focuses on concepts. There
is no truth value without concept application, the argument goes. But to apply a concept is to
make a judgment about class membership, and to make a judgment about class membership
always involves relating the thing about which the judgment is made to other paradigm members
of the class. These judgments of relevant similarity will minimally involve beliefs about the past,
and thus be inferential in character. Our reply to the argument is straightforward. To make a
judgment, say that this is red, involves having the thought that this is red, but the thought does
not involve relating this to some other thing. Indeed, it is in principle possible to produce a
thought of red in the mind of someone who has never experienced a red thing. Since language is
only a secondary and conventional means of representation, it goes without saying that the
inferential character of our judgments concerning the linguistically correct way to express a
thought are neither here nor there when it comes to the question of whether the thought
expressed can be noninferentially justified.

The intelligibility of the above account does rest on the intelligibility of a world that has structure
independent of any structure imposed by the mind. Without nonlinguistic facts that are
independent of the thoughts that represent them, one could not make sense of a relation of
acquaintance between a self and a fact, a relation that grounds direct knowledge. Indeed, I
suspect that it is concern with this idea that lies at the heart of much dissatisfaction with
traditional foundationalism. Since Kant there has always been a strong undercurrent of anti-
realism running through philosophy. The metaphor again is that of the mind imposing a structure
on reality. And there is an intuitively plausible sense in which one can genuinely wonder whether
it makes sense to ask about the number of colors that are exemplified in the world independently
of some framework provided by color concepts. But despite the periodic popularity of extreme
nominalism and anti-realism, it is surely absurd to suppose that it is even in principle possible for
a mind to force a structure on a literally unstructured world. There are indefinitely many ways to
sort the books in a library and some are just as useful as others, but there would be no way to
begin sorting books were books undifferentiated. The world comes to us with its differences.
Indeed, it comes to us with far too many differences for us to be bothered noticing all of them.
And it is in this sense that the mind does impose order on chaos. Thought is abstract in the sense
that many different actual properties can all correspond to a single thought of red. And it is up to
us how finely we want to draw our color concepts. Although I understand that the empirical
evidence is at best questionable, it is common for philosophers to call our attention to the alleged
fact that some cultures have far more finely grained color concepts than our culture. If one
distinguishes color concepts from linguistic terms to express those concepts, the empirical claim
is difficult to assess, but one must surely admit that the alleged phenomenon is in principle
possible. Given the above framework for understanding thought and truth, there would be a sense
in which the one culture would entertain truths about colors that the other culture would be
causally unable to accept. But the fact that there is good sense to be made of the relativity of
conceptual frameworks should not mislead one into thinking that the properties exemplified in
the world depend for their existence on concepts.
I have presented at length a view that I take to be the most plausible version of classical
foundationalism. In what sense, if any, is it internalist? Well, on the specific version of the view I
defended, thought is an internal property of the mind, if by "internal" one means "nonrelational."
The crucial concepts of acquaintance and correspondence, however, are relational. It is true that
given my own views in normative epistemology, it turns out that it is always a mental state or
feature of a mental state with which we are acquainted, and so the complex act of being
acquainted with X will involve constituents all of which are "internal" to the subject. But it
should be emphasized that the metaepistemological acquaintance theory of noninferential
justification does not by itself entail any position with respect to what might be the objects of
acquaintance. In previous discussion of the attempt to define internalism in terms of internal
states being sufficient for justification, I noted that one might be a sense-datum theorist who
thinks one can be directly acquainted with the fact that the surface of a physical object
exemplifies a certain property. One might think that there are mind-independent universals and
claim to be acquainted with them. One might think that there are mind-independent,
nonoccurrent states of affairs and claim to be acquainted with logical relations that hold between
them. It is at best unclear as to whether or not any of the above acts of acquaintance should be
called internal states, and thus equally unclear as to whether a foundationalism defined using the
concept of acquaintance is always going to be a species of "internal state" internalism.

As we saw earlier, being internal might also be understood in terms of access. Does an
acquaintance theory hold that when one is noninferentially justified in believing P one has access
to—that is, knowledge or justified belief about—the fact that one has such justification? No. But
as we saw earlier it's good that the answer is "No". In the paradigm case, I am noninferentially
justified in believing P when I have the thought that P and am simultaneously acquainted with
the thought that P, the fact that P, and the relation of correspondence holding between them. To
have noninferential justification for believing that I am noninferentially justified in believing P, I
must have that rather complex thought and simultaneously be acquainted with its correspondence
to an equally complex fact. And for me to be noninferentially justified in believing that I am
noninferentially justified in believing that I am noninferentially justified in believing that P, I
must be acquainted with facts so complex as to boggle my poor consciousness. The position that
in order to have a noninferentially justified belief on an acquaintance theory one must be
noninferentially justified in believing that one has such justification invites a vicious regress of
infinitely many, increasingly complex conscious states.

If acquaintance foundationalism is neither a form of internal state internalism nor strong access
internalism, why call it internalism at all? The answer I suggested earlier points to its reliance on
the sui generis concept of acquaintance that is fundamental to epistemology and that cannot be
reduced to nonepistemic concepts, particularly the nomological concepts upon which all
externalists build their analyses.

Foundationalism and Skepticism:

Many critics of traditional foundationalism will no doubt be delighted if a defense of

foundationalism turns on a defense of the acquaintance theory sketched above replete with its
unfashionable ontological commitments. A detailed defense of foundationalism, however,
requires a detailed foundationalist view. I want to close, however, by at least mentioning another
reason so many philosophers abandoned foundationalism, one that has very little to do with the
intrinsic plausibility of an account of noninferential justification. It has long been clear to me that
the primary reservations many anti-foundationalists have with foundationalism, particularly
foundationalism coupled with inferential internalism, is their conviction that the view will
inevitably lead to skepticism. This is really the subject of an entirely different talk--one I
considered giving. Since I didn't give it, I have time only for a short answer to the objection that
traditional foundationalism leads to skepticism. The answer is that it probably does, but that it is
matter of considerable philosophical hubris to reject an epistemological view for that reason. If
we had the time, I would try to convince you that the principle of inferential justification states
standards of justification that are reflected in all kinds of commonplace inquiries. One cannot
simply reject it because we can't figure out how to secure noninferential knowledge of
probabilistic connections at those fundamental levels where our habits of inference are
unreflective--the levels at which philosophical inquiry begins. But epistemology is a distinctly
philosophical enterprise. We are asking odd questions because they are so fundamental and we
are insisting on rigorous standards for what constitutes an acceptable answer to these questions.
In commonplace inquiries we simply assume knowledge of the past which we remember, the
future based on inductive inference, the external world based on perception. The question is
whether we can move from this data to other conclusions employing inferences we give
ourselves as legitimate. When we start doing philosophy, we stop getting gifts. We must justify
what we normally don't bother to justify and it may not be possible to do it. To paraphrase
Hume, it may be that nature has been wise enough not to leave our ability to form beliefs about
the world around us to "uncertain" reason.

1 See Fumerton (1995), especially chapters 3 and 4.

2 While many acquaintance theorists held that one is only acquainted with one's internal states,
there is nothing in the view that precludes the possibility of being acquainted with external
objects, surfaces of objects, for example. Moore certainly toyed with the idea of taking sense-
data to be mind-independent entities with which we can nevertheless be acquainted. Bergmann
specifically endorsed such a view.

3 Let's say that P formally entails Q if it is tautology that if P then Q; P analytically entails Q
when the proposition that if P then Q can be turned into a tautology through the substitution of
synonymous expressions; P synthetically entails Q when the proposition that if P then Q is true
in all possible worlds but is neither a tautology nor an analytic truth. It should go without saying
that these are not sketches, not satisfactory analyses of these problematic philosophical concepts.

4 Ayer (1956), p. 19 presents this argument.

5 See BonJour (1985), 58-78.

6 This argument is given in Sellars (1963), 131-32 and also in BonJour (1985), Chapter 4.
VERITAS Porto Alegre v. 50 n. 4 Dezembro 2005 p. 153-172
INFINITISM’S TAKE ON JUSTIFICATION, KNOWLEDGE, CERTAINTY AND
SKEPTICISM
Peter D. Klein
Department of Philosophy – Rutgers University.

ABSTRACT – The purpose of the paper is to show how robust accounts of justification and
certainty can be developed within infinitism. First, I explain how the infinitist conception of
epistemic justification differs from both the foundationalist and coherentist conceptions. Second,
I explain how the infinitist can provide a solution to the epistemic regress problem. Third, I
explain how infinitism, per se, is compatible with both the views of those who hold 1) that
knowledge requires certainty and that such high-grade knowledge is possible as well as those
who deny either or both conjuncts in 1). In other words, infinitism neither endorses nor rejects
skepticism, taking that view to mean that we do not have knowledge in those areas commonly
thought to be within our ken.
KEY WORDS – Certainty. Coherentism. Foundationalism. Infinitism. Pyrrhonism. Epistemic
regress.

I want to thank Anne Ashbaugh, Claudio de Almeida, Alex Jackson and Jason Stanley for their
discussions and emails about this paper. In addition, and in particular, I am indebted to Jeremy
Fantl for his paper, “Modest Infinitism”, Canadian Journal of Philosophy, 33.4, December 2003,
537- 562. In that paper he proposes a way in which infinitism can account for both absolute and
relative certainty and argues that foundationalism cannot do so. Although my treatment of the
issue, relying as it does on the distinction between propositional and doxastic certainty, is
different from his, it was his paper that suggested to me that infinitism can develop a satisfactory
account of certainty.

Introduction

There are multiple purposes of this paper. First, I will explain how the infinitist conception of
epistemic justification differs from both the foundationalis and coherentist conceptions. Second,
using that conception of epistemic justification, I will explain how the infinitist can provide a
solution to the epistemic regress problem which is, as I see it, how to avoid arbitrariness in belief
without falling into dogmatism, where dogmatism is taken to be the acceptance of beliefs beyond
what is justified by the reasons for the belief. Third, I will explain how infinitism, per se, is
compatible with both the views of those who hold 1) that knowledge requires certainty and that
such high-grade knowledge is possible as well as those who deny either or both conjuncts in 1).
In other words, infinitism neither endorses nor rejects skepticism, taking that view to mean that
we do not have knowledge in those areas commonly thought to be within our ken. That’s a good
thing because both the skeptic and the epistemist (a defender of knowledge) can accept infinitism
as a proper account of justification and knowledge and, then, focus on whether the conditions of
justification and knowledge can be fulfilled. If they did not employ common definitions of those
concepts, then there would be no real disagreement between them because they would be talking
past each other as did the people in James’s famous squirrel-tree example.
1 Some Preliminary Remarks about Infinitism, Justification and Knowledge, Certainty and
Skepticism

Infinitism: I have developed and defended a form of infinitism in other papers.1 I will not be
repeating that defense here in any great detail; however, I will be arguing that infinitism can
provide a good account of some features of justification and knowledge, and that it remains
neutral with regard to some forms of skepticism. In so far as those are desiderata of a good
account of justification, this paper constitutes a further defense of a view that has been neglected,
to say the least.

A central claim made by infinitism is that there is an important kind of epistemic justification
that requires that there be available to the reasoner a set of reasons that neither ends with a so-
called “foundational” belief nor is circular. Put positively, if we are to have beliefs that are
comprehensively justified, meaning a belief that has all of the good epistemic features a belief
can have, then there must be reasons offered for it that do not beg the question. We want our
beliefs not to be held arbitrarily from our own point of view. We seek epistemic reassurance that
our beliefs are true, or at least likely to be true. Being able to give reasons for our beliefs is an
important epistemically good feature of our beliefs that makes them comprehensively justified;
and, hence, infinitism is required – or so I will argue.

Infinitism has been deemed a non-starter by virtually all philosophers who have considered it. If
they were right, it would leave some form of foundationalism or some form of coherentism as
the only viable candidates for resolving the epistemic regress problem. But foundationalists and
coherentists do not appear to me to be right in their rejection of infinitism. Their mistake is that
they do not fully appreciate the resources available to the infinitist. The reason for that will, I
hope, become clear as the paper develops.

Knowledge and Justification: Since Plato, philosophers have been seeking to correctly
characterize knowledge – that form of true belief that is most highly prized.2 I am going to
assume that there is a perfectly legitimate sense of “knowledge” in which S is said to have
knowledge that p only if S’s belief that p is sufficiently comprehensively justified. It might be
that doors that open automatically when they are approached know that they are being
approached or that dogs know their master’s voice. But the kind of knowledge I am concerned
with here is the kind that requires that we have comprehensively justified beliefs.3

Some people think dogs reason. For example, Sextus Empircus cites a case in which a dog
apparently reasons by way of disjunctive syllogism.4 In addition, I suppose, that some people
would stretch the concepts of belief and reasons so far that they would say that the devices in
automatic doors have beliefs and reasons. I won’t take issue with those anthropomorphisms here
– just as long as it is clear that ‘knowledge’ and, hence, ‘comprehensively justified beliefs’ are
paradigmatically predicated of humans whose beliefs are backed by good reasons.

Further, I am going to assume that, although knowledge entails true, justified belief, those are not
sufficient conditions for knowledge.5 In other places, I have argued that the defeasibility theory
of knowledge can provide the missing condition which, when added to true, justified belief,
provides a set of necessary and sufficient conditions. I am going to assume, here, that the
defeasibility theory is correct and that it is well enough known so that it is not necessary to
provide a full articulation of that view.6 However, where some of the details of that account will
prove useful, I will provide them. But for our purposes, we can say that some subject, S, knows
that p iff:

1. p is true.
2. S is justified in believing that p.
3. S believes that p (supported by whatever justifies the belief that p).
4. There is no genuine defeater of S’s justification of p.

I said that S is comprehensively justified (from now on I will usually just say “justified”) in
believing that p only if S’s belief is based upon good reasons, but we must be very careful not to
beg a fundamental issue at stake. I am using “reason” here in a very broad sense including such
things as when we say “the reason the pen dropped is that she opened her fingers”. For ‘reasons’
must, at least at this point in the discussion, include such mental states as experiences,
perceptions and memories as well as non-mental facts because foundationalists would take it that
in some circumstances what justifies a belief is an experience, perception, memory or nonmental
fact – rather than a further belief. That is, a foundationalist might say that my experience of
seeing (or my experience of seeming to see) a red, tomato-shaped thing (or sense-datum) is the
reason I believe that there is a tomato. Further, some foundationalists, might take the reasons to
be “external” states in the world. They would say that my reason for, or basis for, believing that
there is a chair includes the fact that there is a chair. Thus, if we were re to limit “reasons” at this
point in the discussion to a type of belief, then infinitism is being assumed at the outset.

‘Justified belief’ inherits an ambiguity from ‘belief.’ The latter refers either to a belief state
(whether occurrent or dispositional) or to the propositional content of a belief state. For example,
when we say of a belief that it is true, we are referring to the propositional content of the belief
state; and when we say that Sally has had a belief for three weeks, we are referring to the belief
state. Hence, when we say that Sally is justified in believing that p, we can mean either that her
believing that p is justified or merely that the proposition, p, is justified for her. Following Firth,
I will refer to the former as doxastic justification and the latter as propositional justification.7 If a
belief that p is doxastically justified for S, then the proposition, p, is propositionally justified
because the belief can be justified only if there is an appropriate set of reasons (in the broad
sense) that justifies the propositional content of the belief. But what is crucial to note for our
purposes is that since a proposition can be justified for a person without the person believing it, a
proposition could be justified without a belief having that propositional content being justified.
When Sally has good enough reasons for believing the proposition (and no overriding
contravening reasons for disbelieving or withholding it), the proposition could be justified for her
even though she is not doxastically justified because she doesn’t have the belief. As we say,
Sally might just not have put two and two together. In addition, Sally might have the belief that p
and p might be propositionally justified but she might believe p for the “wrong reasons.” In that
case her belief (i.e., the believing) would not be justified.

Thus, although the justification condition in knowledge refers to doxastic justification, the no-
genuine-defeater condition implicitly makes use of propositional justification. The defeater
condition means, roughly, that there is no true proposition, d, which is such that if conjoined with
the propositional content of the reasons that justify p, the resulting conjunction no longer
provides an adequate reason for p (unless d defeats only by rendering plausible a false
proposition).8 The qualification mentioned in parenthesis is designed to handle the problem of
misleading defeaters. It is an important qualification in the defeasibility theory, but it will play
no role in this paper.

Justification has been the subject of so much discussion in epistemology that although the
distinction between doxastic and propositional justification preserves a customary distinction, it
is far from clear that the way that philosophers employ ‘justification’ remains faithful in all ways
to the ordinary way of talking. That’s not necessarily a bad thing. After all, ‘acceleration’ in
physics-talk does not remain true to the ordinary notion of mere increase of velocity. Indeed, the
revised concept made possible a major advance in the study of motion by recognizing that both
change in direction and increase in velocity were correlated with the presence of a force.

Given that ‘justification’ has become a term of art, it is important to lay one’s cards on the table
at the outset. Justification, as I understand it, is one of the important components of warrant –
where warrant is whatever has to be added to true belief to convert it to knowledge.9 The other
component is the no-defeater clause. I take ‘justification’ to be a catchall term indicating a large
array of epistemic good making features of beliefs. That array includes both internal features
having to do with the reasons one has for believing as well as external features having to do with
the process that produced the belief. Beliefs produced by reliable processes, including the skills
of the believer and the environmental conditions beyond the skin of the believer, are important
ingredients of comprehensively justified beliefs. But it is not those external features that will
play a role in this paper. It is the reasons that we appeal to in order to justify a belief that will be
crucial. Note a crucial assumption I am making, namely, that justifying a belief is something that
we do. It gets to be comprehensively justified only when we justify it. More about that shortly.
Optimists among us hope that the reasons we have for our beliefs often are their causes – or more
precisely, that the beliefs that have those reasons as their propositional contents are the causes of
our justified beliefs. Perhaps. But I wouldn’t want to die in the last ditch clinging to the hope that
the causal history of our beliefs mirrors in any significant way the reasons that we offer for them.
The pessimist will hold that the scientific image of man and the philosophical image of man
diverge at just this juncture.10 For our purposes, I will not make either the optimistic or
pessimistic conjecture. When something is so clearly an empirical matter, better to wait until the
evidence is in!

Although the causal etiology of a particular belief remains mysterious, at least at this point,
whether a belief is epistemically defensible does not seem mysterious at all. To be defensible,
there must be better reasons for believing it than there are for denying it or withholding it. I take
it that a comprehensively justified belief is one that the agent can rationally defend. If asked, the
agent can give a good answer to this question: What entitles you to believe that x? If the agent
has no good answer available, then there is a clear sense in which the agent does not have a
comprehensively justified belief. A comprehensively justified belief, like earned income, is the
result of our actions. Both require our labor. The belief might be properly caused, but if we have
no reason for thinking that the belief is true, there is something important lacking. It’s like
having acquired an inheritance without any reason for thinking that we have acquired it. Of
course it’s a good thing to have acquired the inheritance, but from our point of view it is useless
since we don’t know we have the money in the bank.

There is one further comment about doxastic justification that needs to be made at this point.
Consider two people who at some specified time have identical reasons for believing that Dunnit
is the murderer. Both of them believe that Dunnit’s fingerprints are on the gun, that she had a
motive and the opportunity to kill Victim, etc. Both of them offer those reasons for their beliefs.
But let us suppose further that one of them, Mr. Subjective, can only trace his reasons back to
those propositions that confirm the belief that Dunnit did it. He can’t provide reasons for those
reasons. (Maybe he suffers amnesia, maybe he acquired them while under hypnosis, maybe he
acquired them because he dislikes Dunnit and heard someone speculating about her involvement
but doesn’t realize that’s how he acquired them, maybe he’s just forgotten the reasons, etc.). Add
to the case that Mr. Subjective has no idea that he couldn’t trace his reasoning back any further.
It never occurred to him to try. His belief is not comprehensively justified, but in spite of that,
there is a clear sense in which his belief that Dunnit did it is at least partially justified. After all,
he has very good reasons for that belief. What else is he supposed to believe? And withholding
wouldn’t be correct since all of his reasons point towards Dunnit being the murderer.11

Compare Subjective’s situation with that of Objective who acquired the beliefs in the good old
fashion way – she earned them through careful investigation and analysis of the relevant data.
She can give good reasons for believing Dunnit did it, and reasons for those reasons. I think we
can say that both Subjective and Objective are subjectively justified – that is, given what they
both have already done to justify their belief that Dunnit did it, there is a clear sense in which
each is entitled to his/her belief that Dunnit did it. But Subjective’s justification stops way too
short. There is no way available for him to continue the process of justifying his beliefs. Only
Objective has that possibility. I think we commonly refer to one belief as subjectively justified
and the other as (at least possibly) objectively justified since only Objective’s beliefs could fulfill
what is required to be comprehensively justified.

Foundationalists, coherentists and infinitists will differ on what those requirements are.
Foundationalists will require that our beliefs are anchored in foundational ones. That is, they will
require that we be able to trace our reasons back to foundational reasons if our beliefs are to be
comprehensively justified. Coherentists will require that a belief be a member of a set of
coherent beliefs; and infinitists – well, let me put that off for the moment.

Certainty and Skepticism: Again there are some ambiguities that must be addressed. We can talk
about a person being certain or a belief being certain or a proposition being certain. The former,
psychological certainty, is not the target of this paper. Doxastic certainty and propositional
certainty are the proper objects of my concern here and they are meant to be the highest degree
of justification that can be obtained by a belief or a proposition. I will have more to say about
them as the paper proceeds. Finally, a few comments about philosophical (as opposed to
mundane) skepticism. Philosophical skepticism comes in two basic forms: Academic skepticism
and Pyrrhonian skepticism.12 The former holds that we do not have knowledge in those areas
commonly thought to be within our ken. It can be a restricted form of skepticism; for example,
the claim could be that we do not know that there are other minds, or it can be a global form in
which it is held that there is no knowledge of any proposition normally thought to be within our
ken. There are arguments for such views, but I will not consider them here. I have done so in
other places.13

Here the issue will be whether infinitism leads to, or even provides any evidence for, such a
view. It might appear to do just that because it might seem to be requiring that in order to have
knowledge our beliefs must be objectively comprehensively justified and, in doing so, it might
seem to be requiring that we must have produced an infinite number of reasons for our beliefs.
But, so the objection continues, that is impossible for creatures like us, with minds that are
finitely limited in many ways. I will show that this reasoning rests upon a misunderstanding of
infinitism because infinitism does not require that we actually provide an infinite number of
reasons for the target belief to be objectively comprehensively justified.

A related ground for rejecting infinitism is that some might hold that knowledge entails certainty
and, since we can never complete the process of providing reasons, no belief can be doxastically
certain. A primary task of this paper is to argue that such a view is mistaken. Infinitism can
provide a good model of beliefs that rise to the level of certainty.

In short, my claim is that infinitism does not lead to Academic skepticism. But does it endorse
Pyrrhonism? The answer is a bit more complicated. Pyrrhonism can be characterized in at least
two ways. If it is characterized as a view that holds that it is reasonable to believe that no non-
evident proposition (a proposition requiring a reason to be accepted) is ever epistemically more
worthy than its contraries, then infinitism is not compatible with that form of Pyrrhonism since
that form is inherently self-contradictory (and nothing can be consistent with a contradiction).
But if Pyrrhonism is not taken to be a reflexively self-contradictory proposition about the power
of reasoning but rather as a characterization of a way of life – a kind of life that takes non-
evident propositions as never settled – then infinitism can be seen as a form of, or at least a close
cousin of, Pyrrhonism.14

2 The Infinitist Solution to the Regress Problem

Here is the regress problem as presented in Sextus Empiricus’s Outlines of Pyrrhonism:

The later Skeptics hand down Five Modes leading to suspension, namely these: the first based on
discrepancy, the second on the regress ad infinitum, the third on relativity, the fourth on
hypothesis, the fifth on circular reasoning. That based on discrepancy leads us to find that with
regard to the object presented there has arisen both amongst ordinary people and amongst the
philosophers an interminable conflict because of which we are unable either to choose a thing or
reject it, and so fall back on suspension. The Mode based upon regress ad infinitum is that
whereby we assert that the thing adduced as a proof of the matter proposed needs a further proof,
and this again another, and so on ad infinitum, so that the consequence is suspension [of assent],
as we possess no starting-point for our argument. The Mode based upon relativity [...] is that
whereby the object has such or such an appearance in relation to the subject judging and to the
concomitant percepts, but as to its real nature we suspend judgment. We have the Mode based
upon hypothesis when the Dogmatists, being forced to recede ad infinitum, take as their starting-
point something which they do not establish but claim to assume as granted simply and without
demonstration. The Mode of circular reasoning is the form used when the proof itself which
ought to establish the matter of inquiry requires confirmation derived from the matter; in this
case, being unable to assume either in order to establish the other, we suspend judgement about
both.15

The Regress Problem can be put as follows: Is there any account of justification with its
attendant view about the required series of reasons that can provide a model for the conditions in
which a belief is comprehensively justified? My claim here is that only infinitism can provide
such a model.

Note that there are five modes. Three of them are ways of resolving the differences of opinion
that do, or at least can, arise because of the relativity of perception and the credible discrepancies
in our initial evaluations about the truth of a proposition. What is crucial for our purposes is that
the modes of relativity and discrepancy are general recipes for cooking up ways to avoid
dogmatism about any proposition about which presumptively credible differences of opinion can
arise.

If we let b1 be any proposition about which there can be plausible differences of opinion, the
regress can be seen as beginning this way:

I believe b1, and my reason for b1 is b2, and my reason for b2 is b3, etc. Now, what, if anything,
can legitimately end the regress?

There is one obvious way to legitimately end the regress. We can grow tired of producing
reasons or we can have other pressing or not so pressing things to do. In the broad sense of
“reason” mentioned above, those are legitimate reasons for ending the giving of reasons. But
those are not good epistemic reasons. To do so requires some epistemic basis – there must be
something about the justificatory status of the last offered reason that makes it proper to end the
regress, if, indeed, there is any such proper ending point.

Foundationalism holds that there are such proper ending places. They typically will support their
view in two ways.

First, they point to examples of regress-ending beliefs and claim that it is disingenuous to deny
that there are foundational propositions. Here is what Aristotle says:

[There are some people who] [...] demand that a reason shall be given for everything, for they
seek a starting point and they seek to get this by demonstration, while it is obvious from their
actions that they have no conviction. But their mistake is what we have stated it to be: they seek
a reason for things for which no reason can be given; for the starting point of demonstration is
not demonstration.16

Indeed, in Sextus’s original formulation of the regress problem he assumes that, at least from the
point of view of the dogmatist, all good reasoning must have a starting point. Recall that this is
what he claims:
The Mode based upon regress ad infinitum is that whereby we assert that the thing adduced as a
proof of the matter proposed needs a further proof, and this again another, and so on ad
infinitum, so that the consequence is suspension [of assent], as we possess no starting-point for
our argument.

This is not an argument against infinitism. Of course, in any discussion of an issue we must
begin somewhere. We must, as Plato said, act like a mathematician and postulate something
from which other things will follow.17 Whether explicitly or implicitly we begin by stating
something that we can agree on as a starting point – otherwise we would be talking past each
other. But the issue here is whether every actual starting point (or ending point if one is thinking
about how the regress progresses) can legitimately be questioned. If we assume that there must
be a starting point for which no further reason can be given, then we have already eliminated
infinitism as a plausible account. For infinitism is just the view that 1) there is no
unchallengeable starting point that, once properly challenged, does not require a reason and 2)
reasons cannot be repeated. It is not an objection to infinitism to point out that it does not include
a provision for such a starting point.

Second, foundationalists will at least appear to argue (as opposed to presupposing) that there
must be a proper starting point if we are to have knowledge because otherwise the regress will go
on indefinitely or circle back on itself. Circling back is not a viable option because doing so will
beg the question, and if the regress continues indefinitely, we could not know anything because
“one cannot traverse an infinite series.” (Post. An. 72b10)

I have examined this “argument” for foundationalism in detail elsewhere.18 Here I merely wish to
underscore the point that it, too, presupposes a foundationalist picture of justification. Aristotle
says, “the same things cannot be simultaneously both prior and posterior to one another, so
circular demonstration is not possible” (72b25-28).19 He is imagining a situation in which the
posterior belief must derive its justification from the prior belief. Champions of circular
reasoning – at least in some forms – would not accept that assumption. Rather they would, or at
least could, claim that justification is essentially a property of the set of coherent beliefs rather
than a property of one belief that is transferred to another. When a particular belief is justified,
they would continue, it is so only because it is a member of a set of beliefs that are coherent.
That is, beliefs are justified in virtue of being a member of a set of beliefs that has the requisite
inferential structure. Justification supervenes on the set of beliefs and their internal relations, not
on the conditions that give rise to a particular belief.20 So, Aristotle’s objection fails to recognize
that coherentists do not share the foundationalists’ conception of justification in which
justification is a property that can be transferred from one belief to another.

Nevertheless, as I and others have argued, this form of coherentism is a one-step form of
foundationalism.21 The property in virtue of which all beliefs in a set of beliefs are justified is
coherence, in whatever way coherence is characterized. A belief in such a set is not justified
because it inherits its justification from other beliefs. All beliefs are foundational because they
are justified but not because they derive their justification from other beliefs. It’s like being a
sibling. A person does not acquire the property of being a sibling from a sibling. Both siblings
are a sibling because of their relationships.
So, if there is a general problem with the foundationalist answer to the regress because it
employs basic propositions in that answer, that problem will apply to this form of coherentism as
well because all propositions are basic, or at least relatively basic. They are more basic if they
enjoy the status of being more inferentially connected to other beliefs. I will mention that general
problem with foundationalism in a moment.

The infinitist’s view of justification does not employ any notion of basic beliefs, except a notion
of contextual basicality to be discussed shortly. In addition, infinitism does not think of
comprehensive justification as a property that is transferred to an as yet non-comprehensively
justified belief. A belief becomes comprehensively justified only after we have justified it. As we
produce more and more reasons for it, the belief becomes more and more justified. A reason,
though not itself yet comprehensively justified – because we haven’t done anything to justify it –
can be employed to comprehensively justify that for which it is a reason. Justifying a belief is an
activity. It is something like gaining a good reputation. We may begin with no reputation
whatsoever but through our efforts (and some good luck) we gain the reputation, say, of being
honest.22 Analogously, it is through the process of producing reasons that we contribute to
making a belief (not a proposition) justified.

Now, of course, the reason must be available to us. That is, there must be a set of propositions of
the right sort that we can call upon to justify a belief, and the infinitist holds that the set had
better be endless and non-repeating.23 But it is absolutely crucial to note that we needn’t
“traverse” the infinite set in order for a belief to be comprehensively justified. That would be
required only if comprehensive justification were an all or nothing thing. If we had to complete
the process of providing reasons in order for a belief to be somewhat comprehensively justified,
then no belief could ever be so justified. But just as gaining the reputation for being honest is
acquired step by step, the justification of a belief is increased as we provide more and more
reasons. We have never completely justified it in the sense that further justification is not
possible. But we have justified it to some extent, and in some cases, perhaps enough to rise to the
level of knowledge. The reason for the qualification, “perhaps”, is that knowledge requires more
than adequately justified belief. The justification must also be non-defeated.

Aristotle’s solution to the epistemic regress problem is foundationalism. He says that his “own
doctrine is that not all knowledge is demonstrative; on the contrary, knowledge of the immediate
premisses is independent of demonstration”. (72b18) Foundationalists hold that some beliefs
have some degree of autonomous justification – a type of justification that is not gained by
transfer from other beliefs. They will argue that doxastic justification comes in two forms. Some
beliefs are justified because they are based (properly inferred from) other beliefs. To use a
familiar Aristotelian analogy, something that is not moving by nature acquires motion from
something already moving. Just as some moving objects acquire their motion from objects
already possessing motion, some justified beliefs acquire their justification from beliefs already
justified. But, he would continue, this can’t go on indefinitely – the motion and the justification
have to originate in some first mover or first principle which does not require another mover or
another belief in order to move or be justified. In other words, there must be some beliefs that are
not justified by other beliefs but rather are self-justified, in the sense that they are justified but do
not inherit their justification – or at least some of their justification – from another belief.
One might have hoped that when the Aristotelian model of motion was discarded, the
foundationalist model of justification would also have (at least) been questioned. But the
foundationalist picture (including coherentism understood as one-step foundationalism) has
remained the dominant, if not exclusive, account of justification.

I have argued elsewhere that foundationalism cannot solve the regress because, as the
Pyrrhonians correctly saw, there is an important sense in which the “basic” proposition is
arbitrary.24 This is not the place to rehearse that argument in any detail, but it depends upon the
claim that comprehensively justified beliefs, i.e., beliefs satisfying all of the vast array of good-
making epistemic features, have been justified, at least in part, by giving reasons. If no reason for
a belief has been provided, then there is a clear sense in which it remains arbitrary. It might have
a property, P, such that any belief with P is basic, but two questions can (and sometimes do) arise
as to whether 1) we can determine whether that belief has P and 2) whether beliefs with P are,
thereby, likely to be true.

Consider one example: Suppose it is claimed that the “basic” belief is one about a first-person
psychological state to which we have “privileged access.” The question can be asked: Are those
psychological states (as opposed to other types of states) such that we are likely to be correct
when we claim that we are in them? The foundationalist has only three possible answers: “Yes,”
“No,” “I withhold.” The “yes” answer, once amplified, provides a reason for thinking the belief
is correct, and, the regress has continued. Once the question has been raised and understood, both
the “no” and the “withhold” responses would expose the arbitrariness of accepting the so-called
privileged access proposition and of basing all of one’s other beliefs on such a basis.

For that very reason, no foundationalist will give those answers. They will seek to explain why
we are at least very likely to correctly detect the psychological states to which we are
“privileged.” Thus, they are implicitly recognizing the need for continuing to provide reasons for
our beliefs beyond the so-called “foundational” reason. If we can legitimately ask what reason
we have for thinking that we can correctly detect facts other than mental facts (i.e., facts in the
“external world”), then it seems to me that we can equally well ask what reason we have for
thinking that we can correctly detect facts about our mental states. What is it about them that
makes them transparent? My claim here is not that there are no good answers to that question;
rather, my claim is that the question is a legitimate one and answering it continues the regress of
justifying our beliefs.

To say much more here about the way in which the “mode of hypothesis” leads to the mode ad
infinitum would distract us from my central tasks, namely showing how infinitism can provide
for a robust sense in which a belief can be certain and that it remains neutral about skepticism. It
is to those central tasks that I now turn. With the background sketch of infinitism given above, I
think it is relatively easy to accomplish those two, related tasks.

3 Skepticism and Certainty

As discussed above, there are two basic types of certainty that form our explanandum: doxastic
certainty and propositional certainty. Within the infinitist conception of justification, we can
define the latter as follows:
A proposition, p, is certain iff:
1. there is an infinite, non-repeating set of propositions beginning with p each member of which
is such that if it is the nth member in the series, there is an (n +1)th member that is a good reason
for it; 25 and
2. there is no true proposition, d, which is such that the conjunction, [d & the (n +1)th
proposition] fails to provide a good reason for nth proposition.26 (Call ‘d’ a defeater.)

This is not the place to give a full defense of this characterization of propositional certainty
because my main point is that there is an available infinitist characterization of certainty that is
robust and both the skeptic and epistemist can accept it. But a couple of comments are in order.
The first clause simply spells out the infinitist’s requirement for a proposition to be
comprehensively justified. Hence, if a proposition is certain, it is justified. The second clause
indicates what has to be added to justification to raise it to the highest level of epistemic merit. It
is needed because most reasons are defeasible and if a proposition is certain it must not be
subject to defeat by the truth. But the second clause also guarantees that p and every other
proposition in the set is true, for if there is a false proposition, say f, at the nth member then there
is a true proposition, ~f, which conjoined with the (n +1)th member fails to provide a good
reason for the nth member. Thus, if a proposition is certain it is true. That seems completely
appropriate since we want justification to be truth conducive and the highest degree of
justification should provide the tightest connection to truth – and a guarantee of truth would
seem to be the tightest connection. Thus, if a proposition is certain, it is fully grounded in the
truth.27

This is what one might call a person-neutral definition. A proposition is certain regardless of
whether anyone ever thinks of it – or for that matter, whether there are any persons.

Now, are any propositions certain for us? I do not know. Some evidence that there are comes
from the progress in science. It does seem that we get better and better able to answer skeptical
challenges by appealing to hitherto unconsidered propositions. Of course, there are moments in
our acquisition of knowledge when we are stymied and no further proposition is readily available
to us. As the Pyrrhonians noted, that should help us avoid dogmatism and cause us to continue
our inquiry, perhaps at a later time. It would not show, however, that the proposition is not
certain – especially if we have often found further reasons for our beliefs upon additional
reflection and new experience.

I would like to consider one possible objection because doing so will help to clarify the infinitist
account of propositional certainty. The primary worry being considered in this paper is whether,
given infinitism, any propositions are certain. But now I would like to consider the objection
that, given infinitism, there are too many propositions that are certain. There are many ways to
put this objection and I have considered some in other places.28 Here is a way I have not
discussed before.

I just showed that on this account if a proposition is certain it is true, but it might also be thought
that on this account if a proposition is true it is certain. For suppose someone were to say that for
every true proposition, there is, automatically, a series of propositions that satisfies the two
conditions for propositional certainty displayed above. Let the nth proposition, p, be true. Now, it
can be shown that simply because p is true, there is another proposition, p*, namely “p” is true,
that is a good reason for p; and there is another proposition, p**, which is a good reason for p*,
namely “p*” is true, etc.

As stated, this objection overlooks the second requirement of propositional justification, namely,
that the path of reasons must be non-question begging. However, I grant that in some cases, “p”
is true can provide a non-question begging reason for p. For example, consider this path of
reasons (n + 2, n + 1, n) where the following propositions are instantiated:

n + 2: Sally endorses p and whatever Sally endorses is true;

n + 1: “p” is true;
n: p.

But the path is question begging, where the following propositions are instantiated:

n + 2: p;
n +1: “p” is true;
n: p.

In other words, whether the step between (n + 1) and n is question begging will depend upon
what precedes (n +1).

So, is the proposed path envisioned in the objection one that begs the question or not? Is there an
implicit appeal to a lower numbered proposition in a higher numbered one? I think the answer is
that it is not entirely clear. To see that, suppose someone held a theory of truth like this:

T: “p” is true =df (“p” means p) & p.

One might have to chisholm that a bit, but I’m one who thinks that T is roughly correct. If so,
then the path would automatically repeat the proposition, p, at every stage of the regress and the
path of reasons would beg the question.

What makes the answer to the question above somewhat unclear is this: It seems dialectically
unfair to saddle the objector with my definition of truth. So, let us suppose that there is some true
proposition of the form “x” is true which is not strictly equivalent to a conjunction, one conjunct
of which is x. If that account of truth were correct, then I think it must be conceded that there
could be a non-question begging path that satisfies condition 1.

But is condition 2 in the definition of propositional certainty satisfied? Can we be sure that there
is no defeater of some step in the envisioned path? Just as the step from “Sally endorses p and
whatever Sally endorses is true” to “‘p’ is true” could be defeated, perhaps some steps in the path
generated by the truth theory could be defeated.

The answer is that we have no way of telling whether the path could be defeated until we see the
fully developed theory of truth because if a proposition of the form “x” is true is not strictly
equivalent to a conjunction one conjunct of which is x, then it is not clear that there won’t be a
defeater of the inference from “x” is true to x.

Perhaps there is some imaginative way around my reply to the objection by developing a theory
of truth that guarantees the inference is not defeated without question begging. So, for the sake of
the argument, I am willing to grant that such a theory is possible. And in so doing, I will also be
granting that it is possible that every true proposition is certain. Now, the question becomes: Is
that a bullet or a piece of cake? I think it is clearly not a bullet and that it might very well be a
piece of cake.

It surely would be a bullet if infinitism resulted in all true beliefs being doxastically certain.
Hunches that happened to be correct would be doxastically certain. True beliefs based upon false
testimony would be doxastically certain. Beliefs based upon what politicians say that just happen
to be true would be certain. And surely that last one is going too far! The reason for dwelling on
this objection is to underscore the difference between propositional and doxastic certainty. So far
we have been concerned with mere propositional certainty. Is there anything wrong with it being
the case that every true proposition has a non-question begging and undefeated path of
supporting reasons. Indeed, isn’t it somewhat re-assuring that there is such a path? All we have
to do is to discover it. Nevertheless, it is crucial to note that the path just considered – containing
increasingly complex propositions, p, p*, p**, etc – would not be helpful in making a belief
justified, much less certain, because we would rather quickly arrive at a proposition that, because
of its complexity, would be beyond our ability to grasp; and what is beyond our grasp could not
be employed to justify a belief.

Nevertheless, suppose that there were at least one path of suitable reasons such that each step
was within our ability to grasp. Wouldn’t that be a piece of cake because it makes it possible that
some of the propositions that we believe to be true are certain – and absolutely so? And, hence, if
someone (perhaps a skeptic like Peter Unger in Ignorance29 or a non-skeptic like G. E. Moore)
were to require that at least some forms of knowledge required a form of certainty, infinitism
seems (at least so far) able to provide a model in which their views could be developed. Whether
there really are such paths is a question beyond infinitism that, at least so far, appears
uncommitted to an answer one way or the other.

Now let’s turn to doxastic certainty. Under what conditions would one of S’s beliefs be
objectively certain? To begin to answer that question note a parallel with objective doxastic
justification. We said earlier that objective doxastic justification of the belief that p requires
propositional justification of p, and by extension, objective doxastic certainty would require
propositional certainty. In other words, if the belief that p is certain, then the proposition, p, is
certain. That is, there is a non-defeated path of suitable reasons for p. So far, infinitism is
compatible with a belief being certain. Further, if the belief that p is comprehensively certain for
S, S must do some justifying of the belief. Indeed, S must do “enough” justifying to make the
belief certain.

I suggest that we take it that a belief that p is objectively certain (for S) iff:

1. The proposition, p, is certain; and

2. S has justified the belief sufficiently (i.e., S has provided enough reasons along the non-
defeated path of propositions that makes p certain).30

I think it will become clear in a moment what I mean by a belief being “sufficiently” justified or
the provision of “enough” reasons. But first, recall that my central claim is that the infinitist
characterization of certainty provides a robust account of certainty that is acceptable to skeptics
and epistemists. To that end it will be useful to make a few general comments about certainty.

Peter Unger pointed out that there is a clear sense in which “certain” is an absolute term.31 If
something is certain then nothing can be more certain than it. But, as he also recognized,
“certainty” is a relative term or gradable adjective because we commonly do say that some things
are more certain than other things.32 That seemingly set of incompatible claims requires some
explanation. Unger suggested that, at base, certainty is an absolute term and offered a way of
paraphrasing the relative expressions employing only the absolute notion, namely, x is more
certain than y iff x is more nearly certain than y. The same paraphrasing would apply to terms
like “clear” and “empty” and “full.” Other explanations have been proposed. For example, it has
been suggested that when we say something is certain, full, empty or clear, we are leaving out a
parameter that is contextually supplied.33 For example, the proposition is certain (for our
purposes) even though for other purposes it is not certain, the refrigerator is full (of food) even
though there is some empty space in it, or the glass is empty (of water) even though there is air in
it. Still others might suggest that when we say that something is certain, full, etc., we are
speaking loosely. Strictly speaking the glass is not empty. When we say it is empty, what we
mean, and what is fully understood by our hearers, is that for all relevant intents and purposes
(e.g., drinking some water) the glass is empty.34

I do not want to engage in a debate about which, if any, of those proposals is correct because
there is an alternative plausible explanation of the relative and absolute sense of ‘certainty’ that
employs the already useful distinction between propositional and doxastic justification. Just as in
‘x is justified’ the ‘x’ can range over propositions or beliefs, I think it is clear that when we say
‘x is certain,’ the ‘x’ can range over either the propositional contents of belief states or over
belief states themselves. Given that, it is natural to suggest that it is propositions that are
absolutely certain (or not at all certain), and that is it beliefs that are relatively certain depending
upon the contextually determined level of justification necessary. One belief, B, is more certain
than another belief, b, because the contextually determined threshold applied to B’s being certain
is higher than the contextually determined threshold applied to b’s being certain, and both beliefs
have reached their respective thresholds. Beliefs can never rise to the level of absolute certainty
since the contextually determined threshold can be raised as high as we please even if we are
content to stop it at some point.

An infinitist would hold that, although the propositional content of a belief can be absolutely
certain for S (if there is an infinite, non-repeating, non-defeated chain of reasons for it available
to S), the belief can never be absolutely certain for S because the contextually determined
bedrock propositions can, themselves, be challenged. Once we have provided the contextually
determined basic reason (in the narrow sense) thus making the belief relatively objectively
certain (if the proposition is certain for S), the bar can always be raised by asking whether 1) we
have a good reason for thinking that we are reliable foundational proposition detectors and 2)
whether the property in virtue of which a proposition is basic is truth conducive. Or if the
contextually determined foundation is not a proposition, but rather a reason in the broad sense,
we can ask 1) what reason we have for thinking that, in general, we are reliable foundational
state detectors or 2) what reason we have for thinking that on this particular occasion we have
correctly detected the foundational state. In either case, if we reply that we have no reasons (in
either sense) for answering those questions affirmatively, we should become epistemically
queasy because we would not be able to distinguish, at least for the time being, our situation
from ones in which a context sanctions the use of the magic eight ball or consulting a holy text
for answers concerning empirical matters. The Pyrrhonian modes remain powerful tools that
save us from falling into dogmatism!

Consider the belief that there is an oak tree in the yard. Typically, that belief is at least
subjectively doxastically certain (enough) once i) we have formed beliefs about its leaf, bark,
limb and root structures, and offered those contextually determined basic beliefs for believing
that it is an oak tree, and ii) we are aware of no defeaters of that reasoning. It would not be
objectively doxastically certain were the proposition not certain; however, nothing in the theory
precludes the possibility that propositions are certain. In other words, such a belief can be more
or less objectively certain depending upon how high the reached threshold is. But, of course, we
can raise the bar and ask whether we are good detectors of leaf patterns, bark patterns, etc.

Note that this account of doxastic certainty and the defeasibility account of knowledge are
closely related. The defeasibility theory holds that if a proposition is known, then it is objectively
doxastically justified and there are no genuine defeaters. Given the proposed account of
propositional and doxastic certainty, if knowledge entails certainty, then the required level of
objective doxastic justification is certainty. I have argued that infinitism can provide a model of
certainty which is such that if knowledge requires certainty, infinitism can supply a good model
of objective doxastic certainty and that it remains an open question as to whether the necessary
conditions of propositional and objective doxastic certainty are satisfied. Are there propositions
which are certain? Perhaps. Are there beliefs which are relatively objectively certain? Perhaps.
Put succinctly, objective doxastic certainty is neither precluded nor necessitated by the robust
account of certainty that infinitism can provide.

What is not possible, given infinitism, is absolute doxastic certainty because we have “finite
minds” and, hence, we can not perform an infinite number of discrete reason givings. That is the
lesson to be learned from the Pyrrhonian discussion of the mode ad infinitum. Dogmatism results
if reasoning is taken to have settled matters once and for all.

However, it is crucial to note that absolute doxastic certainty is also not available to the typical
foundationalist with regard to non-foundational beliefs if at least some of the inferences from the
foundational beliefs to the non-foundational beliefs are not completely truth-preserving. (I say
“typical” because these comments will apply only to non-coherentist forms of foundationalism.)
If foundationalism’s account of doxastic certainty is acceptable even though the non-basic beliefs
are not absolutely certain, why would infinitism’s account of doxastic certainty not be acceptable
if none of the beliefs are absolutely certain? Indeed, the infinitist’s account of relative doxastic
certainty contains a more robust notion of certainty than that incorporated in the typical
foundationalist’s account. For the foundationalist picture of warrant transfer from foundational
propositions to non-foundational ones has the associated worry that the more inference steps
taken from the foundational proposition to the non-foundational ones, the less justified the non-
foundational proposition becomes because the inferences are not completely truth-preserving.
There is the distinct possibility that the more reasons one has for a belief, the less justified it
becomes, unless one throws in some deductive reasoning and/or grants coherence the ability to
raise the warrant of beliefs! On the other hand, the infinitist’s account of reasongiving is such
that the more reasons one has for a belief, the greater the degree of doxastic justification.

What about coherentism? Even though coherentism is a one-step form of foundationalism

because it takes all beliefs to be justified in virtue of the inferential relationships among their
propositional contents, and, hence, it takes all beliefs in the coherent set of beliefs to be prima
facie justified, it is that very fact that prevents the comments about typical foundationalism from
applying here. In coherentism, justification is not a property that arises in one belief and then is
transferred to another by inference. Hence it does not incur the liability that doxastic certainty
diminishes with the number of inferences. In fact, coherentism shares with infinitism the
possibility that doxastic justification can always be increased by adding new beliefs whose
contents cohere with the already coherent sets of believed propositions. But that similarity also
reveals that coherentism does not endorse absolute doxastic certainty anymore than does
infinitism because it is always possible to add new beliefs that increase the degree of coherence,
thus augmenting the degree of doxastic justification of all of the beliefs in the new set. So,
coherentism has to eschew absolute doxastic certainty as well, at least for beings like us, whose
belief sets are constantly being updated as we gain new experiences.

4 Conclusion

Infinitism’s solution to the regress problem employs a concept of comprehensive justification

that incorporates a robust account of both propositional and doxastic certainty. That beliefs ever
do rise to the required level of objective doxastic certainty is neither required nor precluded by
infinitism. That is a welcome consequence since it makes the debate between the skeptics and
epistemists possible. Finally, the discussion here has suggested that infinitism captures the
intuitive plausibility of the Pyrrhonian take on the power and limits of reasoning.

1 Please see the following: “Foundationalism and the Infinite Regress of Reasons”, Philosophy
and Phenomenological Research, 58.4, 1998, 919-925; “Human Knowledge and the Infinite
Regress of Reasons”, Philosophical Perspectives, 13, J. Tomberlin (ed.), 1999, 297-325; “Why
Not Infinitism?” in Epistemology: Proceedings of the Twentieth World Congress in Philosophy,
Richard CobbStevens (ed.), 2000, vol. 5, 199-208; “When Infinite Regresses Are Not Vicious”,
Philosophy and Phenomenological Research, 66.3, 2003, 718-729; “What IS Wrong with
Foundationalism is that it Cannot Solve the Epistemic Regress Problem”, Philosophy and
Phenomenological Research, 68.1, 2004, 166-171; “Infinitism Is the Solution to the Epistemic
Regress Problem”, Contemporary Debates in Epistemology, ed. Matthias Steup and Ernest Sosa,
Blackwell Publishers, 2005, 131-40; “Reply to Ginet”, Contemporary Debates in Philosophy, op.
cit., 149-52; “Human Knowledge and the Infinite Progress of Reasoning”, Philosophical Studies,
forthcoming; “How to be an Infinitist about Doxastic Justification”, Philosophical Studies,
forthcoming.
2 Plato, Meno, 97a-98d.

3 Sosa puts the point this way:

Admittedly, there is a sense in which even a supermarket door “knows” when someone
approaches, and in which a heating system “knows” when the temperature in a room rises above
a certain setting. Such is “servomechanic” knowledge. And there is an immense variety of
animal knowledge, instinctive or learned, which facilitates survival and flourishing in an
astonishingly rich diversity of modes and environments. Human knowledge is on a higher plane
of sophistication, however, precisely because of its enhanced coherence and comprehensiveness
and its capacity to satisfy self-reflective curiosity. Pure reliabilism is questionable as an adequate
epistemology for such knowledge. [Ernest Sosa, Knowledge in Perspective (Cambridge:
Cambridge University Press, 1991), p. 95.]

4 Sextus Empiricus, Outlines of Pyrrhonism, I. 69-70.

5 Gettier, E., “Is Justified True Belief Knowledge?” Analysis, 23, 1963, 121-3.

6 Please see the following for my version of the defeasibility theory: “A Proposed Definition of
Propositional Knowledge”, Journal of Philosophy, 67 (16), 1971, 471-482; “Knowledge,
Causality and Defeasibility”, Journal of Philosophy, 73 (20), 1976, 792-812; “Knowledge is
True, Non-defeated Justified Belief”, Essential Knowledge, Longman Publishers, ed. Steven
Luper, 2004, 124-135; Certainty: A Refutation of Scepticism, University of Minnesota Press,
1981, especially, pp. 137-172.

7 This distinction was first introduced by Roderick Firth in “Are Epistemic Concepts Reducible
to Ethical Concepts?” in Values and Morals (Dordrecht, Holland: D. Reidel Publishing Co.,
1978), edited by Alvin Goldman and Jaegwon Kim.

8 Note that given the broad reading of “reasons” mentioned above, we will have to take the
representational states as having propositional content and include propositions corresponding to
the worldly facts which serve as reasons. But this is, I think, common practice.

9 Alvin Plantinga defines warrant as “that, whatever precisely it is, which together with truth
makes the difference between knowledge and mere true belief” in Warrant: The Current Debate
(Oxford University Press, 1993), p. 3.

10 For a full discussion of this issue see “Philosophy and the Scientific Image of Man” in Sellars,
W., Science, Perception and Reality (New York: Humanities Press, 1963), 1-40.

11 For a more full discussion of this point see Richard Feldman, “Justification is Internal”,
Contemporary Debates in Epistemology, op. cit., p. 282.

12 See Klein, “Skepticism”, Stanford Encyclopedia of Philosophy,

http://plato.stanford.edu/entries/skepticism/
13 See Certainty, op. cit.; “Skepticism and Closure: Why the Evil Genius Argument Fails”,
Philosophical Topics, 23.1, 1995, 213-236; “The Failures of Dogmatism and a New
Pyrrhonism”, Acta Analytica, 15.24, 2000, 7-24; “Skepticism”, The Oxford Handbook of
Epistemology, Paul Moser (ed.), Oxford University Press, 2002, 336-361; “How a Pyrrhonian
Skeptic Might Respond to Academic Skepticism” in The Skeptics: Contemporary Essays,
Ashgate Press, ed. Steven Luper, 2003, 75-94; ‘There is No Good Reason to be an Academic
Skeptic”, Essential Knowledge, Longman Publishers, ed. Steven Luper, 2004, 299-309; “Closure
Matters: Skepticism and Easy Knowledge”, Philosophical Issues, 14, 2004, 165-184.

14 See “The Failures of Dogmatism and a New Form of Pyrrhonism”, ibid.

15 Sextus Empiricus, Outlines of Pyrrhonism, I, 166-169.

16 Basic Works of Aristotle, Richard McKeon, ed., 1011a1-14.

17 Meno, 86e-87b.

18 See the papers mentioned in note 1.

19 Aristotle does say that there is a “qualified sense” in which a proposition might be both prior
and posterior to itself. A proposition would be “prior for us” if we learn it first, but it would not
be “prior in an unqualified sense” – the sense in which a proposition is prior in the order of
demonstrations based upon first principles or what he calls “immediate premisses” (72b18). So
this qualified sense in which a proposition might be both prior and posterior to itself is not
relevant. We are concerned only with what Aristotle calls “prior in an unqualified sense.”

20 Laurence BonJour, The Structure of Empirical Knowledge (Cambridge: Harvard University

Press, 1985).

21 See Ernest Sosa, “The Raft and the Pyramid”, Midwest Studies in Philosophy, 5, 1980, 3-25
and my “Human Knowledge and the Infinite Regress of Reasons” and “Human Knowledge and
the Infinite Progress of Reasoning”, op. cit.

22 There is an analogous sense in which we are lucky if we have a belief available to assist us in
making a belief comprehensively justified. We might not have had such a belief. I take it that this
kind of “epistemic luck” is not the kind that precludes knowledge. The impermissible kind of
luck has to do with the accidental connection between the belief and the truth condition (if you’re
a reliabilist) and/or the justification and the truth condition (if you’re a defeasibility theorist).

23 The qualification “of the right sort” is crucial if we are to avoid a problem proposed by I. T.
Oakley in “An Argument for Skepticism Concerning Justified Beliefs”, American Philosophical
Quarterly, XIII, no. 3, (1976), 221-228, especially pp. 226-227 and by John Post, “Infinite
Regress of Justification and of Explanation”, Philosophical Studies, XXXVIII, (1980), 32-37,
especially pp. 34-35. Post’s argument, in a slightly revised form appears in Post’s book, The
Faces of Existence, (Ithaca: Cornell University Press, 1987), pp. 84-92. I reply to those
arguments in “Human Knowledge and the Infinite Regress of Reasons”, op. cit., especially 311-
312.

24 See the papers in note 1.

25 For the purposes of this paper, we can take the notion of “being a good reason” as a primitive
not needing further explication. My favorite account of it is that x is a good reason for y iff if x is
true, then it is likely that y is true. The likelihood I have in mind is objective likelihood. But
other notions of ‘being a good reason’ could also be employed. Foundationalism, coherentism
and infinitism will all have to appeal to some notion of x’s being a good reason for y, and I am
prepared to accept whatever account turns out to be the best one.

26 I am here ignoring the complication of so-called misleading defeaters. That is a difficult issue
and I have discussed it elsewhere (see the papers mentioned in note 6), but I don’t think it bears
directly on the issue at hand. Since the certain proposition, by definition, must be true (since it is
a member of the series and all members are true) there will always be a way to construct the
series so that it is not subject to defeat from the misleading defeater by adding what I’ve called a
‘restoring proposition’ to the series.

27 I am using “fully grounded” as did Michael Clark in “Knowledge and Grounds: A Comment
on Mr. Gettier’s Paper ‘Is Justified True Belief Knowledge?’” Analysis, 24, 1963, 46-48.

28 See note 23.

29 Peter Unger, Ignorance: A Case For Scepticism, (Oxford University Press, 1975).

30 There is one qualification here that would be necessary if this were to be a full account of
certainty. I have argued elsewhere that knowledge can be based upon false beliefs in an essential
way – that is, that some false beliefs can play an essential role in the justificatory reasons. See
my “Useful Falsehoods”, Epistemology: New Essays, Quentin Smith, ed., Oxford University
Press, forthcoming, 2006.) But that is such a highly controversial matter, it seemed best in this
paper not to be detoured in our main tasks by discussing useful falsehoods. The paper mentioned
above does explain what modifications to the defeasibility theory are required in order to
accommodate useful falsehoods and those modifications could be readily transferred to the
account of doxastic justification, propositional justification, doxastic certainty and prepositional
certainty developed here.

31 Unger, Ignorance, op. cit., especially p. 60 ff

32 For a philosopher who has emphasized the relative use of “certainty”, see Harry Frankfurt,
“Philosophical Certainty”, Philosophical Review. 62; 1971, 303-327.

33 Jason Stanley mentioned this possibility to me in conversations.

34 Earl Conee develops this line in “Contextualism Contested”, Contemporary Debates in

Epistemology, op. cit., p. 47-55 and “Contextualism Contested Some More, idem., 62-66.
International Phenomenological Society
What "IS" Wrong with Foundationalism Is That It Cannot Solve the Epistemic Regress Problem
Author(s): Peter D. Klein
Philosophy and Phenomenological Research Vol. LXVIII, No. 1, January 2004 Rutgers
University

There are many things that could be wrong with foundationalism. For example, some have
claimed that a so-called basic belief cannot be both 1) a reason for non-basic beliefs and 2) such
that it cannot be provided with at least prima facie justification.1 If something is a reason, they
say, then that something has to be a proposition (or sufficiently proposition-like) and if it is a
proposition (or sufficiently proposition-like), then it is the kind of thing that requires a reason in
order to be even prima facie justified.2 Another reason that some give for rejecting normative
foundationalism is that it leads directly to skepticism.3 There is no way, they claim, to move from
so-called basic propositions (typically given as first person introspective reports) to "external
world" propositions by employing normatively acceptable principles of reasoning.4 Still others
have thought that the invention of a nonreasoned reason was as ad hoc as the invention of an
unmoved mover.5

There criticisms were aimed at showing that foundationalism is false.6 Regardless of the merits
(or lack thereof) of those types of criticisms of foundationalism, my claim is that even if
foundationalism is true, it cannot solve the regress of reasons problem.7 My argument was that
once the regress of reasons begins, a self-conscious foundationalist is forced into a form of
unacceptable arbitrariness in her reasoning. I sought to capture what I took to be the point lying
behind the Pyrrhonian response to a foundationalist.8 Michael Bergmann9 replies that either (1)
my arguments are question-begging or (2) I have misunderstood what a "sensible foundationalist
will say in response to socratic interrogation by a skeptic." (WN, 165)

His central claim, I think, is that a self-conscious, sensible foundationalist need not fall into an
unacceptable form of arbitrariness. Examining that claim will be the primary focus of this short
response. But let me address the first disjunct. He says that my argument is question-begging
because I simply assume that in order for any belief, p, to be justified for someone, S, there must
be a reason for p available to S. Thus, basic propositions - propositions for which no reason is
necessary in order to be at least prima facie justified - could not be justified. I claimed that the
following principle applies to the kind of justification produced by reasoning:10

PAA: For all x, if a person, S, has a justification for x, then there is some reason, r19 available to
S for x; and there is some reason, r2, available to S for r2; etc. (HK, 299)

I would have begged the question if I had not given an argument for PAA because it does make
it a necessary condition of a person's having a justification for a proposition in the relevant sense
- the sense in which justification results from reasoning - that the person has an endless series of
reasons available for the proposition. I did give some arguments for PAA, and unless those
arguments contained the assumption that foundationalism could not solve the regress problem, I
did not beg the question. Nevertheless, Bergmann's insightful criticism points to the crucial point
in my argument for the claim that foundationalism cannot solve the regress problem, so let me
try to reformulate that argument more carefully in the context of his criticism and leave it to you,
the reader, to judge whether I begged the question against the foundationalist.

I take it that foundationalists hold, at a minimum, that some propositions are prima facie justified
in the absence of further reasons. A modest foundationalist could hold that such prima facie
justification does not rise to a level sufficient for knowledge.11 Perhaps coherence with other
propositions is also required to give the added degree of epistemic warrant required for
knowledge. Or perhaps the propositions need to be non-defeasible. My claim was that even the
practicing, self-conscious, modest foundationalist would fall into an unacceptable kind of
arbitrariness once the regress of providing reasons begins.

I envisioned a type of socratic questioning that lies behind the regress of reason-giving that
begins with the assertion of some proposition, say p, followed by the question, "Why do you
believe that p?" The regress continues: Because I believe that q. "And why do you believe that
q?' Because I believe that r. "And why do you believe that r? Etc.12 Bergmann and I agree that
the assertion/question dialogue continues until the practicing, self-conscious foundationalist
replies with what she takes to be a basic proposition - that is, a proposition, say b, that she takes
to be at least prima facie justified but not so on the basis of a reason contained in a further
proposition. It might be prima facie justified because of the causal history of the belief that has
the proposition as its content13 or it could be prima facie justified simply because of its content14,
or it could be prima facie justified simply because it is believed and there are no reasons that are
easily (or even possibly) conceivable that could count as evidence against it.15 Foundationalists
will have all sorts of accounts of what feature, say F, makes a proposition prima facie justified.

I thought (and still do) that "[w]hen faced with the skeptic's question, rationality requires that the
foundationalist attempt to show why the offered basic propositions are true or somewhat likely to
be true." (FD, p. 17) Bergmann thinks that the foundationalist, call him Fred, has a way of
replying to the skeptic, call her Sally, that avoids this requirement by explaining to the skeptic
why some beliefs - basic beliefs having some feature, F - are justified in the absence of a reason
for thinking that they are true.

Let us pick up the conversation as Bergmann envisions it at the critical point - the point at which
Fred has said that he believes X, where X is "b has some feature F such that beliefs having F are
noninferentially justified." (WN, 164) Here's how Bergmann thinks the discussion will proceed.

Sally: But that just shows that b is not basic for you.

Fred: Not true. My belief that X plays no role in the justification of my belief that b. Before I
ever grasped or considered or held beliefs like X, 1 justifiedly held beliefs like b. Those beliefs
like b were justified noninferentially in virtue of their having F, not inferentially on the basis of
beliefs like X. The same is true of my current belief b.

Sally: What reason do you have for thinking that X is true?

Fred: We can discuss that if you like. But let's be clear that you have not given me any reason for
thinking that that question is in any way relevant to the justification of b. It's my position that the
justification of b depends upon the truth of X, not on my justifiedly believing X or on my having
a reason for X. (WN, 164)

The point, according to Bergmann, is that in giving this kind of argument, the foundationalist is
not offering a reason for thinking that b is true. What the foundationalist is doing is explaining
"why it is that lacking a reason for a belief is not sufficient for a beliefs being arbitrary." (WN,
164-165)

Bergmann's foundationalist thinks that to hold b is not arbitrary because b's possession of F
makes it not arbitrary to hold it. My claim is that unless Fred also thinks that the possession of F
makes it likely or somewhat likely that b is true, Fred's reasoning that ends with b will not have
achieved its intended goal, namely providing some basis for Sally (and Fred) to believe that the
original proposition, p, ought to be accepted.

Let us grant that foundationalism is true. That is, let us grant that there are some propositions that
possess some property, F, such that we are entitled to believe them in the absence of a reason.
We are granting that there are ways in which a proposition acquires non-inferential, prima facie
warrant. Call that type of warrant, autonomous warrant.

But the question here is: Once we start the reason-giving game can we achieve the intended
purpose of providing reasons by stopping the regress with a proposition that has some
autonomous warrant?

Why does the reason-giving game begin? Presumably because Fred accepts Sally's challenge that
some belief, p, requires a reason in order to have the appropriate kind of epistemic warrant -
warrant required for Sally to be epistemically responsible in believing p to some degree higher
than she would were she not to have a reason. More simply put, the reasoning is supposed to
increase the warrant that Sally has for thinking that p is true. Fred accepts that the reason given
for p, namely q, has to have a reason in order to have the appropriate kind of epistemic warrant.
Fred, being a foundationalist, finally arrives at b and claims that b does not need a reason - it has
autonomous warrant. And perhaps it does. Fred also thinks b is true. But, this is the point where
Sally should ask whether the possession of autonomous warrant is truth-conducive. Does Fred
think that propositions that are basically-warranted are likely to be true in virtue of being
basically warranted? To use Bergmann's terminology, does the truth of X give Sally (or Fred)
any reason for thinking that b is true?

Fred has three (and only three) possible answers: Yes; No; I do not know. If he says that
autonomously warranted propositions are likely to be true in virtue of having the property, F,
then Sally can point out that he has a (very good) reason for thinking that b is true and the regress
has not stopped. If he says that basically warranted propositions are not likely to be true in virtue
of being basically warranted, then Sally should point out that from her point of view Fred has not
given her any reason to believe that p - since p was supposed to gain its needed, additional
warrant from the reasoning that Fred was furnishing. Because Fred's reasoning has not given
Sally any basis for accepting b, believing anything based upon b, including p, would be arbitrary.
If Fred withholds judgement regarding the truth-conductivity of propositions that are basically
warranted, then Sally should again point out that Fred's reasoning has not provided any basis for
believing that p is true.

Perhaps, as Bergmann would claim, there is a good sense of "arbitrary" in which believing b is
not arbitrary. After all, Fred is entitled to believe b (or "b is true") because b has F. True, but that
will not help in resolving any misgivings about p's truth unless one thinks that Fred's entitlement
to believe that b is connected to b's truth. By way of analogy: Descartes saw that even though he
was entitled to believe some proposition because it was clear and distinct, he had to show that
clarity and distinctness was truth conducive before he should accept that the proposition was
true. Descartes' worry was that even though he was entitled to believe that a proposition was true
because it was clear and distinct, perhaps that entitlement was not in any way really connected to
the truth of the proposition. In our terminology, Fred must be able to give some reason to Sally
for thinking that autonomously warranted propositions are at least somewhat likely to be true if
such propositions are to provide a basis for Sally to think that the warrant for p has been
increased through the reasoning. In other words, Descartes' worry applies even if one is a modest
foundationalist because the point of the why-game is to give Sally a basis for believing p. Ending
with what from Sally's point of view is a mere assertion of b and a claim that b is warranted in
the absence of a reason cannot provide Sally with a reason for believing p unless being so
warranted is a reason for Sally to believe that b is true.

As Descartes saw, the same is true if the dialogue is an inner, solipsistic one - a meditation.
Suppose I think I need a reason for p and I give (myself) the reason q...and finally arrive at a
basic proposition, b. Even if I were entitled to believe that p, why should I think I have provided
any warrant for p by my reasoning unless I thought that b's possession of basic-warrant was
truth-conducive? To be entitled to believe that p is true does not, by itself, imply that it is at all
likely that p is true unless that entitlement is truth conducive anymore than being entitled to wish
that p is true or to fear that p is true implies that it is at all likely that p is true. Typically, of
course, the foundationalist will have some, perhaps quite plausible, story to explain why basic
propositions are at least likely to be true. But if so, then they have a plausible reason for
believing that the basic propositions are true and could (and often will) use that story to avoid the
charge of arbitrariness.

To sum up: My point is not that foundationalism is false. What is wrong with foundationalism is
that it cannot provide the self-conscious foundationalist with a satisfactory way of addressing the
regress problem. That strikes me as a formidable problem since furnishing such a solution is
touted as a primary virtue, if not the primary virtue, of foundationalism.16 I think the only
account of warrant that will provide its self-conscious practitioner with an acceptable way of
addressing the regress problem is infinitism. But that's another story.

Thanks to Anne Ashbaugh for her help with this paper.

1 See Wilfrid Sellars, "Empiricism and the Philosophy of Mind," in Science, Perception and
Reality (London: Routledge and Kegan Paul, 1963), especially p. 128f.
2 Others, of course, have not accepted that claim. See, for example, Matthias Steup,
"Foundationalism, Sense-Experiential Content, and Sellars's Dilemma,"
http://condor.stcloudstate.edu/~steup/SellDilem.html

3 W.V.0. Quine, "Epistemology Naturalized" in Ontological Relativity and other Essays (New
York: Columbia University Press, 1969). The view is perhaps traceable to Aristotle, especially in
the Posterior Analytics in which he seems to be arguing that first principles are not arrived at
through a process of reasoning from experience but rather through the activity of some non-
inferential mental mechanisms.

4 Others, of course, have not accepted that claim. See, for example, Jonathan Vogel, "Cartesian
Skepticism and Inference to the Best Explanation," Journal of Philosophy, 87(11), 1990,658-
666.

5 Laurence Bon Jour, The Structure of Empirical Knowledge (Cambridge, Mass.: Harvard
University Press, 1985).

6 Some have held that the distinction between Foundationalism and Coherentism is bogus to
begin with. See, for example, Ernest Sosa, "Two False Dichotomies." This paper will appear in a
collection entitled Pyrrhonian Skepticism edited by Walter Sinnott-Armstrong, Oxford
University Press, 2004.

7 See my papers "Human Knowledge and the Infinite Regress of Reasons," Philosophical
Perspectives, 13, J. Tomberlin (ed.), 1999, 297-325 (HK); "The Failures of Dogmatism and a
New Pyrrhonism," Acta Analytica, 15.24, 2000, 7-24 (FD). Also see "Why Not lnfinitism?" in
Epistemology: Proceedings of the Twentieth World Congress in Philosophy, Richard Cobb-
Stevens (ed.), 2000, vol. 5, 199-208 and "When Infinite Regresses Are Not Vicious," Philosophy
and Phenomenological Research, 66, 718-729.

8 Sextus Empiricus, Outlines of Pyrrhonism, I, 166-169.

9 Michael Bergmann, "What's NOT Wrong with Foundationalism," Philosophy and

Phenomenological Research, 68, 161-165. (WN)

10 PAA, together with the following principle entails infinitism:

PAC : For all x, if a person, S, has a justification for x, then for all y, if y is in the evidential
ancestry of x for S, then x is not in the evidential ancestry of y for S.

Although I argued that the best reason for thinking that infinitism was true is the intuitive appeal
of these principles (HK, 299), I did give several reasons for thinking that PAA was true (HK,
301-306).

11 See, for example, Robert Audi, "Foundationalism, Epistemic Dependence, and Defeasibility,"
Synthese 55, 1983, 1 19-139.
12 Of course, the reasons could be, and typically are, conjunctions.

13 Alvin Goldman, "What is Justified Belief?" in Justification and Knowledge (ed.) G. Pappas
(Dordrecht: Reidel, 1979).

14 I take it the locus classicus for this view is Descartes' account of judgments comprised solely
of clear and distinct ideas.

15 James Pryor, "The Skeptic and the Dogmatist," Nous, 2000, 34(4), 517-549.

16 See, for example, Robert Audi, The Structure of Justification (New York: Cambridge
University Press, 1993), p. 127-8. Also see William Alston who presents the regress argument
and, although he does not examine it in detail, says "I do not claim that this argument is
conclusive; I believe it to be open to objection in ways 1 will not be able to go into here. But I do
feel that it gives stronger support to foundationalism than any other regress argument," Epistemic
Justification (Ithaca: Cornell University Press, 1989), p. 55.
EPISTEMIC PROBABILITY1
Richard Fumerton
The University of Iowa
Philosophical Issues, 14, Epistemology, 2004

In this paper I want to explore both the conceptual and epistemic place of epistemic probability
within a foundationalist theory of justification. After arguing that one should take the relation of
making probable holding between propositions to be one of the most important conceptual
building blocks in epistemology, I examine different ways of trying to understand this key
concept and explore the question of how our understanding of it will affect our prospects for
finding a defensible version of inferential internalism.

Regress Arguments for Foundationalism:

As I indicated, I’m primarily interested in the place of probability within the framework of
foundationalism. More specifically still, I’ll focus on the way in which our understanding of
probability is involved in the foundationalist’s understanding of inferential justification. Before
setting to that task, it might be helpful, however, to very briefly travel some familiar terrain and
remind ourselves of the ways in which foundationalists typically argue for their view about the
structure of justification. As we shall see later, the desire to avoid regress that so influences
foundationalists might bear directly on the advantages that one particular theory of probability
might have for certain versions of internalism.

The standard argument for foundationalism is the regress argument. But there are, however, two
importantly different versions of that argument, an epistemic version, and a conceptual version.

The Epistemic Regress Argument for Foundationalism:

The more familiar of the two regress arguments for foundationalism is probably the epistemic
version. Let us say that S’s belief that p is inferentially justified if S’s justification for believing P
is constituted in part by the fact that S is justified in believing some proposition other than P, a
proposition on which S at least partially bases his belief that P. The following principle seems
patently obvious to both externalists and internalists alike. For S to be justified in believing P by
basing that belief on some other proposition believed, say E1,S must be justified in believing E1.
But if all justification were inferential, then the only way for S to be justified in believing E1
would be for S to be justified in believing yet another proposition E2, and so on ad infinitum.
Finite minds cannot complete an infinitely long chain of reasoning, so if all justification were
inferential we would be forced to the absurd (and epistemically self-defeating) conclusion that
no-one is justified in believing anything at all to any extent whatsoever.2

The Conceptual Regress Argument:

The epistemic regress argument rests on a number of controversial presuppositions. As the

coherence theorist points out, the foundationalist assumes that justification is linear in character
and, as Peter Klein (1998) argues (in defending the view he calls infinitism), it also presupposes
that having justification can only be constituted by a chain of reasoning that one actually
completes, as opposed to one each link of which could be completed. If one allows that E’s
status as justifier requires only that one be capable of coming up with an appropriate argument
for P having E as a premise, then even as finite epistemic agents we may have the potential to
come up with an infinite number of different arguments, one for every link in an infinitely long
chain of reasoning.

These sorts of responses to the regress argument presuppose that the problematic nature of the
regress derives from the need to have an infinite number of different beliefs to serve as justifiers.
As I have argued elsewhere (1995), there is another way of construing the problematic nature of
the regress foundationalism seeks to avoid. Why is it a necessary truth that to be justified in
believing P on the basis of E one must be justified in believing E? The most obvious answer,
perhaps, is that one’s inferential justification for believing P on the basis of E is partially
constituted by one’s justification for believing E. It is an analytic truth that one can be
inferentially justified in believing P on the basis of E only if one is justified in believing E. But if
this is correct, then the above account of inferential justification is viciously circular if it is
intended to be an account of justification in general. Our understanding of inferential justification
presupposes an understanding of the concept of justified belief (in just the way, for example, our
understanding of instrumental goodness presupposes an understanding of intrinsic goodness).
Put another way, reflection on the concept of inferential justification suggests that any plausible
definition of inferential justification will be recursive. Our understanding of the recursive
definition of inferential justification will require an understanding of some base clause that
invokes a concept of justification that is not inferential.3 My desire to end a threatening
conceptual regress is the basis of one of my fundamental concerns with Klein’s infinitism.

To be sure, the argument is not decisive. Klein would argue that one might have some generic
understanding of justification that we could employ in our attempt to understand inferential
justification. Perhaps, for example, we could try some deontic conception of justified belief as
belief that one ought to have, given certain epistemic goals or ends.4 Employing that generic
concept one could still insist that all justification is inferential. But the virtue of a recursive
analysis of justification is that one successfully completes two tasks at once. One shows how one
ends vicious epistemic regress while analyzing the very concept of justification.

The Principle of Inferential Justification:

S’s having a justified belief in E might be a conceptually necessary condition for S’s justifiably
believing P on the basis of E, but it obviously isn’t conceptually sufficient. At the very least S
must base his belief that P on E. The analysis of the basing relation is problematic. Some insist
that it is exhausted by causal connection between beliefs; others insist that it involves something
else. In any event, it is again obvious that S’s justification for believing P on the basis of E
requires more than that S be justified in believing E and that S base his belief that P on E.5 The
most obvious candidates for the missing condition are the following:

1) There is a correct epistemic rule sanctioning the move from believing E to believing P
2) S is aware of, or has a justified belief that there is a correct epistemic rule sanctioning the
move from believing E to believing P,
3) There is an appropriate logical or probabilistic connection between the propositions E and P,
or
4) S is aware of, or has a justified belief that there is an appropriate logical or probabilistic
connection between E and P.

Now on analysis there may be no difference between the contents of 1) and 3) (and therefore
between 2) and 4)). One might argue that to claim that E makes probable P, for example, in the
sense relevant to epistemic justification, is just a way of acknowledging that there is an epistemic
rule licensing the move from believing E to believing P. Conversely, one might argue that all this
talk about the correctness of epistemic rules is itself a convoluted way of talking about
relationships between propositions. So, for example, there are rules of deductive logic that permit
certain sorts of inferences, say modus ponens, but the most obvious answer to the question of
what makes modus ponens a correct rule of inference is that there is a relation of entailment
holding between the premises and conclusions of arguments having a certain form, where this
entailment exists independently of any rule acknowledging it. Similarly, in the case of inductive
reasoning, one might argue that to say that the premises of an inductive argument make probable
its conclusion is just a way of pointing out that there is a correct rule licensing the inference to
the inductive conclusion from the premises describing observed correlations. Conversely, one
might argue that the rule itself merely acknowledges the relevant relation of making probable
that holds between the premises and conclusion of the argument whether or not there exists some
rule that takes account of that relation.

While 1) and 3) may be alternative ways of making a common claim, it is important, I think, to
decide the direction of the reduction. We need to figure out what the relevant conceptual building
blocks are in trying to understand inferential justification. I’ll return to this issue later. But,
however, one decides the issue of conceptual priority, one still needs to decide whether it is the
existence of the probability connection/correct rule or our awareness of the relevant
connection/rule that is crucial to the possession of inferential justification. It is disagreement on
this last point, I have argued elsewhere (1995), that gives us one (though only one) way of
isolating a point on which self-proclaimed internalists and externalists often disagree. One might
label the view that insists that for one to be inferentially justified in believing P on the basis of E
one must be justified in believing that E makes probable P (where entailment can be viewed as
the upper limit of making probable) or, alternatively, that the inference from E to P is sanctioned
by a correct epistemic rule, inferential internalism.

Leaving aside for a moment the correct analysis of probabilistic connection, what arguments can
we offer for inferential internalism? It seems obvious that something like the view was simply
taken for granted, explicitly or implicitly, by most of those in the history of philosophy who
either argued for skepticism or took the problem of skepticism seriously. But given the enormous
difficulties of meeting the skeptical challenge within the constraints of inferential internalism,
why accept the view? We are surely rather liberal in our allowing that all sorts of people,
including of course the philosophically unsophisticated (and the philosophically sophisticated
with false views about justification) have all sorts of perfectly rational beliefs. Given that we
seem not the least bit inclined to abandon the view that people have justified beliefs about the
external world, the future, and the past, despite the difficulty they have coming up with any sort
of reasons indicating the legitimacy of their inferences, should we not at least suspect that the
ordinary understanding of justification requires nothing as strong as what is proposed by the
inferential internalist?6 Perhaps, but there also appears to be prima facie evidence to indicate
widespread acknowledgement of the inferential internalist’s requirements for inferential
justification.

Consider the palm reader who predicts that I will have a long life based on the belief that I have
the proverbial long ‘‘life line’’ on the palm of my hand. It seems obvious that a sufficient
condition for rejecting the palm reader’s inference as rational is that the palm reader has no
reason to believe that the length of a line on one’s palm has anything to do with the probability
of one’s living to a ripe old age. If my high priest predicts that the war I am planning will go
badly based on the observation that the entrails of a recently dissected bird are bloody, an
epistemically rational person will surely demand evidence for supposing that features of entrails
are correlated with success in battle before conceding the rationality of the priest’s prediction.
These commonplace examples and indefinitely many others like them surely indicate that we do
in fact embrace the inferential internalist’s account of what is necessary for inferential
justification. We may pick and choose when we decide to make an issue of someone’s lacking
reason to believe in a legitimate evidential connection. We do sometimes take for granted the
justification of certain beliefs and the legitimacy of certain inferences when we are primarily
concerned with the justification of other beliefs and inferential connections. But once we take
seriously a skeptical challenge to a commonplace inference, we must apply the same standards
that we insist on applying to astrologers and fortune tellers.

Mike Huemer (2002) objects to the above argument for inferential internalism. He argues that
the examples I use to make initially attractive the principle are misleading in that they
inappropriately characterize the evidence from which one infers the relevant conclusion. Even
palm readers don’t think that they can legitimately infer their predictions from propositions
describing the character of a person’s palm and from that information alone. I suspect that the
priests at Delphi didn’t think of themselves as inferring truths about battles from the appearance
of entrails and from that alone. It should be a truism that much of the argument we actually give
outside of a philosophical context is highly compressed, highly enthymematic. As we ordinarily
use the term ‘‘evidence’’, we certainly do characterize litmus paper’s turning red in a solution as
evidence that the solution is acidic. The approach of very dark clouds is evidence that there will
soon be a storm. Footprints on a beach is evidence that someone walked on the beach recently.
But it is surely obvious upon reflection that one’s evidence for believing that the solution is
acidic, for example, is not the color of the litmus paper by itself. To legitimately draw the
conclusion one would need an additional premise, most likely a premise describing a correlation
between the color of litmus paper in a solution and the character of that solution.

Once one realizes that the reasoning in the examples I discussed above is enthymematic, one is
positioned to respond to that appearance of an argument for inferential internalism. For the
reasoning described above to be legitimate, it is indeed necessary to have some justification for
believing that there is a connection between palm lines and life expectancy, the approach of dark
clouds and storms, footprints on a beach and the recent presence of people; but only because
propositions describing connections or correlations of the relevant sort are implicitly recognized
as critical premises from which the relevant conclusions are drawn. As we saw earlier,
internalists and externalists alike typically share the foundationalist’s insight that inferential
justification is parasitic upon the justification we possess for believing the relevant premises of
our arguments. If the palm reader is relying on an unstated, but critical premise describing
correlations between palm lines and length of life in reaching her conclusion, she will, of course,
need justification for believing that premise (in exactly the same unproblematic sense that she
will need justification for believing the premise describing the length of the palm line itself). But
that in no way suggests that when we have fully described all of the relevant premises from
which a conclusion is drawn, we should require that the person who draws that conclusion needs
additional evidence for believing that the premises make probable the conclusion. The existence
of the relevant connection between premises and conclusion is enough.

While Huemer is right, I think, in arguing that the examples discussed above do not support
inferential internalism, it is not clear to me that one can’t make just as strong a case for
inferential internalism focusing on nonenthymematic reasoning. Consider the case of someone
who infers P from E where E logically entails P. Is the inferential internalist right in maintaining
that in order for S to believe justifiably P on the basis of E,S must be aware of the fact that (or at
least have a justified belief that) E entails P (or alternatively, that the inference in question is
legitimate)? The answer still seems to me obviously ‘‘Yes.’’ We can easily imagine someone
who is caused to believe P as a result of believing E where E does in fact entail P, but where the
entailment is far too complicated for S to understand. Unless S ‘‘sees’’ that P follows from E,
would we really allow that the inference in question generates a justified belief? Or to make my
case a bit stronger, would we allow that the person who reaches the conclusion has
philosophically relevant justification or ideal justification—the kind of justification one seeks
when one searches for philosophical assurance.7

A great deal hinges on how we understand the critical concept of someone’s believing one
proposition on the basis of another. There is some plausibility to the claim that genuine inference
involves more than mere causal connection. But the source of this intuition will, I think, provide
little solace for the inferential externalist. In fact, I suspect that we may not concede that there
has been a genuine inference unless there has been a veridical or nonveridical ‘‘perception’’ of a
connection between that from which P is inferred and P. But, of course, this ‘‘perception’’ will
be just what the inferential internalist claims is the awareness of connection that is necessary for
inferentially justified belief.

There is at least some concern that inferential internalism will lead to regress. The view does at
least remind one of Carroll’s (1895) famous dialogue between the Tortoise and Achilles.8
Paraphrasing liberally, the Tortoise admits at one point that P is true and also that if P is true then
Q is true but doesn’t see why that’s a reason to believe Q. Obligingly, Achilles plays the game
and adds an additional premise: If P is true and if it is true that if P is true then Q is true then Q is
true. Even so, wonders the Tortoise, why does that premise coupled with P and (if P then Q) give
one reason to believe Q. Isn’t the inferential internalist in the position of the Tortoise who keeps
insisting that even when one has in one’s possession evidence that entails one’s conclusion,
epistemically rational belief in the conclusion requires yet additional reason for supposing that
the entailment holds? Like the Tortoise the inferential internalist does require for justified belief
in a conclusion something other than mere justified belief in premises which do, in fact, entail
the conclusion. The inferential internalist does insist that the person possessing the justification
be aware of the entailment. But that requirement for inferential justification does not, obviously,
suggest that the argument needs additional premises in order to be good argument. But even if
we make that distinction, we should recognize that to avoid vicious regress the inferential
internalist may need to ensure that the relevant awareness of connection between premises and
conclusion does not itself require inference from still additional premises. But we’ll have more to
say about that shortly.

I have dwelt on Huemer’s argument against inferential internalism not just because I’m
interested in the question of whether or not the argument succeeds. It seems to me that Huemer’s
insightful argument reminds us of certain features of our talk of evidence that will be important
to keep in mind when we evaluate the plausibility of certain views about the nature of
probability.

The Analysis of Epistemic Probability:

Whether or not we adopt inferential internalism, we need an analysis of the probability

connection that by itself or as the object of awareness is partially constitutive of inferential
justification. Earlier we suggested two ways in which one might think of the epistemic claim that
one proposition or conjunction of propositions makes probable another. We might think that the
truth of the probability claim derives from the existence of a correct rule sanctioning the
inference in question. Or we might think that our understanding of the probability connection is
prior to and legitimizes our putting forth the rule that sanctions the relevant inference.

The rule-oriented approach itself can be thought of in quite different ways. On the most radical
view, the rules of inference might be thought of as analogous to the rules of a game, rules that
are themselves neither true nor false (though the claim that they are the rules of the game is
either true or false). Rules thought of this way are more like imperatives that prescribe, permit,
and prohibit certain inferential ‘‘moves.’’ What we obviously need from the proponent of such a
view, though, is some answer to the question of what makes a given rule ‘‘correct’’, or the right
rule to employ.9 There are many possible responses to this question. On one extreme one might
advance a thoroughgoing subjectivist/relativist position, according to which there is no possible
non-question-begging answer to the question of how to choose rules, since any answer would
itself presuppose the legitimacy of certain rules. Such a view seems to lead inevitably to a kind
of philosophical anarchy.

Chisholm and his followers are inclined to take as basic certain epistemic concepts that apply to
beliefs and would employ these concepts in turn in defining the correctness of epistemic rules,
and more generally the concept of epistemic probability.10 So on Chisholm’s view there are facts
about what it is reasonable to believe that supervene on certain properties of believers. The
noninferential justification of certain beliefs supervenes on non-epistemic features of various
conscious states (such as what I seem to remember or seem to perceive). But once we get
justified beliefs, there are additional synthetic necessary truths that tell us which other beliefs are
justified provided that we have certain justified beliefs. We can, if we like, summarize these
truths with rules that permit, require us, or prohibit us from forming certain beliefs when we have
others.

The pragmatist might suggest that the legitimacy of a rule is a function of how well we get on
employing the rule. But depending on how one understands getting along well, that view might
collapse into a view I’ll talk about shortly, a view that understands the critical concept of
probability in terms of frequency. If, for example, a rule of inference is legitimate only in so far
as it generates more true beliefs than false beliefs (when the input beliefs are true), then the
theory that appears to take rules as its conceptual cornerstone is better understood as a version of
the reliabilism that has become so popular primarily through the writings of Alvin Goldman.

Finally, one might take the legitimacy of a rule to be a function of internal relations of making
probable (more about this shortly) that hold between propositions. But if one adopts this
approach one will again be in a position to eliminate reference to rules in the analysis of
inferential justification. It will be our understanding of the relations that hold between
propositions that is key to our understanding inferential justification.

In what follows I want to focus on the last two approaches to understanding epistemic
probability. I’m not sure many are willing to embrace the philosophical anarchy that
accompanies a view in which arbitrary epistemic rules occupy center stage. And while there may
be synthetic necessary truths about what one is justified in believing when one is justified in
believing certain other propositions, it is hard for me to believe that the justificatory status of
inferentially justified beliefs is not fundamentally derived from relationships between that which
is believed. Put another way, it is surely a feature of the arguments whose premises and
conclusions are believed that is key to understanding the justificatory status of the beliefs formed
in the conclusions as a result of justified belief in the premises.

Suppose we agree that a key to understanding inferential justification is an understanding of the

relation of making probable that holds between the premises and conclusions of arguments.
What’s the best way of understanding that relation? That debate has a long history, one that
predates, but in many ways foreshadows, the now more familiar internalist/externalist
controversies in epistemology.11 Painting with a very broad stroke, one can attempt to analyze
probability claims in epistemology on the well-known model of relative frequency that is offered
as a way of interpreting claims about the probability of an individual or event having a certain
characteristic. On a very crude interpretation of the frequency theory, to say of something that it
is probably G is always elliptical for a more complex relativized claim of probability. One must
refer the individual about which the probability claim is made to some reference class, say F. The
more perspicuous statement of the probability claim is one about the probability of a’s being G
relative to its belonging to the class F. On the crudest and least plausible version of the view, the
truth conditions for the claim of relative probability are determined by the percentage of F’s that
are G. The higher the percentage of F’s that are G, the more likely it is that something is G
relative to its being F. Of course, we very often don’t explicitly supply the relevant reference
class for a probability claim. A great deal of time and energy has been spent by philosophers
trying to figure out what reference class is the appropriate default for ordinary probability claims.
Is it a class that is ontologically homogeneous, or one that is epistemically homogeneous?12 If
epistemically homogeneous, with respect to whose knowledge is the homogeneity defined? I’m
not sure that there are unambiguous answers to these questions, and in any event I’m not
concerned with these questions now. It is also fairly obvious that the relative frequencies that
constitute the truth conditions for probability claims are not actual frequencies—one will
inevitably need to turn to counterfactuals with all of the problems that that move meets. My main
concern, here, however, is the extent to which one can incorporate the alleged insight of a
relative frequency theory of probability into an analysis of the epistemic probability that we are
assuming holds between propositions.

One could borrow at least the spirit of the relative frequency interpretation of probability and
apply it to relations between propositions in the following way. We could suggest that in
claiming that P is probable relative to E we are simply asserting that E and P constitute a pair of
propositions, which pair is a member of a certain class of proposition pairs such that, when the
first member of the pair is true, usually the second is. Thus in saying that a’s being G is probable
relative to its being F and most observed F’s being G,I could be construed as claiming that this
pair of propositions is of the sort: Most observed X’s are Y and this is X/This is Y, and most
often it is the case that when the first member of such a pair is true, the second is. Similarly, if I
claim that my seeming to remember eating this morning (E) makes it likely that I did eat this
morning (P), I could be construed as asserting that the pair of propositions E/P is of the form S
seems to remember X/X, such that most often when the first member of the pair is true, the
second is.

The above view obviously resembles, at least superficially, the reliabilist’s attempt to understand
justified belief in terms of reliably produced beliefs. And it encounters many of the same
difficulties. Just as the relative frequency theory of probability must inevitably move beyond
actual frequencies in defining probability, so both the above account of epistemic probability and
the reliabilist will inevitably be forced to move beyond actual frequencies in order to define the
relevant epistemic probability/ reliability. Just as reliabilism must deal with the generality
problem, so the above approach to understanding epistemic probability as a relation between
propositions must deal with the problem of how to choose from among alternative ways of
characterizing the class of propositions pairs to which a given pair belongs. In evaluating the
reliability of beliefs produced by memory, for example, the reliabilist must decide whether or not
to lump together faint and vivid apparent memories, apparent memories of events in the distant
past and events in the recent past, apparent memories of emotions and apparent memories of
memories, vivid memories that occur in young people and vivid memories that occur in old
people. A frequency approach to understanding epistemic probability can make the same sorts of
distinctions between pairs of propositions, and consequently has the same sorts of decisions to
make. Just as many reliabilists are troubled by the implications of their view for what to say
about worlds in which demons consistently deceive epistemically ‘‘faultless’’ believers, so a
frequency theory of epistemic probability must deal with similar alleged counterintuitive
consequences about what is evidence for what in demon worlds. Lastly, both reliabilism and the
frequency theory of epistemic probability will be anathema to the inferential internalist who is
convinced that one needs access to probability connections in order to gain philosophically
satisfying inferential justification. The inferential internalist who is a foundationalist will need to
end a potential regress when it comes to gaining access to probabilistic connections. If one’s
model for foundational knowledge is something like knowledge of truths made true by facts with
which one is directly presented, there seems no hope that one will get that kind of access to
either the reliability of a belief-forming process or a probability relation (understood in terms of
frequency) holding between propositions.

One of the historically most interesting alternatives to the frequency interpretation of epistemic
probability is a view developed some eighty years ago by Keynes (1921). Keynes wanted to
model epistemic probability on entailment. He held that just as one can be directly aware of
entailment holding between two propositions, so one can also be directly aware of a relation of
making probable holding between two propositions. There are, of course, obvious differences
between entailment and making probable. From the fact that P entails Q it follows that the
conjunction of P with any other proposition entails Q. From the fact that P makes probable Q, it
doesn’t follow that P together with anything else makes probable Q. But for all that, we could
still take making probable to be an a priori internal relation holding between propositions (where
an internal relation is one that necessarily holds given the existence and nonrelational character
of its relata). P and Q being what they are it cannot fail to be the case that P makes probable Q.
(It might also be true that P, R, and Q being what they are it cannot fail to be the case that (P and
R) makes probable not-Q).

Which view of probability is correct? One might approach an answer to this question by looking
at the most uncontroversial upper limit of making probable—entailment. But it quickly becomes
apparent that entailment is a double-edged sword when it comes to serving as a paradigm for
understanding probability. The Keynesean will, of course, be right to stress the fact that
entailment is an internal relation knowable a priori. But the frequency theorist (or the reliabilist)
can equally stress that valid deduction is a paradigm of a conditionally reliable belief-producing
process (a paradigm of pairs of proposition types such that when the first member of the pair is
true, the second is as well).

Against the Keynesean, one might argue that it is patently absurd to suppose that making
probable is an internal relation holding between propositions. Such a view yields the absurd
consequence that claims about evidential connections are necessary truths knowable a priori. If
anything is obvious it is that the discovery of evidential connections is a matter for empirical
research. But while the objection might seem initially forceful, one must remember the point we
conceded in considering Huemer’s objections to inferential internalism. There is certainly no
necessary connection between the litmus paper’s turning red in a solution and the solution’s
being acidic, between dark clouds and storms, between footprints on a beach and the prior
presence of people. But then, on reflection, we decided that it is misleading to characterize the
litmus paper, dark clouds and footprints as the evidence from which we infer the respective
conclusions. What we call evidence in ordinary parlance is just a piece of the very elaborate
fabric of background information against which we draw our conclusions. So we shouldn’t
expect to find Keynesean probabilistic connections holding between, for example, the
proposition that the litmus paper turned red and the proposition that the solution is acidic. Where
should we look for a plausible example of Keynes’s relation of making probable?

The obvious, though perhaps not all that helpful, answer is that we should look for it wherever
we have what we take to be legitimate, nonenthymematic and non-deductive reasoning. One
needn’t, and probably shouldn’t, insist that even if probability connections between propositions
are knowable a priori they are easy to know a priori. On some views, all mathematical truths are
knowable a priori but as we painfully learned in math classes, their a priori character doesn’t
necessarily make the final for the course easy. Keyneseans have been given considerable grief
for the fact that they may have come up with bad examples of alleged necessary truths about
probability.13 Various formulations of the principle of indifference, for example, are notoriously
seductive but also notoriously problematic. The difficulty has always been to find the ‘‘right’’
way to characterize the continuum of alternative hypotheses whose probability can then be
‘‘divided’’ equally. If I know that something is either red or not-red, but don’t have any evidence
that bears on the thing’s specific color one might suppose that it is just as likely relative to that
evidence that it is red as that it is not-red. But a bit of reflection tells us that it is unreasonable to
treat being red and being not-red the same way. There are many more ways of being not-red than
there are ways of being red. The examples that give rise to paradox suggest, however, that there
may be no unproblematic way of finding the appropriate way to divide up the properties along a
continuum.14 If it should turn out that there is no useful principle of indifference available to the
epistemologist, it doesn’t follow, of course, that a Keynesean conception of epistemic probability
is doomed. The Keynesean should simply look elsewhere for plausible examples of propositions
standing in the relation.

The trouble, of course, is that philosophers don’t agree with each other about what constitutes
legitimate but deductively invalid reasoning. Notwithstanding difficulties posed by Goodman’s
new (now not so new) riddle of induction, one might look at the relationship between the
premises and conclusion of an enumerative inductive argument. Less plausibly, perhaps, one
might think about the connection between the proposition that I seem to remember having an
experience and the proposition that I had the experience. Still more problematically, we might
suggest that there it is some sort of synthetic necessary truth that when I seem to see some
physical object that is red and round that makes likely that there is some object that is red and
round. How plausible is it to suppose that there are necessary truths asserting that our putative
evidence in the above examples makes at least prima facie probable the conclusions?

Well how do we generally assess the plausibility of the claim that a certain proposition is
necessarily true? We often start by asking ourselves whether we can conceive of a situation in
which the proposition in question is false. And here, it seems we are in a position no more, but
no less, plausible than that critic of externalist analyses of justification who invokes demon
world scenarios in order to cast doubt on the externalist’s view. The critic of reliabilism, for
example, asks you to consider a possible world in which our sensory experiences have been
produced by a demon bent on inducing in us a massively mistaken system of beliefs. The victim
of the demon, the argument goes, surely has as much reason to believe propositions about the
external world as do we with our phenomenologically indistinguishable sense experience.
Because the reliabilist seems committed to the view that the unreliable belief-producing process
in the demon world yields unjustified belief, while the reliable belief-producing process in the
world as we take it to be yields justified beliefs, the reliabilist has an implausible view.

This objection to reliabilism is actually neutral with respect to the epistemic status of beliefs
about the external world based on perception. It states only that whatever epistemic status such
beliefs have in the world of veridical perception, they surely have that same status in the demon
world. But if one adds the premise that both our beliefs and the beliefs of the demon’s victims
are justified, then one seems to be very close to endorsing the view that one can’t conceive of
having the kind of perceptual evidence we have without that evidence making probable the truth
of what we believe, at least in the sense of ‘‘making probable’’ relevant to the possession of
epistemic justification. Since many internalists (and even some externalists) have felt the force of
the demon world objection to reliabilism, the view that making probable is an internal relation
between propositions should be at least initially attractive to many epistemologists.15
At present, I am arguing only for the very modest conclusion that the Keynesean approach to
understanding epistemic probability is a view worth considering seriously. When one
distinguishes partial ‘‘evidence’’ from genuine evidence (the body of propositions in its entirety
from which we infer conclusions), and when one keeps firmly in mind the ways in which making
probable would differ from entailment (even if it is an internal relation between propositions), it
is not that hard to take seriously the idea that one couldn’t seem to remember having done X, for
example, without that rendering probable having done X.

The view that there is an internal relation of making probable that holds between propositions is
also just what the inferential internalist desperately needs in order to avoid vicious regress. While
the classic foundationalist recognizes the need to cauterize the chain of reasoning that threatens
to extend infinitely into the past, the inferential internalist needs to fear not one but an infinite
number of infinite regresses. Just as one’s inferential justification for believing P must be traced
ultimately to something one is noninferentially justified in believing, so one must find evidential
connections one can justifiably accept without inference. If one infers P from E one must not
only be justified in believing E, but one must be justified in believing that E makes probable P.
One might be able to infer that E makes probable P from some other proposition F, but then one
must not only be justified in believing F, one must be justified in believing that F makes probable
that E makes probable P. If inferential internalism and foundationalism are true, then unless we
are to embrace a fairly radical skepticism, we must find some proposition of the form E makes
probable P that we can justifiably believe without inference. Since most foundationalists will
concede that there are at least some propositions of the form E entails P that one can know
without inference, the closer we can make our analysis of making probable resemble our analysis
of entailment, the more plausible will be the claim that we can know without inference
propositions of the form E makes probable P.16

Notes

1. I read a version of this paper at an APA symposium and profited greatly from the comments of
Jim Van Cleve and James Joyce.

2. The conclusion of an argument can be said to be epistemically self-defeating if its truth entails
that no-one could have justification for believing that it is true.

3. This is an argument Goldman made in ‘‘What is Justified Belief?’’ Both classic

foundationalists and their newer externalist/reliabilist cousins hold very similar views about the
need to embrace a foundationalist structure for justification.

4. Foley (1987) pursues a variation on this strategy.

5. As I’m using the locution, S can be justified in believing P on the basis of E even though S is
not justified in believing P on the basis of his entire body of evidence.
6. Of course, one might argue that their difficulties reflect an implicit awareness that talk of
inference in this context is highly misleading. One might also argue that one can ‘‘see’’ a
probabilistic connection without being able to articulate it.

7. I argue in ‘‘Achieving Epistemic Ascent’’ that one might accommodate at least some
externalist intuitions by allowing derivative concepts of inferential justification. Perhaps in
certain contexts we will concede the justificatory status of a belief if it was inferred from
premises that do make probable the conclusion, particularly if we think that the existence of the
probability connection is causally connected to the person’s willingness to form belief in the
conclusion. Relaxing our standards still further, I suggest, we may even allow that a belief is
justified if it is caused by a fact that is the truth-maker for a proposition that makes probable the
proposition believed.

8. Jim Van Cleve suggested to me that inferential internalism might give rise to the regress.

9. There is an exactly analogous question that arises for rule utilitarians. One can perhaps define
morally justified and morally unjustified action by reference to rules that require, permit, and
prohibit certain sorts of actions, but we are in desperate need of a criterion for choosing between
alternative rules.

10. So, for example, in Chisholm (1977),the primitive ‘‘more reasonable to believe than’’ is used
to define ‘‘beyond reasonable doubt,’’ ‘‘has some presumption in its favor,’’ ‘‘is acceptable,’’
‘‘is certain,’’ and ‘‘is evident.’’

11. One of the most interesting debates that has clear implications for the internalism/externalism
controversy can be found in Keynes (1921) and Russell (1948) Part V.

12. Let us say that a reference class F is ontologically homogeneous with respect to some
characteristic G when there is no way of dividing the class further such that the frequency with
which things are G relative to membership in the subclass is different from the frequency with
which things are G relative to F. A reference class F is epistemically homogeneous with respect
to G when as far as we know the class is ontologically homogeneous.

13. Gillies (2000) seems to rely on counterexamples to putative probability connections to attack
the Keynesean view.

14. Consider a well-known example. You know that I drove the mile between point A and point
B traveling somewhere between 30 mph and 60 mph and thus taking somewhere between 1
minute and 2 minutes to make the trip. So what’s the probability that I was going between 30 and
45 mph? It’s surely just as likely as that I was going between 45 and 60 mph—the probability
must be .5. And what’s the probability that it took me somewhere between a minute and a minute
and a half. Also the same as that I took between a minute and a half and two minutes—.5. But it
turns out that you can’t assign .5 probability to both the hypothesis that I was traveling between
30 and 45 mph and the hypothesis that it took me between a minute and a minute and a half.
15. Goldman himself was clearly troubled by the problem in his early paper on reliabilism
(1979). Concern with the general problem led him first to ‘‘normal worlds’’ reliabilism
(1986),and ultimately to bifurcate the concept of justification (1988).

16. Notice that the Keynesean approach to understanding probability does not require that for
one to be inferentially justified in believing P on the basis of E one must know or be able to
formulate general principles of probability. One might be able to see the connection between
particular propositions without seeing how to generalize. An analogous point holds of
entailment. One can see that P entails Q without being able to see that entailment as an instance
of modus ponens, modus tollens, or any other general kind of entailment.
CLOSURE MATTERS: ACADEMIC SKEPTICISM AND EASY KNOWLEDGE
Peter Klein
Rutgers University

…2. Availability of Reasons

What does it mean for a reason to be “available” to S? Consider a case: Suppose S believes that
silver is more dense than copper. In order for that belief to be knowledge, we have said that there
must be good reasons available to S for that belief. And the question is this: How readily
available must they be? Must they be armchair available, such that on mere careful reflection,
ceteris paribus, S will produce them? Put another way, must a proposition be entailed or
otherwise implied by the content of S’s current beliefs in order for that proposition to be
available? In the case under consideration here, that requirement would be fulfilled if, for
example, S believed that a cubic centimeter of silver weighs approximately 10.5 grams and that a
cubic centimeter of copper weighs about 8.3 grams. But requiring armchair availability might
strike some as too strict.

One could take a more lenient view and hold that a proposition, x, is available to S just in case
there is an epistemically credible way of S’s coming to believe that x given S’s current epistemic
practices.8 In the case under consideration, one could hold that if S’s epistemic practices are such
that she would consult a reliable source of information about relative densities and if that reliable
source were to list silver as more dense than copper, then one could deem that there are adequate
reasons available for S to believe that silver is more dense than copper. This might seem too
lenient until it is remembered that the mere availability of reasons is not sufficient for
knowledge. The reasons must be deployed by S.

It is not easy to determine the appropriate standards for “availability of reasons.” But luckily we
need not solve that problem here so long as we employ the same standards in both the antecedent
and consequent of CJ.

I argued earlier that closure does not dictate an epistemic priority among reasons. On one
occasion, x could be available as a reason for y, and on another occasion, y could be available as
a reason for x. What is ruled out is that there is an occasion on which each is available as a
reason for the other. Otherwise, the rules would permit begging the question.

What differentiates one occasion from another? My answer is neo-Wittgensteinian. Being able to
provide reasons requires that some propositions are occasion-relative up-for-grabs (the ones
requiring the reasons) and others are occasion-relative bedrock propositions (the ones that serve
as reasons and are not up-for-grabs). Of course the occasion can change and something that was
bedrock relative to one occasion can be up-for-grabs relative to another occasion. If, for
example, on one occasion I am asked what reason I have for believing that Socrates is mortal, I
could answer that all humans are mortal and Socrates is a human. On another occasion, if I were
asked why I believe that all humans are mortal, I could give Socrates is human and mortal
among my reasons.
A crucial question to ask in a discussion of global academic skepticism is this: Is there some
occasion on which everything is up-for-grabs? I think the answer is clearly “no,” if what would
be required on such an occasion is that no reasons are available to provide a good argument for
the propositions that are up-for-grabs. For if everything were up-for-grabs, nothing could be
available as a reason because nothing could qualify as an occasion-relative bedrock proposition.
On the other hand, if what is meant is that once a reason is given, it can be challenged, and the
reason then offered can be challenged, etc., then that seems perfectly possible.9 Put another way,
the question “Is there a reason for everything I believe?” is ambiguous. If it means “Is there some
reason or other available for each thing I believe?”, the answer could be “yes.” If it means “Is
there one reason for all the things I believe?”, the answer is “no.”

…I have argued against foundationalism on other occasions.17 And I have already suggested a
reason in this paper for rejecting foundationalism—namely that foundationalism holds,
illegitimately, that there is a fixed epistemic priority among propositions which on every
occasion must be mirrored in what leads to a belief being doxastically justified…

8. Further, it seems plausible to suggest that S might even develop new concepts when seeking
reasons. But a discussion of that would take us too far afield. I have discussed that elsewhere,
most recently in ‘‘Human Knowledge and the Infinite Regress of Reasons,’’ Philosophical
Perspectives, 13, J. Tomberlin (ed.), 1999, 297–325 and ‘‘Human Knowledge and the Infinite
Progress of Reasoning’’ (forthcoming).

9. See endnote 8 for a list of some articles in which I have defended such a view. In addition it is
defended in ‘‘When Infinite Regresses Are Not Vicious,’’ Philosophy and Phenomenological
Research, 66.3, 2003, 718–729,‘‘What IS Wrong with Foundationalism is that it Cannot Solve
the Epistemic Regress Problem,’’ Philosophy and Phenomenological Research, 68.1, 2004, 166–
171 and in ‘‘Infinitism is the Solution to the Regress Problem’’ and ‘‘Response to Ginet’’
Contemporary Debates in Philosophy, (Blackwell), ed. Matthias Steup, forthcoming.

17. See endnotes 8 and 9.

The Justification of Deductive Inference and the Rationality of Believing for a Reason
Gian-Andri Töndury

1.5 Specific issues

There are the four following main obstacles for a proponent of Inferential Internalism:

Classical Regress: First, the requirement imposed by Inferential Internalism seems prone to the
classical regress problem exploited in traditional skeptical arguments. If the inference is only
justified, if something else is justified (inferential internalism), then, if this something else is
inferentially justified, as well as everything that is required by any further inference
(inferentialism about the reflective acceptances), then the initial inference is only justified, if
infinitely many things are justified. There are a number of responses to this regress problem: (i)
One might reject inferentialism about the reflective acceptances and claim that this acceptances
are the result of and justified by non-inferential a priori insight [Bonjour, 1998]; but this is less
plausible for induction than for deduction. (ii) One might accept the conclusion and adopt
infinitism, the view that (inferential) justification obtains only once a certain infinite number of
acceptances is acquired (in the right way) [Klein, 1999]; but this faces the problem to have to
account for the fact that there is a psychological dependence, i.e. the basing relation, between an
infinity of states.

Justificatory Lewis-Carroll Problem: Second, there is a problem concerning the justificatory

import of the acceptance of the proposition of the form [premises] support [conclusion]
(henceforth “reflective acceptance”). (This is one aspect of the Lewis-Carroll-Problem) How
does the (granted to be justified) reflective acceptance bear on the justificatory status of the
belief in the conclusion? This must be by way of inferential support, for the propositional content
of the reflective acceptance is not identical to the content of the conclusion. But if the support is
inferential, then according to inferential internalism a further justified acceptance is required, and
again a regress ensues [Boghossian, 2001, 2003]. This problem needs a careful evaluation,
because it probably applies only to some variants of Inferential Internalism. Inferential
Internalism need not countenance the claim that the reflective acceptance partially constitutes (or
explains) inferential support. But even if it does apply, it has been suggested that the acceptance
need not be propositional and hence need not support the inference inferentially [Wright, 2001,
Bonjour, 2005].

Psychological Lewis Carroll Problem: Third, there is also a problem concerning the
psychological import of the reflective acceptance. (This is another aspect of the Lewis-Carroll-
Problem, [Cf. Stroud, 1979]) Even if the reflective acceptance is itself not based on any other
belief, one may ask how it partially explains the belief in the conclusion. If it is inferentially, in
the way a belief in a premise explains a belief in a conclusion, then a regress looms. For a further
reflective acceptance would have to explain this further inferential connection, and so on. A
defense of Inferential Internalism must be able to secure for the reflective acceptance a
distinctive (not premise-like) psychological role.

Alternative strategies: Fourth, even if none of the aforementioned problems proofs to be

insurmountable, one may ask whether Inferential Internalism is necessary in order to take the
intuitions on responsibility into account. There are proposals to capture such intuitions in other
ways. For instance, it has been argued that —even lacking the view on her reasons demanded by
Inferential Internalism— a subject is inferring in a responsible manner when behaving in
accordance with the inference-pattern is constitutive of the subject’s possessing some of the
concepts involved in the premises[Boghossian, 2003]. Or, it may be claimed that if the
inferential dispositions qualify as epistemic virtues, then it can thereby be said that the reasoner
behaves in a responsible manner [Greco, 1999]. Or, one may claim that in some cases it is
practically rational (and thereby not irresponsible) to indulge in an inference in the absence of a
reflective acceptance [Enoch and Schechter, 2008]. Inferential Internalism must compare
favourably with these views.

…Very briefly, the trouble with thoroughgoing exuberance is that it is very implausible, at least
at the face of it, that the further propositions are all non-defeasibly justified. And if they are
defeasibly justified, then there are further propositions which must be justifiably believed. And
they will also be defeasibly justified, hence there are further propositions to be justifiably
believed, and so on. There is an imminent danger of a vicious regress. If exuberance could be
limited to L, it is plausible that the regress can be avoided. If the justification for L were
deductively inferential, then the danger of a regress would arise. But it is plausible that a belief in
L can be non-inferentially justified by logical intuition. Hence, there is no further L-proposition
to be justifiably believed. Unfortunately I cannot go into more detail here.

I have much less ideas about how a minimalist could respond to Greco’s point. However Greco’s
passage does not provide a worked out argument against minimalism, it merely suggests that it is
false. Perhaps the minimalist can provide a full description of the difference between the two
cases.

Or perhaps, she could argue that something that is plausible for the case of a deduction in the
sense of a manipulation of schemata does not apply to deduction in general. She could hold that
if I deduce schemata or mathematical propositions, it is true that I have to know the entailment
relations in order to be justified. But in the case of a simple modus ponens such as Wet Roads
that is not required.

But minimalists generally think that they do not need to go into arguments about their opponents
positive points. For they think that exuberant views are fatally flawed. The following passage by
Paul Boghossian [2001] gives an idea of the problem they raise against exuberant views:

[. . . ] at some point it must be possible to use a rule in reasoning in order to arrive at a justified
conclusion, without this use needing to be supported by some knowledge about the rule that one
is relying on. It must be possible simply to move between thoughts in a way that generates
justified belief, without this movement being grounded in the thinker’s justified belief about the
rule used in the reasoning.

Preceding this passage Boghossian actually gives an argument in order to support what is here
claimed. He thus does not simply attempt to elicit an intuition. Nevertheless I think one can
intuitively grasp the worry concerning exuberant views from the passage. The worry can be
expressed simply thus: What is the subject to do with her insight into L? It can not be that for
every inference, one must use ones knowledge of L inferentially. For if one uses one’s
knowledge of L inferentially, one uses L as a premise in an inference. But then one must use
further knowledge for this inference, just as the knowledge of L is needed with respect to Wet
Roads. And so on ad infinitum. So, at some point, as Boghossian says, one must stop using
further knowledge and just draw the inference. This is all very loose and rushed, but there is at
least an initial suspicion that there is some such fatal problem for exuberance.

…The argument against strong inferential internalism

As said, according to strong inferential internalism, the positive epistemic status of L explains
why the subject has a reason to believe the conclusion on the basis of the premises. But what is it
to “explain why one is justified”? It’s not an explanation of some belief-forming process. It’s an
explanation of something normative, namely that the subject has a reason to infer.31

I take all the following sentences to express the same or a similar question: “In virtue of what
does one have justification?", “What are the justification-conferring facts?”, “Why does one have
justification (on a given occasion)?”. Accounts of epistemic justification provide general answers
to these questions. Such accounts have certain ambitions. For instance, reliabilism wants to give
a reductive account of justification in naturalistic (non-normative) terms. Other naturalistic
accounts take the route of non-factualism about normativity.32 Ambitious forms of internalism
want the account of justification to be such that it is knowable via reflection and a priori
reasoning whether we have justification or not.

But these general questions about justification are not at issue here. Strong inferential
internalism, although perhaps committed to certain views on these issues, attempts to explain
much less. Consider again the example given above about the object I perceive in the wood.
Given only the information that I see the object, it is in a very straightforward and
uncontroversial sense mysterious why this contributes to my having a reason to believe that X
was in the wood. And one reveals part of the mystery by giving the additional information that I
have a reason to believe that the object I see belongs to X.33

As said, according to strong inferential internalism it is the fact that my justification for the
proposition X was in this wood rests on my justification for the proposition The object I see
belongs to X that carries the explanatory burden. For the argument that follows, it is important to
insist on this double aspect of strong inferential internalism. This view does not merely claim
that the justification for (3) rests on the justification for L, but also that this explains why one has
inferential justification for L in N. The two theses are the following:

(A) The justification of (3) in N rests on the justification for L. (dependenceclaim)

(B) That in N one has justification for (3) can be partly explained by appeal to A. (explanatory
claim)

According to one reading of Carroll’s dialogue, it shows that one cannot explain why one has
justification for a given fundamental kind of deductive inference by appeal to the justification for
the logical belief, because such an explanation is circular. Consider the case of Wet Roads. There
is an important disanalogy between an appeal, in the course of an explanation, to the fact that the
subject has justification for (2), the conditional, and an appeal to the fact that the subject has
justification for the proposition that (1) and (2) entail (3), i.e. L. When one wonders why a
subject having justification for (1), i.e. It is raining outside, in a given circumstance has
justification for (3), i.e. The roads are wet, the information that in the circumstance the subject
has justification for (2), i.e. the conditional, reveals part of the mystery why she has justification
for (3).

It is difficult to say why this explanation is successful. One aspect of it is that, given the
additional information about (2), the set34 of justified propositions that confer justification upon
another, i.e. the set of the premises, bears a necessary relation to the proposition upon which
justification is conferred, i.e. the conclusion. That is, the truth of L seems to make such an
explanation possible. Without (2), the contents (1) and (3) are unrelated. With (2) premises and
conclusion are related in the manner expressed by L. So, one aspect of the explanation is that it
reveals a necessary connection among the (or some)35 propositions involved in the inference.
Furthermore there is something general about the connection between premises and conclusion.
The propositions are necessarily connected due to logical form. (1), (2), (3) have the following
logical form (“p” etc. are proposition-schemata): p, If p, then q, q. So, the explanation allows to
see the connection between the three propositions as an instance of a general connection among
propositions with a certain logical form.

Another aspect of the explanation seems to be that the additional information allows to see the
subject’s transition to conform to a general principle of reasoning. Again, it is in virtue of the
logical form of the three propositions that the transition conforms to the principle. The principle
can be stated thus:

MPP-reasoning: Believe q, if you believe p and if p, then q! (Or else give up p or If p, then q!)

Of course the two mentioned aspects of the explanation, i.e. on the one hand that there is a
necessary connection between all or some of the involved propositions, on the other hand that
there is a general principle linking the beliefs in the premises with the belief in the conclusion,
are connected. But there is no simple correspondence. It is not sufficient for a transition to
conform to a principle of reasoning that the propositions stand in a necessary connection. As has
been noted many times, there are propositions that stand in a necessary connection to each other,
even in virtue of logical form, although there is no justified direct inference from the one to the
other. That is, there is no principle of reasoning linking the one directly, i.e. without any further
propositions, to the other. Take any example of a proposition entailing another where even the
best logician needs to construct a proof in several steps. The direct step from the first to the
second is not justified and there is no general principle of reasoning linking the beliefs in them.
Furthermore, it is generally agreed that there are principles of non-deductive reasoning that link
beliefs in propositions that do not stand in a necessary relation to each other. So there is no
simple correspondence between the two aspects of the successful explanation. I will content
myself with pointing to these two aspects without claiming that one or the other is more
fundamental in understanding why the explanation is successful. For the argument against strong
inferential internalism it is not necessary to endorse any further claim. For, according to that
argument, both aspects are missing in the strong inferential internalist’s alleged further
explanation by appeal to L.
The strong inferential internalist’s opponent claims that no similar advance in understanding can
be achieved by an appeal to the justification for L. It is not possible to subsume Wet Roads under
a more general, or in some other way explanatorily more fundamental principle in virtue of the
subject’s further belief in L. And it is not possible to establish any tighter connection between the
set {(1), (2), L} and (3) than the relation that obtained between the set {(1), (2)} and (3). The
relation in both cases is entailment. Surely there is no sense in which including L makes the
connection between the set of the premises and the conclusion tighter.36

Nothing here speaks against claiming that a belief in L is involved. But on the supposition that
MPP-reasoning is one of our fundamental forms of correct reasoning, the belief can at best be
involved as part of reasoning of the form corresponding to that very same principle or a no more
fundamental other principle. For instance, someone could correctly infer from L the proposition
If (1) and (2), then (3) (by another principle of reasoning)37. From If (1) and (2), then (3) and (1)
and (2), she could then infer (3). But obviously this last step is just another instance of MPP-
reasoning. So, this alleged explanation is circular.38

And even if it is false that MPP-reasoning describes one of our fundamental principles of
reasoning, the same Carrollian argument would apply to whatever is a fundamental principle. As
Boghossian says, at some point, we have finished explaining by appeal to further beliefs about
the relation between the contents. At some point the principles by which reasoning is evaluated
are reached. And while one can still try to explain why the principle plays a role in determining
what a subject has an epistemic reason to believe, one cannot do so by appeal to further beliefs
of the subject and more general principles. I have given the argument in a form that directly
challenges the strong inferential internalist’s explanatory claim (p. 39). In this form the argument
refutes the idea that the information that the justification for (3) rests on the justification for L
can explain why (3) has positive epistemic status in N. The moral to be drawn is that the strong
inferential internalist’s explanation is circular. But often when the Carrollian argument is
presented or mentioned the accent is put on a regress. This is of course in line with Lewis
Carroll’s original dialogue which, it is suggested, is never ending. If the accent is put on the
regress, one is led to think that the problem for the attacked view, here strong inferential
internalism, is that it commits the proponent of the view to the claim that a subject who
justifiably infers must have an infinity of justified beliefs. It is possible to illustrate the fact that
an explanation is circular by giving the basically same explanation over and over again.39 But
one should not conclude from this that the problem for strong inferential internalism lies in the
fact it is committed to the claim that a subject who justifiably deductively infers must have an
infinity of justified beliefs or justification for an infinity of propositions, i.e. the propositions
corresponding to L on infinitely many further levels.

It is in the aim of keeping the problem of circular explanation and the problem of requiring
justification for infinitely many propositions apart that I have distinguished on page 39 between
the strong inferential internalist’s explanatory claim and her dependence claim. Her motivation
for the dependence claim is the explanatory claim. Suppose there is another motivation for the
dependence-claim. The justification for (3) rests on the justification for L. Suppose further for the
sake of illustration that it also rests on the justification for the infinitely many propositions of the
form of L but taking as antecedent always the set of premises one level below (the consequent is
always (3)). For instance, at the level next to one where one appeals to L the proposition is the
following: (1), (2) and L entail (3). Given these suppositions, a subject only has justification for
(3), when it has justification for an infinity of further propositions. This is problematic, but not as
obviously problematic as a circular explanation. It is not generally impossible to have
justification for infinitely many propositions. After all, when one has justification for a general
proposition about an infinite population (e.g. numbers), then one has justification for an infinity
of propositions. Now, the issue here is not only having an infinity of reasons, but furthermore
having an infinite chain of reasons each resting on the preceding one. When I have one (or a
finite number) of reasons justifying a general proposition about an infinite population, I have an
infinity of reasons. But it is the one (or the finite number of) reason(s) for the general proposition
on which all the others rest. However, it is not wholly implausible to suggest that one and the
same capacity provides me in a sort of “inductive way” (in the mathematical sense of
“inductive”, not the epistemological) with the infinity of reasons for the infinity of propositions
and with an infinity of reasons for the general proposition. The reasons for the individual
propositions do hence not rest on the reason(s) for the general proposition. If there is a one way
dependence, then rather in the other way: the reason for the general proposition depends on the
infinity of reasons for the individual propositions. In the same vain, it is not wholly implausible
that whatever capacity allows the subject to have a reason for L, also provides a reason for any
similar proposition “higher up”. Of course, I cannot vindicate here the claim that reasons can be
arranged in such a way. Nevertheless, I think one can conclude that the regress of reasons is not
obviously vicious.40 But even if such a regress of justifications is not vicious, this does not help
the strong inferential internalist. For his explanation of the inferential justification of (3) is still
circular. So, it is not so much the dependence claim as the explanatory claim which in my eyes
seriously undermines strong inferential internalism.

My rendering of the argument explicitates the traditional diagnosis of the Lewis-Carroll-

Problem: L should not play the role of a further premise. It seems at first, that strong inferential
internalism does not make this mistake. For, he attributes to L not the same, but a more important
or fundamental role than that of the premises. That the subject has a reason for L is supposed to
explain why the reasons for the premises provide a reason for (3). However, it turns out, that the
only way the justification for L could help explain why (3) is justified is in the way (2) (or in the
example about the object in the wood the proposition The object I see belongs to X) helps to
explain it, and this is as a premise. So, the traditional diagnosis seems right to me; but sometimes
people think that it is the Tortoise’ avoidable mistake to think of L’s role as that of a premise.
This line is not available to the strong inferential internalist, for as a premise seems to be the
only way L could play the role strong inferential internalism reserves for it.

The Lewis-Carroll-Problem as developed here poses a serious threat to strong inferential

internalism. I have not provided a full discussion of the problem. There might be some further
moves and objections a strong inferential internalist would want to consider.41 But the view I
want to defend is the psychological view. Therefore it is more important to determine whether a
version of the Lewis-Carroll-Problem poses a similarly strong threat to the psychological view.

…An infinite regress of beliefs

But let us return to the (psychological) question whether the linking-proposition is always
believed for a reason. This question hints at the possibility of a regress of beliefs as follows:

(I) The linking-proposition is always believed for a reason.

(II) Believing for a reason always implies believing some linking-proposition.
(III) If one believes for a reason, then one has an infinity of beliefs and one single belief depends
on an infinity of further beliefs.

We can now apply the discussion of foundations to the case of the linking-belief. In fact, we find
the proponent of the belief-proper account in a similarly uneasy position. For it might initially
seem that if the linking-belief, at least at some point, were a foundation, then the regress could be
halted. However, this move merely brings again the problems with foundations to the fore.

It was said that it is hard to see how foundations could be believed for a reason where this
involves a linking-proposition. But even if this idea could be made plausible, such foundations of
course would not help avoiding the regress. If the linking-belief were such a foundation believed
for a reason (thus assuming (I)), it would still imply that there is a further linking-belief (thus
implying (II)).

However, if the linking-belief is a foundation believed for a reason where this does not involve a
linking-proposition, we could deny (II). But then, as said, a new kind of believing for a reason
must be introduced. Furthermore this puts at risk the motivation for introducing the linking-belief
requirement in the first place. In general the belief-proper account seems to be ill-motivated if
believing for a reason is in some cases not in virtue of a linking-belief. This would mean one has
to accept (II). The linking-belief cannot be a foundation believed for a reason in a way not
involving a further linking-belief.

Again if the linking-belief is a foundation not believed for a reason we face the slumping
problem. In general, if one does not accept (I), then one seems to face a variant of the slumping-
problem. For instance, if the linking-belief is not even a reaction to a reason, then one must ask:
How worthy can it be to believe for a reason as opposed to responding to a reason, if believing
for a reason involves a belief that is not even a response to a reason? And if the linking-belief is a
reaction to a reason but merely a response, a similar question can be asked: How worthy can it be
to believe for a reason as opposed to responding to a reason, if believing for a reason involves a
belief that is no more than a response to a reason?

It is not too implausible, at least not for the case of deduction, that the linking-belief is a
foundation. Of course, it can be excluded that linking-beliefs are perceptual foundations. They do
not have the sort of content that can be directly perceptually supported. The question is whether
they could be some other sort of foundation. In particular, on some theories of justification a
priori intuitions can support belief. A priori intuitions are (defeasible) direct, i.e. non-inferential,
insights into certain truths. There are two ways in which a proponent of the belief-proper account
could incorporate a priori intuitions in order for the linking-proposition to play the role of a
foundation.44 She could take them to be seemings, i.e. perhaps non-propositional experiential
(non-sensory) states which (perhaps partially) provide a reason to believe the linking proposition
—a bit like (according to some at least) non-propositional perceptual states provide a reason for
perceptual belief. Or, she could take them to be a kind of propositional attitude and claim that the
linking-proposition must not be believed properly, but “intuited” (thus rendering the name
“belief-proper account of deduction” somewhat inappropriate).45

But the problem is that even if the linking-belief is a foundation, this must be in either one of the
three ways alluded to above:(i) A foundation believed for a reason where this involves a further
linking-belief. (ii) A foundation believed for a reason where this does not involve a linking-
belief. (iii) A foundation not believed for a reason. The first does not avoid the regress. The
second undermines the belief-proper account. The third faces the slumping problem. So
foundations do not seem to bring the solution to the regress.

But if the linking-beliefs are always believed for an inferential reason, then we have in fact not
only an infinity of linking-beliefs, but also an infinity of beliefs providing the reasons for the
infinitely many linking-beliefs. For each linking-proposition is believed for a reason, where this
reason is provided by a further belief and believing for this reason therefore involves a further
linking-proposition.

In summary, a proponent of the belief-proper account faces a trilemma.

(i) Either the beliefs in the linking propositions are not foundations, then they are believed for an
inferential reason, and then there is a regress of linking beliefs as well as of linked (supporting)
beliefs;

(ii) or they are (at some point) foundations and foundations are not believed for a reason, and
then we face the slumping-problem;

(iii) or they are foundations and foundations are believed for a reason. But it is hard to explain
how foundations could be believed for a reason, in a way that takes the tracing-back intuition on
board. And if this means that there must be a linking proposition, then there is an infinite regress
of linking-propositions.

It is to be noted that this trilemma does not depend on adopting psychological foundationalism.
For the first horn is the problem as it arises for a psychological coherentist.

It is not my aim to present this dilemma as a fatal argument against the belief-proper account. It
is here rather meant as an indication that there is no obviously unproblematic path to accounting
for the manner in which the linking-belief is psychologically embedded with the subject’s other
states. Perhaps one or more of the horns of the trilemma turns out on scrutiny not to be so
worrisome.

For instance, it is not clear that horn (ii) is so problematic. For the slumping problem mentioned
before in the discussion of foundations is not exactly the slumping problem that arises in
connection with the linking-belief. We can call the former problem “the slumping problem for
foundations”, the latter “the slumping problem for the linking-belief”. As said, a given
“foundation” in the psychological sense I use this term can be justified or unjustified. But
foundations can in principle be justified. And for a foundationalist about good epistemic reasons,
some foundations are justified. Furthermore he holds that the justification of all non-foundational
beliefs rests on the justification of the foundations. Thus, for him the slumping problem arises as
follows. The kind of justified belief one has, when one believes for a good epistemic reason, is
better than the kind of justified belief one has, when one responds to a good reason. So, suppose
a given justified foundation is not believed for a good reason (as none are on the view under
consideration). If now someone believes for a reason partially provided by that justified
foundation, the justification obtained by her belief rests on an instance of the lower-kind of
justification obtained by the foundation. That the belief is held46 for a reason ensures that the
subject traces back her commitments. But, if foundations are not believed for a reason, she traces
back these commitments to beliefs that are crucial for her justification and there stops to trace
further back. This seems not to be a much more responsible behavior than when one does not
trace back one’s commitments from the start. This is the slumping problem that arises for this
way of dealing with foundations.

The slumping problem for the linking-belief is not exactly the same. For suppose the linking-
belief is such a foundation not believed for a reason. Suppose it is justified in virtue of being a
response to a good reason. Does in this case the justification obtained by the the belief in the
conclusion rest on the lower kind of justification obtained by the linking-belief? No, for the good
reason for which the conclusion is believed is not provided by the justified linking-belief at all.
This is different from the case where the foundation provides the reason for which one believes.
In fact, it is not even clear that the linking-belief must be justified in the first place. This is the
question to which I now turn.

The justificatory status of the linking-belief

The second questioned asked at the beginning of this section was whether the linking-belief must
be believed for a good reason. Or we can now ask, whether it must be believed for a good reason,
or at least “be a response to” a good reason.

We have already seen a way in which the psychological view imposes constraints on N that can
be qualified as “normative”. If believing for a deductive reason involves having some
propositional attitude, e.g. as according to the psychological view, an attitude towards L, then the
subject must not have conflicting attitudes, e.g., as seen, disbelieving L. When this happens a
(normative) rational requirement is violated. This was claimed in chapter 2.4. The questioned
asked here is whether the attitude towards L postulated by the psychological view must in
addition to not violating a rational requirement be supported by reasons.47

Believing (3) for the reason provided by (1) and (2) implies that the subject believes a linking
proposition L. Suppose that the belief in L is completely unjustified, for instance because the
subject has strong evidence against L. It sounds false to say that in that circumstance the belief in
(3) is justified. Hence, the linking-belief must be justified.

However, it is not clear that this describes a coherent situation. For evidence against L will defeat
the otherwise good epistemic reason. Thus, the subject does not believe for a good epistemic
reason. What we must imagine is rather a situation in which the linking-belief L is unjustified,
but not because there is evidence against it, but because either there is no reason in favor of it or
because it is not believed as a response to such a reason. Only in this circumstance could the
reason for (3) provided by (1) and (2) be good, i.e. remain undefeated.

It seems that in such a situation the belief in (3) is unjustified, in spite of the fact that there is a
good reason for it and in spite of the fact that it is believed for that reason. This means that
believing for a reason has or lacks an evaluative property which is relevant for the epistemic
evaluation of the belief held for that reason, e.g. the belief in (3). As argued in chapter 2.4, it can
have the evaluative property of violating a rational requirement. Now it can be claimed in
addition that the linking-belief must be justified. And this even though the reason for believing
(3) does not depend on the reason for believing L. And even though the reason for L does not
even seem to be required for the subject to be in the psychological state of believing (3) for the
reason provided by (1) and (2). Even if the linking-belief is unjustified, according to the belief-
proper account if the belief in (3) depends on it in the way specified in (B), then it is believed for
the reason provided by (1) and (2).

This either means that it is believed for a good epistemic reason, or that it is held in mere
response to a good epistemic reason. Again, the proponent of the belief-proper account faces a
dilemma: Either the linking-belief (at some point) is a mere response to a good reason and this
risks to slump the status of the belief held for a reason (believed for a reason). Or it is believed
for a reason throughout and there must then be an infinity of good epistemic reasons. I will
allude to the first horn as the “slumping problem with the linking-belief”, to the second as the
“linking-belief regress-objection”.

However, even if

L must have positive status: i.e. being justified in believing (3) for the reason provided by (1) and
(2) implies being justified in believing L (for a reason or in response to a reason) (implication-
claim),

the justification for (3) does not rest on the justification for L. Here is again the strong inferential
internalist’s dependence-claim from page 39:

J(3) rests on J(L): If S didn’t have her justification for L, then S wouldn’t have her justification
for (3), but not vice versa.

The belief-proper account does not imply the dependence claim. This means that the belief
proper account leaves it open that the justification for L rests on the justification for (3). Because
the reason for (3) does not depend on the reason for L, it is possible that the reason for (3)
supports the reason for L. Perhaps there does not happen to be a reason provided partially by the
deduced belief in (3) supporting the belief in L, but the implication-claim leaves it open. Such a
justification of L would not beg the question, for the justification for (3) does not rest on the
justification for L.

That the belief-proper account does not imply the dependence-claim has important consequences
for the present problem. With respect to the linking-belief regress-objection the following can be
said: Even if there is an infinite number of justified linking-beliefs, this does not entail that there
is an infinite number of reasons. The reasons for the linking-beliefs or the linked premises do not
rest on one another. Thus reasons used earlier in the chain of reasons can reappear at a later
stage. The implication-claim together with the claim that each linking-belief is believed for a
good epistemic reason does not entail a structure of reasons such that each reason obtains only if
an infinity of further reasons obtains. Rather there could be reasons (themselves not depending
on any further reason) supporting the structure of infinitely many beliefs at various points.

In general, the implication-claim obtains more as a matter of internal coherence than as a matter
of there being something bad about the reason for believing (3) when the linking-belief is
unjustified. One could say that when the linking-belief is unjustified, then the way the subject
believes for a (let us suppose) good epistemic reason produces some “collateral damage”. The
implication-claim thus points to a collateral requirement on believing for a good epistemic
reason.

This collateral status of the implication-claim not only shows the alleged regress in a more
benign light, I think that it also supports the view that the slumping problem with the linking-
belief is much less of a problem than it might appear. For that the linking-belief is a mere
response to a reason is not obviously as bad as when the belief in (1) or (2) that provides the
reason for believing (3) is a mere response to a reason. The justification for L in
N is a justification for a belief with a very specific function. Namely, the function of establishing
for which reason the subject believes (3). Whether the reason is good or bad does not depend on
the justification for that belief (although if it’s negation is justified, then the reason is defeated).
It seems to me plausible to claim that when the justification for a belief with this function is of a
lower kind, this does not lower the justification for the belief for which a reason is provided.

I think of the two horns of the dilemma, I prefer the slumping problem with the linking-belief.
Intuitively the idea is this —expressed in terms reminiscent of the quotation from Boghossian
[2001] in chapter 2.1:

In believing for a reason one has to start with taking some things to be reasons. One has to take
certain linking-contents for granted. One cannot indefinitely take things for epistemic reasons
only because one takes some other things to be reasons for thinking that the first things are
reasons. At some point one has to start with taking things for reasons.

Against taking the regress route speaks here simply the fact that I would have an infinity of
linking-beliefs. (As suggested, the infinity of beliefs could perhaps be supported by a finite
number of reasons.) It is true that passages such as these illicitly suggest that there would have to
be ever a temporally antecedent linking-belief and this might be partially responsible for the
impression that there can not be infinitely many (for that would take an infinite amount of time).
The belief-proper account does not imply that the linking-beliefs are temporally antecedent. Still
the mere fact that the subject must believe an infinity of contents, where there does not appear to
be a clear way how the beliefs are extended from a finite set of beliefs —as it is the case with
beliefs about numbers— does speak against there being an infinite number of linking-beliefs.

So, if we take the slumping-horn, then one somehow “starts” with certain linking-beliefs without
believing them for a reason. This does not mean that it is completely arbitrary which linking-
beliefs one has. One is by certain factors such as one’s psychological make-up, one’s upbringing
in a certain community etc. lead to taking certain things for reasons. Furthermore, these factors
happen to shape one in such a manner that one starts with taking certain things for reasons in
response to the actual presence of a reason (for believing the relevant linking-proposition). The
fact that one does not believe the linking-proposition for these reasons does not lower one’s
justification.

…Why is it reasonable to be confident in L in the absence of a reason for believing L? In

response we can draw upon an idea enunciated in Wright [2004b] and in Dretske [2000a]. The
regress-problem itself provides a reason for merely being confident! Believing each defeater to
be false for a reason would involve a new set of defeaters corresponding to this new reason. And
if these new defeaters must be believed to be false for a reason, then there are new defeaters
again, and so on. Thus, if you want, the best way to construct an epistemic agent that can believe
for defeasible reasons is to give her confidence in the falsehood of what she takes to be defeaters.
An agent who would try to settle that each defeater is false would never achieve being in the
state of believing for a reason. Respective to the aim of believing for a good epistemic reason,
being merely confident in some propositions is instrumentally more valuable than trying to
believe everything for a reason.

… But both ideas are false. First, if reflection can take antecedently justified beliefs as premises,
these premises are the content on which one reflects. And such premises can have all sorts of
contents, for instance typical empirically justified contents such as Induction has proved reliable.
So, reflection is not only on mental properties.13

… The corresponding argument (A2) with H instead of H’, is structurally different:

(C1) The justification for H must be partially deductively inferential.

(C2) The subject’s justification for believing Hrests on her justification for believing a
corresponding proposition H1 for the reflective inference.
(C3) The justification for H1 must be partially deductively inferential, hence it rests on a
justification for H2, and so on.
(C4) There is no infinite chain of (epistemically) antecedent justification, hence the subject does
not have justification for H.

It is to be noted that (C1) can be weakened in a certain way without much consequence. Strong
inferential internalism is here taken as a thesis about deduction. But it seems clear that a
proponent of explanatory inferential internalism about deduction would hold a similar view
about any kind of inference. (In fact, in section 4.2 I explained inferential internalism as a view
on inference in general.) This means that even if the inference to H is e.g. an inference to the best
explanation, a explanatory inferential internalism would hold that its justification rests on the
justification for a corresponding H1 linking the abductive premise to the abductive conclusion.
This would lead to the same regress, no matter what kind of inferences are further involved.

Since (C2) is the explanatory inferential internalist’s claim, the argument can come under attack
by (C1) or (C3) and by (C4). One can either claim that the justification for H or one of its
“successors” H1, H2 etc. is non-inferential or that there is an infinite chain of justifiers. It seems
to me that the case for (C4) is prima-facie weaker than the case for (D3). It is less controversial
that the involved kind of circularity in (D3) is bad, than that there can not be an infinite chain of
justifiers as claimed in (C4). However, it is clear that infinitism is initially problematic. In
particular, it appears to me to be intitially more problematic than not adopting explanatory
inferential internalism, i.e. (C2).32

So the weakest links seems again to be (C1) or (C3). Even in the remote past people have
thought that on a particular instance of inference, it is possible to non-inferentially justifiably
come to rationally intuit that in this particular instance one has a reason.33 In the recent past, this
response has been considered and endorsed precisely in response to a version of the foregoing
argument (C1)-(C4).

Boghossian [2003] develops an argument against (A2) as follows: The justification for H cannot
rest on an inferential justification for some other proposition. Thus, if it rests on some other
proposition this proposition must be non-inferentially intuited. For instance, one could perhaps
non-inferentially intuit L, i.e. (1) and (2) entail (3). However, L can support H only in a further
deduction, namely from L and something like If my premises entail the conclusion, then they
support my conclusion. Thus, according to (A2) this deduction requires it’s own justification for
a corresponding H1, and so on.34 35

But Bonjour [2005, 100] who can be interpreted as endorsing (A2) recently says the following:36

For a variety of reasons, but most fundamentally because of the role that such [a priori] insights
are supposed to play in deductive inference, it is often and quite possibly always a mistake to
construe them as propositional in form. The problem here is essentially the one pointed out long
ago by Lewis Carroll: at least in the most fundamental sorts of cases (think of modus ponens),
the application of a propositional insight concerning the cogency of such an inference would
require either a further inference of the very sort in question or one equally fundamental, thereby
leading to a vicious regress. Instead, I suggest, the relevant logical insight must be construed as
non-propositional in character, as a direct grasping of the way in which the conclusion is related
to the premises and validly flows from them.37

We can interpret this passage as proposing a way to deny (C1): the reflective justification for H
is non-inferential. Bonjour thinks that if the reflective reason for H is provided by a non-
propositional state, then it will not be an inferential reason. This seems very plausible. For it is
plausible that non-propositional states do not enter into computational processes. They are not
combined with and compared to propositional states, rather they give directly rise to
propositional states.

Let us first sum up what I call the “sceptical argument” against explanatory inferential
internalism, and then give an overview of the options with which the argument leaves a
proponent of this view.

• First, a certain claim is ascribed to explanatory inferential internalism: the justification of a

deductive conclusion rests on the reflective justification for the normative linking-proposition H.
Then, it is argued that from this assumption —absurdly— scepticism follows.
• The sceptical conclusion is reached via the claim that the reflective justification for H is
inferential. In that case a regress of justification for normative linking-propositions H1, H2 etc.
obtains.

• If the reflective justification for H rests on the inferential justification of a proposition about
inference in general, then the justification is not merely regressive, but even viciously circular.

If this argument is sound, then there appear to be only three options for rebutting it. A proponent
of explanatory inferential internalism could claim some of the following:

• That he is not committed to the claim that the justification for a deductive conclusion rests on
the reflective justification for H.

But in that case he is hard pressed to give his explanation of the deductive justification by appeal
to the reflective justification for H. Strong inferential internalism is the view that the fact that (1)
and (2) provide a reason to believe (3) in N can be explained by appeal to the fact that the subject
has a reflective reason to believe H. How else could such an explanation go than by claiming that
H antecedently supports believing (3) on the basis of believing (1) and (2).38 Or, he could claim:

• That the reflective justification for H is non-inferential.

Finally, his claim could be:

• That infinitism is true, and that the inferential justification for H does not rest on an inferential
justification of a proposition about inference in general.

I think the case against against explanatory inferential internalism can be strengthened. In section
4.5 I present an argument that focuses directly on the explanatory claim, thus avoiding the need
to argue that explanatory inferential internalism implies the resting-on claim (A2). Furthermore
the argument does not depend so much on the claim that the reflective justification for H is
inferential.

4.4 Tortoise’ argument against explanatory internalism

Here is again the defining thesis of explanatory inferential internalism:

Explanatory inferential internalism: That (1) and (2) provide a reason for believing (3) is
partly explained by the fact that there is a reflective reason to believe H.

As mentioned this closely resembles the thesis of strong inferential internalism. The argument to
be presented is a generalization of the argument against strong inferential internalism form
chapter 2.3. This was then called the “Lewis-Carroll-Problem for strong inferential internalism”
in virtue of the fact that the argument can be read into a dialogue presented by Lewis Carroll
[1895]. I will here present the argument in close connection to the original dialogue.
Roughly, the argument is that any way in which we could appeal to the subject’s epistemic
position towards H in an explanation of why she has a reason to believe (3) provided by (1) and
(2), would rely on our having antecedently an explanation of why deductive reasons are good
epistemic reasons.

The original dialogue

In order to briefly summarize the note by Carroll, let us assume that it is about Wet Roads instead
of the inference involving propositions about (simple) geometry given by him. Let us further call
the following conditional constructed out of the propositions involved in Wet Roads “the
logically true hypothetical”:

C If (1) and (2), then (3).

This proposition is logically true, although this is not part of its content. In the latter respect it is
to be distinguished from the proposition L which does not play a role in Carroll’s dialogue.

The note by Carroll features a dialogue between Achilles and the Tortoise in the course of which
Achilles is lured into a regress. At the crucial stage of the dialogue the Tortoise asks whether
someone might not accept the premises of Wet Roads and yet not be “logically forced” to draw
the conclusion. In their diagnosis of this possibility, they agree that the subject must therefore
accept the logically true hypothetical, in order to be forced to draw the conclusion. However as
the Tortoise is not forced into accepting the conclusion at the start, accepting the logically true
hypothetical no more “logicallly forces” it to accept the conclusion. The initial premises entail
the conclusion just as well as the new set of premises including the logically true hypothetical.
At this point the dialogue continues potentially infinitely as ever new and more complex
hypotheticals are added to the initial premises in the hope of reaching a point where the “logical
force” is irresistible.

The dialogue exploits the fact that the two sets of premises bear the same logical relation to the
conclusion. Here is again the representation of Wet Roads compared to its “enhanced version”
(to stress that it is of the MPP form also represented as a two-premise-inference):

(1) It is raining outside.

(2) If it rains outside, the roads are wet.
(3) The roads are wet.

(1*) It is raining outside and If it rains outside, the roads are wet.
(2*) If It is raining outside and If it rains outside, the roads are wet, then The roads are wet.
(3*) The roads are wet.

It is not immediately clear what lesson Carroll intends one to learn from the dialogue. How
should one understand the expression “logically forced” (to the conclusion)? Depending on the
interpretation, the dialogue raises somewhat different general issues. On a first reading, the
expression means the same as “committed” or “epistemically required”. The dialogue could then
be taken as providing an argument against the view that one is epistemically required to take the
MPP-step, because one accepts —and only insofar as one does— the logical principle, i.e.
because one believes in the hypothetical which is a logical truth. Against such a view the
dialogue purportedly shows that the commitment to the conclusion of the MPP is not in any
sense stronger given the hypothetical is added as a further premise. The conclusion is still no
more (!) than entailed by the premises. If one is committed to believing what is entailed (or at
least obviously entailed, see below) by what one believes, then the initial MPP and the inference
from the set of premises including the hypothetical are on a par.39

On another plausible reading Carroll means by “logically forced” something like

“psychologically forced” in the sense that a subject can not psychologically avoid believing the
conclusion. On this reading Achilles and the Tortoise agree that it is psychologically possible for
subjects to believe the premises but not the entailed conclusion (while considering,
understanding and so on, all three propositions). They then conclude that what explains why the
subject believes in the conclusion when she does are not merely the beliefs in the premises but a
further belief in the logically true hypothetical, and they are thus lead into trouble. It seems that
if subject’s are psychologically constituted in such a manner that believing the premises of an
MPPnever alone and in itself explains why she believes what follows from it, then she is so
constituted that believing the premises and in addition the logically true hypothetical will never
explain why she believes what follows from the premises neither.40

But in fact, the two readings are not wholly independent, for (i) epistemic commitments do not
simply correspond to logical relations between propositions, but depend on some features of the
(kind of) subjects which are under the commitment, and (ii) the psychological explanation in
question (supposedly not applying to the recalcitrant subject) is a special sort of psychological
explanation, namely one in terms of motivation by reasons, which is relevant to the question of
normative commitment.41 On (i): No one is held responsible for not believing, even when
presented with the proposition, something which is in a very non-obvious manner entailed by
what one believes. Epistemic commitments, even those about deductive inference, do not simply
correspond to logical relations.42 On (ii): If “logically forced” is not concerned with motivation,
it is not concerned with “real psychology”. The problem is not simply to give any sort of
psychological explanation (somehow involving the beliefs in the premises) of the belief in the
conclusion. The point cannot be that belief in the premises and the logically true hypothetical can
no more explain the belief in the conclusion than simply the beliefs in the premises. For as far as
just any psychological processes are concerned, nothing speaks against the view that there is a
psychological process from the first set of beliefs to the conclusion but not from the second set
(perhaps neither speaks anything in favor of that view). Rather when the concern is motivation,
then the psychology under consideration is ideal. Such a subject fully considers, understands and
believes the premises and considers one in the light of the other(s), is not mislead by disturbing
features, can deploy enough resources to the task etc. The possibility Achilles and the Tortoise
envision is that such a subject would not be moved to the conclusion. The issue, on this reading,
is not whether the subject is committed to the belief. That may be granted. What they do not
grant is that the subject would be forced to move to the conclusion, could not help but to be
moved to it —except some mistake in processing, understanding and so on. They then proceed to
exclude the possibility to be unmoved by ascribing further beliefs to the subject.
The argument against explanatory inferential internalism is an adaptation of the first reading. It is
to be noted that Achilles and Tortoise are not involved in the activity of justifying (3). Rather, on
both readings they are trying to explain something. This is the reason why I take the argument to
be provided to be closer to the dialogue by Carroll, than the skeptical argument from section 4.3.
The sceptical argument would better correspond to a regressive dialogue in which someone is
trying to justify his belief in (3). But in the Carrollian dialogue Achilles is trying to give an
explanation from a third-person point of view of why the subject should or why she does believe
(3).

… This means that we reach the same conclusion as we reached by the skeptical argument of
section 4.3. The sceptical argument however, left the proponent of explanatory inferential
internalism with two options that are no longer available if the Carrollian argument of this
section is sound. These options were: (i) Adopt infinitism, and (ii) deny that explanatory
inferential internalism implies that the justification for (3) rests on the justification for H. The
first option is no longer available, for we needn’t assume that infinitism is false for the purpose
of the argument. The Lewis-Carroll-Problem exposed here is not that each inferential step
implies and antecedently justified other inferential step. Infinitism in this sense would be
compatible with it being explainable (i.e. it not being a primitive normative fact) why this
infinite chain of inferential dependence provides a reason to believe (3).46 The argument here is
that the explanation by appeal to inferentially justified H would be a circular explanation of
inferential justification. The second option, (ii) is not available to the explanatory inferential
internalist either. We are presently considering an explanation of why an epistemic reason for (3)
obtains that appeals to there being an epistemic reason for H, and that the latter reason is
provided by propositional contents (resp. the justified beliefs in them). Perhaps this explanation
does not imply the claim that the justification for (3) rests on the justification for H. But it would
have to appeal to some kind of support-relation between propositions, perhaps some sort of
holistic support by coherence among all believed contents. But even so, it is hard to see what
kind of more fundamental support-relation could play a role in an explanationof the principle of
MPP-reasoning. These considerations seem to leave a proponent of explanatory inferential
internalism with the only remaining option: the reason for H is provided by a non-propositional
seeming.

… As said, the reason why it is reasonable for the subject to be merely confident in H in the first-
order belief-formation is that she would otherwise be engaged in a regress. This consideration
does not speak in favor of the truth of H, therefore it does not provide or constitute evidence (or
an “evidential reason”) in favor of H.

… Foundations were introduced in order to stop a regress (or avoid circularity) of subjective
truth-connections. All this requires is that the foundations are not subjectively truth-connected. It
does not imply that the foundations’ epistemic responsibility does not depend on the standing of
any other acts at all. Foundations can depend on the standing of acts that do not provide a
subjective truth-connection. And it is clear from the vindication that the piece of practical
reasoning does not establish that the employment of IBE is truth-conducive.

However, the proponent of (PF)’s fear of regress and circularity, which motivates his
endorsement of foundationalism, also motivates a requirement to the effect that the foundations
are in some wider sense fundamental —in other words, that there is a sense in which the
foundation’s responsibility does not depend on further acts. It will be argued that if a
foundational act is pragmatically defensible as required by the Defensibility Strategy, then it
cannot be fundamental in the relevant sense.

Consider a given candidate for being a foundational act, e.g. an instance of IBE. A proponent of
the Defensibility Strategy is committed to the claim that there is a relevant difference between

(a) requiring that this act is subjectively truth-connected and

(b) requiring that it is pragmatically defensible.

He is committed to claiming that while the former gives raise to a a vicious regress or malign
circularity, the latter does not. But pace that, it will be argued, there is no reason to expect that
there is a relevant difference.

Consider an instance of IBE, called now for simplicity’s sake an “IBE-act”. Suppose that it is
subjectively truth-connected. According to how this notion has been introduced, this implies that
the subject “imminently” accepts that the IBE-act is truth-preserving. Let us say that the the
subject therefore has an available truth-connection defense. So, it is possible to compare the
available truth-connection defense with an available pragmatic defense. The pragmatic defense
might be one along the lines of the vindication given by Enoch and Schechter. The proponent of
the Defensibility Strategy denies that the subject has an available truth-connection defense. But
his reason for denying this should not be a reason also to deny that she has an available
pragmatic defense.

One possible worry concerning a further subjective truth-connection is the following: We are
about to give an account of epistemic responsibility, that is, we are to explain why a given
cognitive act is epistemically responsible. If the account satisfies the relevant explanatory
standards, then for some cognitive acts the epistemic responsibility must be explanatorily
fundamental. This means it must be fundamental in the sense that it’s responsibility is not
explained in terms of the responsibility of some other of the subject’s act. Now we are
accounting for the responsibility of a given IBE-act. But if we appeal to an available truth-
connection defense, we appeal to the responsibility of a potential, imminent cognitive act. But it
is not the case that that further act (or, if any, the further acts it relies on) are explanatorily more
fundamental. On the contrary these acts are explanatorily less or equally fundamental than an
IBE-act. Perhaps some of them are IBE-acts themselves. Therefore our account risks to fall prey
to an explanatory regress explaining something in terms of what is to be explained. At the very
least the appeal to an available truth-connection defense does not allow us to make any
explanatory progress. Everything remains to be explained.

But if this is a legitimate worry (the objector can remain neutral on this), then the Defensibility
Strategy commits one to the following: the potential act of the pragmatic defense and, if any, the
acts it itself relies on are explanatorily more fundamental than the foundational cognitive acts.
It cannot here be given a precise account of what it means to be explanatorily more fundamental
(with respect to epistemic responsibility). A plausible requirement is that if an act’s epistemic
responsibility is explanatorily more fundamental than another’s, then one can understand why it
is responsible without first understanding why the less fundamental act is responsible. This
requirement is obviously violated when the supposedly more fundamental act is of the same kind
than the less fundamental act. Perhaps one could argue that a defense of IBE along a subjective
truth-connection would necessarily be circular relying at some point at least on IBE itself, while
the pragmatic defense does not. That the weighing consideration above does not rely on IBE is
not that clear, since we do not know how the premises could be supported. At least the argument
form does not seem to be an IBE. But be that as it may, it is very difficult to contend that a
similar pragmatic defense for the basic deductive method would not rely on that method itself.
(The parallel pragmatic defense of MP —assuming that this is one of our basic deductive
methods— would be a similar weighing consideration. The only difference between the
vindication for IBE and MP is that they are the best options with respect to different rationally
required project.) It is very plausible that practical reasoning exploits the same connections
between contents that deductive reasoning does. For instance, means-end reasoning can be
brought in MP-form.

And even if all the support relations involved in the pragmatic defense were of a different kind,
the question whether the reliance on these methods is responsible does not seem to have an easier
answer than the original question about IBE, modus ponens, perception, and so on. I can see no
reason why the acts and support relations involved in the pragmatic defense would not give rise
to an explanatory regress as well.

This obviously does not definitively refute the claim that there is any relevant difference. But it
raises a challenge for a proponent of this line of thought. She must provide good reasons to think
that the pragmatic defensibility is less subject to an explanatory regress than a truth-connection
defensibility. Perhaps there are such reasons, but they can only lie in a detailed account of
practical reasoning. For it seems that the intuitive view of the interrelation between theoretical
and practical reasoning does either not give priority to either one, or, if it gives a priority, then it
lies with theoretical reasoning. Practical reasoning is not, it seems, more explanatorily
fundamental or self-standing, than theoretical reasoning.

Consider now another worry about a further subjective truth-connection. This would involve, as
said, an imminent act of acceptance that the IBE-act is truth-preserving. This act is imminent in
virtue of the subject’s further mental states, that is experiential states, cognitive acts already
taken and dispositions to engage in cognitive acts. If this imminent act is to help explain
epistemic responsibility —as subjective truth-connections do in the case of non-foundational
acts, then there must be a psychological explanatory connection between the imminent defense
and one’s taking the IBE-act. For if the potential defense were psychologically ineffective, then
how could it account for responsibility? It must be in virtue of the defensibility that the subject
takes her act. Therefore the further mental states which constitute the defensibility’s
psychological implementation must psychologically explain (in part) the IBE-act. But this
precludes the possibility that the IBE-act is psychologically fundamental in the following sense:
An act is psychologically fundamental if and only if it is not taken because of any further actual
or potential cognitive act. It may be argued that some cognitive act must be psychologically
fundamental. For if there are not, then a vicious regress of cognitive acts ensues. It requires that
there is an infinity of actual or imminent cognitive acts. And this, it may be argued, is impossible
for limited human minds.

Whatever may be thought of this worry, it again does not seem to allow to point to a relevant
difference between the availability of a truth-connection defense and the availability of a
pragmatic defense. For if the pragmatic defensibility is to account for responsibility then it has to
be psychologically implemented too. And if psychological implementation poses a problem for
truth-connection defensibility, why should not there be the same problem for pragmatic
defensibility.

In sum, it is hard to see a sense in which pragmatic defensibility is less problematic for an
account of epistemic responsibility than a thoroughgoing truth-connection defensibility. If the
latter gives raise to a regress or circularity problem, then so does the former. Enoch and
Schechter clearly agree, as is witnessed by the following passage:

[. . . ] A thinker may, for instance, be justified in relying upon her favorite thermometer because
she is antecedently justified in believing that it reliably indicates the prevailing temperature. But
holding a belief about the reliability of a belief-forming method cannot be a necessary condition
for being justified in employing it as basic [. . . ]. [. . . ] since the relevant reliability beliefs must
themselves be justified, the threat of an infinite regress or vicious circularity looms large. And
analogous problems face views that ground the justification of our employment of basic belief-
forming methods in the justification of other beliefs. [. . . ] we account for the justification of
employing basic belief-forming methods directly, and not via the justification of any belief.
[emphasis added] [Enoch and Schechter, 2008, 549]

… The focus of much of the discussion in this section has been on Enoch and Schechter’s
account. But it is clear that the problem raised is a problem for anyone who endorses the
Defensibility Strategy. Having an available defense implies that the defensible act’s standing
depends on the standing of some further actual and imminent acts. But it is not clear at all why
this should not lead to the same regress problems that motivated the appeal to pragmatic features
in the first place.

If what has been argued in this section is right, then the following claim should be added to the
pragmatic foundationalist’s set of claims, (PF1) – (PF4):

(PF5) The pragmatic feature of the foundational act does not depend on the subject’s other actual
or potential cognitive acts (where this includes instances of practical reasoning).

… It is an assumption of pragmatic epistemic foundationalism that there is “a bottom”. Some

standing which conveys the standing associated with subjective truth-connection is not itself a
standing associated with subjective truth-connection. The motivation for this assumption is
evidently a classic worry about regresses [fn 5 Dretske, 2000a, cf.],[549 Enoch and Schechter,
2008, cf.]. Notoriously there are several such worries.4 One is that finite minds as ours could not
“contain” the infinity of states necessary to provide a subjective perspective on the acts, a
subjective perspective on the subjective perspective and so on. Another worry is that some
standing has to be underived, least we are incapable of giving an illuminating account of
epistemic standing. An account of a kind of standing of act α in terms of the same kind of
standing of act ẞ can presumably not be the last word on the matter.

… As said, the conception of epistemic responsibility currently under consideration does justice
to the involuntariness of cognitive acts. And it still associates blame with a choice. Of all the
conceptions of epistemic responsibility discussed in this chapter it is therefore the one that is
most robustly deontological. For blame plausibly implies that there has been a decision to act in
one way or another. But for the same reasons for which the defensibility strategy failed, that
conception does not serve the pragmatic epistemic foundationalist’s purpose. For it is
incompatible with his fear of regress and circularity. If there is a regress problem concerning
subjective truth-connections, then there is a regress problem concerning evidence-related
activities which subjectively appear to be beneficial.

… No completely satisfying conception of epistemic responsibility has emerged from the

discussion. It seems that we are lead to a difficult choice between the following: Allow epistemic
responsibility to have a non-foundational structure (for instance by endorsing infinitism)16, or
give up on a notion of epistemic responsibility with robustly deontological connotations.

31 The most fully developed version of the argument I am about to give can be found in
Boghossian [2003]. My version differs from his in some respects, mostly a matter of emphasis.
First, his characterization of the view the argument is supposed to refute is compatible with both
strong inferential internalism and a variant of the psychological view. Second, I put the emphasis
on the argument’s power to refute an explanatory claim. The Carrollian problem in my opinion is
not so much that the attacked view gives rise to a regress of reasons, as that it attempts to
circularly explain why one has a reason. This will become clearer in the text.

32 If non-factualism about epistemic justification is true, then all that can be said in explaining
why certain kind of (fundamental) inferential steps are justified, is that these are the fundamental
modes of reasoning we accept. Nothing further is expressed by these epistemic evaluations than
that we fundamentally accept the mode in question [cf. Field, 1998, 2000, Gibbard, 1995].

33 There are cases where it is in fact a controversial issue whether there is a mystery about why
one has a reason or not. For instance, proponents of inferential accounts of perceptual
justification would perhaps claim that it is mysterious how the mere perception of an object gives
me a reason to believe that there is such an object. They would argue that just as in the example
above one needs some additional information, in particular something like the information that
the subject has a reason to believe that she has a perceptual state and that this state indicates the
truth of its content. Their opponents would argue that it is not at all mysterious in the first place
how the perceptual state can contribute to the subject’s having a reason. They would perhaps
claim that this appears mysterious only for those with misguided general views on justification.
Similarly in the case of deduction, minimalists would perhaps argue against the strong inferential
internalist that it is not mysterious how the justified premises can confer justification upon the
conclusion. However, the example of the object I see in the wood is not such a controversial
case. In that case the additional information is clearly required to understand why I have a reason
to believe that X was in the wood.
34 Or a subset of that set. See footnote 35.

35 According to strong inferential internalism, L is also involved in the inference. In order to

remain neutral at this point one must say that it reveals the connection among some of the
involved propositions.

36 In a relatively early paper on the Lewis-Carroll-Problem by J.F. Thomson [1960] entitled

“What Achilles should have said to the tortoise” the observation that no “tighter” logical relation
than entailment can obtain is claimed to be sufficient to avoid the problem. This seems to miss
the aspect to which I allude here as that of conforming to a principle of reasoning. A step to a
conclusion entailed by a set of premises can fail to conform to a principle of reasoning. In that
case addition of further premises, although not altering the kind of logical relation between
premises and conclusion, may well have the effect that the step now conforms to a principle of
reasoning.

37 Something like: modal reasoning: Believe q, if you believe 2q! (Or else give up 2q!)

38 The following adaption of the Carrollian dialogue illustrates the predicament of strong
inferential internalism as I understand it:

A: Why is S’ step from p to q justified?

B: Because S’ also believes that if p, then q.
A: I see, S infers q from p and if p, then q. This seems OK.
A: But wait, thinking of it, why is S’ step from p and if p, then q to q justified?
B: Oh, because S also justifiably believes that that step preserves the truth.
A: I see, S infers q from p, if p, then q and if ‘p’ and ‘if p, then q’, then q.
A: But I still don’t get it. Why is S justified in taking the last inferential step?
B: (in a mood of despair) Because S believes that that last step preserves the truth.

Here B and A do not directly appeal to the proposition corresponding to L, but rather the one
corresponding to If (1) and (2), then (3). There is a difference, but it is ultimately irrelevant to the
argument against explanation. There does not seem to be an explanatorily more fundamental
principle linking belief in L, (1) and (2) and the conclusion (3).

39 I do this in footnote 38.

40 That that particular regress, i.e. a regress concerning propositions such as L in the course of
inferential justification, is not vicious is suggested by C. Wright [2001]. Wright does, following
Boghossian, not distinguish between the explanatory regress and the regress of dependences. For
other epistemic regresses and the claim that one can have justification for an infinity of
propositions, see [Sosa, 1980, Klein, 1999].

41 I discuss in more detail a version of the Lewis-Carroll-Problem as it pertains to a view similar

to strong inferential internalism, but appealing to a proposition about epistemic support instead
of L, in chapter 4.4.
44 That there are a priori foundations is a very traditional and influential idea. In recent times it
has been endorsed among many others by Chisholm, Bonjour [1998, 2003, 2005],Burge
[1993],Bealer [1992], Goldman [1999a].

45 For accounts of a priori intuitions, see Goldman and Pust [1998], Bealer [1998, 1999, 2004],
Sosa [1998].

46 I use “held for a reason” as a stylistic variant of “believed for a reason”.

47 For the distinction between support by reasons and rational requirement, see there in chapter
2.4.

13 There is an additional requirement which would yield the result that reflection is only on
mental properties. Suppose in addition to (HO-R) the following more general higher-order
requirement:

general higher-order requirement (HO-G): Every belief is justified, only if the subject is in a
position to reflectively justifiably come to believe that it is justified.

Since antecedent beliefs used in reflection have to be justified, every antecedently justified belief
implies that the subject is in a position to reflectively justify that the antecedent belief is justified.
Since one cannot draw on infinitely many antecedent beliefs (let us suppose), the ultimate
reflective reasons are provided by introspection and a priori intuition. (In fact, the requirement,
as is widely believed, can not avoid the regress anyway, for the introspective, as well as the
reflectively justified beliefs are under the very same requirement.) This would have the result
that facts about justification would be ultimately known by appeal to reasons provided by
introspection and a priori intuition. Some have been happy to endorse this claim, e.g. Russell
[1997].

32 Furthermore, infinitism can avoid the circularity only on the following condition:

(GP) The justification for H does not rest on the justification for any proposition about inference
generally, i.e. H’ extended to all kinds of inference.

For in that case, the circularity with respect to the justification for the generalized H’ obtains
again (on the assumption that the justification for the generalized H’ is inferential).

33 For instance Goodman [1955]. However the immediate judgment is according to him further
supported by some general judgment (which is in turn inductively supported by the different
instance-judgments. He thus adopts some form of holistic coherentism. Still the (first) instance-
judgment must have some initial justifying force by itself.

34 Boghossian does not focus on H but rather on The conclusion is justified. I come to this
question below.
35 To my knowledge the first version of a similar argument is Van Cleve [1984].

36 To my knowledge the suggestion to appeal to non-propositional states in connection with the

kind of problem exposed here was first made by Wright [2001].

37 Bonjour [2005] and Boghossian [2001, 2003] both claim that the argument stems from Lewis
Carroll. I call “Carrollian argument” the distinct argument of section 4.4. For the differences see
below in the text.

38 One might perhaps think that there could be some holistic support relation between H and (3),
so that the deductive justification for (3) does not asymmetrically depend on the justification for
H. But if explanation is asymmetric, then the epistemic dependence must plausibly be
asymmetric too. And I conjecture that this is sufficient for the skeptical argument.

39 In fact, according to many, one can distinguish between two sorts of epistemic commitments:
one corresponding to a commitment because one has a good epistemic reason, the other to a
commitment because one is rationally required. For the distinction, see chapter 2.4.

40 This reading corresponds to what I have called the “Lewis-Carroll-Problem for the
psychological view” in the preceding chapters.

41 “motivation” is often reserved for the practical case as motivation for action. I use it
freely for the move to a belief which one so comes to believe for a reason.

42 This point is made by Gilbert Harman in a number of places.

46 Klein [1999] argues against Sosa [1980] that infinitism is compatible with a reductive
naturalistic account of epistemic justification.

4 For a classic paper in which some of these regresses are disentangled, see Sosa [1980]. See
chapters 2 and 4 of this thesis for a disentanglement of the issues with respect to deductive
justification.

16 Such a view is held by Klein [1998, 1999]

THREE ARGUMENTS AGAINST FOUNDATIONALISM: ARBITRARINESS, EPISTEMIC
REGRESS, AND EXISTENTIAL SUPPORT
forthcoming in Canadian Journal of Philosophy
Daniel Howard-Snyder and E.J. Coffman

Abstract. Foundationalism is false; after all, foundational beliefs are arbitrary, they do not solve
the epistemic regress problem, and they cannot exist without other (justified) beliefs. Or so some
people say. In this essay, we assess some arguments based on such claims, arguments suggested
in recent work by Peter Klein and Ernest Sosa.

A particular belief of a person is basic just in case it is epistemically justified and it owes its
justification to something other than her other beliefs or the interrelations of their contents; a
person’s belief is nonbasic just in case it is epistemically justified but not basic. Traditional
Foundationalism says that, first, if a human being has a nonbasic belief, then, at bottom, it owes
its justification to at least one basic belief, and second, there are basic beliefs. Call the second
thesis Minimal Foundationalism. In this essay, we assess three arguments against Minimal
Foundationalism which we find in recent work of Peter Klein and Ernest Sosa.1

1. Foundationalism and Arbitrariness

Peter Klein puts his case against Foundationalism succinctly as follows:

[F]oundationalism is unacceptable because it advocates accepting an arbitrary reason at the base,

that is, a reason for which there are no further reasons making it even slightly better to accept
than any of its contraries. (Klein 1999, 297)

The argument suggested here is plain enough:

The Argument from Arbitrariness

1. If Foundationalism is true, then there are basic beliefs of the following kind: S’s basic belief
that p is justified although there are no further reasons that make it even slightly better that S
believe p rather than any of p’s contraries.

2. There can be no such beliefs.

3. So, Foundationalism is false. (1, 2)

A question arises: what sort of thing is a “reason,” according to Klein? As it turns out, Klein uses
“reason” to refer both to beliefs, which are a certain sort of mental state, and to propositions,
which are not.2 Let’s not worry about which Klein meant and consider both options.

Suppose reasons are propositions. Then premise 1 reads:

1a. If Foundationalism is true, then there are basic beliefs of the following kind: S’s basic belief
that p is justified although there are no further propositions that make it even slightly better that
S believes p rather than any of p’s contraries.

Unfortunately, 1a is false. There are versions of Foundationalism according to which there are
further propositions the truth of which make it (at least) slightly better that S believes p rather
than one of its contraries, even if S’s belief that p is basic and justified. To illustrate: suppose that
Evan’s (allegedly) basic belief that the ball is red owes its justification to the ball’s looking red to
him and not to any other beliefs of his. This supposition is compatible with a version of
Foundationalism according to which Evan’s belief owes its justification to his visual experience
because, in part, these two propositions are true:

(A) When one’s belief that a ball is red is caused in normal circumstances by the ball’s looking
red to one, it is very likely that the ball is red, much more likely than that the ball is yellow or
blue, for example.

(B) Evan’s belief that the ball is red was caused in normal circumstances by the ball’s looking
red to him.

So, on this version of Foundationalism, there are some further propositions, namely (A) and (B),
the truth of which makes it (at least) slightly better that Evan believes the ball is red rather than,
say, yellow or blue. Other versions of Foundationalism have the resources to make the same
point, mutatis mutandis. Indeed, it’s difficult to think of one that lacks them; after all, each
version of Foundationalism has some story to tell about how basic beliefs are justified and that
story will consist of some propositions the truth of which would make it the case that for each
basic belief that p, it is (at least) slightly better that the person who holds it believes that p rather
than any of p’s contraries.

Perhaps Klein’s line of thought will fare better if we take “reasons” to refer to beliefs. In that
case, premise 1 of The Argument from Arbitrariness reads:

1b. If Foundationalism is true, then there are basic beliefs of the following kind: S’s basic belief
that p is justified although there are no further beliefs that make it even slightly better that S
believes p rather than any of p’s contraries.

Is this an improvement? We think not. There are versions of Foundationalism according to which
there are further beliefs the truth of which make it (at least) slightly better that S believes p rather
than one of its contraries, even if S’s belief that p is basic and justified. To illustrate: consider
Evan again, and suppose someone else, say William, correctly believes both (A) and (B). This
supposition is compatible with a version of Foundationalism according to which Evan’s belief
owes its justification to his visual experience because, in part, what William believes, namely

(A) and (B), are both true. So on this version of Foundationalism, Evan’s belief can be basic
even if there are some further beliefs—e.g., William’s belief that (A) and (B)—the truth of which
makes it at least slightly better that Evan believes that the ball is red, rather than yellow or blue.
No doubt, the same point can be made by other versions of Foundationalism, mutatis mutandis.
Perhaps it matters whether Evan himself is the one who believes (A) and (B). If so, then
Klein’s purposes might be better served by

1c. If Foundationalism is true, then there are basic beliefs of the following kind: S’s basic belief
that p is justified although there are no further beliefs of S that make it even slightly better that S
believe p rather than any of p’s contraries.

Is this a change for the better? No; 1c is false. That’s because there are versions of
Foundationalism, like the one sketched above, according to which S’s belief that p can be basic
and justified even though there are further beliefs of S the truth of which makes it (at least)
slightly better that S believes that p rather than one of its contraries. To illustrate: return to Evan,
and suppose that he correctly believes both (A) and (B). That supposition does not imply that
Evan’s belief that the ball is red is not basic. For the mere fact that Evan believes (A) and (B)
does not imply that his belief that the ball is red owes its justification to those two beliefs of his.
To suppose otherwise is to fail to distinguish two states of affairs:

• Evan believes that the ball is red and Evan believes (A) and (B).
• Evan’s belief that the ball is red owes its justification to his belief that (A) and (B).

While the second implies the first, the first does not imply the second. Thus, so long as Evan’s
belief that the ball is red does not owe its justification to his belief that (A) and (B), his belief
that the ball is red might well be basic.

Perhaps one will object: “But surely, if Evan believes (A) and (B), then his belief that (A) and
(B) must be, at least in part, his grounds for his belief that the ball is red. In that case, Evan’s
belief that the ball is red cannot be basic.”

We deny both claims here. The first claim—that if Evan believes (A) and (B), then his belief that
(A) and (B) must be, at least in part, his grounds for his belief that the ball is red—fails to
distinguish two states of affairs:

• Evan believes that the ball is red and Evan believes (A) and (B).
• Evan’s belief that (A) and (B) is, in part, Evan’s grounds for believing that the ball is red.

While the second implies the first, the first does not imply the second. That’s because, in general,
one can believe the premises of an argument and yet not believe the conclusion on that basis but
rather on the basis of something else. The second claim—that if Evan’s belief that (A) and (B) is,
at least in part, his grounds for his belief that the ball is red, then the latter belief cannot be basic
—is false. That’s because a basic belief can have multiple sources of justification; when it does,
its epistemic status is overdetermined. For example, suppose that Evan believes that the ball is
red on the basis of both its looking red and an inference from his belief that (A) and (B).

So long as Evan’s belief that the ball is red would be justified on the basis of the ball looking red
to him absent his inference from (A) and (B), it is basic.3
Some critics of Foundationalism have failed to recognize that basic beliefs can be epistemically
overdetermined. Their failure has led them to specious objections. A recent case in point is Susan
Haack’s objection to what she calls “weak foundationalism,” according to which a basic belief is
defeasibly justified by something other than a belief (Haack 1993, 16), say, an experience. At
first blush, says Haack, weak foundationalism is a sensible account of the common view that, for
example, if Peter believes that there is a dog before him on the basis of it’s phenomenally
appearing to him as if a dog is before him, then Peter’s belief is justified, but only defeasibly
justified since the appearances might be misleading, etc.. “At second blush, however,” Haack
continues,

an awkward question arises: would not [Peter] be more justified, or more securely justified, in
believing that there is a dog before him if he also justifiedly believed that his eyes are working
normally, that he is not under the influence of post-hypnotic suggestion, that there are no very
lifelike toy dogs around, etc., etc.? Surely, he would. But the weak foundationalist cannot allow
for this, for his story is that basic beliefs get their justification exclusively from something other
than the support of further beliefs;... (Haack 1993, 31, our emphasis)

Exclusively?! The problem with the argument is evident: weak foundationalism states that a basic
belief is defeasibly justified by something other than a belief, which does not imply that there
cannot be anything else that contributes to its justification. In particular, it does not imply that
“basic beliefs get their justification exclusively from something other than the support of further
beliefs”.

How could Haack make such a mistake? It is instructive to see how. The passage just quoted
above continues as follows:

…to allow that they [basic beliefs] get some justification from experience and some from the
support of other beliefs would violate the one-directional character of justification, on which,
qua foundationalist, [the weak foundationalist] insists. (Haack 1993, 31-32)

But why suppose that the weak foundationalist, qua foundationalist, insists that the character of
justification is “one-directional” in such a way that a basic belief cannot derive justification from
both experience and other beliefs? Here’s Haack’s answer:

[H]ere and throughout the book [(Haack 1993)], ‘foundationalism’ will refer to theories of
justification which require a distinction, among justified beliefs, between those which are basic
and those which are derived [nonbasic], and a conception of justification as one-directional, i.e.
as requiring basic to support derived beliefs, never vice versa. (Haack 1993, 14, our emphasis)

So, according to Haack, the weak foundationalist, like all foundationalists, is committed to the
claim that basic beliefs cannot receive any justification from nonbasic beliefs—and so cannot be
epistemically overdetermined. Why? Because, well, that’s the way she’s going to use the word
‘foundationalism’. We leave it to the reader to discern the merits of an argument against
Foundationalism that, at bottom, relies on a critic’s stipulative definition of ‘foundationalism’.4
Returning to the main thread of discussion, one might suggest that we have missed Klein’s point.
Klein should be understood as asserting

1d. If Foundationalism is true, then there could be basic beliefs of the following kind: S’s basic
belief that p is justified although there are no further beliefs of S that make it even slightly better
that S believe p rather than any of p’s contraries.

Since Foundationalism implies that there could be basic beliefs of this kind, and there couldn’t
be, Foundationalism is false.5

What should we make of this variation on Klein’s theme? It seems to us that what’s at issue here
is not Foundationalism but epistemic externalism and internalism, both of which are compatible
with Foundationalism. We do not have the space to go into the matter in depth here, and so we
will content ourselves with a general observation. Externalists tend to allow for cases in which
one’s belief that p is justified even if one has no further beliefs that support p over its contraries,
whereas (some) internalists do not. So if you are an externalist, you will probably not be much
impressed by the premise above that there could not be a case of the sort in question.6 However,
if you are an internalist, and, more importantly, if your particular brand of internalism implies
that one’s belief that p is justified only if one has some further beliefs that support p over its
contraries, you can still be a foundationalist (and hence deny 1d) provided that you allow that in
some cases, namely the case of basic beliefs, one’s belief does not owe its justification to those
further beliefs. (See our discussion of 1c above.) Of course, you will need to explain why one
must have those further beliefs in the first place, but that explanatory demand does not arise from
Foundationalism; it arises from your brand of internalism.

So far as we can see, there is but one way to modify premise 1 of Klein’s Argument from
Arbitrariness so that it is clearly true:

1e. If Foundationalism is true, then there are basic beliefs of the following kind: S’s basic belief
that p is justified although there are no further beliefs of S that both justify it (in large part) and
make it even slightly better that S believe p rather than any of p’s contraries

But in that case, premise 2—the claim that there can be no such beliefs—is nothing more than a
stylistic variant on the sentence “Foundationalism is false”. No good argument has a premise
which is a mere stylistic variant of its conclusion.

We see no way to rescue Klein’s Argument from Arbitrariness. Let’s turn to his contention that
basic beliefs do not solve the epistemic regress problem.

2. Foundationalism and Epistemic Regress

Basic beliefs have a theoretical role to play. Their entrée onto the epistemological stage is
arguably exhausted by that role. If they cannot perform it, they do not exist.7 What is that role?
Well, basic beliefs are supposed to solve the epistemic regress problem. But what is that
problem, exactly?
Before we answer that question, it is essential to understand that any discussion in this area must
be conducted in terms of individual persons and their beliefs; for Foundationalism is a thesis
about the beliefs of individuals and the conditions under which their beliefs are justified. That is,
it is a thesis about doxastic justification, a thesis about the conditions under which a person’s
beliefs, as opposed to propositions or persons are justified.8 With this in mind, we can describe
the regress problem that Foundationalism aims to solve.

Suppose that S’s belief that q is justified; furthermore, suppose that S believes that q on the basis
of an inference from two other beliefs of hers, her belief that if p, then q, and her belief that p.
Finally, suppose that S’s belief that q owes its justification to these other beliefs of hers via this
inference. Then, S’s belief that q is inferentially (or mediately, or indirectly) justified.
Now, a simple question arises: How can S’s belief that q be justified on the basis of an inference
from the contents of other beliefs of hers? How, that is, can S’s belief that q be inferentially
justified? Could it be that S’s belief that q is inferentially justified on the basis of her belief that
if p, then q, and her belief that p, while neither of those beliefs is justified—the unjustified
justifier option? Or, could it be that S’s belief that q is inferentially justified on the basis of her
justified belief that if p, then q, and her justified belief that p, and that these latter beliefs of hers
are inferentially justified in their turn on the basis of some further beliefs of hers, and so on, so
that, ultimately, her belief that q owes its justification to itself—the circular justification option?
Alternatively, could it be that S’s belief that q is inferentially justified on the basis of her justified
belief that if p, then q, and her justified belief that p, and that these latter beliefs of hers are
inferentially justified on the basis of some further beliefs of hers, and so on, ad infinitum, for
infinitely many non-repeating beliefs of hers—the infinite regress option? Finally, could it be
that S’s belief that q is inferentially justified on the basis of her justified belief that if p, then q,
and her justified belief that p, and that these latter beliefs of hers are non-inferentially justified,
that is, justified but not on the basis of any other beliefs of hers—the basic belief option? The
regress problem consists in explaining which of these options is correct and why.9 The traditional
foundationalist contends that the basic belief option is the only feasible option.

Now, according to Klein, what’s wrong with Foundationalism is that it cannot solve the regress
problem. So we have the following argument:

The Argument from Failure to Solve the Regress Problem

1. If Foundationalism is true, then basic beliefs can solve the regress problem.
2. Basic beliefs cannot solve the regress problem.
3. So, Foundationalism is false. (1, 2)

We concur with premise 1, but what about 2? What does Klein have to say on its behalf? We
discern three lines of thought in his writing.

2.1 Worries about meta-justifications

We can get at the first line of thought by returning briefly to an earlier matter in section 1. One
version of the Argument from Arbitrariness affirmed premise 1a, which states that if
Foundationalism is true, then there are basic beliefs of the following kind: S’s basic belief that p
is justified although there are no further propositions that make it even slightly better that S
believe p rather than any of p’s contraries. We argued that Foundationalism’s basic beliefs are
compatible with there being some further propositions of the sort in question, and we illustrated
the point with a case involving (A) and (B). Many foundationalists want to go a step further.
They insist that something like those two propositions must be true for each basic belief. Indeed,
according to them, any viable version of Foundationalism requires that, for any human person S
and proposition p, S’s belief that p is basic only if there is some property F such that S’s belief
that p has F and B’s having F makes p likely to be true. Such pairs of propositions are standardly
called “meta-justifications”. Suppose, if not because you believe it then just for the sake of
argument, that these foundationalists are correct.

With this supposition in place, we are in a position to hear directly from Klein why basic beliefs
cannot solve the regress problem.

[C] Can [the foundationalist] avoid advocating the acceptance of arbitrary reasons by moving to
meta-justifications?…. Pick your favorite accounts of the property, [F]. I think…that the old
Pyrrhonian question is reasonable: Why is having [F] truth-conducive? Now, either there is an
answer available to that question or there isn’t…. If there is an answer, then the regress
continues—at least one more step, and that is all that is needed here, because that shows that the
offered reason that some belief has [F] or some set of beliefs has [F] does not stop the regress. If
there isn’t an answer, the assertion is arbitrary. (Klein 1999, 303, emphasis added)

[D] To generalize: Foundationalism…cannot avoid the regress by appealing to a meta-claim that

a belief having some property, [F], is likely to be true. That claim itself requires an argument that
appeals to reasons…. For surely a reason is required to justify the belief that propositions with
property, [F], are likely to be true; and whatever justifies that claim will require a reason; and—
well, you get the point. (Klein 1999, 304, emphasis added)

[E] My point is merely that moving to the meta-level, that is, arguing that such beliefs [i.e. basic
beliefs] are likely to be true because they possess a certain property, [F], will not avoid the
problem faced by foundationalism. Either the meta-justification provides a reason for thinking
the base proposition is true (and hence the regress does not end) or it does not (hence, accepting
the base proposition is arbitrary). (Klein 1999, 304, emphasis added)

What is the argument here?

Klein means to pose a dilemma: either arbitrariness or failure to stop the regress. But how do
these unpalatable options arise? We discern two answers in the texts above. One focuses on
whether the meta-justification provides a reason for the proposition believed—see quotation [E],
the italicized bit. The other focuses on whether there is a further reason for the metajustification
—see quotations [C] and [D], the italicized bits. Klein does not distinguish these answers but we
shall; for whether there is a reason for something is not the same issue as whether it provides a
reason for something else. So we have two arguments to assess.
For ease of exposition, let ‘there is an MJ’, which stands for ‘there is a metajustification’,
abbreviate the following sentence: ‘There is some property F such that S’s belief B that p has F
and B’s having F makes p likely to be true’. Here is the argument in [E]:

The Does-MJ-provide-a-reason-for-p? Argument

2a. S’s basic belief B that p can solve the regress problem only if B can stop the regress and B
owes its justification to something other than S’s other beliefs.
2b. B owes its justification to something other than S’s other beliefs only if there is an MJ.
2c. If there is an MJ, then either it provides a reason for p or it does not.
2d. It is false that MJ does not provide a reason for p. (Otherwise, arbitrariness.)
2e. If MJ does provide a reason for p, B cannot stop the regress.
2f. So, if there is an MJ, B cannot stop the regress. (2c-2e)

Premise 2 follows.10 What should we make of this argument?

Given our assumption that 2b is true, the argument is above reproach, with one exception—2e is
false. To see why, note that if 2e is true, then, if MJ provides a reason for p, B owes its
justification to some other beliefs of S. But B owes its justification to some other beliefs of S
only if her grounds for B are other beliefs of hers. So if 2e is true, then, if MJ provides a reason
for p, B’s grounds are other beliefs of S. This implication (the sentence in italics) is false,
however. An MJ is just a pair of propositions, existing in the abstract. Obviously, they can exist
and “provide a reason” even if S is completely unaware of them.

Perhaps one will object: “But what if S is aware of an MJ? Indeed, what if she believes it,
perhaps even justifiedly? In that case, surely, B does not stop the regress.” Two replies. First, the
objection is not responsive to the case offered; it leaves it open that basic beliefs can solve the
regress problem in those cases in which S is not aware of or does not believe an MJ. Second, the
conclusion does not follow. It follows that B does not stop the regress when S is aware of or
believes an MJ only if S’s being aware of or believing an MJ entails that B owes its justification
to her awareness or belief; no such entailment holds, however.

Perhaps Klein’s other argument, the one in [C] and [D], will fare better. Here it is:

The Is-there-a-reason-for-MJ? Argument

2a. S’s basic belief B that p can solve the regress problem only if B can stop the regress and B
owes its justification to something other than S’s other beliefs.
2b. B owes its justification to something other than S’s other beliefs only if there is an MJ.
2c*. If there is an MJ, then either there is a further reason why B’s having F makes p likely to be
true or there is not.
2d*. It is false that there is no further reason why B’s having F makes p likely to be true.
(Otherwise, arbitrariness.)
2e*. If there is a further reason why B’s having F makes p likely to be true, B cannot stop the
regress.
2f. So, if there is an MJ, B cannot stop the regress. (2c*-2e*)
Is each new premise true? We think not.

Consider premise 2d* first. Recall that the schema for a meta-justification is this: there is some
property F such that B has F and B’s having F makes p likely to be true. For any candidate for a
basic belief that p, that schema can be satisfied by a substitution instance for F for which there is
no further reason why B’s having F makes p likely to be true; so, 2d* is false. Let us explain.

For any candidate for F that you might think, at first blush, satisfies the schema—e.g., one
natural candidate for F in the case of Evan’s belief that the ball is red is the property of being a
belief that is caused in normal circumstances by a ball’s looking red (of course, there are other
candidates)—there will be some reason why B’s having F makes p likely to be true that can be
specified in terms of the relevant fundamental features of the world and the laws of nature that
govern them. Call this sort of reason a Fundamental Reason, and call the sort of complex
property specified by a Fundamental Reason a Fundamental Property. Now, premise 2d* is
equivalent to the proposition that, for any F and for any belief B that p, there is a further reason
why B’s having F makes p likely to be true. But suppose the substitution instance for F is a
Fundamental Property. 2d* implies that in that case there would have to be a further reason why
B’s having the Fundamental Property makes p likely to be true. But that implication seems false.
Perhaps it is a brute contingent fact that B’s having the Fundamental Property makes p likely to
be true. Or perhaps the Fundamental Property is specified in such a way that it is a necessary
truth that if B has it, p is likely to be true. If either of these suggestions is correct (and surely one
of them is), there is an MJ for S’s belief B that p, but there is no further reason why B’s having
the Fundamental Property makes it likely that p. All one can say is “that’s just the way it is”. For
this reason, 2d* is false.

As for 2e*, suppose that there are infinitely many non-repeating reasons why B’s having F
makes p likely to be true. Even so, it does not follow that B cannot stop the regress since it does
not follow that what S goes on in believing p is any member of this infinite set of nonrepeating
reasons; indeed, it does not follow that B cannot stop the regress even if she justifiedly believes
every member of it.

2.2 Klein’s Infinitism and Foundationalism

Elsewhere in (Klein 1999), Klein suggests a different argument for premise 2 of the Argument
from Failure. He introduces two principles for one’s having a justification for p:

Principle of Avoiding Circularity (PAC): S has a justification for p only if p is not in its own
evidential ancestry for S,

and

Principle of Avoiding Arbitrariness (PAA): S has a justification for p only if there is some
reason, p*, available to S for p; and there is some reason, p**, available to S for p*; etc.

He then tells us that

the combination of PAC and PAA entails that the evidential ancestry of a justified belief must be
infinite and non-repeating… It is the straightforward intuitive appeal of these principles that is
the best reason for thinking that if any beliefs are justified, the structure of reasons must be
infinite and non-repeating. (Klein 1999, 299, his emphasis)

One might see in these principles an argument for premise 2 of the Argument from Failure, along
the following lines:

The Infinitism Argument

2a. S’s basic belief B that p can solve the regress problem only if B is justified and B owes its
justification to something other than S’s other beliefs.
2b. B is justified only if S has a justification for p.
2c. S has a justification for p only if there is available to S an infinite, non-repeating series of
propositions beginning with p such that each (non-initial) proposition in the series is a reason for
believing its predecessor. (2c combines PAC and PAA.)
2d. If there is available to S an infinite, non-repeating series of propositions beginning with p
such that each (non-initial) proposition in the series is a reason for believing its predecessor, then
B does not owe its justification to something other than S’s other beliefs.
2e. So, B is justified only if B does not owe its justification to something other than S’s other
beliefs. (2b-2d)
2f. So, B can solve the regress problem only if B both does and does not owe its justification to
something other than her other beliefs. (2a, 2e)
2g. It is false that B both does and does not owe its justification to something other than her other
beliefs.
2. So, S’s basic belief that p cannot solve the regress problem. (2f, 2g)

Is this a good argument? Three observations will prove helpful in answering that question.

First, notice that the antecedent of premise 2b uses the locution ‘B is justified’, and the
consequent uses the locution ‘S has a justification for p’. The former refers to doxastic
justification, the latter to propositional justification. So 2b assumes that doxastic justification
requires propositional justification. If this assumption is false, 2b is false. Our main concerns lie
elsewhere, so we will grant Klein both 2b and its assumption.

Second, notice that premise 2c is ambiguous. Its ambiguity stems from the term ‘available’. One
way a proposition p can be available to S is for p to be the content of an “already formed” belief
of S’s. A different way p can be available to S is for p to be such that S is merely disposed to
believe p. So 2c might mean

2c1. S has a justification for p only if S has a belief with content p1 that constitutes a reason for
believing p; and S has a distinct belief with content p2 that constitutes a reason for believing p1;
etc., ad infinitum,

or
2c2. S has a justification for p only if S is (at least) disposed to believe a proposition p1 that is a
reason for believing p; and S is (at least) disposed to believe a proposition p2 that is a reason for
believing p1; etc., ad infinitum.

Which of these does Klein endorse? Pretty clearly, he has 2c2 in mind. Here’s what he says:

I have already said that the infinitist is not claiming that during our lifetime we consciously
entertain an infinite number of beliefs [i.e. propositions]. But what might not be so obvious is
that the infinitist is also not even claiming that we have an infinite number of…“unconscious
beliefs” if such beliefs are taken to be already formed dispositions... Consider the following
question: Do you believe that 366+71 is 437? I take it that for most of us answering that question
brings into play some of our capacities in a way that answering the question “Do you believe that
2+2=4?” does not. For I simply remember that 2+2=4... By contrast,... [w]e do not simply
remember that 366+71=437. Rather,... [w]e are disposed to think that 366+71=437 after a bit of
adding... We have a second order disposition—a disposition to form the disposition to think
something. Thus, there is clearly a sense in which we believe that 366+71=437. The proposition
that 366+71=437 is subjectively available to me because it is correctly hooked up to already
formed beliefs. (Klein 1999, 308)

Here Klein recognizes the difference between 2c1 and 2c2, and seems to endorse 2c2.

Third, notice that since premise 2 is disambiguated as 2c2, the validity of the argument is
preserved only if premise 2d is disambiguated accordingly:

2d2. If S is (at least) disposed to believe a proposition p1 that is a reason for believing p; and S is
(at least) disposed to believe a proposition p2 that is a reason for believing p1; etc., ad infinitum,
then B does not owe its justification to something other than S’s other beliefs.

With these observations in hand, we turn to an assessment of the argument.

Consider 2c2. Given the assumption that doxastic justification requires propositional justification
(without which 2b is false), 2c2 implies that

• S’s belief that p is justified only if S is (at least) disposed to believe a proposition p1 that is a
reason for believing p; and S is (at least) disposed to believe a proposition p2 that is a reason for
believing p1; etc., ad infinitum.

One might worry that while some beliefs of ours are justified, none of us is disposed to believe
infinitely many propositions. But the latter is not obviously true. For, if there were infinitely
many medium-sized octagonal objects, there would be infinitely many objects of which we are
disposed to believe that it is octagonal; and, if God were to create infinitely many angels, there
would be infinitely many objects of which we are disposed to believe that it is a creature (Klein
1999, 307-309).
More troubling is the fact that human beings have justified beliefs but they are not disposed to
believe each member of a relevant infinite, non-repeating series of propositions. Suppose Peter
believes that he’s appeared-to-greenly on the basis of an appropriate experience; and suppose
that the only relevant infinite, non-repeating series of propositions such that he is disposed to
believe even some of its non-initial members is the one whose second member is the proposition
that it seems to him that he’s appeared-to-greenly, whose third member is the proposition that it
seems to him that it seems to him that he’s appeared-to-greenly, and so on. Now, although Peter
is disposed to believe each of the propositions just mentioned and their like for some relatively
small number of reiterations, he is not disposed to believe that it seems to him that it seems to
him that…it seems to him that he is appeared-to-greenly, where the number of reiterations of ‘it
seems to him that’ filling the ellipsis is the number of nanoseconds since the Big Bang. Peter is
not unusual. No human being is even disposed to grasp such ‘long’ propositions, as we might
call them, much less believe them. Nevertheless, or so we say, despite Peter’s failure—a failure
each of us shares in—his belief that he is appeared-to-greenly may well be justified.

“But,” Klein might object, “even if Peter presently lacks the disposition to believe such long
propositions, his belief that he is being appeared-to-greenly is still justified—provided he can
develop his capacities so that he becomes disposed to grasp them. In that case, he has a second
order disposition, a disposition to form the disposition to grasp long propositions.”11

The main thing we want to say in reply is this: who do you think you’re kidding? We defy
anyone to exercise their alleged ability to develop their capacities in such a way that they are
disposed to grasp long propositions. Just try it. (We’ll wait while you give it a shot….) The truth
is you cannot. You have no second order disposition to form the disposition to grasp long
propositions and hence none to believe them.

Klein’s second order disposition objection encourages us to think that our present dispositions to
believe propositions supervene on what would be the case if we were, in effect, cognitively
divine. But that’s just plain wrong. Our present dispositions to believe supervene on our present
cognitive capacities, not on what they would be like if they were enhanced to Olympian
proportions, or embellished with a godlike vocabulary and conceptual repertoire.

Suppose that what we’ve just said about 2c2 is wrong. Suppose we have the relevant dispositions
and/or second-order dispositions. Still, 2d2 is false. For consider its contrapositive:

2d2cp. If B owes its justification to something other than S’s other beliefs, then the following
conjunction is false: S is (at least) disposed to believe a proposition p1 that is a reason for
believing p; and S is (at least) disposed to believe a proposition p2 that is a reason for believing
p1; etc., ad infinitum.

A question arises: Why couldn’t B owe its justification to something other than S’s other beliefs
—say, an experience and certain facts relating experiences to beliefs grounded on them— even if
S is disposed to believe each (non-initial) member of the relevant infinite, non-repeating series of
propositions? S’s being so disposed would spell trouble for B’s status as basic only if such
dispositions were beliefs and B owed its justification to them. But a disposition to believe is not a
belief;12 and, even it were, the fact that B is basic only if S has infinitely many beliefs of the
relevant sort does not entail that B owes its justification to them.

Now, we do not aim here to defend the claim that S has a basic belief only if S is disposed to
believe each (non-initial) member of a relevant infinite, non-repeating series of propositions; nor
do we aim to defend the claim that S has a basic belief only if S has infinitely many beliefs of the
relevant sort. Rather, we aim to point out, first, that Foundationalism is compatible with each of
these claims and, second, that at least one of those claims is constitutive of Klein’s Infinitism. So
Klein’s Infinitism is compatible with Foundationalism, and so no reason to deny it.

2.3 The reasons-giving regress problem

In more recent work, Klein offers a different line of thought in defense of premise 2 of the
Argument from Failure.13 The following quotation is representative of this strand of his
antifoundationalist thinking:

The so-called "regress problem" can be stated briefly in this way: There are only three possible
patterns of reasoning. Either the process of producing reasons stops at a purported foundational
proposition or it doesn't. If it does, then the reasoner is employing a foundationalist pattern. If it
doesn't, then either the reasoning is circular, or it is infinite and non-repeating. There are no other
significant possibilities. Thus, if none of these forms of reasoning can properly lead to assent,
then no form can….

…Suppose that an inquirer, say Fred D'Foundationalist, has given some reasons for his beliefs.
Fred offers q (where q could be a conjunction) for his belief that p, and he offers r (which could
also be a conjunction) as his reason for q. Etc. Now, being a foundationalist, Fred finally offers
some basic proposition, say b, as his reason for the immediately preceding belief. Sally
D'Pyrrhonian asks Fred why he believes that b is true. Sally adds the "is true" to make clear to
Fred that she is not asking what causes Fred to believe that b. She wants to know why Fred
thinks that b is true…. Being true to his foundationalism, he must think that there is some
warrant that each basic proposition has that does not depend upon the warrant possessed by any
other proposition.

The crucial point to note here is that Sally can grant that the proposition has autonomous warrant
but continue to press the issue because she can ask Fred whether the possession of autonomous
warrant is at all truth conducive. That is, she can ask whether a proposition with autonomous
warrant is, ipso facto, at all likely to be true. If Fred says "yes," then the regress will have
continued. For he has this reason for thinking that b is true: "b has autonomous warrant and
propositions with autonomous warrant are somewhat likely to be true." If he says "no" then Sally
can point out that he is being arbitrary since she has asked why he thinks b is true and he has not
been able to provide an answer….

The dilemma is that either Fred has a reason for thinking that proposition is true or he doesn't. If
he does, then the regress has not stopped—in practice. If he doesn't, then he is being arbitrary—
in practice.
Once again, it is crucial to recall that Pyrrhonians are not claiming that foundationalism is false.
They could grant that some propositions do have autonomous warrant which is truth-conducive
and that all other propositions depend for some of their warrant upon those basic propositions.
What lies at the heart of their view is that there is a deep irrationality in being a practicing self-
conscious foundationalist. The question to Fred can be put this way: On the assumption that you
cannot appeal to any other proposition, do you have any reason for thinking that b is true? Fred
not only won't have any such reason for thinking b is true, given that assumption, he cannot have
one (if he remains true to his foundationalism). Arbitrariness seems inevitable. Of course,
foundationalists typically realize this and, in order to avoid arbitrariness, tell some story (for
example, about privileged access) that, if true, would provide a reason for thinking basic
propositions are at least somewhat likely to be true. But then, the regress of reasons has
continued. (Klein 2003, sections 7 and 8)

What should we make of this passage?14

While it bears a strong affinity to (Klein 1999), it indicates a significant change, signaled by
Klein’s astonishing confession that he does not mean to argue that Foundationalism is false, but
rather that one cannot practice Foundationalism. It does not permit the “practicing, self-
conscious foundationalist” to play the reasons-giving game; a “deep irrationality” attends those
who attempt it. Given our aim to discover a good argument for the denial of Foundationalism,
we’re a bit nonplussed by Klein’s (dis)claim(er). Is he being serious? Or is he just being coy?
Perhaps we can make headway in achieving our aim if we approach his (dis)claim(er) like this:
although Klein (perhaps) doesn’t mean to argue that Foundationalism is false, what he says may
nevertheless be a good argument for that conclusion. So in what follows we will try to find in
what Klein says the premises of another argument against the truth of Foundationalism.

Fortunately, this is not difficult to do. For, unlike (Klein 1999), (Klein 2003) and its successors
emphasize the activity of justifying a belief—of producing, providing, and giving reasons for a
proposition one believes. And this is linked to his conception of the regress problem, according
to which the regress problem is the problem of explaining how “the processof producing
reasons” for a proposition can result in its being justified for one. The link suggests the following
train of thought: S’s allegedly basic belief that p can solve the regress problem only if S can
justify it, and S can do that only if S can give good reasons to think that p is at least likely to be
true; but in that case, S’s basic belief that p cannot stop the regress.

We find the train of thought here quite puzzling. For the regress problem that, according to
Klein, basic beliefs cannot solve is not the regress problem that, according to Foundationalism,
basic beliefs are supposed to solve. The regress problem, according to Klein, is the problem of
explaining how “the regress of providing reasons,” “the regress of reasons-giving,” or “the
process of producing reasons” can result in a proposition being justified (for a person). The
regress problem, according to Foundationalism, is the problem of explaining how there can be
inferentially justified beliefs. These are not the same problem.15 Furthermore, Foundationalism
only aims to solve the latter, not the former. Just as the theory of elements cannot be faulted in
virtue of its failure to help the practicing self-conscious chemist find water on Chikamin Ridge in
the Central Cascades, so Foundationalism cannot be faulted in virtue of its failure to help Fred
D’Foundationalist give reasons in accordance with the rules of the game that Sally D'Pyrrhonian
invites him to play. Just as the theory of elements is not about finding water, so Foundationalism
is not about giving reasons.

We can put our puzzlement more clearly if we recall The Argument from Failure, which went
like this:

1. If Foundationalism is true, then basic beliefs can solve the regress problem.
2. Basic beliefs cannot solve the regress problem.
3. So, Foundationalism is false. (1, 2)

If we read the argument without equivocation, insisting on Klein’s focus on the reasons-giving
regress problem, premise 1 is false—Foundationalism has no implications for solving that
problem. On the other hand, if we insist on Foundationalism’s focus on the inferential-
justification regress problem, premise 2 is not even mentioned in Klein’s writings. Indeed, the
premises are both true only if “the regress problem” refers to the inferential-justification regress
problem in premise 1 and the reasons-giving regress problem in premise 2.

The charge of equivocation can be avoided on the assumption that one’s belief can be justified
only if one can show that it is likely to be true. If this assumption is correct, then, since showing
that one’s belief is likely to be true is what the reasons-giving regress problem is all about, it
follows that basic beliefs can solve the inferential-justification regress problem only if they can
solve the reasons-giving regress problem. If we add this last italicized claim as a premise to The
Argument from Failure, the equivocation disappears.

Unfortunately, this additional premise is no better than the assumption that underlies it, namely
this:

The Assumption. S’s belief B that p is justified only if S can show that p is at least likely to be
true.

We think that The Assumption is false, for at least three reasons.

First, think of moral analogues. Obviously, one’s actions can be morally justified even if one is
unable to show that one’s action is not wrong. Indeed, consider any sort of justification you
please—legal, pragmatic, moral, what have you—and, clearly, an action’s being justified in that
way does not require that the one who performed the action be able to show that it passes muster
with the relevant standards. Why should epistemic justification be any different? The onus is on
those who opt for The Assumption to give a compelling reason to suppose that there is a salient
difference here. No one in the literature has succeeded on this score.

Second, epistemic nihilism looms. S can show that p is at least likely to be true only if S holds
justified beliefs whose contents make it at least likely that p is true. In that case, The Assumption
applies to the beliefs to which S must be able to appeal in order to show that p is at least likely to
be true—and we’re off to the races. While Klein qua infinitist has no problem with this, as we
argued above there is much to worry about here. In particular, no human being can so much as
grasp the contents of the beliefs generated by this regress. The result is epistemic nihilism: no
human justifiedly believes anything.

Third, counterexamples are a dime a dozen. Being able to show that a proposition is true
involves being able to give a good argument for its truth. Philosophers are too prone to take for
granted the conceptual sophistication and skills needed to give a good argument, sophistication
and skills not shared by the class of those who have justified mundane beliefs, past and present.
We need not think only of higher non-human animals in this connection. Consider young, normal
humans, like our little kids. Surely their mundane beliefs are justified—but they do not have the
conceptual sophistication and skills to argue that what they believe is true. Or consider cases in
which those who can ordinarily argue in the required way temporarily lose that capacity. Normal,
mature adults occasionally experience mental ‘seizures’—episodes that render them cognitively
‘tongue-tied’ and which temporarily render them unable to give good arguments. The causes of
such episodes are various but include shyness, insecurity, anger, shock, and the like. Even
mature, philosophically adept adults who are not in the grip of a mental seizure are sometimes
unable to produce a good argument for what they believe, even though their belief is justified.16

The upshot is this. On one reading of Klein’s Argument from Failure, it equivocates. On
the only other available reading, it relies on a false claim, The Assumption. Either way, the
Argument from Failure fails.17

3. Foundationalism and Existential Support

According to Ernest Sosa,

[A] Given that beliefs would not so much as exist without an extensive supporting cast of related
beliefs, there is an air of unreality about the foundationalist claim that beliefs might nevertheless
be justified independently of other beliefs. It is hard to conceive of the hypothetical cases that
one would naturally invoke in support of such a claim, for these would be cases where one held
the target beliefs along with the supporting conscious states but without the supporting cast of
other beliefs. But you could not possibly so much as host the target belief without a lot of the
relevant supportive beliefs. [B] Nor does it seem that you could enjoy justification for the target
belief in the absence of justification for a good number of those supporting beliefs, absent which
you could not hold the target belief at all. [C] If one nevertheless insists that, despite this [i.e.,
despite the fact that one has a justified belief, Y, only if one has other justified supporting
beliefs], in some sense the other beliefs do not help justify the target belief, one will surely be
asked to explain this special relation of justifying that can fail to relate a belief X and a belief Y
even when belief Y would not be justified (not as fully, or as well) in the absence of support by
belief X. (E. Sosa 2003, 208-09)

On the face of it, this passage argues against Foundationalism (“there is an air of unreality about
the foundationalist claim that…”). We have learned through personal correspondence however,
that Sosa did not mean to argue against Foundationalism here. Instead, he meant to request an
explanation of the “special relation of justifying” countenanced by Foundationalists. Sections [A]
and [B] were intended to contextualize Sosa’s request, which is issued in [C]. In what follows,
we try to honor that request. In order to better position ourselves to do so, we will first
reconstruct and assess the argument against Foundationalism naturally suggested by Sosa’s
words, even if they were not thus intended. Doing so will highlight some matters to which we
will later appeal.

3.1 The Argument from Existential Support

Here is the argument suggested to us by [A] and [B]:

The Argument from Existential Support by Other (Justified) Beliefs

1. If Foundationalism is true, there are basic beliefs.

2. If there are basic beliefs, there is a justified belief that is not supported by any other (justified)
beliefs of the person who has it.
3. It cannot be that there is a justified belief that is not supported by any other (justified) beliefs
of the person who has it.
4. So, Foundationalism is false. (1-3)18

What should we make of this line of thought?

Two preliminary observations are in order. First, [A] states that a basic belief requires a
“supporting cast of other beliefs”. [B] goes beyond [A] by stating that “a good number of those
supporting beliefs” must be justified.19 We could treat these as distinct arguments, but for
brevity’s sake we will assess them simultaneously, as indicated by the parenthetical “justified” in
premises 2 and 3. Second, and most importantly, notice that the term ‘supported by’ is
ambiguous. It might express the relation of justificatory support, where X justificatorily supports
Y if and only if Y owes its justification (at least in part) to X. Alternatively, it might express a
different relation, the relation of existential support, where X existentially supports Y if and only
if Y would not exist in the absence of X.20

Now to our assessment. Either premises 2 and 3 both express the relation of justificatory support
or they both express the relation of existential support. (Otherwise, the argument commits the
fallacy of equivocation.) Let’s explore each option in turn. If premises 2 and 3 both express the
relation of justificatory support, premise 3 is this:

3j. It cannot be that there is a justified belief that does not owe its justification to any other
(justified) beliefs of the person who has it.

3j is synonymous with the sentence ‘Foundationalism is false’ and thus it is unavailable as a

premise in a good argument against Foundationalism. Exploring the second option will take
considerably longer.

If premises 2 and 3 both express the relation of existential support, we have these premises:

2e. If there are basic beliefs, there is a justified belief, Y, and the person who has Y has no other
(justified) belief, X, such that Y would not exist in the absence of (justified) X.
3e. It cannot be that: there is a justified belief Y and the person who has Y has no other
(justified) belief, X, such that Y would not exist in the absence of (justified) X.

3e is fine, but why should we suppose that 2e is true? It isn’t just obvious. Passages [A] and [B]
jointly suggest this argument for 2e:

2e1. If there are basic beliefs, then there is a justified belief Y that is not supported by any other
(justified) beliefs of the person who has it.

2e2. If there is a justified belief Y that is not supported by any other (justified) beliefs of the
person who has it, then that person (the person who has Y) has no other (justified) belief X such
that Y would not exist in the absence of (justified) X.

2e follows.

Or does it? Perhaps not; perhaps the argument equivocates on the term “supported by”. After all,
2e1 is uncontentiously true only if “supported by” expresses justificatory support; and, 2e2 is
uncontentiously true only if “supported by” expresses existential support. Equivocation can be
avoided, however. There are two options to explore on that score: first, read “supported by” in
both 2e1 and 2e2 as expressing justificatory support; second, read “supported by” in both of
those premises as expressing existential support. On the first option, we disambiguate the
argument for 2e as follows:

2e1j. If there are basic beliefs, then there is a justified belief Y that lacks justificatory support
from other beliefs.

2e2j. If there is a justified belief Y that lacks justificatory support from other beliefs, then the
person who has Y has no other (justified) belief X such that Y would not exist in the absence of
(justified) X.

On the second option, we disambiguate the argument like this:

2e1e. If there are basic beliefs, then there is a justified belief Y that lacks existential support from
other beliefs.

2e2e. If there is a justified belief Y that lacks existential support from other beliefs, then the
person who has Y has no other (justified) belief X such that Y would not exist in the absence of
(justified) X.

Notice two things about these two arguments. First, premise 2e1j of the first argument is true by
definition of “basic belief” (epistemic overdetermination aside), and premise 2e2e of the second
argument is a tautology. Second, premise 2e2j of the first argument is equivalent to premise 2e1e
of the second argument; both say, in short, that a justified belief’s lacking justificatory support
from other beliefs entails that it also lacks existential support. Hence, if one premise is false, so is
the other. With that in mind, we choose to focus on premise 2e2j. What should we make of it?
At [C], Sosa expresses puzzlement over how it could be false. He says, in effect: if one insists
that justified Y lacks justificatory support from (justified) X yet enjoys existential support from
(justified) X, one has some explaining to do. But why does this position so puzzle Sosa? On the
face of it, 2e2j’s antecedent entails the denial of its consequent, given Sosa’s [A] and [B], the
upshot of which is 3e. According to 3e, if there is a justified belief Y, the person who has it must
have some other (justified) belief X without which Y would not exist. If, like Sosa, we endorse
3e, then shouldn’t we be at least somewhat sympathetic to the denial of 2e2j? We think so.
Hence, we are puzzled by Sosa’s puzzlement.

Of course, a proponent of 2e2j who holds that its antecedent entails the denial of its consequent
may also hold that its antecedent also entails its consequent. This would nicely echo the common
refrain that Foundationalism entails a contradiction. By our lights, though, it is such a proponent
of 2e2j who bears the explanatory burden here, not one who withholds on or denies 2e2j. After
all, why suppose that Y’s lacking justificatory support from other (justified) beliefs entails that it
also lacks existential support? Or, to put it slightly differently: why suppose that Y’s owing its
existence to other (justified) beliefs entails that it owes it justification to other (justified) beliefs?
We think that’s the real puzzle to dwell on.21

3.2 The Bigger Picture

Let’s return to Sosa’s question: How can it be that justified Y could not exist in the absence of
the (justified) Xs while it does not owe its justification to them? And let’s approach this question
by stepping back and taking a look at the bigger picture.

Like many of us, Sosa takes it that, in general, “beliefs would not so much as exist without an
extensive supporting cast of related beliefs.” Thus, for any allegedly basic belief, it could not
exist without an extensive cast of related beliefs. Moreover, like many of us, Sosa assumes that,
in general, given that a belief could not exist without an extensive cast of related beliefs, it
cannot be justified “in the absence of justification for a good number of those supporting
beliefs”. Thus, for any allegedly basic belief, it could not be justified in the absence of
justification for a good number of those beliefs that existentially support it. Suppose Sosa is
right. Why is he right? Why suppose that, in general, a belief could not exist without an extensive
cast of existentially supporting beliefs? And, assuming that’s right, why suppose that a good
number of those existentially supporting beliefs must themselves be justified? We hypothesize
that any plausible explanation that is neutral between Foundationalism and its denial will not
imply that a belief that is existentially supported by other justified beliefs must also be justified
by them.

We have space for only one illustration of this point. Our explanation is contentious of course, as
any other explanation would be.22 But, be that as it may, it illustrates the point we wish to make
while honoring Sosa’s assumptions and his request for an explanation.

When one forms a belief, even the most mundane belief, one applies at least one concept. But
one can apply a concept only if one has a grasp of it, at least somewhat. Grasping a concept, even
somewhat, requires the ability to discriminate between things to which it applies and things to
which it fails to apply. A disposition to perfect application is not necessary; but with respect to a
wide variety of things to which it does and does not apply, one must be disposed to apply it
correctly. Now, this capacity to discriminate consists in, among other things, the possession of
certain beliefs: non-occurrent beliefs (say, expectations) about how things to which the concept
applies would tend to behave under various conditions, as well as non-occurrent beliefs about its
interrelations with other concepts. Call such beliefs “concept-possession beliefs”. So, to return to
our earlier example, if Evan has the belief, Y, that the ball is red, then Evan has the relevant
concept-possession beliefs, the Xs, beliefs without which Y would not exist, beliefs associated
with his grasp of the concepts of a ball, redness, and the relation expressed (here) by ‘…is…’. So
it is that, in general, a belief could not exist without an extensive cast of existentially supporting
beliefs.

But what if Evan’s concept-possession beliefs are all unjustified? What if his beliefs about the
correct application of the concepts of a ball, red, etc. are all unjustified? What if his beliefs about
the relations between those concepts and other concepts are all unjustified? Then Evan is
completely ignorant of how to use those concepts, in which case he has no grasp of them at all.
So, if Evan has a grasp of the concepts required for his belief, Y, that the ball is red to exist, the
relevant concept-possession beliefs of his, the Xs, must be justified. What holds for Evan’s belief
holds for all. So it is that the existentially supporting beliefs must themselves be justified, or at
least a good number of them.

But if, in accordance with the story we just sketched, belief Y could not exist in the absence of
the (justified) Xs, doesn’t it seem possible for Y to be justified without owing its justification to
any other beliefs of S, including the (justified) Xs? Is there a puzzle with this possibility? We fail
to see one. At any rate, nothing in the story we sketched reveals why a belief that is existentially
supported by other justified beliefs must also be justified by them.

The upshot, then, is this. The distinction between existential support and justificatory support
provides the conceptual space for the possibility that, even though Y cannot exist in the absence
of the (justified) Xs, Y does not thereby owe its justification to the (justified) Xs. Y owes its
justification to them only if they are a reason or ground for which S holds Y. And that condition
is not satisfied merely in virtue of S’s having them. If, furthermore, the justificatory relation
holds between Y and some non-doxastic state of S (i.e., if some non-doxastic state of S is the
reason or ground for which S holds Y), then, given some facts relating those states and Y, Y can
be justified without owing its justification to the (justified) Xs even though Y owes its existence
to the (justified) Xs.

Naturally, we might well wonder what, exactly, those facts are that are supposed to relate those
non-doxastic states and Y. As everybody already knows, different foundationalists offer different
substantive answers to that question. It is no part of our aim here to defend any substantive
version of Foundationalism. We have been concerned only to explain at a more abstract level
how the possibility of basic belief is not vitiated by the mere fact that a basic belief cannot exist
unless the person who has it has some further (justified) beliefs. We hope to have achieved this
aim to the satisfaction of Sosa, thereby honoring his request for an explanation. No doubt there
are other explanations as well.

4. Conclusion
Neither Klein’s arguments nor the argument suggested by Sosa’s words renders Foundationalism
dubious. Of course, Foundationalism might be false all the same. All we can conclude here is
that if we are to reject it for good reasons, we must look elsewhere for them.23

1 Four notes in one. (1) Hereafter, we will leave the modifier “Minimal” in “Minimal
Foundationalism” implicit. (2) We add “…or the interrelations of their contents” to the definition
of basic belief to rule out Pure Coherentism as a version of Foundationalism. We leave this
clause implicit in the text. (3) Epistemic statuses other than justification are amenable to the
basic/nonbasic distinction, but to make this explicit would needlessly complicate the text. (4)
Our definition of basic belief needs to be qualified; see note 3 and the text to which it is
referenced.

2 The paragraph straddling (Klein 1999, 298-99), uses 'reason' to denote a belief (where belief is
a kind of mental state), while the very next paragraph uses 'reason' to denote a proposition.

3 This is the qualification referred to at note 1, part (4). Here is a more accurate statement of
what a basic belief is:
A particular belief of a person is basic just in case it is epistemically justified and it owes its
justification, at least in large part, to something other than her other justified beliefs or the
interrelations of their contents, where the italicized qualification signals that if a basic belief
owes its justification to some other belief of the person or the interrelations of their contents,
then, it would remain justified even if it did not owe its justification to those other beliefs or
interrelations. While it is crucial to recognize that basic beliefs can be epistemically
overdetermined, we will not import that point into the definition of basic beliefs that we will use
in the text.

4 Twenty years before the publication of (Haack 1993), Frederick Will, another critic of
Foundationalism, argued against the epistemic overdetermination of basic beliefs (Will 1973,
200-201). (Unlike Haack, he didn’t just stipulate away the possibility.) Three years after Will
published his book, and seventeen years before the publication of (Haack 1993), William Alston
published these words:

To say that a belief is immediately justified is just to say that there are conditions sufficient for its
justification that do not involve any other justified beliefs of that believer. This condition could
be satisfied even if the believer has other justified beliefs that could serve as grounds.
Overdetermination is an epistemic as well as a causal phenomenon. What fits a belief to serve as
a foundation is simply that it doesn’t need other justified beliefs in order to be justified itself.
(Alston 1989b, 45)

And later, ten years before the publication of (Haack 1993), Alston reiterated the point (Alston
1989c, 64). Neither of these works is cited in the Bibliography of (Haack 1993).

5 Ryan Wasserman brought this way of casting Klein’s argument to our attention.
6 In this connection, compare how externalists reply to BonJour’s case of Norman the
clairvoyant (BonJour 1985, 41-45) and (BonJour 2002, 230-31).

7 This is not to say that the only reason to think that there are basic beliefs is that they solve the
epistemic regress problem. Perhaps, as James Pryor has reminded us recently, the best reason
comes from considering examples. See (Pryor 2005, 184-85).

8 We can distinguish doxastic, propositional, and personal justification as follows. Doxastic

justification is a property of beliefs that is expressed by locutions like ‘B is a justified belief of
S’s’ and ‘S’s belief that p is justified’. Propositional justification is a property of propositions
that is expressed by locutions like ‘p is justified for S’ and ‘S has a justification for p’. Personal
justification is a property of persons (or, more broadly, cognitive subjects) that is expressed by
both belief-entailing and non-belief-entailing readings of ‘S is justified in believing p’. For
interesting work on the logical properties of and relations among these different kinds of
justification, see (Bach 1985), (Engel 1992), (Kvanvig and Menzel 1990), and (D. Sosa
unpublished).

9 There are permutations on these options, of course, as well as other options to consider. Other
options include Epistemic Nihilism, which can be seen as denying the initial supposition that S’s
belief that q is justified, and Pure Coherentism, which can be seen as denying the linear
conception of justification which is embodied by the initial supposition that S’s belief that q
owes its justification to the inference in question.

10 Like this:

2g. So, B owes its justification to something other than S’s other beliefs only if B cannot stop the
regress. (2b, 2f)
2h. So, S’s basic belief B that p can solve the regress problem only if B both can and cannot stop
the regress. (2a, 2g)
2i. It is false that B both can and cannot stop the regress.
2. So, S’s basic belief that p cannot solve the regress problem. (2h, 2i)

11 This reply is fashioned along the lines of the paragraph straddling (Klein 1999, 308-09). Note
an important difference between the “finite vocabulary objection” that Klein considers there and
the one that we have put forward. The concepts and vocabulary in our case are minimal. To
grasp long propositions of the sort we have in mind, we do not need to develop or invent new
concepts and ways of specifying them.

12 See (Audi 1994).

13 (Klein 2003), (Klein 2004, 168-69), and (Klein 2005, 132-35).

14 For a critical assessment of Klein’s line of thought that differs from the one to follow, see
(Bergmann 2002).

15 See (Alston 1989a, 26-32) and (Audi, 1993, 118-25).

16 William Alston writes:

We frequently take ourselves to know things with respect to which we have no such capacity [i.e.
the capacity to produce adequate reasons]. I often suppose myself to know that my wife is upset
about something, where I would be hard pressed to specify how I can tell, that is, hard pressed to
specify what makes it reasonable for me to believe this. The same goes for much of our supposed
knowledge about history, geography, and physical regularities. In the face of all this, why should
we accept the thesis that justification essentially involves the capacity to demonstrate
reasonableness? (Alston 1989b, 70)

Compare (Alston 1989e, 335). For more on the “show-be” distinction, see, among other places,
(Alston 1989b, 44-45), (Alston 1989c, 73-74), and (Alston 1989d, 197-98).

17 For a recent, interesting, and unsuccessful defense of The Assumption, see (Leite 2004).

18 Others who argue in this way use the word “depends” and its cognates. They say: if
Foundationalism is true, then there are basic beliefs, beliefs that can be justified without
depending on other (justified) beliefs; but that’s impossible—a belief could not so much as exist
without depending on other (justified) beliefs. See, e.g., (Davidson 1986, 311), “Emphasis on
sensation…flashing green light?,” and (Lehrer 1990, 73-75). The distinction we draw in the next
paragraph with respect to “supported by” can be applied to “depends on”. For critical discussion
of Davidson and Lehrer on this score, see (Howard-Snyder 2002, 546-53) and (Howard-Snyder
2004, 52-61), respectively.

19 Actually, [B] goes beyond [A] by stating that, according to Foundationalism, “you can enjoy
justification” in holding a basic belief only if you have some other justified beliefs (emphasis
added). But Foundationalism is first and foremost a thesis about doxastic justification, not
personal justification. We cannot just assume that the first implies the second; indeed, there are
powerful reasons to suppose otherwise. See (Bach 1985, 251-52), (Engel 1992), and (D. Sosa
unpublished).

20 We find this distinction in, among other places, (Alston 1989b, 63-4) and (Audi 1993, 151).

21 Compare (Pryor 2005, 183): “the fact that you have immediate justification to believe P does
not entail that no other beliefs are required for you to be able to form or entertain the belief that
P.”

22 Mark Moffett and Ryan Wasserman made this clear to us. We were happy to discover that at
least one other epistemologist agrees with our illustration, and uses it to similar effect. See (Pryor
2005, 198, note 5). The point was also made in (Howard-Snyder 2002, 550-53).

23 We are grateful to Peter Klein, Adam Leite, Mark Moffett, Ernest Sosa, Ryan Wasserman and
two anonymous referees for useful and interesting comments. Work on this paper was done, in
no small part, with the support of a Summer 2004 Research Grant from Western Washington
University.
Synthese (2008) 164:117–139
DOI 10.1007/s11229-007-9219-0
Foundational beliefs and the structure of justification
Kenneth Hobson

Abstract I argue that our justification for beliefs about the external physical world need not be
constituted by any justified beliefs about perceptual experiences. In this way our justification for
beliefs about the physical world may be nondoxastic and this differentiates my proposal from
traditional foundationalist theories such as those defended by Laurence BonJour, Richard
Fumerton, and Timothy McGrew. On the other hand, it differs from certain non-traditional
foundationalist theories such as that defended by James Pryor according to which perceptual
experience is sufficient to justify beliefs about the external world. I propose that justification for
propositions describing our perceptual experiences partially constitutes any justification we may
possess for beliefs concerning the external world. In this way, our justification for beliefs about
the physical world may only be inferential since it is grounded in any justification we have for at
least one other proposition. This theory occupies an intermediate position between the two
aforementioned foundationalist accounts, which allows it to sidestep problems that confront each
of them.
Keywords Foundationalism · Justification · Epistemology · Belief

1 Introduction

Traditional foundationalist theories have generally held that our ordinary beliefs about physical
objects and events in the external world are justified by justified beliefs about our sensory
experiences. I will refer to these respectively as external world beliefs and experiential beliefs.1
To avoid ambiguity between state and content, I use (a) ‘belief’ to refer to a state of someone
believing something and (b) ‘external world proposition’ and ‘experiential proposition’ to refer
to the respective contents of believing. These are our two categories of propositions and beliefs.
We might call this claim that any justification we have for believing external world propositions
must be constituted by justified experiential beliefs the justified experiential belief requirement.2

Traditional foundationalism and its justified experiential belief requirement have been criticized
for a number of reasons, but I will give attention to one criticism recently put forward by Richard
Feldman (2004a). Feldman argues against traditional foundationalism as follows. First,
according to traditional foundationalism, John knows that there is a tree there only if he believes
that he is having a treeish experience. Second, cognitive science could discover that we do not
form experiential beliefs without its also thereby showing that we lack knowledge of the mind-
independent world.3 Therefore, experiential beliefs are not necessary for empirical knowledge.
So cognitive science could show that John does not have beliefs about his treeish experience
without thus showing that John does not know that there is a tree. Therefore, John could know
that there is a tree without believing that he is having a treeish experience. Since traditional
foundationalism denies that this is possible, traditional foundationalism is false.

Feldman’s argument makes an assumption that will be the focus of our attention. Before
mentioning this assumption, we need to review a couple important epistemological distinctions.
First, we can have justification for a proposition even if we do not actually believe that
proposition. In such cases, we say that we have propositional justification for that proposition.
When our belief is justified, we will say that we have doxastic justification for that belief.4 The
distinction is necessary because, as I mentioned, we can have justification for propositions about
which we have no beliefs. More importantly, philosophers cite the distinction because we can
possess justification for a proposition in situations in which, even though we also believe that
proposition, we do not justifiably believe it. That is, the belief may not be appropriately based on
or grounded in the states that constitute our justification for it.5 The concept of propositional
justification is prior to that of doxastic justification. Second, we should also distinguish between
inferential and noninferential justification. Justification is inferential when it is at least partially
constituted by justification one has for believing some other proposition6 and noninferential
when it is not constituted—even partially—by justification one has for any other proposition.7
The concept of noninferential justification is prior to that of inferential justification.

Feldman assumes that foundationalist theories are to be distinguished according to whether they
affirm or deny that justification for external world propositions must be constituted by justified
experiential beliefs where the denial is taken to be equivalent to the claim that justification for
external world propositions is noninferential. It is this assumption that allows Feldman to
conclude that if traditional foundationalism is false, then our external world beliefs may be
noninferentially justified. One meets this assumption so often in discussions of foundationalism
that we might call it the standard categorization of foundationalism. According to this standard
categorization, in traditional foundationalism any justification we have for external world beliefs
is always at least partially constituted by justified experiential beliefs. Non-traditional
foundationalism would be said to hold that such justification is noninferential because perceptual
experience is epistemically sufficient for justification.8

In what follows, I dispute this assumption by showing that the standard categorization is
incomplete. This is because the reasons we have for taking either external world or experiential
propositions to be foundational are logically separate from the reasons we have for taking either
external world or experiential beliefs to be foundational. Consequently, there is logical space for
a theory in which these issues diverge. If the case can be made for this separation, then simply
supposing (as Feldman does) that we may have justified external world beliefs in the absence of
experiential beliefs does not thereby show that our justification is noninferential. As a
consequence, a more penetrating taxonomy classifies theories according to whether or not
justification of external world propositions may be noninferential. Inferential foundationalism
denies this while noninferential foundationalism endorses this possibility. Traditional
foundationalism, then, is just one version of inferential foundationalism. There is room for a
theory that is non-traditional but also inferential. On this theory, justification for external world
beliefs is constituted by our justification for experiential propositions even when we have no
corresponding experiential beliefs. Call this theory nondoxastic inferential foundationalism. It is
‘nondoxastic’ because it denies the justified experiential belief requirement.9 It is ‘inferential’
because it denies that we can have noninferential justification for external world propositions or
noninferentially justified external world beliefs. Inference in the sense that I have given it here
refers to whether the justification that S must have for believing a proposition must be
constituted by the justification S has for another belief and not in its ordinary sense in which we
might be said to infer consciously one proposition or statement from another.10 We might call the
former notion constitutive inference and the latter procedural inference. This should not cause
any concern since epistemologists have long divorced the notion of inference from any
connection with explicit, conscious step-by-step inferring activity—even if such activity serves
as an idealization.

If I am correct in suggesting that the standard categorization is mistaken, then nondoxastic

inferentialism is at least conceivable.11 Despite the commonality shared by inferentialist theories,
we still need some way of marking the distinction between endorsement and rejection of the
justified experiential belief requirement. I retain the traditional/non-traditional label to mark this
distinction.12

I also have a deeper motivation for establishing the plausibility of the nondoxastic inferentialist
theory aside from simply pointing out this largely ignored possibility.13 Simply put, the demand
of the justified experiential belief requirement together with the likelihood that we do not have
the requisite experiential beliefs appears to lead to what we might call reflective ignorance. We
should be cautious of appeals to introspection: it is not always reliable guide to what beliefs we
have, so it cannot definitively determine that we lack experiential beliefs (as a routine cognitive
practice). Nonetheless, the results of a cursory inspection of my own mental life are not
encouraging: I do not discover experiential beliefs in the numbers and variety that traditional
foundationalism requires for justification of external world beliefs. To be sure, if someone were
to ask me whether or not I believe that I perceive the table in front of me (not just that I believe
that there is a table in front of me), then I would certainly answer that I do. But that response
fails to differentiate between several equally plausible alternatives and the mere ease with which
I can answer the question does not reveal that I believed this all along. It may well be that, in this
particular case, I do have the belief in question and all that has happened is a realization of a
disposition to become aware of it when questioned. However, there are at least two other
possibilities that do not presuppose the presence of a belief. We can say either that I simply have
a mere disposition to believe or that I have the mere capacity to believe given presence of
circumstances that trigger its formation. If the justified experiential belief requirement is true,
then we likely fail to know what we ordinarily take ourselves to know. While we should not let
pre-critical conceptions of what we take ourselves to know dictate the acceptability of an
epistemological theory, it would be a dialectical advantage of a theory if it allowed that (a) we do
have justification for believing much of what we pre-critically take ourselves to know (b) in a
way that accurately reflects what we can easily discover to be the case concerning our routine
cognitive practices. At the same time, while I do not argue for this claim here, I am convinced
that perceptual experiences by themselves cannot provide one with even prima facie justification
for external world propositions. Thus, I think that the noninferential theory is inadequate. The
nondoxastic inferential theory is a way of sidestepping the unpalatable practical consequences of
traditional inferential foundationalism and the epistemological inadequacy of noninferential
foundationalism.14

Feldman’s paper will be the focus of the first half of this paper. If I am correct in suggesting that
there is available a version of inferentialism that is not committed to the justified experiential
belief requirement, then although noninferentialism would gain some degree of support from any
successful argument against this requirement, it does not yet demonstrate that the noninferential
theory actually prevails. To be fair, Feldman explicitly says that he is not offering an argument
for noninferentialism merely by attempting to show that traditional foundationalism is untenable.
However, he does take the dialectical advantage to shift toward noninferentialism. I hope to
show that nondoxastic inferentialism is in a superior position since it is able to retain some of
what makes traditional foundationalism inviting to so many while omitting those components of
it that worry Feldman and other proponents of noninferentialism.

2 First criticism: Feldman’s argument is not decisive

In this section, I evaluate Feldman’s argument against traditional inferential foundationalism. In

the next section, I show how nondoxastic inferentialism is obscured in Feldman’s very
characterization of the difference between the traditional and nontraditional theories.

First, as I suggested, introspection is not a reliable means by which to determine whether or not
we have experiential beliefs. For that matter, I do not find myself forming beliefs of the sort
Feldman himself suggests, such as “I see something red.” Rather, the sorts of beliefs I find
myself forming when looking around are beliefs like “there is a red thing” and “that thing there
is red.” But, as I said, none of this is very conclusive. Assuming, as most philosophers do, that
beliefs are dispositional states of minds, they are not easily available for cursory introspection.
(Of course, they are not mere dispositions: a conscious assenting may be the realization of a
disposition or at least the acquisition of one. However, such realizations are the exception rather
than the norm for most of our beliefs.) It is a virtue of Feldman’s argument that he does not
invoke these sorts of phenomenological considerations. Others, though, have cited the apparent
absence of experiential beliefs (as a routine component of human cognition) as the basis for a
criticism of traditional foundationalism.15 If the claim made by these philosophers is only
supposed to be that we do not normally, in our unreflective daily activity, consciously formulate
experiential beliefs, then traditional foundationalists would hardly dispute this claim. The
traditional foundationalists might then offer the following point in connection with this
agreement. The justified experiential belief requirement is a philosophical claim and not an
empirical claim about what most people do most of the time; rather, it is a reconstruction or
idealization, a claim about what it would take to have justification for our beliefs about the
external world. That is, traditional foundationalists may never have taken this requirement to be
descriptive of the everyday cognitive life of all or most people.16 This is not (yet) to say that
philosophical skepticism—the claim that we cannot know what we take ourselves to know about
the external world—is the outcome. Consequently, they can simply conclude (if it turns out to be
true that we do not ordinarily form experiential beliefs) that most people do not have justification
for (and consequently knowledge about) the many things they generally take themselves to
know. This is what I earlier dubbed reflective ignorance (what we might also call practical
skepticism).

As I mentioned, Feldman does not appeal to the surface phenomenology of our conscious life.
Nonetheless, the tactic he relies on is an appeal to philosophical intuition. There is nothing
inherently unacceptable about this strategy; indeed, I will use the same approach later in this
paper. However, its effectiveness is quite limited. Proponents of traditional foundationalism are
likely to reply as follows: to suppose that the absence of all experiential beliefs is consistent with
the possession of ordinary knowledge of or justified beliefs about the external world is already
to require that traditional foundationalism be false. When Feldman says that cognitive science
cannot discover that we lack knowledge even while discovering that we lack experiential beliefs,
traditional foundationalists will reply that, of course, qua scientist, one cannot discover the
absence of knowledge in examining the brain, but neither could one discover the presence of
knowledge in that same manner. Without importing epistemological intuitions—ones perhaps
not shared by traditional foundationalists—about whether or not the particular details of a
thought experiment are relevant to ascertaining its epistemic status, no hypothetical discovery
could show either that John knows or that John fails to know that there is a tree.

On the other hand, Feldman grants that, in other kinds of circumstances, cognitive scientists
could in fact show that we do not know by showing, say, that the causal route from perception to
belief is fantastically deviant (2004a, 144–145). Thus, on the quite plausible intuition—one that
traditional foundationalists will likely share—that deviant causal connections of this sort rule out
these as cases of knowledge, we can conclude that we do not know when these sorts of
circumstances obtain. But even here the results are driven by intuitions. It is just that in this latter
case, the intuitions are more widely shared among philosophers in the same way that
philosophers almost universally agree that Gettier-style counterexamples show the traditional
analysis of knowledge to be inadequate.

One final issue: Feldman takes it as given that we do have knowledge of the external world and
that proponents of the traditional view agree in this assumption. Call this the dogmatic
assumption. With this assumption, Feldman rules out tout court certain versions of philosophical
skepticism. While I am sympathetic with nonskeptical epistemological theories, such theories
should be the upshot of philosophical labor and not assumed at the outset. While Feldman’s
comparison of the two versions of foundationalism with respect to how they fare given the
dogmatic assumption is fair enough within a very limited dialectical context, it does not, I think,
respect the proper order of assessment. Foundationalist theories are best seen as competing
claims about the structure of justification, so they should be assessed independently of
substantive claims about the content of justification. That is, the comparative plausibility of
specific versions of foundationalism should be considered independently of what we take
ourselves to know.17

I will now set aside these concerns with Feldman’s argument from cognitive science. My
concern henceforward is to consider whether or not Feldman’s claim that we can have
noninferential justification for external world propositions follows at all from the supposed
demise of traditional foundationalism. I argue that it does not given the availability of
nondoxastic inferentialism. I will also explain how the nondoxastic inferentialist view is
obscured by Feldman’s characterization of non-traditional foundationalism.

3 Second criticism: Feldman’s taxonomic division is not exhaustive

Before looking at Feldman’s characterization of the distinction between traditional and non-
traditional foundationalism, I want to introduce two epistemic principles that I will try to
substantiate throughout the rest of the paper.

First, propositional justification we have for believing p can be constituted by propositional

justification we have for believing q. John’s justification for believing that there is a red object
there can be constituted (in part) by John’s justification for believing that he is being appeared to
redly. Call this the weak transfer principle. I will develop this more shortly, but otherwise I take
this principle to be fairly uncontroversial.18

The second principle, however, will be controversial: our justified belief that p may be
constituted by our justification for q even when we do not believe q. That is, we can justifiably
believe p when we properly base it on our justification for q.19 John’s belief that there is a red
object there is justified on the basis of his justification for the proposition that he is being
appeared to redly. Call this the strong transfer principle. This raises a very important question
about what constitute the relata of the basing relation. I will return to this important issue in the
final section.

These principles are incomplete as stated. We must include propositions concerning the
evidential connection between p and q.20 There are three ways to go here. First, the principles can
simply include the requirement that there be an evidential connection between p and q. Second,
the principles can include the requirement that S have justification for some proposition
describing the evidential connection. Third, the principle can include the requirement that S
justifiably believes that there is an evidential connection. A theory might make evidential
connections constitutive of justification in any one of these three ways by incorporating them
into the strong and weak transfer principles. Noninferentialism will eschew these principles
entirely.21

Now we move to Feldman’s discussion. Compare these three claims (2004a, 134–135):

(A) If John is appeared to redly, then the proposition that he is appeared to redly is justified for
him,

(B) If the proposition that John is appeared to redly is justified for him, then, provided that he has
no defeaters, the proposition that he sees something red is justified for him.

(C) If John is appeared to redly, then, provided he has no defeaters, the proposition that he sees
something red is justified for him.

According to Feldman, while traditional foundationalism is committed to (A), it is only

mistakenly thought to be committed to (B). This is a mistake (he claims) because (A) and (B)
together entail (C). But (C) is also endorsed by non-traditional foundationalism, a view that is
supposed to be at odds with the traditional view on this very issue. Thus, traditional
foundationalism appears to be committed to (C) as well. Further, the non-traditional view has no
problem accepting (A) and (B). Feldman (135) reasons that non-traditional foundationalists have
no qualms about (A) and (B) because (i) experiences can also justify experiential propositions
and (ii) experiential propositions are justified whenever external world propositions are justified.
Thus, the non-traditional view “need not deny that propositions about the external world are
justified for a person only if propositions about internal states are also justified for that person”
(141, italics added). In light of this, Feldman suggests traditional and non-traditional
foundationalism—which are supposed to be mutually inconsistent positions—end up looking
indistinguishable. I will return to this important point shortly in order to examine Feldman’s use
of “only if” here. However, I first want to note a correction required for Feldman’s attempted
solution to this apparently inadequate way of distinguishing traditional and non-traditional
foundationalism.

Feldman suggests that this indistinguishability is only apparent because traditional

foundationalists should not accept (B) as it has been stated. Traditional foundationalism,
Feldman suggests, is most commonly understood as the claim that one must “‘go through’
beliefs about experience in order to get justification for external world propositions (or beliefs)”
because one “needs additional evidence . . . that comes . . . from forming the belief that one is
having an experience of red” (136, italics added). Modifying (B) brings out this defining
characteristic of the traditional view:

(B∗) If John is appeared to redly and believes on this basis that he is appeared to redly, then,
provided he has no defeaters, the proposition that he sees something red is justified for him.
(136)

However, even this is not enough to get Feldman what he wants. Feldman seeks a pair of claims
that will provide us with a way of distinguishing traditional from nontraditional foundationalism,
claims that make them inconsistent with each other. It was in accordance with such a purpose
that (B) was deemed inadequate. However, by stating (B*) with “if” rather than “if and only if,”
Feldman still fails to identify such a claim. All that he achieves is making traditional
foundationalism inconsistent with (C) on the basis of its commitment to (A) and (B*). He does
not succeed in generating any inconsistency between non-traditional foundationalism and either
(A) or (B*). Take conditionals generally: p implies q entails that p and any other proposition r
implies q. This is what we have with (C) and (B*). (C) is of the form p⊃q. (B*) is of the form (p
· r)⊃q. When (B*) is taken to be a necessary condition for justification, then (C) is insufficient
because it is too weak. However, when (C) is taken to be sufficient, then since it is a weaker
requirement, it remains consistent with (B*). As a result, non-traditional foundationalism can
affirm everything that traditional foundationalism affirms. Consequently, although Feldman
identifies a claim distinctive of non-traditional foundationalism—claim (C)—he has not
succeeded in identifying a claim distinctive of traditional foundationalism. But if we reformulate
(B*) to be a claim about the only way that we can have justification for believing propositions
about the external world, we will have succeeded in rendering these views mutually
inconsistent:22

(B∗∗) If and only if John is appeared to redly and believes on this basis that he is appeared to
redly, then, provided he has no defeaters, the proposition that he sees something red is justified
for him.

(B**) properly states the justified experiential belief requirement. Now since (C) does not entail
the conjunction of (A) and (B**), non-traditional foundationalists are no longer able to affirm
everything that traditional foundationalism affirms and Feldman can now distinguish the two
views.

However, I want to suggest that an unmodified rendering of (A)–(C) can be used to distinguish
inferentialism and noninferentialism. Returning to Feldman’s contention that non-traditional
foundationalists have no trouble accepting (A) and (B) because they “need not deny that
propositions about the external world are justified for a person only if propositions about internal
states are also justified for that person.” But as I have suggested, Feldman takes non-traditional
foundationalism to be equivalent to noninferentialism when, in fact, non-traditional
foundationalism is also consistent with the view that I have called nondoxastic inferentialism. So
while Feldman is correct in saying that the nondoxastic inferentialist and the noninferentialist
views both agree that John has justification for believing that there is something red only if John
has justification for believing that he is appeared to redly, this agreement depends on
equivocation in the sense of “only if.” We can get at this equivocation by looking at Feldman’s
response to a suggestion by McGrew (1999) that arguments like the one Feldman presents
confuse “psychological priority with epistemic priority:”

One possibility [that McGrew has in mind] is that psychological priority has to do with the order
in which things occur in the mind, while epistemic priority has to do with epistemic dependence
relations. One might hold that the justification of propositions about the external world depends
upon the justification of propositions about internal states, yet deny that one must actually form
beliefs about those internal states. As noted earlier, however, defenders of nontraditional
foundationalism need not deny that propositions about the external world are justified for a
person only if propositions about internal states are also justified for that person. If this is what
epistemic priority amounts to, then it is unclear that it brings out any point of disagreement
between the two kinds of foundationalism. (141)

Note that McGrew seems to be suggesting that justification for external world propositions is
epistemically constituted by justification for experiential propositions. Feldman, however,
reformulates his statement with the more imprecise claim that external world propositions are
justified “only if” experiential propositions are justified: his use of “only if” muddles an
important distinction. Note that Feldman’s weaker claim is entailed by McGrew’s stronger
claim; however, the former does not entail the latter and this is what is problematic in Feldman’s
weaker restatement of the stronger claim.23

There are two ways that we might read Feldman’s use of “only if:” a constitutive reading and an
entailment reading. Consider these three pairs of propositions:

(r1) I am appeared to redly.

(r2) There is a red object there.
(b1) The beagle sitting here has fewer than six legs.
(b2) The beagle sitting here has fewer than five legs.
(g1) The gas gauge reads “empty.”
(g2) The gas tank is empty.24

For each pair of propositions, it is true in some sense that we have justification for the second
member of each pair—(r2), (b2), and (g2)—only if we have justification for the first member of
each pair—(r1), (b1), and (g1), respectively. If we take “only if” as invoking epistemic
dependence (which is what McGrew seems to have had in mind), then—contrary to Feldman’s
suggestion—noninferentialists will deny that John has justification for believing that something
is red “only if” he has justification for believing that he is appeared to redly.
It is important to keep in mind that this discussion about an epistemic dependence of our
justification for believing the second member of each pair on the corresponding justification for
the first member of each pair occurs in the context in which our justification for the latter is in
some way the only justification we have for the former. Further, for (r1) it is important that the
justification for this proposition itself be derived from perceptual experience. We may very well
have other ways of obtaining justification for the second member of each pair that does not
require justification for the first member of each pair. Without this stipulation, then it is false that
we have justification for believing that there is a red object there only if we have justification for
believing that we are being appeared to redly. There can be situations where we are appeared to
bluely but where we have justification for believing that a pill we just took makes all red objects
look blue and all blue objects look red. Consequently, in this case we have justification for
believing that there is a red object there but we do not therefore have justification for believing
that we are appeared to redly. Likewise, it may be dark and I cannot see the gas gauge, but Joan,
who is a trustworthy source of information who is wearing her night vision goggles, tells me that
the gauge reads “empty.”

Noninferentialists will model their understanding of the epistemic connection between (r1) and
(r2) on that which arises between (b1) and (b2). While proponents of the noninferential view can
accept that Kelly, when observing a four-legged beagle sitting next to her, has justification for
believing (b2) only if she has justification for believing (b1), they can go on to point out that her
justification for (b2) need not be constituted by any justification she has for (b1): justification for
both propositions rests on a common evidential base that justifies both or neither.25 Likewise, the
noninferential view could accept such a connection between justification for (r1) and (r2). Upon
being appeared to redly, John would be said to have justification for believing that something is
red only if he also has justification for believing that he is appeared to redly, but only because
both have a common justificatory grounding in the experience.26 But this is not a relation of
epistemic dependence between (r1) and (r2).27

On the other hand, if we try to understand Feldman’s “only if” claim in a way that
accommodates the noninferentialist mere entailment conception of the connection between (r1)
and (r2) (which is what Feldman seems to have in mind), we end up ignoring the epistemic
dependence that nondoxastic inferentialists say obtains between our justification for (r1) and
(r2). They would certainly agree with noninferentialists that the epistemic connection between
(b1) and (b2) is merely counterfactual, but they would go on to argue that this is not the right
way to conceive of the connection between (r1) and (r2). Rather, nondoxastic inferentialists will
model their understanding of the epistemic connection that obtains between our justification for
(r1) and our justification for (r2) on that which obtains between our justification for (g1) and our
justification for (g2). Upon viewing the gas gauge, Mike has justification for believing (g2) only
if he has justification for believing (g1) where the justification for the former is partially
constituted by (depends on, has its source in, derives from) the justification for the latter.
Likewise, John has justification for (r2) only if he has justification for (r1) and the justification
for the former is constituted by justification for the latter. So while proponents of nondoxastic
inferentialism may affirm (C), it is because they suppose that justification transfers from John’s
being appeared to redly, through John’s justification for the proposition that he is appeared to
thus, to his justification for believing that something is red.28
Earlier I mentioned that inferentialist theories might also include propositions describing
evidential connections between experiential propositions and external world propositions in the
structure of justification. So now one might suppose on behalf of Feldman that by including
reference to an evidential proposition in the preceding discussion on we can make sense of the
difference between nondoxastic inferentialism and noninferentialism exclusively in terms of the
entailment sense of “only if.” However, I want to suggest that even here we need to distinguish
the entailment and constitutive sense of the phrase. It is still possible that there be mere
counterfactual dependence of our justification for external world propositions on our justification
for believing that evidential connections obtain between experiential and external world
propositions. This would not yet be the constitutive epistemic dependence invoked by
nondoxastic inferentialism. Such a modified mere entailment claim would need to make
reference to John’s having justification for

(r3) the proposition that one is appeared to redly (r1) evidentially supports the proposition that
something is red (r2).

However, to return to our previous example, while we can certainly say that Kelly has
justification for believing (b2) “only if” she has justification for believing both (b1) and

(b3) the proposition that x has fewer than six legs evidentially supports the proposition that x has
fewer than five legs,

there need be no constitutive dependence of her justification for (b2) on her justification for the
conjunction of (b1) and (b3).29 Here is why. Kelly may acquire justification for all three
propositions on the same basis: perception of the four-legged beagle. Just because certain logical
relations obtain among the group of propositions for which one has justification, it does not
follow that the justification one has for them maps onto that logical structure. One might have
acquired justification for each of the members of the group in some other way. This is what I
suppose may be the case with (b1)–(b3) and what the noninferentialist may suppose to be
possible with (r1)–(r3). The conceptual sophistication needed to grasp basic numerical and
anatomical concepts (five legs, six legs) is arguably sufficient to grasp comparison relations
between them (four is fewer than five and six). Even though no beliefs asserting (b1) or (b2) are
being invoked by either the noninferential or the nondoxastic foundationalists, some degree of
conceptual sophistication is needed to have justification for such propositions. It is not
necessarily that the subject must have the conceptual capacity to formulate in thought all
propositions for which he has justification; rather, the capacity to formulate related thoughts or
employ neighboring concepts in recognizing objects may be enough.30 Thus we might be able to
suggest that we cannot have the conceptual sophistication to have justification for p and for q
unless, as a matter of contingent psychological fact, we are also able to grasp that q evidentially
supports p—yet all this can be true without its being the case that the justification we have for p
is in any way constituted by the justification we have for q and the evidential connection between
p and q. Grounds independent of the logical relations obtaining between p and q may constitute
our justification for p. So Kelly, upon observing the four-legged hound, may not even be able to
have justification for believing that the beagle has fewer than five legs unless she is also has
justification for believing both that the beagle has fewer than six legs and that the proposition
that x has fewer than six legs evidentially supports the proposition that x has fewer than five legs.
(Keep in mind that (b1) and (b2) are more sophisticated than the simpler proposition that ‘that
beagle has thus (indicated by ostension) number of legs.’) Observing the beagle puts her in an
epistemic position to have justification for all three propositions.

Likewise, upon being appeared to redly, John may not even be able to have justification for
believing that there is something red there unless he also has justification for believing both that
he is being appeared to redly and that propositions describing red appearing states evidentially
support propositions describing red objects in the vicinity. Merely observing the red object puts
him in an epistemic position to have justification for all three propositions. Now, the parallel
between (b3) and (r3) is not quite exact since I want to allow that we can have justification for
propositions about basic sensible properties even if we do not possess such concepts prior to the
experience. This can be taken care of by noticing that it is our experiences themselves that make
possible our conceptual grasp of experiences and their objects. Being appeared to redly may be
sufficient to acquire the concepts of ‘red,’ ‘red object,’ ‘appeared-to-redly,’ and by virtue of that,
grasp evidential connections between propositions describing red appearances and propositions
describing red things in the world.

Summing up, we see that on a constitutive reading of “only if,” the noninferentialist cannot
accept that we have justification for external world propositions only if we have justification for
experiential propositions. For the noninferentialist, justification for external world beliefs is
constituted solely by perceptual experiences. On the other hand, a mere entailment reading fails
to do justice to the way in which the nondoxastic view understands the connection between
having justification for experiential propositions and justification for external world propositions.
For the nondoxastic view, justification for external world beliefs is not simply counterfactually
dependent on justification for experiential propositions: it is partially constituted by it.
Consequently, although the nondoxastic and noninferential versions each accept (A), (B), and
(C), the epistemic structure of the relationships among the propositions are different. So we do
not need to import (B**) and the justified experiential belief requirement to generate a
distinction between inferential and noninferential versions of foundationalism.

4 Defending nondoxastic inferential foundationalism

Recall that merely having a belief for which one has justification is not sufficient to make it a
justified belief: the belief must be based properly on its justificatory grounds. At this point, most
philosophers suggest a causal notion of basing to underwrite the basing connection.31 Typically,
epistemologists will say that S’s believing p is justified on the basis of a further belief q only if
(i) S’s believing q is itself justified and (ii) S’s believing p is based on S’s believing q (and
perhaps S’s believing that q evidentially supports p), where the basing relation is understood (at
least partially) as a causal relation. It should be apparent by now that nondoxastic inferentialism
cannot accept a causal conception of basing—at least not without some serious modification—
since it permits justified external world beliefs where this justification is constituted by (based
on) justification for experiential propositions alone—that is, no experiential beliefs.32

Our options, then, are (a) developing a modified causal basing account or (b) offering an entirely
non-causal basing account. I develop option (b) below. I am aware that the suggestions that
follow will likely fail to persuade anyone who not at least somewhat amenable to a non-causal
conception of the basing relation. What follows is merely a sketch and not a full defense of the
view, so it will not convert those convinced that it is incontestable that doxastic justification
minimally requires some causal connection between the justificatory grounds and the belief in
question. But I want to reiterate that this is not the only way open to the proponent of
nondoxastic inferentialism. We could instead follow option (a). Here are a few suggestions for
doing that. First, we could say that the external world belief must be caused in the right way by
the very appearance state that is the subject matter of the experiential proposition for which one
has justification (but no belief).33 Second, we could say that it must be caused by the ground of
one’s capacity to entertain and form beliefs about appearances.34 Third, we could say that it must
be caused by the ground of one’s disposition to believe the experiential proposition.35

Returning now to the noncausal basing account, the thrust of the theory is this: what must be
added to the believing p when one has justification for p to make that belief justified is mere
introspective awareness of that belief —which I am not conceiving of as a second-order belief or
a belief-like state of any sort. This may appear to be a rather thin requirement, but once I
articulate what I think is required to constitute propositional justification for such beliefs in the
first place, one might be more inclined to grant that nothing further is needed for doxastic
justification.

To start with, consider our propositional and doxastic justification for experiential propositions
and beliefs. The motivation for the nondoxastic inferentialist theory has been to offer an
alternative to traditional foundationalism and its justified experiential belief requirement. As I
suggested, we have little to no evidence for supposing that we have experiential beliefs in the
number and complexity required by traditional foundationalism. However, I never ruled out our
having such beliefs at least on occasion and it will be helpful to see how such beliefs may be
justified if we omit a causal basing requirement. My suggestion is that noninferential justification
for the proposition that one is appeared to redly is constituted solely by direct awareness of the
fact that one is appeared to thus. Second, when one justifiably believes noninferentially that one
is appeared to redly, this justification is constituted by direct awareness of the fact that one is
appeared to redly, and of the belief that one is so appeared to, and of the correspondence relation
between them.36 In short, nondoxastic inferentialism holds that a non-deviant causal connection
between the belief and its justifying grounds is conceptually unnecessary. The reason is this:
given that one is directly aware of the experiential fact, of the belief about that fact, and of the
correspondence between them, any possibility of causal deviance seems irrelevant.37 The fact
that these three relations of direct awareness obtain in a unified way in consciousness is
sufficient for a justified belief. (It is important to remember that awareness is being thought of
here as a relational state that holds between subject and object and not an introspective belief, a
second-order belief, or a representational state of any sort.)

Turning to our central concern with external world propositions and beliefs, proponents of
nondoxastic inferentialism must decide whether the proper view requires (a) justification for
propositions describing evidential connections between experiential and external world
propositions or (b) its merely being the case that there be an evidential connection. I will
concentrate on versions opting for (a). Justification for the proposition that there is something red
there will be constituted by (1) direct acquaintance with the fact that one is appeared to redly and
(2) direct acquaintance with the probability relation between propositions describing perceptual
experiences and propositions describing the external world. This is an application of the weak
transfer principle.

When an external world belief that there is a red object is justified, this doxastic justification is
constituted by the factors involved in justification for the proposition coupled with an awareness
of the belief state itself. This is an occurrence of the strong transfer principle. Again, there is no
appeal to a causal connection between the justifying grounds and the external world belief. As
with noninferentially justified beliefs, given that one is directly aware of the fact that one is
appeared to redly and directly aware of the evidential connection between the relevant
propositions, any bizarre causal source generating or sustaining the external world belief
(however improbable) is epistemically inconsequential.38

While many philosophers may accept the weak transfer principle, I doubt that the strong transfer
principle will be received in the same manner given the widespread commitment to a causal
notion of epistemic basing: we cannot have an inferentially justified belief unless it is causally
connected to the justifying belief(s).39 However, if we can be dissuaded of our adherence to a
causal notion of basing, there may be fewer objections to the strong transfer principle. In the
remainder of the paper, I sketch a positive case for its acceptance (and, so, for nondoxastic
inferentialism). However, to reiterate my earlier disclaimer, the argument that follows is unlikely
to dislodge hardened proponents of the causal conception of basing, so I repeat my suggestion
that they consider instead the modified causal basing account mentioned earlier. If that modified
causal basing account is plausible, then the strong transfer principle can be vindicated on those
terms.

Call this the epistemically disparate cases argument.40 We might suggest that there is a real
epistemic difference between the following four cases, in each of which John is appeared to redly
and is directly aware of his being appeared to redly:

• The mere true belief case: He is appeared to redly, is directly aware of this appearing state, and
believes that there is a red object in front of him.

• The mere false belief case: He is appeared to redly, is directly aware of this appearing state, but
believes that there is a blue object in front of him (though he is not appeared to bluely at all).

• The color fixation case: He is appeared to redly, is directly aware of this appearing state, is
directly aware of the evidential connection between red experience propositions and red object
propositions, and he even believes that there is a red object in front of him—but he does so
because he has fond memories of objects with this color and unconsciously forms this belief
whenever he thinks about it.

• The justifying case: He is appeared to redly, is directly aware of this appearing state, is directly
aware of the evidential connection between red experience propositions and red object
propositions, and he believes that there is a red object in front of him.

Proponents of nondoxastic inferentialism suggest that in the justifying case John is doxastically
justified in his belief that something is red and that contrasting this case with the other cases
shows this. The difference between the mere true belief case and the justifying case illustrates the
contrast between nondoxastic inferentialism and noninferentialism. Nondoxastic inferentialists
argue that John’s belief that something is red is epistemologically better off in the justifying case
than it is in the mere true belief case. It is not just that John has justification for the proposition
that something is red in the justifying case but not in the mere true belief case: the contrast being
drawn is between actual beliefs and their respective relations to the evidence in each case. If one
is willing to grant this observation—even though no mention has been made either of a belief
that John is appeared to redly or of the causal source of the belief about the red object—then
proponents of nondoxastic inferentialism may take this to be a concession that the causal source
of beliefs is not constitutive of our being justified in holding them. This is because the epistemic
difference will be explainable only in terms of the belief being justified in the justifying case and
not in the mere true belief case.41 By contrast, (i) proponents of traditional inferential
foundationalism will say that the epistemic standing of the beliefs in both cases are the same
because they are both unjustified while (ii) noninferential foundationalists will say that they are
the same because they are both justified.42 When we compare the justifying case and the mere
false belief case, the proponent of nondoxastic inferentialism will suggest that it is evident that
the epistemic credentials of the belief that something is red in the justifying case outweigh that of
the belief that something is blue in the mere false belief case and this difference is best explained
in terms of the belief in former case being justified.

One might object at this point that we cannot say whether or not the belief in the justifying case
is justified because it is under described in just that respect in which we require more detail: the
absence or presence of a non-deviant causal connection between the belief and some or all of the
justificatory grounds. Without filling in that missing piece of information, we cannot judge
whether or not the belief is justified, so we cannot say whether or not the belief is epistemically
better off in that case than are the beliefs in either the mere true belief case or the mere false
belief case. However, this is just to fault the argument at the very point at which it aims to ignore
the requirement of the causal basing account.

Now compare the justifying case and the color fixation case. Any claim that they are
epistemically disparate will likely rest on the difference in the causal credentials of the beliefs in
each case as they have been described. But, the justifying case intentionally contains no mention
one way or the other of the causal connection between the belief and its justificatory grounds.
However, since causal relations were not an issue in comparing the justifying case with either the
mere true belief case or the mere false belief case, why should it now become an issue with the
color fixation case? Compare the color fixation case with both the mere true belief case and mere
false belief case. To be sure, there is something amiss with the subject in the color fixation case,
but perhaps we can account for this with an independent notion of rationality. This notion of
rationality, as distinct from that of justification, might involve an assessment of the cognitive
functioning of the subject where this minimally involves something like assessing the degree and
strength of correspondence between the structure of justification for the beliefs one has and the
structure of the causal relations among those same beliefs. Whatever we say about the subject in
the color fixation case, we cannot compare the causal source of the beliefs in the justifying case
and the color fixation case because the causal story in justifying case has been suppressed.
However, since we did not require causal details when comparing the justifying case with either
the mere true belief case or the mere false belief case, this should indicate that it was not
essential to our assessment of the epistemic standing of the beliefs in those comparisons. Further,
the same accounting can be given of the absence of beliefs about one’s being appeared to redly.
No mention has been made of these in our comparison of these situations.

Nondoxastic inferentialism occupies a logical middle ground between traditional and

noninferential foundationalism. It is dialectically attractive because its splits the difference
between those two theories. It parts ways with traditional foundationalism in holding that
justification for external world propositions need not be constituted by justified experiential
beliefs. It parts ways with noninferentialism in holding that justification for experiential
propositions partially constitutes our justification for external world beliefs. Consequently, even
if Feldman shows that the justified experiential belief requirement is untenable, it does not
follow that we have noninferential justification for propositions about the external world.

One final note: in some cases, noninferentialism and nondoxastic inferentialismwill coincide in
their determination of which beliefs are justified. In fact, for each external world belief that is
justified according nondoxastic inferentialism, that same belief will also be justified according to
noninferentialism. But the converse is not true: some external world beliefs will fail to be
justified according to nondoxastic inferentialism (namely, those where we lack awareness of the
appropriate evidential connection) but will be justified according to noninferentialism.
Nondoxastic inferentialism is more demanding in this respect.

Acknowledgments An earlier version of this paper was read at the 2006 Pacific APA meeting. I
would like to thank the commentator, Stephen Blatti, and several of the audience members. Their
comments and objections spurred me to think more about some of the issues raised in the paper. I
would like to express my appreciation to Richard Fumerton for many hours of conversation
about the topics surrounding this paper and for reading and commenting on several earlier drafts.
Evan Fales and Diane Jeske also provided helpful feedback on an earlier version. Kelly Hobson
patiently read several versions of the paper and corrected me in several places. Finally, I would
like to thank two reviewers for Synthese who provided extremely helpful comments and
suggestions.

1 It is not easy to provide an uncontroversial distinction between experiential and external world
beliefs. By ‘experiential’ I do not have in mind only beliefs that refer explicitly to the subject and
the experiential state (“I am being appeared to bluely”). I intend the term to apply more loosely
to include statements that refer to how putatively physical objects and events in the world appear
or look to be (where ‘looks’ and ‘appears’ statements are taken as non-comparative
phenomenological descriptions and not tentative assertions). Thus, concerning the blue wall
when viewed in normal conditions, while we ordinarily believe that the wall is blue and we can
also believe that the wall appears blue or that there is a blue appearance in my visual field. If one
accepts certain versions of direct realism, then the blue in these cases will be a property of the
wall itself. Nonetheless, the senses of the statements just given emphasize the experiential or
presentational aspect of that property.

2 Some traditional foundationalists have also endorsed a further claim that external world beliefs
are typically acquired through and sustained by experiential beliefs. This psychological claim has
not always been carefully distinguished from the justified experiential belief requirement.
3 Or, alternatively, it might show that even if we do form experiential beliefs, these beliefs are
not psychologically antecedent to or the causal basis of the formation of external world beliefs.
This would be the case because either (a) the experiential beliefs are not psychologically
antecedent to the external beliefs or (b) the external world beliefs are not ordinarily based on the
experiential beliefs.

4 The terms “propositional” and “doxastic” are Feldman’s and I will adopt them throughout.
Feldman also refers to those beliefs for which we have doxastic justification as “well-founded”
beliefs. Robert Audi (1993, 274–296) goes further and distinguishes four senses of justification.
In addition to the two kinds mentioned, he also distinguishes two others that he calls structural
justification and what he calls “propositional” justification. The latter differs from the sense in
which Feldman (and I) use the term. Audi uses instead the phrase situational justification to
denote our ‘propositional justification.’ Let us call Audi’s sense of “propositional” justification
public justification, instead, for that is more in line with Audi’s meaning: rather than being
justification for believing a proposition from within a subject’s perspective, it is the sense in
which a proposition is justified for someone in that evidence is available for discovery if they
were to seek it out where this availability has nothing to do with anything within the subject’s
cognitive system. It is “public” evidence, so to speak. Structural justification concerns the sense
in which evidential grounds are available in the subject’s cognitive structure in the form of states
and dispositions that are not properly integrated into the structure. He writes of structural
justification: “I in some sense have a justification, but it is buried in my cognitive inventory, not
registered in premises I believe, or displayed in my perceptual consciousness, or in any other
way ready to ground my believing the proposition.” Again, “I have the makings of a justification,
but they are unintegrated and so do not amount to a justification as they stand in my cognitive
system: roughly, in my body of beliefs, dispositions to believe, inferential tendencies, memory,
and consciousness” (277). There is some overlap between Audi’s conception of structural
justification and two principles that I will introduce further in the paper.

5 The situations typically cited include (a) believing p on the basis of non-justifying grounds, (b)
believing p as a result of deviant causal connections between it and the justifying grounds, or (c)
believing p in situations where there is no causal connection with the justifying grounds at all.
Situations like (b) and (c) play a prominent role in discussions of causal conceptions of epistemic
basing relations; however, in developing my preferred view, I will suggest abandoning this
(widely accepted) causal notion of basing altogether. But whatever conception of basing is
correct, justifiably believing is believing that p appropriately on the basis of justifying grounds.
We also need the distinction to make sense of situations where we are attempting to determine
whether we have justification for believing that p or not. Since we have yet to make up our
minds, it does not make sense to talk about the epistemic status of an actual psychological state.
Nor is it advisable to attempt to state it in terms of a counterfactual such as what I would believe
were certain circumstances to obtain. On a counterfactual analysis, in the situation where I am
standing at a door and, unknown to me an elephant is on the other side, then I would now have
justification for believing that there is an elephant on the other side of this door because, were I
to have opened the door, I would have formed the quite justified belief that there is an elephant
on the other side.
6 The “at least partially” qualification is added because of the possibility that a proposition can
have inferential justification where this justification is constituted, not only by (ultimately)
noninferential justification for other propositions, but also by relations of coherence and
consistency between this proposition and other propositions for which we have inferential
justification. I will ignore the qualification through the remainder of the paper.

7 We can have noninferential justification for a proposition even if we actually believe that
proposition on the basis of other justified beliefs and, consequently, the justification upon which
the belief is actually based is inferential. So, if we lacked the other beliefs upon which we
actually base the belief, we would still possess noninferential justification for that same belief.
Also, having noninferential justification does not preclude the possibility that in order to have
noninferential justification for believing p, we must satisfy one or more of the following
contingent causal conditions involving a belief in a further proposition q: (a) believe q, (b) have
justification for believing q, or (c) justifiably believe q. With each of these possibilities, the
obtaining condition would merely be a causal precondition for possessing justification for
believing p. So long as the justification S has for believing p is not constituted by any of the
justification S has for q, then S’s justification for the p is noninferential.

8 This is not to claim that perceptual experience is psychologically sufficient because

presumably further cognitive capacities and beliefs are causally necessary for grasping external
world propositions.

9 My use of ‘nondoxastic’ should not be confused with the way John Pollock (2001) employs the
term. His sense of the term is broader and includes both traditional inferentialism as well as
noninferentialism as I have described them here. Pollock thinks that these theories exhaust the
field of nondoxastic theories in his sense of the term, so his discussion follows the script of the
standard categorization story. For him, the contrasting “doxastic” view is one that makes beliefs
the foundation of justification where these beliefs are ones for which we have justification that is
not constituted by any other state—experiential, doxastic, or otherwise.

10 George Pappas (1982) states it like this: “Perhaps we have gone wrong [with earlier
explications of non-inferential knowledge] by taking the term ‘inferential,’ and thus the term
‘non-inferential,’ too literally. What is meant by the claim that some knowledge is inferential,
one might argue, is that it is based on some evidence. And, non-inferential knowledge would
then be that knowledge that is not based on any evidence . . .” (87).

11 I am taking for granted that one can have justification for a proposition without needing to
have the capacity to believe that proposition. That is, one can have justification for believing p
even when one does not even have the conceptual sophistication to formulate the thought that p.
However, none of this is meant to rule out the quite plausible constraint that some level of
general conceptual sophistication is necessary for having justification for a proposition. Even if
we need not have the conceptual capacity to formulate in thought the propositional content of p
in order to have justification for p, there maybe some more minimal threshold of conceptual
sophistication that we need to satisfy, some other related concepts that are needed. But this is all
a question of causally necessary conditions and does not concern the constitutive basis of
justification.
12 There are various ways of making the distinction among versions of foundationalism,
between those that often go under the banner of ‘traditional,’ ‘classical,’ ‘strict,’ or ‘strong’ and
those that often go under the banner of ‘moderate,’ ‘modest,’ ‘nontraditional’ or ‘weak.’ I use the
contrast between ‘inferential’ and ‘noninferential’ versions to highlight the question at issue in
Feldman’s argument and in my response concerning whether or not we must have justification
for propositions about our sensations. A second issue that often overlaps this one concerns
whether or not noninferential justification for basic empirical propositions—whether experiential
propositions or external world propositions—must be infallible or indefeasible or certain. The
two issues are closely related though since one reason often given for thinking that noninferential
justification must concern experiential propositions is that they are the only kinds of empirical
propositions that are even possible candidates for infallibility. On the other hand, fallibilist or
moderate foundationalists often cite as a virtue of their view the fact that, by allowing that we
can have noninferential justification for propositions that are not infallible, we can admit external
world propositions as candidates for foundational level propositions. See McGrew (1999) for a
discussion that combines the inferential–noninferential issue with the infallibility-fallibility issue.
The traditional foundationalism of concern in this paper is not driven toward the justified
experiential belief requirement because of a fixation on infallibility and absolute certainty of
noninferential justification; rather, it is motivated by the search for more substantial justificatory
grounds for external world propositions beyond the mere having of perceptual experience. See
also Fumerton (1995, 2001a), Bonjour (1999, 2001a), Pryor (2000), and Pollock (2001) for more
discussion.

13 This theory is not as novel as I originally supposed. Robert Audi (1993, 282–283) and David
Chalmers (2003) each suggest it as an option. However, neither attempts any defense of it.

14 I have purposely left the concept of ‘belief’ unanalyzed here. It cannot simply be analyzed as
being disposed to assent to the proposition in the appropriate circumstances. By itself, this fails
to differentiate believing from merely being disposed to believe. This latter state can be realized
in the formation of a belief, an event that could involve an act of conscious assent. This makes it
difficult to differentiate from the manifestation of a dispositional belief in an occurrent act of
assenting or affirming. Nonetheless, I think that there are plausible grounds for maintaining this
distinction, so any analysis of belief that conflates this distinction is prima facie problematic. See
Robert Audi (1982) for the argument that simple affirmation or assent to a proposition does not
imply a pre-existing belief. He mentions there the distinction between belief as a disposition and
a mere disposition to believe. This distinction is further explicated in his (1994). Further, more
work needs to be done to distinguish dispositions to believe from having a mere capacity to
believe. We do not have a disposition to believe that there is an elephant outside the room simply
because it is true that were we to leave the room and encounter one, we would form the belief
that there is an elephant there. In this case, we clearly have a capacity to believe that an elephant
is outside the room, but since we lack any “adequate psychological basis for it,” it would not be
correct to say that we have a disposition to believe that there is one (Robert Audi, 1994, 426).

However, having made this distinction, I realize that even with it in place, there is still room for
disagreement about what constitutes having a belief and this has relevance to the
phenomenological worry that I raise in the paper. Some philosophers may be quite liberal
concerning what must be satisfied in order to say that S believes that p while others may endorse
a more restrictive criterion. This difference may even complicate the taxonomy that I have
suggested in this paper and that is because it may separate traditional foundationalists in a way
that somewhat blurs the distinctions I have made. For example, Timothy McGrew
(1999, 229–230) is willing to accept that (what he calls) a “referentially formed belief” about
perceptual experience can be construed so that it “is separable from and does not entail the
existence of explicit verbalized judgment.” On such a view “[h]aving an experience at all . . .
might come out to be equivalent to having a tacit referential belief.” Contrast this with
Fumerton’s suggestion that we do probably lack beliefs about our sensory experiences—at least
as a general rule (2001b; 2004a). This appears to involve a more restrictive notion of belief.
Note, however, that Fumerton’s conception of believing is at least permissive enough to allow
for dispositional background beliefs that guide the formation of external world beliefs on the
basis of our sensory experiences (1998). Further, he grants that certain types of regress
arguments for foundationalism are inconclusive because he supposes that proponents of
Infinitism (such as Peter Klein) are correct in suggesting that we do have an infinite number of
certain kinds of iterative beliefs (2001a). This concession, too, exhibits a somewhat less
restrictive view of believing. I think that this latter phenomenon is better explained by supposing
that we have a limited number of quite general beliefs that allow for a disposition to form an
infinite number of such iterative beliefs. See Robert Audi (1982) for a defense of this claim.
Now, on a more permissive view of the sort suggested by McGrew, we do have large numbers of
tacit beliefs about our perceptual experiences, so the phenomenological worry that I raise in the
paper may not count against this sort of view as much as it would against a more robust view of
belief. We would not expect to find beliefs of these sorts in introspective examination.
Nonetheless, we still have no evidence that we do have them (at least in the numbers and variety
required). Having said that, the difference between (1) traditional inferential foundationalism,
when paired with a permissive view of what constitutes having a belief, and (2) nondoxastic
inferential foundationalism, when paired with a more restrictive view, may be a verbal dispute,
in the end. Both can agree, though, that we ordinarily do not have experiential beliefs of the more
robust kind. See Feldman (2004b), Fumerton (2004a), Fales (2004), and Robert Audi (1994) for
discussion of this issue and its bearing on epistemology.

15 James Pryor (2000, 536) and John Pollock (2001, 43–44, 55–56) (and with Iris Oved (2005,
311) have offered versions of such an appeal.

16 Evan Fales (2004) also makes this point. He warns against proponents of internalist and
externalist approaches to justification or warrant thinking of their accounts as “providing a
plausible account of the actual formation of our ordinary noetic structures, with a view toward
certifying ordinary knowledge claims” (378). Fales defends traditional foundationalism against
the charge that its idealized structure in any way is supposed to reflect actual human cognition,
insisting that it is, rather, an attempt to set forth “ways in which a set of beliefs could properly be
built up so that the result is a body of knowledge” (374). See also BonJour (2001b, 79–80).

17 I realize that these comments will hardly move Feldman or any other proponent of a
particularist approach to constructing an epistemological theory. The objection I have raised
here is essentially to criticize Feldman for operating with a particularist approach and not a
generalist approach. Commitment to one or the other approach is usually deep seated and not
likely to be shaken by my few comments in the text. Nonetheless, it is important to note that
these methodological assumptions underlie the argumentative strategy.

18 James Pryor (2000, 525), for one, affirms just such a principle. He gives the following
example: “In this example . . . it seems plausible to say that I have justification for believing that
I’m out of gas, even though I do not actually form this belief. What’s more, there is an obvious
sense in which my justification for believing that I’m out of gas rests on my justification for
believing that the gas gauge reads ‘E’—even though I do not actually form either belief, and so
do not base the one belief on the other” (525). If by “rests on” Pryor means “constituted by” or
“derives from” rather than either merely “a necessary logical condition for” or “causally
sustained by,” then his claim agrees with mine. Audi appears to deny this principle with regard to
propositional justification (what he calls situational justification). He appears to hold that S’s
justification for believing that p must be constituted by having a justified belief that q. However,
as I noted in endnote 4, Robert Audi (1993, 277–278) also acknowledges a sense of justification
that he calls structural justification, which approximates what I have in mind with the weak
transfer principle. While there may be some overlap between what he and I each have in mind
with these distinctions, the primary difference appears to be this. Audi wants to say that a
proposition p is merely justifiable for S in cases where S has structural justification for believing
p but that S does not have justification for believing p. I want to say that where the weak transfer
principle is operative, S has justification for believing p that is constituted by the justification S
has for believing q (that is, p is not merely justifiable for S). Feldman (2004b) seems to express
agreement with Audi.

19 Recall the discussion of the previous endnote: Audi also seems to entertain the strong transfer
principle with regard to structural justification.

20 See Fumerton (2004b) for more on epistemic probability. Feldman omits any discussion of
evidential connections between beliefs about being appeared to redly and beliefs about red things
in the world. This might provide an additional explanation for his failure to acknowledge views
between traditional inferential and non-traditional noninferential foundationalism beyond the
explanation that I suggest in the main body.

21 Pryor (2000) is clear about this. He writes, “The dogmatist about perceptual justification says
that when it perceptually seems to you as if p is the case, you have a kind of justification for
believing p that does not presuppose or rest on your justification for anything else, which could
be cited in an argument (even an ampliative argument) for p. To have this justification for
believing p, you need only have an experience that represents p as being the case. No further
awareness or reflection or background beliefs are required” (519). Of course Pryor’s statement is
ambiguous between further justification or awareness being necessary in the sense counterfactual
sense and its being necessary in the constitutive sense. I take it that Pryor has the latter in mind.
However, I am not so sure about Feldman’s appeal to Pryor. This is because Feldman seems to
articulate a more qualified view in the postscript to his essay with Earl Conee, “Evidentialism”
(2004, 107). They indicate that the principle that if one is appeared to redly then one is justified
in believing that there is something red there (provided one has no defeaters for that proposition)
is not true. I am unable to read this assertion in a way that is consistent with Feldman’s apparent
endorsement of a straightforward noninferentialism in the essay under consideration here.
22 This is a rendering of (B*) that was tacitly presupposed by Feldman is his argument. I am not
disputing that Feldman understood (B*) in the way that I have reformulated it as (B**). It is
merely that it was not formally sufficient as formulated.

23 To be fair to Feldman, he does acknowledge that McGrew and others may have in mind some
relation of epistemic dependence not captured by his (Feldman’s) weaker claim. He simply is not
sure what it would be and invites those who suppose that something like the stronger claim is in
play to say more about it. This paper is an attempt to answer his request.

24 The gas gauge and gas tank example are borrowed from Pryor (2000).

25 Or, at least, for both or for (b1) only. See the next footnote.

26 The noninferentialist can even say that the justification for the experiential proposition is
stronger and less fallible—that it is epistemically better situated—than the justification for the
external world proposition. This admission does not by itself make the justification for the
external world proposition constitutively dependent on the justification for the experiential
proposition. This parallels the case of the quantity of beagle legs. Kelly’s justification for the
proposition that the beagle has fewer than six legs may be epistemically better off than her
justification for the proposition that the beagle has fewer than five, but it does not follow from
this that the justification for the latter constitutively depends on the justification for the former.
(One may be more apt to err in noticing that something has fewer than five than noticing that it
has fewer than six. This difference is more apparent if we contrast it with the justification we
have for believing that it has fewer than five and the justification we have for believing that it has
fewer than 50. Again, the difference is less apparent if we contrast it with the comparative
justification we have for believing it has fewer than 49 and the justification we have for believing
that it has fewer than 50. When observing the four-legged beagle, ‘fewer than 50’ is safer than
‘fewer than 5’ and about as safe as ‘fewer than 49.’) Likewise, the justification John has for
believing that there is a red object there is arguably less safe than the justification he has for
believing that he is being appeared to redly but arguably safer than the justification he has for
believing that is the red table he was given by his grandmother last year.

27 Pappas (1982) makes a similar point. He distinguishes between (a) S’s justification for
believing that p being counterfactually dependent on his background knowledge or justified
beliefs as well as concept possession and (b) S’s justification for believing that p being
evidentially dependent on his background knowledge or justified beliefs as well as concept
possession. He writes: “If S’s justified belief that P is evidentially dependent on Q, then it is also
counterfactually dependent on Q; but the converse relation fails to hold. And it is evidential
dependence or the lack of it that is critical in any determination of whether some instance of
knowledge is non-inferential or not” (95). Pappas exploits this distinction to argue that
justification for beliefs about physical objects in optimal conditions is noninferential because in
such cases the justification we have for background beliefs is only counterfactually necessary
and not evidentially necessary. Consequently, he, too, falls into the noninferentialist camp.
28 This is what Pappas (1979) calls “inferential basing” as opposed to either “belief basing” or
“justification basing.” Inferential basing concerns the logical, probabilistic, or epistemic
connection between a proposition p that S believes and either the evidence or reason S has.
Pappas uses “evidence” to mean propositions that S believes (justifiably believes, knows) while
he uses “reason” to mean non-doxastic states of S such as perceptual states. S’s belief that p is
inferentially based on reason states when propositions describing these reason states evidentially
support p.

29 It may appear odd to say that ‘having fewer than six legs’ evidentially supports ‘having fewer
than five legs’ since the converse is a stronger connection. But I need to state it like this to
maintain the parallel with the evidential connection between ‘being appeared to redly’ and ‘there
is a red object here.’ While ‘having fewer than five’ certainly entails ‘having fewer than six,’
‘having fewer than six’ does make more probable the truth of ‘having fewer than five.’ (Contrast
this with the more minimal degree of probability that ‘having fewer than five’ receives given the
truth of ‘having fewer than fifty.’) Likewise, while ‘a normal observer in normal conditions
seeing a red object’ certainly entails ‘being appeared to redly,’ ‘being appeared to redly’
evidentially supports ‘seeing a red object.’

30 See footnote 11. Nothing said here should be seen as conflicting as what I say in that endnote.

31 Robert Audi (1993) offers an extensive development and defense of the casual understanding
of the basing relation. See “Belief, Reason, and Inference” (233–273) and “The Causal Structure
of Indirect Justification” (214–232).

32 There is another logically possible theory lurking in here, one that sides with nondoxastic
inferentialism concerning the justification of external world propositions, but then suggests that
when it comes to having justified external world beliefs, we need justified experiential beliefs.
We might call this restricted nondoxastic inferentialism. I will ignore this possibility. Its main
motivation would be an inability of the unrestricted version I sketch in the paper to say just how
external world beliefs can be based on justification for experiential propositions unless we have
experiential beliefs in the picture as well.

33 Fumerton (2004a, 83) suggests a derivative conception of propositional justification similar

to this. The difference between his proposal and mine, of course, is that the strategy I suggest
involves a derivative conception of a causal basing relation and, consequently, of what it takes to
have a justified belief. The justification that one has for the proposition is not in any way
derivative. Fumerton stresses that it is not just any description of the appearance state that
legitimately can be brought in at this point because the description must be limited to properties
of the state that can be causally productive of the ensuing belief state: “We can take the
properties that are causally relevant to be those constitutive of the fact that is the truthmaker for
the relevant evidential proposition. My belief that there is a table in front of me is caused by a
visual experience that may have the property of being the kind of experience usually caused by
tables under these sorts of conditions, but reflection on standard epistemological problems
strongly suggests that it is only the non-relational intrinsic character of the experience that is
causally relevant to producing the belief. I’d believe precisely the same thing about the table if I
lived in a world in which demons typically produce hallucinatory experience. So on the account
I’m suggesting the only experiential proposition relevant to assessing the epistemic contribution
of the experience would be the proposition made true by the exemplification of the non-relational
(intrinsic) properties of the experiential state” (84).

34 People differ in their capacity for conceptualization of sensory experience, so this view would
lead to graded degrees of justification for external world beliefs depending on the extent to
which the capacity is developed and, perhaps, its propensity to causally produce or sustain
justified external world beliefs. Despite a painter’s keenly developed capacity to attend to subtle
details of a scene, she still may not generally form experiential beliefs upon which to base her
external world beliefs. Nonetheless, since her capacity to form experiential beliefs is more
developed than that of a non-painter, resulting in a stronger disposition to form justified beliefs
about her perceptual experiences, then we might say that she has more doxastic justification for
external world beliefs grounded in that capacity.

35 Robert Audi (1993, 282–283) mentions this possibility, but he does not elaborate. He writes,
“This is not to imply that a disposition to believe . . . a proposition that is adequate evidence for
p cannot itself justify one’s belief of p. I leave this difficult question open . . . If the disposition
did justify q . . . this would not . . . yield a case of inferential justification of the belief that p; one
would be disposed to infer p from q but, not having formed a belief that q, one would not believe
p on the basis of believing q.” However, we could say that one would believe p on the basis of a
disposition to believe q and count that as a kind of inferential justification. Note that Audi has in
mind here the possibility of a disposition to believe that q doxastically justifying the belief that p.
While he is noncommittal about that possibility, the context (the previous paragraph on 282 and
later on 286) makes it clear that he thinks that a disposition to believe that p can constitute
structural justification for believing the proposition that p—a conception that mentioned in an
earlier footnote that is quite similar to the conception of propositional and doxastic justification
that I develop in this paper.

36 This is essentially Fumerton’s proposal concerning noninferential propositional justification

for experiential beliefs (1995). He would add that for the belief to be justified, it must be causally
based on some or all of the justificatory grounds.

37 I assume here that a belief about some state of affairs A and direct awareness of A are distinct,
so it is conceivable that one can exist without the other. We can have false beliefs about being in
pain and even about being aware of being in pain. Likewise, we can be in pain and even be
aware our pain without believing either that we are in pain or that we are aware of our pain.
(What is not possible is being aware of our pain without being in pain: the awareness is a
complex state that includes the pain as a constituent.) These are certainly extreme possibilities
and not ones that normally arise in ordinary subjects nor, once generated, would survive long in
their noetic structures.

38 Compare Richard Foley (1987, 185). Though his discussion here includes beliefs concerning
all the relevant propositions and not just appearing states of which one is aware and which
correspond to propositions describing those states, I think that his suggestion works just as well
for the view that I am attempting to develop. He writes that “if a person who rationally believes
both p and if p then q and who rationally believes that modus ponens is truth preserving is
nonetheless caused to believe q in some unusual way (perhaps due to some recent neurological
malfunction about which he knows nothing), then it is no longer clear that his belief is irrational
in any epistemic sense.”

39 While the causal conception of the epistemic basing relation seems to be the dominant view
among epistemologists, it has had its detractors. Richard Foley (1987, 174–186), Keith Lehrer
(1971, 2000, 195–197) and Kvanvig (2003) are three proponents of the non-causal view.

40 This is the converse of Foley’s “epistemic twins” strategy (1987, 185–186). It also resembles
arguments suggested by Kvanvig (2003) and Lehrer (1971). Actually, only several of the
situations are epistemically disparate. The final two are “epistemic twins.”

41 I am assuming here that while the justification one has can come in degrees, justifiably
believing is not something that can come in degrees.

42 I set aside here the question as to whether or not proponents of noninferentialism will retain
some causal connection between perceptual experiences and external world beliefs.
To appear in Botany and Philosophy: Essays on Ivar Segelberg. Edited by Helge Malmgren. To
be published by Nya Doxa.
Segelberg on Unity and Complexity
Anna-Sofia Maurin
University of Gothenburg

… Bradley (and Segelberg) on the Problem of Unity in Complexity

In the beginning of Appearance and Reality (1908 [1893]), F. H. Bradley asks of a particular
lump of sugar what it means to say of it that it is white, hard and sweet. I will continue as before
and instead ask of our rectangle R what it means to say of it that it consists of three squares x, y,
and z. According to Bradley, there are only three things that this might possibly mean:

1. ...that R is identical with each of x, y, and z taken separately;

2. ...that R is identical with each of x, y, and z taken collectively, or;

3. ...that R is identical with each of x, y, and z related.

If one object could consist of its many constituents, it would therefore have to do so in one of the
above listed ways. But, Bradley argues, neither is a possible way in which one object could
consist of its many constituents. Therefore, complexity (of both the complex unity- and the
collection kind) is impossible.7

The reasons why Bradley does not think that any of the above listed answers to the question
‘what does it mean to say of R that it is/contains x, y, and z?’ is a possible answer, are the
following:8 the object, first of all (and obviously), cannot be identical with square x and then also
with square y (and then with z), unless of course squares x, y, and z are identical, which they ex
hypothesi are not. Nor, Bradley tells us, can the object be identical with each of x, y, and z taken
collectively for then, to use a Segelbergian term, the object would no longer be a complex unity,
different from a mere collection. We are left, therefore, to consider the option according to
which, when we say of R that it is (contains) x, y, and z, we mean by this that R is identical with
x, y, and z related. But what does this mean? Bradley, once more, surveys the field of
possibilities, and comes (on my interpretation) up with the following two options:

1. It means that x, y, and z are by their nature such that they are related to each other.

2. It means that x, y, and z are not by their nature such that they are related to each other, but that
there is a relation r which appears with them and which relates them, and “makes” them into R.

The first option is ruled out because Bradley believes that it, just like the option according to
which R is identical with each of x, y, and z taken collectively, fails to distinguish the mere
collection of x, y, and z, from those three squares united. This leaves Bradley with only one more
option: When we say of R that it is (contains) x, y, and z, we mean by this that x, y, z, and a
relation r, exist, and that r relates x, y and z. This, of course, is the collectionist view. The trouble
with this view is that it seems to lead us into a vicious infinite regress. In Bradley’s own words
(1908 [1893]: 21):

The relation C has been admitted different from A and B, and no longer is predicated of them.
Something, however, seems to be said of this relation C, and said, again, of A and B. And this
something is not to be the ascription of one to the other. If so, it would appear to be another
relation, D, in which C, on the one side, and, on the other side, A and B, stand. But such a
makeshift leads at once to the infinite process. The new relation D can be predicated in no way
of C, or of A and B; and hence we must have recourse to a fresh relation, E, which comes
between D and whatever we had before. But this must lead to another, F; and so on, indefinitely.

It is the Bradleyan regress which Segelberg alludes to when he states that: “A collection of x, y
and one or more relations z can never be the same as a unity xy. For, the collection x+y+z exists
whether x, y and z are ‘dispersed’ or form a unity, but the unity xy does not exist if x, y and z are
‘dispersed’” (1999 [1945]: 55). It is clear, therefore, that Segelberg accepts the Bradleyan
argument. He cannot, however, accept every aspect of it. Bradley’s argument condemns not only
the collectionist view; it condemns the very existence of complex unities. Segelberg, on the other
hand, is a firm believer in the existence, not only of complex unities, but of collections as well.
But, then, which part(s) of Bradley’s argument does Segelberg disagree with?

As we have seen, Segelberg accepts at least the following Bradleyan claims explicitly (and I will
assume in what follows that he also, albeit implicitly, accepts the claim that a complex unity
cannot be identified with each of that which constitutes it taken separately):

1. A complex unity ≠ its non-relational constituents plus a relation (reason: infinite regress)

2. If there are complex unities, then the collection which now constitutes a certain complex unity
could exist and not constitute that particular complex unity (DISTINCTION).9

But now things no longer seem to add up. If Segelberg accepts that a complex unity is not to be
identified with its constituents taken separately (on pain of contradiction), and if he also accepts
that it cannot be identified with its non-relational constituents plus a relation (because this leads
to a vicious infinite regress), then it would seem that only one option remains: the complex unity
is identical with its constituents taken collectively. But this option is also rejected by Segelberg.
A complex unity cannot be identified with the collection of its constituents because this would
contradict DISTINCTION. Segelberg seems to have run out of options.

Segelberg and States of Affairs

Segelberg, it seems, wants to eat his cake and have it too. On the one hand, he does not want to
argue that the difference between a complex unity and a collection is yet another entity. He does
not, after all, want to be a “collectionist” (1999 [1945]: 55):

The difference between...a unity xy and the corresponding collection x+y does not consist in the
presence of certain relational moments in xy. There is altogether no “difference” between xy and
x+y, if by ‘difference’ one means a content which belongs to one object but not the other.
On the other hand, he does not want to equate the complex unity with the collection of its
constituents. He does not want to give up DISTINCTION (ibid.):

On the other hand, there is, of course, a difference between them in the sense of an incongruence
relation.

But how can Segelberg, without contradicting himself, claim that although the collection and the
complex unity are different, there is nevertheless nothing that a complex unity contains that its
corresponding collection does not also contain, and vice versa? If there is a difference there must
be something that makes the difference (DIFFERENCE). Can Segelberg reconcile what appears
to be irreconcilable? Perhaps.

According to Segelberg, a complex unity and a collection, although not differently constituted,
are nevertheless different in that, as soon as R exists, not only does the collection x+y+z that R
contains exist, so does the following state of affairs:

S: That the squares x, y, and z lie next to each other.

Taking this state of affairs into account means that in the actual world (the world in which, ex
hypothesi, the rectangle exists) we can now distinguish between, not just two, but in fact three
different complexes:

1. The state of affairs S: that the squares x, y, and z lie next to each other (or, more generally: that
x, y, and z, form a unity);

2. The complex unity R: the rectangle consisting of squares x, y, and z, and;

3. The collection C: the squares x, y, and z

We can now see how R and C can be indistinguishable when it comes to that which constitutes
them, yet one – C – may exist, even if the other – R – does not. The difference between a world
which contains R (and, therefore C), and one which contains only C, is S. In a world which
contains R (and so C), S obtains. In a world which contains C, but not R, S does not obtain.
Therefore, although there is no content which belongs to the complex unity which does not also
belong to the collection (and vice versa) we can still say that the two are different. The
difference-maker is S, which is external to R. Surprisingly, therefore, it seems as if you can have
your cake and eat it too.

States of Affairs and the Bradley Problem

Not so fast. If the difference between the actual world, which contains the rectangle, and some
possible world, which contains only the collection, is S, we are obligated to ask (just as we once
did about the complex unity): How does S manage to turn what is a “mere” collection into a
complex unity? According to Segelberg, when S exists (or, obtains) there exists a higher order
complex object, consisting of x, y, z and (surprisingly) r. More precisely (1999 [1947]: 221):
A state of affairs is a complex constituted in a certain way. If one wants to discover the
characteristics of such a complex which distinguishes it from complexes of the first order, one
ought to compare a state of affairs with a first order complex that has, as far as is possible, the
same components as the state of affairs. We have made such a comparison earlier, when we
compared a rectangle H, consisting of two squares a and b, with a state of affairs S: the squares a
and b lie next to each other. We found one similarity to be that H and S are complexes which
have a and b as components. The crucial difference turned out to be that S contains a relation
while H does not.

And again (ibid):

The essential characteristic of an object of higher order (a state of affairs) seems to be just that it
contains one or more relations as components

So, the state of affairs S, just like R and C, consists of x, y, and z. With the help of S, we can now
distinguish between a world in which only C exists and one in which also R does (in the latter, S
obtains). But, clearly, we can only do this if we can distinguish between a world in which S
obtains and one in which it does not. If S is constituted by precisely the same entities as is C it is
(once more) unclear how a distinction can be made (and, so, it is unclear how Segelberg’s view
avoids both accepting and rejecting DISTINCTION). But, as we have seen, S is not constituted
by precisely the same entities as is C. The difference between a world in which the state of
affairs obtains and one where it does not is that in the former world, not only do the non-
relational constituents of the state of affairs (and of the collection and the complex unity) exist:
so does a (uniting) relation. But how is this not a “collectionist” answer? That is, how can
Segelberg say this and not end up in vicious infinite regress?

On External vs. Internal Infinite Regresses10

According to Segelberg, it is because it is a fact that x, y, and z lie next to each other (Rxyz) that
xyz is a complex unity and not a mere collection. But then, it seems, Segelberg must admit that it
is because it is a fact that a unifying relation relates the ‘lying next to each other’ relation to x, y,
and z (R′Rxyz) that it is a fact that x, y, and z lie next to each other. And so it does not seem as if
Segelberg’s account escapes the infinite regress. There is a difference, however, between this
regress, and the regress which is generated if you are a collectionist. Or so Segelberg wants to
argue. Rxyz and R′Rxyz (as well as the infinitely many further states of affairs that Segelberg’s
account commits him to) are distinct from one another, just as xyz is distinct from Rxyz. In fact,
although on Segelberg’s account, infinitely many entities (states of affairs) are regressively
produced, this does not mean that anything that is infinitely complex is thereby produced. Due to
an external regress, like the one Segelberg accepts, there is no infinite complexity in that which
generates the regress; rather, there is an infinity of “entities” (in this case, states of affairs), each
of which is itself finite. In the case of the collectionist’s regress the situation is the opposite: each
step of the regress adds a further entity to the original ones, with the result that there is more and
more entities, but no unity.
An infinite regress, Segelberg claims, can never be such that one object ends up with an infinity
of components. This would be the case if we agreed that the original (first-order) complex unity
contained some relation r (i.e., if we were collectionists). An infinity of entities of increasingly
higher order, however, is not likewise a problem. Or so Segelberg wants to claim. It is not
enough, however, to simply state that an external regress is unproblematic whereas an internal
regress is not. This is far from obvious. What reasons could Segelberg have for believing that an
infinite regress that commits one to an infinity of components in an object is a serious problem,
whereas a regress that commits one to an infinity of objects (distinct from one another) is not?
Unfortunately, Segelberg does not say much more on this topic. Fortunately, others do.
One contemporary “Segelbergian” is Francesco Orilia. Orilia is interested in what makes a state
of affairs an entity that exists over and above its constituents. His answer is that it is a further
state of affairs that accomplishes this feat (2006: 229):

What makes Fa an entity that exists over and above F and a is the state of affairs E2Fa,
understood as different from Fa, in that E2 [i.e. dyadic exemplification] is taken to be the really
attributive constituent of the former, whereas F is taken to be the really attributive constituents of
the latter.

Orilia is well aware that his proposal gives rise to an infinite regress of states of affairs (and he
therefore appropriately names his suggestion fact infinitism). But as the regress in question is
external he, just like Segelberg, takes it to be harmless.

Infinite Regress – Vicious or Benign?

To be able to judge whether Segelberg and Orilia have solved the problem of the unity of a
complex unity (or of a state of affairs in Orilia’s case), we must investigate further what
distinguishes an infinite regress that is vicious from one that is harmless or benign.11 We could,
of course, simply equate this distinction with that between internal and external regresses, and
hold that an internal regress is the same as a vicious regress, whereas it is always the case that an
external regress is harmless or benign. This would be putting the cart before the horse, however.
Whether every internal regress is vicious (and every external regress is benign), is something we
will have to wait and see until we have formulated a more substantial criterion by which to make
the relevant distinction.

As a first approximation, we can say that a regress is vicious if its existence somehow makes that
which triggers the regress (I will call this ‘the theory’) impossible, or at least improbable. In this
sense an infinite regress argument, featuring a vicious infinite regress as its main-premise, is a
reductio against the position from which the regress is generated. This may sound clear-cut, but
it is not. In order to decide whether or not an infinite regress contradicts the position from which
it is generated it is not enough simply to look at the regress, and then at the triggering theory, to
see if they match. A theory is the same as an answer to a question (albeit a very complex
question). It is a solution to a problem, an account of a particular phenomenon, an explanation. In
the case of an infinite regress, it is our theory (answer, solution, account, explanation) which
automatically and necessarily gives rise to an infinity. This infinity contradicts, or makes highly
implausible, the theory (answer, solution, account, explanation) from which it is produced if it
robs this theory of its status as an answer, solution, account, or explanation. To be able to see if a
particular regress in this sense contradicts the theory from which it is generated we must
therefore also know what question this theory is supposed to be an answer to, or, what
phenomenon it is supposed to explain. To complicate things further, to be able to judge whether
or not a particular infinity demotes the triggering theory from its status as an answer – if it robs it
of its ‘explanatory value’ – we must also decide what might constitute such a ‘demotion’. This
involves deciding what makes a theory explanatory, as well as what might turn an explanation
into a non-explanation. Big and difficult decisions indeed! Fortunately, for the purposes of this
paper, none of the really big decisions need to be made. Some smaller ones cannot be avoided,
however.

In the spirit of (relatively speaking) ‘small’ decisions, the first thing we need to decide is whether
it can never be the case that our triggering theory somehow depends, for its status as an answer,
on whatever comes next in the regress, if the regress is to count as benign. If this is never
acceptable, it follows immediately that the external regress, pace Segelberg and Orilia, is
vicious. It is after all because R(xyz) obtains that xyz is a complex unity. But this criterion of
viciousness seems much too strong and would be hard to justify. A more reasonable criterion is
instead one which requires that there must be some point at which the triggering theory obtains
its status as a full answer, or a complete explanation. That is, although it could be that the answer
our triggering theory pertains to be, depends (for its status as an answer) on what comes next in
the regress, it seems likely that it cannot be that its status as a full answer is never obtained
because it is infinitely deferred. In other words, it can be that every step in the regress requires
some further step to ‘make’ it a complete answer – but it probably cannot be that no step in the
regress in facts furnishes the relevant completion; completion can most likely not be allowed to
be infinitely deferred. If completion is infinitely deferred, therefore, the regress is most likely
vicious. This gives us the following (I believe relatively neutral and reasonable) account of what
distinguishes a regress that is vicious from one that is benign:

Vicious Regress

A regress is vicious if the answer which the theory from which it is generated purports to be, is
incomplete and requires completion in the next step of the regress, but where the answer
furnished in this step is also incomplete, and requires completion in the following step, etc. ad
infinitum

Benign Regress

A regress is benign if the answer the theory from which it is generated purports to be, is not
incomplete and so the existence of the regress does not demote the theory from its status as an
answer/explanation (although it does force it to give its answer/explanation with (great)
redundancy).

Do Explanations Have to Ground Out?

Now, what difference does it make if, instead of requiring that our answer in no way depends for
its status as an answer on what comes next in the regress, we require that, although it could
depend on what comes next, it cannot do so ad infinitum? In particular, does it make a difference
to how we judge the regress to which Segelberg and Orilia are willing to commit themselves?
No. On either account, the regress comes out as vicious.12 That this is so is clear at least to Orilia,
who in fact explicitly admits that an external regress is precisely one which defers explanation
ad infinitum:13

The externalist regress contradicts the thesis that an explanatory chain cannot go ad infinitum
without reaching a bottom line.

But this is no problem. For, according to Orilia, although the viciousness of a vicious infinite
regress does have something to do with explanation (a vicious regress certainly does not increase
the explanatory value of a theory), it has nothing to do with not being able to come up with a
complete explanation or a full answer. To the contrary, he tells us, in the case of the external
Bradley regress, it is precisely the regress’s postponing full explanation ad infinitum that counts
as an increase in explanation.14 But this means that the first question we need to ask ourselves in
order to be able to evaluate Segelberg’s (and so Orilia’s) answer to the problem of complex
unities is the following:

Q: Should we, or should we not, accept the assumption that explanation must ground out?

It should be obvious to all, first of all, that the assumption that explanation must ground out, and
so cannot be infinitely deferred (call it GROUNDING), is highly intuitive. Orilia is certainly
sensitive to this fact (2006: 232):

...intuitively it seems correct to say that we have an explanation for P only insofar as there is, so
to speak, an increase in our knowledge/understanding, when we contemplate P. But, one could
argue, if in an attempt to explain P I begin an explanatory task wherein at every stage I must
presuppose a succeeding stage, then there is no increase. For any such increase is an
approximation to the final stage and if there is no such stage, then there is no explanation. And
thus there cannot be infinite explanatory chains.

Intuitions may be misleading however, and only a little bit of reflection is required to convince
Orilia that GROUNDING is an assumption that should be given up (ibid.):

But in fact good motivations in favour of such chains can be offered. That at any given stage we
can continue the explanatory task does not show that no knowledge or no understanding is
provided at any stage. It merely shows that at no stage do we know/understand everything that
there is to know/understand about the explicandum which gives rise to the explanatory chain.
And noting that the explanandum in question gives rise to such an infinite chain may be
considered part of our understanding of it.

To argue for or against fundamental methodological principles is always a tricky thing. After all,
methodological principles are the framework against (and with the help of) which we measure
our theories. They are not, or are at least very rarely, explicitly introduced as substantial theses
up for evaluation in and of themselves. But even so, if asked to choose between two opposing (or
mutually excluding) methodological principles, there are some things we value more than other
things, reference to which can make our decision at least reasonable. Intuitiveness is of course
one such value, and it is clear that intuitiveness is on the side of keeping rather than rejecting
GROUNDING. Orilia is however right to point out that intuition is very often misleading and so
can certainly not be our only reason for making that choice. Another important value, of
relevance here, is however intelligibility. If faced with two options, where one makes sense,
whereas exactly what the alternative entails is far from clear, it seems as if we ought (pending
more information) to go for the comprehensible over the incomprehensible or the less
comprehensible option. That being said, it is hard to understand how “that the explanandum in
question gives rise to such an infinite chain may be considered part of our understanding of it.” If
this means anything at all, it means that Orilia measures explanatory value in terms of our
understanding. But surely, to find out that something has no explanation could also increase our
understanding of it. Explanation and understanding, although perhaps related to each other, are
therefore separate, and they should be treated as such. Moreover, if the external regress is
benign, it is hard to understand what that which triggers the regress really is. It is, after all, our
answer to that very question which is being infinitely deferred. For reasons of intelligibility,
therefore, we should prefer to keep GROUNDING over rejecting it. Whether there can be
explanation even where the explanatory chain never “grounds out”, and exactly what that means
for that matter, remains to be explained. And the burden of proof is on whoever wishes to reject a
perfectly comprehensible and deeply intuitive principle, not on the one who wishes to keep it.
Orilia is simply saying that there can be infinitary explanatory chains. This does not constitute
the requisite defence.15 We should therefore (pending arguments able to convince us otherwise)
accept GROUNDING as a matter of ‘methodological’ law. But then Segelberg as well as Orilia
fail to solve the problem at hand.

Taking Stock

Segelberg wants to claim that there are both collections and complex unities, and that the
difference between the two consists in the presence in worlds in which the complex unity (and
not just the collection) exists of an infinity of states of affairs of increasingly higher order.
Segelberg’s suggestion is interesting, but problematic. I have argued that it can succeed only if
we give up a fundamental and deeply intuitive assumption: GROUNDING. As it is not clear
what might replace this assumption, it is not clear in what sense Segelberg’s account actually
manages to solve the problem at hand. For this reason his suggestion should be rejected. Not
even Segelberg gets to eat his cake and have it too.

Acknowledgements

I would like to thank the participants (and especially the organizers) at the Philosophy and
Botany conference (September 2008) for helpful and inspiring suggestions. I also would like to
thank Francesco Orilia for reading an earlier version of this paper. Last but not least, I would like
to thank Johan Brännmark, for his help as well as his (non-viciously) infinite patience.

7 In Bradley’s own colourful terms: (1908 [1893]: 33) “We want to take reality as many, and to
take it as one, and to avoid contradiction. We want to divide it, or to take it, when we please, as
indivisible; to go as far as we desire in either of these directions, and to stop when that suits us.
And we succeed, but succeed merely by shutting the eye, which if left open would condemn us;
or by a perpetual oscillation and a shifting of the ground, so as to turn our back upon the aspect
we desire to ignore.”

8 A more thorough explication of Bradley’s reasoning can be found in Maurin 2012.

9 It is easy to see that DISTINCTION is very much the driving force behind Bradley’s
reasoning. It is because he accepts this assumption that he rejects the alternative according to
which the complex unity is (nothing but) its constituent collection, and the same assumption
leads to his rejection of the alternative according to which the constituents of a complex unity
are, by their nature, such that they are related. Evidence that Segelberg accepts DISTINCTION
can be found throughout his writings. It is, for instance, clear that he does so from this, by now
familiar, quote: “the collection x+y+z exists whether x, y and z are ‘dispersed’ or form a unity,
but the unity xy does not exist if x, y and z are ‘dispersed’.” (ibid.) And when he claims that:
“One can imagine several situations in which the collection x+y+z exists without the rectangle
xyz being present: (a) the squares x, y, and z are included in o without touching each other; (b)
the squares x, y, and z form a rectangle but the consecutive order between the parts is not xyz but
xzy or yxz.” (ibid.: 49). Or, again: “If the rectangle xyz contains x, y, z and an arbitrary object u,
disparate from these, one can always imagine a collection x+y+z+u, incongruent to xyz, in which
x, y, z, and u do not form a rectangle.” (ibid.: 50)

10 For more on the difference between internal and external infinite regresses, see Maurin
2011b.

11 I discuss the distinction between a vicious and a benign regress at length in my (2013).

12 That both external and internal regresses are vicious is argued in Hochberg (1999).

13 In connection with the 11th eidos meeting in Geneva on the topic of Bradley’s Regress
(December 2008), William F. Vallicella and Francesco Orilia engaged in an interesting and
clarifying discussion on the topic of the regress via Vallicella’s philosophically rich blog
(Maverick Philosopher). To follow the discussion, go to:
http://maverickphilosopher.typepad.com/maverick_philosopher/bradley_and_his_regress/
(henceforth: ‘the maverick blog’) The above quote comes from that discussion. For Vallicella’s
views on how to solve the Bradleyan problem cf. also his (2004).

14 Richard Gaskin is if possible even more positively inclined towards this sort of regress.
Gaskin is interested in how to unite the different parts of a proposition so as to produce
something that can carry a truth-value. He offers a solution to the problem where the unity of the
proposition is catered for by another proposition (the proposition that the parts of the original
proposition are united), and so ends up with an analogous external regress. When discussing the
viciousness of this regress, he says (1995: 176): “It follows that Bradley’s regress is, contrary to
the tradition, so far from being harmful that it is even the availability of that regress which
guarantees our ability to say anything at all. Bradley’s regress is the metaphysical ground of the
unity of the proposition. Reverting to the original terms of the problem /.../ we might say,
somewhat paradoxically, that what stops a proposition from being a ‘mere list’ is that it is an
infinite list (of the specified kind).” Cf. also his (2008).
15 Cameron (2008) also thinks that we should keep GROUNDING. His reason is however
different from mine. He argues (2008: 12): “I am denying that the intuition can be justified by
any more basic metaphysical principle, and so it is a mistake to attempt to justify it in any way. I
suggest trying instead to justify it by appeal to theoretical utility. If we seek to explain some
phenomenon, then other things being equal, it is better to give the same explanation to each
phenomenon than to give separate explanations of each phenomenon. A unified explanation of
phenomena is a theoretical benefit. This seems to provide some evidence for the intuition under
discussion. For if there is an infinitely descending chain of ontological dependence, then while
everything that needs a metaphysical explanation (a grounding for its existence) has one, there is
no explanation of everything that needs explaining.” I, however, do not agree that, in the
objectionable kind of infinity “everything that needs a metaphysical explanation (a grounding for
its existence) has one”. On the contrary, I believe that in the objectionable kind of infinite
regress, nothing is really explained (because for a real explanation, we need at some point to
have a full explanation). I also suspect that Cameron’s argument might be an instance of the kind
of fallacy which Russell once accused Copleston of committing when arguing for the existence
of God: that of conflating the properties of the parts, with those of the whole (Cf. Russell
(1948)).
FOUNDATIONALISM AND THE IDEA OF THE EMPIRICAL
CIARA FAIRLEY
University College London

… 2. The Epistemic Regress Argument

Now that we have some sense of what foundationalism is supposed to be we can discuss whether
there are any good arguments for that view. I said that I would be taking foundationalism to be a
view about justification that is motivated at least in part by the ‘epistemic regress argument’.
What is that argument and how does it motivate the position I’m calling foundationalism?4 While
there is broad agreement on the importance of this argument in motivating foundationalism there
is less consensus on the form it should take.5 In the literature several different arguments purport
to be the epistemic regress argument.

Sometimes people have in mind a dialectical regress in which subjects are invited to defend their
beliefs in the context of an argument. This is how BonJour presents the argument in the
following passage:

The most obvious, indeed perhaps the only obvious way to show that an empirical belief
is adequately justified (in the epistemic sense) is by producing a justificatory argument:
the belief that p is shown to be justified by citing some other (perhaps conjunctive)
empirical belief, the belief that Q, and pointing out that P is inferable in some acceptable
way from Q. Proposition Q, or the belief therein, is thus offered as a reason for accepting
proposition P…[But] for the belief that P to be genuinely justified by virtue of such a
justificatory argument, the belief that Q must itself already be justified in some fashion;
merely being inferable from an unsupported guess or hunch, for example, can confer no
genuine justification. Thus the putative inferential justification of one empirical belief
immediately raises the further issue of whether and how the premises of this inference are
justified…empirical knowledge is threatened with an infinite and apparently vicious
regress of epistemic justification. Each belief is justified only if an epistemically prior
belief is justified, and that epistemically prior belief is justified only if a still prior belief
is justified, and so on, with the apparent result, so long as each new justification is
inferential in character, that justification can never be completed indeed can never even
really get started – and hence that there is no empirical justification and no empirical
knowledge (BonJour 1985: 18-19).

The dialectical regress, as I will call it, is concerned with what it takes for subjects to show that
they are justified. You challenge me to defend my belief and in order to respond successfully to
that challenge I must adduce some considerations in its defence thereby showing you that my
belief is justified.6

This is not the only target of epistemic regress arguments. Other presentations of the argument
focus more on what it takes for subjects to be justified in believing what they do. They aim to
find out whether or not a subject is justified in a given belief and if so what makes it the case that
she is justified.7 This is how DePaul presents the argument in the following passage:
 According to this ancient argument, when we consider a belief that is justified and ask
how it is that the belief is justified, we are typically led to another belief that supports the
first. When we ask about the second belief, we may well be lead to a third. The third may
in turn lead to a fourth, and so on. But how long can things go for in this fashion? There
would seem to be only three possibilities: the chain of beliefs either goes on forever,
circles back upon itself, or stops. Finding the first two possibilities unacceptable,
Foundationalists opt for the third, holding that there are some beliefs that are justified, but
that are not justified by any further beliefs. The regress stops with such basic or
foundational beliefs, and any other beliefs that are justified must be supported by the
foundational basic beliefs (DePaul 2001: vii).

I’m going to call this the justification-making regress to distinguish it from the dialectical
regress.8 Very often, however, one finds elements of both in the context of a single presentation.
This is evident in the quotation from BonJour and it is also clear in the following passage from
Dancy:

Suppose that all justification is inferential. When we justify belief A by appeal to beliefs
B and C, we have not yet shown A to be justified. We have only shown that, it is justified
if B and C are. Justification by inference is conditional justification only; A’s justification
is conditional upon the justification of B and C. But if all inferential justification is
conditional in this sense, then nothing can be shown to be actually, non-conditionally
justified. For each belief whose justification we attempt, there will always be a further
belief upon whose justification that of the first depends, and since this regress is infinite
no belief will ever be more than conditionally justified…The regress argument therefore
drives us to suppose that there must be some justification which is non-inferential if we
are to avoid the sceptical consequence of admitting that no beliefs are ever actually
justified (Dancy 1985: 55-6).

On the face of it, Dancy confuses the two different things just distinguished. He starts with a
claim about what it takes to show that a belief is justified and concludes with a claim about
whether or not the belief in question is justified. This is an easy mistake to make, however, since
the word ‘justify’ is ambiguous and can be used to refer to both. Pryor calls attention to these
different uses of the verb in the following passage:

On the first construal, ‘justifying’ a belief in P is a matter of proving or showing the

belief to be just (or reasonable or credible). (Here we can include both arguments whose
conclusion is P, and arguments whose conclusion is that your belief in P is epistemically
appropriate, or is likely to be true.) By extension, we can also talk about things justifying
beliefs; in this extended sense, a thing counts as justifying a belief if it’s something
you’re in a position to use to prove or show your belief to be just…There’s also a second
way to construe the verb ‘justify’, which sees it as akin to the verbs ‘beautify’ and
‘electrify’. When a combination of light and colour beautifies a room, it’s not proving
that the room is beautiful; rather, it’s making the room beautiful. Similarly, on this
understanding, justifying a belief is a matter of making a belief just or reasonable, rather
than a matter of showing the belief to be just (Pryor 2005: 194).
We have seen that it is possible to initiate a regress argument using either notion. The important
point is that the regresses thereby initiated will differ. One will concern whether a given belief
has been shown to be justified; the other, whether a given belief is justified.9

Given that they differ, which do we have reason to prefer? When I talk about foundationalism as
a position that is motivated in part by the regress argument what I mean is that it’s a position
motivated by the justification-making version of that argument. This is the more fundamental
version of the argument for several reasons. It is the more prevalent, and we’ll see shortly that it
is by far the more plausible of the two arguments, but the most important reason is that it is only
the justification-making version of the argument which threatens us with a truly unacceptable
epistemological conclusion. It is only the justification-making regress which threatens to show
that none of our beliefs is justified. The dialectical regress, in contrast, only promises to show
that we cannot, or have not, shown that our beliefs are justified. The latter conclusion, while
somewhat counter-intuitive, is not one we must do everything we can to avoid.10

Things would be different if we had some reason to link these two notions of justification and tie
the conditions under which a belief can be justified with the conditions under which one has
shown that it is justified. On certain views of justification one cannot be justified unless one can
show that one is justified. Dancy’s mistake may not then be a mistake but a substantive claim
about the nature of justification. This is a view about justification with eloquent exponents. Thus
McDowell describes ‘the time honoured connection in our discourse between reasons the subject
has for believing as she does and reasons she can give for thinking that way’ and criticises
writers like Peacocke for having to ‘sever’ that connection (McDowell 1994:162-166).

There is clearly something to the picture of justification that McDowell recommends here. It is
certainly true that when we have reasons for our beliefs, we can very often give them. It is also
true that we normally expect other people to be able to give us their reasons (the very young
think that we should be able to do that indefinitely) and we frequently take the fact that others
can’t give us their reasons as a sign that they do not really have any. That is partly why, as
Austin pointed out, the question of why the subject believes as she does can be asked not just out
of “respectful curiosity”, but pointedly; her inability to answer can only reveal that she ought not
to have been so bold (Austin 1979:78). These are no doubt some of the reasons the dialectical
regress can strike us as plausible. When one has reasons, the normal expectation is that one will
be able to give them.

However, this is just an expectation and it is defeasible. At least, that is how I will be talking
about justification. When I talk about justification I am going to allow that a subject can have
reasons or justification for believing what she does without necessarily being able to show that
she has them, much as a person can be honest or funny without necessarily being able to defend
the claim that she is honest or funny when under attack. This seems the more natural usage and
the more plausible. At the most basic level talk of justification is a way of appraising someone
who is doing well in her thoughts; there is no reason to think that necessarily brings with it the
ability to show that you are justified.11 The latter requires that you recognise that you are justified
and have the ability to articulate and perhaps even defend the grounds upon which that is so. This
looks like a more sophisticated cognitive achievement. Even where we are able to defend the
claim that we are justified it is still important to distinguish between what it is that shows that we
are justified and what it is which makes it the case that we are justified. Not everything that plays
one role may be capable of playing the other. In order to show that you are justified you have to
adduce claims to that effect, which claims are expressive of your beliefs. Showing that you are
justified may therefore always involve appeal to beliefs. We will see shortly that it is crucial to
foundationalism that being justified does not.

None of this to deny that there is something gripping about, what we might call, the ‘internalism’
McDowell here expresses - the idea that when one is justified one’s justification ought to be
somehow ‘available’ to one. It is very plausible to think that for something to be your reason as
opposed to just a reason, or for it to be what makes you justified in believing, it has to accessible
to you: it has to be a basis upon which you can justifiably form a belief, not just a basis upon
which a belief could be so formed. But since it is not in virtue of your ability to articulate your
reasons as your reasons that they count as yours we can acknowledge what is right about this line
of thought without going as far as McDowell. They can be your reasons or what makes you
justified in believing what you do even if you cannot state that fact when pressed, much less
defend it under questioning.12 It is also plausible to think that reasons must possess a certain sort
of ‘evidence’ or perspicuity. They must be the sorts of things of which you are somehow aware:
if they do not consciously reveal the world as being a certain way, what leads you to believe it to
be one way rather than another? Again though that doesn’t warrant going as far as McDowell.
What reasons make manifest, first and foremost, is the layout of the world, not their own status
as reasons. It is the latter, however, that one would require for showing that one is justified.

As I understand it, then, justification is the sort of thing you can have without necessarily being
able to show that you have it however often the two may in fact accompany one another. And
when I talk about foundationalism as a position motivated in part by the epistemic regress
argument I mean it’s a position motivated by the justification-making version of that argument.

So how does that argument motivate what I am calling foundationalism? Recall that the
argument starts from reflection on cases of inferential justification – that is, cases in which one’s
justification derives from other things that one has justification for believing. This is how many
of our beliefs do seem to be justified. For instance, I may believe that Tony Blair will not win
another election because I believe his policy in Iraq has been so unpopular with the electorate.
That belief may be what makes it the case that I am justified in my belief about his electoral
prospects. Or I may believe that Socrates is mortal because I believe that Socrates is a man and
that all men are mortal. Some people call that ‘mediate justification’ since other beliefs, in this
case my beliefs about foreign policy and mortality, mediate my justification. I’m going to stick
with inferential though, since it is an important part of the reason why those further beliefs
justify me that there is an acceptable inference between them and the belief they are claimed to
justify. The fact that Blair’s policies have been unpopular with the electorate makes it likely that
he will not win: my beliefs thus stand in a relation of probabilification. In other cases the relation
will be one of implication: that Socrates is a man and all men are mortal implies that Socrates is
mortal.13

According to the epistemic regress argument where we have a belief that is justified we can ask
what makes it the case that belief is justified: we can ask why the subject believes as she does or
what justifies her in that belief. So in answer to the question what justifies my belief about
Blair’s electoral prospects we can appeal to my belief about his foreign policy. But what justifies
this belief? If I don’t have any reason to believe his policy has been unpopular then I won’t have
any reason to believe that he will not win. So what justifies me in the belief about his foreign
policy? Here again the answer may involve appeal to beliefs and what justifies these beliefs may
be yet further beliefs still. But how far can things carry on in this fashion?

On the face of it there seem to be only four ways in which the justificatory regress can pan out:

1. The regress ends with a belief that is not justified. While it is not justified it is still able
to justify other beliefs.

2. The regress goes on forever: the belief that p is justified by the belief that q, and the
belief that q is justified by the belief that r, and the belief that r is justified by….and so
on, ad infinitum.

3. The regress circles back upon itself: the belief that p is justified by the belief that q,
and the belief that q is justified by the belief that r, and the belief that r is justified by the
belief that p.

4. The regress ends with a belief that is non-inferentially justified: while it is justified, it
does not draw it’s justification from other justified beliefs.

If this is the choice with which the regress presents us, what would be a good response?
According to the position that I am calling foundationalism it is only if the fourth option is
correct, and some of our beliefs are non-inferentially justified, that any of our beliefs are justified
at all. This is how the regress argument motivates foundationalism. According to this argument
foundationalism is the only alternative to the view that none of our beliefs is justified. This ought
to make it irresistible to all but the most sceptically
minded.14

Now we know what the regress argument is supposed to be, we can ask whether it is any good.
The argument is an argument by elimination so the case in favour of foundationalism is only as
good as the case against scepticism and the other options (1-3). I am not going to defend the
rejection of scepticism, but I will now defend the rejection of the three alternatives. While they
all enjoy support in some quarters, a strong case can be made against each of them. The case
against them is strong insofar as it relies upon assumptions about justification that it is
overwhelmingly plausible to make. It may be possible to give them up, but why do so unless we
really have to?

The first option claims that the regress ends with an unjustified belief. While it is unjustified it is
still able to justify other beliefs. How can a belief that is not itself justified, justify other beliefs?
Many writers take that to be obviously impossible. Thus, Susan Haack simply states without
further ado:

If A believes that p on the basis of his belief that q, then he is not justified in believing
that p unless he is justified in believing that q (Haack 1993: 22). Haack is not alone.
The idea that beliefs must be justified in order to justify other beliefs is intuitive and it is
grounded in a picture of the way in which beliefs confer justification that makes a lot of broader
sense.

Take any central case in which one belief justifies another belief and it seems to do so in virtue
of inferential relations between the propositions believed. The reason why my belief about
Blair’s unpopular foreign policy justifies my belief that he will not win another election is that it
stands in a relation of probabilification to the latter. Similarly, the reason why my beliefs that
Socrates is a man and all men are mortal justifies me in believing that Socrates is mortal, is that
the former imply the latter. Inferential relations are therefore an important part of the story as far
as beliefs go. But it is obviously not enough for one belief to justify another that it merely stand
in inferential relations like these to it. That would make it far too easy to be justified since every
belief stands in an infinite number of such relations to all manner of other beliefs (including
itself).

Of course, such relations might explain why I believe certain things, given what else I believe.
They might in that sense make it rationally intelligible that I believe as I do. But they do not, by
themselves, give me any justification to believe those things, since they do not, by themselves,
give me any reason to suppose that things actually are as my beliefs represent them as being This
is what Laurence BonJour is getting at in the following passage. He writes:

For the belief that p to be genuinely justified by virtue of such a justificatory argument,
the belief that q must itself already be justified in some fashion; merely being inferable
from an unsupported guess or hunch, for example, can confer no genuine justification
(BonJour 1985: 18).

BonJour’s thought is that merely appealing to inferential relations will not do. This seems right;
we want justification to be a guide to how things actually are in the world. We want it to have a
connection with truth and mere inferential relations do not secure that. So we need a further
constraint and this is precisely the role played by the requiring the beliefs involved be justified.
By specifying that the ‘inputs’ to this potential inference be justified we plug the intuitive
justificatory gap that merely believing something leaves open.15 This gives us a picture of the
way in which beliefs confer justification according to which justificatory status is inherited.
Beliefs justify other beliefs to which they are suitably related and they do so by passing on their
own justification. This is a compelling picture and it explains why an unjustified belief cannot
make another belief justified. A belief cannot pass on justification it does not itself possess, just
as I cannot inherit your car if you do not yourself possess one.16 If so then the first option
according to which the regress ends with an unjustified belief is a non-starter.

Up until now I have talked rather loosely about the way in which beliefs confer justification.
Actually the claim being made here is a claim about inference and the conditions under which
the fact one’s beliefs stand in inferential relations to one another is capable of conferring
justification. I am claiming that inference, in this sense, is a conditional vehicle of justification; it
only confers justification where the input beliefs are already justified. Dancy makes a similar
point in the following passage:
 Inference is basically a matter of moving from premises to conclusion along an
acceptable path. If the premises are unjustified there will be no justification for the
conclusion - at least not by this inference (Dancy 1985: 55).

This makes inference an essentially dependent form of epistemic justification: its existence and
functioning as a source of justification depends upon the existence and functioning of some other
source of justification.17

This is why we need to distinguish between the conditions under which justification is
inferential, and the conditions under which it derives from beliefs. Beliefs may be capable of
furnishing us with a different sort of justification. In that case it may not be true that they must be
justified in order to confer justification.18 However, this justification will not be inferential
justification. This response will then no longer be a version of the first option; it will be a version
of the fourth. It is hard to imagine what that role might be and that makes it tempting to frame
the conclusion that I have drawn about inference as a conclusion about beliefs more generally
and the conditions under which they are capable of furnishing us with justification. Strictly
speaking, though, I have only argued for the claim about beliefs insofar as they confer
justification by standing in inferential relations to other beliefs. I claim that in those
circumstances, beliefs can only make other beliefs justified where they are themselves justified.

This is a very plausible idea and it rules out the first option according to which the regress ends
with an unjustified belief. Notice though that is not part of a general claim about justifiers; it is a
specific claim about inferentially justified beliefs. It says they can only confer justification where
they are themselves justified. For all that’s been said there may be other things which can make a
subject justified – other things which can confer justification upon a belief – about which it
doesn’t even make sense to wonder whether or not they are themselves justified. As we will that
possibility is central to foundationalism.

What about the second option? Is there anything wrong with supposing that the regress goes on
ad infinitum? Since we do have actually have an infinite number of beliefs supporting each and
every one of our beliefs, this response cannot be the only non-sceptical way of terminating the
regress.19 Assuming it is unacceptable to claim that none of our beliefs is justified, the possibility
of an infinite regress cannot show that foundationalism doesn’t accurately describe the actual
justificatory structure of our beliefs. This may be all the foundationalist needs, but I think we are
entitled to a stronger conclusion in any case since ‘infinitism’ does not succeed in articulating a
justificatory structure our beliefs could possibly enjoy (whether or not they actually enjoy it). Or
so I will now argue.

The reason is very simple and takes us back to the point made at the end of the previous
discussion in connection with the proposal that the regress ends with an unjustified belief. We
are now considering a response according to which the regress goes on ad infinitum with each
belief inferentially justified by some further belief. However. I have already argued that
inference is an essentially dependent source of justification: its existence and functioning as a
source of justification depends upon the existence and functioning of another source of
justification. Given that is so, it is not possible that inference could be the only source of
justification. It is not intelligible that it could be the only way in which justification is conferred
upon a subject’s beliefs as the present response envisages since it is not an autonomous source of
justification in that sense. If that is right, then the second option according to which the regress
goes on ad infinitum is no good either.20

What about the third option according to which the regress circles back upon itself? According to
this option what justifies my belief that it rained last night might be my belief that the grass is
wet and what justifies my belief that the grass is wet might be my belief that it rained last night.
Does this make sense? It looks like a case in which a belief effectively justifies itself and that
seems to defeat the whole point of requiring that beliefs be justified in the first place.

The obvious inadequacy of this response brings out even more clearly the underlying inadequacy
of inference as a source of justification and helps explain why we should think that it is a
fundamentally dependent form of epistemic justification. If inference were not a dependent
source of justification as I have claimed it is unclear why circularity of this sort wouldn’t be
acceptable. After all, one of the beliefs that every belief stands in inferential relations to is itself;
the belief that p stands in a relation of implication to the belief that p. So if inferential relations
between one’s beliefs were sufficient for justification (if inference could in that sense be one’s
sole source of justification) then one could be justified in believing anything whatsoever
provided only that one does believe it.

This is totally unacceptable: beliefs are not justified simply in virtue of being held and they do
not in that way justify themselves.21 Indeed, this misses the whole point of requiring that beliefs
be justified in the first place. The original idea was that there should be something which
functions as a reason why the belief is likely to be true and makes it something you ought to
believe, something over and above the mere fact you do believe it. Thinking of inference as a
dependent source of justification as I have done enables us to explain why circularity of this sort
is unacceptable. Given that is so the third option according to which the regress goes round in
such circles is also no good.

Having just said that I want to acknowledge that this is not the only way that people have
understood the suggestion that the regress ‘circles’ back upon itself. According to a position
known as coherentism the objection just raised goes wrong in assuming a ‘linear’ conception of
conception. According to the coherentist’s holistic alternative we are not to think of justification
being passed from one belief to the next, eventually landing up back with the belief with which
the series begun. Rather, justification is a property of an individual’s entire set of beliefs.
Specifically, it is that property the set enjoys when it’s individual members ‘cohere’ with one
another and it accrues to each individual belief in virtue of its membership of such a set of
beliefs.

This is a very implausible account of what justifies our beliefs, but I do not want to take issue
with it here. The important point is that this strategy effectively appeals to beliefs that are non-
inferentially justified in terminating the regress. In fact, it says that all of our beliefs are non-
inferentially justified, since it says that they are all justified by the fact they belong to a coherent
set of beliefs and that is not where justification comes from in standard cases of inferential
justification.
What I mean by this is that the relations the coherentist appeals to are not inferential relations of
the usual sort - they are not like the relation that Socrates is a man and all men are mortal stand
in to the belief that Socrates is mortal - they are far more extensive and encompass all of an
individual’s beliefs. There is also a difference in the sort of access the subject has to those
relations and the role they play in getting her to be justified. In the Socrates case, there’s a fairly
robust sense in which I believe that Socrates is mortal because I believe that he is a man and that
all men are mortal; the latter beliefs really are operative in getting me to believe as I do, and if I
had different beliefs on that front, I’d adjust my beliefs about his mortality accordingly. This is
quite unlike the holistic case, where it is doubtful whether I am aware of the relevant facts about
all my beliefs and their inter-relations (or could even easily become aware of them), and doubtful
these facts are in any way operative in getting me to believe as I do. This may be the basis on
which certain coherentist epistemologists form their beliefs, but it is clearly not the basis upon
which most of us do so.22

Finally, though, even the coherentist doesn’t just appeal to facts about the inferential relations
between our beliefs, however extensive we take the set to be. She also has a story to tell about
why the fact that one’s beliefs cohere makes them likely to be true, and hence why it should be a
justification that they provide. For instance, Davidson appeals to the fact that beliefs are by
nature veridical (Davidson 2000). This ought to remove any remaining temptation to call this a
case of inferential justification, in the ordinary sense. We certainly do not ordinarily appeal to
facts about the nature of beliefs in making sense of inferential justification. Coherentism is
therefore a version of the fourth strategy which I have associated with foundationalism, rather
than the third.

I am not claiming that coherentism is a particularly plausible version of the fourth strategy or that
there is nothing awkward about so characterising it. I am just claiming that overall it is better
seen as a version of the fourth strategy which appeals to a non-inferential source of justification,
than to one which maintains that all justification is inferential in the relevant sense. In the end,
though, it may be more accurate to say the coherentist simply rejects the framework of inferential
and non-inferential justification within which the regress is set. After all, that position is one
according to which all beliefs have the same source of justification; they are all justified by the
fact they are members of a coherent set. Ironically, it may turn out that the best way of bringing
out the difference between coherentism and more traditional versions of the fourth response is
not by stressing the idea that some justification is non-inferential (since that is something the
coherentist thinks is true of all beliefs) but by stressing the idea that some justification is actually
inferential (some of it really does derive from other things that we have justification for
believing). Intuitively, not all beliefs have the same source of justification and this is something
the coherentist denies.

So I have claimed that inferential justification is a fundamentally dependent form of epistemic

justification and I have rejected options (1-3) on that ground. However, I have also suggested
that subtle variations of those options ought to count as versions of the fourth strategy. It is now
time to consider this option – the option according to which the regress ends with a belief that is
non-inferentially justified. Would this be a good response to the regress?
On the face of it, yes it would. Unlike the other three options there is nothing immediately
implausible about the suggestion that the regress ends with a belief that is non-inferentially
justified.23 Some people deny that a belief can be justified by anything other than its inferential
relations to other justified beliefs. But there is nothing intuitive about this view. It rests entirely
upon philosophical arguments that we will later find wanting. Given that options (1-3) are no
good, and that there is nothing intuitively problematic with the fourth option, the regress
argument does look like a good argument in favour of thinking that there must be such a thing as
non-inferential justification.

But is it a good argument in favour of foundationalism? It is certainly not a good argument in

favour of foundationalism in the traditional sense. Nothing in this argument supports the demand
for a layer of non-inferentially justified or epistemically basic beliefs distinguished in terms of
their content or enjoying the sorts of strong epistemological privileges that Williams and Sosa
describe. All this argument supports is the claim that there must be some beliefs that are non-
inferentially justified; there must be some beliefs that do not draw their justification from their
inferential relations to other justified beliefs. It doesn’t tell us anything about the sorts of beliefs
that can be non-inferentially justified or what it is about them that enables them to play that role.
Any substantive claims on that score are totally unmotivated. The regress argument is therefore
not a good argument in favour of foundationalism as traditionally conceived.

Is it a good argument in favour of foundationalism in my sense? I have claimed that

foundationalism is a view about the structure of justification (a) and a view about its sources (b)
& (c). With respect to (a) I have claimed that foundationalism is a view about the structure of
justification that is motivated by the epistemic regress argument. The foundationalist claims that
we must acknowledge the existence of non-inferentially justified beliefs in order to respond
satisfactorily to that argument. This is precisely what we have just seen the regress does
establish, given that the other three alternatives are no good and that scepticism is false. So the
regress argument is a good argument in favour of (a). That is hardly surprising though since (a)
simply commits the foundationalist to whatever it is one needs to solve that problem. Once one
sees foundationalism as a response to the regress argument it is easy to see that my response is a
better response to that argument than the traditional view, since my view is actually motivated by
that argument. The traditional view, by contrast, commits itself to all sorts of things that just
aren’t relevant to solving that problem.

Of course someone might now say: why call that view ‘foundationalism’? In a way this is a good
question. As we have just seen the regress argument only gets you as far as thinking there must
be some non-inferential justification. This cannot be all there is to foundationalism; it is far too
permissive. We have seen that slightly modified versions of options (1-3) all appeal to non-
inferential sources of justification on one reading, yet it would be wrong to think of those
positions as forms of foundationalism in any serious sense. This is why (a) is only one
component of the position I call foundationalism. Foundationalism isn’t just a view about the
overall structure of justification. It is also a view about the sources of justification. This is the
point of (b) and (c). A foundationalist also thinks that perception is a distinctive and privileged
source of non-inferential justification.
So is the regress argument a good argument in favour of these claims? No. The regress argument
doesn’t say anything about where justification actually comes from; it just tells us where it
doesn’t come from. So the regress argument is therefore not a good argument in favour of
foundationalism in my sense sense since it doesn’t on its own get you all three components of
that position. It is a good argument for (a) but not for (b) or (c).

Of course, once we get as far as acknowledging the existence of non-inferential justification an

overwhelmingly natural question presents itself: where does such justification come from? There
is then a very natural progression from this view to full-blown foundationalism in my sense,
since the natural answer to this question is to advert to the senses. This is an independently
plausible claim about where justification actually comes from and the beauty of this response is
that it looks like it provides just the kind of non-inferential justification that we need – the sort of
justification, that is, which doesn’t land us back with a form of the regress. Unlike inference,
perception isn’t an essentially dependent form of epistemic justification. It doesn’t merely spread
around justification that is already there or require antecedently justified beliefs as inputs. It can
be what, in the first instance, gives us justification to believe; it is a source of new justification.

But this response, however natural, is not mandated by the regress; it is not something that
argument establishes. This is not to deny that the story one tells about how or why the senses are
a privileged source of justification might not be tacitly informed by the picture of justification
underlying the regress, or that rivals like coherentism might not run counter the spirit of that
argument. I have already pointed out that coherentism sits oddly next to the idea that there are
both inferential and non-inferential sources of justification, and it repudiates the linear
conception of justification underlying that argument by denying that we can meaningfully ask
after the justification of an individual belief without settling the status of all the rest. We will see
in later chapters how foundationalism is more in keeping with the spirit of the regress. My point
is merely that the foundationalist’s claims about the primacy of perception and about where
noninferential justification comes from are not an inevitable consequence of the regress
argument.

This is a strength rather than a weakness in the foundationalist’s argument since it depends upon
considerations that even those who claim not to be moved by the regress or its linear conception
of justification ought to take seriously. Views like coherentism, which fail to do so, and which
deny that there is any such thing as distinctively perceptual justification are therefore doubly
wide of the mark. The fact that they aren’t ruled out by the regress is not enough to save them
given that this is not the only thing to be said in favour of foundationalism. Foundationalism is
also motivated by the desire to accommodate the obvious fact that perception plays a
justificatory role and that is something the coherentist denies. This is a reason not to take that
position seriously quite aside from concerns about the regress.

3. Conclusion

In conclusion, I have argued that the epistemic regress argument doesn’t establish as much as
some have thoughts since it doesn’t establish foundationalism. It is a good argument in favour of
one component of that view, but not the other. This is not a problem for my view, however, since
I am not taking foundationalism to be motivated solely by the regress argument. The regress
argument motivates the foundationalist’s claims about the overall structure of justification. But
foundationalism is also a view about its sources. Specifically, it is the view that perception or
observation is the basic source of justification. This is not motivated by the regress argument and
nor is it meant to be. It is an independently plausible claim about where justification actually
comes from. The next chapter will explore two very different ways in which a foundationalist
can try and hold onto it.

… Things might be different given a different understanding of the regress argument. Sometimes
people think the problem that argument raises is temporal. If for every belief that is justified
there must be some further belief that one is be justified in believing before one can be justified
in the first, how could justification ever get started.35 They conclude that there must be some
beliefs that can be justified prior to any other beliefs being justified; there must be some beliefs
that we can in that way start from. The problem with beliefs about objects in the world external
to our minds is that it doesn’t look like a subject can have those beliefs (let alone be justified in
them) unless she has lots of other beliefs. I can’t believe that the squirrel is on the fence unless I
have the concept squirrel and that plausibly requires me to have certain sorts of beliefs about
squirrels. I may have to believe that squirrels are animals. This requires I have the concept
animal and that plausibly requires me to have yet further beliefs. Beliefs about objects in the
world external to our minds therefore do not look like they can stop the regress, if the regress is
understood in temporal terms. They aren’t beliefs that we can in that way start from.

… we have no reason to accept this view of concepts and no reason to take the temporal regress
seriously. Maybe you can’t be justified in believing anything about the world around you until
you believe lots of things about it – maybe justification, to that extent, emerges en masse. So
what? In order for there to be any justified beliefs there don’t need to be any beliefs that are
justified before all the rest. This is just a separate issue from the issue of whether or not all your
justification could derive from the justification you have to believe other things. Your
justification can be non-inferential, even if it’s not possible to be justified in that way without
also being justified in believing lots of other things.36 In that case your justification will not
derive from other beliefs, though it will in a sense depend on them (since it will require their
existence). This is not a problem though, since it is only the former that we have reason to worry
about. It is the derivation of justification that the epistemic regress worries about and we avoid
that by requiring that your justification not derive from other beliefs (whether or not it depends
upon them).37

… there is nothing objectionable about dependence per se. As we saw in the first chapter it is the
derivation of justification from one belief to the next with which the epistemic regress is
concerned. This is what the regress argument focuses on since it is only in these cases that we are
threatened with a vicious regress of justification. If one’s justification for believing p derives
from the belief that q, then not only must q be justified, but any justification one has for
believing q must be antecedent to one’s justification for believing p. If it is not, then one’s
justification will be vitiatingly circular and this is precisely what makes the regress vicious.

4 Regress arguments occur in other philosophical settings. Gilbert Ryle uses one to draw a
conclusion about the nature of voluntary action viz. that voluntary acts can’t be acts caused by a
prior act of will if acts of will are themselves voluntary (Ryle 1949). Searle uses one in
connection with intentionality (Searle 1983).

5 Evidence of its importance is legion. Thus, Alston claims that the main reason for being a
foundationalist is “the seeming impossibility of a belief’s being mediately justified without
resting ultimately on
immediately justified belief” (Alston 1976: 182); Pryor calls it “the most famous argument in
favour of non-inferential justification” (Pryor, 2005: 184); BonJour claims “the main reason for
the impressive durability of foundationalism is not any overwhelming plausibility attaching to
the main foundationalist thesis in itself, but rather the existence of one apparently decisive
argument, which seems to rule out all non-sceptical alternatives to foundationalism” (BonJour
1978:1); and Bernecker and Dretske maintain “The driving force behind foundationalism has
always been the threat of an infinite regress” (Bernecker and Dretske 2000: 231).

6 This is even more explicit in Peter Klein’s presentation of the argument. He asks us to imagine
Fred and Doris in conversation “Fred asserts some proposition, say p. Doris says something –
who knows what – that prompts Fred to believe that he had better have reasons for p in order to
supply some missing credibility. So, Fred gives his reason, r1, for p. Now Doris asks why r1 is
true. Fred gives another reasons, r2. This goes on for a while until Fred…arrives at what he takes
to be a basic proposition, say b’. (Klein 2005: 133). Presentations of the regress do not often
explicitly claim it concerns what it takes to show that a belief is justified. The argument is merely
presented in such a way that is what it concerns, whether or not that is acknowledged.

7 The regress argument is often thought to be an ancient argument. Ancient presentations are
equally ambiguous. Thus, Sextus Empiricus asks whether reasoning can ever legitimately lead to
assent and writes: ‘the mode based upon regress ad infinitum is that whereby we assert that the
thing adduced as a proof of the matter proposed needs a further proof, and this again another, and
so on ad infinitum, so that the consequence is suspension [of assent], as we possess no starting
point for our argument...we have the mode based upon hypothesis when the Dogmatists, being
forced to recede ad infinitum take as their starting point something which they do not establish
but claim to assume as granted simply and without demonstration. The mode of circular
reasoning is the form used when the proof itself which ought to establish the matter of inquiry
requires confirmation derived from the matter; in this case being unable to assume either in order
to establish the other we suspend judgement about both” (Sextus Empiricus 1967: 166-9). In
contrast, Aristotle employs a version of what I am calling the justification-making regress to
show that some justification must be non-inferential (Aristotle, 1993: A3).

8 Similarly, Susan Haack claims: “Suppose A believes that p. Is he justified in believing that p?
Well, suppose he believes that p on the basis of his belief that q. Then he is not justified in
believing that p unless he is justified in believing that q. Suppose he believes that q on the basis
of his belief that r. Then he is not justified in believing that q, and hence not justified in believing
that p, unless he is justified in believing that r. Suppose he believes that r on the basis of his
belief that s. Then he is not justified in believing that r, and hence not justified in believing that
q, and hence not justified in believing that p unless…. Now either (1) this series goes on without
end; or (2) it ends with a belief which is not justified; or (3) it goes round in a circle; or (4) it
comes to an end with a belief which is justified but not by the support of any further beliefs.”
Haack claims that if (1-3) is the case, then A’s belief that p is not justified and goes on: “If (4),
however, if the chain ends with a belief which is justified but not by the support of any further
belief, A is justified in believing that p. So, since (4) is precisely what Foundationalism claims,
only if Foundationalism is true is anyone ever justified in any belief. (Foundationalism is the
only tolerable – non-sceptical-alternative.)” (Haack 1993: 22). For other statements of the
justification-making regress see (Quinton 1973: 119) and (Pryor 2005).

9 In requiring that the subject respond to the question about what makes it the case she is
justified it is easy to miss the fact that the questioning effectively goes higher-order: if she
replies, she gives expression to a belief about what she thinks makes it the case that she is
justified. To ask what justifies this belief is therefore to ask what justification she has for
believing that she has a given justification. This is easy to miss, since it’s easily confused with
the case in which one appeals to beliefs as a source of justification. In the latter case, the beliefs
one appeals to must be justified. To ask what justifies those beliefs is therefore not to ask a
higher-order question. But that is that quite unlike the dialectical case; in that case I merely give
expression to a belief in saying what makes it the case that I am justified.

10 This is especially plausible when one thinks about what it takes to show that something is the
case. Alston draws attention to this in the following passage “showing by its very nature requires
the exhibition of grounds. Furthermore, grounds must be different from the proposition to be
shown. (This latter follows from the pragmatic aspect of the concept of showing. To show that p
is to present grounds that one can justifiably accept without already accepting p. Otherwise
showing would lack the point that goes towards making it what it is” (Alston 1976: 178-9).

11 Alston claims this as an “elementary point” (Alston 1976: 178).

12 If one models accessibility as, in effect, belief – so that for a given fact to be accessible is for
you to believe that fact obtains –there would be a much tighter connection between what is
accessible and what is capable of being articulated (assuming beliefs are capable of being
articulated). But we have no reason to model accessibility in that way.

13 It is important to what I am calling inferential justification that the inference be available

between one’s beliefs. Some people think that all justificatory relations obtain in virtue of
inferential relations between a subject’s attitudes, whether or not the attitudes involved are
beliefs. This will not make all justification inferential in the sense in which I am interested. I’ll
return to this issue in chapter 3.

14 It is also possible to frame the argument in terms of knowledge. Just as we ask why one
believes p, so too we also ask how one knows p. It is easy to use the latter to initiate a regress:
much of what we know we know because we have inferred those things from other things that
we know, but could all our knowledge be like this? Not according to the foundationalist.

15 In other cases perhaps the sort of ‘attitude’ involved might plug that gap. But this is not
plausible in the case of belief: the mere fact that one believes something, together with the fact
that what one believes implies or probabilifies something else, isn’t enough to confer
justification on the latter. Here, we must specify that the attitude (viz. belief) have a certain
additional property, namely that of being justified.

16 In a similar vein, Jose Zalabardo claims “When a proposition p obtains warrant inferentially,
it inherits it from other propositions to which it is suitably related. And p cannot inherit from
other propositions warrant that the latter don’t possess” (Zalabardo, unpublished).

17 See also (Ginet 2005: 148-9).

18 Suppose I have the unjustified belief that Fino is matured in contact with air. Can’t that belief
still make me justified in believing that I have at least one belief? If so it is not true that only
justified beliefs can make other beliefs justified, since my belief about Fino is not justified.
Perhaps this makes sense; either way it would not be a case of inferential justification since there
are no appropriate inferential relations between these two beliefs. There are appropriate
inferential relations between the belief that I believe that Fino is matured in contact with air and
the belief that I have at least one belief. But it is not obvious the former belief is unjustified; it is
my views about Fino that are unjustified, not my views about what I believe about Fino.

19 Some self-styled ‘infinitists’ are not committed to thinking we must actually have an infinite
number of beliefs. Peter Klein thinks our beliefs merely become more justified, the greater the
number of beliefs we have in support of them. This is not a version of ‘infinitism’ as I
understand that position; it is a version of coherentism (as Klein himself acknowledges) (Klein
2005).

20 Here is another way to see that: arguments are only as good as their starting points. This is
what I mean when I say that inference is a conditional vehicle of justification - whether it
succeeds in conferring justification depends on whether the starting points are any good. An
infinite regress is compatible with the starting points being all good or being all bad. The fact
that, for every belief there is some further belief that would support it doesn’t suffice to
determine whether they do support it. All it rules out is the possibility that if any of the beliefs in
the series is justified, then they aren’t all justified given that each has successors that would
justify it. But if inferential relations don’t suffice to determine whether any is good though then
something else must be necessary.

21 Thus Quinton writes: ‘For a belief to be justified it is not enough for it to be accepted, let
alone merely entertained: there must also be good reason for accepting it’ (Quinton 1973: 119).

22 Intuitively to be justified in believing p is to believe p on the basis of the facts which give you
justification to believe p. This is hard for the coherentist to make sense of, since it is very hard to
see how a belief’s inter-relations to all other beliefs could be the ‘basis’ on which you adopt it.
The only obvious way of making sense of this is to suppose you believe p on the basis of a
coherentist meta-argument claiming that the belief that p coheres with the rest of one’s beliefs,
and that beliefs which cohere are likely to be true. This makes sense of how such facts could
intelligibly be the basis upon which one believes but it is clearly very implausible as a
description of the basis upon which most people form their beliefs. Coherentism therefore leaves
most people’s beliefs unjustified and that isn’t much of a recommendation.
23 Of course, it is much harder to see how one would stop the dialectical regress in this way.
How can one show that a belief is justified other than by producing a justificatory argument in its
favour? See also (Pryor 2005: 193-4).

35 For a presentation of the argument along temporal lines see (Moore 2002: 122-3).

36 The following example from James Pryor nicely illustrates this point: ‘Consider: in order to
have the concept of a unicorn I may need to believe (i) that unicorns have hooves, and (ii) that
unicorns have horns. Now suppose I acquire evidence that a virus has killed all hoofed creatures.
Since I believe unicorns to be hoofed creatures, I form the belief (iii) that no unicorns currently
exist. It is clear that (ii) plays no role n justifying this belief. This shows that there can be
propositions that you need to believe in order to have certain concepts (you need to believe (i) in
order to have the concept of a unicorn) without those propositions mediating your justification
for every belief involving the concept Now (iii) is not an immediately justified belief. But it
serves to make my point. We can see the same phenomenon with beliefs that are good candidates
to be immediately justified like (iv) If any unicorn exists, it is identical with itself. (ii) plays no
more role in justifying that belief than it plays in justifying (iii)’ (Pryor 2005: 198).

37 Cf. (Burge 2003: 503-48).

63 …Why should it be acceptable to allow that one’s grounds for ruling out the possibility of
mistake, do not themselves rule out the possibility of mistake, when it is not acceptable to claim
that is so in the original case? Thus this line is immediately regressive. It could only be stopped
by appealing to infallibly justified beliefs.

99 …The regress assumes that justification is a positive epistemic status. In contrast, a belief
might be ‘rational’ (or not irrational) just as long as it doesn’t conflict with other things you
believe.

114 …The epistemic regress argument merely tells us that some of our beliefs must be non-
inferentially justified, so there must be sources of such justification. That doesn’t tell us anything
about which beliefs are non-inferentially justified or what the sources of such justification are.
For an answer to that question we must look to examples
Reasons, Regresses, and Tragedy:
The Epistemic Regress Problem and the Problem of the Criterion
Forthcoming in American Philosophical Quarterly
Andrew D. Cling
Department of Philosophy
The University of Alabama in Huntsville
Huntsville, Alabama 35899

Abstract

The epistemic regress problem is akin to the problem of the criterion. Each is posed by
propositions that are independently plausible but jointly inconsistent. Each problem makes
assumptions that imply (i) that belief in a proposition P can have an epistemically valuable
relational property—being supported by evidence or being authorized by a criterion of truth—
only if P is the first member of an endless sequence of propositions each of which stands in that
relationship to its successor, and (ii) that this prevents any proposition from having that property.
The best explanation for these similarities is that each problem is a special case of a general
paradox about reasons: it seems that we can have reasons for belief, that reasons must be based
on reasons, but that this requires an endless regress that is incompatible with having reasons.
This problem is especially acute for reasons that are non-arbitrary from our own point of view
and for reasons that are required for rational intellectual autonomy, for these kinds of reasons
must be based on reasons. To respond to the problem we must either accept the idea that endless
regresses are required for reasons, or recognize, in a tragic spirit, that having fully satisfactory
reasons is a worthy but impossible ideal.

1. Introduction

The epistemic regress problem is about the possibility of having beliefs that are based on
evidence. The problem of the criterion is about the possibility of having beliefs that are based on
general standards for distinguishing what is true from what is false. These problems are similar.
Each is constituted by a set of propositions about epistemically valuable relational properties—
being supported by evidence and being authorized by a criterion of truth—that are individually
plausible but jointly inconsistent, a paradox. The propositions that constitute one problem have
the same forms as the propositions that constitute the other. In particular, each problem turns on
the claim that a belief can have a valuable relational property only by means of an endless
regress and that this requirement cannot be satisfied. Finally, since evidence and criteria are
kinds of reasons, each problem is about the possibility of having beliefs that are based on
reasons.

The best explanation for the similarities between these problems is that each is a special case of a
more general problem, the paradox of reasons. It seems that we can have beliefs that are based
on reasons. It seems that we can base the belief that P on a reason P only if P is based on a
1 2 2
reason. That, however, implies that having a reason always requires an endless regress of reasons
—either a circle of reasons or an infinite sequence of distinct reasons—but it seems that this
requirement cannot be satisfied.
I begin with the epistemic regress problem. I state the problem and use it as a template for
understanding the problem of the criterion. Next I show that each of these problems is a special
case of the paradox of reasons. Finally, I give two arguments to show that if we value beliefs that
are based on reasons, those reasons must be based on reasons. One argument proceeds from the
assumption that reasons must be non-arbitrary from one’s own point of view. The other is based
on the demands of rational intellectual autonomy. I conclude that if we accept these values, the
best response to the paradox of reasons is either to reject the idea that endless regresses block
reasons, or to accept, in a tragic spirit, that having fully satisfactory reasons is a worthy but
impossible ideal.

2. The Chain and the Wheel

The epistemic regress problem is the ancient skeptical paradox of the chain. To know a
proposition we must believe it and base that belief on evidence. So, since supporting evidence
must be supported, knowledge requires an endless chain of reasons: either an infinite regress of
different reasons or a sequence of finitely many reasons that loops back endlessly through the
same proposition, a circle. Knowledge is secured only by an unanchored chain or by a noose.
Since no one can know a proposition if that requires an endless regress, knowledge is impossible.
The problem of the criterion is the ancient skeptical paradox of the wheel. To know a proposition
we must first know a standard for distinguishing true from false propositions—a criterion of truth
—but we can know a criterion only by means of propositions or criteria we already know. These
must be authorized either by the criterion in question—a circle—or by different propositions or
criteria—an infinite regress. So, again, knowledge is secured only by an unanchored chain or by
a noose. Since no one can know a proposition if that requires an endless regress, knowledge is
impossible.

These problems are strikingly similar. Each is about conditions that beliefs must satisfy in order
to have an epistemically valuable relational property: being supported by evidence or being
authorized by a criterion of truth. Each assumes that a proposition can stand in this relationship
to another only if the latter stands in the same relationship to a proposition. In each case,
therefore, an epistemically desirable relational property requires an endless sequence of
propositions each of which stands in that very relationship to its successor. Finally, each problem
assumes that this requirement cannot be satisfied. The best explanation for the similarity of these
problems is that each is a special case of a general problem about the possibility of having beliefs
that are based on reasons of any kind, the paradox of reasons: if reasons must be based on
reasons but reasons cannot depend upon an infinite regress of reasons, having beliefs that are
based on reasons is impossible.

3. The Epistemic Regress Problem

The epistemic regress problem is posed by three propositions that are individually plausible but
jointly inconsistent. I shall give each proposition a nickname and state it both in (philosopher’s)
English and in a first-order language with modality:1

(R1) Evidential Support is Possible. It is possible that some proposition is evidentially supported
by a proposition. ◊(∃x)(∃y)Sxy.
(R2) Supporting Propositions are Supported. Necessarily, if a proposition P1 is evidentially
supported by a proposition P2, then there is a proposition P3 that evidentially supports P2.
􀀀(∀x)(∀y)(Sxy → (∃z)Syz).
(R3) Regresses Block Support. Necessarily, if it must be that any proposition P1 is evidentially
supported by a proposition P2 only if P1 and P2 are the first two members of an endless
sequence of propositions each of which is evidentially supported by its successor, then no
proposition can be evidentially supported by any proposition.
􀀀[􀀀(∀x)(∀y)(Sxy → ERxy) → ~◊(∃x)(∃y)Sxy].2

Variations on the epistemic regress problem result from different specifications of the conditions
under which beliefs are evidentially supported by propositions. Because it results in a powerful
version of the problem, let us suppose that P is evidentially supported by P for person S if, and
1 2
only if, (i) S believes P , (ii) P is evidence for P , and (iii) S’s belief that P is based on P ’s
1 2 1 1 2
being evidence for P . On this interpretation, the problem is about the possibility of having
1
beliefs that are based on propositions that provide evidence for them. The conditions under
which belief in a proposition is based on a proposition are disputed, but it is plausible that a
person S’s belief that P is evidentially based on a proposition P only if S believes P , P is
1 2 1 2
evidence for P , S believes P in part because P is evidence for P , and S would tend not to
1 1 2 1
believe P if P were not evidence for P . By requiring only that S would tend not to believe P
1 2 1 1
if P were not evidence for P , and not requiring that S would not believe P if P were not
2 1 1 2
evidence for P , it is left open that a belief can have more than one basis. If S’s belief that P is
1 1
based on P and is also, independently, based on P , then S need not fail to believe P if P
2 3 1 2
were not evidence for P . This account of evidential support does not require that supporting
1
reasons be believed, but it does require that only beliefs are supported.
(R1) asserts that a belief can be based on the evidence provided by a proposition. (R2) says that a
belief can be based on the evidence provided by a proposition only if that proposition is believed
and is supported by evidence. This, however, entails that a belief that P can be evidentially
1
supported by a proposition P only if P and P are the first two members of an endless
2 1 2
sequence of propositions each of which is evidentially supported by its successor. (R3) says that
this requirement cannot be satisfied: if having evidential support always requires an endless
regress of reasons, no belief can be evidentially supported. Thus (R2) and (R3) jointly entail that
(R1) is false: no belief can be evidentially supported.3

4. The Problem of the Criterion

The problem of the criterion is about the possibility of having beliefs that are authorized by
criteria of truth. Criteria of truth are general principles that distinguish true from false
propositions, so the problem of the criterion is different from the epistemic regress problem.
Because of complications surrounding the ways in which criteria might acquire the epistemic
standing they need in order to authorize propositions, the specific forms of the propositions that
pose the problem of the criterion are more complex than the forms of the propositions that pose
the epistemic regress problem. Still, the problems have the same general form and each is a
problem about the possibility of having an epistemically significant reason for a belief.

A criterion of truth is a general principle according to which propositions with a detectable

property are true or likely to be true. An ideal criterion would have this form:

(C) For all P, P has M if, and only if, P is true.

In (C), ‘P’ ranges over propositions and ‘M’—for ‘mark of truth’—expresses a detectable
property of propositions.

A detectable feature M need not be observable, but must be such that we can be sensitive to the
presence or absence of M, whether by sensory observation, introspection, rational reflection, or
in some other way. To be sensitive to a property is to be disposed, under the right conditions, to
have a representation of the property that occurs fairly regularly depending on the instantiation of
the property. Perfect sensitivity is too much to ask, while getting it right only sometimes is not
enough. Sensitivity to a mark of truth must be strong enough to make it possible to base beliefs
on the presence or absence of M.

In an ideal criterion of truth, the connection between M and a proposition’s being true must be,
at least, a lawlike fairly necessary or fairly sufficient condition of being true.4 Ideal marks of
truth are like the touchstones used in ancient times to assess the presence and quality of precious
metals depending upon the type of visible mark made on the stone by scratching the metal
against it. M is a lawlike fairly necessary condition of truth just in case propositions lacking M
would tend to be false. M is a lawlike fairly sufficient condition of truth just in case propositions
with M would tend to be true. There are helpful criteria in which there is no lawlike connection
between mark and truth, however. For suppose it happens that everything S’s mother says is true,
though this is an accident because it is relevantly possible that S’s mother says something false.
Under these conditions, S might use what S’s mother says as a useful mark of truth by basing
beliefs on S’s mother’s having made the corresponding statements. It is good to be liberal in our
accounts of detectability and of criteria of truth, for the problem of the criterion remains deeply
serious no matter how many propositions we allow to be potential criteria of truth. Placing
additional conditions on what can be a criterion of truth will make it increasingly difficult to
have criteria that can be used as reasons for belief, but such conditions do not affect the key
skeptical problem of accounting for how any criterion, however understood, can acquire the
standing it needs in order to authorize any proposition.5

Consider some examples of criteria of truth. Theorems of logic are criteria of truth because each
is a general principle to the effect that any proposition with a specified, detectable form must be
true. For example, persons are able to identify infinitely many truths if they know that every
proposition with the form P or not-P is true. Theorems with a conditional form such as the one
underwriting modus ponens—if (P and (if P, then Q)), then Q—not only put us in a position to
identify infinitely many necessary truths, they put us in a position to identify contingent truths
that stand in the relevant relations to other truths. Thus, having identified it is raining and if it is
raining, then the grass is wet to be true, modus ponens puts us in a position to identify the grass
is wet to be true.

Some criteria would enable persons to identify true propositions by means of detectable features
of the mental states that are, in a specified way, related to those propositions. Descartes’
principle of clear and distinct perception—everything I clearly and distinctly perceive is true—is
a principle of this sort, for it states that when one’s awareness of a proposition P has a detectable
characteristic, P is true.6 A principle of induction such as the unobserved resembles the observed
is a criterion of truth that would enable persons to identify some propositions to be true on the
basis of their relationship to propositions that describe states of affairs they have experienced.7
Again, is good to be liberal in counting a wide range of principles as criteria of truth, whether
they specify only sufficient, only necessary, or only probabilistic marks of truth, and whether
they specify that it is characteristics of propositions themselves or characteristics of one’s
awareness of those propositions that are symptomatic of their truth.8

Taking the similarities between the epistemic regress problem and the problem of the criterion as
a clue and using the template provided by (R1)−(R3) as a guide, we might conjecture that the
problem of the criterion is posed by these propositions:9

(C1) Authorization is Possible. It is possible that some proposition is authorized by

a criterion of truth. ◊(∃x)(∃y)Axy.

(C2) Authorizing Criteria are Authorized. Necessarily, if a proposition P is

1
authorized by a criterion C , then there is a criterion C that authorizes C .
1 2 1
􀀀(∀x)(∀y)(Axy → (∃z)Ayz).

(C3) Regresses Block Authorization. Necessarily, if it must be that any proposition P1 is

authorized by a criterion C1 only if P1 and C1 are the first two members of an endless sequence
of propositions each of which is authorized by its successor, then no proposition can be
authorized.
􀀀[􀀀(∀x)(∀y)(Axy → ERxy) → ~◊(∃x)(∃y)Axy].10

Variations on the problem of the criterion result from different accounts of the conditions under
which a proposition is authorized by a criterion of truth. Because it results in a powerful version
of the problem, let us suppose that P is authorized by criterion C for person S if, and only if, (i) S
believes P, (ii) C is a correct criterion of truth that sanctions P, and (iii) S’s belief that P is based
on C’s sanctioning P.

(C1)–(C3) capture a version of the problem given by Sextus Empiricus:

… [I]n order to decide the dispute that has arisen about the criterion, we have need of an
agreed-upon criterion by means of which we shall decide it; and in order to have an
agreed-upon criterion it is necessary first to have decided the dispute about the criterion.11

On a natural reading, the problem raised in this passage has the same structure as the epistemic
regress problem. A belief can be authorized by a criterion only if that criterion is itself authorized
by a criterion. Therefore, a belief can be authorized by a criterion only if that belief and that
criterion are the first two members of an endless sequence of propositions each of which is
authorized by its successor. So no belief can be authorized by any criterion of truth.

Whether or not they capture Sextus’ problem, (C1)–(C3) do not quite capture the most general
version of the problem of the criterion, for the problem need not assume (C2) Authorizing
Criteria are Authorized. The problem is that (C2) does not capture the possibility that a criterion
might acquire the standing it needs to authorize propositions by being supported by propositions
that are not criteria of truth.12 “Track-record” arguments, for example, are intended to show that
principles designed to distinguish true from false propositions are correct because, in a
significant range of cases, the propositions they sanction are actually true—the principles have
good track records.13 If, however, the premises of track-record arguments must themselves be
authorized by criteria of truth or supported by propositions that are authorized by criteria of truth
—and how else could we decide that these propositions are true?—a more complex regress
arises. Such a regress begins with a proposition that is authorized by a criterion that is, in turn,
either authorized by a criterion or supported by authorized propositions, ad infinitum.

To capture these possibilities, we should modify (C2) to read:

(C2*) Necessarily, if a proposition P is authorized by a criterion of truth C ,

1 1
then either C is authorized by a criterion of truth C , or C is supported by a
1 2 1
proposition P that is authorized by a criterion of truth C .
1 2
􀀀(∀x)(∀y)[Axy → (∃z)(Ayz ∨ (Syz & (∃w)Azw))].

With corresponding changes made to (C3), the problem of the criterion is posed by these
propositions:

(C1*) Authorization is Possible. It is possible that some proposition is authorized

by a criterion of truth. ◊(∃x)(∃y)Axy.
 (C2*) Authorizing Criteria are Authorized or Supported. Necessarily, if a
proposition P is authorized by a criterion of truth C , then either C is
1 1 1
authorized by a criterion of truth C , or C is supported by a proposition P that
2 1 1
is authorized by a criterion of truth C .
2
􀀀(∀x)(∀y)[Axy → (∃z)(Ayz ∨ (Syz & (∃w)Azw))].

(C3*) Regresses Block Authorization. Necessarily, if it must be that any

proposition P is authorized by a criterion C only if P and C are the first two
1 1 1 1
members of an infinite sequence of propositions each of which is authorized or
supported by its successor, then no proposition can be authorized.
􀀀[􀀀(∀x)(∀y)(Axy → ERxy) → ~◊(∃x)(∃y)Axy].14

5. Reasons and The Paradox of Reasons

The similarities between the epistemic regress problem and the problem of the criterion are best
explained by the fact that each is a special case of a problem about having reasons for belief that
I call ‘the paradox of reasons.’ I shall briefly survey some kinds of reasons for belief and then
state the paradox of reasons. Many types of reasons are vulnerable to the paradox of reasons,
especially reasons that allegedly provide a basis for belief.

There are many kinds of reasons for belief. Reasons vary depending upon the type of proposition
that can function as a given kind of reason—must an explanatory reason, for example, be a
general principle?—and the conditions that must be satisfied in order for that sort of proposition
to stand in the appropriate reason-providing relationship to the belief in question.

‘Reason for belief’ is ambiguous. ‘Belief’ can refer to a proposition—the content of the
psychological state that consists of a person’s accepting that proposition—or to the psychological
state as a whole. These are connected, because reasons for a proposition can also be relevant to a
person’s accepting that proposition. I propose to be liberal in counting a wide range of
propositions as reasons for belief: evidence that a proposition is true; a criterion of truth that
sanctions a belief; an account of how or why a proposition is true; an epistemic principle that
implies that a belief would be a case of knowledge or justified belief; an account of how the state
of affairs described by a proposition is possible; factors that would make a belief valuable
whether or not it is true, likely to be true, or reasonably believed to be true. Given the many
values propositions can have and the many interests beliefs can serve, it is highly unlikely that
this list is even close to being exhaustive.

Reasons for belief also vary depending upon the kind of relation that reason-providing
propositions must bear to beliefs in order to be reasons of the relevant type. To be a reason, a
proposition not only must be of the right type, it must stand in the appropriate relationship to the
belief for which it is a reason. There is, for example, an important difference between a belief for
which there merely happens to be a reason and a belief that is based on a reason. Basing one’s
belief that P on a reason R is a matter, at least in part, of believing P because R is a reason for it.
If S bases S’s belief that P on reason R, then S believes that P because R is a reason (of the
relevant sort) for P. Such basing reasons are particularly vulnerable to the paradox of reasons,
for it seems that many propositions can be basing reasons only if they are believed and are
themselves based on reasons.

An essential feature of reasons for belief is that they be non-arbitrary. It seems, furthermore, that
part of what makes a reason non-arbitrary is that there is a reason for it. If so, however, then a
proposition can be a reason only if it is the first member of an endless regress of reasons. Since it
seems that this endless regress requirement cannot be satisfied, we have the paradox of reasons:

(1) Reasons are Possible. It is possible that a proposition P is a reason for a

2
proposition P . ◊(∃x)(∃y)Rxy.
1

(2) Reasons Need Reasons. Necessarily, if P is a reason for P , then there is a

2 1
15
proposition P that is a reason for P .
3 2
􀀀(∀x)(∀y)(Rxy → (∃z)Ryz)

(3) Endless Regresses Block Reasons. Necessarily, if it must be that any proposition
P is a reason for a proposition P only if P and P are the first two members of
2 1 1 2
an endless sequence of propositions each of which is such that its successor is a
reason for it, then no proposition can be a reason for any proposition.
􀀀[􀀀(∀x)(∀y)(Rxy → ERxy) → ~◊(∃x)(∃y)Rxy].

Because it poses a particularly serious version of the problem, let us assume that P2 is a reason
for P1 just in case P2 is a basing reason for P1. A proposition P2 is a basing reason for a
proposition P1 for a person S just in case (i) S believes P1, (ii) P2 has the objective
credentials necessary to be a reason of the relevant sort for P1, and (iii) S bases S’s belief that P1
on P2. (1)–(3) constitute a paradox for any kind of reason that makes (1)–(3) individually
plausible, including the epistemic regress problem and the problem of the criterion. Basing
reasons are particularly vulnerable to the paradox of reasons because it seems that there are some
kinds of purposes for which it is arbitrary, hence unreasonable, to base one’s beliefs on reasons
that are not, in turn, based on reasons.

6. Reasons Need Reasons

If we suppose that (1) Reasons are Possible, then we must respond to the paradox of reasons, in
any of its guises, by rejecting either (2) Reasons Need Reasons or (3) Endless Regresses Block
Reasons. Some philosophers have argued that (2) is not true of all reasons. Whether or not they
are correct, (2) is true of some important kinds of reasons. There are two powerful arguments for
(2). The first proceeds from the value of having the kinds of reasons for belief that are non-
arbitrary from the believer’s own point of view. The second is based on the value of rational
intellectual autonomy. If either of these arguments is correct, then either (3) is false or (1) is
false.
Some philosophers think that (2) is not true of all reasons. For example, foundationalist accounts
of evidential support claim that some empirical beliefs can be evidentially supported by sensory
states of appearing that need not be evidentially supported by beliefs. (e.g. Moser 1985) Some
philosophers have responded to the problem of the criterion by claiming that since correct
criteria can authorize propositions for a person S provided only that S’s reasoning conforms to
those criteria, S need not base S’s beliefs on criteria or have reasons for criteria in order to have
authorization for S’s beliefs. (e.g. Van Cleve 1979 and 1984) Similarly, some philosophers think
that since S can have knowledge or justified belief provided only that S’s beliefs satisfy the
conditions specified by correct epistemic principles, S need not believe the relevant epistemic
principles or base S’s beliefs on epistemic principles in order to have knowledge.16 These
accounts may succeed in showing that there are some kinds of reasons for belief that do not
satisfy (2). There are, however, some valuable reasons—including important basing reasons—
that must satisfy (2). There are two powerful arguments for this.

Reasons must, by their very nature, be non-arbitrary in the relevant respect. It follows that
basing reasons must be non-arbitrary:

(NA) Necessarily, if P is a basing reason for P , then P is not arbitrary in the relevant
2 1 2
respect.
􀀀(∀x)(∀y)(Rxy → NAy).17

Some reasons are non-arbitrary in the relevant respect only if they are non-arbitrary from the
believer’s own point of view. For example, if we want to conduct an investigation in light of
considerations that we recognize to be good reasons, then we need to base our beliefs on reasons
that are non-arbitrary from our own point of view. A basing reason P2 for a belief that P1 is non-
arbitrary from S’s point of view only if S bases S’s belief that P1 on P2, S believes P2, and S
bases S’s belief that P2 on a reason. Thus:
(PoV) Necessarily, if P1 is a basing reason that is non-arbitrary from S’s own point of view, then
there is a proposition P2 that is a basing reason for P1.
􀀀(∀x)(NAx → (∃y)Rxy).

(NA) and (PoV) jointly entail this version of (2):

(2B) Necessarily, if a proposition P is a basing reason for P , then there is a proposition

2 1
P that is a basing reason for P .
3 2
􀀀(∀x)(∀y)(Rxy → (∃z)Ryz).

(2) is therefore true of the kinds of basing reasons that are non-arbitrary from the believer’s own
point of view. This includes the supporting evidence that is the focus of the epistemic regress
problem and the authorizing criteria that is the focus of the problem of the criterion. Responses
to these special cases of the paradox of reasons that do not require that reasons be basing reasons
or that do not require that basing reasons be non-arbitrary from the believer’s own point of view
do not solve the paradox of reasons.

A different argument for (2B) is based on the value of rational intellectual autonomy. This value
has two components: rationality and intellectual autonomy. To be rational with respect to a
given kind of reason is to be such that one bases one’s beliefs on reasons of that sort:

(R) Necessarily, if S believes P1, then there is a P2 such that P2 is a basing reason for P1.
􀀀(∀x)(Bx → (∃y)Rxy).

So, for example, evidentialism is the view that epistemic rationality—forming beliefs in a way
that contributes to having knowledge or epistemically justified beliefs—is a matter of basing
one’s beliefs on evidence. A related position about criteria of truth is that one should believe
only propositions that are, or can be, based on correct criteria of truth.

Intellectual autonomy is the ideal of controlling what one believes by conducting one’s inquiry in
such a way that one’s beliefs are caused or otherwise controlled by aspects of one’s self, perhaps
aspects of one’s self that one identifies with or that are in some other way genuine. Intellectual
autonomy thus requires that if our beliefs are based on reasons, we believes those reasons:

(A) Necessarily, if P is a basing reason for S’s belief that P , then S believes P .
2 1 2
􀀀(∀x)(∀y) (Rxy → By).

(R) and (A) jointly entail (2B). For rationality requires that we base our beliefs on reasons and
intellectual autonomy requires that we believe those reasons. Taken together, these imply that
only propositions based on reasons can be reasons and this requires an endless regress of reasons.
Responses to the paradox of reasons that do not require that persons base their beliefs on reasons
or that do not require that persons have reasons for their reasons cannot explain how rational
intellectual autonomy is possible.

7. Endless Regresses or Tragedy

If we value reasons that are non-arbitrary from our own point of view or we value rational
intellectual autonomy, we must either reject (3) and claim that reasons always require endless
regresses of reasons, or we must accept, in a tragic spirit, that it is impossible to have fully
adequate reasons. Since (3) is true, I fear we must accept tragedy.

(2) Reasons Need Reasons entails that whenever P is a reason for P , then P and P are the
2 1 1 2
first two members of an endless sequence of propositions each of which is such that its successor
is a reason for it. Such an endless regress either must be such that no proposition appears both
sooner and later—it can be a sequence of infinitely many reasons—or it must be such that some
proposition appears both sooner and later, a circle. If there can be reasons, as (1) says, and
reasons require reasons, as (2) says, then there can be reasons even though all reasons require
endless regresses of reasons.
It is hard to see how this endless regress condition can be satisfied, for it seems that (3) is true: if
reasons always require endless regresses of reasons, there can be no reasons. It seems that some
circular arguments might be good.18 It also seems that infinite regresses of reasons are
conceptually possible and that an actual infinite regress of reasons might enhance the strength of
its component reasons. What is false, however, is a proposition implied by (1) & (2) but
incompatible with (3): that it is possible that there are reasons but that all reasons require endless
regresses of reasons.

Here is why. If endless regresses of reasons are not always vicious, then there is a distinction
between endless sequences of genuine reasons and endless sequences of arbitrary propositions.
Let the bold arrows ‘→,’ ‘←,’ ‘↑,’ and ‘↓’ indicate that a proposition at the pointed end of an
arrow is entailed by the proposition at the other end. Now consider the following two schemata
for endless, circular sequences of propositions:

(I) P → P
1 2
↑ ↓
P ←P
4 3

(II) ∼ P ← ∼ P
1 2
↓ ↑
∼P →∼P
4 3

(I) and (II) represent endless sequences of propositions connected by logical relations that are
relevant to some kinds of reasons, yet (I) and (II) are not by that very fact reason-providing for
the same person at the same time. So, if either (I) or (II) is an endless sequence of genuine
reasons and the other is not, there must be something that distinguishes (I) from (II) and that
explains why one is a sequence of genuine reasons and the other is not. It follows that in addition
to being a component of an endless sequence of propositions—circular or infinite—a genuine
reason R must have an additional property Φ that, taken together with being a component of the
endless sequence of propositions, makes R a reason. Φ cannot, on pain of circularity, be a reason
that requires an endless regress of reasons, but Φ must distinguish reasons from propositions in
otherwise similar endless sequences of propositions that are not reasons.

This cannot be done. There are two possibilities. First, Φ may be or imply the existence of a
proposition that is a reason for R that does not require an endless regress of reasons. Second, Φ
may be a non-propositional factor that accounts for R’s having the standing necessary to be a
reason. If Φ is or implies the existence of a proposition that is a reason for R that does not require
an endless regress of reasons, then (2) is false because reasons do not require an endless regress
of reasons. If Φ is a non-propositional factor that accounts for R’s having the standing required to
be a reason, then (2) is gratuitous because Φ is a reason that is not a proposition. So reasons
cannot satisfy the requirement that all reasons require endless regresses of reasons. The upshot is
that even if there can be endless sequences of genuine reasons, these can be reason-providing
only because of reasons that do not require endless regresses.19
So the prospects of solving the paradox of reasons by rejecting (3) are dim. Still, given the value
of basing beliefs on reasons that are non-arbitrary from one’s own point of view or the value of
rational intellectual autonomy, the most promising way to solve the problem by rejecting (3) is to
claim that reasons require circles, not infinite regresses, of reasons. The need for infinite
regresses of reasons would push us beyond the capacity of our occurrent or dispositional beliefs20
and thus would not give us the kinds of basing reasons that are non-arbitrary from our point of
view or that are required for rational autonomy. By keeping the number of required reasons
finite, the requirement that all reasons require circles of reasons is, at least in this way,
compatible with having non-arbitrary reasons and with rational intellectual autonomy. A major
task for this response to the paradox of reasons is to show why the argument for (3) given in the
previous paragraph is mistaken and to explain how it is possible that all genuine reasons require
circles of reasons.

The only other way to respond to the paradox of reasons in a way that preserves the value of
having reasons that are non-arbitrary from one’s own point of view or the value of rational
intellectual autonomy is to concede that (1) is false—we cannot have reasons; at least, we cannot
have some of the kinds of reasons we want—but to claim that this is compatible with seeking
and valuing reasons. According to this view, our predicament is tragic because it is impossible
for us to satisfy the conditions that important values impose upon us. It is valuable to base our
beliefs on reasons that are non-arbitrary and it is valuable to believe our reasons, but we cannot
do both. Similarly, rationality is valuable and intellectual autonomy is valuable, but these values
impose conditions that cannot be jointly satisfied. It is important to strive to have the kinds of
reasons for beliefs that are required by these values, but the reasons that we acquire in the
attempt cannot be fully satisfactory by these measures because we must run out of reasons for
our reasons in ways that appear arbitrary to us and that are symptomatic of the limits of our
rational autonomy. It is impossible to satisfy the ideals that impose (2) on us, but those ideals are
still valuable.

That there might be value in such a tragic predicament was recognized by the legendary football
coach Vince Lombardi who said, “Perfection is not attainable, but if we chase perfection we can
catch excellence.” Lombardi’s claim implies that although perfection is impossible, it is a worthy
ideal because the pursuit of perfection makes it possible to become excellent. Lombardi’s
aphorism might apply to our predicament with respect to reasons. We cannot have some of the
kinds of reason to which we aspire, but perhaps in pursuing them we can acquire excellent
reasons. A promising area for further investigation is to explain how this position is possible by
thinking about reasons in general, and epistemic reasons in particular, in this tragic light.
8. Conclusion

At the cost of tedium, I have tried to be as rigorous and complete as possible in stating the
relevant paradoxes and in presenting my arguments. I hope that this will sharpen future
discussion of these problems and make it easier for others to uncover any mistakes I have made.
I have argued that the best explanation for the striking similarities between the epistemic regress
problem and the problem of the criterion is that each is a special case of the paradox of reasons.
Because the value of having reasons that are non-arbitrary from our own point of view and the
value of rational intellectual autonomy imply that (2) Reasons Require Reasons, the best non-
skeptical response to the paradox of reasons, in any of its guises, is to reject (3) Endless
Regresses Block Reasons. There is a fatal problem with this response, however, for (3) is true. I
reluctantly conclude that we should reject (1) Reasons are Possible in a tragic spirit that
recognizes the value of pursuing reasons. Each of my arguments ultimately rests on assumptions
for which I have given no reasons. I take cold comfort in the fact that although running out of
reasons is a mark of my imperfection, it is also my fate.21

NOTES

1 A related formulation of the problem is given in Cling 2008. The version presented here differs
slightly from that one, however. These differences are discussed in Note 2.

2 In all numbered principles, the domain is propositions that are indexed to persons and times.
‘Sxy’ = x is evidentially supported by y. ‘ERxz’ = x and y are the first two members, in that order,
of a sequence of propositions 〈x, y, …〉 each of which has a successor that evidentially supports
it.

As stated in the first-order language with modality, (R1)–(R3) are not formally inconsistent. The
following additional assumption is needed to guarantee formal inconsistency:
 (R4) Supporting Propositions are Supported Requires Endless Regresses. Necessarily, if
it is necessary that only evidentially supported propositions provide evidential support,
then it is necessary that any proposition is evidentially supported by another only if they
are the first two members of an endless sequence of propositions each of which is
evidentially supported by its successor.
􀀀[􀀀(∀x)(∀y)(Sxy → (∃z)Syz) → 􀀀(∀x)(∀y)(Sxy → ERxy)].

This principle is necessarily true and (reasonably) evident.

Every circular sequence is endless, though some circular sequences contain only finitely many
members. Suppose, for example, that P is evidentially supported by P , that P is evidentially
1 2 2
supported by P , and that P is evidentially supported by P , closing the circle. It is possible to
3 3 1
represent these propositions as a finite sequence every component of which is evidentially
supported by its successor, if it has one, thus: σ = 〈P , P , P 〉. Since, however, we are given
1 2 3
that P is evidentially supported by P and we are assuming that only evidentially supported
3 1
propositions provide evidential support, the given conditions can be satisfied only if there is an
endless sequence with finitely many components σ = 〈P , P , P , P , P …〉 each component
E 1 2 3 1 2
of which has a successor that evidentially supports it. Indeed, for any sequence of propositions
each of which is evidentially supported by its successor, if it has one, such that some component
appears both sooner and later—that is, for any circular sequence of reasons—there is an endless
sequence of propositions having only finitely many components each of which has a successor
that evidentially supports it.

(R1)–(R3) differ in some ways from the propositions used to pose the problem in Cling 2008.
For one thing, (R1)–(R3) include modal operators. This is necessary in order and
to avoid some counterexamples. More importantly, the original version of (R3)—(∀x)[(∀y)(Sxy
→ ERxz) → ~(∃z)Sxz)], which is equivalent to (∀x)[(∃y) Sxy → (∃z)(Sxz & ∼ERxz)]—is
replaced by a weaker proposition (ignoring modal operators). The original principle is true if,
and only if, any proposition evidentially supported only by endless regresses of reasons is not
evidentially supported. (R3), however, allows that some propositions may be evidentially
supported only by endless regresses of reasons. It requires only that not every proposition is
evidentially supported only by an endless regress. In addition to being simpler and weaker than
its ancestor, (R3) is just sufficient to guarantee that (R1)–(R3) are jointly inconsistent. With
modal operators inserted at the relevant places, the ancestor of (R3) implies (R3), but not vice
versa. (R3) follows from an even stronger injunction to the effect that there can be no infinite
regress of reasons—~◊(∃x)(∃y)ERxy—but does not imply it. For it is compatible with (R3) that
there be an infinite regress of reasons that provides evidential support for each of its components.
(R3) is also compatible with the view that some supported propositions are evidentially
supported only by infinite regresses of reasons. (R3) rules out only the possibility that there is
some evidentially supported proposition but no proposition that can be evidentially supported
without an infinite regress of reasons.
Finally, note that the epistemic regress problem is primarily about evidential support and not
about epistemic justification or knowledge. (R1)–(R3) pose a problem for epistemic justification
and knowledge only given the further assumption that these require evidential support.

3 Similarly (R1) & (R2) entails that (R3) is false and (R1) & (R3) entails that (R2) is false.

4 ‘Fairly necessary condition’ is taken from Hacking 1995, p. 82 though I do not assume that he
would accept my gloss of it.

5 There are also interesting questions about the conditions under which one detects a mark of
truth. Suppose, for example, that S believes something that S’s mother in fact has said as the
result of a sequence of events that masks her having said it. Perhaps S believes P in a fit of
credulousness upon finding a sentence expressing P written on a scrap of paper that, unknown to
S, was S daughter’s transcription of a statement expressing P that was made by S’s mother. In
such a case, S believes a proposition that is sanctioned by a mark of truth, but it does not seem
that S believes it because S has detected that mark, for S’s belief is insensitive to the mark and,
by hypothesis, is not sensitive to any mark of truth. It seems plausible to conjecture that one
detects M only if one is in a relevant mental state D (for ‘detector’)—for example, the state of
believing that P has M—that is fairly sensitive to the presence or absence of M. Thanks to an
anonymous referee for calling attention to this problem about detection and detectability.

6 See Descartes 1644/1985, pp. 207–208. See Coffa 1991, p. 10 for an interesting discussion of
this standard in early modern philosophy.

7 Indeed, the problem of induction is a special case of the problem of the criterion, as shown by
Weintraub 1995.

8 If the mark of truth that figures in the criterion is connected in the right sort of way to the
property of being true, that criterion will be explanatory. Such a criterion would not only provide
a means of identifying true propositions, it would provide the basis for an account of why those
propositions are true. Not all explanatory features need be marks of truth, however, for there may
be undetectable properties of truth. If there is a conceptual connection between the mark of truth
and the property of being true, an understanding of the criterion would put a person in a position
to see that it is a correct criterion. This observation is adapted from a discussion of ethical
standards in Schneewind 1998, pp. 522–523.

9 As in the epistemic regress problem, knowledge and justification are affected only given the
additional assumption that knowledge or justification requires authorization by a criterion of
truth.

10 ‘Axy’ = y is a criterion of truth and x is authorized by y; ‘ERxz’ = x and y are the first two
members of a sequence of propositions 〈x, y, …〉 each of which has a successor that authorizes it.

11 Sextus Empiricus 1996, pp. 128-129.

12 This is illustrated by one of Chisholm’s statements of the problem:

To know whether things really are as they seem to be, we must have a procedure for
distinguishing appearances that are true from appearances that are false. But to know whether
our procedure is a good procedure, we have to know whether it really succeeds in
distinguishing appearances that are true from appearances that are false. And we cannot
know whether it does really succeed unless we already know which appearances are true and
which ones are false. And so we are caught in a circle. (Chisholm 1973, p. 3.)

Chisholm overlooks the possibility that Sextus assumes must hold: that an authorizing criterion
acquires the standing it needs because it is authorized by a criterion of truth. Chisholm is also
wrong to think that the requirement that authorizing criteria be supported requires a circle. It
requires either a circle or an endless sequence of distinct propositions or criteria, an infinite
regress. A fully general account of the problem must recognize all of these possibilities.

13 Alston 1993, Chapter 2.

14 ‘ERxy’ = x and y are the first two members of an endless sequence of propositions each of
which is authorized or supported by its successor.

(C1*)–(C3*) provide a basis for describing the range of possible responses to the problem using
Chisholm’s (1973) terminology. Skeptics reject (C1*) and claim that the falsehood of (C1*)
implies that knowledge or justification is impossible. (One might, however, reject (C1*) but
avoid skepticism about knowledge or justification by claiming that knowledge or justification
does not require authorization.) Particularists and methodists reject (C2*). Particularists and
methodists agree that it is possible that a proposition P be authorized by a criterion C—they
accept (C1*)—and that this does not require that C be authorized by a criterion or supported by
an authorized proposition. Particularists, but not methodists, hold that an authorizing criterion
can be supported by an unauthorized proposition. Methodists, but not particularists, claim that a
criterion C can authorize a proposition without being authorized by a criterion or supported by an
authorized proposition. This exhausts Chisholm’s terminology for describing the possible
responses, so (C1*)–(C3*) show that additional positions need to be added to Chisholm’s list.
Coherentists reject (C3*) on the grounds that endless regresses with finitely many members—
circles—can result in both authorized propositions and criteria of truth that are authorized or
supported. Infinitists reject (C3*) on the grounds that non-circular endless regresses can result in
authorized propositions and criteria of truth that are authorized or supported. Dialethists—
persons who think that some contradictions are true—might hold that although (C1*)–(C3*) are
inconsistent, they are all true. Contrary to Chisholm, then, there are more than three possible
responses to the problem of the criterion. The present version of the problem of the criterion
improves, in many ways, on the version given in Cling 1994.

15 (2) presupposes that only propositions can be reasons, an assumption made throughout this
paper. This is controversial. Some philosophers believe, for example, that non-propositional
states—states of sensory appearing, for example—can be evidential reasons for belief. Cling
2008, pp. 408–412 argues that evidential reasons for belief not only must be propositions but that
they must have the same hyperintentional propositional content that beliefs have. The central
idea of the argument is that if reasons lack propositional content—indeed, if they lack the
hyperintentional propositional content that beliefs have—then there will be too many reasons.
For suppose that a property Φ is conceptually or nomologically identical to a property Ψ. Under
these conditions an allegedly non-propositional reason can exemplify Φ if, and only if, it
exemplifies Ψ, yet a person can have a reason for Φ-propositions but not for Ψ-propositions.
This can be explained only by supposing that reasons have the hyperintentional propositional
content that beliefs have. This argument generalizes to show that all reasons for belief are
propositional. Those who believe that there can be non-propositional reasons can take the
relevant conclusions of the present paper to be conditional: if all reasons are propositions,
then . . . Given the skeptical conclusion of this paper, such philosophers will have an additional
incentive to try to make accounts of non-propositional reasons work.

16 Van Cleve 1979, e.g. pp. 77–78 discusses several arguments like this. These arguments are
discussed in Cling 1997.

17 ‘Rxy’ = y is a basing reason for x; ‘NAx’ = x is non-arbitrary (in the relevant respect). The
arguments for (2B) given in this section assume that basing reasons are propositions. They show,
but do not assume, that reasons must be believed.

18 See Cling 2002. The conclusion of the argument in that paper is only that essentially circular
arguments can enhance the justification (better: evidential support) persons have for their
conclusions, not that persons can acquire justification (better: evidential support) for propositions
by means of essentially circular arguments.

19 This argument is a generalization of an argument against an infinitist account of evidential

justification that is presented, in much greater detail, in Cling 2004, pp. 118–120.

20 Peter Klein, the foremost defender of infinitism, thinks that infinite regresses require only the
second-order dispositions to form the relevant occurrent beliefs. See Klein 1999, pp. 308–309.

21 I am very grateful to Scott Aiken, Eric Loomis, Brian Martine, John Post, Ted Poston, Jeffrey
Tlumak, the Department of Philosophy at the University of South Alabama, and the participants
at the 2008 meeting of the Tennessee Philosophical Association for very helpful comments on
ancestors of this paper. I am especially grateful to my superb colleague Nicholaos Jones for his
very helpful comments at every stage of the paper, his sage instruction about the semantics for
quantified modal logic, and his Herculean last-minute proofreading of the entire paper that, in the
event, uncovered philosophical as well as typographical errors. I give special thanks to two
anonymous referees for this journal. One referee’s apropos “so, what?” to an ancestor of this
paper provoked me to think about what the similarities of the epistemic regress problem and the
problem of the criterion might mean. I want it on the record that the referee only caused, and did
not recommend, my move toward skepticism. I am also grateful to the editor, Paul Moser, not
only for supporting this paper, but for probing discussion and wise philosophical counsel given
over many decades. Because the task of setting me straight is endless, I alone bear responsibility
for the errors that remain.
Epistemic levels, theProblem of Easy Knowledgeand Skepticism
Tito Flores*
VERITAS Porto Alegre v. 54 n. 2 maio/ago. 2009 p. 109-129

Abstract– The Problem of Easy Knowledge has recently been defined as arising in two different
forms: one connected with some version of the epistemic closure principle and the other with
bootstrapping. This essay shows that these supposed two forms actually generate two quite
different problems. One of them is related with the supposed easy (unacceptable) acquisition of
first-order knowledge, and the other with the supposed easy (unacceptable) acquisition of
second-order knowledge, each of them requiring different solutions. It is also presented how
Infinitism, the view that the structure of justificatory reasons is infinite and non-repeating, can
deal with these two different challenges.
Keywords– Skepticism. Problem of the Criterion. Closure Principle. Problem of Easy
Knowledge. Epistemic Level Confusion.

The so-called Problem of Easy Knowledge has recently become a great focus of interest in
epistemology. This is so, mainly, I think, due to Stewart Cohen’s paper “Basic Knowledge and
the Problem of Easy Knowledge”.1 Since then, much has been said about this problem and
different solutions have been offered.2

Despite Cohen’s claims that “the Easy Knowledge Problem arises in two related ways”,3 there is
neither one problem arising in two distinct manners nor a new epistemological problem being
presented. Rather, what Cohen calls “the problem of easy knowledge” seems to be more the
result of a couple of misunderstandings, a combination of a high degree of epistemic level
confusion and the renaming of very ancient philosophical challenges.

In the course of this paper I will try to show that Cohen is dealing with not one, but two quite
different problems. One problem has to do with an unacceptable way to come to know that we
know things (and that looks pretty much like one alternative to resolve what in the pyrrhonian
tradition has been called “the problem of the criterion”). The other involves our capacity to
reason from appearances in order to establish what things really are.

1. Epistemic levels

Level confusion is not at all a new phenomenon in epistemology.4 I do not intend, though, to
examine it in detail. I’m concerned with epistemic levels only to the extent it may cause
difficulties for understanding the easy knowledge problem.

The first epistemic level, the level of first-order knowledge, is pertinent to our discussion as it
deals with the conditions according to which a person S can know that some proposition p is true.
This level could be understood as the most elementary one, the object of the larger part of
epistemological parley.

When we get through the first level we do no more than take into account the conditions
according to which a subject S knows that a proposition describing empirical events is true.
The first level in epistemology epitomizes the relation between a person’s beliefs about empirical
events and the content of these beliefs. In other words, the first level is confined on the affairs
concerning the suitable connection between the beliefs and the truth, regardless the believer’s
awareness of her believing, knowing or capturing the truth.

All Gettier cases, for example, are reduced to the epistemic first level since they are intended to
show nothing else than that mere true justified belief does not count as (first order) knowledge.
The Gettier cases deal with first order knowledge as they tackle only the subject’s oblivious
competence in avoiding the accidental coincidence between justification and truth, i.e. the
subject’s capacity of knowing some proposition that describes the world.

The second level, the level of second-order knowledge or meta-knowledge, on the other hand,
concerns the conditions according to which a person S can know that she knows some
proposition p is true. Now the picture changes radically. To begin with, one of the necessary
conditions for second-order knowledge is first-order knowledge since it is not possible for any
subject to know that she knows p is true if she doesn’t know p is true in the first place. The
requirements for second order knowledge may very well include some (or even all) conditions
for first order knowledge, but by necessity they will never be the same.

At the second level in epistemology we’ll emphasize the subject’s reflections on her knowing –
or believing – that a proposition describing empirical events is true.

One way to describe the second level in epistemology is to say that it embodies the relation
between a person who has beliefs about empirical events and her own believing. This relation
brings in the person’s reflection on her specific epistemic condition. The second level is, in one
way or another, related to the affairs concerning the connection between beliefs and truth as well,
but in this case regarding the awareness of the believer on her believing, knowing or capturing
the truth.

Another way to figure out the epistemological meta-level is to say that this level stands for the
justification of knowledge claims. So, when we talk about meta-justification, we are talking
about the justification of what we believe to be justified.

As a consequence, when we assume the conditions for first and second-order knowledge are
equal, or we do not see the distinction between these two levels, we incur in epistemic level
confusion.

The conclusion of any particular theory of knowledge may require a particular articulation
between first and second epistemic levels. This doesn’t mean, however, that we are free to mix
up the levels, since it is not clear at all that as soon as a person knows p is true she is
automatically in the position to know that she knows p is true. That’s one of the main reasons
why we always have to distinguish between levels clearly.

Also depending on theory, huge differences in terms of the attributed importance to the levels
will appear. Even while recognizing the possibility for second order knowledge, some
epistemological theories are consistent with the idea that second order knowledge is simply not
too important – at least, not essential. Accordingly, once the right conditions for first order
knowledge are determined, the foremost job in epistemology is done and all the rest can be kept,
with no cost whatsoever, in the realm of skepticism. These sorts of epistemological theories may
end up having no qualms to attribute first order knowledge to anyone capable of meeting these
conditions, like very young children, animals and virtually anything that may be capable of
believing.5

On the other side of the spectrum, some theories will state that if a person knows p is true but
doesn’t know that she knows p is true, she will find herself with empty hands. This is so,
according to these theories, because the important philosophical aspects of knowledge are linked
with some sort of appreciation that we know things. This way, knowing without knowing that
one knows amounts to pretty much nothing. Expectedly, these theories will not be eager to
concede knowledge to neither dogs nor children.6

Different theories of knowledge will define in their own terms not only its subject of analysis but
also the importance of theirs and other possible subject matters. In the same way a particular
theory will define its scope and determine what is philosophically relevant to be analyzed and
defined.

I believe that some of the recent epistemological debate about two senses of knowledge has to do
with the difference about epistemic levels7. It has been said that one sense of “knowledge”, less
philosophical and more mundane, could be completely analyzed in the terms suggested by some
externalist-type theories. In contrast, the sense of “knowledge” more articulated and intelligent,
the one menaced by skepticism, runs off from any externalist examination and could be grasped
only by internalist-type theories.

This debate shows, again, the importance of keeping the distinction between epistemic levels up
and clear. Even if we decide to break epistemology in two, each part dealing with a different
sense of knowledge, we still can’t forget that these two senses are but two distinct levels of
knowledge, one exclusively portraying the connection between one’s beliefs and the truth and
the other depicting the appreciation of the believer about this supposed connection.

It seems quite reasonable to me to believe that part of the misunderstandings regarding

skepticism – particularly pyrrhonian skepticism – are due to the fact the epistemic levels are not
clearly distinguished. As we’ll see a little latter, Sextus Empiricus himself seems to be one of the
main sources for epistemological level confusion. Once we realize that some of the skeptical
challenges are directed to meta-knowledge, as it seems to be the case with the problem of the
criterion, and others to first-order knowledge, as it seems to be the case with the epistemic
regress brought about by Agripa’s trilemma, we’ll not only have a better understanding over
skepticism, but also will be in a better position to deal with them.

2. The closure principle

As it was said at the beginning, according to Cohen “the Easy Knowledge Problem arises from
two related ways”.8 As I understand it, it means that there exist two ways to raise the same
problem. One of them is via bootstrapping; the other one is through the closure principle. We’ll
see that the problem with bootstrapping has to do with illegitimately acquiring second-order
knowledge. But first, we will examine what kind of problem, and what kind of easy knowledge,
may emerge from the closure principle.

When we consider the closure principle under the strict optic of epistemology, all we say is that
knowledge is “closed” under logical implication. This roughly means that if we know some
proposition p, and this proposition logically implies another proposition, q, we also know q. The
idea is that we can increase the set of proposition we know to the extent of what is implied by
this original set of known propositions.

I do not intend to offer a flawless formulation of the closure principle here. Rather, I want to
discuss a few points that will give us a clear idea about what is epistemologically important
about the closure principle. I think that, at least for now, more important than reaching an
unblemished presentation of the closure principle is to understand what is behind this principle
and why sometimes it is thought that we will be better off rejecting it.

In a first attempt, the closure principle could be expressed like this: if a person, S, knows p, and
p implies q, then S knows q.

Even though this first formulation captures the key idea behind the closure principle, we can add
to it some clause that closes the closure principle under the realm of S’s conscience. This way,
we can avoid unnecessary complications that we’ll go over briefly. So we can reformulate the
closure principle like this: if a person, S, knows a proposition p, and S implies q from p, then S
knows q.

Depending on the way we formulate the closure principle, we can make it either obviously false
or even more strikingly appealing. We can see that when we change the predicate that works as
its operator. If we say, for example: if p is true and p implies q, then q is true, the principle will
maintain the same structure as epistemic the closure principle but will appear to be much beyond
reasonable complains. This is so because nobody will be ready to deny that the predicate “truth”
is closed under logical implication.

However, if we change the epistemic operator, “to know” for “to believe” things will seem
completely different. If we formulate the principle like this: if a person, S, believes a proposition
p and p implies q, then S believes q, the closure principle will be bluntly false. This is so
because, as a matter of fact, our believing is not closed under logical implications.

It is possible to point out problems with closure based on the inapplicability of closure to
believing. For if knowledge requires, among other things, a belief and the believing is not closed
under logical implication, so knowledge cannot be closed under implication as well.

This is probably true, but if we continue in this direction we’ll end up lost in epistemologically
irrelevant problems. The same happens when we dispute over the fact that S must somehow
“see”, or perform, the logical implication between p and q. All these non-epistemic problems
with the closure principle can be settled by adding some restrictions and constraints over the
applicability of the principle. In my opinion, this is a fine and important task. However, we can
keep all these additional complications aside.

The important aspect is this: what is distinctively relevant about knowledge, that of which our
knowledge is made of, epistemic justification or whatever it may be, seems indeed to be closed
under logical implication. There is something deeply intuitive about the fact that if I have
epistemic justification for believing a proposition that describes empirical events – even if I do
not have the belief itself already formed – and this first proposition logically implies another one
of the same type, I do not lose any justification for believing in the implied proposition. In this
respect, the closure principle poses that, in the same way as truth, something important about our
knowledge is preserved along with logical implication.

It is worthwhile to mention that one reason that has been put forth to reject the closure principle
is the alleged fact that it would lead to skepticism. In the one version of Cartesian (or academic)
skepticism an instance of the closure principle is used as premise for the conclusion that
knowledge in unattainable.

The skeptical argument goes like this: 1) If I know some ordinary proposition p – that I’m in
Rosario, for example – then I know that q, a proposition that is inconsistent with p, but that could
be true regardless the empirical evidence I have for p – that I’m in Philadelphia dreaming that
I’m in Rosario, for example – is false; 2) I don’t know q is false – I don’t know that I’m not in
Philadelphia dreaming that I am in Rosario; therefore 3) I don’t know p – I don’t know I’m in
Rosario.

I will not discuss the skeptical argument in any detail now. I just want to point out that the first
premise is an instance of the closure principle and the skeptical argument can be easily
generalized in order to include any descriptive proposition like p.

Since the argument is valid, the only way to reject the skeptical conclusion established in 3 is to
refuse one of the premises. If the chosen premise is 2, the dispute will be directed over the
closure principle.

That strategy for dealing with this type of skeptical argument is one of the main reasons for
rejecting the closure principle. I think, however, that that strategy is misguided. This is so
because the closure principle is neutral regarding the skeptical conclusion. It can
straightforwardly be seen once we change the form of this kind of skeptical argument from a
modus tollens to a modus ponens. This way we’ll observe that the potentially skeptical premise is
2, not 1. Differently from the first premise, it is the second one that hides regulations on our
evidences to believe descriptive propositions, and it is this sort of evidential restrictions that,
bottom-line, makes the skeptical conclusion possible.

If we keep the first premise, and consequently do not quarrel over the closure principle, but
change the second one, we can have an argument like this: 1) If I know some ordinary
proposition p – that I’m in Rosario, for example – then I know that q, a proposition that is
inconsistent with p – hat I’m in Philadelphia dreaming that I’m in Rosario, for example – is false;
2) I do know that p is true; therefore 3) I do know that q is false – I do know that I’m not in
Philadelphia dreaming that I’m in Rosario.

The latter argument shows that the closure principle can be used for skeptical and non-skeptical
purposes as well. Hence, it is false that there is something in closure that motivates skepticism.

In the following discussion no details about the closure principle will be necessary. As we will
see, the closure principle has little to do with the kind of troubles Cohen envisions with his easy
knowledge problem and a preliminary understanding of the closure principle will be enough for
this purpose.

3. Closure and easy knowledge

Cohen’s argument against the closure principle has a form of a reductio ad absurdum. Since we
have unacceptable results from using the closure principle, we should reject it. According to
Cohen, the closure principle would help us to increase our knowledge however in an
unacceptable fashion, since the kind knowledge we can acquire from using the closure principle
is “easy”.

He describes his concerns as follows:

“If I know the table is red on the basis of its looking red, then it follows by the closure principle
that I can know that it’s not the case that the table is white but illuminated by red lights.
Presumably, I cannot know that it’s not the case that the table is white illuminated by red lights,
on the basis of the table’s looking red.”9

Apparently, the problem is that it would be all too easy to come to know that the table is not
white but illuminated by red lights on the basis of its red looking. The point for him, it seems so,
is to try to show that there is something wrong with the closure principle, something that would
allows us to know that the table is not white but illuminated by red lights on the basis of the
table’s looking red. It’s very important to notice that Cohen is arguing that the reason for
rejecting the closure principle is that this principle allows us to come to know the table is not
white but illuminated by red lights on the basis of its looking red.

This rejection of the closure principle is anchored on the indisputable fact that if the table were
not red but rather white illuminated by red lights, the table would continue to look red. In this
case, we would have a white table that looks red and, according to him, the closure principle
would allow us to conclude, on the basis of the red looking of the table, that the white table is not
really white but red.

What Cohen calls, then, “easy” is our knowledge that the table is not white but illuminated by
red lights, and what makes things appealing is that this easy knowledge can be attained because
of closure, a very popular principle in epistemology.

We should keep in mind at this moment that (1) Cohen thinks that the closure principle should be
rejected for it permits us to conclude that one table that looks red is not white but illuminated by
red lights; that item of knowledge is illegitimate easy knowledge and (2) whether Cohen is
misguided or not, that item of illegitimate easy knowledge is clearly first-order knowledge.

In “Closure Matters”,10 Peter Klein argues that the question is rather whether it is possible to
know that the table is red by reasoning from the table looks red, and not the following move –
which involves the closure principle – from the table is red to the table is not white but
illuminated by red lights.

If we consider the three following steps, (1) it looks red, (2) it is red, and (3) it is not white but
illuminated by red lights, the controversial move is from the first step to the second, and not from
the second to the third, the one that actually includes the closure principle.

This is so because if I can know that the table is red on the basis of its looking red, there is no
problem with the inference from “it’s red” to “it’s not white but illuminated by red lights”.
Closure just allows us to conclude: “the table is not white but illuminated by red lights” from
“the table is red”. The problem here is to decide if it is possible (and, of course, specify under
which conditions this would be possible) to reason from “the table looks red” to “the table is red”
and not the further step, when we use the closure principle to increase our knowledge, from “it’s
red” to “it’s not white but illuminated by red lights”.

If that is right, the only easy knowledge problem is that it is not easy to see what could be the
problem with closure in this context. The idea of a reductio argument showing that we have
problems with the closure principle ignores the most important point: the problem pointed out by
Cohen is rather related to deciding how much a theory will allows us to trust in the way things
appear to us.

Cohen’s idea, that the closure principle should be rejected because it would allows us to
conclude that something is not white but illuminated by red lights on the base of the red looking
of the thing, seems just wrong. This is so because closure does not permit that. Closure leads to
the conclusion that something is not white but illuminated by red lights from “the thing is red,”
not from “the thing looks red.” The epistemological discussion here has little to do with the
closure principle, but rather with whether we can know the table is red based on the red looking
of the table.

The only possible easy knowledge is that the table is red based on the table’s looking red.
Cohen’s insistence on the third part of the argument–the part that goes from “it’s red” to “it’s not
white but illuminated by red lights” – only shows that there may be something wrong with the
first step of the argument – the step that goes from “it’s look red” to “it’s red”. But this, again,
shows nothing epistemically wrong about the closure principle.

In pointing out that there should be something wrong with the closure principle in the acquisition
of illegitimate easy knowledge, Cohen somehow misplaced the real philosophical discussion
and, at the same time, he called attention to his easy knowledge problem. Despite of other
differences of opinions, the real disagreement between Klein and Cohen is much less about the
closure principle than it is about the first step in the reasoning, the one that goes from “it looks
red” to “it’s red”.
Klein claims that there is nothing incorrect, in any respect, with the closure principle. However,
much more interesting in this context is that he seems to think that the conclusion that something
is red on the basis of its red looks may be, under certain circumstances, just a fine way of
reasoning. The red appearance of a table can be seen, according to Klein, as a clue about its
redness, since, he claims, things that are red used to have this red appearance.11

It could be said that we may know the table is red on the basis of its looking red if we have no
defeaters. So, the red appearance of the table may be an appropriate way to come to know the
table is red, under certain circumstances – namely, no defeater circumstances. Klein assumes that
the red appearance of the table, under the no-defeaters circumstance, is a good indication of the
redness of the table – because red things have this “tendency” to appear red. How does he know
that red things have a tendency to appear red? Well, that’s another story. What is essential now is
that he is trying to provide a rationale according to which we can reason from “the table looks
red” to “the table is red”.

If we assume for a minute that Klein is in the right path when he says that the controversial move
from “it’s look red” to “it’s red”, but this move may not be problematic at all, it would be odd to
expect him to blame the closure principle for any sort of expansion of our knowledge. For if one
sees nothing absurd in going from “the table looks red” to “the table is red”, one will be ready to
accept the conclusion “the table is not white but illuminated by red lights”.

On the other hand, Cohen seems to think that we can’t possibly know that the table is red based
on its red appearance. He seems to be bothered with the fact that we have no reason to think that
red things have tendencies to look red. However, stressing the last part of the argument, the part
where the closure principle plays a role, can only make case for the implausibility of the first
step of the argument. The best that could be said would be something like this: “look, we can’t
accept to go from “the table appears red” to “it is red” because once we do that we may also go,
using the closure principle, to “the table is not white but illuminated by red lights”. But, again,
the last part of the argument is just a freeloading discussion derived from the main disagreement
exposed in the first part, the one in which we reach implications from “it is red”.

Thus, the real disagreement is about whether we can know things are based on the way they
look, a disagreement as old as philosophy itself, not about the closure principle. What Cohen
seems to do is to use the closure principle to try to make a stronger case about the impossibility
of knowing the table is red based on it’s red appearances.

It’s not my purpose to discuss here whether we can know how things are based on the way they
appear to us – that would be way too much for this paper. What is important at this time is to
realize that the so-called problem of easy knowledge neither involves closure nor presents any
epistemological novelty. There is nothing in the problem of easy knowledge that even calls for a
defense of the closure principle as well, since the epistemological question at stake does not
involve the closure principle. The only “easy” thing is that is too “easy” to know how things are
based on the way things look and not the logical consequences that follow from our knowledge
the way things are, which does involves the closure principle.
Wherever this discussion ends, what is decisive here is the fact that all knowledge implicated so
far cannot be considered, by any means, second-order knowledge. Even if the analysis of this
part of Cohen’s problem of easy knowledge is wrong, there is no possible way we could consider
the easily acquired knowledge in the case involving the closure principle second-order
knowledge. That is important for in what follows we’ll see that the other way the problem of
easy knowledge arises relates to a very different kind of knowledge, one kind that involves
knowledge of knowledge claims.

4. Bootstrapping and easy knowledge

The starting point of what is now called “problem of easy knowledge” was aimed to be a
criticism against some forms of externalism. This seems to have been one of Fumerton’s points
in “Metaepistemology and Skepticism” and, more recently, in Vogel’s “Reliabilism Leveled”.
We could say that in both cases the objection against externalism was based on the idea that
theories like Reliabilism would permit what was called “bootstrapping”.

Bootstrapping was considered then a strongly counterintuitive strategy. To sum up the idea,
bootstrapping serves to transform the belief that results from a process of belief formation into
the very proof of the reliability of that process.

We can start thinking of bootstrapping through the following example: Michael is driving his car
and then he watches the gas gauge’s needle. The gauge’s needle says that the tank is full. Then,
he comes to believe that the tank is full. This belief is based on a process supposedly quite
reliable, the gauge’s needle indicating that the tank is full. From his belief that the tank is full he
comes to believe that the gauge is reliable. Now, he is basing the belief about the reliability of
the gauge in another belief, which results from the pure observation of the gauge’s needle. In
doing so, he comes to know that he knows the tank is full, since he now knows the gauge is
reliable. The question is: what is the problem with that?

The problem is that the same strategy could be used even if the gauge was broken. Of course, if
the gauge were broken it would seem wrong to allow using a belief that is formed from what the
gauge’s needle indicates to support a belief about the reliability of the gauge. In both cases, when
the gauge’s needle is working well and when the gauge’s needle is broken, the strategy to
bootstrap is the same, but the results are quite different – in one case he knows that the tank is
full, but in the other he doesn’t. The problem with bootstrapping is that we will get the same
result “the gauge is reliable” independently of the reliability of the gauge. That is the reason we
must refuse bootstrapping. Thus, bootstrapping delivers second-order knowledge we actually do
not have.12

The gas tank example is original from Michael Williams’s “Unnatural Doubts”. The example
emerges precisely from a discussion about externalism. Williams states the same conclusion
although only thinking on externalism. According to him, “It is perfectly true that I do not learn
that the gauge is working properly simply by learning that the tank is full. But what I do learn,
that the tank is full, does not entail what I don’t, that the gauge is working properly. If the needle
is stuck on ‘F’, the tank may or may not be full. What is true is that, in particular circumstances, I
may have no way of distinguishing the case in which my belief that the tank is full amounts to
knowledge from that in which, even if it happens to be true, it does not”.13 So, the problem with
bootstrapping is that we justify a knowledge claim in a vicious way.

The feature that appears to permit bootstrapping would lie in the very core of reliabilism: we
could come to know some proposition, p, using some reliable method, m, with no necessity for
us to either know or to be justified in believing (or to have some justification for believing or
even identify the method used) the method used were in fact reliable.

In “Externalism and Skepticism”, Michael Bergman pointed out that bootstrapping could be a
problem not only to externalist theories like reliabilism, as was the original idea, but to internalist
theories as well. Cohen in his “Basic Knowledge and the Problem of Easy Knowledge” perhaps
made this point even more explicit.

Cohen claimed that the problem of easy knowledge threatens all theories that adopt a “basic
knowledge structure”. By Cohen’s definition, a basic knowledge structure is one that permits that
“a belief source can deliver knowledge prior to one’s knowing that the source is reliable and hold
that reliability knowledge is based on basic knowledge”.14 He then claimed that, in principle, a
Basic Knowledge Structure could shape not only externalist theories but also internalist ones.

This idea seems quite exciting. The intended criticism seems to be directed less toward some
possible distinction between internalism and externalism in epistemology and more toward the
following, and more interesting, question: should any epistemological theory assume that the
individual’s awareness of her own evidential basis plays a decisive role in the justification of
whatever beliefs the individual has?

One of Cohen’s merits is the way in which he unifies the criticism for both externalism and
internalism. According to him, many externalist as well as internalist theories do not accept what
he calls KR Principle. This principle states that: “a potential knowledge source K can yield
knowledge for S only if S knows K is reliable”.15

Of course, despite the vocabulary used, such a principle will suit any sort of internalist theory.
KR principle can be read like this: a potential evidence E can justify a belief B for S only if S
knows (or has evidence) E is good evidence for B. The main idea KR principle conveys is that
the mere fact that E is good evidence, or that source K is reliable, isn’t sufficient for S to be
justified, or to know, some proposition on the basis of that evidence or knowledge source. For
using a belief that results from certain process to prove the reliability of that process is equivalent
of using a belief p, that is justified by the reason r, to show that r is a good reason to believe p.
Such a move is possible because, according to some internalist theories, it is not necessary for S
to have justification to believe that r is, in fact, a good reason to believe p in order to have
justification to be believe p on the basis of r in the first place.

As I think is clear, when we have a true belief that results from reliable method of belief
formation, such belief can be prized as “knowledge”. And when we know the reliable method
used is in fact reliable, we know that we know that proposition. In a passage from Goldman’s
“Epistemology and Cognition”, that says, “to know that we know we would have to know that
we use reliable processes of belief formation”.16
In denying KR principle, a theory assumes a basic knowledge structure. Consequently, a
principle like KR would be assumed just for those theories according to which it is impossible to
be justified – or to know something – unless we have some justification that the reasons, or
methods of belief formation, we use are epistemically appropriate.

In this sense, the dispute over KR principle involves theories according to which we can know
things without knowing that we know and theories according to which we can’t know without
knowing that we know.

If that is right, what Cohen calls KR principle could be seen as the pyrrhonian assumption that
poses the problem of the criterion. If the problem of the criterion is “a meta-epistemological
problem concerning the justification of first order knowledge claims among disagreeing
disputants”,17 the KR principle seems to be a clear way to make this disagreement explicit. In
other words, when disputants disagree about the justification of first-order knowledge claims,
they disagree over the acceptance of a principle like KR.

It seems quite patent, then, that KR Principle and problem of the criterion have a meta-
epistemological nature. The problem of the criterion is a meta-epistemological problem, in the
sense that it is not a problem about whether we can know but rather it is a problem about whether
we can justify our first-order knowledge claims.

This important feature of the problem of the criterion was made clear by Chisholm, when he
revitalized it in contemporary epistemology some years ago. As it was considered in the
pyrrhonian tradition, Chisholm describes the problem of the criterion as one concerning
“decisions”. According to him, the problem of the criterion was “how do we decide, in any
particular case, whether we have a genuine item of knowledge?”18. The problem of the criterion
is not about the specification of the necessary and sufficient conditions for knowing. Again, the
trouble is rather about whether we can know that we know.

After all, the meta-epistemological nature of the problem of the criterion and, consequently, of
the bootstrapping problem–given that the former seems to be a predicament derived from an
alternative response to the problem of the criterion – is in perfect harmony with the original
inspiration of the criticism against reliabilism. For Vogel, for example, “[a form of reliabilism
that he called Neighborhood Reliabilism] is too weak. It would allow us higher-level knowledge
we do not have”.19 If this is right, the bootstrapping problem must be understood as the problem
of the easiness of a certain strategy of acquiring second-order knowledge.

So, it is correct to say that bootstrapping problem is a consequence of a certain strategy to deal
with the problem of the criterion. More specifically, and using Cohen’s terms, the bootstrapping
problem is a consequence of trying to resolve the problem of the criterion avoiding the KR
principle by means of a basic knowledge structure view.

In this sense, the problem with bootstrapping emerges when using a certain strategy to resolve a
second-order problem, which could be well represented by the knowledge of the reliability of a
knowledge source, provided that knowing the source used to come to know a proposition p is
reliable means knowing that we know p.

When we bootstrap we manage a certain way to come to know that we know – e.g. to come to
know we use a reliable process. The relation between the bootstrapping problem and the problem
of the criterion is fruitful because it reassures the meta-epistemological character of the problem
of easy knowledge. Out of this relation we can then better understand what kind of knowledge is
easy knowledge.

In order to notice why bootstrapping is a strategy to deal with the problem of the criterion, we
have to understand what is the pyrrhonian view on bootstrapping.

According to Sextus:

“For the proof always requires a criterion to confirm it, and the criterion also a proof to
demonstrate its truth; and neither can a proof be sound without the previous existence of a true
criterion nor can the criterion be true without the previous confirmation of the proof. So in this
way both the criterion and the proof are involved in the circular process of reasoning.”20

Granted, we can’t accept anything when its proof hangs on circular reasoning. And what seems
to be the main problem here is the impossibility of establishing a criterion using whatever the
criterion itself proves to be the case. To judge our knowledge claims, we ought to have a
criterion – in order to know whether the belief B is a genuine item of knowledge, we need a
criterion according to which we will decide such matter. We can know that we know p if we
know that the belief formation process is in fact reliable. But how do we know the belief
formation process is in fact reliable? If we answer this question pointing out the belief B, a belief
that is sanctioned by the criterion as a case of knowledge, we will be caught into the diallelus.

On the other hand, the typical circular reasoning is meant to be close up under the same
epistemic level (as it is showed, for example, by another pyrrhonian argument, Agripa’s
trilemma or the infinite regress argument). The main idea of Agripa’s trilemma is that in order
for S to have justification to believe Bon the basis of a reason R, S will also have to have
justification to believe R. Granted, Band R must be related in a very distinctive form so R can be
a proper epistemic ground for B. Yet nothing is said, at least in this kind of regress, about the
necessity for S to have any reason to believe that R confers a proper epistemic ground for B.

The circle shaped by Agripa’s trilemma materializes when I use a reasoning that goes like this:
my justification to believe B is R; my justification to believe R is S and my justification to
believe S is B. This way I’m looping inside the same epistemic level.

In the bootstrapping case, the kind of circularity involved has to do with proving the reasons
used are appropriate, and in this way we jump up between levels. In this meta-regress, I use my
belief B, which is epistemically grounded by my other belief R, to prove that R is a proper
epistemic base for B.
As it looks, there are two types of circular reasoning, a first-order circular reasoning and a
second-order circular reasoning. In order to understand the difference between them we have to
remember the epistemological level distinction. Albeit both types have the same characteristic of
avoiding the onus of proving the premise basing it on the previous acceptance of the conclusion
– what makes them fallacies of petitio principii – the result of each circular reasoning will
change. In one case, my belief B is part of a reasoning chain that justifies B itself; in another
case, B is used in a reasoning chain that is intended to demonstrate that the grounds for Bare
epistemically suitable. In one case B is used to provide justification for itself; in the other, B is
used to provide meta-justification for itself.

The circular reasoning that has to do with the problem of criterion – and, hence, bootstrapping –
is, lets say, vertical. It relates to a movement from first to second level. In this form of circular
reasoning the circle goes from first to second level, and we try to establish that we have a
genuine item of knowledge using a criterion that is proved by a knowledge (claim) that was, in
turn, legitimated by the criterion itself.

This significant difference about the two types of circularities is not detected by Sextus. That’s
why he is one source for level confusion. Even though he provides, two distinct arguments, he
fails to notice the difference – at least, he falls short of making it unambiguous.

A different way to understand these two different circular reasoning is to consider Fumerton’s
Principle of Inferential Justification (PIJ): to be justified in believing one proposition p, on the
basis of another proposition E, one must be (1) justified in believing E and (2) justified in
believing that E makes probable p.21

There would be many things to say about this principle. However, the idea now is just to give an
example of how create two different kinds of epistemic regresses that, by consequence, can
create two different kinds of circular reasoning.

According to PIJ, in order to have an appropriate evidence to believe B, or to have justification

to believe B, we must have to deal with two different chains of reasoning: one implied by the
first clause, that will trigger a first-order regress, and another implied by the second clause, that
will start a second-order, or meta-regress.

One noticeable aspect of PIJ is that it requires meta-justification as a condition for first order
justification. Thus, in order to have justification to believe B we’ll altogether need some sort of
meta-justification for B.

The first clause of PIJ triggers a chain of reasoning such as Agripa’s trilemma. In order to have a
justified belief that owes its justification to some reason – or evidence, etc. – we’ll be forced to
go further ad infinitum, or reasoning in a circle, or stopping at a certain point. All of these
possibilities, though, are inserted in the same epistemic level.

The second clause, however, sets off a slightly different array of possibilities. The type of regress
we deal with then, the regress of meta-justification, forces us not to put forward reasons for the
beliefs we have, but reasons for the epistemic quality of our reasons.
Thus, in the same way we can create either an infinite, more complex, regress of meta-
justification or an infinite, less complex, regress of justification, we can create either a circular
reasoning of justification or a circular reasoning of meta-justification. Again, in one case we are
offering reasons to believe a certain proposition; in the other we are presenting a reason that we
do have good reasons to believe what we believe.

So, what has been called “bootstrapping” is but an attempt to justify knowledge claims through a
form of a circular reasoning that switches from the first to the second epistemic level. That’s why
bootstrapping is one strategy – unacceptable for both Cohen and the pyrrhonians – to deal with
the problem of the criterion. Of course, any explanation we may find of bootstrapping is much
sharper and sophisticated than anything we can find in Sextus Empiricus. Then again, this does
not opaque the fact that circular reasoning for the problem of the criterion and bootstrapping are
the same thing.

On the same token, what Cohen calls easy, the pyrrhonian see as plainly inadequate. The only
way to understand the meaning of “easy knowledge” is that it designates the inappropriateness of
one sort of knowledge acquisition. The idea of “easy knowledge” hides the conception that this
“easy” means “too easy,” “unbearably easy.” In one word, “unacceptable.”

Summing things up: when we talk about the reliability of a belief process formation, we could
bootstrap in order to come to know that the process is reliable; when we talk about reasons, we
could bootstrap in order to come to know that we have adequate reasons to believe the things we
do. In both cases the problem seems to reside on the limitation of the requirements of what
would be necessary in order to be justified in believing some proposition. Exactly because it is
not required for S to have any clue about the epistemic quality of the reasons (or methods) she
uses to believe the things she does, she will be able to use the epistemic good (or bad!) resultant
beliefs she has to prove the epistemic quality of the reasons or (methods) she uses to believe.
And this may be done independently of the fact that the reasons are adequate or the methods are
reliable. That’s the problem with what Cohen calls basic knowledge structure views. The
skeptical conclusion would be that just a reliable method, or just a proper reason, seems not to be
enough to deliver justified beliefs.

Whether we talk about either reasons or methods of belief formation, we will be trapped into a
circular form of reasoning that establishes unacceptable bridges between first and second
epistemic levels.

There are a number of forms to manifest dissatisfaction with this kind of restrictions. I think it is
exactly what Lehrer was saying with “examples of alleged knowledge in which a person does
not know that the information he accepts is correct my be of some philosophical interest but such
knowledge falls outside the concern of knowledge used in a way that is characteristically human
in critical reasoning and the life of reason”.22

In some sense, the problem of the criterion and bootstrapping can be seen as demonstrating what
is wrong with this sense of “knowledge” when the person ignores the information accepted is
correct – or ignores that the reason is epistemically good, or that the process is reliable.23
I’ll repeat, it doesn’t matter if the processes are in fact reliable or if the reasons are in fact good.
One could think that using the results of some process to prove the reliability of it is legitimate
just in case the process is in fact reliable. Such strategy fails because it is in itself inadequate, and
this is so because such a strategy is trivial, it will always give the same result: the processes are
reliable, even when they aren’t.

So, this is the problem of bootstrapping. There are theories that independently of their
internalistic or externalistic brand will face the problem of allowing this kind of meta-circularity,
the bootstrapping strategy, and bootstrapping has to do with acquiring second-order knowledge
through a form of reasoning that inadequately connects first and second epistemic levels.
If we think about Goldman’s suggestion that “a plausible theory ought to have the property that
knowing that one knows is more difficult than simply knowing”,24 we will find bootstrapping, or
circular reasoning, as a way to make second-order knowledge somehow less difficult than first-
order knowledge.

Conclusion

One optimistic way to describe what we have seen would be to say something like this: instead
of one problem that emerges in two different forms, Cohen has two quite different problems at
stake when he discusses the easy knowledge problem. One of them is the meta-epistemological
question posed by the KR principle, which is represented by the inadequate acquisition of
reliability knowledge through bootstrapping. The other is the supposedly inadequate knowledge
acquired through the closure principle.

Accordingly, the difference between closure and bootstrapping is a fundamental one: closure
does not motivate any “easy” second-order knowledge acquisition. He apparently failed to notice
that there exist two different things subsumed in what he called “the problem of easy
knowledge.” Differently from closure, bootstrapping is a way to come to know that our belief
sources are reliable, that our reasons are epistemically appropriate. Then, it has a meta-
knowledge nature. This difference is crucial.

The closure principle does not motivate any easy knowledge problem as bootstrapping does. In
the worst hypothesis, closure is related to the problem of deciding if we can reason like this: it
seems red, thus it is red. Bootstrapping, on the other hand, is related to the problem of coming to
know that we know in a way in which we can reach the conclusion that we know that we know
even if we do not know in the first place. Those are quite different problems.

Cohen – and many epistemologists after him – does not observe the important difference
between the two supposed ways in which the problem of easy knowledge arises. He didn’t notice
that there exists a difference regarding epistemic levels. If this is right, there is some level
confusion lurking in his talk about the easy knowledge problem. He should have observed that
there were two different problems – one related to the easy acquisition of first-order knowledge
by closure, and the other one related to the easy acquisition of second-order knowledge by
bootstrapping – that require two quite different analyses.
On a less optimistic description – the one I think is correct – we’ll have to say that there is no
easy knowledge problem. To begin with, there is no problem at all with the closure principle.
Thus, the closure principle does not motivate any easy knowledge problem. The real problem
Cohen is dealing with is to decide – how, how much, to what extent, etc. – the appearances of
things can be the base for our knowledge of the world.

In the bootstrapping case, Cohen is, again, battling against a very old and complicated problem:
namely, the problem of the criterion. However, differently from the case about the closure
principle, his description of the problem seems to be quite interesting. Nevertheless, he is
mistaken to imagine a new problem, the easy knowledge problem.

* Doutor em Filosofia pela Pontifícia Universidade Católica do Rio Grande do Sul.

1 See Cohen, Stewart. “Basic Knowledge and the Problem of Easy Knowledge”. In: Philosophy
and Phenomenological Research , 65 (2002), p. 309-329; “Why Basic Knowledge is Easy
Knowledge”. In: Philosophy and Phenomenological Research, 70 (2005), p. 417-430.
2 E.g. Markie, Peter. “Easy knowledge”. In: Philosophy and Phenomenological Research, 70
(2005), p. 406-416; Neta, Ram. “A contextualist solution to the problem of easy knowledge”. In:
Grazer Philosophische Studien. Internationale Zeitschrift für Analytische Philosophie, 69
(2005), p. 183-206; van Cleve, James. “Is knowledge easy – or impossible? Externalism as the
only alternative to skepticism”. In: Luper, Steven (ed.). The skeptics: contemporary essays, 2003,
p. 45-49; Black, Tim. “Solving the easy knowledge problem”. In: The Philosophical Quarterly,
58 (2008), p. 597-617.
3 See Cohen, Stewart, 2002, p. 312.
4 See Alston, William. “Level confusions in epistemology”. In: Midwest Studies in Philosophy,
(1980), p. 135-150.
5 See Fred Dretske, Perception, Knowledge and Belief, 2000, p. 23.
6 E.g. Peter Klein, Human Knowledge and the Infinite Regress of Reasons, in: Philosophical
Perspectives, 13 (1999), p. 297-325; and Keith Lehrer, Theory of Knowledge, 2000, p. 55.
7 See Ernest Sosa, Two False Dichotomies: Internalism/Externalism and
Foundationalism/Coherentism, in: Walter Sinnott-Armstrong (ed.), Pyrrhonian Skepticism, 2004,
p. 146-160.
8 See Stewart Cohen, 2002, p. 312.
9 See Stewart Cohen, 2002, p. 310.
10 Cf. Peter Klein, Closure Matters: Academic Skepticism and Easy Knowledge, in:
Philosophical Issues, 2004, p. 165-181.
11 Id. ibid. p. 170.
12 For a detailed version of bootstrapping cf. Tito Flores, Infinitism, knowing that one knows
and the problem of the criterion, in: Veritas, 50 (2005), p. 109-128.
13 Cf. Michael Williams, Unnatural doubts, p. 347.
14 Cf. Stewart Cohen, Why Basic Knowledge is Easy Knowledge, op. cit., p. 311.
15 Cf. Stewart Cohen, Basic Knowledge and the Problem of Easy Knowledge, op. cit., p. 309.
16 Cf. Alvin Goldman, Epistemology and Cognition, p. 56.
17 Cf. Roberto Amico, The Problem of the Criterion, p. 143.
18 Cf. Roderick Chisholm, The Foundations of Knowing, p. 62.
19 Cf. Jonathan Vogel, Reliabilism Leveled, in: Journal of Philosophy, p. 612.
20 Cf. Sextus Empiricus, Outlines of Pyrrhonism, 1.114-117.
21 Cf. Richard Fumerton, Metaepistemology and Skepticism, p. 36.
22 See Keith Lehrer, op. cit., p. 41.
23 I guess to Lehrer the question is a bit different, since he seems to reject the very idea that we
could gain knowledge in the way described: “if I read some gauge or meter and believe the
information I received, though I have no idea whether the instrument is functioning properly, I
may thus acquire information, but this is not knowledge.” See Keith Lehrer, op. cit., p. 6.
24 See Alvin Goldman, op. cit., p. 57.
A Quintet, a Quartet, a Trio, a Duo? The Epistemic
Regress Problem, Evidential Support, and Skepticism
Timo Kajamies
Department of Philosophy, University of Turku, Turku, Finland

Abstract In his topical article, Andrew Cling claims that the best extant formulation of the so-
called epistemic regress problem rests on five assumptions that are too strong. Cling offers an
improved version that rests on a different set of three core epistemic assumptions, each of which
he argues for. Despite of owing a great deal to Cling’s ideas, I argue that the epistemic regress
problem surfaces from more fundamental assumptions than those offered by Cling. There are
ultimately two core assumptions—in fact two contradictory strands within the concept of
epistemic support—which jointly create a powerful challenge for our pursuit of paramount
epistemic values.
Keywords Epistemic regress problem. Epistemic justification. Epistemic support. Skepticism

The Epistemic Regress Problem: an Outline

Hankinson (1995, 189) writes as follows:

Consider a train of infinite length, in which each carriage moves because the one in front of it
moves. Even supposing that fact is an adequate explanation for the movement of each carriage,
one is tempted to say, in the absence of a locomotive, that one still has no explanation for the
motion of the whole. And that metaphor might aptly be transferred to the case of justification in
general.

Hankinson’s metaphor is fairly easy to grasp. Suppose I sit on a carriage whose apparent
movement I want to explain. I don’t know how to do it, so I ask my fellow traveler. ‘Oh, you
need to know why this carriage is moving? Well, the carriage in front of this carriage is moving,
so this one moves as well.’, she tells me. I ask another question, and I get a new answer
pertaining to the movement of yet another carriage. After the twentieth question and answer I
express my frustration: ‘Now what the heck is this? Always another carriage!’ And then she tells
me: ‘I understand that you are a bit bewildered, but rest assured, we are moving. You see, we are
riding an endlessly long train, so there is no locomotive to move this baby.’ I get the picture now;
the movement of each carriage is conditional on the movement of the next carriage. And I think
to myself: ‘Very well, this carriage gets its movement from the next carriage—which gets its
movement from the next carriage—which gets its movement from the next carriage, and so on
with no end to it. But really, where does that movement come from in the first place? What we
would really need is a locomotive, even if this train extended right up to Mizar.’ My fellow
traveler then interrupts my contemplation: ‘Remember, this is an infinitely long train.’ Not
wanting to challenge her any further, I decide to go to sleep.

Claiming that there is an analogy between trains and epistemic justification of course is not
enough.1 The issue with respect to epistemic justification must be studied in its own right. Here
is Aikin’s (2005, 191) rough characterization of the problem:
Insofar as we strive to be rational, we strive to believe on the basis of good reasons. For those
reasons to be good, they must not only support our first belief, but they themselves must also be
believed for good reasons. This is where we begin to see a disturbing pattern. If that first belief is
to be held on the basis of good reasons, it seems we are in need of a very long chain of reasons.
This is a rough and ready picture of the regress problem. It seems endemic to the project of
believing on the basis of reasons. And thereby, it seems endemic to the very project of being
rational.

While pursuing truth in a rational way, we adopt an evidentialist outlook, according to which
belief requires good reasons. However, believing on the basis of good reasons threatens to launch
an infinite regress of reasons. We get caught up in a trouble, since there is something very
suspect about an infinite regress of reasons. Therefore, the project of being rational becomes
problematic.

Why does a rational pursuit of truth pull us towards endless sequences of supporters? And, even
if I was troubled on the infinitely long train, why would an infinite sequence of supporters be so
very problematic? In order to answer questions such as these, we need to get beyond a rough and
ready characterization of the epistemic regress problem. We need to discover the key factors
involved.

The Quintet

According to Cling, the crux of the epistemic regress problem is a set of core epistemic
assumptions that are jointly inconsistent. The problem thus is a paradox; the assumptions are
individually plausible but jointly inconsistent. Now, the question naturally arises as to what
exactly these assumptions are. Cling (2008, 401) provides two answers to this; one that he calls
“the best extant statement of the epistemic regress problem” and another he calls “the improved
version”. The best extant statement consists, he argues, of five assumptions (R1)–(R5) that are
jointly incompatible but, Cling claims, too strong. Let ‘Sxy’ mean ‘x is supported by y’; ‘Jx’
mean ‘x is justified’; and—condensing Cling’s discussion a bit—‘IRJxy’ mean ‘x and y are the
first members in an infinite sequence of justified propositions, the sequence being such that its
every member is supported by its successor’ (Cling 2008, 402–403). We then get the following
inconsistent quintet of assumptions (Cling 2008, 403–404).

The Best Extant Statement of the Epistemic Regress Problem:

(R1) Justification Requires Justified Support.

(∀x)(Jx → (∃y)(Jy & Sxy)).

(R2) Some Proposition is Justified.

(∃x)Jx

(R3) Support is Irreflexive.

~ (∃x)(Sxx)

(R4) Support is Transitive.

(∀x)(∀y)(∀z)((Sxy & Syz) → Sxz).

(R5) No Infinite Regress of Justified Propositions.

~(∃x)(∃y)IRJxy.

Cling shows that the assumptions are jointly inconsistent. However, he claims that these
assumptions are too strong, and that the formulation of the epistemic regress problem can be
improved by enunciating a more meager set of assumptions. (Cling 2008, 405.)

The Trio

Cling offers an improved version of the epistemic regress problem. Having already specified
‘Sxy’, let ‘ERSxy’ mean—condensing Cling’s discussion a bit again—‘x and y are the first
members of an endless sequence of propositions, the sequence being such that its every member
is supported by its successor’ (Cling 2008, 402–403). We then get, Cling (2008, 404–405) states,
the following inconsistent trio of assumptions.

The Improved Version of the Epistemic Regress Problem:

(1) Reasons are Supported.

(∀x)(∀y) (Sxy & (∃z(Syz)).

(2) No Proposition is Supported only by Endless Regresses.

(∀x)[(∀y)(Sxy → ERSxy) → ~(∃z)Sxz].

(3) Some Proposition is Supported.

(∀x)(∀y)Sxy.

According to Cling, each of the assumptions (1)–(3) is very plausible. Hence, they suffice to
create “a powerful challenge to the consistency of some of our core assumptions about evidential
support” (Cling 2008, 420).

I do not enter into the examination of Cling’s (2008, 405–408) arguments in favor of the trio. All
in all, I agree with Cling that all three assumptions are initially very plausible, and if they are
jointly inconsistent, a paradox definitely ensues. Instead, I shall concentrate on the view that (1)–
(3) are inconsistent. Here is Cling’s (2008, 405) argument for the inconsistency claim:
Assumptions (1)–(3) are jointly inconsistent. Reasons are Supported implies that any proposition
P0 is supported by a proposition P1 only if P0 and P1 are the first two members of an endless—
infinite or circular—regress of reasons. Given No Proposition is Supported only by Endless
Regresses it follows that that [sic] no proposition is supported. This contradicts Some
Proposition is Supported.

The first step in Cling’s argument, i.e. that (1) launches an endless regress, is very interesting.
Someone might find it questionable. Suppose we have a proposition P0 which is supported by
another proposition P1. Hence, SP0P1. Now, there has to be a proposition that confers support
on P1: it is true by (1) that SP0P1 → (∃x)(SP1x). And hence, there has to be yet another
proposition P2, such that SP0P1 and SP1P2. And of course, by (1), SP0P1&SP1P2 → (∃x)
(SP2x), and again it seems to then follow that SP0P1 and SP1P2 and SP2P3. It is easy to see that
an endless regress ensues. But, one might ask, is it ultimately correct to state that SP0P1 implies
that there has to be yet another proposition P2, such that SP0P1 and SP1P2? What about a
situation such as SP0P1 and SP1P1? Isn’t it in full harmony with (1)? If the train metaphor is
allowed, we now have a locomotive that moves the whole train, including the locomotive itself.
And one might point out that a train having a locomotive is not an endless train; why should a
situation in which a proposition supports itself launch an endless sequence of supporters?

One might thus claim that (1)–(3) are consistent if one does not assume that the relation of
evidential support is irreflexive. One might maintain that some proposition is supported in a way
that does not launch an endless regress of supporters, for the regress is prevented by there being a
self-supporting proposition down the line.2 And, the critic might continue, a contradiction
follows only if the irreflexivity assumption is borrowed from the quintet of too strong
assumptions and plugged into the improved version, thus expanding the trio into a quartet. Only
if a case of evidential self-support launches an endless regress, (1) launches an endless regress.
We shall see that there is a compelling reason to think that it does, but let us first see how the
irreflexivity assumption surfaces again within the argument that foundationalism does not
provide a satisfactory solution to the regress problem.

Foundationalism and the Improved Version

Cling argues that the improved version of the epistemic regress problem is powerful enough to
resist a foundationalist solution. However, Cling’s argument is convincing only if irreflexivity of
evidential support is assumed. Here is why.

According to Cling, foundationalism minimally holds that “(1) some propositions are basic
because they are justified without the support of other propositions and that (2) any non-basic
proposition is justified only if it is supported by a basic proposition, or supported by a
proposition that is supported by a basic proposition, and so on.” Furthermore, Cling holds that
foundationalists are forced to reject Reasons are Supported because they hold that basic
propositions can be reasons that are not supported by other propositions. Now, the rejection of
Reasons are Supported, according to Cling, commits the foundationalist to the view that basic
propositions are either unsupported or supported by some non-propositional states. (Cling 2008,
408.) Unsupported propositions, Cling (2008, 409) holds, are epistemically arbitrary, and non-
propositional states, Cling (2008, 409–411) argues, are insufficient for support. Hence, the
foundationalist is not able to solve (the improved version of) the epistemic regress problem.

Cling’s argumentation against foundationalism includes a move that someone might find
questionable. Why should the rejection of Reasons are Supported follow from the foundationalist
thesis that some propositions are justified without the support of other propositions? The
rejection follows only if justification without the support of other propositions boils down to a
lack of support. And it does boil down to a lack of support only if propositions cannot be self-
supporting. To use the train metaphor once again in order to make the case more vivid, if there is
a locomotive—understood as a carriage that moves itself—there is a carriage that moves without
a further carriage moving it. This train does not violate the principle that all moving carriages are
being moved.

According to the present criticism, irreflexivity of evidential support is required again.

Therefore, Cling’s argument to the effect that foundationalism cannot solve the epistemic regress
problem is convincing only if an assumption can be borrowed from the quintet of too strong
assumptions, making a quartet out of a trio. Or else, there should be an argument which does not
involve invoking the irreflexivity assumption but which, however, blocks out cases of self-
support.

Piggybacking

Supposing that we are ready to accept the trio, is it fair to borrow assumptions from the quintet?
Cling’s (2008, 402) statement that “the regress problem is not just a problem of knowledge and
justification, it concerns evidential support, a more basic epistemic value” arguably involves the
idea that the quintet is too strong precisely because it involves appealing to a notion that is not
needed for the formulation of the problem. That is, the quintet would be too strong because it
involves appealing to the notion of justification, even though the problem can be formulated just
by considering the relation of evidential support and its characteristics. The trio would, thus, well
involve implicit accompanists borrowed from the quintet, as long as justification keeps away.

Which assumptions can be borrowed, then? The assumptions (R3) and (R4) will presumably
qualify, since they just express the logical properties of evidential support, and contain no trace
of the notion of justification. However, Cling (2008, 415) says that “no consistent subset of (1)–
(3) implies Support is Transitive.” Hence, Cling wants to exclude (R4) from the trio. But if he
does this, what reason could there be to borrow (R3) from the quintet? For no consistent subset
of (1)–(3) implies Support is Irreflexive either.

It may well be the case that there is a compelling reason to accept the thesis that evidential
support is irreflexive.3 But if the irreflexivity thesis is assumed, it should be as convincingly
argued for as the rest of the assumptions in the improved version of the epistemic regress
problem. Even if Cling does not offer a direct argument in favor of the irreflexivity thesis, his
views help to construct a defense of it. The defense stems from the view that genuine evidential
support is ultimately unconditional. As a result, we reach a very fundamental level of epistemic
intuitions—a deeper level than that on which Cling’s trio performs.
Genuine Evidential Support

The notion of arbitrariness is crucial for Cling’s arguments against a foundationalist approach
(unsupported basic propositions are arbitrary), coherentist approach (circular support is
arbitrary), and a contextualist approach (the foundations of inquiry are unsupported and hence
arbitrary) to the epistemic regress problem. Cling does not offer a clear cut characterization of
arbitrariness, but one of his key ideas seems to be that arbitrariness has to do with conditionality.
For instance, when he argues against coherentism, he states that members in a circular sequence
of purportedly support-ordered propositions are “arbitrary: I have no unconditional reason to
believe any of them” (Cling 2008, 407). The members in such sequences are not supported, as
Cling (ibid.) writes, per se.

So, Cling holds, I believe incorrectly, that if there is no unconditional reason to believe a
proposition, the proposition is arbitrary. If Cling was right, support in an purportedly support-
ordered endless sequence of propositions would be hopelessly arbitrary, since no proposition in
such sequence is unconditionally supported. Nevertheless, Klein, for one,4 maintains that of all
theories it is precisely infinitism that avoids arbitrariness. Klein’s “Principle of Avoiding
Arbitrariness” (1999, 299) may be put by stating that if P0 is supported, it has a supporter P1
which has a supporter P2, and so on ad infinitum. No proposition in the infinite sequence is
without a reason, so no proposition is arbitrary. But if one accepted P0 without any reason, i.e.
that P0 would get its support neither from itself nor from any further proposition, P0 would be
arbitrary precisely because one seems to be no better off than if one accepted ~P0 without any
reason. It is therefore far from clear that conditional support implies arbitrary support.5

Nevertheless, purportedly support-ordered endless sequences of propositions still face a serious

problem. The problem is precisely the conditional character of support, not its arbitrariness. In
order to get a clearer view of the issue at stake, let me enunciate a distinction between incurably
conditional (and therefore spurious) and genuine (and therefore ultimately unconditional)
support. As understood here, the support for a proposition is incurably conditional when there is
no unconditional reason to believe the proposition. For instance, the support for proposition P0 in
the purportedly support-ordered sequence S1=〈P0, P1, P2…〉 is incurably conditional provided
the sequence has no unconditionally supported proposition down the line. The sequence must
have such unconditionally supported proposition in order that the
support for P0 is ultimately unconditional.

There is something very appealing about the view that incurably conditional support is not
genuine support. If the support for each proposition Px in the sequence S1 is incurably
conditional, i.e. if its each proposition is conditionally supported by its successor, no satisfactory
answer can be reached as to whether these conditions are satisfied, even in principle. This idea, I
believe, is reflected in Alston (1989) criticism of an infinite sequence of mediately justified
beliefs:

If there is a branch [of mediately justified beliefs] with no terminus, that means that no matter
how far we extend the branch the last element is still a belief that is mediately justified if at all.
Thus, as fas as this structure goes, whenever we stop adding elements we have still not shown
that the relevant necessary condition for mediate justification of the original belief is satisfied.
I believe we have a deep intuition according to which the question ‘Is Pi supported?’ cannot be
satisfactorily answered by ‘Yes it is if Pj is supported’. We should be able to reach (at least in
principle even if no one ever traversed their reasons for Pi) a ‘Yes’ or a ‘No’, but no such answer
is reachable if the support for each Px is incurably conditional. It is reasonable to suppose that
having a reason for Pi does not entail traversing those reasons (see Klein 1999, Aikin 2005), but
if support for Pi is incurably conditional, there is no unconditional reason there to be traversed
even in principle. Our deep intuition that genuine epistemic support is ultimately unconditional
may of course be false (and hence an infinite series of supporters could provide genuine support),
but this intuition, I claim, nonetheless is of utmost importance in explaining why we think there
is such a problem as the epistemic regress problem.

Notice that the view according to which incurably conditional support is not genuine dictates that
self-support is not genuine. As already seen, self-support satisfies Reasons are Supported. Self-
supporting propositions are not unsupported, but supported by themselves; there is always a
proposition supporting a self-supporting proposition, namely that proposition itself. Hence, a
purportedly support-ordered sequence that has a self-supporting proposition down the line can be
represented as an sequence S2=〈P1, P2, …, Pn, Pn, Pn, …〉—a sequence that winds up endlessly
repeating Pn. And it is easy to see that the support for the proposition Pn is not genuine precisely
because the support is incurably conditional. Ask the question ‘Is Pn supported?’, and you don’t
reach a ‘Yes’ or a ‘No’. The best that can be said as an answer is ‘Yes if Pn is supported’, and
this is not satisfactory. Therefore, if incurably conditional support is not genuine support, self-
support is not genuine support either.6

Supposing, as I do, that one of the desiderata operating here is the unconditionality of epistemic
support, we may add a fourth assumption to Cling’s trio and see where it takes us.7

(1) Reasons are Supported

(2) No Proposition is Supported only by Endless Regresses
(3) Some Proposition is Genuinely Supported
(4) Incurably Conditional Support is not Genuine Support

For the time being, let us not worry over whether there are redundant assumptions in the above
list. We at least see that the quartet is inconsistent. According to (3), there is a proposition Px
that is genuinely supported. There is, then, a proposition Py that confers support on Px. And
according to (1), Py has a supporter Pz which, again, has a supporter Ps, and so on. Propositions
Pz and Ps, for instance, may of course be identical (a case of self-support), but (1) dictates that a
proposition gains its support from a proposition only. We now have the purportedly support-
ordered endless sequence S3=〈Px, Py, Pz, Ps, …〉, and the support for Px is thus incurably
conditional. However, according to (4), the support for Px is, then, not genuine. In other words,
accepting (1) commits one to accept the view that support for any proposition is always
incurably conditional. And this is incompatible with the assumption that some proposition is
genuinely supported.
It is important to note that the assumption (2) was not needed for the inconsistency to arise. No
Proposition is Supported only by Endless Regresses is now redundant. We only require the
following three assumptions:

(1) Reasons are Supported

(2) Some Proposition is Genuinely Supported
(3) Incurably Conditional Support is not Genuine Support

We thus have a trio after all, but a different one than Cling’s.

From a formal point of view, the epistemic regress problem can be formulated in any of the ways
that are designed to capture those structural elements of justification that jointly create the
problem. However, I strongly believe that the fundamental epistemic desiderata should drive our
formal accounts and not the other way around. I believe this is the main motive for Cling as well,
since he strives to improve the “best extant formulation” with his new version. Therefore, I
believe both Cling and I attempt to explain why the problem arises, not merely to describe how
the problem can be formalized.

In the trio I proposed above, the driving force is the unconditionality of genuine support, and this
notion is of utmost importance in explaining why Cling’s trio applies. We are reluctant to accept
endless sequences of support as providing genuine support precisely because we are reluctant to
accept incurably conditional support as genuine support. Also, we are strongly inclined to hold
that (genuine) reasons must have (genuine) support. These two fundamental convictions fly on
the face of our conviction that some proposition is genuinely supported.

At the core level, then, the epistemic regress problem can be explained in terms of there being a
strong pull into two incompatible directions within the concept of evidential support. On the one
hand, we are inclined to hold that a genuinely supported proposition must be supported by a
genuinely supported proposition; hence the threat of an infinite sequence of supporters. On the
other hand, we are inclined to hold that genuine evidential support is ultimately unconditional.
These two strands within the concept of evidential support jointly create a fundamental problem,
since no proposition will go with them both.

Conclusion

The epistemic regress problem is a major issue in epistemology. As Cling argues, it has to do
with the incompatibility of our core assumptions about evidential support. However, Cling does
not, I believe, get to the innermost core. I have argued that on the most fundamental level we
recognize that our concept of evidential support pulls into two contradictory directions. The first
direction takes us to endless sequences of propositions having incurably conditional support, and
in the second direction we are bound to hold that incurably conditional support is no genuine
support at all.

The duo plays heavily. It presents a powerful skeptical challenge to our pursuit of paramount
epistemic values. How can we ever successfully carry out our rational pursuit of truth, if no
genuine epistemic support is attainable? And, perhaps most notably, we may ask how knowledge
can be acquired if no genuine evidential support is reachable. It is a remarkable task for
epistemologists to show how knowledge is possible without flying on the face of our core
assumptions about evidential support.8

1 Aikin (2005, 195-196) argues that the analogy is not very apt. See, however, Sanford (1975)
for an interesting way of uncovering the general structure of regress arguments. According to
Sanford, regress arguments dealing respectively with an infinite extent of time and with the
infinite divisibility of time and space share a similar structure. Arguably, arguments dealing
respectively with an infinite stretch of train carriages and with epistemic justification may have
the same structure as well. Sanford’s solution to the regress arguments is based on the idea of
conceptual vagueness. I shall not enter into this approach here, since its application to the
epistemic regress problem does not sound too promising to me.

2 Here I agree with Klein (1999, 299), who holds that some foundationalists could consistently
maintain both that some propositions are self-supporting reasons and that reasons must be
supported. An additional argument is needed to block out self-support. One might, for instance,
invoke the assumption that genuine support is not circular, or that genuine support is irreflexive.
Either assumption suffices to preclude situations in which a supported proposition is included in
its own evidential ancestry.

3 Aikin (2005, 198) has reported that a specific kind of simplification argument, due to Oakley
(1976, 227-228) and Foley (1978, 313), works as an argument against reflexivity in support-
relations. If reflexivity is allowed, repeated application of a simplification rule for conjunction
leads to an absurdity. Applying the simplification rule, we get the sequence S4=〈P0, (P0 and P1),
(P0 and P1 and P2), …〉. However, we can replace P0 with ~P0 in S4 without any trouble. Hence
irreflexivity of evidential support. The force of this argument might be objected, though, by
arguing that implication is not sufficient for support (e.g. Cling 2008, 402).

4 In fact Klein is, to my knowledge, one of the three philosophers explicitly defending epistemic
infinitism, the others being Jaquette (1996), Fantl (2003), and Aikin (2005).

5 Bergmann has argued that the lack of reasons is not sufficient for arbitrariness. According to
Bergmann, basic beliefs—beliefs that share some feature F—can be non-inferentially justified.
Suppose I believe that P0, and P0 in fact has the feature F. According to Bergmann (2004, 164),
my belief is justified provided P0 in fact is basic; no inference needed. I may even believe that
all beliefs having F are non-inferentially justified, but this belief plays no role in the justification
of my belief P0. The justification of my belief P0 depends on the truth of the claim that all
beliefs having F are non-inferentially justified, not on my justifiedly believing this claim or
having any reason for it. (Bergmann 2004, 164.) Klein has replied that non-inferentially justified
propositions, those that purportedly have autonomous warrant, do not terminate the threatening
regress in a satisfactory way. Perhaps P0 in fact has autonomous warrant, but this
possibility does not increase the warrant that I have for P0 if I doubt whether P0 is true. (Klein
2004, 169-171.) I agree with Klein, at least insofar as I believe that if I am skeptical about P0,
the possibility that P0 enjoys autonomous warrant does not give me any comfort whatsoever. In
that case P0 is, from my point of view, arbitrary.
6 Incidentally, Cling (2008, 415) sets forth an argument to the effect that self-support is not
really support affording by appealing to the assumption that support is irreflexive. It is not very
clear to me whether Cling himself defends this argument, but if he does, the trio expands into a
quartet again.

7 Of course, the assumption (4) in this quartet may be worked into the assumption (1), but for the
sake of clarity I shall here formulate an additional assumption to Cling’s trio.

8 I would like to thank Olli Koistinen, Juho Ritola, Krister Talvinen, and two anonymous
referees of Philosophia for criticism and comments. My work on this paper has been financially
supported by the Academy of Finland (grant 8114178).
ARGUMENTATION AND ADVOCACY
44 (Winter 2008): 130-142
THREE OBJECTIONS TO THE EPISTEMIÇ THEORY OF ARGUMEMT REBUTTED
Scott F. Aikin
I Abstract: Three objections to epistemic theories of argument are briefly presented and rebutted.
In light of this reply, a ease for argumentative epistemic eclecticism is made. Key words:
argument, epistemics

EPISTEMIC THEORIES OF ARGUMENT: AN OVERVIEW

The epistemic theory of argument is the view that arguments are to be evaluated in terms of their
comprising epistemic reasons. This is to say, good arguments are those that are conducive of or
pursuant of knowledge. Epistemic theories of argument vary according to how knowledge and
epistemic reasons are delineated-from, for example, the veritistic and social in Goldman's
analysis (1999, 2003) to the evidentialist and individual in Feldman's (1994, 2005). What makes
these widespread forms of a family is the central role that the concepts of knowledge and
epistemic justification play in the analysis of what constitutes good arguments. What follows in
this section is a rough map I of the dialectical terrain ¡around epistemic theories of argument. My
overall objective' is to provide defenses for epistemic theories of argument as a family from
objections arising from the rhetorical tradition.

The appeal of epistemic theories can be captured by the axiological and the constitutive norm
arguments. The axiological argument is that since arguments are to be normatively evaluated, a
theory of argument must provide criteria for those evaluations. Epistemic theories provide
normative criteria for good arguments and may be deployed to explain why fallacies are
fallacious: they fail in some way or other to provide epistemic support. The alternatives, as the
argument goes, fail to provide such explanations. Rhetorical theories provide criteria for
evaluation (that of eliciting assent), but then cannot address the problem of fallacies (they
convince, but shouldn't). This, again, is a rough challenge for the rhetorical theories ¡of
argument, one that stretches all the way back to Socrates' concerns about rhetoric in the Gorgias
(465 a-d). Pragma-dialectical strategies evaluate arguments on their procedural correctness in
rationally reducing conflict, but they ¡leave open the question of why the procedures should be
rational and what the nature of that rationality is. On the axiological argument, epistemic
theories are the last standing (cf. Biro & Siegel, 1992, 1997; Feldman, 1999; Freeman, 2006).

The constitutive norm argument is that so long as arguments are supposed to achieve any change
in view from audiences, as the competing theories hold, they must do so on (or on what passes
for) good epistemic grounds. Listeners don't knowingly change their minds about things unless
they think that adopting the new view puts them in a better cognitive position with regard to the
truth of what is believed. Epistemic reasons provide that connection between belief and truth, so
arguments, by their bearing on the truth of their conclusions, must be epistemically bounded (cf.
Aikih 2006, 2008a; Cherwitz, Í977; Cherwitz & Darwin, 1995; Cherwitz & Hikins, 1986;
Heysse, 1998; Scott, 1967, 1976; Stark, 2000; Zaner, 1968). This is to say that so long as one
changes ¡one's mind about a matter only under the conditions that one takes the new view as
more likely true than its competitors, the reasons for this comparative judgment must bear on and
be productive of knowledge of the truth of those theses. Those reasons are, by definition,
epistemic reasons. As a consequence, epistemic theories are of a broader family with logical
theories of argument—that one constitutive objective of arguments is arriving in a manner that
confers the committed subject with a warrant for her conclusion. Epistemic theories assess the
connection between premises and conclusions as argumentative products in a similar, but
broader, fashion compared to logical theories. But these theories, again, broadly take arguments
as the primary object of evaluation, and are posited on the assessment of the connection between
reasons proposed or presumed and the conclusion according to general rules of good reasoning.

There has been a measure of resistance to epistemic theories. A number of lines of argument
have come out, and here I will respond to three I take as connected and widespread. I will term
them the contestability, practicability, and dignity objections. What connects these objections, as
I take them, is that they proffer a critique of epistemic goals and criteria from a rhetorical
perspective, from that of the process elements of argumentation. In what follows, I will present
these three arguments (section II), briefly defend the epistemic theory (section III), and survey
the case for what I will call epistemic argumentative eclecticism that arises from the defenses.

Three objections

The contestability objection runs that, given the variety of views and debates in epistemology,
there will be a variety of competing accounts of the epistemic norms bearing on arguments. If we
are to evaluate an argument by the appropriate epistemic norms, we must determine the norms
first. Epistemologists have been working full-bore on that for quite a while, and it looks like no
one view is winning out. As a consequence, when we evaluate an argument, we are likely to
introduce a contestable criterion for judgment, and in so doing, we risk gerrymandering the
axiology for one side of the case or another. First-order natural theological arguments like the
design argument inexorably drive the discussion to second-order arguments about the epistemic
principles driving them—how acceptable are presuppositions about God's likely designs, how
strong are analogies between designed machines and solar systems, is faith a legitimate source of
data for these arguments, who has the burden of proof in natural theology? These second-order
discussions hardly shed any more light than generate greater heat, and this is a consequence of
the contestedness of the epistemic principles behind the first-order theological discussions. One
might go further and, on the analogy with the cynical induction, take the current state of
dialectical play in epistemology generally to be evidence that we don't know what epistemic
principles are true (Kaplan, 2000, p. 283; Neilson, 2007, p. 142; Rorty, 1967, pp. 1-2, Rorty,
1991, p. 23; Rosenbaum, 2002, p.69). Consequently, we have no criteria for argument
evaluation. Hoffman captures the difficulty of the situation with regard to our argumentative
criteria as follows:

It might be possible that the evaluation standards I am using in my particular situation happen to
be "universal" standards, but how do I know that? And how could it be possible for anyone to
justify the claim that his or her standards are in fact "the" universal standards? (2005, p. 248)

In similar fashion, Tindale rejects any non-relativist account of truth and reasonability in
argument, and notes:
People from different perspectives can dispute the reasonableness of their judgments. The
rhetorical perspective on argumentation facilitates this. As long as any position is assumed to
hold the truth ... the exercise of reasonable disputation is undermined. (1999, p. 98)

As a consequence, Tindale reasons, criteria beyond those recognized by the audience are useless.
In its place he proposes that the rhetorical notion of audience-acceptability is all there is to
epistemic assessment:

[W]e must evaluate the acceptability of a premise according to whether it would be accepted
without further support by the audience that is to consider it, the immediate or intended audience.
Let us call this an epistemic condition for audience acceptability. (1991, p. 243)

In a similar rhetorical vein, Levi argues that the search for criteria, even independent of their
contestability, yields a structural problem for logical theories generally: ' '

That there must be criteria for argument correctness is logic's article of faith, and explains why it
does not see that the assumption that something must make an argument correct is unwarranted.
If criteria were needed, then why not criteria for the criteria? A vicious regress seems
inevitable. . . . The real problem is with the assumption that criteria are needed. (1975, p. 266-7)

The obvious self-defeat of this commitment should not be lost on us here, as Levi is criticizing a
theory of argument for having as an article of faith that there must be criteria for good argument.
Surely if he's right, then there are no grounds to criticize the theory that there are grounds.
However, the point here is not to bring the charge a self-defeat problem for rhetorical theories,
but to provide a defense of their competitors in the epistemic theory (cf. Aikin 2008b, in press;
Rowland, 1995, for pressing this line of self-defeat reasoning). What is crucial from Levi's
argument is that given the structural ¡problem for determining criteria, we have grounds for
presuming that there are not any. ; |

The contestability objection, then, comes in two strengths. Weakly, the view is that given the
contestability of epistemic principles, we have no justification for introducing them to evaluate
arguments. Strongly, the view is that there are no such principles or standards beyond those that
arguers hold (cf. Ede, 1981, p. 125; Harpine, 2004, p. 335; Perelman & Olbriechts-Tyteca, 1969,
p. 66).

The practicability objection follows hard on the heels of the contestability objection. A
desideratum of a theory is not only that it get what we are theorizing about right, but in
proceeding, it should provide good advice as to how to manage ourselves in relation to it.
Brummett notes rightly that theories of argument and rhetoric must "apply or die" (1990, p. 71).
Theories of argumentation, then, should have practical payoff, but epistemological theories are in
a bad place to provide those goods. If the contestability argument goes through, epistemic
theorists aren't in any position at all to provide any criteria for arguments, so they have no advice
beyond empty slogans like: construct arguments that provide good epistemic reasons.

This said, the practicability objection need not depend on the contestability objection. Huss
(2005) presents the following version of the practicability problem for epistemic theories
independently of the problem of contestation. Let us grant that there are some principles that are
not contestable, say, the basic principles of probabilistic reasoning that jointly explain why the
gambler's fallacy is a fallacy. So far, the epistemic theory's advice; is to avoid the gambler's
fallacy. Huss then considers a group of gamblers who have been told not to make use of the
argument form. But what if they fail to see why the fallacy is a fallacy? If they don't see, then the
advice on offer will fail to motivate them, and if they aren't motivated by the advice, they will
continue to use the gambler's fallacy (2005, p. 267). As it is taken, an epistemic theory's advice is
only the introduction of criteria for judgment,' not the introduction of the means to motivate or
clarify those criteria for those who deploy the arguments. As a consequence, epistemic theories
provide advice, but it is not what Huss calls 'followable advice.' Alternately, Huss proposes, on a
consensus theory, we seek such motivation. As a consequence, the gamblers, when given the
right motivations, "come together as truth-seeking rational agents and agree that the inference is
likely to yield epistemically justified beliefs. It is this agreement... that motivates them to both
avoid the gambler's fallacy and continue with the discussion" (p. 267). What's needed, Huss
contends, is not just a theory that provides criteria, but one that gives us a means to achieve the
necessary results. Consensus theories, because they are focused on those sorts of means and
results, are better designed to provide this sort of advice (cf. Burke, 1984, p. 23; Govier, 1987, p.
46; Grootendorst, 1991, p. 113; Sillince & Minors, 1991, p. 282).

The critical edge of the practicability objection is that epistemic theories are devoted to looking
at arguments, but they ignore arguers, listeners, and the various other aspects of the acts of
arguing. One feature that seems to drop out prominently with epistemic theories is consideration
for the autonomy of those involved within a dispute. Why, one may ask, from an epistemic
perspective, should we be open to challenge from all quarters, instead of from only those
recognized as the competent? Don't all deserve consideration? Don't all deserve a response?
Don't all who don't agree deserve arguments addressed to them? Johnstone notes there that "the
issue is really moral; it is only apparently epistemic" (1968, p. 166). Commenting on the
demands of universality in argument, Crosswhite argues that, in light of a rhetorical conception
of argument (and universality), there are special requirements in place:

[W]hen an argument is known to project its claims to a universal audience, critics can raise
objections that certain groups of people or certain features of their identities have been left out—
that the reasoning does not have the scope imagined. (1996, p. 159)

Later, noting the claims of logic in arguments, Crosswhite notes that his students often have
"philosophical difficulties": they "are not clear just what the claim of logic is, or who is making
the claim. More specifically, they are not clear about why they should take the logical point of
view" (1996, p. 161). In a similar vein, commenting on Perelman's emphasis on argument being
addressed to listening audiences, Tindale argues:

First-hand recognition of something is likely more compelling than a second-hand relating of it,
because the person "sees" the point and invests in the idea. Self-persuasion, insofar as it is
explicitly encouraged here, indicates further the non-exploitative sense of rhetoric that governs
the proceedings (2006, p. 344).
Rhetorical theories, because they require the appeal to the perspectives of those addressed by the
argument and arguer, embody the aspiration of respecting the dignity of one's listeners. Tindale
calls it "cooperation in a shared community of mutual regard" (2006, p. 344). Perelman himself
calls this requirement "the rule of justice" (Perelman & Olbrechts-Tyteca, 1969, p. 464).
Epistemic theories have no obvious requirement for audience acceptability. Audience-
acceptability is not forbidden, but it is not a criterion for argumentative success. However, with
rhetorical theories, it is required. As a consequence, with rhetoric, we respect the dignity of our
audience, whereas in epistemic contexts, those with the wrong beliefs or with the wrong
epistemic principles do not need to be addressed. On the epistemic theory, you only need correct
epistemic commitments, not corrections of wrong ones. Epistemic theories, then, leave too many
behind in critical discussions, and that's not right.

A rebuttal and the case for eclecticism

The epistemic theory of argument can accommodate these objections. This is achieved by having
the exclusivity of evaluation by epistemic justification weakened to the primacy of evaluation by
epistemic justification, consequently, a form of eclecticism. What follows is a sketch of a path
through the brambles.

The contestability objection runs that since epistemic theories are contested, we have no
unproblematic criteria for adjudging arguments. What follows is skepticism of some degree
regarding argumentative validity. The first thing to distinguish is the concept of correct epistemic
principles (whatever they may be) and our conceptions of them. The epistemic theory is only that
arguments should be adjudged in terms of the correct epistemic principles. Of course, it follows
that if we do not know what principles are correct, we are not in a position to judge the validity
of an argument. But, this looks more like a reason why epistemology is important than a reason
why we should not care about the epistemic principles at work in the argument. The first lesson
of the contestability argument is that we should get to work in epistemology, not yet turn to
rhetoric.

Moreover, the contestability of epistemic principles is overstated. It is not that epistemologists

disagree on the justification yielded by, say, experience. There is plenty of disagreement about
the nature of that justification. But epistemologists do not disagree on whether experiences are
sources of epistemically reasonable beliefs. They may not yet count as knowledge, they are
defeasible, and so on. However, the core epistemic principle, that if a subject (S) has some
experience (e.g., a visual experience) that has some representational content (e.g., that of having
one's hand in front of one's face), then S has' a reason to; believe that she has her hand in front of
her face. Epistemological theories here are devoted to explaining why so., or how widely to
construe the representational content, how easily S's reasons can be defeated, whether S need
further reasons in addition to the representational content, and so on. That's where the
disagreement resides not about the core principle. The same goes for a majority of the epistemic
principles first-order arguments work on-the importance of reliable sources of information, good
track records for truthfulness, the default justification for what we see and remember, and the
transmission of justification over truth-preserving inference.
The practicability objection runs that epistemic theories are either too thin to answer the needs of
offering advice or offer useless advice, because epistemic theories risk ignoring the attitudes of
those who need correction. In effect, then, the practicability objection amounts to the same
complaint that the dignity objection keys on - that the epistemic theory leaves out, or does not
give sufficient emphasis to the thoughts and inclinations of audiences. Practically, it is the
writing teacher's refrain: remember your audience. Morally, it is the requirement of being
considerate.

The first thing to be said to the practicability objection is that the epistemic theory i as plenty of
advice to contribute to argumentation. The most obvious place to start' is with Huss's own
employment of the practicability objection—he notes that if we were to motivate the gamblers in
his example to avoid the gambler's fallacy, it would be on the basis of bringing them to see that
the inference form "is not likely to yield epistemically justified beliefs" (2005, p. 267). If
epistemically justified beliefs were not the goal for such reasoning, but instead bare agreement,
assent, or reduced conflict, this point would not make any difference. Feldman makes precisely
this point in his reply to Huss, by pointing out that his own textbook. Reason and Argument
(1993) was written from such a perspective (2005, p. 280). But we do not have to go beyond
Huss's own example, as it is on Huss's story, epistemic justification is the ultimate end here.

A further point to be made here is that the epistemic theory has plenty of other sources for advice
- consider the argumentative options that the epistemic regress problem poses with its various
answers. The epistemic regress problem can be captured by the tension between three apparently
correct, but mutually inconsistent epistemic principles:

(1) The Principle of Inferential Justification: If some subject (S) is justified in believing that
something (p) is true, then S must have some further reason (q) for believing p.

(2) The Principle of Justified Justifiers: If S is justified in believing that p on the basis of q,
then S must be justified in believing q.

(3) The Impossibility of a Justifying Regress: No infinite chain of reasons provides

justification. (Cf. Cling 2008)

The problem is that 1 and 2 require that any justified belief will require an infinite chain of
reasons, which is inconsistent with 3. Foundationalist strategies with the inconsistent set are
posited on revising 2, so that there is a special class of reasons, basic beliefs, that do not require
further inferential justification. Contemporary defenses of foundationalism can be found in Audi
(2001a), Bonjour (2002a), Fumerton (1999) and McGrew (1995). Take, for example, beliefs
such as: I have a headache, 2+2=4, and All things are identical to themselves-all you have to do
is understand those sentences, and you're in a position to adjudge their justificatory status.
Coherentist strategies revise 1, so that justification need not derive from serial chains of
inference, but rather may supervene on coherent systems of truths. Contemporary defenses of
coherentism can be found in Bonjour (1985), Haack (1993), Rosenberg (2002), Sellars (1997),
and Thagard (2002). For example, consider the way explanations fit large sets of data—in order
to assess the justification we have to believe, say, that it was the cat that knocked over the vase,
you have to have that belief fit coherently with a large body of other knowledge (e.g., where the
vase was, the cat's usual activities around it, that nobody else had access to it, etc.). Infinitist
epistemic theories revise 3, so that only infinite series of serial inferences yield justification. The
case for infinitism has been made by Aikin (2005), Fantl (2003), and Klein (1999). Knowledge
may require that we be able to answer all the questions, and there may be no end to them.
Consequently, inquiry and critical discussion have no in principle stopping points. Reliabilist
theories of justification reject 1 with certain classes of belief— ones that are produced by reliable
sources under the right circumstances (Goldman, 1986). For example, take beliefs yielded by
visual perception in good lighting they are reliably produced, so those who believe them are
justified in so doing. Importantly, on this rubric, even rhetorical theories of argument may be
classified as forms of epistemic theories, as they can be taken to be a revision of 2 such that one
need not offer further justification for commitments not challenged in a context—namely, that
audience acceptance confers justification. Again, the point here is not to show that rhetorical
theories cannot avoid epistemic work, but that epistemic structures come part and parcel with our
views about good reasoning.

Turning to how these alternative epistemic theories yield advice, one need only trace each
theory's take on the structure of justification as a blueprint for how to make an argumentatively
legitimate case. Foundationalism provides the argumentative strategy of proffering basic
premises from which to erect arguments. Basicality, surely, may be something at issue with some
starting points, but many are broadly acceptable (truths of logic and mathematics, present and
accessible empirical truths, truths of self-awareness). Coherentism offers the strategy of setting
issues in relevant and explanatory connection with broader truths—you know how things hang
together. The infinitist suggests that one be ready to answer challenges until there simply are
none. It takes only a little imagination to turn meta-epistemological theories into bits of
argumentative advice.

The second thing to be said is that internalist (or subjective) epistemic theories are bound to
address the attitudes of arguers. Broadly, internalist epistemic argumentative theories are
proposed by Aikin (2006, p. 99; 2008a, p. 243), Feldman (1994, p. 196), freeman (2005, pp. 73-
6) and Lumer (2005, p. 196). Internalist theories of epistemic justification require, that for
subjects to be justified in their beliefs (or for the beliefs to be justified for them), the justifying
reasons must be available and recognizable as justifying ¡reasons for the subject (cf. Audi,
2001b; Bonjour, 2002b; Chisholm, 1966; Feldman & Conee, 2001). That is, you are justified
only if you can explain how you are. So, returning to Huss's example of gamblers with no
motivation to follow the advice of avoiding the gambler's fallacy, we see that the case works
against the epistemic theory only if the gamblers do not see the reasons for why the fallacy is a
fallacy. But internalist theories of justification would riot hold that these gamblers are| in the
right epistemic relation to the rule-Huss's gambler case shows the epistemic theory doesn't work
only because the people in the example don't live up to the epistemic demands of the theory. The
internalist epistemic theorist would give the advice: give the gamblers a demonstration of why
the fallacy is a fallacy so that they] have justifying epistemic reasons (and hence a motive) to
avoid the inference form.

This brings us to the question of why we must address those who are not motivated by the right
epistemic norms. On the epistemic theory, arguments deployed according to the right epistemic
norms are correct. Correcting those who do riot have the right norms or even addressing them is
not necessarily a desideratum of an argument. One might say that all of the pragmatic use of
argument drops out of the epistemic consideration. This may be right for arguments considered
as sets of premises supporting conclusions, but only on the thought that the norms are considered
very strictly. There are plenty of broader cognitive norms that, though not in the service of
constituting knowledge (hence, epistemic), are nevertheless correct rules for the management of
one's intellectual life. Two joint goals of reasoning are that of securing the truth and
understanding what those truths are. Having no answer to people with whom one disagrees
strikes me as a compelling reason to think that someone even with good prima facie reasons
doesn't know in a way that satisfies our epistemic duties of understanding the things we know. I

Take, for example, my daughter (a first grader). She can't answer my question as to why we don't
count the grouping of 2 and 2 when we add them. If we count the grouping, then 2-I-2,=5, but
she nevertheless knows in some attenuated way that 24-2=4. However, someone who can answer
the question understands the notions of addition and number better than someone who cannot.
And consequently, one may say this person knows it better than someone who does not.
Attending to those with whom we disagree is a cognitive duty (or I prefer to say a broader
epistemic duty), one that comes part' and parcel with the commitment to what we believe—we
are committed to the truth and to| its intelligibility. If we care about understanding the things we
think we know, having satisfying answers to those with whom we disagree is a positive duty.

One way to see these broader epistemic norms of engagement is in cases of deep disagreement,
where a disagreement between two parties is not held against a broader agreement in many other
matters, but is against a backdrop of wider disagreement. Take, for example, the perennial
debates concerning evolution. Denyse O'Leary, a Canadian author and blogger critical of
evolution, tells the following story:

A couple of years ago, after I had been following the controversy for several years, I found
myself listening to a long lecture by a Darwinist, replete with bafflegab and pretty lame
examples. Finally, sensing (correctly) that I was unconvinced, he proclaimed to me, "You just
don't understand how natural selection works, do you?"

And suddenly, the penny dropped. What he meant was that I just don't believe in magic. I can't
make myself believe in magic; I haven't been able to since I was a child. And I was no longer
going to give the matter any attention. (O'Leary, 2009)

P.Z. Myers, a biology professor at the University of Minnesota in Morris and Pharyngula
blogger, responds to O'Leary by first correcting the assumption that evolution is magic, but then
notes an inconsistency in O'Leary's own position:

Natural selection is not magic; there are no miracles, no unexplained steps in the process, and
once you grasp it, it's simple and obvious. That O'Leary equates the two means the correct
answer to the question was "yes".

The real funny part, though, is that O'Leary is an intelligent design advocate and ardent Catholic.
She does believe in magic! (Myers, 2009a)
Two important norms are in sharp focus here. One is deliberative honesty-namely, that in
argumentative exchanges, one must (unless overtly taking on the viewpoint of an unrepresented
side in playing the devil's advocate) present as honest and thorough version of one's
commitments as the circumstances allow. Disagreements will not be resolved if one fails to be
honest about what produces them. Moreover, one will not learn anything if one will not be
honest about what one does not know. Myers' charge is that O'Leary hasn't done that she has
inverted the dialectical relationship between intelligent design/creationism and evolution by
saying that it is evolution that is the magical explanation. Not only has O'Leary misrepresented
the dialectical situation, she has misrepresented on what side of the disagreement she really is on
—she is the one who accepts supernatural explanations, not the evolutionists. Until it is clear
what the issues are and how those in the argument are coming down on them (and until they can
be consistent in reporting their views) no progress in resolution or inquiry is likely.

The second norm O'Leary's post and Myer's response brings out is that there are responsibilities
of clarity. O'Leary notes the 'bafflegab' of the 'Darwinists,' and thereby, she makes the charge
that the failure of clarity on behalf of the 'Darwinists' is reflective of a kind of evasion. The
question, of course, is whether jargon in the papers she heard was there to prevent outsiders from
understanding and contributing, or to facilitate insiders' dialogue. If the former, then the regular
charge of academic and scientific elitism is appropriate-something implicit in O'Leary's terms,
namely, that the big technical words are simply to baffle and thereby cow their non-specialist
audience. But the test for this is whether the terms can be explained and made accessible in other
contexts. And further, whether accessible answers to creationist challenges are available.
Responding to a different but direct charge of academic elitism, Myers notes:

Read any of Stephen Jay Gould's books . . . and you'll find them to be lucid and enthusiastic and
eager to explain. Even more so, crack open one of Richard Dawkins' books—they are
exceptionally clear. Heck, just walk into your bookstore, find the tiny, narrow little shelf where
the science books are hidden, and you'll find lots of plain-spoken exposition.

Science papers tend to be heavy on the jargon because they are tightly condensed. It's a highly
refined format designed to facilitate communication between knowledgable people in the field.
It's not that hard, though: we teach ^undergraduates how to read and write science papers, and
although admittedly they find them difficult at first, it only takes a little knowledge to be able to
work through them. (Myers, 2009b)

Myers' point is that dialogue requires work from both sides, and one cannot expect to step
directly into the cutting edge of a technical discipline without at least a modicum of training.
Technical papers come across as 'bafflegab' because hearers often do not know how little they
know. Making the accusation of evasive and overbearing language (and the attendant elitism)
requires that one have done one's homework and know the clearer and more accessible
alternatives. But without having done any of that work, O'Leary and most elitism chargers fail to
make their case.

But the two-way element of dialectic also puts an epistemic burden on those who| defend
evolution from these charges, too. Writing popular science requires a knowledge of widespread
audience temperament, temptations to misunderstand, hotspots for further disagreement, terms
that enrage, and so on. The first reason why this is important is that one understands the issue
best when one knows the other side's case and can answer it. The second reason why this is so
important is that knowledge cannot be transferred or understood if it is not in a form that is
accessible to its audience. Again, epistemic theories of argument require that arguments be
evaluated by their comprising epistemic reasons, and if a given argument fails to give its
audience epistemically good reasons because the audience fails to understand and thereby
believe on their basis, the argument fails to be epistemically successful. The argument's author
may understand and successfully believe on the basis of those reasons, and thereby be justified
and herself know. But if the argument is not conducive of knowledge in its audience in a similar
fashion, the argument is a failure, and not for moral reasons, but for epistemic reasons - it fails to
transmit knowledge.

Respecting the dignity and perspectives of audience is epistemically important, because a failure
to do so reflects a form of short-sighted dogmatism. Many issues are vexing and difficult, and
any theory worth its salt must be capable of accommodating the possibility (every likelihood)
that people can be reasonably engaged in a dispute and that one side (or both) can be wrong.
That is, it is a desideratum of a theory of argument and rationality generally that it be possible
that people can be rational but wrong. 'Fallibilism' is broadly the term denoting this view, and it
is a positive epistemic failure of advocates of views to mistake the perceived errors of those with
whom they disagree to be indicators of their irrationality, stupidity, or duplicity. As a
consequence, eagerness to clarify, enthusiasm, for dispute, and willingness to revise one's views
in light of criticism are reflections not only of our respect for the dignity of others in a
discussion, it is a necessary component for epistemically responsibly holding our beliefs.
Consequently, the way we respect the dignity of those with whom we disagree is in giving them
arguments. Precisely, we respect their dignity in giving them arguments that (a) they can see
from their own internal reasons are good (so that they will come to have epistemic justification),
(b) they will come to understand the disagreement and its resolution, and (c) that have space for
their response and for their own case to be brought to bear on ours (so that if they have the better
argument, we'll come to have justification for changing our minds and understanding the
resolution).

These goals are primarily cognitive, but they clearly have a meet-up with pragmatic rhetorical
norms. We should give epistemically good arguments, because those are the arguments that are
most likely to withstand criticism, and we should give accounts" that address widely flung forms
of critique, because these strategies yield more stable commitments. That is, if we argue to
change someone's mind as the rhetorical and dialectical theories run, it seems right that we
should proffer arguments that not only resolve the disagreements but do so with some measure of
stability to that resolution. Epistemically good arguments (ones that both provide justification
constituting knowledge but also provide broader understanding of the issue) are going to be the
ones that best achieve these goals. Cognitive stability is one sort of good, but cognitive stability
when we have the truth seems much better. So, in giving these arguments, a broader epistemic
theory does promote the moral goods of respecting the dignity of fellow arguers, since it is
pursuant of bringing them to know and to understand a point at issue by their own lights. And if
the arguments we proffer don't work but our interlocutor's do, then it's our duty to change our
views. This requires that we make room for objections from far-flung quarters, for extended
discussion, and for sometimes leaving things open. For those who come out of argumentative
contexts with the best epistemic reasons, we must run the argumentative context fairly and be
attentive to the reasoning on the various sides of the debate. This requires that we treat them with
dignity, but it needs to be noted that this dignity is mediated by cognitive goals-we have the
dignity we do because we are honest inquirers.

This defense of epistemic theories of argument has required that the exclusivity of epistemic
norms be weakened to their primacy. We primarily seek good knowledge-conducive arguments,
but there are other goods to be achieved by argument. These are precisely those pursued in the
act of arguing, resolving disagreement, improvement of understanding of the issue, and
promoting a stable dialectical situation (one constituted by the exchange of good arguments
instead of vicious means of resolving conflict). Classically, these process elements of arguments
have been considered under the rubric of rhetoric. Admittedly, these ancillary goods are ones that
are not exclusively epistemic (as ones directly pursuant of the goal of knowledge), but they are
broadly cognitive in that they are reflective of our general epistemic responsibilities, and they
additionally reflect our practical-rhetorical interests in argument. As a consequence, the best
defense for epistemic theories of argument is for them to be parts of wider cognitively eclectic
theoretical programs wherein the variety of goods aimed at in arguing can be accommodated.

This epistemic eclecticism amounts to a broadly epistemic theory of arguments not only as
products but also as bearing on processes of argumentation. As a consequence, it bears a
resemblance to earlier much stronger epistemic theories of rhetoric, namely, the thesis that all
rhetoric is epistemic. In fact, I think it is fair to classify the view defended here, in the end, as a
mitigated form of an epistemic theory of rhetoric. However, there are two features distinguishing
it from the standard views in the epistemic theory of rhetoric.

First, nothing here requires that all rhetoric is epistemic. Cherwitz and Hikins, for example,
defended this strong view, and though my view bears a strong resemblance to it, I do not hold
that all forms of rhetoric must be devoted to being rational representations of reality.
Bullshitting, for example, is often effective rhetoric, but it is overtly non-representational (cf.
Frankfurt's (1988) "On Bullshit"). Second, the view defended here is not posited on any
substantive theory of epistemic justification, truth, or knowledge. One can be agnostic about
what the ultimate analyses of the main epistemic concepts are and yet still hold that those
concepts constitute the norms of argument-assessment and argumentation generally. This
contrasts starkly with the main competitors in the realm of epistemic rhetoric, as Scott's classic
case for epistemic rhetoric was on the grounds of a form of relativism (1967, p. 13; 1976, p.
261). Cherwitz and Hikins' case for epistemic rhetoric proceeds from a linguistic theory of
content and an ontology of relations (1986, pp. 41, 150). Farrell's case depends on the thesis that
all knowledge "depends on human consensus" (1976, p. 4). Finally, Foss's connection between
epistemology and rhetoric is posited on the inference that, "in the field of communication, the
idea that rhetoric creates reality is ¡known as the notion that rhetoric is epistemic, which simply
means that rhetoric creates knowledge" (1989, p. 122). My own attitude regarding these
epistemic theories of rhetoric is that they are more rhetorical theories of epistemology, but that is
not the issue. Rather, the point is that a thin epistemic theory of argument (one that takes no
substantive stand on the issues internal to the analysis of epistemic terms) is what is on offer, as
opposed to a thick one (that requires a substantive stand).
Contrast the thin view on offer here with the contentious views that previous forms of the case
for epistemic rhetoric. The defense of epistemic theories on offer here is in a better dialectical
position. This is for two reasons. First, because criticisms of epistemic theories of argument, at
least the ones surveyed here, are targeted at the family. I've argued that the criticisms of the
family of views can he rebutted, and so it is best not to adopt any one epistemic position to
provide this case. Second, because epistemological theories are highly controversial within
epistemology, it is best to get as much as one can for argumentation theory independently of the
controversial theories. Having one's entire case for an epistemic theory of argument hang on a
highly controversial premise in, say, meta-epistemology is recipe for a theoretical disaster.
Moreover, it runs headlong into the contestability objection. Finally, a theory of argument should
be able to apply to how epistemologists themselves do epistemology, which isn't that they
assume their epistemic theories are right in their details, and criticize the arguments of others
accordingly. Instead, they do their best to argue in fashions that are neutral with regard to which
theory is correct. Or at least, they should.

Instead, if it is clear that however one comes down on the concepts knowledge, belief,
justification, truth, and so on, one can see them bearing relevantly similar fruit in argument
evaluation, then a much more defensible form of the epistemic theory of argument is in the
offing. I, myself, defend a mixed view of epistemic infinitism, foundationalism, and
contextualism, but I do not think that any of those commitments are necessary for the defense ¡of
the family of epistemic theories provided here.

A final worry must be addressed, as it has been charged; by one of the blind reviewers for this
paper that epistemic theories of argument (and epistemology generally) suffers: from a hasty
generalization from the validity of epistemic norms in argument evaluation to their primacy. An
unjustified privileging, it is charged, occurs when one takes the formal norms of knowledge-
assessment and their demands into an area of informal reasoning. We argue, often, not just to
know, but also to cajole, to explore, to tease, to make nuisances of ourselves, to pass the time, to
edify, or to humiliate. These are not overtly epistemic employments ¡of argument, and so goes
the objection, the centrality of epistemology to these employments of argument is an illusion of
taking one case as paradigm.

This is a serious challenge for the epistemic theory on offer. I cannot answer the objection
completely in this short space. However, there are two rebuttals to be made. The first is
stipuiative epistemic theories of argument are about how arguments improve our cognitive
position on an issue. As such, they are devoted to capturing the norms of public reasoning in
inquiry, critical discussion, and deliberation. If people perform combative or argumentative
speech acts for other purposes, they may look like arguments, and maybe they are in some^ cases
to be evaluated according to epistemic rubrics in addition to other criteria (aesthetic, moral,
agonistic, etc.), then these are arguments in a derivative sense. Again, this point is admittedly
stipulative, and it isn't designed to move the dialogue any further' on this point beyond clarifying
the issue. But the stipulative point isn't just out of the air. There's a thought behind it - namely,
that in all these alternative cases of argument, epistemic elements are clearly significant, if not
still central. It surely seems that if one is merely exploring an issue, one would want
epistemically better rather than worse reasons comprising that exploration. Were one out to be a
nuisance or hurt people's feelings, the best means to do that would be to provide arguments that
have epistemic weight behind them. The more likely that the things one says are true (or pass for
true) makes them more likely to cause actual difficulty or hurt those challenged. The same goes,
as far as I can see, for teasing, cajoling, and so on. The quality of a good tease depends on how
well it purports to reflect or transmit knowledge of a situation—you can only tease people for
things that they and you take yourselves to know to be true about them.

The broad epistemic view on arguments has been, I take it, reasonably criticized by those in the
rhetorical tradition. The process elements of argumentation and the demands of addressing
audiences should be reflected by a theory of argument. However, I've argued here that epistemic
norms bear on the process elements of argumentation and that they require attentiveness to
disagreement. Consequently, we can see epistemic norms of argument as a constitutive part of a
wider set of cognitive and practical responsibilities.

A Pragmatic Phenomenalist Account of Knowledge

Byeong D. Lee
Dialogue / Volume 47 / Issue 3-4 / January 2008, pp 565 - 582

… Fifth, as Brandom (1994, esp. pp. 204-206) insists, our social practice of demanding
justification and responding to such demands requires the default and challenge structure of
justification, for the infinite regress of justification is impossible.6 Suppose that a claimer defends
his claim by offering a ground, A. A challenger can call this ground into question by saying,
“Why A?” To meet this challenge, the claimer might provide another ground, B. The challenger
can, in turn, call this ground into question by saying, “Why B?” Here it should be noted that if
the challenger were entitled to keep raising a question “Why is that?” to any reply of the claimer,
there would be no claim that the claimer can ultimately justify. Thus, in order for it to be possible
that one can defend something, some claims must be treated as having default justification unless
challengers provide positive reasons to doubt them. That is, there must be some claims for which
the burden of proof is shifted to challengers. This is similar to a criminal case in which a
defendant is presumed “innocent until proven guilty.” In addition, this social practice model of
justification requires that something play the role of evidence or ground. If nothing played such
role, one could not defend one’s claim in response to a challenge. What, then, plays such role?
Without reliable perceptual judgements, we would have no evidence for how things are in the
world. Thus, our social practice model of justification requires that such non-inferential
judgements have default justification. In other words, we are justified in holding such
noninferential judgements unless we are given positive reasons to doubt them;
e.g., we are having a hallucination under the influence of a drug.

From these considerations, let me add one more point in my justification condition on
knowledge. If S’s belief that p is non-inferentially justified, S’s adequate ground for believing
that p is simply the fact that such a non-inferential belief has default justification.

6 Peter Klein (2005) argues for infinitism, according to which inferential justifications can
ramify without end and must do so for a belief that is truly justified. He defends this infinitism
by claiming that every justification must be inferential, but that justification can emerge when
inferential chains are long enough. But I am not persuaded by his arguments. I agree with Carl
Ginet (2005) that Klein fails to provide a real example of justification having such an endless
ramifying structure. I also agree with Ginet that a formally good inference can only transfer
justification from premises to conclusion so that it cannot create justification even if its
inferential chain is long enough.

Contemporary Pragmatism Editions Rodopi

Vol. 6, No. 2 (December 2009), 71–87 © 2009
Prospects for Peircean Epistemic Infinitism
Scott F. Aikin
Epistemic infinitism is the view that infinite series of inferential relations are productive of
epistemic justification. Peirce is explicitly infinitist in his early work, namely his 1868 series of
articles. Further, Peirce’s semiotic categories of firsts, seconds, and thirds favors a mixed theory
of justification. The conclusion is that Peirce was an infinitist, and particularly, what I will term
an impure infinitist. However, the prospects for Peirce’s infinitism depend entirely on the
prospects for Peirce’s early semantics, which are not good. Peirce himself revised the semantic
theory later, and in so doing, it seems also his epistemic infinitism.

1.

On the classical conception of knowledge, in order to know, you’ve got not only to believe
something, what you believe has got to be true, and you’ve got to have a good reason as for why
you believe it is true. But how do you know that’s a good reason? It itself has not only got to be
true, but you’ve got to have another reason for why you hold it is. And then for that reason,
another. Then another. Then we’re off to the races.

Virtually nobody thinks we should run that very long race. There is a variety of theories of how
to shorten the distance. Foundationalism is the thesis that we may stop giving reasons with a
special set of beliefs called basic beliefs. Coherentism is the thesis that at a certain level of
reason-giving, we may stop when the beliefs in question have a special fit with the rest (or some
set) of our beliefs. Contextualism is the thesis that we may stop with beliefs appropriate for the
given circumstances. Externalism is a rejection of the requirement that subjects need to give or
be aware of further reasons when they obtain.

Examples of philosophers and whole traditions abound for each of these theories of ways to
leave off giving reasons, and it is easy to find many prime examples of and extended exchanges
between these traditions. But it is difficult to find real examples of the tradition that we might
call infinitist, those committed to the thesis that reasons may iterate infinitely. The old
Pyrrhonean skeptical tradition has a version of accepting infinitism as a requirement on
knowledge, but thereby they reject the project of knowing. This seems to be a plausible reading
of the Agrippan modes and Sextus’ use of them, but it is not a satisfying epistemological theory,
since we are here looking for a theory of justification that dovetails with a theory of actual
human knowledge.1 In passing, Peirce is mentioned as a candidate for holding the theory, but
little more is said. Recently, Peter Klein (1999, 2000, and 2004) Jeremy Fantl (2003) and I (2005
and 2008) have argued for the view, and they have not only given accounts of the view’s
intelligibility but also its resistance to traditional arguments against it.

In light of these new developments in articulating and defending infinitism, I will retrieve the
question as to whether Peirce was an epistemic infinitist. This is for two reasons. First, it is
important that we have an adequate textual backing for these classifications, even if they are
tentative and little hangs on them for current epistemological theorizing. We are scholars, and we
should get it right. Second, in pursuing another example of a philosophical view, we might
deepen and broaden our conception of it. Peirce’s work is innovative in a number of areas, and it
stands to reason that his views relevant to current issues in metaepistemology will be helpful.
I will argue for the following: First, that Peirce is explicitly infinitist in his early work, namely
the 1868 series of articles, “Questions Concerning Certain Faculties Claimed for Man,”
“Consequences of Four Incapacities,” and “The Grounds of Validity.”2 Second, that Peirce’s
phenomenology of the relations between firsts, seconds, and thirds favors an interpretation of
Peirce’s infinitism as a mixed theory of justification, where infinite series of inferences are also
supported by non-cognitive input from seconds. The conclusion, then, is that Peirce was an
infinitist, and particularly, what I will term an impure infinitist. However, the prospects for
Peirce’s infinitism depend entirely on the prospects for his early semiotics, which are not good.
Peirce himself revised the semiotic theory later, and in so doing, it seems also the motives for his
epistemic infinitism.

2.

Broadly, epistemic infinitism is the view that infinite series of inferential relations are necessary
for epistemic justification. The view may be construed as exceedingly demanding in that it may
be an exclusivist thesis regarding sources of justification – namely that only these series are
productive of justification. Such would be a pure infinitism, as is exemplary in Klein’s 1999 and
Fantl’s 2003, where all the standing alternatives are shown insufficient against the backdrop of
criteria for a sufficient theory of justification.3

The model for successful justification for pure infinitism is that of an infinite chain of justifying
relations between commitments. However, infinitism may alternately be a mixed theory of
justification, so that it is an ecumenical take on sources of justification. Mixed theories of
justification are not exclusive with regard to what sources are allowable for epistemic
justification, so, for example, Susan Haack’s foundherentism makes room for both direct
justification and justification from coherence (1993).

For impure infinitism, these series are productive of justification, but they are not the only
sources, and they may work in tandem with other sources of justification. The model for
infinitism on an impure theory would be an infinitely long chain of reasons supported by direct
epistemic support at various nodes. Such support itself need not be infallible, and, in fact,
precisely because of its fallibility, it must be supplemented by further reasons. Such would be an
impure infinitism, as is exemplary in my own 2005 and 2008. The core of the infinitist view,
however, is the same: the response to the regress problem is that of having reasons on to infinity.

An epistemic infinitism is at work in Peirce’s famous 1868 Journal of Speculative Philosophy

series “Questions Concerning Certain Faculties Claimed For Man” and “Consequences of Four
Incapacities.” Peirce’s main purpose is to show that all cognitions are inferential, not immediate
or intuitional. At first, Peirce seems sympathetic to the issues motivating the foundationalist-
intuitionist program in the regress problem. Question 7 of the “Questions” essay is whether there
are any cognitions not determined by previous cognitions. In short, are there intuitions? Peirce
notes that:

It would seem that there is or has been; for since we are in possession of cognitions, which are all
determined by previous ones and these by cognitions earlier still, there must have been a first in
this series or else our state of cognition at anytime is completely determined according to logical
laws, by our state at a previous time (CP 5.259).

But he then notes:

But there are many facts against this last supposition.... [It] is impossible to know intuitively that
a given cognition is not determined by a previous one, the only way in which this can be known
is by hypothetic inference from observed facts (CP 5.260 emphasis in the original).

The Peircean definition of intuitions on the semiotic theory is that they are signs that refer
immediately to their objects. The epistemic corollary is that an intuition is a “premiss not itself a
conclusion” (5.213).

The question of determination of one thought by another, though, complicates the matter, since
these determinations may be semiotic, epistemic, or merely causal in nature. I will take it here
that so long as Peirce’s arguments are that these semiotic determinations are interpretive, they
are jointly causal and epistemic. Call this the dual role of semiotic interpretation: the reasons one
has for an interpretation are both explanations for how one arrived at an interpretation and
justifications for that interpretation.

Peirce further refines his notion of an intuition such that intuitions are (i) first cognitions
determinable by logical laws (5.259), (ii) premises not themselves conclusions (5.213), and (iii)
something about which it can be a question as to whether they are intuitions, even if one has
them (5.260). Peirce synthesized the semiotic features of such judgments (their inferential role as
antecedents to chains of further inferences) with their epistemic features (the justification one has
in having them and the justification they confer through inference). So the question of the regress
problem is whether there can be immediate reference to something that both gives a cognition its
content and its epistemic status as an immediately justified belief.

Peirce’s requirement for immediate reference is that if reference is immediate, it must be

immediate that it is immediate. That is, we must intuitively know when we have intuitions. Let
us call this the meta-requirement. So the difference between intuitive and non-intuitive
knowledge itself must be intuitive.4 Peirce frames the argument for the requirement as such:

[I]t is plainly one thing to have an intuition and another to know intuitively that it is an intuition,
and the question is whether these two things, distinguishable in thought, are, in fact, invariably
connected, so that we can always intuitively distinguish between an intuition and a cognition
determined by another.(5.214)

Peirce, then, from his conception of cognition, derives the following argument that intuitions and
their immediate knowability as intuitions are connected:

[A]ll the cognitive faculties we know of are relative, and consequently their products are
relations. But a cognition of a relation is determined by previous cognitions. No cognition not
determined by a previous cognition can be known. It (intuition) does not exist, then, first,
because it is absolutely incognizable, and second, because cognition exists so far as it is known.
(5.262)

Peirce’s first case is that, given that there has been (up to Peirce’s time and up to now, too)
debate about what is intuitive knowledge and not (the rationalism-empiricism debate, the current
arguments against consciousness), it seems clear, on the meta-requirement, that such intuitions
do not obtain. Additionally, given the meta-requirement, for the second-order-intuition to be
intuitive, we would need a third intuition, and a fourth for that. As a consequence, it would be
unclear how intuitions could solve the regress problem at all, since their introduction yields a
new meta-regress.5

Clearly, Peirce’s argument here depends crucially on the meta-requirement for intuitions. We get
an extended argument for the meta-requirement and the conception of cognition behind it in the
1868 sequel to the “Questions” paper, “Some Consequences of Four Incapacities.” The thesis of
the paper is that all mental events are results of the manipulation of signs. Thoughts are
combinational; they, to be thoughts at all, must be well-formed-formulae, and the model for the
movements of thought must be that of valid inference (CP 6.338).

The components of these formulae are signs: feelings, images, emotions, representations. For
these thoughts to have content and for them to be of something, they must not only be well-
formed, but they must be meant (CP 5.289). How thoughts have intentionality, on this account, is
that the signs comprising thought have objects because they are interpreted as being of
something by other signs. Without the interpretant, the original sign loses its meaning. It is
therefore essential that signs come not just as sets of well-formed formulae, but as sets that mean
by their interrelations. So thoughts, as combinatorial, are like sentences, and sentences do not
mean anything unless they are read and interpreted.

So it goes for thought. The rationale for the meta-requirement, then, is that the condition for a
cognition to exist as a cognition is for it to be known to exist as such (CP 5.262). As long as
there is a problem with nailing down whether or not there are intuitions, they are not known to
exist. So they don’t. Pierce anticipates a familiar regress of analysis problem (an analogue of the
epistemic regress) in light of this requirement. If we cannot have thoughts without signs, then it
seems for all interpretations, there must be interpretations of them, and then interpretations of
them, and we’re off to the races, but this time on a semantic or semiotic level.

Yet Peirce accepts infinitism on this level also, as he notes that the alternative requires that there
be intuitions that play non-derived inferential roles. To hold that there must be such notions
“assumes the impossibility of an infinite series. But Achilles, as a fact, will overtake the tortoise.
How this happens, is a question not necessary to be answered at present, as long as it certainly
does happen” (CP 5.250, emphasis in original). So Peirce is not only committed to there being a
requirement of infinite analysis, he is committed to the thesis that these infinite analyses are
actual. Just as Achilles actually catches the tortoise, and in so doing traverses an infinity of
spaces, we mean and reasonably believe, and in so doing, are involved in an infinity of
interpretations and justifications. The question, despite the fact that Peirce tables it, is how
exactly this is brought to actuality.
Let us take Peirce’s analogy with the Achilles case more seriously. One of the principles behind
the paradox is that Achilles cannot make an infinite number of journeys in finite time. But the
question with this requirement is why not. As the distances become shorter and shorter between
Achilles and the Tortoise, the time it takes for Achilles to traverse the distance grows
progressively shorter. As such, the times, instead of stretching out to infinity, contract to the limit
of Achilles overtaking the tortoise. Time, however, does not start and stop with each journey.
Neither does Achilles. The question is whether thoughts have the same underlying continuity that
time and Achilles’ movement have. On Peirce’s model, they do.

Peirce’s phenomenology provides an account of the continuity of thought in interpretation. The

levels of sign-interpretation are triadic. First, are emotional, atomic, interpretants; second,
energetic interpretants; and, third, are logical interpretants (CP 5.475). The triadic scheme
classifies any object of thought, and presumably, thoughts themselves (as they may be objects of
thought, too). For our purposes, the most useful version of the view is rendered in terms of the
logic of relations, where the relata stand in one of three kinds of irreducible relations: monadic,
dyadic, and triadic. For example, some x’s redness is monadic, some x bumping into some y is
dyadic, and some A giving some z to B is triadic.6

Firsts are simple monadic elements, immediate qualities of feeling, non-comparative properties.
Seconds are actions and reactions, causes and effects, the brute force of resistance. As Peirce
describes secondness it is that “which the rough and tumble of life renders most familiarity
prominent. We are continually bumping up against hard fact” (CP 1.324). Thirds are lawlike
mediated relations. Signs and significant behavior are clearly the best candidates, for their status
requires the relations of who-the-sign-is-for, what-the-sign-is-of, and the sign itself. Thoughts
themselves, again on the semiotic theory, are thirds, as they have their combinatorial elements,
but they are interpretations of other thoughts and experiences, and for them to mean, they must
be interpreted.

Firsts, as feelings, are the primary constituents of consciousness – they are the data that give rise
to empirical judgments. They, however, do not function as sense-data or immediate objects of
awareness. Rather, they are the raw fluid material of awareness. Because they are prior to
interpretations, we are (on Peirce’s theory) not aware of firsts as such. We come to awareness of
them as discrete seconds – the fact of some quality bearing on us, the world bumping back, the
resistance of things to our will. Facts impose themselves on us in experience, and that pressing is
announced with the presence of sensory properties. But for us to recognize those facts as such, to
understand and make inferences on the basis of our experiences, they must be taken as being
distinct law-like events indicative of their correlate facts (CP 1.420). As such, thirdness of
thought is the abduction of understanding of what is happening in the world on the basis of the
secondness of experience. Returning to the analogy with Achilles and the tortoise, the distinct
movements of thought with thirds depends on the fluid movements of awareness in firsts. The
continuity of consciousness in firstness, however, is the ground for thirds. To see the situation as
paradoxical is to take thirds to be the antecedent elements.

Most certainly thirdness is our primary object of interest for epistemology, and Peirce’s
argument against intuitionism is clearly made on the level of thirds. It is only when thoughts are
interpreted on this level that they may bear inferential relations with each other and are
candidates for the meta-requirement. However, these relations between thirds are not the only
epistemically relevant relations. And though firsts solve the problem of the continuum, there is a
final category that plays an epistemic role.

Secondness has a place in the Peircean epistemic program, especially if we are considering the
possible place of non-inferential justification for intuitions. If an intuition is a thought that
immediately refers to its object, instead of looking at the mediacy of reference on the theory of
signs, the possibility of direct reference may be found in the ontology of the signs themselves.

The place of seconds in the semiotics is the requirement that for thought to have the determinant
logical content it must have to be the kind of thoughts capable of us in a position where we can
have knowledge: something must exist, there must be something that can be knocked up against,
and therefore something in consciousness that cannot be thought away. There must be something
that forces our determinate acknowledgement (CP 2.358). The place of secondness in our
interpretive triads is that it is a condition not just for the being of the objects of our thought, it is
a condition for the existence of the thoughts themselves:

For as long as things do not act upon one another, there is no sense or meaning saying that they
have no being, unless it be that they are such in themselves that they may perhaps come into
relation with others (CP 1.25)

Guy Debrock (1997, 27) has termed this requirement “the Peircean ontological principle”: if
there is no secondness, there is nothing at all. Thirdness, as habits of thought requires that there
be occurrences causally suitable to form those habits.

Consequently, thoughts about the seconds that give rise to them are themselves referentially
direct. A person may say, “I am being appeared to thusly” or “He said that the car wouldn’t look
like that,” or “Oh, this is how it feels to win the race.”7 In making such references, the speakers
must be both aware of the objects of reference and capable of articulating what such references
entail. But the awareness of what they are experiencing is not provided by their awareness of the
context or inferential and practical commitments or consequences. They are aware of the causal
connection between the experiences and their actions. One takes aspirin because headaches hurt,
not vice versa. These seconds, the pressure of the headache, and its correlate effects in our
behavior and our beliefs, have a force that we feel and articulate with assent to proper
descriptions of them. Most certainly there is another interpretive element to bringing this force to
full belief, but the interpretation depends on the material component of the sign for its impetus
(CP 5.292).

So there is a component of thought, on the Peircean theory, wherein there must be non-doxastic
support for beliefs. Seconds in experience impel us to formulate commitments. These
formulations themselves, on what I’d termed the dual role of semiotic interpretation, with causal
and epistemic reasons functioning in tandem, require further inferences, but the impetus itself is
not inferential. It is the condition for inference.8

Peirce’s argument from “Questions” depends on a robust requirement for intuition: namely, that
the intuition provides justification that itself is perspicuous to the subject that has it, and further,
that the subject’s justificatory status for her beliefs must not only be available to the subject’s
reflection, but it must be a constitutive component of the belief. From this thought, Peirce derives
the meta-requirement. The mistake interpretively is to take the requirement unqualifiedly to
entail the justification assessment must obtain for all intuitive or direct epistemic support. Direct
support may come in a variety of forms, and Peirce’s own semiotic is posited on requiring that
some non-inferential support function at the heart of the interpretive enterprise. The conception
of the progress of inquiry itself requires some feature of non-inferential, defeasible support:

Besides positive science can only rest on experience; and experience can never rest on absolute
certainty, exactitude, necessity, or universality (CP 1.55).

And in the 1903 Harvard lectures, Peirce is clearly committed to the thesis that “perceptual
judgments are the first premises of all our reasonings” (5.116), and that “all our knowledge rests
on perceptual judgments” (5.142). Peirce was (at least by 1903) committed to there being
asymmetric and non-doxastic support, but he denied that the criteria for such a role was to be the
restricted categories arising from the project of refuting skepticism. It does require that if there is
knowledge, there is non-inferential support.

However, it seems clear that the argument from the meta-requirement in “Questions” and
“Consequences” is still in play. Beliefs bearing this indexical relation to the seconds from which
they arise must still have further beliefs mediating their inferential relations with other beliefs.
Sensations so far as they represent something are determinate according to a logical law by
previous cognitions (CP 5.291). That we have such and such an experiential event may be
provided by the index, but what the sensation is of must be provided by the event’s relation to
other beliefs. We can have direct non-inferential access to our experiences as seconds, but for us
to know what those experiences are of as thirds requires a measure of interpretation.

We have, then, a model for why Peirce would say that there are no intuitions in the sense of first
thoughts, ones that do not depend on others for their content or justificatory status – for the
thoughts to be thoughts, they must play be the consequence of inferences. The requirements for
justification, further, are that premises may function as premises only if they themselves are
conclusions of further arguments.

However, this does not preclude the direct support provided by indexicals in experience. In fact,
without the secondness of these indexicals, the arguments themselves could never be formulated.
We make, as Peirce frames the alternative to intuitions, “hypothetic inferences from observed
facts” (CP 5.260) This is not to say that this support is indefeasible. In fact, quite to the contrary
– the inferences are hypothetic. The support provided by direct justification is foundationalist in
one way and non-foundationalist in another. In the way it is foundationalist, the direct support
provided by seconds are brute and give rise to the content of interpretants. The support is non-
foundationalist in that these events enter into consciousness and play the roles they do because
they are interpreted and the consequences of related inference. Seconds cause us to believe
things spontaneously, but we arrive at their content by interpretation – the beliefs are direct in
one way, but indirect in another.
Peirce’s early theory of justification, then, is committed to two seemingly inconsistent theses. On
the one hand, there must be an inferential or interpretive feature to all epistemic support.10 This
drives Peirce’s epistemology in the direction of infinitism. On the other hand, there can, and it
seems must, be a non-doxastic support for premises for them to even exist and be justified at all.
This drives Peirce’s epistemology back in an empiricist-foundationalist direction. The result is
the seeming contradiction of on the one hand requiring that all justificatory support be inferential
but on the other hand allowing seconds to justify.

This contradiction, though, arises only if one takes the theories of sources of epistemic
justification in tension here to be exclusivist – namely, that if one allows in foundationalism or
infinitism, one cannot allow in the other. But impure theories are non-exclusivist, and it seems
that Peirce is working with a mixed theory of justification. Seconds may play a role in a series of
reasons, but on the requirement that they do so in combination of a series of interpretations that
give seconds epistemic-semiotic significance. By this, our series of inferences may also be held
accountable to our experience, so that there are observational judgments justified by the contents
of the experience. One has the experiences, and with them, one is justified in having beliefs
depending on those correlate experiences. But these experiential beliefs do not play a singular
role in cognitive life, as their relation to other beliefs is mediated by other beliefs. For a subject’s
belief that she is being appeared to like this to be relevant to her other beliefs, the appearance
must be interpreted as a representation of something, or that it is a symptom of being in some
situation that gives rise to certain representations.

Given that all inferential justification is mediated by interpretation, no indexical belief about
one’s experience (that one is experiencing this) is ever sufficient by itself for the justification of
another belief, but it may play a role in the determination of the other belief’s status (especially if
that other belief entails that one instead be experiencing that). However, whether or not one has
that experience does not depend on other interpretations or inferences, but only on the brute fact
of experience.

Peirce explicitly rejects the notion that the regress argument shows that there must be a first
cognition. His arguments in the 1868 series are precisely designed to show that there are not first
cognitions, because inference is necessary for any cognitive element of life to have semantic or
epistemic value.

However, Peirce also takes himself to have shown how it is possible for these series to begin,
given this requirement. Peirce shows that cognition arises by a process of “unfolding” (CP 5.263)
in which the non-inferential material features of signs are the explanatory, and eventually
justificatory, features for cognitions. He notes in the third paper of the 1868 series, “The Ground
of Validity”:

[I]t does not follow that because there has been no first in a series (of premises), therefore that
series had no beginning... for the series may be continuous, and may have begun gradually... (CP
5.327).

This is to say that the continuity of thought arises less like the beginning of a song with a first
note, but more like someone progressively turning up the volume, and the song fades in. If there
were a first thought, it could neither have content nor normative status, on the Peircean semiotic.
However, thought does arise for individuals, and it seems to come about gradually as a whole.
Again, this is because of the requirements of the semiotic here with the meta-requirement for
first thoughts and the ties between a cognition’s content and its epistemic status. The dual role of
semiotic interpretation as both causally determinative and epistemically justificatory is the locus
of this connection. For a thought to mean, it must be interpreted. Some interpretations are better
than others by their responsiveness to the signs and their syntax, and as such, some are justified
or not. Interpretations have both semiotic-semantic and epistemic roles to play.

One question is which role is conceptually prior. Would Peirce, say, be a semiotic infinitist
independently of his epistemic infinitism, or vice versa? One thought regularly expressed among
contemporary epistemologists is that epistemological theories cannot be articulated
independently of a theory of meaning. If we are interested in inferential justification, then some
notion of the meaning of the variables between the logical functions must be implicated.
Inference is a semantic notion. A justified inference, then, is a conceptually dependent notion –
we’ve got to get our semantics nailed down to do epistemology properly.

Peirce’s project is structurally similar – it is Peirce’s semiotic commitment that thought must be
understood as inferential that yields (a) the meta-requirement for intuitions, (b) the intelligibility
of a regress of interpretations, and (c) the demonstrative relation between thoughts and the
seconds that give rise to them. Peircean infinitist epistemology rides piggyback to Peircean
infinitist semiotics.

3.

The problem is that Peirce’s 1868 semiotics were flawed on three fronts. First, Peirce’s theory of
indexicals was inadequate to the task. Or, at least, Peirce later thought so. All the demonstrative
indexes are limited to seconds and thirds in consciousness. This leaves Peirce’s semantics
incapable of analyzing reference to objects in the world.

Peirce’s case for demonstratives in “Consequences” is that they bear a “real physical connection
of a sign with its object, either immediately or by its connection with another sign” (CP 5.287).
The question, then, is how these connections are established. On Peirce’s semantics, it could only
occur as a static system of relations between judgments. That is, if intuitions have the meta-
requirement, then demonstratives do too. As a consequence, on the 1868 Peircean model, there is
no escape from the circle of judgments.11 Peirce must have realized this problem, as “Fixation” in
1877 has it such that “external objects affect our senses according to regular laws, and that we
may, on the basis of this knowledge, come to know them.” (CP 5.384). The mediacy of
judgments and interpretations of knowledge of things other than judgments had been dropped for
a causal theory of reference.12 Given that Peirce embraced infinitism on the basis of the
requirement for interpretive mediacy for reference, his later acceptance of such reference seems
to withdraw motivation from taking the view on. Peirce still could have been an infinitist in the
1877–78 papers (“Fixation” and “How to Make our Ideas Clear”) and later, but he had no
positive reason to be so.
Second, Peirce would also have had to explain how demonstrative reference to seconds provides
epistemic support. Insofar as Peirce’s requirements for support run that only inference can
provide support, it seems that Peirce’s theory may suffer from a familiar dilemma: either the
demonstrative reference is cognitive or it isn’t. If it is cognitive, then it must be so because it has
been interpreted. If so, then it cannot provide epistemic justification (since that status is possible
only if it is conferred by a further cognitive state). If the reference is not cognitive, then it cannot
provide support, because it cannot function in inference.13

So long as Peirce is committed to the requirement that inference is the sole means of support,
then it is still unclear how seconds may justify any other state. They may cause them (which is
how they arise), but without the inferential-interpretive component, they cannot, on Peirce’s
requirements, play an epistemic role. This, again, is a consequence of what I earlier termed the
dual role of semiotic interpretation as both causally determinative of content and epistemically
justifying. It seems with seconds, we have a question as to how the epistemic element arises
without further inferences.

The third problem is that it is unclear how Peirce’s account of interpretation in the 1868 essays
has anything to do with truth. We know that signs, in order to mean, must be interpreted by other
signs. But we have no criteria for what a correct or incorrect interpretation is. Peirce must be
committed to this difference, as it is crucial to save his theory that all thought is modeled on
valid inference (CP 6.338). He explains cases of fallacious reasoning as cases where one
misinterprets or confuses a rule of inference, and thereby uses the wrong one (CP 5.282).

But what is this notion of misinterpretation? On the semiotic theory, inferences and content of
the thoughts they move from are established by other inferences. How, then, could they be
distinct from the wrong ones, since ex hypothesi they are both logically determinate enough to be
cognitively significant? If it is a brute fact of the matter which is right, then there are components
of consciousness that have their cognitive contents independently of interpretation. The theory of
seconds earlier is supposed to explain how interpretations come about, but it is silent on whether
those interpretations are correct or incorrect. Consequently, if the 1868 theory is to be salvable, a
more robust theory of seconds is necessary.

A closing question is whether another Peircean epistemic infinitism could be abstracted from the
model for inquiry and truth that is incipient in the 1868 articles:

[T]he real ... consists of those which, at a time sufficiently future, the community will always
continue to re-affirm (CP 5.312).

And the theory runs through the later work (1901):

Truth is that accordance of the abstract statement with the ideal limit towards which endless
investigation would tend to bring scientific beliefs (CP 5.565).

The problem for such a model for the purposes of epistemic infinitism is that it underdetermines
the theory of justification necessary for the task. Let us take the theory of reality and truth here
on an ideal convergence model of infinitely extended investigation. This is admittedly a
contentious interpretation. However, if we do not take the model as requiring infinite
investigation, there is clearly no tie to epistemic infinitism. The question is whether epistemic
infinitism can be tied to the theory of truth as interpreted as requiring infinite inquiry. If the
theory of truth articulated here is not a convergence theory, then it will not support an infinitist
epistemology. That seems clear. However, if the convergence theory of truth is right, the
question is what follows for Peirce’s epistemology?

On the model of ideal convergence of infinite inquiry, does epistemic infinitism follow? The
answer is that endless investigation does not mean infinite series of justifying inferences. Endless
investigation of a proposition’s truth, at any time, may change which other propositions are in
the series, change their order, or add new inferences. Or it may reveal that the series does not
provide the requisite support for p by defeating some branch necessary for p’s justification. Or it
may add new nodes to the series infinitely, so that the series may stretch to infinity. But there is
nothing to the notion of infinitely extended inquiry that means that the series of justificatory
inferences it yields are infinite or otherwise.

Endless inquiry, I believe, is reflective of the fallibilist component of Peirce’s epistemology and
its requirement of revisability. The notion of truth at the end of endless inquiry is more a model
for an ethic of well-regulated inquiry than a model for epistemic justification, infinitist or not.
We may be justified now, but that does not make our beliefs immune to correction when new
information comes in. We may be open to infinite iterations of correction and revision, but such
a structure itself does not require that the supporting reasons be infinite. In this respect, Peirce’s
later theory of inquiry may be called epistemically infinitist, because his epistemology does not
rule out the possibility of infinitely extended chains of inferences, but such an infinitism (as
opposed to his views in 1868) is hardly explicit enough to merit little more than a qualified
attribution of the view. The later view is, on what’s given, consistent with infinitism, but hardly
an endorsement.

Peirce was an infinitist in his early work, namely in the 1868 series of articles in The Journal of
Speculative Philosophy. I’ve argued that the infinitism may be understood on what I’ve termed
impure infinitism. However, Peirce’s infinitism was held on the basis of a flawed semiotic
theory. Peirce himself later recognized these defects, and in correcting them, it seems he also
took leave of the epistemic view. In any case, he no longer felt it necessary to explicitly endorse
it.

ACKNOWLEDGMENTS

Thanks to Caleb Clanton, Allen Coates, Charlie Hobbs, Cornelis de Waal, Michael Hodges, John
Lachs, Joseph Margolis, Brian Ribeiro, Andrew Sergienko, Aaron Simmons, and Robert Talisse
for their many helpful comments on and discussions regarding this paper.

NOTES

1. Peter Klein makes a case for this reading of Sextus in 2004. See Sextus’ use of the regress
horn of the Agrippan modes (PH I.168).
2. References to Peirce’s Collected Works will be designated: (CP, Volume #, p#).

3. Klein and Fantl use different criteria for their arguments. Klein’s requirements are those of
non-arbitrariness and non-circularity (1999, 222), and Fantl’s are intelligibility of complete and
degreed justification (2003, 538).

4. See Davis (1972, 8) for an account of the meta-requirement. Note, also, that this requirement
is analogous to the Pyrrhonean requirement for regress-ending hypotheses (PH 164–177) and
also has found other forms in BonJour’s meta-justifications argument (1985, 30–33), and have
figured widely in response to foundationalism. See Oakley (1976, 222–223), Possin and
Timmons (1989, 206) and Klein (1999, 277–279).

5. It seems though, that if Peirce is willing to accept the first order level with the regress, then
why should a regress of intuitions be objectionable at the second order? Couldn’t it be intuitive
that it is intuitive that ... that it is intuitive that p? So long as intuition functions, for those who
have them, as transparent (i.e., that one simply grasps the content, not the embedding), such
embeddings would not be an issue.

6. See Cheryl Misak’s discussion of Peirce’s categories (2004, 22).

7. See Peirce’s discussion of pure demonstrative applications of thought-signs (CP 5.296).

8. Cf. Delaney (1993, 89 and 111–118) who takes this point as indicative of a weak
foundationalism in Peirce’s early works.

9. This strategy is noted by Floridi (1997, 54), who extends Perice’s response to skeptics as an
argument for convergent realism.

10. See CP 5.318 for Perice’s explicit endorsement of this thesis.

11. See Thompson (1978, 79) and de Waal (1996, 436) for versions of this problem for Peirce’s
early semantics.

12. Hookway (2002, 18 and 29–30) notes this difference and attributes Peirce’s change of mind
to Royce’s influence. See also Short (2004) for an account of the development of Peirce’s
semiotics and theory of indexes.

13. Versions of this familiar dilemma can be found in Sellars (1963, 146–147), Davidson (1986,
311), Rorty (1979, 95–96), and BonJour (1985, 69).

Klein’s Case for Infinitism

Ram Neta
University of North Carolina, Chapel Hill
Infinitism is both a theory of epistemic justification and a solution to the regress problem. To
specify the content of infinitism more precisely than this requires some stage-setting. I devote
section I of this paper to that stage-setting, and to stating the content of infinitism more precisely.
Section II gives a sympathetic rendering of Klein’s argument for infinitism. Section III rebuts
what I take to be the most compelling objections to that argument, and shows how we need to
interpret Klein in order to issue those rebuttals. In the concluding section, I will briefly sketch
what seems to me to be a more promising strategy for rebutting Klein’s argument.

Section I: What is Infinitism?

Are there propositions that you can be justified in believing even without having a reason to
believe them? It may seem that there are. For instance, you are currently justified in believing
the proposition that 1>0, and it may seem that you need not have a reason for believing that
proposition in order to be justified in believing it. Or, to take another example, you are currently
justified in believing the proposition that you are conscious, and yet, once again, it may seem
that you need not have a reason for believing that proposition in order to be justified in believing
it. Let us use the term “basic” to denote just the propositions in this category: those such that
your justification for believing them does not require you to have a reason for believing them.
Our initial question, then, may be put as follows: are there any basic propositions?

Of course, even if it is true that some of the propositions that you are currently justified in
believing are basic for you, it’s also true that many of the propositions that you are currently
justified in believing are not basic for you. For instance, I am currently justified in believing that
the high temperature today will be 61 degrees Fahrenheit, but my justification for believing this
depends upon my having a reason to believe it (viz., that this is what the weather forecast said
this morning). So whether or not there are basic propositions, there surely are non-basic
propositions.

Suppose, then, that you are justified in believing a non-basic proposition p. Then you have a
reason for believing that p – call this reason r1. Suppose also that you can be justified in
believing p by virtue of having r1 only if r1 is a proposition that you are justified in believing.
Then, either r1 is basic for you or it is not. If it is not basic for you, then you have a reason to
believe r1 – call this reason r2. Again, suppose that you can be justified in believing r1 by virtue
of having r2 only if r2 is a proposition that you are justified in believing. Then, either r2 is basic
for you or it is not. If it is not basic for you, then you have a reason to believe r2 – call this
reason r3. Again, suppose that you can be justified in believing r2 by virtue of having r3 only if
r3 is a proposition that you are justified in believing. Then, either r3 is basic for you or it is not.
If it is true in general that justifying reasons for belief must be propositions that the believer is
justified in believing, then this regress ends only once we arrive at a proposition that is basic for
you. But must this regress end?

Most philosophers assume that it must, and are thereby committed to the claim that either there
are propositions that are basic for you (foundationalism), or else that justifying reasons for a
proposition need not be other propositions that the believer is justified in believing (externalism).
The coherentist denies both of these disjuncts, and claims that the regress never ends, but only
goes round in a circle.

But infinitism denies all of these claims. According to infinitism, the regress does not end, nor
does it go around in a circle. On the contrary, the regress goes on forever, and no proposition
reached in the regress is ever reached again at another stage of the regress. This is what I will
call the infinitist’s “solution to the regress problem”.

Klein sometimes describes the regress problem as the question “which type of series of reasons
and the account of warrant associated with it, if any, can increase the credibility of a non-evident
proposition?”1 where a “non-evident proposition” is one concerning which there could be
credible disagreement. This particular way of describing the regress problem is both historically
unfamiliar, and also seemingly quite different from the problem raised by the regress that I’ve
just described above. But we will see in section III, when we try to defend Klein’s argument for
infinitism, why it is that Klein describes the regress problem in this way.

As I said above, infinitism is not merely a solution to the regress problem. It is also a theory of
epistemic justification. Not every theory of epistemic justification also serves as a solution to the
regress problem. Consider, for instance, the process reliabilist theory of justification. According
to such a theory, a belief’s being justified consists in its being formed by means of a reliable
process. What does this theory imply about whether or not the regress of justifiers must end, or
whether it can go in a circle? Nothing. The process reliabilist theory of justification is consistent
with the view that the regress must end, and it is also consistent with the view that the regress
does not end. (It is open to a process reliabilist to claim, for instance, that a belief-forming
process can reliably form the belief that p only if p is the terminus of a regress of reasons that
stretches infinitely far back, each of which is itself reliably formed.) Process reliabilism is
consistent with the view that the regress goes in a circle, and it is also consistent with the view
that the regress does not go in a circle. In short, a process reliabilist theory of justification has no
implications whatsoever for the problem. Neither does a reliable indicator theory of justification.
Neither does a virtue theory. In short, a theory of justification need not be a solution to the
regress problem. And the converse is also true: a solution to the regress problem need not be a
theory of justification. But infinitism, unlike the other aforementioned theories of justification, is
simultaneously a theory of justification and a solution to the regress problem.

The infinitist gives a theory of propositional justification, and also of doxastic justification. The
most succinct statement of each theory is given by John Turri, as follows:

“Infinitist propositional justification (IPJ): The proposition Q is propositionally justified for

you just in case there is available to you at least one infinite non-repeating series of propositions
(or reasons) such that R1 is a good (and undefeated) reason to believe Q, R2 is a good (and
undefeated) reason to believe R1, R3 is a good (and undefeated) reason to believe R2, … Rm+1 is a
good (and undefeated) reason to believe Rm, for any arbitrarily high m.
“Infinitist doxastic justification (IDJ): Your belief that Q is doxastically justified just in case Q
is propositionally justified for you, and you have provided enough reasons along at least one of
the infinite non-repeating series of reasons, in virtue of which Q is propositionally justified for
you, to satisfy the contextually determined standards.”2
IPJ and IDJ are both very unorthodox claims. What could Klein have to say on their behalf?
Roughly, the grounds on which Klein recommends both forms of infinitism are that only they
can explain how it is possible for someone to have a non-question-begging and non-dogmatic
answer to any question of the form “why think that p?”, when such a question is directed towards
a person who regards the particular value of p at issue to constitute a reason for something else
she believes. According to Klein, the coherentist can offer only question-begging answers to
such questions, and the foundationalist can offer only dogmatic answers to some such questions
(i.e., those that concern what she herself counts as foundationally justified propositions). But
how can Klein justify these bold claims? We turn to that question in the next section.

Section II: What is Klein’s Case for Infinitism?

Klein’s clearest statement of his argument for infinitism occurs in Klein 1999, and is an
argument for IPJ specifically. Klein can argue from IPJ to IDJ as follows:

(i) IPJ

(ii) Your belief that Q is doxastically justified just in case Q is propositionally justified for you,
and your belief that Q is properly based on what propositionally justifies Q for you.

(iii) Your belief that Q is properly based on what propositionally justifies Q for you just in case
you have provided enough reasons along at least one of the infinite non-repeating series of
reasons, in virtue of which Q is propositionally justified for you, to satisfy the contextually
determined standards.

(iv) IDJ

This argument is valid, and premise (ii) is uncontroversially true. While premise (iii) is highly
controversial, it is plausible that if (i) is true, and if we have any doxastically justified beliefs at
all, then (iii) is true. That’s because doxastic justification comes not simply from holding a
propositionally justified belief, but from holding such a belief on the basis of what
propositionally justifies it. However precisely this “basing” requirement is to be understood, it is
quite plausible that, if (i) is true, then the condition laid down in (iii) is necessary to satisfy that
requirement. And so the fundamental controversy surrounding the argument above for IDJ
concerns step (i), viz., IPJ. In the rest of this paper, I focus exclusively on Klein’s argument for
IPJ.

In order to understand his argument for IPJ, I must first introduce and define a term that Klein’s
argument employs. This is the term “evidential ancestry”. Here’s what Klein says by way of
explaining his use of the term: “By ‘evidential ancestry’ I am referring to the links in the chains
of reasons, sometimes branching, that support beliefs. For instance, if r is a reason for p, and q is
a reason for r, then r is in the evidential ancestry of p, and q is in the evidential ancestry of both p
and r.”3 For a reason to “support” a belief, in the sense that Klein has in mind in this passage, the
reason must be both objectively and subjectively available to the believer. Klein does not
commit himself to any specific account of what is involved in either objective or subjective
availability, but roughly, a reason r is objectively available for a proposition p only if r increases
the credibility of p, and r is subjectively available for a proposition p only if the believer has the
appropriate sort of subjective relation to r.

We can use the term “evidential ancestry” to distinguish foundationalist theories of propositional
justification (FPJ), coherentist theories of propositional justification (CPJ), and IPJ as follows:
FPJ is the view that the evidential ancestry of any proposition that is justified for a person at a
time is finite and contains no element more than once. (FPJ can treat foundationally justified
propositions as having an evidential ancestry that contains no elements at all, and so has a finite
cardinality of zero.) CPJ is the view that the evidential ancestry of any proposition that is
justified for a person at a time contains some elements more than once. And IPJ is the view that
the evidential ancestry of any proposition that is justified for a person at a time is infinite and
contains no element more than once.
Klein’s argument for IPJ is an argument by elimination. It proceeds from two premises, which
Klein states as follows:

“PAC: For all x, if a person, S, has a justification for x, then for all y, if y is in the evidential
ancestry of x for S, then x is not in the evidential ancestry of y for S.

PAA: For all x, if a person, S, has a justification for x, then there is some reason, r1, available to
S for x; and there is some reason, r2, available to S for r1; etc.”4

If PAC is true, then no evidential ancestry can contain the same element more than once, and so
CPJ is false. If PAA is true, then no evidential ancestry can be finite, and so FPJ is false. (As
Turri 2009 shows, FPJ does not imply that every evidential ancestry is finite, but it does imply
that some are, and so PAA entails that FPJ is false.) If – as Klein assumes – IPJ is the only
logically possible alternative to FPJ and CPJ, then it follows that infinitism is true. If I had space
here, I would call into question Klein’s assumption that IPJ is the only logically possible
alternative to FPJ and CPJ. But that is an assumption that might plausibly seem unassailable to
Klein since none of Klein’s critics to date have assailed it. Indeed, Klein’s critics have all
seemed willing to grant this assumption, and also to grant PAC. What Klein’s critics have done
is, in one or another way, to attack Klein’s case for IPJ by attacking PAA. What I want to do in
this section of the paper is to spell out Klein’s argument for PAA as sympathetically as I can.
Then, in the next section, I will show that Klein can rebut all of the best objections to PAA. If
Klein’s argument for IPJ is flawed, then the flaw may lie elsewhere than in PAA.

So now let’s turn to the issue: what does Klein have to say on behalf of PAA? What he says on
behalf of PAA is, for the most part, not contained in the 1999 paper in which he originally
defends IPJ, but rather in a 2005 paper that initiates a fruitful exchange with Ginet.5 There,
Klein defends PAA by arguing that, if it were false, then a particular kind of obviously illicit
epistemic practice would be licensed:

“imagine a dialogue between Fred, the Foundationalist, and Doris, the Doubter. …Fred asserts
some proposition, say p. Doris says something …that prompts Fred to believe that he had better
have reason(s) for p in order to supply some missing credibility. So, Fred gives his reason, r1, for
p. (r1 could be a conjunction.) Now, Doris asks why r1 is true. Fred gives another reason, r2.
This goes on for a while until Fred (being a practicing foundationalist) arrives at what he takes to
be a basic proposition, say b.

“Doris will, of course, ask Fred for his reason for b. But Fred, being a self-conscious,
circumspect foundationalist will tell Doris that b doesn’t need a reason in order to possess the
autonomous bit of warrant. He will say that her question ‘Why do you believe that x?’ though
appropriate up to this point is no longer appropriate when ‘b’ is substituted for ‘x’ because b is
basic. There is no reason that supplies the autonomous warrant that b has.

“Grant that foundationalism is true; b has some autonomous bit of warrant that arises because b
has some foundational property, F, such that any proposition having F is autonomously
warranted, and every non-basic proposition that depends upon b for its warrant would lose some
of its warrant were b not autonomously warranted.

“Doris should say to Fred, ‘I grant that b has autonomous warrant. But what I want to know is
whether autonomously warranted propositions are, in virtue of that fact, somewhat likely to be
true.’ Here worry becomes a ‘meta’ worry. But she went meta, so to speak, because Fred went
meta first.

“Given that with regard to any proposition, once we consider whether it is true, we must hold it,
deny it, or withhold it (i.e. neither hold nor deny it), Fred is now faced with a trilemma:

1 He can hold that autonomously warranted propositions are somewhat likely to be true in
virtue of the fact that they are autonomously warranted.
2 He can deny that autonomously warranted propositions are somewhat likely to be true in
virtue of the fact that they are autonomously warranted.
3 He can withhold whether autonomously warranted propositions are somewhat likely to be
true in virtue of the fact that they are autonomously warranted.

“If he takes alternative 2, then using b as a reason for the first non-basic proposition in the series
is arbitrary. Holding b is not arbitrary. Doris has granted that b is autonomously warranted and
she could grant that it is not arbitrary to hold a proposition that has autonomous warrant. But if
Fred believes that such propositions were not even somewhat likely to be true in virtue of being
autonomously warranted, how could he think that b could provide a good reason for thinking that
the penultimate proposition was likely to be true? Fred thinks that the warrant for all of his
beliefs rests on basic propositions. If he thought that b’s possession of F was not the least bit
truth conducive, then why is he using b and all the other basic propositions on which the warrant
for his non-basic beliefs rests?
“The same applies to alternative 3. Doris has asked whether the fact that b is autonomously
warranted makes it at all likely that b is true. Fred responds that he doesn’t have an opinion one
way or the other. Fred thinks b is true but her neither has a reason for thinking it is true nor does
he thinks that basic propositions are somewhat likely to be true because they are autonomously
warranted. So, from Fred’s point of view and Doris’s, Fred ought not to use b as the basis for
further beliefs. The mere fact that he thinks b is true is not sufficient for him to use b as a reason,
unless he thinks that b is true somehow makes it likely that b is true.

“If he takes alternative 1, then using b as his reason for the penultimate proposition is not
arbitrary, but that is because the regress has continued. Fred has a very good reason for
believing b, namely b has F and propositions with F are likely to be true. Fred, now, could be
asked to produce his reasons for thinking that b has F and that basic propositions are somewhat
likely to be true in virtue of possessing feature F.

“Therefore: foundationalism cannot solve the regress problem, even if it were true. A practicing
foundationalist cannot increase the rational credibility of a questioned proposition through
reasoning.”6

It’s not obvious how to interpret the argument given in the passage just quoted. Klein concludes
by saying that foundationalism cannot solve the regress problem; his grounds for this conclusion
seem to concern what a foundationalist (Fred) would be committed to thinking if he were to
consider the issue of whether autonomously warranted propositions are likely to be true. But just
how is the latter issue related to the issue of whether foundationalism can solve the regress
problem? Could it be that, by considering the issue of whether autonomously warranted
propositions are likely to be true, we alter our justificatory situation in such a way that
propositions that were heretofore foundationally warranted are no longer foundationally
warranted? Michael Williams has developed a response to the regress problem according to
which such changes in justificatory status are to be expected.7 An argument for infinitism should
not simply assume that Williams’s epistemological views are false.

I propose then, that we interpret Klein’s appeal to what Fred would be committed to thinking if
he were to consider a particular issue to be an implicit invitation to reflect on what Fred is
already committed to thinking about a particular issue, whether or not he ever thinks about that
issue.

But if we interpret Klein in that way, then we face another difficulty. The issue that Fred is
invited to consider is the issue of whether “autonomously warranted propositions” are likely to
be true. Of course, since Fred is a “practicing” foundationlist, he will already have the concept
of an “autonomously warranted proposition” – but in this respect he is unlike the rest of us, and
so it is not clear just what we can inferences we are entitled to make from his justificatory
situation to our own. Since very few people have the concept of an “autonomously warranted
proposition”, it is not clear that we have any rational commitments to holding any views about
the very general issue of whether such propositions are likely to be true. But this is a problem
that we can avoid as follows: what matters for Klein’s argument is not what Fred (qua
representative thinker) is committed to thinking about the likelihood of truth of all the
propositions in some general category (e.g., the “autonomously warranted” ones). Rather, what
matters is what Fred (qua representative thinker) is committed to thinking about the likelihood of
truth of the particular proposition that he takes to be foundational, and to which he appeals in the
course of defending one of his beliefs. It is the issue of that likelihood that matters for Klein’s
argument. Of course, it might be that the likelihood of a property being instantiated by a single
case is always relative to, or dependent upon, some reference class to which the single case
belongs. But the reference class need not be the particular one that Klein picks out: it could be
lots of others.

In light of these considerations, I believe that Klein’s reasoning in the passage quoted above can
be most charitably rendered as follows:

Premise 1: If PAA is not true, then there are possible cases in which someone can rationally
defend a belief of hers by appeal to what she takes to be a foundationally justified proposition.

Premise 2: Someone cannot rationally do both of the following two things: defend a belief of
hers by appeal to what she takes to be a foundationally justified proposition, and also deny that
that proposition is likely to be true.

Premise 3: Someone cannot rationally do both of the following two things: defend a belief of
hers by appeal to what she takes to be a foundationally justified proposition, and also neither
believe nor deny that that proposition is likely to be true.

Premise 4: If someone rationally defends a belief of hers by appeal to what she takes to be a
foundationally justified proposition, then she is committed to believing that that foundationally
justified proposition is likely to be true. (from 2 and 3, by the general principle that, if ∂ and π
are mutually exclusive and logically exhaustive options, and no one can rationally ø and π, then,
if someone rationally ø’s, she is thereby committed to ∂’ing.)

Premise 5: If someone is committed to believing that a foundationally justified belief (B) is

likely to be true, then she has an available reason to hold belief B.

Premise 6: If someone has an available reason to hold a belief, then that belief is not
foundationally justified for her.

Premise 7: If someone rationally defends a belief of hers by appeal to what she takes to be a
foundationally justified proposition, then that proposition is not foundationally justified for her.
(from 4, 5, 6)

Premise 8: There are no possible cases in which someone rationally defends a belief of hers by
appeal to what she correctly takes to be a foundationally justified proposition. (from 7)

Premise 9: If there are no possible cases in which someone rationally defends a belief of hers by
appeal to what she correctly takes to be a foundationally justified proposition, then there are no
possible cases in which someone can rationally defend a belief of hers by appeal to what she
takes to be a foundationally justified proposition. (This premise is not made explicit anywhere in
Klein’s passage, but Klein must hold something like this premise in order for his argument to be
valid.)

Conclusion: PAA is true. (from 1, 8, 9)8

Although this argument is valid, there appear to be grounds for doubt concerning some of its
premises. One might question premise 1: couldn’t PAA be false, and there be some
foundationally justified propositions, even if nobody took any propositions to be foundationally
justified? At first blush, this seems clearly possible: foundationalism could be true even if
everyone consistently denied it, and so even if no one ever took any proposition to be
foundationally justified. One might question premise 3: couldn’t someone defend a belief of
hers by appeal to what she took to be a foundationally justified proposition, even if she held no
view on the issue of whether that proposition is likely to be true? Once again, this seems to be
clearly possible: the issue of likely truth might never have occurred to the agent in question, and
this wouldn’t seem to impugn her ability rationally to defend one of her beliefs by appeal to what
she took to be a foundationally justified proposition. One might question premise 5: couldn’t
one be committed to believing that foundationally justified propositions are likely to be true, and
yet still have no available reason for believing any particular proposition that she took to be
foundational? Her belief that foundationally justified propositions are likely to be true might
seem not to provide her with such a reason if, say, that belief is itself based on the foundationally
justified proposition in question. And finally, one might question premise 9: couldn’t there be
circumstances under which one could rationally defend one’s beliefs by appeal to what one took
to be foundationally justified propositions, even if one could not rationally defend them by
appeal to what one correctly took to be foundationally justified propositions? If any of these four
premises are false, then the argument above is not sound, even if it is valid.

In the next section, I will address the doubts just raised concerning each of these four premises,
and reformulate Klein’s argument in a way that renders it invulnerable to those doubts. But
before I do that, I want to say something about why it is that Klein describes the regress problem
as the rather unfamiliar sounding question “which type of series of reasons and the account of
warrant associated with it, if any, can increase the credibility of a non-evident proposition?”
This unorthodox description of the regress problem may suggest that Klein means for infinitism
to solve a problem that is very different from the historically familiar regress problem described
in section I above. But this suggestion would be misleading: Klein clearly does think of
infinitism as, among other things, a solution to the historically familiar regress problem. But
then why does he offer such an eccentric description of the regress problem? In the next section,
we will see that the answer to this question helps us to understand how Klein can rebut the four
objections issued above against his argument.

Section III: Defending Klein’s Argument from Objections

As I said, there are apparent grounds for doubting premises 1, 3, 5, and 9 of Klein’s argument for
infinitism. In this section, I would like to assess those grounds for doubt, and elaborate Klein’s
argument in a way that renders it invulnerable to them. But in order to do this, I must first
address a question that arose above, when we noted the difference between the orthodox
presentation of the regress problem above, and Klein’s highly idiosyncratic presentation of the
regress problem. The question is: why does Klein present the regress problem in such an
idiosyncratic way? Once we know the answer to this question, we will be in a position to see
how Klein can defend premise 1. Then we can go to see how he can defend premises 3, 5, and 9.

So how does Klein present the regress problem? He begins with a quote from Sextus Empiricus:

“The locus classicus of the regress problem is to be found in Sextus Empiricus’s Outlines of
Pyrrhonism:

‘The later Skeptics hand down Five Modes leading to suspension, namely these: the first based
on discrepancy, the second on the regress ad infinitum, the third on relativity, the fourth on
hypothesis, the fifth on circular reasoning. That based on discrepancy leads us to find that with
regard to the object presented there has arisen both amongst ordinary people and amongst the
philosophers an interminable conflict because of which we are unable either to choose a thing or
reject it, and so fall back on suspension. The Mode based upon regress ad infinitum is that
whereby we assert that the thing adduced as a proof of the matter proposed needs a further proof,
and this again another, and so on ad infinitum, so that the consequence is suspension [of assent],
as we possess no starting-point for our argument. The Mode based upon relativity… is that
whereby the object has such or such an appearance in relation to the subject judging and to the
concomitant percepts, but as to its real nature we suspend judgment. We have the Mode based
upon hypothesis when the Dogmatists, being forced to recede ad infinitum, take as their starting-
point something which they do not establish but claim to assume as granted simply and without
demonstration. The Mode of circular reasoning is the form used when the proof itself which
ought to establish the matter of inquiry requires confirmation derived from the matter; in this
case, being unable to assume either in order to establish the other, we suspend judgment about
both.’

“Although the three alternative strategies for solving the regress will be the focus of this essay, a
brief discussion of the other two modes will be useful in understanding what initiates the regress.

“The Modes were recipes for avoiding dogmatism, i.e. the disposition to assent to non-evident
propositions when it is not settled whether they are true. One could locate such a non-evident
proposition either by noting that there was credible disagreement about it or by merely
recognizing that there could be credible disagreement. For in order to avoid epistemic hubris,
the recognition that our epistemic peers could sincerely disagree with us about the truth of some
proposition forces us to regard it as requiring reasons in order to rise to the desired level of
credibility.

“The Regress Problem can be put as follows: Which type of series of reasons and the account of
warrant associated with it, if any, can increase the credibility of a non-evident proposition? Can
a series with repeating propositions do so? Can one with a last member do so? Can one that is
non-repeating and has no last member do so?” (Klein 2005a, 131 – 2)

It may seem puzzling that Klein describes the Sextus passage as the “locus classicus of the
regress problem”, since it is not clear from that passage what “the regress problem” could be.
Sextus is simply mentioning five ways in which the later skeptics attempted to induce suspension
of judgment in their audience, but he does not, in that passage, describe anything that he takes to
be a problem. It is Klein who identifies what he calls the “the regress problem”. But what is the
relation between the problem that Klein identifies as such, and the five modes that Sextus
mentions?

Klein points out that the five modes were intended as correctives to the tendency to assent to
propositions the truth of which is not evident. Whenever it’s possible for epistemic peers to
sincerely disagree as to the truth of a proposition, the truth of that proposition is not evident. But
if the five modes lead us to suspend judgment concerning the truth of all such propositions, then
must our judgment remain confined to those few propositions concerning which disagreement
among epistemic peers is impossible (e.g., 1 + 1 = 2, I am conscious now)? The only way to
answer this question in the negative is to find a way to guide our judgment concerning the truth
of non-evident propositions. Klein takes the regress problem to be the problem of finding such
ways, i.e., finding ways to guide our judgment concerning the truth of non-evident propositions.
More precisely, since we cannot expect to find some small set of procedures to guide our
judgment concerning the truth of any and all non-evident propositions, the problem is rather to
specify the constraints on any such procedure: what would a procedure have to be like in order
for it to guide our judgment concerning the truth of non-evident propositions? That is what
Klein takes the regress problem to be.

But how is this very practical sounding “regress problem” related to the more familiar-sounding
but highly theoretical regress problem described back in section I: the problem of understanding
the structure of propositional justification? And why does Klein himself say nothing about how
they are related? The more familiar-sounding problem is a theoretical problem concerning the
structure of propositional justification. But what is propositional justification? We cannot
explain it in terms of what a person would be justified in believing were she to believe it: such
an account would have the absurd consequence that no one could ever be propositionally
justified in believing that they don’t hold beliefs about their own beliefs. We also cannot explain
propositional justification in terms of what justifies a person in holding a belief, for someone
could have propositional justification for a belief that she does not hold. While it seems doubtful
that we can give any non-circular account of propositional justification, the following at least is
plausible: propositional justification is that in light of which a belief is justified.

Some internalists think that a belief can be justified in light of some factor or other only in virtue
of the believer’s being somehow aware of that factor, and of its serving to justify the belief. But
this form of internalism is highly controversial. There is, however, a weaker internalist thesis
that, while still controversial, is more generally plausible: this is the thesis that a belief can be
justified in light of some factor or other only in virtue of the believer’s being capable, by means
of reflection, of becoming aware of that factor, and of its serving to justify the belief. Whatever
justifications we possess cannot be beyond our reflective reach, even if they are not currently
within our reflective grasp. I will henceforth call this the “reflective accessibility” constraint on
propositional justification. As we will see, if Klein accepts this reasonable (albeit not
uncontroversial) reflective accessibility constraint, then this would help to explain why he frames
the regress problem in the way that he does, and also why he takes the argument for PAA offered
above to be sound. For the remainder of this paper, I will therefore assume, for charity’s sake,
that Klein accepts the reflective accessibility constraint on propositional justification.
Notice that the reflective accessibility constraint on propositional justification is compatible with
a wide variety of views, including radically externalist views, concerning propositional
justification: it is compatible with process reliabilism, indicator reliabilism, various versions of
virtue theory, and even the recently propounded view that a belief is justified only if it is
knowledgeably held. Each of these views is compatible with the claim that what gives a subject
justification for holding a belief does so only by virtue of its being recognizable, upon reflection,
as doing so. To say that justifiers need be recognizable as such is not to imply that they do not
have some epistemically inaccessible properties as well.

We can take it, then, that if something constitutes propositional justification to believe that p,
then it does so by virtue of the epistemic agent’s ability to recognize it (upon reflection) as
constituting such justification. And so, for instance, a series of reasons can constitute a
justification for S to believe that p only if that series of reasons is recognizable by S as
constituting a justification to believe that p. But if S knows that her epistemic peers – those who
share all of S’s reasons – could reasonably disagree with her about whether p is true, then S
cannot also rationally regard her current reasons for believing that p to constitute a justification
to believe that p: those reasons don’t suffice to rule out reasonable denial of p. (I assume, on
Klein’s behalf, that, even if we can rationally regard reasons as somewhat permissive, we cannot
rationally regard them as radically permissive. That is, we cannot rationally allow that a body of
reasons sufficient to justify one in believing that p is also, on its own, sufficient to justify one in
denying that p.) More generally, if S knows that p is not evident to her (if, say, all the evidence
on the basis of which she believes p is evidence on which she could, by her own lights, equally
reasonably believe something incompatible with p), then S cannot rationally take her current
reasons for believing that p to constitute a justification for believing that p. In order for S to
acquire something that she can recognize as a justification for believing that p, she must gain
additional support for p, and this additional support must be in the form of something that S can,
at least upon reflection, recognize as bolstering her justification for believing that p. And so,
Klein’s own statement of the regress problem is simply the form that the familiar theoretical
problem assumes when it concerns propositions that we know not to be evident and that we want
to gain justification for believing. Klein’s regress problem is not an altogether different problem
from the more familiar one stated in section I above, but rather a version of it that arises as a
practical matter for agents who want to have justification for believing things that are not
initially evident.

We established this conclusion by assuming that Klein accepts the reflective accessibility
constraint on propositional justification. But, if propositional justifications are recognizable as
such to the epistemic agent, then each of the person’s justifications must be recognizable to her
as such. And if each of a person’s justifications is recognizable to her as such, then this suggests
(though it does not imply) that, if the structure of an agent’s propositional justifications is finite,
then that structure may perhaps be recognizable by her upon reflection. Just as my ability to see
each Lego in a tower of Legos typically suffices for me to be able to see the whole tower as such,
and my ability to hear each note in a sonata typically suffices for me to be able to hear the whole
sonata as such, so too my ability to recognize each justification in a finite structure of
justifications might suffice for me to be able to discern that structure as such. In other words,
perhaps my reflective access to my justifications gives me reflective access to the structure
formed by those justifications, at least assuming that structure is finite. While this claim, which
I’ll heretofore call “structure accessibility” is far from obvious, it is not entirely implausible
either. And, as we will see, attributing this assumption, along with the assumption of reflective
accessibility, to Klein helps us to understand how he can rebut all of the objections raised above
to his argument for PAA.

Recall that the objection to premise 1 was this: couldn’t PAA be false, and there be some
foundationally justified propositions, even if nobody took any propositions to be foundationally
justified? To see how Klein should respond to this objection, let’s first consider more carefully
what is involved in someone taking a proposition p to be foundationally justified. Presumably,
this need not involve something as conceptually sophisticated as believing that p is
foundationally justified – that is a belief that very few people have the conceptual sophistication
to hold, since very few people have the concept of foundational propositional justification.
Taking p to be foundationally justified must be more conceptually primitive than that, if premise
1 is to be at all plausible. Could it involve as little as simply this: confidently believing that p,
and also not regarding anything distinct from p as one’s reason for believing that p? No. While
confidently believing that p might plausibly be regarded as a way of (implicitly) taking p to be
justified, we cannot identify one’s failing to regard anything distinct from p as one’s reason for
believing that p with taking one’s justification for p to be foundational: one might fail to regard
anything distinct from p as one’s reason for believing that p simply because it doesn’t occur to
one to think about one’s reason for believing that p. It would be more plausible to think of one’s
modally robust disposition – even upon reflection – not to regard anything distinct from p as
one’s reason for believing that p as a way of taking one’s justification for p to be foundational. It
is not easy to see how a creature lacking the concept of foundational justification could take a
particular belief of theirs to be foundationally justified by virtue of anything less than having
such a modally robust disposition with respect to one of their own confidently held beliefs.

But notice that, if PAA is not true, then the structure of a person’s reasons for any particular
proposition is finite. And if that structure is finite, then, according to structure accessibility, that
structure is reflectively accessible as such to that person. Furthermore, if PAA is not true, then
there is at least one proposition such that S is justified in believing that proposition but S has no
reason to believe it: the proposition is foundationally justified. Finally, if structural accessibility
is true, and if some proposition is foundationally justified for me, then I should be able to
recognize, upon reflection, that that proposition is foundationally justified for me.

Thus, I suggest that Klein can defend premise 1 of his argument above as follows:

(a) If PAA is not true, then there is some proposition p that is foundationally justified for
someone S.
(b) If PAA is not true, then S’s structure of reasons is finite.
(c) If structural accessibility is true and S’s structure of reasons is finite, then S can recognize
her structure of reasons upon reflection.
(d) If S can recognize her structure of reasons upon reflection, and there is some proposition
p that is foundationally justified for S, then S can recognize upon reflection that p is
foundationally justified for S.
 (e) If PAA is not true, and if structural accessibility is true, then there is some proposition p
that S can recognize upon reflection to be foundationally justified for S. (from a, b, c,
and d)
(f) If S can recognize upon reflection that p is foundationally justified for S, then S can
rationally defend some beliefs of hers by appeal to what she takes to be a foundationally
justified proposition.
(g) Structural accessibility is true. (Assumption, attributed to Klein for sake of charity.)

Premise 1: If PAA is not true, then there are possible cases in which someone can rationally
defend a belief of hers by appeal to what she takes to be a foundationally justified proposition.

Notice that Premise 1 of the argument above follows from a – g, and the only controversial move
in that argument is g (viz., structural accessibility). If we attribute to Klein an acceptance of
structural accessibility – rendered at least somewhat plausible in light of the reflective
accessibility constraint on justification – then Klein is in a position to establish premise 1 on the
basis of what he is committed to regarding as a sound argument. This would suffice to rebut the
envisaged objection to premise 1.

Now recall that the objection to premise 3 was this: couldn’t someone rationally both defend a
belief of hers by appeal to what she took to be a foundationally justified proposition, and also
have no view on the issue of whether foundationally justified propositions are likely to be true?
Once again, this seems to be clearly possible: the issue of likely truth might never have occurred
to the agent in question. But the question is not whether a particular psychological state is
possible; the question is rather whether a particular psychological state can be rational. Can it be
rational simultaneously to defend a belief by appeal to what you take to be a foundationally
justified reason, and also to refrain from regarding foundationally justified reasons as probably
true? No; here’s the argument:

(a’) If you can rationally defend a belief of yours by appeal to some reason r, then you
can rationally take r to be your reason for holding that belief. (This follows from the
accessibility constraint on justification.)

(b’) You can rationally take r to be your reason for holding a belief only if you can
rationally take r to be a good reason for holding a belief. (If you took r to not be a good
reason for the belief, but to be your one and only reason for the belief, you could then no
longer clearly count as holding the belief; how can you endorse the content of a belief
that you hold for what you yourself take to be a bad reason?)

(c’) You can rationally take r to be a good reason for holding a belief only if you can
rationally take r to be at least probably true.

(d’) If you can rationally defend a belief of yours by appeal to some reason r, then you
can rationally take r to be at least probably true. (by a’ – c’)

But notice that, if this argument from (a’) – (d’) is sound, its soundness can be recognized simply
by reflection on the argument and its premises. And reflection on the argument reveals that its
soundness is independent of the value of r. Thus, by reflection on the argument above, we can
reach the following general conclusion:

(e’) For any reason r, if you can rationally take r to be a good reason for holding a belief,
then you can rationally take r to be probably true.

In other words, what you rationally regard as your reason for belief is such that you are rationally
required to regard it as probably true. This conclusion holds independent of whether the reason
at issue is foundationally justified or not: the foundationally justified reasons are not going to be
different from other reasons in this respect. And so, in general:

(f’) What you rationally regard as your foundationally justified reasons for belief are also
rationally regarded as probably true.

In short, what you rationally regard as a foundationally justified proposition is also something
that you rationally should regard as probably true. Failing to regard it in this way would be
failing to believe a conclusion that you could know to be true by reflection alone, given the
accessibility constraint on justification. Such a failure would be a failure of rationality (even if
not a gross or culpable form of irrationality). Thus, Klein can rebut the envisaged objection to
premise 3.

Recall that the objection to premise 5 was as follows: couldn’t one rationally believe that a
foundationally justified proposition is likely to be true, and yet still have no available reason for
believing that proposition? Her belief that the proposition is likely to be true might seem not to
provide her with such a reason if that belief is itself based on that very proposition; in that case,
by PAC, her belief that the foundationally justified proposition is likely to be true could not be a
reason for her to believe the foundationally justified proposition in question. But recall, from our
discussion of the objection to premise 3, that her belief that the foundationally justified
proposition is likely to be true is justified by dint of her reflection on an argument (a’) – (f’).
And none of (a’) – (f’) can themselves be identical to the foundationally justified proposition in
question. So the envisaged objection to premise 5 fails: you have a reason for believing that the
foundationally justified propositions is likely to be true that is not itself directly justified on the
basis of that foundationally justified proposition (even if its justificatory chain eventually
terminates in some foundationally justified proposition – a possibility that the present argument
is supposed to rule out eventually, not to assume away).

Finally, recall the objection to premise 9: couldn’t there be circumstances under which one
could rationally defend one’s beliefs by appeal to what one took to be a foundationally justified
proposition, even if one could not rationally defend them by appeal to what one correctly took to
be a foundationally justified proposition? The structural accessibility constraint on finite
justifications dictates a negative answer to this question. If that constraint holds, then, even if
you don’t know, or have any views about, what justifications you have to believe various
propositions, these facts about what justifications you have are at least in principle accessible to
you upon reflection. If there are no possible cases in which you rationally defend a belief of
yours by appeal to what you correctly take to be a foundationally justified proposition, then, by
structural accessibility, this cannot be simply because there are no possible cases in which you
correctly identify those propositions that are foundationally justified for you. It also cannot be
because the correctness of your identification of a foundationally justified proposition as such
somehow prevents you from rationally defending a belief of yours by appeal to that proposition:
the correctness of your identification could not make it any less rational for you to defend a
belief of yours by appeal to that proposition than it would be otherwise. So, if there are no
possible cases in which you rationally defend a belief by appeal to what you correctly take to be
a foundationally justified proposition, that can only be because there are no possible cases in
which you rationally defend of belief of yours by appeal to what you take – correctly or not – to
be a foundationally justified proposition. The envisaged challenge to premise 9 thus fails.

Given the unfamiliar way in which Klein frames the regress problem, it is charitable to interpret
him as accepting the accessibility constraint on justification. And if he accepts the accessibility
constraint, then he is also likely to accept the structural accessibility constraint on finite
justification: while the latter does not follow from the former, it is made plausible by it. But if
Klein accepts both the accessibility constraint on justification and the structural accessibility
constraint on finite justification, then he can rebut all of the envisaged objections to his argument
for PAA. It seems, therefore, that Klein’s argument for PAA is as solid as the accessibility
constraint and the structural accessibility constraint on finite justifications. While neither of
these last two constraints is obvious, neither is implausible. Klein’s case for PAA, while not
airtight, is nonetheless much more plausible than has generally been assumed, especially to
someone inclined to accept the accessibility constraint on justification.

Of course, much of the work that I’ve done has been done for the sake of charitable
interpretation, and perhaps I have not succeeded in correctly interpreting Klein’s argument. But
even if that is so, I have still set out an argument against PAA, and so for IPJ and IDJ, and this
argument is itself worth taking seriously, even if it is not an argument that Klein himself wanted
to espouse.

Conclusion: A Better Reply to Klein?

I have defended (what I take to be the best version of) Klein’s argument for PAA, and if PAA is
true, then FPJ is false. Since PAC is so widely accepted, Klein does not argue for it, and if PAC
is true, then CPJ is false. But if we grant Klein that FPJ and CPJ are both false, must we then
accept IPJ as the only possible alternative? Klein thinks so, and nobody, so far as I know, has
challenged him on this point. But I think that this is where his argument for infinitism might be
most fruitfully challenged. Recall our earlier statement of the options:

FPJ is the view that the evidential ancestry of any proposition that is justified for a person at a
time is finite and contains no element more than once. CPJ is the view that the evidential
ancestry of any proposition that is justified for a person at a time contains some elements more
than once. And IPJ is the view that the evidential ancestry of any proposition that is justified for
a person at a time is infinite and contains no element more than once.

But, though we followed Klein in using the term “evidential ancestry” in stating these views, we
also followed Klein in not explaining what the term denotes. Could it be that the term
“evidential ancestry”, as Klein uses it, is ambiguous, and that on one reading FPJ is true and on
the other reading CPJ is true, but that on neither reading are any of FPJ, CPJ, or IPJ all true?
That is the proposal that I would like to defend.

Recall Lewis Carroll’s story of the Tortoise and Achilles; one moral of the story is that we must
distinguish the premises of an inference from the rule governing the inference. For a conclusion
to be established inferentially, not only must the premises be true, but the inference must also be
valid. The truth of the premises is not sufficient to establish the conclusion, unless the inference
is valid. The justification of the conclusion thus depends not simply on the reasons given in
support of that conclusion (i.e., the premises cited); it depends also upon the validity of the
inference from those reasons.

Just as there are these two distinct components in the inferential justification of a conclusion, I
suggest that all propositional justification contains two such distinct components. There are the
reasons in light of which a claim is justified – reasons that must themselves be, as Klein says,
objectively and subjectively available if they are to justify. And then there is the relation –
whether logical, probabilistic, or what have you – between those reasons, on the one hand, and
the claim made on their basis, on the other: a relation that must be suitable for the claim to be
justified in light of those reasons. If the term “evidential ancestry” denotes only the first set of
factors, I suggest, FPJ is true: reasons eventually bottom out in justifiers that do not require, and
cannot receive, justification. Such justifiers might include, say, its intuitively seeming to me as
if addition is commutative (an intuitive seeming that I have by virtue of my possessing some
mastery of the skill of adding), or its visually seeming to me that there are clouds in the sky now
(a visual seeming that I have by virtue of my possessing a mature visual system). Even if the
possibility of my having justification by virtue of these seemings requires me to have
justification for believing that these seemings are themselves veridical, the latter justification is
not part of the “evidential anecestry” of the propositions justified by these seemings. My
justification for believing that my seemings are veridical is a necessary condition of those
seemings serving to justify various propositions for me, but that does not imply that it is itself
one of the reasons for believing those various propositions.

If, however, the term “evidential ancestry” denotes a broader category that includes the second
set of factors mentioned above, then CPJ is true: what serves as a justificatory relation between a
premise and a conclusion can also be made explicit as a proposition concerning the validity of
that relation, and then that proposition can be used as a reason to believe the premise. For
example, suppose I can prove that modus ponens is a valid rule of inference, and my proof of this
conclusion does not itself proceed by means of modus ponens. Still, the proof must proceed by
means of some rules or other. Then, if I prove the validity of those rules, that proof will itself
have to proceed by means of some rules or other. While there need not be any rules that must
necessarily be used in order to prove their own validity, any proof of the validity of any rule
must use some rules or other. And there does not seem to be any problem with using one rule in
a proof of the validity of another, and then using the latter in a proof of the validity of the former.
Coherence here is a virtue, not a vice. And so, when “evidential ancestry” is interpreted in the
broader way, there is no reason to accept PAC.

When Fred appeals to the foundational proposition f, and then Doris asks Fred whether
autonomously warranted propositions are likely to be true, Fred should answer “yes, they are, but
that is not my reason for believing f, for f is justified foundationally, not on the basis of any
reason at all.” And when Doris asks Fred why he thinks that autonomously warranted
propositions are likely to be true, Fred should not be worried about answering that question by
appeal to considerations to which he appealed earlier, and in the defense of which he earlier cited
f: such appeal would not constitute circular reasoning, since Fred is not reasoning both from f to
those considerations, and also back again. Fred is reasoning from f to those considerations, but
then he’s reasoning from the latter considerations to the propositions that autonomously
warranted propositions (like f) are likely to be true. He needs to be justified in believing that
last proposition in order to be justified in believing f, but that doesn’t imply that that last
proposition – or any other proposition – is what makes him justified in believing f.

To sum up: Klein’s argument for PAA is much more plausible than it is typically understood to
be. Nonetheless, his argument from PAA and PAC to IPJ is unsuccessful. There is a narrow
interpretation of “evidential ancestry” on which PAC is true and PAA is false: on this narrow
interpretation, evidential ancestry might well have a foundation. There is also a broader
interpretation of “evidential ancestry” on which PAA is true and PAC is false: on this broader
interpretation, evidential ancestry might well go in a circle. But on neither reading are both PAA
and PAC true, and so on neither reading does Klein’s argument for IPJ succeed.9

1 Klein 2005a, 132.

2 Turri, 2009, 210

3 Klein 1999, 298

4 Klein 1999, 298 – 9

5 That exchange consists in Klein 2005a, Ginet 2005a, Klein 2005b, and Ginet 2005b.

6 Klein 2005a, 133 – 4

7 See Williams 1996.

8 Note that this particular argument for PAA (and thus against the possibility of foundational
justification) is not addressed by the considerations that Turri 2009 uses to defend
foundationalism. Turri argues that the foundationalist can do just as well as the infinitist can in
explaining how it is possible to give a non-question-begging defense of any non-evident
proposition that one believes. But Turri does not engage with Klein’s argument, laid out above,
to the effect that it is impossible for any proposition to be foundationally justified. Of course if
we can prove that it is impossible for any proposition to be foundationally justified, then it
doesn’t matter whether foundationalism has the resources to solve this or that philosophical
problem, since it is anyway demonstrably false.
9 I am grateful to John Turri for his very helpful comments on an earlier draft of this paper.
A reply to Cling’s ‘‘The epistemic regress problem’’
William A. Roche
Springer Science+Business Media B.V. 2011

Abstract Andrew Cling presents a new version of the epistemic regress problem, and argues that
intuitionist foundationalism, social contextualism, holistic coherentism, and infinitism fail to
solve it. Cling’s discussion is quite instructive, and deserving of careful consideration. But, I
argue, Cling’s discussion is not in all respects decisive. I argue that Cling’s dilemma argument
against holistic coherentism fails.
Keywords Circular regresses of reasons, Cling, Coherence, Epistemic regress problem, Holistic
coherentism

1 Introduction
Andrew Cling (2008) presents a new version of the epistemic regress problem, and argues that
intuitionist foundationalism, social contextualism, holistic coherentism, and infinitism fail to
solve it.1 Cling’s discussion is quite instructive, and deserving of careful consideration. But, I
argue, Cling’s discussion is not in all respects decisive. I argue that Cling’s dilemma argument
against holistic coherentism fails. In Sect. 2, I explain Cling’s dilemma argument. In Sect. 3, I
argue that Cling’s dilemma argument is unsuccessful. Last, in Sect. 4, I conclude.

2 Cling’s dilemma argument explained

I follow Cling in the following preliminaries (pp. 402–403). Implication is necessary for support:
A proposition P1 supports—i.e., is a reason, or is evidence, for—a proposition P0 for a subject S
only if P1 implies P0, by standing in a (relevant) logical or quasi-logical relation, e.g. logical
entailment, to P0.2 But, implication is not sufficient for support. Consider the propositions I have
$10 million and I have at least $5 million. The former logically entails the latter, and so implies
the latter. But, alas, the former does not support the latter for me; the former does not provide me
with a reason, or evidence, to believe the latter. Things would be different if I had a reason to
believe I have $10 million. However, I have no such reason. A sequence of propositions σ =<P0,
P1, …, Pn (…)> is I-ordered (implication ordered) just in case (i) σ has propositions in at least
its first two places and (ii) every member of σ is implied by its successor, if it has one.3 A
sequence of propositions σ = <P0, P1, …, Pn (…)> is S-ordered (support ordered) for a subject
S (at a time t) just in case (i) σ is I-ordered and (ii) every member of σ is supported, for S (at t),
by its successor, if it has one.4 An endless regress of reasons is an S-ordered sequence of
propositions every member of which has a successor. An endless regress of reasons is circular
(or noninfinite) just in case it has infinitely many filled places but not infinitely many
components, and is infinite just in case it has infinitely many components, not just infinitely
many filled places.5 The sequence σS = <I have three sisters, The number of my sisters = √9>, I
have three sisters, …[has infinitely many filled places but not infinitely many components. The
sequence σT = <I am at least 7 feet tall, I am at least 8 feet tall, I am at least 9 feet tall, …> has
infinitely many filled places and infinitely many components.6

Cling argues that the epistemic regress problem assumes just three theses:

(1) Reasons are Supported: Only supported propositions provide support.

(2) No Proposition is Supported only by Endless Regresses: Propositions supported only by
endless S-ordered sequences are unsupported.
(3) Some Proposition is Supported: At least one proposition is supported by a proposition.7

(1)–(3), Cling argues, are jointly inconsistent:

(1)–(3) are jointly inconsistent. Reasons are Supported implies that any proposition P0 is
supported by a proposition P1 only if P0 and P1 are the first two members of an endless—
infinite or circular—regress of reasons. Given No Proposition is Supported only by Endless
Regresses it follows that no proposition is supported. This contradicts Some Proposition is
Supported. (p. 405).
Yet, if Cling is right, each of (1)–(3) is independently plausible (pp. 405–408). Thus the problem
—the epistemic regress problem.8

It might seem that (2) is obviously false. For, it might seem that (2) implies: If P0 is supported
by an endless S-ordered sequence of propositions, hence, is supported, and P0, though, is not
supported in some other manner, then P0 is not supported. But this is not how Cling means (2) to
be understood. Cling means (2) to be understood as saying: If support requires an endless regress
of reasons, then there is no support—no proposition is supported. Cling’s formal expression of
(2) is:

(∀x)[(∀y)(Sxy → ERSxy) → ~ (∃z)Sxy].

The variables ‘‘x,’’ ‘‘y,’’ and ‘‘z’’ range over propositions. ‘‘Sxy’’ means that x is supported by
y. ‘‘ERSxy’’ means that x and y are the first two members, in that order, of an endless (circular or
infinite) regress of reasons. So, formally speaking (2) says: For any x, if for any y, x’s being
supported by y requires that x and y be the first two members, in that order, of an endless regress
of reasons, then there is no z such that x is supported by z.9

Cling characterizes Laurence BonJour’s coherentism (1976, 1985) as a form of holistic

coherentism, on which ‘‘a believed proposition is justified for a person S just in case the
propositions S believes are coherent’’ (p. 417). Cling argues that holistic coherentism (so
understood) faces a fatal dilemma:

Are justified beliefs supported (Justification Requires Support)?… If justification requires

coherence and coherence requires that the believed propositions in a coherent set be supported by
supported propositions, then justification requires an endless regress. If, on the other hand, a set
of beliefs can be coherent without every member of the set being supported by a supported
proposition, then either support is not required for justification— Justification Requires Support
is false—or unsupported propositions can provide support—Reasons are Supported is false. So
either holistic coherentism does not solve the regress problem or it is a version of
foundationalism, contextualism, or some other theory incompatible with Reasons are Supported
(p. 417).

Let’s grant the second horn of this dilemma argument.10 I want to consider the question: Does
Cling establish the first horn?

Implicit in the first horn is No Proposition is Supported only by Endless Regresses. The charge,
more fully put, is this: If, on holistic coherentism, justification requires coherence and coherence,
in turn, requires that each of the propositions believed be supported by a supported proposition,
then justification requires an endless regress—that is, justification requires support and support,
in turn, requires an endless regress—and so holistic coherentism runs afoul of No Proposition is
Supported only by Endless Regresses, and, thus, holistic coherentism fails to solve the epistemic
regress problem. So, to establish the first horn, Cling needs to establish No Proposition is
Supported only by Endless Regresses.
Cling’s argument for No Proposition is Supported only by Endless Regresses has two main sub-
arguments. The first is meant to show that No Proposition is Supported only by Endless
Regresses holds for the case of circular endless regresses (p. 407). The second, in turn, is meant
to show that No Proposition is Supported only by Endless Regresses holds for the case of infinite
endless regresses (p. 407). We can set aside the second main sub-argument. If holistic
coherentism requires, for support, an endless regress, it requires a circular, not an infinite,
endless regress.11

The first main sub-argument of Cling’s argument for No Proposition is Supported only by
Endless Regresses can be seen as having three main parts. In the first, Cling argues, by example,
that a circular I-ordered sequence of propositions is not S-ordered per se and so some further
condition must be met (p. 407). Then, in the second main part of the argument (which is given in
the final two sentences of the passage below), Cling argues that the further condition in question
must include an independent reason to believe some member of the sequence:

Consider, then, a circular I-ordered sequence of propositions σC = <P0, P1,…, P0>. Each
member of σC is conditionally supported: Pn is supported if Pn+1 is, but σC is not S-ordered per
se. The members of σC must satisfy some additional condition. This condition must include an
independent reason to believe a member of σC. For without a reason P1’—a reason not already a
member of σC—to believe some member of σC, the sequence will be arbitrary from the
believer’s own point of view (p. 407, emphasis Cling’s).

Last, in the third main part of the argument, Cling moves to the overall conclusion that No
Proposition is Supported only by Endless Regresses holds for the case of circular endless
regresses:

If the sequence of propositions that conditionally supports P1’ is itself circular, the problem
arises again. So no proposition is supported only by circular sequences of propositions.
Therefore No Proposition is Supported only by Endless Regresses holds for the case of circular
sequences of reasons (p. 407).

I aim to show that the first main sub-argument of Cling’s argument for No Proposition is
Supported only by Endless Regresses fails. If I succeed, it follows that Cling fails to establish the
first horn of his dilemma argument against holistic coherentism.

3 Cling’s dilemma argument critiqued

Let’s consider a form of holistic coherentism on which Some Proposition is Supported is correct,
and the following thesis holds:

(4) P is supported for S if and only if (i) there is a circular I-ordered sequence of propositions σC
such that (a) P is a member of σC and (b) S believes each of the propositions in σC and no other
propositions, and (ii) S’s belief system is coherent.

Let’s call this theory, viz., the conjunction of Some Proposition is Supported and (4), ‘‘(HC).’’ If
(HC) is correct, the epistemic regress problem is to be answered by affirming Reasons are
Supported and Some Proposition is Supported, and denying No Proposition is Supported only by
Endless Regresses.12

Several comments are in order. First (HC), unlike the holistic coherentist theory at issue in the
first horn of Cling’s dilemma argument, is a theory of support, not a theory of justification. I
focus on (HC) because the epistemic regress problem, as construed by Cling, concerns (first and
foremost) support. Cling’s charge against (HC) would be: (HC) denies, and, so, runs afoul of, No
Proposition is Supported only by Endless Regresses, and, thus (HC) fails to solve the epistemic
regress problem. Second (4) should be understood so that if, say, S believes P0, P1, and P2, and
no other propositions, and if, in fact, P0 is implied by {P1, P2},13 P1 is implied by {P0, P2}, and
P2 is implied by {P0, P1}, then, with respect to each of P0, P1, and P2 (i) in (4) is satisfied. It
might be that Cling’s machinery for constructing I-ordered sequences of propositions will need
to be modified.14 But Cling should allow for this. A situation of the sort just described, where S
believes P0, P1, and P2, and no other propositions, and where P0 is implied by {P1, P2}, P1 is
implied by {P0, P2}, and P2 is implied by {P0, P1}, is (an oversimplified example of) just the
sort of situation holistic coherentists have in mind in speaking of ‘‘mutual’’ or ‘‘reciprocal’’
support. If Cling denied that in such a situation (i) in (4) is satisfied, Cling would be in no
position to charge, as he does in the first horn of his dilemma argument, that holistic coherentism
runs afoul of No Proposition is Supported only by Endless Regresses.15 Third (4) is too stringent,
in requiring that P be believed by S. This deficiency can be remedied by modifying (a) in (4) to
say ‘‘P is a member of σC or P is implied by a member of σC.’’16 I focus on (4) unmodified
solely for simplicity. Fourth, it is no simple matter to specify with precision just what it takes for
a belief system to be coherent.17 I leave it open that not all circular I-ordered sequences of
propositions are coherent (to the requisite degree or in the requisite sense), and, thus, that there
can be cases in which (i) in (4) is satisfied but (ii) is not. Fifth (4), as currently stated, is perfectly
‘‘egalitarian’’: (4) entails that some of the propositions S believes are supported only if all such
propositions are supported. One way to make (4) less egalitarian would be to modify (b) in (4) so
that S is permitted to believe propositions in addition to the propositions in σC, and so that what
matters for support is the coherence not of S’s belief system as a whole, but just of a certain
subset (or ‘‘module’’) of S’s belief system.18 Nothing in what I argue below, in reply to Cling,
requires that (4) not be modified in this fashion. I focus on (4) unmodified solely for simplicity.

(HC) can be glossed as saying that support is a matter of coherence. This claim (that support is a
matter of coherence), though, can be easy to misunderstand. The claim is not that P, when
supported, is supported by the coherence of S’s belief system, or by the fact that S’s belief
system is coherent. Cling is right that coherence itself cannot provide support: ‘‘Coherence
cannot… support any proposition in the relevant sense, for coherence per se is not the sort of
thing that could be a reason’’ (p. 417). Nor is the claim that P, when supported, is supported by
the proposition (believed by S) that S’s belief system is coherent. (HC) requires not that S believe
that his belief system is coherent, but just that S’s belief system be coherent. (HC) thus avoids
Cling’s point that ‘‘requiring belief in the coherence of one’s beliefs for justification or support
would be both too strong—one must have the relevant concept of coherence—and too weak—to
be a reason, this proposition needs support, and the regress returns’’ (p. 418). The claim, on
(HC), is that P, when supported, is supported by virtue of, or because of, the fact that S’s belief
system is coherent. S’s reason for P is some other proposition he believes, say, Q. But what
makes it the case that Q is a reason for P, that is, what makes it the case that Q supports P, and
does not merely imply P, is the fact that S’s belief system is coherent.19

I want to stress the point that (HC) requires not that S believe that his belief system is coherent,
but just that S’s belief system be coherent. It might be that, if (HC) is correct, S needs to believe
that his belief system is coherent in order for certain epistemic propositions, e.g., P is supported
for me, to be supported for him. But that would be another matter. Where P is a nonepistemic
proposition, then, if (HC) is correct, the requirement is just that S’s belief system be coherent
(and there be a circular I-ordered sequence of propositions σC such that P is a member of σC,
and S believes each of the propositions in σC and no other propositions).20

I have clarified (HC) and noted some respects in which (HC) can be refined. I will now argue
that the first main sub-argument of Cling’s argument for No Proposition is Supported only by
Endless Regresses fails.21

The central issue concerns the notion of an independent reason. Suppose there is a circular I-
ordered sequence of propositions σC such that P0 is a member of σC, and S believes each of the
propositions in σC and no other propositions. Suppose S’s belief system is coherent. Then, by the
second main part of Cling’s argument for No Proposition is Supported only by Endless
Regresses, it follows that P0 is supported for S only if S has an independent reason—a reason not
in σC —to believe some member of σC. If S has no such reason, then, despite the fact that P0 is
a member of σC, and S believes each of the propositions in σC and no other propositions, and
despite the fact that S’s belief system is coherent, P0 is not supported for S: P0 is arbitrary from
S’s point of view.22 By contrast, if (HC) is correct, then, since (i) and (ii) in (4) are satisfied, it
follows that P0 is supported for S, by P1 (which, too, is supported for S, by P2, and so on), and,
hence, P0 is not arbitrary from S’s point of view—this despite the fact that S has no independent
reason to believe some member of σC.23 The question is: Does Cling show that an independent
reason is needed?

Note: (HC) agrees with Cling that a circular I-ordered sequence of propositions σC is not S-
ordered per se and so some further condition must be met. If (HC) is correct, the further
condition in question is: S believes each of the propositions in σC and no other propositions, and
S’s belief system is coherent.

Cling, of course, would object that this condition is not enough, since it does not require that S
have an independent reason to believe some member of σC. But, again, the question is whether
Cling shows that an independent reason is needed.

Cling argues, by example, that a circular I-ordered sequence of propositions σC is not S-ordered
per se and so some further condition must be met. Cling holds that the further condition in
question must include an independent reason to believe some member of σC. Why follow Cling
in this? Cling states that if S has no such reason, then the members of σC are arbitrary from S’s
point of view. But why believe this? Cling does not say. He simply moves on to (what I am
calling) the third main part of the argument, saying that ‘‘[i]f the sequence of propositions that
conditionally supports P1’ [where P1’ is an independent reason to believe some member of σC]
is itself circular, the problem arises again’’ (p. 407).24 I suspect that Cling takes the point to be
obvious—that if S has no independent reason to believe some member of σC, then the members
of σC are arbitrary from S’s point of view. If, though, holistic coherentism, in at least some of its
varieties, e.g. (HC),25 is correct, it follows that: It is false that if S has no independent reason to
believe some member of σC, then the members of σC are arbitrary from S’s point of view. I find
it far from obvious that holistic coherentism is false.

Consider the case of Norman, which I adapt from BonJour (1985, p. 41):

Norman believes the President is in New York City. This belief was produced by Norman’s
process of clairvoyance, under circumstances in which this process is highly reliable. Norman,
though, has no belief as to whether he has a highly reliable process of clairvoyance. In fact, there
is no proposition P (distinct from the President is in New York City) such that P implies the
President is in New York City, and Norman believes P.

(HC) entails that the President is in New York City is arbitrary from Norman’s point of view—S
has no reason to believe the President is in New York City. This seems to be the right result; the
President is in New York City is not implied by any other proposition Norman believes. Compare
the case of Norman with the case of S:

There is a circular I-ordered sequence of propositions σC such that P0 is a member of σC, and S
believes each of the propositions in σC and no other propositions. S’s belief system is numerous
in beliefs, rich in content, and coherent.

(HC) entails that, though S has no independent reason to believe some member of σC, P0 is
supported for S, and so is not arbitrary from S’s point of view. This result strikes me as being at
least somewhat plausible. P0 is implied by some other proposition S believes,26 which in turn is
implied by some other proposition S believes, and so on without exception. Further, S’s belief
system is numerous in beliefs, rich in content, and coherent.

(HC) does not (explicitly) require that a belief system be numerous in beliefs and rich in content.
But (HC) allows for this, and could be modified so as to require it. Also, Cling’s position on the
need for an independent reason is meant to hold for any circular I-ordered sequence of
propositions regardless of size and richness of content.

Bear in mind that (HC) is just one form of holistic coherentism. (HC) places no content, or
subject-matter, requirements on a belief system. (HC), thus, does not require that S have beliefs
about the ways in which she is reliably connected to the world. Other forms of holistic
coherentism, by contrast, do place content requirements on a belief system, and do require that S
have beliefs about the ways in which she is reliably connected to the world.27 One such view,
‘‘(HC*),’’ is the conjunction of Some Proposition is Supported and the following thesis:

(5) P is supported for S if and only if (i) there is a circular I-ordered sequence of propositions σC
such that (a) P is a member of σC, (b) σC includes propositions about the ways in which S is
reliably connected to the world, and (c) S believes each of the propositions in σC and no other
propositions, and (ii) S’s belief system is coherent.
Consider the case of S set out four paragraphs above. Let P0 be the proposition a cat is before
me. Suppose σC includes propositions about the ways in which S is reliably connected to the
world, including, in particular, the propositions it appears to me visually as if a cat is before me
and usually when it appears to me visually as if such and such is the case, this is because
(causally) such and such is the case. (HC*) entails that a cat is before me is supported for S,
hence is not arbitrary from S’s point of view—S has, at the least, some reason to believe a cat is
before me. This implication strikes me as being at least somewhat plausible.

If we are to side with Cling, and against (HC) and (HC*), on the need for an independent reason,
we should have an argument for doing so.28 Since Cling fails to provide such an argument, Cling
fails to establish No Proposition is Supported only by Endless Regresses.29

Perhaps, though, an argument (for the need for an independent reason) is close at hand. Consider
the argument:

Argument from No Beliefs

Imagine S is in a situation of no beliefs, that is, a situation in which S has no beliefs. Suppose S is
considering whether to believe the members of σC. Suppose S has no independent reason to
believe some member of σC. Then, since a circular I-ordered sequence of propositions is not S-
ordered per se, it follows that the members of σC are arbitrary from S’s point of view. Hence, if,
in a situation of no beliefs, S has no independent reason to believe some member of σC, then the
members of σC are arbitrary from S’s point of view.

Let’s grant that Argument from No Beliefs is sound. Is this enough for Cling’s purposes?

No. The conclusion of Argument from No Beliefs is the claim that:

(6) If, in a situation of no beliefs, S has no independent reason to believe some member of σC,
then the members of σC are arbitrary from S’s point of view.

(6) is about only situations of no beliefs; (6) says nothing about situations of some beliefs (that is,
situations in which S has some beliefs). Cling’s argument for No Proposition is Supported only
by Endless Regresses, though, needs a claim about all situations, a claim, thus, much stronger
than (6). Specifically, Cling’s argument needs the claim:

(7) If, in any situation, S has no independent reason to believe some member of σC, then the
members of σC are arbitrary from S’s point of view.30

(6) leaves it open that (7) is false, by leaving it open that there are situations of some beliefs in
which S has no independent reason to believe some member of σC, and yet the members of σC
are not arbitrary from S’s point of view. Consider (HC). (6) leaves it open that (HC) is correct. In
fact, if (HC) is correct (6) is correct.31 If, though (HC) is correct, there are situations of some
beliefs in which S has no independent reason to believe some member of σC, and yet the
members of σC are not arbitrary from S’s point of view—viz., situations in which S believes the
members of σC, and no other propositions, and S’s belief system is coherent.
(HC) entails that a situation of no beliefs is a situation of no support (a situation in which no
propositions are supported for S). It might seem that this implication is problematic. How, it
might be asked, is a subject to non-arbitrarily build a belief system from scratch, that is, from a
situation of no beliefs, if a situation of no beliefs is a situation of no support? Holistic
coherentists, though, do not mean to be giving a decision-procedure for non-arbitrarily building
a belief system from scratch. And for good reason: It is doubtful that a subject could use a
decision-procedure to nonarbitrarily build a belief system from scratch.32 Holistic coherentists
mean to be giving a theory of support,33 a theory specifying the conditions under which a
proposition is supported for a subject (at a time).

Argument from No Beliefs can be modified so that it covers some situations of some beliefs:

Argument from No Beliefs*

Imagine S is in a situation of no beliefs, or at least of no beliefs in the members of σC. Suppose S

is considering whether to believe the members of σC. Suppose S has no independent reason to
believe some member of σC. Then, since a circular I-ordered sequence of propositions is not S-
ordered per se, it follows that the members of σC are arbitrary from S’s point of view. Hence, if,
in a situation of no beliefs, or at least of no beliefs in the members of σC, S has no independent
reason to believe some member of σC, then the members of σC are arbitrary from S’s point of
view.

The conclusion of Argument from No Beliefs* goes beyond (6), but still falls short of (7).34
Argument from No Beliefs*, thus, even if sound, is insufficient for Cling’s purposes.

What needs to be shown, for Cling’s argument for No Proposition is Supported only by Endless
Regresses to succeed, is that even in a situation in which S believes the members of σC and no
other propositions, and S’s belief system is coherent, S needs to have an independent reason to
believe some member of σC. Put another way, what needs to be shown is that, contra (HC) (i)
and (ii) in (4) are not (together) sufficient for support.35 Cling fails to show this, and so too do
Argument from No Beliefs and Argument from No Beliefs*.

(HC), it must be admitted, is a rather radical theory of support. If (HC) is correct, the only mental
states that matter, epistemically, for support are the subject’s beliefs; the subject’s perceptual
experiences matter not at all. (HC) is thus a form of ‘‘doxastic’’ holistic coherentism.36 I want to
close by considering a less radical, and, arguably, more plausible, form of holistic coherentism—
a form of ‘‘nondoxastic’’ holistic coherentism.

Let ‘‘(HC**)’’ be the conjunction of Some Proposition is Supported and the following thesis:

(8) P is supported for S if and only if (i) there is a circular I-ordered sequence of propositions σC
such that (a) P is a member of σC and (b) S believes each of the propositions in σC and no other
propositions, (ii) there is a circular I-ordered sequence of propositions σC * such that (a) each
member of σC is a member of σC *, (b) each proposition serving as the propositional content of
one of S’s perceptual experiences (that is, one of S’s perceptual experiences at the time in
question) is a member of σC *, and (c) each member of σC * is a member of σC or (inclusive
‘‘or’’) serves as the propositional content of one of S’s perceptual experiences, and (iii) S’s
system of beliefs and perceptual experiences is coherent.37

How is (ii) to be satisfied? I can think of several ways. Perhaps the simplest is where S believes
the members of a circular I-ordered sequence of propositions σC and no other propositions, and
where for each perceptual experience e in S’s system (of beliefs and perceptual experiences),
there is a perceptual belief b in S’s system such that e and b have the same propositional content
(e.g., S has a visual experience with the propositional content a blue object is, now, there, and S
has a visual belief, i.e., a belief produced by vision, with that same propositional content).38
(HC**) agrees with (HC) that Cling is wrong on the need for an independent reason,39 and agrees
with (HC) that No Proposition is Supported only by Endless Regresses is false. But (HC**)
disagrees with (HC) that the only mental states that matter for support are the subject’s beliefs.
Also important, according to (HC**), are the subject’s perceptual experiences; (ii) in (8) requires
that each proposition serving as the propositional content of one of S’s perceptual experiences be
a member of σC*, and (iii) in (8) requires that S’s system of beliefs and perceptual experiences
be coherent. This, arguably, is a respect in which (HC**) is to be preferred to (HC).40

Much more would need to be said to show that (HC**) solves the epistemic regress problem.
Importantly, it would need to be argued that perceptual experiences have propositional content.41
Also (HC**) would need to be refined in various respects. For one thing (HC**) would need to
be refined so that it can account for cases in which a subject has a reason to believe that certain
of his perceptual experiences are illusory. Perhaps, too (HC**) would need to be refined so that
it can accommodate the idea that certain of a subject’s nonperceptual experiences, e.g.,
memorial experiences, are relevant, epistemically, to what propositions are supported for her. My
point is fairly modest: Cling would need to say much more to establish, against (HC**), No
Proposition is Supported only by Endless Regresses, and thus would need to say much more to
show that, because (HC**) runs counter to No Proposition is Supported only by Endless
Regresses (HC**) fails to solve the epistemic regress problem.

4 Conclusion

I conclude that, though Cling’s discussion overall is very much instructive and deserving of
careful study, Cling’s dilemma argument against holistic coherentism is unsuccessful. The first
horn of Cling’s dilemma argument relies on No Proposition is Supported only by Endless
Regresses, i.e., the claim that if support requires an endless regress of reasons, then no
proposition is supported. But Cling fails to establish No Proposition is Supported only by
Endless Regresses.

Acknowledgments I wish to thank Andrew Cling, Michael Roche, and Joshua Smith for very
helpful questions and comments on prior versions of this paper.

1 All references to Cling are to Cling (2008).

2 When is it that P1 implies P0, by standing in a (relevant) logical or quasi-logical relation to

P0? Cling mentions three possible answers: when P1 entails or inductively implies P0; when the
probability of P0 given P1 is sufficiently high; when P1 stands in an irreducibly epistemic
relationship to P0 such that P1 would justify P0 (p. 402). (When Cling gives the second possible
answer, he writes: ‘‘the probability of P1 given P0 is sufficiently high’’ (p. 402). I take it that,
since the question is of when it is that P1 implies P0, not of when it is that P0 implies P1, what
Cling means to say is: the probability of P0 given P1 is sufficiently high.)

3 For any Pi in σ such that Pi has a successor, the successor of Pi is Pi+1.

4 This construal of the notion of an S-ordered sequence differs from Cling’s initial construal (p.
402) in making no use of the notion of relevant accessibility. I take it, though, that, as Cling
understands the notion of relevant accessibility, if every member of σ is supported, for S (at t), by
its successor, if it has one, then the members of σ are relevantly accessible to S (at t).

5 When Cling initially defines the notion of an infinite endless regress of reasons (p. 403), he
does not refer to the other kind of endless regress of reasons as ‘‘circular.’’ But Cling adopts this
terminology elsewhere in the paper.

6 The claim is not that σS is a circular endless regress of reasons, or that σT is an infinite endless
regress of reasons. Perhaps neither σS nor σT is S-ordered. The claim is simply that σS has
infinitely many filled places but not infinitely many components, and σT has infinitely many
filled places and infinitely many components.

7 The names and wording of these theses are Cling’s (pp. 404–405). I clarify (2) below.

8 Cling does not claim that there is no solution to the epistemic regress problem. Cling’s claim is
more modest (though still quite strong): Each of intuitionist foundationalism, social
contextualism, holistic coherentism, and infinitism fails to solve the epistemic regress problem.
Cf. Cling (2009).

9 Cling’s formal expressions of (1) and (3) are, respectively, ‘‘(∀x)(∀y)(Sxy → (∃z)Syz)’’ and
‘‘(∃x)(∃y)Sxy.’’ Cf. Cling (2009).

10 I find it plausible that holistic coherentists as such are committed to Justification Requires
Support and Reasons are Supported. Regardless, my interest is in answers to the epistemic
regress problem on which Justification Requires Support and Reasons are Supported are correct.

11 The situation with respect to holistic coherentism and the first main sub-argument is very
much analogous to the situation with respect to holistic infinitism and the second main sub
argument. Perhaps, then, what I argue (below) on behalf holistic coherentism (against the first
main sub-argument) can be argued, mutatis mutandis, on behalf of holistic infinitism (against the
second main sub-argument). Peter Klein’s ‘‘warrant-emergent’’ infinitism (2005) is a form of
holistic infinitism about warrant. Klein’s theory could be transformed into a holistic infinitist
theory of justification, or a holistic infinitist theory of support.

12 (4) is a component of (HC), and (4) entails Reasons are Supported. (If (4) is correct, then
when P is supported, P is supported by its successor in σC, and P’s successor in σC, in turn, is
supported by its successor in σC, and so on. Thus (4) entails Reasons are Supported.) Some
Proposition is Supported is a component of (HC). (Note: (4) does not entail Some Proposition is
Supported. (4) can be true even if (i) and (ii) are never satisfied.) (HC), thus, is correct only if
Reasons are Supported and Some Proposition is Supported are correct. Since, as Cling holds,
Reasons are Supported and Some Proposition is Supported together entail that No Proposition is
Supported only by Endless Regresses is false, it follows that if (HC) is correct, then No
Proposition is Supported only by Endless Regresses is false.

13 Cling construes support as a two-place relation (p. 403), and notes that: ‘‘In case the evidence
that is available for a proposition consists of more than one proposition, we may represent it as
the conjunction of the relevant propositions or as the set of those propositions’’ (p. 403, n. 3,
emphasis mine). If a proposition can be supported by a set of propositions, then, likewise, a
proposition can be implied by a set of propositions.

14 One way to do this would be to construe I-ordered sequences of propositions as sequences of

ordered pairs, where the first member of a pair is a proposition, the second member of a pair is a
proposition or a set of propositions, and the first member of a pair is implied by the second
member of the pair. A circular I-ordered sequence would be a sequence with just finitely many
pairs, and where each proposition involved in the sequence is the first member of some pair.
Then, the sequence <(P0, {P1, P2}), (P1, {P0, P2}), (P2, {P0, P1})> would be a circular I-
ordered sequence.

15 It would be true that holistic coherentism requires that supported propositions be supported by
supported propositions, as in a case where (according to holistic coherentism) P0 is supported
by, hence is implied by, {P1, P2}, P1 is supported by, hence is implied by, {P0, P2}, and P2 is
supported by, hence is implied by, {P0, P1}. But it would not follow that holistic coherentism
requires circular I-ordered sequences of propositions, and so would not follow that holistic
coherentism requires circular S-ordered sequences of propositions, that is, circular endless
regresses.

16 (b), though, would remain the same.

17 For general discussion of the elements of coherence, see BonJour (1985, Chap. 5, Sect. 5.3).
For discussion of probabilistic conceptions of coherence, see Olsson (2005, Chap. 6, Sects. 6.1
and 6.2).

18 See Lycan (1996) and Olsson (1997).

19 Strictly speaking, on (HC), part of what makes it the case that Q supports P is the fact that
there is a circular I-ordered sequence of propositions σC such that P and Q are members of σC,
Q is P’s successor in σC, and S believes each of the propositions in σC and no other propositions.

20 I am assuming, as seems plausible, that S’s belief system can be coherent even if S does not
believe that his belief system is coherent.
21 Hereafter I will speak simply of Cling’s argument for No Proposition is Supported only by
Endless Regresses, and not refer explicitly to ‘‘the first main sub-argument’’ thereof.

22 Two comments are in order. First, as I understand Cling, a proposition P is supported for a
subject S (that is, S has a reason to believe P) just in case P is not arbitrary from S’s point of
view. Cling clarifies the notion of a proposition’s being arbitrary from a subject’s point of view
on p. 406. Second, Cling sometimes speaks of a sequence of propositions as being arbitrary from
a subject’s point of view. I take it that when Cling speaks in this fashion, he means just that the
propositions in the sequence are arbitrary from the subject’s point of view.

23 If (HC) is correct, S has no reasons in addition to the members of σC. Hence, if (HC) is
correct, S has no independent reason to believe some member of σC.

24 It might be wondered whether I have misread Cling. Consider the following passage (from the
first full paragraph on p. 407): ‘‘To be S-ordered, σC must satisfy some additional condition. If
this condition is not arbitrary from the believer’s own point of view it must include having an
independent reason P1’ to believe some member of σC, that is, it must include having a reason
that does not itself depend on σC’’ (emphasis mine). It might seem that this passage involves a
claim (an implicit claim) not included in my presentation of Cling’s argument for No
Proposition is Supported only by Endless Regresses—viz., the claim that the further condition in
question must itself not be arbitrary from the subject’s point of view. I have no idea, though,
what it would mean to say that a condition, as opposed to a proposition, is not arbitrary from a
subject’s point of view. And Cling himself never specifies how to understand the notion of a
condition’s not being arbitrary from a subject’s point of view. So I read Cling as saying: If the
further condition in question is to make it such that the members of σC are not arbitrary from the
subject’s point of view, then that condition must include an independent reason to believe some
member of σC. This is just the claim: The further condition in question must (to make it such that
the members of σC are not arbitrary from the subject’s point of view) include an independent
reason to believe some member of σC.

25 Below I discuss two additional varieties of holistic coherentism.

26 Or set of propositions S believes.

27 It is not uncommon for coherentists to invoke content requirements. See, e.g., Blanshard
(1939, Chap. 26, Sect. 19), BonJour (1985, Chaps. 6–7), Brink (1989, Chap. 5, Sects. 5–7), and
Lehrer (2000, Chaps. 6–7).

28 Of course, there are well known objections to views such as (HC) and (HC*), e.g., the
‘‘Alternative Systems Objection.’’ For discussion and references, see Kvanvig (2007). Cling
makes no explicit appeal to any such objection. So I presume the basis for Cling’s position on the
need for an independent reason lies elsewhere. Below I discuss a view, ‘‘(HC**),’’ which,
arguably, fares quite well against the various standard objections to views such as (HC) and
(HC*).
29 It might seem that Cling should be read as arguing: ‘‘A circular I-ordered sequence of
propositions is not S-ordered per se. If a circular I-ordered sequence of propositions is not S-
ordered per se, then if S has no independent reason to believe some member of σC, then the
members of σC are arbitrary from S’s point of view. Therefore, if S has no independent reason to
believe some member of σC, then the members of σC are arbitrary from S’s point of view.’’ On
this reading, Cling gives an argument for his position on the need for an independent reason. But
the argument is inadequate. Since (HC) and (HC*) entail the falsity of the second premise, and
since (HC) and (HC*) are at least somewhat plausible, it follows that Cling needs to provide an
argument for the second premise. Cling, though, does not provide an argument for the second
premise.

30 Cling’s argument for No Proposition is Supported only by Endless Regresses needs (7), a
claim about all situations, because No Proposition is Supported only by Endless Regresses is
itself a claim about all situations. Note: If No Proposition is Supported only by Endless
Regresses covered only some situations, then even if No Proposition is Supported only by
Endless Regresses were correct, and even if Reasons are Supported, understood as a claim about
all situations, were correct, it might still be that Some Proposition is Supported is correct. All
that would follow (from the correctness of No Proposition is Supported only by Endless
Regresses and Reasons are Supported) is that in the limited situations covered by No Proposition
is Supported only by Endless Regresses, no proposition is supported. This would leave it open
that in some of the situations not covered by No Proposition is Supported only by Endless
Regresses, some proposition is supported.

31 If (HC) is correct, it follows that if S is in a situation of no beliefs, and S has no independent

reason to believe some member of σC, then, by (4), no propositions are supported for S, hence all
propositions are arbitrary from S’s point of view, hence the members of σC are arbitrary from S’s
point of view. Note: If (HC) is correct, it follows that if S is in a situation of no beliefs, then S
has no independent reason to believe some member of σC—S has no reasons whatsoever.

32 For relevant discussion, see Goldman (1980) and Pryor (2005).

33 Or, more commonly, a theory of justification.

34 The conclusion of Argument from No Beliefs* leaves it open that (HC) is correct. In fact, if
(HC) is correct, the conclusion of Argument from No Beliefs* is correct. If, though (HC) is
correct (7) is false. So, the conclusion of Argument from No Beliefs* leaves it open that (7) is
false.

35 But even if it were shown that (i) and (ii) in (4) are not sufficient for support, and that (i) and
(ii) in (5) are not sufficient for support, it would not follow that Cling is right about the need for
an independent reason. It would still need to be shown (inter alia) that (i), (ii), and (iii) in (8),
below, are not sufficient for support.

36 Doxastic holistic coherentists deny that perceptual experiences can serve as reasons for
beliefs, but do not deny that perceptual experiences can cause beliefs. For defense of the thesis
that only beliefs can serve as reasons for beliefs, see, e.g., BonJour (1985, Chap. 4) and
Davidson (2000). For helpful discussion and references, see Pryor (2005).

37 I noted above (third paragraph of this section) that (HC) can be refined in certain respects.
(HC**) can be refined in those same respects. Also (HC**) can be modified so that, like (HC*),
it is required that S have beliefs about the ways in which she is reliably connected to the world.

38 How, in σC*, are the separate contributions of S’s beliefs and perceptual experiences to be
represented? One way to do this would be to designate the members of σC* as ‘‘doxastic’’ or
‘‘perceptual-experiential’’ as appropriate. If p serves as the propositional content of a belief, this
could be represented as ‘‘pd.’’ If p serves as the propositional content of a perceptual experience,
this could be represented as ‘‘pp-e.’’ Then, if p served as the propositional content of both a
belief and a perceptual experience, both pd and pp-e would have a place in σC*.

39 If (HC**) is correct, then (i), (ii), and (iii) in (8) are sufficient for support. It is thus not
required that S have an independent reason (a reason not in σC*) to believe some member of
σC*.

40 One of the main objections to views such as (HC) and (HC*) is the ‘‘Isolation Objection.’’
This objection (at least in one of its versions) charges that holistic coherentism implies that a
subject’s perceptual experiences are irrelevant, epistemically, to what propositions are supported
for her, and that, because of this, holistic coherentism is open to counterexample. (Strictly
speaking, the objection is typically put in terms of justification.) See, e.g., Feldman (2003, pp.
68–70). (HC**) is not open to this objection. Nor are certain other forms of nondoxastic holistic
coherentism. For further discussion of nondoxastic holistic coherentism, see Cohen (2002),
Horgan and Potrc (2010), Kvanvig (1995), and Kvanvig and Riggs (1992).

41 For defense of the view that perceptual experiences have propositional content, see, e.g.,
Searle (1983, Chap. 2). For an introduction to the relevant literature, see Siegel (2010).
The epistemic regress problem

Andrew D. Cling
Philosophy, University of Alabama in Huntsville, 301 Sparkman Drive, Huntsville, AL 35899,
Philos Stud (2008) 140:401–421

Abstract The best extant statement of the epistemic regress problem makes assumptions that are
too strong. An improved version assumes only that that reasons require support, that no
proposition is supported only by endless regresses of reasons, and that some proposition is
supported. These assumptions are individually plausible but jointly inconsistent. Attempts to
explain support by means of unconceptualized sensations, contextually immunized propositions,
endless regresses, and holistic coherence all require either additional reasons or an external
condition on support that is arbitrary from the believer’s own point of view.
Keywords Regress, Skepticism, Reasons, Foundationalism, Coherence, Contextualism,
Infinitism, Paradox

[I]t is not possible to untie a knot which one does not know.
—Aristotle, Metaphysics b 1995a29–30 (1984, 1572).

Lather, rinse, repeat.

—Shampoo Instructions.

1 Introduction

It seems that we have knowledge. To know a proposition P0 we must have a reason P1 that
supports P0 by providing evidence for it. A proposition, however, is a reason only if there is a
proposition that supports it. This requires that we have a reason P2 that supports P1, and so on.
The resulting sequence of reasons is endless: infinite or circular. We cannot, however, acquire
support by means of endless regresses. Thus we have no knowledge.

That, roughly, is the epistemic regress problem. Some of our core epistemic assumptions are
jointly inconsistent, a paradox. To solve this problem we must understand it. Unfortunately, the
best extant statement of the problem makes unduly strong assumptions and thus does not capture
its deepest challenges. Contrary to the received version, the regress paradox is not just a problem
about knowledge and justification, it concerns evidential support, a more basic epistemic value.

This paper presents an improved version of the epistemic regress problem. I do not claim that
there is no solution to the problem—I hope there is—but I do claim that some popular
epistemologies do not solve it. In particular, I argue that unconceptualized sensations,
contextually immunized evidence, endless regresses, and holistic coherence are insufficient for
support. To provide support, such features must either include propositions that require support
or be external factors that make the propositions that have them arbitrary from the believer’s own
point of view.

2 Two versions

This section presents the best extant version of the problem and my alternative. I begin by
clarifying the crucial concepts of reasons and regresses.

In paradigm cases, a reason for a proposition P0 is a believed proposition P1 that provides the
believer with evidence for P0. A proposition P1 provides supporting evidence for a proposition
P0 only if P1 implies P0 by standing in a relevant logical or quasi-logical relation to P0: P1
entails or inductively implies P0; the probability of P1 given P0 is sufficiently high; P1 stands in
an irreducibly epistemic relationship to P0 such that P1 would justify P0; et cetera.1 The nature
of such relations and which is required for support are contested but that does not affect the
structure of the regress problem.

Implication is insufficient for support since propositions can stand in the relevant implication
relation despite the fact that a person who recognizes that relation has no reason to believe either
one. I see that I have $10 million implies that I have at least $5 million but (alas) I have no
reason to believe either proposition. Implication is necessary but not sufficient for support.2

Call a finite or infinite sequence of propositions σ = <P0, P1, . . . Pn (. . .)> implication-ordered

(I-ordered) just in case σ has propositions in at least its first two places and every member of σ is
implied by its successor, if it has one. Similarly, call a sequence of propositions σ support-
ordered (S-ordered) just in case σ is I-ordered, the elements of σ are propositions relevantly
accessible to a person at a time, and every member of σ is supported by its successor, if it has
one. So a regress of reasons is an S-ordered sequence of propositions.

An infinite regress of reasons is an S-ordered sequence every component of which has a

successor. There are two kinds of such sequences, however, and I shall refine my terminology to
capture this distinction. First, there are sequences with infinitely many filled places. Sequence σS
= <I have three sisters, The number of my sisters = √9, I have three sisters, . . .> and the
sequence σT = <I am at least 7 feet tall, I am at least 8 feet tall, I am at least 9 feet tall, . . .>
have this feature. Call such sequences ‘endless.’ An endless regress of reasons, then, is an S-
ordered sequence of propositions every member of which has a successor. Second, there are
endless regresses with infinitely many components. Endless sequence σT has infinitely many
components, unlike endless sequence σS. I reserve ‘infinite’ for sequences with infinitely many
components. An infinite regress of reasons, then, is an S-ordered sequence of propositions σ
with infinitely many components. Every infinite regress of reasons is endless, but not conversely.

I now state the two versions of the regress problem. The constituent propositions are given
shorthand names and stated in (philosopher’s) English and in a first-order language. Let the
variables ‘x’, ‘y’, and ‘z’ range over the domain P of propositions relevantly accessible to a
person at a time. Assign the one-place predicate ‘J’ to the members of P that are epistemically
justified: the propositions that it is permissible, virtuous, or otherwise good for the person to
accept at that time. Let the two-place predicate ‘S’ express the relation between members of P
that obtains when the first is supported by the second.3

Let the two-place predicate ‘ERS’ (for ‘Endless Regress of Support’) denote the set of ordered
pairs <x, y> of propositions in P such that x is the first and y is the second member of an endless
S-ordered sequence of propositions in P. So the sentence ‘ERSP0P1’ is true just in case P0 and
P1 are the first two components, in that order, of an endless S-ordered sequence. ‘ERS’ is a
device for referring to endless regresses of reasons by enumerating only their first two members.
An ordered pair of propositions <P0, P1> satisfies ‘ERS’ just in case there is an endless S-
ordered sequence of accessible propositions beginning with P0 and P1—<P0, P1; . . .>—an
endless regress of reasons. Similarly, let ‘IRJ’ (for ‘Infinite Regress of Justified propositions’) be
a two-place predicate satisfied by any ordered pair of accessible propositions hx; yi just in case x
and y are the first members of an infinite S-ordered sequence of justified propositions.

The best extant version of the epistemic regress problem is posed as a paradox, a set of plausible
propositions that are jointly inconsistent:4

(R1) Justification Requires Justified Support. Every justified proposition is supported by

a justified proposition.
 (∀x)(Jx → (∃y) (Jy & Sxy)).

(R2) Some Proposition is Justified. There is at least one justified proposition.

(∃x)Jx.

(R3) Support is Irreflexive. No proposition supports itself.

~(∃x)Sxx.

(R4) Support is Transitive. If a proposition x is supported by a proposition y and y is

supported by z, then x is supported by z.

(∀x)(∀y)(∀z)((Sxy & Syz) → Sxz).

(R5) No Infinite Regress of Justified Propositions. There is no infinite S-ordered

sequence of propositions each member of which is justified.

~ (∃x)(∃y)IRJxy).

(R1)–(R5) are jointly inconsistent. Justification Requires Justified Support and Some
Proposition is Justified jointly imply that there is a proposition P0 and a proposition P1 such that
P0 and P1 are justified and P0 is supported by P1. Given Support is Irreflexive, P0 =/= P1. Since
P1 is justified, Justification Requires Justified Support implies that there is a justified proposition
P2 that supports P1 whence, by Support is Irreflexive, P1 =/= P2. Given Support is Transitive it
follows that P0 is supported by P2, hence, by Support is Irreflexive, P0 =/= P2. At each step,
therefore, the justified proposition that is required to support its predecessor cannot be identical
to any proposition earlier in the sequence, so an infinite regress of justified reasons is required.
(R1)–(R4) thus jointly imply that there is an infinite S-ordered sequence of justified propositions
σ = <P0, P1, . . .>. Since this is true just in case P0 and P1 are the first two members of an
infinite S-ordered sequence of justified propositions, there are propositions x and y that are the
first two members of an S-ordered infinite sequence of justified propositions. This is the negation
of No Infinite Regress of Justified Propositions, so (R1)–(R5) are jointly inconsistent.

(R1)–(R5) are too strong. The paradox assumes only that supporting propositions are supported,
that no supported proposition requires an endless sequence of reasons, and that some proposition
is supported:

(1) Reasons are Supported. Only supported propositions provide support.

(∀x)(∀y)(Sxy → (∃z)Syz).

(2) No Proposition is Supported only by Endless Regresses. Propositions supported only by

endless S-ordered sequences are unsupported.
 (∀x)[(∀y)(Sxy → ERSxy) → ~(∃z)Sxz].

(3) Some Proposition is Supported. At least one proposition is supported by a proposition.5

(∃x)(∃y)Sxy.

(1)–(3) are jointly inconsistent. Reasons are Supported implies that any proposition P0 is
supported by a proposition P1 only if P0 and P1 are the first two members of an endless—
infinite or circular—regress of reasons.6 Given No Proposition is Supported only by Endless
Regresses it follows that that no proposition is supported. This contradicts Some Proposition is
Supported.

If we assume that Knowledge Requires Justification:

For all x, x is known only if x is justified,

and that Justification Requires Support:

For all x, x is justified only if there is a y such that x is supported by y,

the problem also threatens core assumptions about justification and knowledge. Support is the
keystone, however.

A set of jointly inconsistent propositions is a paradox only if each is independently plausible.

What can be said for (1), (2), and (3)?

3 The assumptions are plausible

I now give preliminary cases for (1), (2), and (3). Additional support will be given in my
discussion of proposed solutions.

Why accept Reasons are Supported? A proposition P0 is supported by a proposition P1 only if

P1 provides evidence for P0. This requires that P1 implies P0, but this is insufficient. Support
requires that the relevant propositions satisfy an additional condition such as that expressed by
(R1) Justification Requires Justified Support or that expressed by (1) Reasons are Supported.

To require that reasons be known or justified is too strong since unjustified propositions can
provide support. Minimally credible testimony can be evidence for a proposition it implies even
if we are not justified in believing the testimony itself. Suppose that Squealer’s testimony (P1) I
(Squealer) saw Mugsy do it is minimally trustworthy but falls short of the threshold for
justification. Suppose that given all available evidence, (P0) Mugsy did it is no more reasonable
than its negation. Under these conditions, P1 is evidence for P0 though neither P0 nor P1 are
justified.7 Reasons are Supported is compelling, therefore, because it explains the probative
value of supporting propositions: we have reasons for thinking that they are true even if those
reasons do not satisfy all conditions on justification or knowledge.
I assume that a proposition P1 is a reason for a proposition P1 for a person S only if P0 and P1
are accessible to S, P1 implies P0, and P1 is not epistemically arbitrary from S’s point of view.
A powerful consideration in favor of Reasons are Supported is that it implies that reasons are not
epistemically arbitrary from the believer’s point of view. A proposition is epistemically arbitrary
full stop for a person S in a situation just in case there are no characteristics of that situation that
make believing P epistemically preferable to believing not-P. A proposition is arbitrary from the
believer’s point of view just in case there are no such characteristics accessible to the believer
that make P epistemically preferable to not-P. P is epistemically preferable to not-P for a person
S only if P is more likely to be true than not-P given all of the features of S’s situation. A
characteristic Ф is accessible to a believer S in a situation just in case it is possible for S to
distinguish cases in which Ф obtains from cases in which Ф does not obtain. For example, it is
possible for me to distinguish cases in which I seem to see Mugsy confessing from cases in
which I do not. From my own point of view, however, I am unable to distinguish situations in
which my perception both seems and is highly reliable from situations that are otherwise similar
except that my perception is systematically erroneous. We must not confuse questions about the
conditions in which a proposition P1 is a reason for a proposition P0 for S from the conditions in
which S has a reason to believe the higher-level proposition that P1 is a reason for P0. We have
and use reasons long before we are able to recognize them as such so I assume that the
conditions that make a proposition non-arbitrary from one’s own point of view do not require
this ability.

This does not imply that there are no external conditions on reasons. It implies only that the
propositions that are reasonable for a person S have some characteristic detectable by S that
distinguishes them from unreasonable propositions. If Reasons are Supported is true, there are
discernible features of one’s situation—accessible propositional evidence—that make the target
proposition epistemically preferable to its negation. Reasons are Supported is a weak condition
that helps to explain the probative value of reasons and implies that supported propositions are
not arbitrary from the believer’s point of view.

No Proposition is Supported only by Endless Regresses also plausible. It does not imply that
there are no endless regresses of reasons. It says only that support does not require endless
regresses: if every proposition supporting P0 does so only by means of an endless regress, then
P0 is not supported. Perhaps there are endless regresses that can enhance the support one already
has for a proposition. No Proposition is Supported only by Endless Regresses is plausible
because propositions supported only by circles or by infinite regresses are arbitrary from the
believer’s own point of view. I shall first consider the case of propositions supported only by
circles and then the case of propositions supported only by infinite regresses.

Nothing about a circular sequence of I-ordered propositions σC = <P0, P1, . . ., P0> per se
makes it a sequence of reasons, that is, S-ordered. To be S-ordered, σC must satisfy some
additional condition. If this condition is not arbitrary from the believer’s own point of view it
must include having an independent reason P1’ to believe some member of σC, that is, it must
include having a reason that does not itself depend upon σC. By parity of reasoning, P1’ is not
itself supported only by a circular sequence of reasons.
For example, sequence σL = <I was born in 1957 and will live 70 years, I will die in 2027, I was
born in 1957 and will live 70 years> is a circular I-ordered sequence of propositions. Every
member of σL is supported if its successor is. However, σL is not S-ordered for me even if I am
aware that I will die in 2027 implies I was born in 1957 and will live 70 years and vice versa. For
despite this awareness the members of σL are arbitrary: I have no unconditional reason to believe
any of them. So circular I-ordered sequences of propositions are not S-ordered per se. A circular
I-ordered sequence of propositions is S-ordered only if it has some additional feature that
explains why it is a sequence of reasons.

Consider, then, a circular I-ordered sequence of propositions σC = <P0, P1, . . .; P0>. Each
member of σC is conditionally supported: Pn is supported if Pn+1 is, but σC is not S-ordered per
se. The members of σC must satisfy some additional condition. This condition must include an
independent reason to believe a member of σC. For without a reason P1’—a reason not already a
member of σC —to believe some member of σC, the sequence will be arbitrary from the
believer’s own point of view. If the sequence of propositions that conditionally supports P1’ is
itself circular, the problem arises again. So no proposition is supported only by circular
sequences of propositions. Therefore No Proposition is Supported only by Endless Regresses
holds for the case of circular sequences of reasons.

The same problem afflicts infinite sequences of I-ordered propositions. Every proposition in such
a sequence is supported if its successor is. However, no feature of an I-ordered infinite sequence
of propositions per se explains why any of its members is unconditionally supported. Indeed,
every proposition is the first member of an infinite I-ordered sequence.8 So there must be some
additional feature that distinguishes S-ordered infinite sequences of propositions—infinite
regresses of reasons—from merely I-ordered infinite sequences. If this feature is not arbitrary
from the believer’s own point of view, it must include an independent reason P1’—a reason
outside the infinite sequence—to believe some member Pn of the infinite sequence in question.
By parity of reasoning, what makes P1’ a reason for Pn cannot consist merely in P1’’s being the
first member of an infinite I-ordered sequence of propositions, P1’ must satisfy an additional
condition. If it is not arbitrary from the believer’s own point of view, this condition must include
support by an accessible proposition P1’’ that is not a member of the new regress.9 So No
Proposition is Supported only by Endless Regresses holds for infinite and circular sequences of
I-ordered propositions for the same reason. No Proposition is Supported only by Endless
Regresses is therefore plausible because propositions supported only by endless regresses are
arbitrary from the believer’s own point of view.

Finally, consider Some Proposition is Supported. To reject this proposition is to adopt radical
skepticism about support. If Knowledge Requires Justification and Justification Requires
Support are true but Some Proposition is Supported is false, we have no knowledge and no
justified beliefs. We might, however, lack knowledge or even justified belief but have support
for a proposition. For this is true just in case we have at least one reason to believe some
proposition, whether or not that proposition satisfies additional conditions on justification or
knowledge. Giving this up is even less plausible than giving up the claim that we have
knowledge or that we have justified beliefs. Furthermore, Some Proposition is Supported must
be true if some propositions are more reasonable than others. For if Some Proposition is
Supported is false, then no one ever has any reason to believe any proposition, hence no more
reason to believe one proposition than another. If there is a weak link in the regress paradox,
Some Proposition is Supported does not seem to be it.

The regress problem requires only modest and plausible conditions on support that, alas, are
jointly inconsistent. So at least one plausible core assumption about support is false. But which?
To address this question, I turn to a consideration of some attractive responses to the regress
problem. None succeeds. I conclude that it is difficult to solve the problem while maintaining
that reasonable beliefs are not arbitrary from the believer’s own point of view.

4 Foundationalism and reasons are supported

A foundationalist theory of justification, minimally, holds that (i) some propositions are basic
because they are justified without the support of other propositions and that (ii) any non-basic
proposition is justified only if it is supported by a basic proposition, or supported by a
proposition that is supported by a basic proposition, and so on. Foundationalists therefore reject
Reasons are Supported on the grounds that basic propositions can be reasons that are not
supported by other propositions. Either basic propositions are unsupported or they are supported
by non-propositional aspects of accessible states. Since unsupported propositions are arbitrary
from the believer’s own point of view, foundationalists must hold that some propositions are
supported by accessible non-propositional states. States that lack propositional content such as
unconceptualized sensations are reasons. Unfortunately, such states are insufficient for support.

Black rejects (R1) Justification Requires Justified Support since, for example, there is a smell of
cigarettes is a sense experience, not a belief, that can be a reason to believe someone has been in
the room.10 This is compatible with Reasons are Supported, however, for it does not imply that
only believed propositions can be reasons. There is a smell of cigarettes is a state with
propositional content and thus, by Reasons are Supported, must itself be supported if it is to
support someone has been in the room.

Foundationalists must therefore make a choice. Either basic propositions are unsupported or
some non-propositional features of accessible states are reasons for basic propositions. Since the
former implies that basic propositions are arbitrary from the believer’s own point of view, the
second choice is required if support is not arbitrary.

Moser gamely makes this choice.11 His view—intuitionism—is that all regresses of reasons for
justified empirical propositions terminate with propositions supported by sensory apprehensions
or intuitions.12 Since having sensory intuitions does not involve judging or classifying their
contents, these states lack conceptual, hence propositional content.13 Intuitions can, however,
support propositions believed ‘‘in light of’’ them.14 In sensing ‘‘bluely,’’ for example, one has
defeasible, non-propositional sensory support for the proposition that something blue is before
me. Thus Moser can maintain both that justified propositions require support (where ‘support’ is
now construed to include the relevant relations between the non-propositional contents of
sensory apprehensions and the associated propositions) and that regresses of reasons for
empirical propositions terminate with sensory apprehensions, not propositions. Sensory
apprehensions are non-arbitrary and epistemically relevant because they are immediately
apprehended extra-conceptual states, the unspoiled sensory data that constitute the proper
foundation of our supported empirical beliefs. Because they are accessible mental states with
sensory content, intuitions can provide support that is not arbitrary from the believer’s own point
of view. Because they are not interpretations or classifications, they require no support. So
Reasons are Supported is false. Problem solved, no?

No. Assume we have sensory experiences with non-conceptual content. Such states require no
support because we can be in them without having beliefs or other attitudes that commit us to
any propositions concerning their nature or significance. Call mental states that commit the
person in them to propositions ‘assertoric’ since they involve the sort of commitment to the truth
of propositions that sincere assertions do. For example, beliefs, expectations, suppositions, and
takings are assertoric, but desires, hopes, wishes, and ‘‘raw’’ sensory states are not. The problem
with intuitionism is that non-assertoric sensory states such as unconceptualized sensations are
insufficient to support beliefs. In particular, they cannot support beliefs unless they are
accompanied by assertoric states with the hyperintentional propositional content that is
distinctive of the content of de dicto beliefs. If apprehensions were sufficient for support, we
would have support for propositions we have no reason to believe, a manifest absurdity. Let me
explain.

The content of de dicto beliefs is hyperintentional: the belief that a is Ф is not the belief that b is
ψ even if it is nomologically or even conceptually necessary that a is b and Ф is ψ. The
hyperintentionality of the content of beliefs is reflected in the intensionality of belief
descriptions. S believes that a is Ф is not equivalent to S believes that b is ψ even if it is
somehow necessary that a is b and Ф is ψ. The identity of beliefs and the truth of statements
describing them depend upon the ways in which the relevant states of affairs are represented.15

How might sensory apprehensions support beliefs? The unconceptualized sensation of a cold,
pink surface is evidentially relevant to beliefs about cold by being of cold, not by being of pink.
Thus it is only aspects of experiences, not complex sensations themselves, that could support
particular beliefs. Presumably it is these aspects of experience that are the contents of sensory
intuitions. In Moser’s view, to have sensory intuitions and to form basic beliefs ‘‘in light of
them’’ is to form beliefs on the basis of aspects of our sensations by means of directing our
attention upon the relevant sensory features. So the infrastructure of consciousness must perform
significant discrimination and sorting before it yields the sensory intuitions that constitute the
alleged foundations of empirical knowledge.

Even if we suppose that the processes that yield such intuitions do not involve supporting
assertoric mental states, the resulting intuitions cannot support the corresponding beliefs by
themselves. For the content of a sensory apprehension of even a single property is itself thick
with aspects and there can be many distinct, correct ways of conceptualizing those aspects. To
support the propositional content of a belief, an intuition must involve the direction of attention
not just upon a sensed property, but upon the relevant aspect of that property. The intuition of a
sensed property can support a proposition involving that property only if the apprehension is (or
is accompanied by) an assertoric state with the same hyperintentional propositional content that
is characteristic of the content of de dicto beliefs.
Here is why. Even if a sensation represents a property Ф it will not support propositions about Ф
unless Ф is apprehended in a way that picks out Ф from other sensed properties in the same
sensory complex. The apprehension of a sensory complex that happens to include a sensation of
Ф will not support beliefs about Ф unless the apprehension represents Ф in a way that screens Ф
off from other properties in the same complex: we may not recognize a sensation of Ф even if it
happens to be an element of our complex sensory state. So a sensation of Ф need not support
beliefs about U even if the sensation is evidence for the relevant proposition: Ф must be salient
in the apprehension.

Attention therefore plays a crucial role. A sensation of Ф cannot support beliefs about U unless
one’s attention is directed upon the sensation of Ф itself, that is, unless one’s attention picks the
sensation of Ф out of the complex sensory state that includes it. The trouble is that the sensed
property Ф we attend to itself has many aspects and only hyperintentional states can schematize
Ф in a way that brings the relevant aspects to bear on the target belief. The bare apprehension of
a single sensed property would, by itself, support too much—its content is too ‘‘thick’’—since
there can be ungrasped aspects of that property. The content of attending to a sensation of Ф does
not support a proposition about Ф unless it is accompanied by a hyperintentional assertoric
representation of Ф. Contrary to intuitionism, apprehensions of Ф must be (or be accompanied
by) assertoric states with the relevant hyperintentional propositional content. For unless this is
so, apprehensions of Ф would support beliefs about Ф even if the former do not represent the
aspects of Ф that are expressed by the thin, hyperintentional content of the corresponding
believed proposition.

Consider an example. If property W (being water, say) is (or is otherwise necessarily

coinstantiated with) property H (being H2O), then any sensation of W is a sensation of H. Not
every sensation of H supports beliefs about H, however. For beliefs involving properties that
happen to be signaled by sensations need not be held in light of the relevant aspects of the
sensations, be causally sustained (or otherwise objectively controlled) by the relevant aspects of
the properties they represent, or in any other way be appropriately linked to the relevant aspects
of the sensed properties. A sensation of H may be inaccessible under the relevant description. A
person S can have a sensation x of H (=a sensation of W) that signals the truth of propositions
about H, but even if S’s attention is directed upon the sensory content of X, S’s beliefs about H
can be unsupported. For S might not, perhaps cannot, apprehend x as a representation of H. In
apprehending the nonconceptual sensory content of X, S might not apprehend that content as a
representation of H. So if they are to support beliefs, apprehensions must be or be accompanied
by assertoric hyperintentional propositional states. In this case, S must take the relevant sensation
to be of H2O. Even if the unconceptualized sensory state figures somehow in the support of S’s
belief about H, that support also requires a representation of H that need not be a representation
of W even if it is necessary that H is W.16 Thales had sensations of H2O but could not apprehend
them as such so they did not support propositions about H2O for him, even if he could have had
access to such propositions.17

Intuitionism thus faces a fatal dilemma. If apprehensions of sensory content are sufficient to
support beliefs but lack assertoric hyperintentional propositional content, then they would
support propositions a person has no reason to believe, a manifest absurdity. If, however,
apprehensions have such content, then they might support the corresponding propositions, but
the regress problem remains. For, given Reasons are Supported, the assertoric states by means of
which persons schematize and apprehend the relevant aspects of their sensations must
themselves be supported if they are to provide support for other propositions.

Foundationalists have two choices. They can reject Justification Requires Support by claiming
that basic propositions are justified without support because they have a relevant property that
does not involve having reasons for basic propositions. Some hold, for example, that some
beliefs are basic because they are produced by sufficiently reliable processes. The alternative is
to claim that basic propositions are unjustified: Justification Requires Support is true but
Reasons are Supported is false. According to this view, unsupported, hence unjustified,
propositions can be reasons. Any view that rejects Justification Requires Support is incompatible
with the claim that justified beliefs are not arbitrary from the believer’s own point of view. A
tempting view of the second sort is a version of social contextualism.

5 A contextualist objection to reasons are supported

A social contextualist can hold that justified propositions depend upon unsupported, contextually
immunized reasons that are neither justified nor unjustified. Although this position is able to
maintain that justified beliefs are not arbitrary from the believer’s own point of view, it is false.
It presupposes a mistaken view about the connection between legitimate doubts and having
reasons. Furthermore, the unsupported reasons that are basic in contexts of inquiry are
themselves arbitrary from the believer’s own point of view.

Social contextualism holds that Justification Requires Support is true but that there are
unsupported, hence unjustified, propositions that can be reasons because social practices of
inquiry immunize them from criticism. So Reasons are Supported is false. Specifically, social
contextualism holds that a proposition P is justified for a person in context of inquiry C if, and
only if, that person has reasons for P in C that are strong enough to defeat the doubts about P that
can legitimately be raised in C.18 So Justification Requires Support is true. Propositions that
cannot legitimately be doubted in a context are basic but unjustified in that context for they can
provide support but require none. Whether a doubt is legitimate is determined by the social
conventions governing inquiry in a context. Exactly which propositions are basic is determined
by the relevant conventions and therefore varies across contexts. Since doubts about some
propositions are illegitimate because of their role in a particular context of inquiry, they are
basic. Basic propositions can be reasons but are neither justified nor unjustified. Since every
context immunizes some propositions from doubt, each context includes basic propositions that
are, so to speak, unsupported supporters of justified propositions.

Assuming that a social contextualist says about justification what Wittgenstein says about
knowledge,19 the position outlined in On Certainty is a version of social contextualism that
accepts Justification Requires Support but rejects Reasons are Supported: some reasons are basic
since they make inquiry possible but are neither justified nor unjustified. Support for justified
propositions can be provided by propositions that ‘‘stand fast’’ for us but that we cannot
correctly claim to be justified or unjustified.20 For to claim to be justified in believing a
proposition is ipso facto to admit that doubts about it are legitimate and thus to commit oneself to
having reasons for it.
Social contextualists can thus maintain that the value of justified propositions consists (partly) in
our having reasons for them and, contra intuitionism, that only propositions provide support. The
propositions that are justified in a context of inquiry require support, but all sequences of reasons
terminate with propositions that cannot correctly be said to be justified because they are partly
constitutive of the context of inquiry and are thus immune from doubt.21 The basic propositions
that terminate sequences of reasons are partly constitutive of the social infrastructure required for
inquiry. Claims that such propositions are justified and claims that they are unjustified are
equally incorrect or meaningless. G.E. Moore is mistaken when he gives his famous list of some
of the things he takes himself to know despite the fact that it would be equally mistaken to deny
that he knows them.22 Demanding reasons for such propositions expresses inappropriate doubt.

Social contextualism thus implies that there is a connection between the propriety of expressing
doubt about a proposition P and the requirement that P be supported, for according to social
contextualism P requires support for a person only if it is appropriate to demand that the person
provide reasons for P. This confuses the appropriateness of demanding or giving reasons with
the value of having them. In particular, social contextualism presupposes that Reasons are
Supported is incompatible with the view that in any given context of inquiry there are some
reasons for which support cannot legitimately be asked. But these are compatible. For it can be
legitimate in a specific context to require that no one be allowed to demand reasons for a
proposition P because this will prevent the inquiry from proceeding, while it is also true that
unless one has reasons for P, P is arbitrary from one’s own point of view.

Consider this analogy. While a professor is demonstrating a theorem, it is inappropriate for logic
students to demand documentary evidence of the professor’s qualifications for teaching the
course, but this does not mean that the professor need not have the qualifications. Similarly, there
is no connection between the inappropriateness of expressing doubt about a proposition P in
specific contexts and the epistemic requirement that one have reasons for P. Indeed, if we lack
reasons for the assumptions we make in specific contexts we would lack the ability to recognize
them as non-arbitrary, trustworthy guides to the truth.

Inquiries take much for granted since the benefits of cooperation and the shortness of life require
that we rule some questions out-of-bounds in specific contexts. It is bad form if an historian, for
example, objects to an account of a battle on the grounds that the account presupposes that there
are truths about the past. However, unless truth is merely what is or would be defensible given
the assumptions and practices that happen to structure inquiry, it can be legitimate to investigate
whether what we normally accept without demanding reasons is true. Such fundamental
questioning has sometimes revolutionized our understanding and our methods of inquiry. Even
when they are embedded in the conventions that make specific inquiries possible, assumptions
undertake evidential debts. When the epistemic bills come due, if persons have no reasons to
think that their assumptions are true, their enterprise is bankrupt because their assumptions are
arbitrary. In this case, the arbitrariness is not due to an inaccessible difference between believed
propositions and their negations, for it is presumably accessible to persons whether or not there is
a consensus about what is to be accepted without question. The problem is with the other
constraint on arbitrariness: the characteristics that are accessible to the believer must make
propositions epistemically preferable to their negations.23 Unless truth just is what is produced by
consensus, consensus does not make propositions epistemically preferable to their negations.

Social contextualism is a close relative of the most powerful and troubling version of skepticism.
Skeptical conventionalism claims that because we cannot have evidential support for the
propositions we believe, we should adopt the beliefs and practices of our own time and place.
Skepticism arrives at conventionalism by way of the claims that the conditions on evidential
support cannot be satisfied—Reasons are Supported and No Proposition is Supported only by
Endless Regresses, for example—but that we must have beliefs in order to carry on.
Contextualism arrives at conventionalism by means of the analogous claims that there can be no
reasons for our basic assumptions and that what counts as a reason is just what our practices
dictate. Since, according to contextualism, the only concepts of evidence we have are those
embodied in our practices, what we take to be good reasons are good reasons. At bottom, social
contextualism is skeptical conventionalism disguised with the fig leaf of unsupported, merely
conventional reasons.

Social contextualism confuses the conditions in which it is acceptable to demand reasons with
the conditions in which it is valuable to have them. Further, since it holds that the propositions
that constitute the foundations of inquiry are unsupported, they are arbitrary because they lack an
appropriate connection to truth.

6 Transitivity and circularity

Post proposes to solve the received version of the regress problem by rejecting (R4) Support is
Transitive.24 This will not solve the more resilient problem, however. Still, this position deserves
further consideration because Support is Transitive is connected to the important question of
whether circular sequences of propositions can be support-affording. Unlike the received version
of the problem, no consistent subset of (1)–(3) implies that circles of reasons are always vicious.
So it will not solve the regress problem to claim that some circles of reasons are not vicious.

Post argues that the relevant epistemic support relationship is not transitive since there are
propositions P0, P1, and P2 such that P0 is supported by P1 and P1 is supported by P2, but P0 is
not supported by P2. If so, this solves the received version of the problem. It does not solve the
more resilient problem, however, for no consistent subset of (1)–(3) implies Support is
Transitive.

Support is Transitive is, however, relevant to a question that is important to the regress problem:
is circularity always vicious? Support is Transitive and Support is Irreflexive, if true, jointly
imply that the answer is ‘‘yes.’’ For Support is Irreflexive implies that no proposition is
supported by itself. Thus no small circle—a sequence of propositions with the form <P0, P0>—
is support-affording. What about big circles, I-ordered sequences such as σBC = <P0,
P1, . . .Pm=(P0)> such that P0 = P1? A big circle σBC is support-affording only if each of its
members is supported by its successor, if it has one. By m applications of Support is Transitive,
σBC is support-affording only if P0 is supported by Pm (=P0), that is, only if P0 is supported by
itself. This is incompatible with Support is Irreflexive. So Support is Irreflexive and Support is
Transitive jointly imply that no circular sequence of propositions is support-affording. Rejecting
Support is Transitive thus opens the door to the possibility of solving the regress problem by
holding that big circles of reasons can be support-affording. For without the assumption of
transitivity, big circles of reasons need not be cases of self-support.

The concept of the evidential ancestry of a proposition will simplify our discussion of
circularity.25 The evidential ancestry of a proposition can be specified recursively:

(A1) If P0 is supported by P1, then P1 is in the evidential ancestry of P0.

(A2) If P1 is in the evidential ancestry of P0 and P1 is supported by P2, then P2 is in the

evidential ancestry of P0.

(A3) Nothing is in the evidential ancestry of P0 except in virtue of a finite number of

applications of (A1) and (A2).

Support is Irreflexive and Support is Transitive jointly imply:

No Circles. No proposition is in its own evidential ancestry.

So circularity is always vicious.

By contrast, No Proposition is Supported only by Endless Regresses implies:

No Required Circles. If a proposition P0 is supported by any proposition P1 only if some

proposition is in its own evidential ancestry in the corresponding sequences <P0, P1, . . .,
Pm(. . .)>, then P0 is not supported by any proposition.26

No Circles implies No Required Circles, but not vice versa. In particular, No Required Circles is
compatible with the possibility that a circular sequence is S-ordered because it enhances the
support one has for its first member P0 even if P0 is in its own evidential ancestry. No Required
Circles implies that there is no supported proposition that is supported only if it is in its own
evidential ancestry. It does not imply that all circles are vicious, only that supported propositions
must have independent support. In particular, every supported proposition Pn must be supported
by at least one proposition Pm which is such that Pn is not in the evidential ancestry of Pm. So
No Required Circles implies only that circles of reasons are not, by themselves, sufficient for
support. This is compatible with the claim, accepted even by some foundationalists, that
propositions can acquire additional support by being in their own evidential ancestry.27

Consider a circular sequence of propositions σSC = <P0, P1, . . ., Pm = (P0), Pm+1, . . .> such
that neither Pm+1 nor any of its successors is identical to any other member of the sequence
σSC. Although P0 is in its own evidential ancestry, σSC might be S-ordered by bringing
additional evidence—the evidence provided by P1-Pm–1 and the propositions not in σSC that
might support them—to bear on P0. I think that there are support-affording circular sequences of
this sort,28 but their existence does not solve the regress problem. For No Proposition is
Supported only by Endless Regresses does not imply No Circles. To solve the regress problem by
appeal to circles of reasons we would need to show that a proposition can be supported even
though it is supported only by circles of reasons. We have seen that this is false.

I conclude that, unlike the received version, the resilient version of the regress problem does not
imply No Circles. Unfortunately, this makes the problem more difficult to solve, for the mere
existence of S-ordered circular sequences of propositions is not incompatible with No
Proposition is Supported only by Endless Regresses.

My earlier argument for No Required Circles clearly assumes a ‘‘linear’’ conception of support.
In particular, it assumes that if a proposition P0 is supported by P1 and Pn is in the evidential
ancestry of P1, then Pn is in the evidential ancestry of P0. Perhaps no proposition is supported
only by linear circles of reasons, but the regress problem can be solved by showing that the non-
linear, global coherence of propositions is sufficient for support. I now turn to this possibility.

7 Coherence without circles

Until his conversion to foundationalism,29 BonJour held that a believed proposition is justified
for a person S just in case the propositions S believes are coherent.30 BonJour rejected the view
that circles of reasons can provide justification, claiming instead that justification is holistic.
Coherence is not ‘‘linear’’ circular support—which would require circles of reasons or
unsupported supporters—but of ‘‘mutual or reciprocal support.’’31 In particular, coherence
requires ‘‘[t]he inferability of [a] particular belief from other particular beliefs, and further
inference relations among particular beliefs.’’32

This does not solve the problem. Are justified beliefs supported (Justification Requires Support)?
That beliefs in a coherent set must be ‘‘inferable’’ from others does not settle this since
inferability is a logical, not an epistemological, relation. Every raven I have ever seen or heard
about has been black is inferable from all ravens are black and vice versa. If justification
requires coherence and coherence requires that the believed propositions in a coherent set be
supported by supported propositions, then justification requires an endless regress. If, on the
other hand, a set of beliefs can be coherent without every member of the set being supported by a
supported proposition, then either support is not required for justification—Justification
Requires Support is false—or unsupported propositions can provide support—Reasons are
Supported is false. So either holistic coherentism does not solve the regress problem or it is a
version of foundationalism, contextualism, or some other theory incompatible with Reasons are
Supported.

One might reply that this criticism misses the point: the coherence of a set of beliefs itself
supports the members of the set. So Justification Requires Support is true, but support is
provided by the coherence of the set of beliefs itself and not by individual members of the set.

This equivocates on ‘support.’ The coherence of a set of believed propositions might contribute
to the positive epistemic standing of a proposition in the set by, for example, making it more
likely to be true. Coherence cannot, however, support any proposition in the relevant sense, for
coherence per se is not the sort of thing that could be a reason. Coherence cannot support a
proposition in a coherent set any more than inconsistency can support the negation of a
proposition in an inconsistent set: coherence and inconsistency are not propositions, they are
properties of sets of propositions.33 The proposition P is a member of a coherent set of
propositions might support P, but this is a different matter. Furthermore, even if coherence is a
non-arbitrary characteristic that helps make a proposition preferable to its negation, requiring
belief in the coherence of one’s beliefs for justification or support would be both too strong—one
must have the relevant concept of coherence34—and too weak—to be a reason, this proposition
needs support, and the regress returns.

Holistic coherentism thus faces a fatal dilemma: either it does not solve the regress problem or it
denies that Reasons are Supported. In the latter case, either awareness of its coherence with
one’s other beliefs is required for a proposition to be a reason or it is not. If so, then the regress
returns, for belief in the coherence of one’s beliefs will require support. If not, then a proposition
might be non-arbitrary full stop because it has a good-making epistemic property—coherence—
but it will be arbitrary from the believer’s own point of view. For if I am unaware that P is but
not-P is not coherent with my beliefs, I have no reason to prefer one to the other.

8 Infinitism

There remains the possibility that we might acquire support for a proposition by means of an
infinite regress of reasons. Reasons do not Require an Endless Regress does not, however, imply
that infinite regresses are not support-affording, only that supported propositions do not require
infinite regresses of reasons. This is correct, since propositions that are supported only by infinite
regresses of reasons would be arbitrary from the believer’s own point of view.

The conjunction of Reasons are Supported and No Required Circles entails:

Infinite Regresses Required. A proposition P0 is supported only if P0 is the first member

of an infinite sequence of propositions each of which is supported by its successor.

Two roads lead from these assumptions to the negation of Some Proposition is Supported, a very
strong version of skepticism. One assumes that infinite regresses block support, period:

No Infinite Regresses. There is no proposition P0 such that P0 is the first member of an

infinite sequence of distinct propositions each member of which is supported by its
successor.

The more secure road assumes only this consequence of No Proposition is Supported only by
Endless Regresses:

No Required Infinite Regress. For all propositions P0, if P0 is supported by any

proposition Pn only if P0 and Pn are the first two members of an infinite regress of
reasons, then P0 is not supported by any proposition.

No Required Infinite Regresses says that no supported proposition is such that its being
supported requires an infinite regress of reasons. This does not imply No Infinite Regresses. Even
if no supported proposition requires an infinite regress, such sequences of propositions might be
support-affording. Indeed, it seems it would enhance the support one has for a proposition if it
were the first member of such a regress. So the more resilient version of the regress argument
implies No Required Infinite Regresses, but not No Infinite Regresses.

No Required Infinite Regresses is prima facie correct. Indeed, that Reasons are Supported and
No Required Circles jointly imply that support requires an infinite regress of reasons seems
paradoxical. Klein, however, argues that rejecting No Required Infinite Regresses is the key to
solving the regress problem.35 He holds that we can have justified beliefs, that justification
requires support, but that circular reasoning cannot provide it. Hence, there are justification-
providing infinite regresses of distinct propositions. Indeed, support requires infinite regresses.
By making infinite support-affording regresses of propositions necessary, but not sufficient, for
justification and by carefully distinguishing between occurrent and dispositional reasons, Klein
successfully responds to the most popular extant objections to his ‘‘infinitism.’’

Infinitism must, as we have seen, distinguish infinite regresses of reasons from arbitrary, non-
probative sequences of I-ordered propositions. Infinitism therefore requires that, in addition to
being the first members of infinite sequences of propositions, justified propositions have an
additional property Ф that does not, on pain of circularity, require justified beliefs or an infinite
regress, but that does distinguish sequences of genuine reasons from I-ordered, but epistemically
arbitrary, sequences of distinct propositions. This cannot be done, for any plausible account of
the conditions under which P0 is supported by P1 will either require a justified belief (thus
making infinitism circular) or will be incompatible with the infinitist commitment to Reasons
are Supported. Indeed, for the reasons that I have given above, Ф must include having a reason
that does not require an infinite regress, otherwise a proposition ostensibly supported by an
infinite regress of I-ordered propositions will be arbitrary from the believer’s own point of view.
So No Required Infinite Regress is true.36

Even if No Required Infinite Regress were false, that would be a Pyrrhic victory. For few, if any,
of our empirical beliefs are supported by infinite regresses since most of us lack even
dispositional access to reasons for our reasons ad infinitum.37 By explaining the possibility of
justified beliefs at the cost of their actuality, infinitism swallows the spider to catch the fly.

9 Conclusion

The epistemic regress problem is a powerful challenge to the consistency of some of our core
assumptions about evidential support and, by extension, about knowledge and justified belief.
Each assumption is compelling. Taken together they are intolerable. A solution will not come
easily. For to solve the regress problem we must find a reason to reject at least one proposition
that seems to be essential to any concept of support that is compatible with supported
propositions being non-arbitrary from the believer’s own point of view.

Acknowledgements I am very grateful for helpful comments on the topics covered in this paper
to Scott Aiken, Daniel Allen, Curtis Bridgeman, Ed Damer, Catherine Elgin, Deborah Heikes,
Peter Klein, Jonathan Kvanvig, Ben Letson, Stanislav Lioubomoudrov, Peter Markie, Brian
Martine, John Post, Jeffrey Tlumak, and to the audiences at my 1995 Presidential Address to the
Alabama Philosophical Society, ‘‘Epistemology and the Meaning of Life,’’ my presentation of a
related paper at Emory and Henry College in 1997, and my presentation of a distant ancestor of
this paper to the 2001 annual meeting of the Tennessee Philosophical Association. Thanks also
to the students in my Summer 2002 section of Philosophy 101 who were required to read and
comment on a very obscure ancestor of this paper. Special thanks to Aiken, Heikes, Kvanvig,
Markie, Post, and Tlumak for detailed, penetrating comments on what I thought was the
penultimate draft. I alone am responsible for errors that remain.

1 I use ‘implies’ to abbreviate ‘stands in the epistemically relevant logical or quasi-logical

relation to.’

2 A similar point is made by Sosa (1980, p. 13).

3 In case the evidence that is available for a proposition consists of more than one proposition,
we may represent it as the conjunction of the relevant propositions or as the set of those
propositions.

4 Post (1993, pp. 209–212) who follows Black (1988). Another account of the general structure
of regress arguments along similar lines is given by Sanford (1975). I have modified Post’s
formulation of the problem somewhat.

5 The first-order versions of (1)–(3) are formally inconsistent only assuming that Reasons are
Supported Requires an Endless Regress: (∀x)(∀y)(Sxy → (∃z)Syz) = (∀x)(∀y)(Sxy → ERSxy).
This proposition is evident.

6 Indeed, we have the stronger consequence that Reasons are Supported just in case any
proposition’s being supported implies that there is an endless I-ordered sequence:
(∀x)(∀y)(Sxy → (∃z)Syz) = (∀x)(∀y)(Sxy → ERSxy).

7 Elgin (2005, p. 157) presents a case to similar effect.

8 See Post (1980) and Post (1987, pp. 84–92).

9 I spell out this argument in greater detail in Cling (2004).

10 Black (1988, p. 425).

11 Moser (1985, pp. 141–210).

12 I focus on Moser’s view because it is one of the most careful and complete defenses of
intuitionist foundationalism. My argument, however, is intended to show that any view on which
sensory states suffice for support is bound to fail.

13 Moser (1985, p. 164).

14 Moser (1985, p. 184).

15 So beliefs have what Dretske (1981, pp. 172–174) calls the ‘‘second’’ and the ‘‘third orders’’
of intentionality.

16 Cling (1991).

17 Compare to the distinction between ‘‘objective’’ and ‘‘subjective intentionality’’ in

Churchland (1979, p. 14).

18 The contextualism I discuss is suggested in Wittgenstein (1969) and is akin to the view
defended by Annis (1978), though Annis thinks unsupported reasons can be known. A subtle
version of contextualism in the spirit of Wittgenstein is defended by Williams (1991). Some
recent versions of contextualism have not focused on accounts of knowledge or justified belief
per se but on the context-relativity of the truth conditions of sentences that attribute knowledge
to a person, for example ‘Sam knows that the cat is on the mat.’ These accounts attempt to show
that ordinary attributions of knowledge are not threatened by the skeptical doubts raised by
skeptical hypotheses such as I am a brain in a vat because, in most contexts, the truth conditions
for such attributions do not require the subject to be in a position to defeat such hyperbolic
doubts. In the rare contexts in which one needs to be in a position to defeat such hypotheses, then
ordinary knowledge attributions are false but—as I would put it—the kind of knowledge we
would lack is different from the ordinary kind we still have. Such skeptical hypotheses raise the
ordinary standards for knowledge. For such views, see DeRose (1994) and Lewis (1994).
DeRose gives an account of the truth conditions for knowledge attributions in terms of the
sensitivity of beliefs to changes in facts across relevant possible worlds and does not endorse an
evidence requirement on ordinary knowledge while Lewis explicitly rejects an evidence
requirement for knowledge. Even if DeRose succeeds in defending some types of knowledge
against hyperbolic skeptical doubt, either his view does not provide a solution to the regress
problem or, like Lewis’, the kind of knowledge that is immune to skeptical doubts can be
arbitrary from the believer’s own point of view. In any case, since the reason requirement that
generates the regress problem is weak—it can be satisfied without one having the ability to
defeat hyperbolic skeptical hypotheses—it is not clear how semantic contextualism has the
resources to solve the regress problem.

19 I shall not enter the dense thicket of Wittgenstein exegesis. I only investigate an interesting
epistemological view suggested by some of his remarks. I leave it to others to decide what
Wittgenstein believed.

20 Wittgenstein (1969, §151).

21 Ribeiro (2001, §1.8).

22 Moore (1959a, 1959b), Wittgenstein (1969, §58).

23 I am grateful to an anonymous referee for helping me to see this point.

24 Post (1993).
25 The term ‘evidential ancestry’ is due to Klein (1999, p. 298).

26 I am indebted to John Post for pointing out serious limitations of a previous version of No
Required Circles.

27 For example Chisholm (1977, pp. 82–84).

28 Cling (2002).

29 BonJour (2001).

30 BonJour (1976).

31 BonJour (1976, pp. 285–286, and 1985, pp. 89–93).

32 BonJour (1976, p. 287).

33 Sosa (1980, pp. 8–9), Klein (1999, pp. 317–318).

34 Acquiring evidence for such a claim will also be non-trivial given the computational
complexity involved in making determinations of coherence (Milgram 2000).

35 Klein (1999).

36 For more see Cling (2004).

37 BonJour (1976, pp. 23–24) makes the stronger claim that finite minds could not have the
infinite number of beliefs that such a regress requires. Klein (1999, pp. 306–310) shows that
infinitism does not require this.

Epistemic and Dialectical Regress

Michael Rescorla

Abstract: Dialectical egalitarianism holds that every asserted proposition requires defence when
challenged by an interlocutor. This view apparently generates a vicious “regress of
justifications,” since an interlocutor can challenge the premises through which a speaker defends
her original assertion, and so on ad infinitum. To halt the regress, dialectical foundationalists
such as Adler, Brandom, Leite, and Williams propose that some propositions require no defence
in light of mere requests for justification. I argue that the putative regress is not worrisome and
that egalitarianism can handle it quite satisfactorily. I also defend a positive view that combines
an anti-foundationalist conception of dialectical interaction with a foundationalist conception of
epistemic justification.

I. The regress argument(s)

My topic is the perennial “regress of justifications.” If a proposition depends for its justification
upon a second proposition, and that second proposition depends for its justification upon a third
proposition, and so on, then infinite regress seemingly ensues. As Alston [1989: 26-32] and Audi
[1993: 118-120] note, the regress arises in two versions: epistemological and dialectical. The
first, concerning the structure of justification, emerges when we ask what justifies a given belief.
The second, concerning the structure of rational dialectic, emerges when a speaker justifies an
assertion to an interlocutor who persistently challenges her justifications. Recent attention has
focused mainly on the epistemological regress. But a few commentators, such as Brandom
[1994], Klein [1999, 2003], Leite [2005], and Williams [1999, 2004], deploy the dialectical
regress to support sweeping doctrines about dialectical interaction and epistemic justification. I
will argue that this explanatory strategy is misguided. Unlike the epistemological regress, the
dialectical regress is harmless. It does not expose significant features of either dialectical
interaction or epistemic justification. It is a red herring.

Philosophers who study dialectical interaction often hope to derive epistemological conclusions.
Sometimes the conclusions are sceptical, as with the Pyrrhonians, sometimes antisceptical, as
with Austin [1979], Brandom [1994], Leite [2005], Klein [1999, 2003], and Williams [1999,
2004]. I will sketch a view that combines foundationalism about epistemic justification with anti-
foundationalism about dialectical interaction. On the proposed view, doctrines about the
structure of rational dialectic neither entail nor follow from doctrines about the structure of
justification. Thus, investigating dialectical interaction is not a good method for establishing
substantive epistemological conclusions. Alston, Audi, and many others have stressed the same
point for decades. But philosophers seldom note a converse point: misplaced emphasis upon
justification and knowledge yields a distorted picture of dialectical interaction. Specifically, I
will argue that misplaced concern with scepticism leads Brandom, Leite, and Williams to impose
a needless foundationalist structure upon rational dialectic. If we treat epistemic status and
dialectical interaction as too closely connected, we impede our study of both phenomena.

II. The epistemological regress

I begin with the epistemological regress. Because it is so familiar, I will be brief. An evidential
chain is a sequence of beliefs, each justified by the preceding belief in the sequence. How might
evidential chains be structured? The only options are:

(E1) There is an evidential chain that is infinitely long.

(E2) There is an evidential chain that circles back upon itself.

(E3) There is an evidential chain that begins with an unjustified belief.

 (E4) There is an evidential chain that begins with a belief (a so-called epistemically basic
belief) that is justified but that does not depend upon any other belief for its justification.
(E5) There are no justified beliefs, and hence no evidential chains.

Aiken [2005] and Klein [1999, 2003] endorse (E1): infinitism. While I know no conclusive
argument against infinitism, most epistemologists deem it implausible. (E2) boasts some
advocates, especially among coherentists, but most epistemologists again reject it.1 (E3) seems
incredible, although Wittgenstein [1969: §253] is sometimes interpreted as endorsing it. (E5)
strikes virtually everyone as unacceptable. That leaves (E4), epistemic foundationalism, as the
most promising option. But how should we develop it into a convincing position?

Alston [1989], Audi [1993], Peacocke [2004], Pryor [2000], and others argue that perceptual
experiences provide immediate, prima facie, defeasible justification for associated beliefs. For
instance, my perceptual experience of seeing that a red cube is located before me justifies the
belief that a red cube is located before me. The justification does not depend upon ancillary
beliefs, such as the belief that my perceptual systems are functioning reliably. However, suitable
ancillary beliefs might support or defeat it. Memories, such as my memory of seeing a red cube,
can likewise justify beliefs, such as my belief that I saw a red cube. Again, the justification
depends upon no ancillary beliefs, although suitable beliefs might support or defeat it. Thus, non-
doxastic mental states, such as memories and perceptual experiences, can provide justification
without themselves requiring justification. A belief that acquires justification in this way is
epistemically basic.

I will assume that Alston, Audi, Peacocke, and Pryor are correct. Suitable relations to suitable
non-doxastic mental states can confer justification upon a belief, without mediation by other
beliefs.

III. The dialectical regress

To examine the dialectical regress, I begin with Brandom [1994]. Following Sellars [1963],
Brandom treats assertion as a move within “the game of giving and asking for reasons”: the
activity through which we rationally assess propositions by providing one another with
arguments and counter-arguments. I will refer to this activity as reasoned discourse. A key idea
behind Brandom’s account is that assertion involves a commitment to defend what I say in
response to challenges and counter-arguments [1994: 173]. Brandom countenances several ways
I might discharge this commitment: by providing reasons to believe the asserted proposition; by
noting that another speaker asserted the proposition; by citing the reliability of my perceptual
faculties. For our purposes, such details are irrelevant. What matters is that defending an asserted
proposition requires asserting additional propositions.

My interlocutor can challenge those additional assertions, which I may defend with further
assertions. Infinite regress looms. To halt the regress, Brandom proposes that some propositions
do not require defence in light of mere requests for justification [1994: 177].

Say that a proposition is dialectically basic iff it requires no defence in light of mere requests for
justification. Call the view that such propositions exist dialectical foundationalism. A
dialectically basic proposition requires defence, if at all, only when my interlocutor offers special
supporting considerations, such as reasons to doubt the proposition or reasons to doubt that I am
justified in believing it. Dialectically basic propositions halt the regress by shifting the burden of
proof from speaker to interlocutor. They provide a foundation of stable, if not immutable, resting
points within conversation. This response to the dialectical regress stretches back to Aristotle [cf.
Barnes 1990: 120-123]. Besides Brandom, contemporary advocates include Adler [2002], Leite
[2005], Norman [1997], and Williams [1999, 2004].

Which propositions are dialectically basic? Aristotle regarded a privileged class of “first
principles” as furnishing an axiomatic foundation for science. Adler [2002, pp. 159-160] holds
that every proposition is dialectically basic. Brandom, Leite, and Williams regard the line
between basic and non-basic as shifting with conversational context. For instance, Leite claims
that it varies with the speaker’s epistemic circumstances [2005: 405].

Say that a challenge to some assertion is legitimate iff correct participation in reasoned discourse
requires the speaker to meet the challenge by defending her assertion. Following Williams
[2004], say that a challenge to an assertion is brute iff it is a mere request for justification,
accompanied by no supporting considerations. Dialectical foundationalists claim that a brute
challenge to a dialectically basic proposition is not legitimate. In contrast, dialectical
egalitarians claim that all challenges are legitimate, including brute challenges. On this view, all
assertions require defence when faced with brute challenges. Reasoned discourse does not assign
any propositions a default role in our reasoning with one another, even relative to context. There
are no privileged resting points in the game of giving and asking for reasons.

Historical advocates of egalitarianism include the Pyrrhonian sceptics and Neurath [1983].
Contemporary advocates include Klein [1999, 2003] and van Eemeren and Grootendorst [2004:
135-140]. Traditionally, the main motivation for it has been abhorrence of dogmatism. When
arguing with another speaker, it seems arbitrary and dogmatic to refuse to defend some disputed
premise. If there is nothing be to said in favour of p, how can I advance p as a reason for
believing other propositions? And if there is something to be said in favour of p, shouldn’t I say
it? This worry is a constant refrain among egalitarians, from Sextus’s attack on the “mode of
hypotheses” to Klein’s complaint that dialectical foundationalism ‘appears to advocate a process
of reasoning that relies upon arbitrary propositions at the base’ [2003: 81].

Because all arguments employ premises, any theory of reasoned discourse must specify which
premises speakers may invoke without defence during a given conversation. Dialectical
egalitarianism provides an extremely simple answer: one may invoke without defence whichever
premises one’s opponent concedes. Dialectical foundationalism is more complicated. By
insulating certain premises from brute challenges, it introduces additional structure into reasoned
discourse. The question is whether we need this additional structure. Do we require a set of
default propositions to halt the putative regress of justifications? I will argue that we do not.
Dialectical egalitarianism can analyze the putative dialectical regress just as satisfactorily as
dialectical foundationalism.

One might object to egalitarianism in various additional ways. For instance, one might cite a
putative intuition that it is deviant to advance a brute challenge against a speaker who asserts that
she has hands. I have addressed such intuitions elsewhere, arguing that dialectical egalitarianism
can accommodate them quite satisfactorily [Rescorla forthcoming b]. The intuitions are
irrelevant to my discussion here, which focuses solely upon whether the regress of justifications
supports dialectical foundationalism.

IV. Examining the dialectical regress

Let us examine more closely how the regress arises. Say that a speaker vindicates an assertion
during a conversation if, by the conversation’s end, she successfully meets all legitimate
challenges her interlocutor advances to that assertion, all legitimate challenges her interlocutor
advances to propositions asserted while meeting legitimate challenges to the initial assertion, all
legitimate challenges her interlocutor advances to propositions asserted while meeting legitimate
challenges to those latter propositions, and so on. Suppose I encounter an interlocutor who issues
a brute challenge to some proposition I assert, then issues a brute challenge to the propositions I
assert while defending the original proposition, and so on ad infinitum. Following Leite, call this
individual a persistent interlocutor. What is the result of my dialectical interaction with a
persistent interlocutor? The only options are:

(D1) I vindicate my assertion by providing an infinite chain of arguments.

(D2) I vindicate my assertion by providing a circular argument, i.e. an argument that

invokes a previously asserted disputed proposition as a premise.

(D3) I vindicate my assertion by providing a chain of argument that includes an

undefended, disputed premise. The premise is not dialectically basic.

(D4) I vindicate my assertion by providing a chain of argument that includes an

undefended, disputed premise. The premise is dialectically basic.

(D5) I do not vindicate my assertion.

(D1) is impossible, since human life spans are finite. (D2) is dubious. How could a circular
argument discharge my dialectical commitments? Doesn’t it just beg the question? (D3) is a
contradiction. If my argument rests upon an undefended proposition that my opponents disputes,
and if the premise is not dialectically basic, then the premise requires further defence, so I have
not vindicated my original assertion. Thus, we may set aside (D1), (D2), and (D3), leaving only
(D4) and (D5). Dialectical foundationalists affirm (D4). Dialectical egalitarians affirm (D5).

For the regress to support dialectical foundationalism, we require an argument against (D5).
From the egalitarian perspective, however, (D5) is innocuous. If I encounter a sufficiently
recalcitrant interlocutor, then she and I will fail to converge upon mutually acceptable, relevant
premises. Either I provide a circular argument, or I provide no argument for some disputed
premise, or I retract my assertion. What is wrong with saying that, on those occasions when I
encounter a persistent interlocutor, I fail to vindicate my position? Either speaker and
interlocutor agree upon mutually acceptable relevant premises, in which case the regress halts, or
speaker and interlocutor do not agree upon mutually acceptable premises, in which case the
speaker leaves certain dialectical commitments undischarged.

As the foregoing analysis suggests, dialectical egalitarianism entails

The Vulnerability Thesis: When defending an assertion against a persistent interlocutor, one
will fail to vindicate that assertion.

The Vulnerability Thesis is a dialectical analogue to (E5), the sceptical doctrine that no beliefs
are justified. The difference is that, while (E5) is intolerable, the Vulnerability Thesis is
plausible. It reflects the unfortunate fact that there is simply no reasoning with some people.

There are two main ways one might attack our egalitarian analysis of the dialectical regress: by
arguing that the Vulnerability Thesis entails undesirable epistemological consequences, and by
arguing that, independently of epistemological considerations, it yields an untenable conception
of reasoned discourse. In sections V and VI, I consider how egalitarianism relates to various
epistemological doctrines. In section VII, I examine whether egalitarianism yields a viable
conception of dialectical interaction.

V. Scepticism and egalitarianism

It may seem that the Vulnerability Thesis engenders global scepticism, whereby one withholds
judgment regarding virtually all propositions. The Pyrrhonians are often interpreted as arguing
along these lines [Barnes 1990]. On this interpretation, the Pyrrhonians deployed the dialectical
regress to induce global suspension of judgment. Crudely: they inferred global scepticism from
dialectical egalitarianism. This crude description is a bit misleading, since global scepticism
presumably involves withholding judgment from all philosophical doctrines, including
dialectical egalitarianism and even the thesis that one should withhold judgment from all
propositions. Nevertheless, many expositions of Pyrrhonism, including [Barnes, 1990], implicitly
presuppose egalitarianism.2 Dialectical foundationalists usually accept the inference from
egalitarianism to scepticism, but they deploy it against egalitarianism. For instance, after
outlining a Pyrrhonian argument for scepticism, Williams notes that we can defuse the skeptical
argument by adopting a foundationalist conception of reasoned discourse [2004: 133].

The inference from egalitarianism to scepticism, whether deployed to support skepticism or rebut
egalitarianism, presupposes something like:

The Vindication Thesis: One is justified in believing p only if one can, at least in
principle, vindicate p. (The qualification “at least in principle” is meant to handle
“performance” errors stemming from exhaustion, inebriation, etc.)

Dialectical foundationalists usually embrace the Vindication Thesis, either explicitly, as with
Leite [2004] and Norman [1997], or implicitly, as with Brandom [1994] and Williams [2004].
Consider Williams’s diagnosis of how egalitarianism engenders scepticism: egalitarianism
‘allows the sceptic to enter brute challenges: challenges that are apparently presuppositionless…
Since… presuppositionless challenges can be entered anywhere and everywhere… the
impossibility of meeting them shows something about the epistemic standards of all our beliefs’
[2004: 134]. The impossibility of answering iterated brute challenges shows nothing about
epistemic justification unless we assume the Vindication Thesis. Granting that iterated brute
challenges are legitimate engenders scepticism only if we assume that justificatory status is tied
to the ability to meet legitimate challenges.

Although Williams opposes Pyrrhonian scepticism, he shares with the Pyrrhonians a picture of
how justification relates to reasoned discourse, encapsulated by the Vindication Thesis. Given
this picture, the central question becomes which challenges are legitimate and hence what it
takes to vindicate a proposition. An egalitarian answer leads to scepticism, while a
foundationalist answer does not.

The Vindication Thesis is heir to a venerable philosophical tradition that elucidates justification
by studying justificatory transactions between speakers. Proponents of this tradition include
Austin [1979], Sellars [1963], Toulmin [1958], and Wittgenstein [1969]. Not coincidentally, all
four philosophers espouse doctrines resembling dialectical foundationalism. In its most typical
modern form, derived from Sellars, this tradition regards justification as explanatorily derivative
from norms governing reasoned discourse. Williams, crediting Brandom, puts the point as
follows: we should ‘think of “being justified” in one’s beliefs as enjoying a certain normative
status within “the game of giving and asking for reasons”’ [2004: 127]. Such a perspective
renders the Vindication Thesis quite natural.

Yet many contemporary epistemologists reject the Vindication Thesis. For instance, Alston
[1989: 70] and Audi [1993: 145] argue that justification is a positive epistemic status possession
of which does not presuppose the ability to defend one’s beliefs. On this approach, we must
sharply distinguish the state of holding a justified belief from the activity of justifying
propositions to one another.

Following Alston and Audi, non-sceptical egalitarians should reject the Vindication Thesis and
kindred doctrines. They can thereby escape the inference from egalitarianism to scepticism.
When faced with a persistent interlocutor, I fail to vindicate my assertions. It does not follow that
the corresponding beliefs are unjustified or that I should suspend them. From my inability to
defend some proposition with reasons my opponent accepts, it does not follow that I lack any
reason to believe the proposition.

A common objection to the Vindication Thesis is that ordinary speakers are seemingly justified
in believing numerous propositions that they cannot defend against challenges [Alston
1989: 70; Howard-Snyder and Coffman 2006: 556]. For instance, if I learn a proposition through
testimony (Napoleon was defeated at the Battle of Waterloo), I often cannot remember where
exactly I learned it. So I cannot defend it. But I am seemingly justified in believing it. Beliefs
based on perception (There is a red cube in front of me) or proprioception (I am raising my arm),
along with background presuppositions of ordinary conversation (The world has existed for more
than five minutes), offer further compelling examples.

Proponents of the Vindication Thesis can respond to such counter-examples by invoking

dialectical foundationalism. Classify as dialectically basic any proposition that I am justified in
believing but that I cannot defend. Then the proposition is no longer an immediate
counterexample to the Vindication Thesis, because a brute challenge to it does not require an
answer. This manoeuvre does not completely defuse the objection, since it ignores the many
cases where a speaker apparently cannot answer any challenge, no matter how well-motivated, to
a proposition he justifiably believes. But the manoeuvre diminishes the objection’s force. In this
way, the Vindication Thesis and dialectical foundationalism “take in one another’s washing.”
Dialectical foundationalism solves a problem, the inference from the Vulnerability Thesis to
global scepticism, that arises only given the Vindication Thesis. And the Vindication Thesis
becomes remotely plausible only if we accept dialectical foundationalism.

Another familiar objection to the Vindication Thesis is that young children, and possibly also
non-linguistic animals, are justified in believing many propositions, even though they cannot
mount sustained arguments.

In response to this objection, Leite concedes that there may be some sense in which young
children and non-linguistic creatures have justified beliefs, but he insists that this sense is
different than that in which normal adult humans have justified beliefs [2004: 243-245].
Although many philosophers would disagree, let us grant the point for the sake of argument.
Once Leite concedes that there is any sense in which a creature can have a justified belief
without being able to justify the belief, he can retain the Vindication Thesis only in a diluted
form that applies to certain kinds of justification and not others. But then the Vindication Thesis
no longer draws out devastating sceptical consequences from dialectical egalitarianism. At the
worst, it entails that there is one kind of justification we never attain, leaving open that we attain
the other kind already attained by young children and non-linguistic creatures. Thus, even if
Leite’s approach is plausible on its own terms, it undercuts the argument against dialectical
egalitarianism that we have been pursuing.

A final objection to the Vindication Thesis presupposes the version of epistemic foundationalism
from section II. On the proposed foundationalist view, non-doxastic mental states, such as
perceptual experiences and memories, prima facie justify associated beliefs. If we accept this
view, as increasingly many contemporary epistemologists do, then the Vindication Thesis should
not strike us as remotely plausible. A belief is justified if it is caused in a suitable way by a
suitable non-doxastic mental state. This positive epistemic status does not constitutively involve
an ability to defend what one says. It depends solely upon appropriate causal, cognitive, and
epistemic relations between a belief and a non-doxastic mental state. Ability to participate a
certain way in reasoned discourse seems irrelevant.

VI. Epistemic foundationalism and dialectical egalitarianism

The previous section suggests a general moral. All too often, the debate between dialectical
foundationalism and egalitarianism is conflated with some epistemological debate. But
dialectical foundationalism and egalitarianism are rival views about the structure of reasoned
discourse. They entail epistemological claims only when conjoined with additional doctrines
relating reasoned discourse to epistemic status. One can reject most such doctrines, so one can
coherently combine either dialectical foundationalism or egalitarianism with virtually any
conception of epistemic justification.
Neurath combines egalitarianism with coherentism [1983]. Klein combines it with infinitism
[1999, 2003]. Perhaps one cannot coherently combine egalitarianism with global scepticism,
since the latter presumably involves suspending judgment in all propositions. Nevertheless, the
Pyrrhonians employed egalitarianism to motivate global scepticism.

As already indicated, I favour the epistemological thesis that certain non-doxastic mental states
prima facie justify associated beliefs. I now want to sketch an account that combines this
foundationalist epistemology with dialectical egalitarianism.

Suppose John has a perceptual experience that represents a red cube in front of him. The
experience prima facie justifies the belief that a red cube is located in front of him. That belief is
epistemically basic. But the corresponding assertion is not dialectically basic. Suppose John
asserts ‘A red cube is located in front of me’ in the course of defending some other proposition
(e.g. ‘There are red cubes in this building’). If an interlocutor challenges John’s assertion, then
John must provide an argument for it, even if his interlocutor sees the cube. The interlocutor will
not concede the premise, so it would be dogmatic to invoke it without further argument. John is
justified in believing the premise, but that does not excuse him from defending it.

What about the assertion ‘I seem to see a red cube located in front of me’? Surely it requires no
defence in light of a mere request for justification? Admittedly, dialectical egalitarianism looks
less compelling when applied to propositions about one’s own mental states. Yet Klein [2003:
84] argues that even these propositions require defence when challenged, on pain of dogmatism.
I have defended the same conclusion elsewhere [Rescorla forthcoming b]. What matters here is
whether my proposed merger of epistemic foundationalism and dialectical egalitarianism
convincingly handles the regress of justifications. Defending the proposed view against all other
possible objections is a task much too large for a single paper.

Many philosophers seek a privileged class of observation sentences: sentences, about either
one’s observable surroundings or one’s perceptual experiences, directly based upon sensory
input. The hunt for observation sentences preoccupied various logical positivists, especially
Schlick [1979]. More recently, Brandom writes that observation sentences ‘can function as
unjustified justifiers: claimings that are treated as having a defeasible default status as entitled…
So observation provides regress-stoppers, and in this sense a foundation for empirical
knowledge’ [1994: 222]. In contrast, I am proposing a view on which non-doxastic perceptual
experiences can halt the epistemic regress, even though there are no privileged perceptual reports
to halt the dialectical regress. On this view, empirical knowledge rests on a foundation provided
by observation, but empirical discourse does not rest on a foundation of observation sentences.

The proposed marriage of dialectical egalitarianism and epistemic foundationalism may seem
unstable. How can my belief in p depend upon no other beliefs for its justification if my assertion
of p requires backing by additional assertions?

The answer is that cognition features non-doxastic mental states, such as perceptual experiences
and memories, that justify beliefs without themselves requiring justification. These non-doxastic
mental states halt the epistemic regress. Reasoned discourse features nothing that can play an
analogous role. Speakers cannot somehow share the same perceptual experiences or memories. I
have my perceptual experiences, and you have yours. My perceptual experience can justify my
beliefs, but they cannot somehow justify my corresponding assertions. I can describe my
perceptual experience, perhaps saying ‘I see a red cube’ or ‘I seem to see a red cube.’ But these
are just further assertions, which may themselves be challenged.

In this connection, consider how Klein deploys the dialectical regress against epistemic
foundationalism [2003: 82-83]. Klein imagines an epistemic foundationalist, Fred, who takes his
belief that p to be epistemically basic. Fred asserts p, and a persistent interlocutor asks, ‘What
makes you think p is true?’. According to Klein, Fred faces a dilemma. He may provide some
reason for thinking that p is true, in which case ‘the regress has not actually stopped… Fred has
given up his foundationalism’ [2003: 83]. Or Fred may realize that he can offer his interlocutor
no reason for thinking p is true, in which case he should suspend judgment in it. The first horn of
the dilemma should lead Fred to abandon epistemic foundationalism, while the latter should lead
him to abandon his belief in p. As Klein emphasizes, Fred cannot avoid the dilemma by claiming
that his belief in p is epistemically basic. Fred’s interlocutor can reply, ‘So what? Either there’s a
reason for believing p, or there isn’t. If there is, tell me. If not, your position is dogmatic.’

Klein draws several conclusions. First, epistemic foundationalism does not block the dialectical
regress. Specifically, one cannot halt the regress merely by observing that some belief is
epistemically basic. Second, epistemic foundationalism does not ‘provide a model of reasoning
that can be rationally practiced’ [2003: 82]. Fred’s dilemma shows that, on pain of irrationality,
the epistemic foundationalist must either abandon his foundationalism or else abandon his
putative epistemically basic beliefs.

Klein’s first conclusion reflects a broadly egalitarian stance with which I am sympathetic.
However, I reject Klein’s second conclusion, which he does not carefully distinguish from the
first. I think that neither horn of Fred’s alleged dilemma should trouble epistemic
foundationalists. The first horn attacks a straw man. Epistemically basic beliefs possess some
justification that does not depend upon other beliefs. They may also possess further justification
that derives from other beliefs. Thus, Fred can cite reasons for believing p without abandoning
epistemic foundationalism [Alston 1989: 38]. The second horn of Klein’s dilemma moves
without argument from claims about reasoned discourse to claims about what Fred should
believe. From the fact that Fred cannot defend p, it does not follow that Fred should suspend
judgment in p. That conclusion follows only if we accept one or another dubious thesis linking
epistemic status to one’s standing within reasoned discourse. For similar criticisms of Klein, see
[Howard-Snyder 2005; Howard-Snyder and Coffman 2006].

Largely on the basis of Fred’s Pyrrhonian dilemma, Klein rejects epistemic foundationalism in
favour a view that combines infinitism with dialectical egalitarianism: ‘[t]here is always another
reason, one that has not already been employed, that can legitimately be required for each reason
that is given for a belief. Only if there is an infinite set of nonrepeating reasons available for a
belief is it fully justifiable’ [2003: 86]. Klein does not distinguish the infinitist component of his
view (a justified belief rests upon an infinite evidential chain) from the much more anodyne
egalitarian component (one can legitimately require that a speaker defend every asserted
proposition). Once we distinguish the state of holding a justified belief from the activity of
justifying claims to one another, the dialectical phenomena cited by Klein do not support
infinitism over epistemic foundationalism.

What if Fred replicates the Pyrrhonian regress within his own thinking? He can demand of
himself a reason for believing p, where he takes p to be epistemically basic. He can then
recognize that he enters into some non-doxastic mental state, such as a perceptual experience,
and that the state provides reason for believing p. Since the non-doxastic state does not stand in
need of justification, the regress halts. Fred might ask himself the further questions: ‘What is my
reason for believing that I am in a non-doxastic state with certain properties?’ and ‘What is my
reason for believing that the non-doxastic state provides any reason for believing p?’. These are
interesting questions. But Fred does not need to answer them to be justified in believing p or to
be rational in retaining that belief. The non-doxastic state, not the higher-belief that he is in the
state or the higher-order belief that the state provides reason for believing p, is Fred’s reason for
believing p. Even if Fred can think of no reason for holding either higher-order belief, it does not
follow that he should suspend his belief in p. That would follow only if we accepted one or
another dubious thesis linking higher-order justification and first-order justification, such as ‘one
is justified in believing p only if one is justified in believing that one is justified in believing p’
[Alston 1989: 153-171].

Davidson famously claims that ‘nothing can count as a reason for holding a belief except another
belief ’ [2001: 141]. Following Alston, Audi, Peacocke, Pryor, and many others, I reject this
doctrine. Perceptual experiences, among other non-doxastic mental states, can justify beliefs.
However, one can reject the Davidsonian doctrine and simultaneously endorse an analogous
doctrine regarding assertion: only an assertion can serve as a reason for an assertion. More
precisely, although less pithily: only an assertion can serve as a premise in defending another
assertion, and every assertion of a disputed proposition requires some defence.

One discharges a dialectical commitment by providing an argument. How does one provide an
argument? By asserting additional propositions. Thus, as Brandom himself emphasizes [1994:
167], assertion occupies a dual role in reasoned discourse: through it, one both undertakes and
discharges dialectical commitments. (See also [Rescorla forthcoming a].) This dual role
generates a clear threat of regress. Brandom reacts by positing dialectically basic propositions,
thereby qualifying the initial intuition that assertion involves a commitment to defend what one
says. Dialectical egalitarians, untroubled by the regress, retain the intuition in its original form.
Once we disentangle questions about reasoned discourse from questions about epistemic status,
we can capture Brandom’s insights into assertion more faithfully than Brandom’s own theory
does.

Cognition rests upon mental states, like memories and perceptual experiences, that justify while
requiring no justification. Reasoned discourse involves a central speech act, assertion, that both
justifies and requires justification. A given thinker’s space of reasons exhibits a pervasively
foundational architecture, resting upon reason-giving items that do not themselves require further
reasons. The game of giving and asking for reasons incorporates no comparable reason-giving
items. While cognition is rationally constrained by non-doxastic mental states, conversation
lacks any analogous boundary of non-assertoric speech acts. Dialectical foundationalists, seeking
to mitigate this disanalogy between rational thought and rational dialectic, introduce a privileged
class of foundational propositions that justify without requiring initial justification. But perhaps
we should simply acknowledge that assertion occupies a different structural role within reasoned
discourse than belief occupies within cognition.

VII. A defective dialectical practice?

I turn now to non-epistemological arguments that the dialectical regress undermines

egalitarianism. A practice that allows the regress to occur strikes some philosophers as
‘degenerate’ [Norman 1997: 485]. As Brandom puts it, ‘nothing recognizable as a game of
giving and asking for reasons results if justifications are not permitted to come to an end’ [1994:
177]. I will pursue several worries in this vein.

Objection: If dialectical egalitarianism is correct, then it is impossible to vindicate any assertion

within reasoned discourse.

Reply: Dialectical egalitarianism entails that one cannot vindicate any assertion against a
persistent interlocutor. But most interlocutors are not persistent. In practice, speakers usually
agree fairly easily upon many relevant propositions. Even when speakers disagree violently, they
can usually fall back upon the vast range of background beliefs that we all share [Adler 2002:
135-185]. Vindicating a proposition requires responding to actual challenges, not to challenges
someone might potentially offer. Thus, on the egalitarian model, it is both possible and routine to
vindicate assertions.

Objection: ‘If a speaker both accepts a challenge (for free) and follows the rule that challenges
require the production of specific reasons, the speaker will take a large range of his assertions to
be asserted improperly, since he will lack those specific reasons’ [Adler 2002: 183]. So
dialectical egalitarianism entails that most assertions are incorrect. Taken to its logical extreme,
egalitarianism mandates ‘widespread withdrawal from assertion’ [Adler 2002: 181].

Reply: Egalitarianism entails nothing about the correctness or incorrectness of one’s initial
assertion. It describes what one must do after the assertion is challenged. For instance,
egalitarians can endorse Williamson’s [2002] proposal that the conditions for correct assertion
are given by the Knowledge Norm: one should assert p only if one knows p. Assume that I can
know some proposition p without being able to justify it to other speakers. Suppose I assert p. If
another speaker challenges my assertion, then I fail to defend it. The proposed egalitarian view
can say that my initial assertion was perfectly proper and correct, since it conformed to the
Knowledge Norm, even though I subsequently failed to discharge the dialectical commitment I
undertook by asserting p.

Dialectical egalitarianism entails that most (or all) of our assertions are incorrect only if we
accept the following thesis: an assertion is correct just in case the speaker could, in principle,
defend it in response to most (or all) legitimate challenges that might arise. There is no reason for
egalitarians to accept this thesis. For further discussion, see [Rescorla forthcoming a].
Objection: According to egalitarians, assertion involves a commitment to the impossible feat of
performing infinitely many tasks: justifying the asserted proposition, justifying the premises
thereby employed, justifying the new premises, and so on.

Reply: By asserting a proposition, I commit myself to defending it with an argument if

challenged to do so. The commitment is conditional. If I assert a proposition that my interlocutor
does not challenge, my commitment to defend it is not activated. If it is activated, I can discharge
it by providing an argument of some appropriate sort. In doing so, I undertake a new conditional
commitment: to defend the premises asserted during my argument, if those premises are
challenged. This process can continue indefinitely. At each stage, I undertake only a finitary
conditional commitment. I never undertake a commitment whose content is that I perform some
infinite range of tasks.

Objection: Egalitarianism entails that, by asserting a proposition, I implicitly commit myself to

an indefinite series of justifications. If I encounter a persistent interlocutor, then at any stage I
can avoid providing an argument only by shirking one of the commitments entrained by my
initial commitment. Thus, the initial commitment has infinitary implications, even though they
are not part of its explicit content.

Reply: There is nothing wrong with a practice that allows speakers to undertake commitments
with infinitary implications. On the contrary, we know at least one perfectly respectable practice
with this feature: promising. Take the genie who promises to perform any three actions
commanded by his master. The master commands the magical genie to perform two desirable
actions. As his final wish, the master demands that the genie promise to perform any additional
three actions commanded by the master, a demand to which the genie must comply. This
procedure can now iterate indefinitely. Thus, the genie’s original promise contains the seeds of
an infinitary commitment, just as assertion, according to egalitarians, contains the seeds of an
infinitary commitment. A practice that allows an implicitly infinitary commitment is not
necessarily defective.

Admittedly, the genie is a special case. Most promises do not involve an implicitly infinitary
commitment. In contrast, according to egalitarians, every assertion involves an implicitly
infinitary commitment. However, this contrast is not worrisome. It reflects intrinsic differences
between promissory and assertoric commitment. A promise involves a commitment to perform
some action, not necessarily a further promise, under certain circumstances. An assertion
involves a commitment to perform some action, necessarily a further assertion, under certain
circumstances. Thus, the potential for infinitary commitment is essential to assertion in a way
that it is not essential to promising. Each assertoric performance contains the seeds of an
indefinite dialectical commitment, whereas only special promises contain the seeds of indefinite
promissory commitment.

Objection: Dialectical egalitarianism leaves speakers at the mercy of a persistent interlocutor.

As Norman puts it, egalitarianism ‘rig[s] the reason-giving game in the challenger’s favour’
[1997: 487]. Simply by the reiterating the question, ‘How do you know?’, interlocutors can
enforce dialectical stalemate. Intuitively speaking, it should not be so easy to prevent a speaker
from winning an argument.
Reply: The putative “intuition” has little force, since it amounts to an intuition that the
Vulnerability Thesis is false. Dialectical egalitarians claim not to share this intuition. They
acknowledge that their position entails the Vulnerability Thesis, and they accept that
consequence with equanimity. The proposed objection does not so much argue against this
position as announce that the objector disagrees with it. Rather than providing a non-circular
reason to doubt egalitarianism, the objection generates a Kuhnian stalemate of conflicting
theory-laden intuitions.

To bolster this analysis, consider the following intuition: brute challenges to the assertion I have
hands are somehow deviant. This intuition, which apparently supports dialectical
foundationalism, strikes me as powerful and nearly unanimous. Any adequate egalitarian account
must accommodate it, a task that I undertake in [Rescorla forthcoming b]. The intuition is a
datum for philosophical theorizing, a kind of pragmatic analogue to the grammaticality intuitions
that serve as data for generative linguistics. In contrast, a putative intuition that the Vulnerability
Thesis is false amounts to little more than a judgment regarding whether some philosophical
theory is plausible. An ordinary speaker would experience this intuition, if at all, only after
exposure to suitable definitions, elucidations, arguments, and distinctions, that is, only after
initiation into philosophical discourse. Such a speaker would experience the intuition not in her
capacity as an ordinary speaker, but rather in her capacity as an amateur philosopher. Intuitions
of this kind are certainly relevant to philosophical theorizing, but their force and interest rapidly
diminish when they are not widely shared by rival philosophers.

The important issue here is not whether the Vulnerability Thesis strikes some philosophers as
“counterintuitive,” but whether a dialectical practice to which the Vulnerability Thesis applies
must exhibit some incontrovertible defect. We have yet to encounter any compelling argument
that it must. I submit that, once we address the various worries enumerated earlier in this section
and in previous sections, the spectre of a persistent interlocutor should not seem troubling. Even
though such an interlocutor can easily bar speakers from vindicating assertions, that theoretical
possibility has little practical import for quotidian linguistic interaction. Why should we
condemn a dialectical practice as degenerate merely because it countenances this possibility?
Dialectical egalitarianism does leave us at the mercy of a persistent interlocutor, but that does not
prevent ordinary speakers from exploiting reasoned discourse to resolve their disagreements in
light of mutually accepted premises.

Objection: It is intuitively defective to offer iterated challenges without accompanying

explanations or reasons for doubt [Williams 2004: 133]. Egalitarianism does not yield this
intuitive verdict. It treats iterated brute challenges as perfectly acceptable.

Reply: There is something to this objection. But our intuitions are more nuanced than it allows.
Although iterated brute challenges strike us as intuitively defective, we also recognize a sense in
which they are perfectly appropriate. Leite captures the tension well: ‘the persistent
interlocutor’s requests for reasons seem to be simultaneously licensed and misguided’ [2005:
398]. A good account should explain and preserve this apparent tension. It should isolate a sense
in which the interlocutor’s questions are licensed, and another sense in which they are
misguided. Dialectical foundationalists can accomplish the latter task, but it is doubtful that they
can accomplish the former.3 In contrast, as I will now argue, dialectical egalitarians can
accomplish both tasks.

A basic purpose of reasoned discourse is to isolate mutually acceptable premises relevant to the
truth of disputed propositions. Mutually acceptable premises provide a neutral evidentiary base
for adjudicating disputes. A neutral evidentiary base may not decisively resolve a dispute, but it
serves as common ground. By isolating it, speakers achieve what I will call rapprochement. Only
by achieving rapprochement do they engage one another rationally. One might say that, if two
speaker cannot agree on relevant premises, then they succeed only in talking at one another,
rather than reasoning with one another. Thus, a speaker who recognizes no need to achieve
rapprochement does not fully grasp the point of reasoned discourse. In this sense, rapprochement
is a constitutive goal of reasoned discourse. For further discussion of rapprochement, see
[Rescorla 2007; van Eemeren and Grootendorst 2004: 139].

We can now explain why the persistent interlocutor’s conduct seems ‘simultaneously licensed
and misguided.’ Iterated brute challenges are licensed, insofar as they violate no norm of
reasoned discourse. But they are misguided, insofar as they flout a basic point of reasoned
discourse: achieving rapprochement. The persistent interlocutor obeys the norms of reasoned
discourse while subverting one of its goals.

Objection: How can one engage correctly in a practice while subverting one of its basic
purposes or goals?

Reply: This is a widespread, harmless phenomena. Consider a tennis player who deliberately
misses certain shots to let his opponent win. He plays the game in a deviant fashion, because he
does not make a genuine effort to win. Yet his deviant conduct differs markedly from that of a
player who cheats, perhaps by calling a ball out when it is in. He follows the rules of tennis
while subverting the point of the game: winning [Raz 1990: 117-118].

As a closer analogy, consider the filibuster. A United States senator who gains the floor can hold
it by talking continuously. Under Rule 22 of the Senate, no vote may occur until the senator
cedes the floor or until a two-thirds majority votes to end the filibuster. Theoretically, then, a
persistent senator can prolong deliberations indefinitely, destroying a bill’s chance to become
law. Whether pursued nobly or ignobly, the filibuster is a technique for manipulating the
Senate’s rules. The rules promote free and open debate, culminating in a legislative verdict. A
filibuster travesties this design, prolonging debate so as to forestall any legislative verdict. Such
conduct is deviant, because it subverts the basic goal of any legislative body. As Senator Thomas
Eagleton once complained, ‘the Senate is now in a state of incipient anarchy. The filibuster…
has now become a routine frolic in almost all matters. Whereas our rules were devised to
guarantee full and free debate, they now guarantee unbridled chaos’ [Congressional Record,
November 23, 1985, S33453].4

The persistent interlocutor manipulates reasoned discourse, much as filibustering manipulates the
Senate’s deliberative procedures. The norms of reasoned discourse promote free and open
debate, hopefully culminating in rapprochement and possibly resolution of the dispute. The
persistent interlocutor stonewalls, challenging everything his opponent says without asserting
any propositions himself. He enforces dialectical stalemate, in which conversationalists converge
upon no propositions relevant to their dispute. Such conduct is highly deviant, because it
subverts a fundamental goal of dialectical interaction: rapprochement.

Dialectical foundationalism attempts to prohibit this deviant conversational behaviour. By

classifying certain propositions as dialectically basic, it builds safeguards against stonewalling
directly into reasoned discourse. The idea is that, because the persistent interlocutor presses his
challenges too far, he no longer places the original speaker on the defensive. In contrast, I think
that stonewalling places the original speaker on the defensive. That is why we find it so
outrageous. The persistent interlocutor is maddening, not because he fails to raise legitimate
challenges, but because he succeeds in raising legitimate challenges. He remorselessly forces the
original speaker to discharge dialectical commitments, yet he offers no help in doing so, and he
undertakes no analogous commitments of his own. Rather than erecting safeguards against
stonewalling, we should admit that reasoned discourse, like almost any other practice, can be
abused. Structural features of reasoned discourse do not ensure a well-conducted conversation,
any more than the rules of the Senate ensure a well-conducted legislative deliberation.

Objection: A lingering sense may remain that dialectical interaction requires a foundation. How
can we reason with one another if our reasoning does not include a neutral evidentiary base of
propositions upon which we agree?

Reply: Fruitful reasoning requires that conversationalists identify mutually acceptable premises.
Thus, participants in reasoned discourse must jointly strive for rapprochement. Since we share so
many background beliefs, achieving rapprochement is usually quite easy. Occasionally, it is
difficult or even impossible. Either way, reasoned discourse does not itself provide a neutral
evidentiary base of propositions. Locating those propositions is up to the speakers. Rational
dialectic, unlike rational cognition, involves a clash between opposing viewpoints. Nothing
intrinsic to linguistic interaction ensures that the opposing viewpoints will isolate argumentative
common ground [van Eemeren and Grootendorst 2004: 134].

VIII. Reason and reasoned discourse

Philosophers who study how we reason with one another often hope to illuminate reason itself.
They seek to elucidate rational cognition by analyzing rational conversation. Brandom and
Sellars illustrate this approach most clearly, urging that rational thought is constitutively tied to
the game of giving and asking for reasons.

I think that any adequate account of how reason relates to reasoned discourse must cultivate a
healthy appreciation of their differences. To highlight those differences, I have sketched a hybrid
view that combines epistemic foundationalism with dialectical egalitarianism. On the proposed
view, rational cognition exhibits a pervasively foundational structure, while rational conversation
does not. Reasoned discourse is not just rational cogitation transplanted into the public sphere,
and rational cogitation is not just reasoned discourse internalized. They are two distinct modes of
rational activity, with two very different architectures.5 University of California, Santa Barbara
Notes

1 For a sophisticated development of coherentism, see [BonJour 1985]. Rather than embrace
(E2), BonJour rejects the choice between (E1)-(E5) as resting upon a spuriously “linear” and
“non-holistic” conception of justification.

2 According to “urbane” interpretations, Pyrrhonians intend their scepticism to apply only to

philosophical propositions, not to ordinary non-philosophical propositions. Frede [1987]
advocates the urbane approach, and Barnes [1990] opposes it. Fogelin [1994] offers a modern
development of urbane Pyrrhonism.

3 Williams [2004: 135-137] offers an error theory to explain away why we experience the
intuition that iterated brute challenges are licensed.

4 For extensive discussion of the history and politics of the filibuster, see [Binder and Smith:
1997]. This discussion conclusively demonstrates that the rules of the Senate were not designed,
as proponents of the filibuster sometimes allege, to safeguard minority rights by allowing the
minority to filibuster.

5 I am indebted to Jonathan Adler, C. Anthony Anderson, Anthony Brueckner, Kevin Falvey,

Warren Goldfarb, Matthew Hanser, Elizabeth Harman, Richard Heck, Charles Parsons, and
Nathan Salmon for helpful feedback. I am also grateful to Juan Comesaña for his commentary
when I presented this paper at the Pacific APA 2008, as well as to audience members present on
that occasion, especially Selim Berker and Ralph Wedgwood. Finally, I thank two anonymous
referees for this journal for comments that helped improve the paper.

Notre Dame Journal of Formal Logic

Volume 50, Number 2, 2009
Justification by an Infinity
of Conditional Probabilities
David Atkinson and Jeanne Peijnenburg

Abstract Today it is generally assumed that epistemic justification comes in degrees. The
consequences, however, have not been adequately appreciated. In this paper we show that the
assumption invalidates some venerable attacks on infinitism: once we accept that epistemic
justification is gradual, an infinitist stance makes perfect sense. It is only without the assumption
that infinitism runs into difficulties.
1 Introduction

Foundationalism and coherentism come in various sorts and sizes, but the difference between the
two is clear: foundationalists hold that basic beliefs justify nonbasic beliefs while coherentists
maintain that beliefs justify one another and that basic beliefs do not exist. In some versions of
foundationalism basic beliefs depend on basic nonbeliefs, and in some versions of coherentism
some beliefs are more fundamental than others, but this is a fine structure that for our purposes
we may neglect.

To gain a better understanding of the difference, let us consider the simplest of all belief systems,
namely, an epistemic chain. In such a chain, (a belief in) a target proposition E0 is justified by (a
belief in) proposition E1, which in turn is justified by E2, and so on:

E0 ←← E1 ←← E2 ←← E3 ←← . . . , (1)

where En ←← En+1 means ‘En+1 justifies En’. Of course, most belief systems are much richer
and more intricate than this. Nevertheless, the chain is a good starting point: it can help us to
understand more realistic cases, which have been represented as trees, rafts, pyramids, teepees,
houses of cards, cobwebs, or crossword puzzles, all of which have single chains as their
elements.

Now foundationalists will argue that (1) only makes sense if it finally comes to a halt in a basic
proposition corresponding to Es+1, where s is some finite number:

E0 ←← E1 ←← E2 ←← E3 ←← . . . . . . ←← Es+1. (2)

Coherentists, on the other hand, characteristically claim that the chain will eventually make a
loop, thus supporting itself:

E0 ←← E1 ←← E2 ← E3 ←← . . . . . . ←← Es{←← E0} . (3)

Both (2) and (3) presuppose that the chain is finite. This is in accordance with the positions of
both foundationalists and coherentists, namely, that chains of infinite length do not make sense.
Adherents of infinitism, on the other hand, allow s to go to infinity; indeed they argue that
epistemic chains have the form

E0 ←← E1 ←← E2 ←← E3 ←← . . . . . . . . . ∞ (4)

In this paper we are especially interested in the viability of infinitism, and so we will focus on
chains like (4).

Some see justification as a normative relation, claiming that the expression ‘En+1 justifies En’
states something about the logical connection between the justifier and the justified. Others
conceive it as an empirical connection, holding that En+1 brings about En in one way or another.
Still others believe that justification is best captured by a double aspect theory, arguing that it
incorporates both normative and empirical aspects.
Apart from the question of whether justification is normative or empirical, on which we will
remain neutral, there is the issue of strength. How close should the connection between En and
En+1 be in order to say legitimately that the one is justified by the other? Should we take an
austere position and say that En+1 only justifies En if it logically implies or singly causes En? Or
could we be more lenient, admitting that justification is a gradual concept, allowing a more or
less? Today an increasing number of epistemologists opt for the latter alternative. In an attempt
to make the concept of epistemic justification more general and more realistic, they prefer to
speak about ‘justification’ even if En+1 only probabilistically supports En (to a certain degree
above some threshold).2

In this paper we follow suit and regard the justification relation as a probability relation. This
means that, from now on, we interpret (1) as follows: E0, the target proposition, is
probabilistically supported by E1, which in turn is probabilistically supported by E2, and so on.
To symbolize this adoption of probability we replace the double arrow ← of justification tout
court by a single arrow ←, signifying probabilistic support. Thus the chain in which we are
interested, namely, (4), becomes

E0 ← E1 ← E2 ← E3 ← . . . . . . . . . . . . ∞ (5)

It has often been argued that infinitism must be incorrect because an infinite chain of
propositions, the one justified by another, does not make sense. In the present paper we will
show that such arguments are untenable if the idea that justification comes in degrees is taken
seriously. Once justification is interpreted in terms of probabilistic support, then certain
venerable arguments against infinitism become questionable. Among these are the argument that
as yet no example of an infinite epistemic chain has been found (Black [1], p. 436), the argument
that such chains cannot be completed (Klein [10], p. 920; cf. Lehrer [11], p. 155–56), the
argument that the support given by an infinite epistemic chain always culminates in zero (Lewis
[13], p. 172; cf. Lewis [12], pp. 327–28), and the argument that “if all justification is conditional.
. . , then nothing can be shown to be actually, nonconditionally justified” (Dancy [5], p. 55). A
variant of the latter argument has recently been put forward by Gillet, who holds that an infinite
regress of deferred justification cannot yield an unconditional, determinate justification for the
target proposition (Gillet [8]). Against all these arguments we will show that the gradual
character of justification allows the existence of infinite epistemic chains that may indeed be
completed, yielding a final unconditional probabilistic justification that is not zero. In Sections 3
and 4 we will give detailed examples of such infinite chains, showing that they can be perfectly
consistent. But first, in Section 2, we explain more carefully what we mean by “probabilistic
support,” for that will allow us to formulate our examples with more precision.

2 Probabilistic Support

We will say that En+1 probabilistically supports En if and only if En is more probable if En+1 is
true than if it is false. In other words, the conditional probability of En, given that En+1 is true, is
greater than the conditional probability of En, given that En+1 is false:

P(En|En+1) > P(En|¬En+1). (6)

The notion of probabilistic support is closely allied to that of (probabilistic) confirmation. There
are, however, many measures of confirmation currently on the market (Douven and Meijs [6]). A
popular one is the Carnap [3] degree of confirmation that En+1 gives to En. It is defined by

D(En|En+1) = P(En|En+1) − P(En).

Another measure, used by Christensen [4] and Joyce [9], is

S(En|En+1) = P(En|En+1) − P(En|¬En+1). (7)

The condition of probabilistic support (6) is equivalent to S(En|En+1) > 0. In fact, since one can
show that D(En|En+1) = S(En|En+1)P(¬En+1), it follows that (6) is also equivalent to D(En|
En+1) > 0, provided that P(¬En+1) does not vanish, that is, provided that P(En+1) =/= 1.
Actually all the so-called Bayesian measures of confirmation are positive if (6) is satisfied
(Fitelson [7]; Shogenji [16]).

The rule of total probability is

P(En) = P(En|En+1)P(En+1) + P(En|¬En+1)P(¬En+1), (8)

and this may be rewritten in the form

P(En) = P(En|¬En+1) + S(En|En+1)P(En+1), (9)

where S(En|En+1) is the Christensen–Joyce measure (7).

For convenience we will frequently make use of the abbreviations

αn = P(En|En+1)

ẞn = P(En|¬En+1)

γn = S(En|En+1) = αn − ẞn. (10)

This means that equation (9) takes on the form

P(En) = ẞn + γnP(En+1), (11)

which is a compact way of writing the rule of total probability. Evidently the condition of
probabilistic support as expressed in (6) can be written more succinctly in the new notation as

αn > ẞn , or equivalently as γn > 0.

So αn > ẞn together with equation (11) are what we mean by ‘En+1 probabilistically supports
En’ or En − En+1. To indicate this, we introduce the symbol
 En ← (αn>ẞn) En+1.

Here αn > ẞn expresses the requirement that γn = αn − ẞn is positive, the condition that En+1
probabilistically supports En. With our new symbol, the epistemic chain (5) becomes

E0 ← (α0>ẞ0) E1 ← (α1>ẞ1) E2 ← (α2>ẞ2) E3 . . . . . . ∞. (12)

In Sections 3 and 4 we will discuss the viability of (12) in more detail. Our overall aim is to
show that, if one takes seriously that justification comes in degrees, infinitism is less eccentric
than many people think.

3 Finite and Infinite Chains

A standard way of envisaging an infinite chain is first to look at a chain consisting of s + 1 links,
where s is a finite number. So let us consider the finite chain

E0 ← (α0>ẞ0) E1 ← (α1>ẞ1) E2 ← (α2>ẞ2) E3 . . . ← (αs>ẞs) Es+1.

From the rule of total probability in the form of equation (11), we find

with n = 0, P(E0) = ẞ0 + γ0P(E1)

with n = 1, P(E1) = ẞ1 + γ1P(E2),

and on combining these equations, we obtain

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1P(E2).

The iteration can be continued up to n = s, with the result

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 +· · ·+ γ0γ1 . . . γs−1ẞs + γ0γ1 . . . γsP(Es+1). (13)

For a foundationalist, Es+1 would be the basic proposition that forms the starting point for the
entire chain. Often that starting point has been seen as the report of a sensation or an impression.
Traditional foundationalists such as C. I. Lewis hold that Es+1 is absolutely certain, so that the
probability of this proposition equals unity: P(Es+1) = 1. Many contemporary foundationalists,
on the other hand, are what Moser et al. have called modest, and what Bonjour has called
moderate foundationalists (Moser et al. [14], p. 87; Bonjour [2], p. 26). Like their philosophical
forbears, they insist that epistemic chains must be rooted in basic propositions, but unlike them
they hold that a basic proposition need not have probability one. It is enough if it has a definite
(known) probability greater than some threshold of acceptance, say P(Es+1) > τ. We might think
here of the infamous Gedankenexperiment of Schrödinger, in which P(E0) could be the
probability that the cat is alive at time t1, P(E1) the probability that the hammer breaks the vial
of hydrocyanic acid, P(E2) the probability that an alpha particle that leaves the radium actually
enters the Geiger counter window, and P(E3) the probability that at least one radium nucleus
decays between t0, the beginning of the grisly experiment, and t1. Here P(E3) is supposed to
have a value that exceeds the threshold τ, and it could be calculated from the known half-life of
radium and the known number of radium atoms in the sample.

Since we are interested in chains of infinite length, it proves profitable to consider further
examples, namely, ones in which the length of the chain, s, is regarded as a variable, instead of
being fixed, as it was in the case of the cat (there s = 2). The advantage of having s as a variable
is that we will be able to investigate what happens as s becomes very large and ultimately tends
to infinity. Below we shall in fact provide three examples with variable s. In the first example the
conditional probabilities are uniform, that is, αn == α, independently of n, and similarly ẞn == ẞ
and so γn == γ = α − ẞ. In the second example there is nonuniformity: _n and _n do depend
nontrivially on n, although n does not, while in the third αn, ẞn and γn all depend in an essential
way on n.

Here is the first example. Imagine colonies of a bacterium growing in a stable chemical
environment known to be favorable to a particular mutation of practical interest. The bacteria
reproduce asexually so that only one parent, the ‘mother’, produces ‘daughters’. The probability
that a mutated daughter descends from a normal, not mutated, mother is known to be very small
(say, 0.02), but the probability that a mutated daughter descends from a mutated mother is, on the
other hand, high (say, 0.99). Let En be the proposition: ‘the ancestor in generation n, reckoned
backward from the present, was a mutant’. In this case the conditional probabilities are αn =
P(En|En+1) = 0.99, ẞn = P(En|¬En+1) = 0.02 and so γn = αn − ẞn = 0.97. We are told further
that each batch develops from a single, mutant ancestor. In this situation, in which all the ẞn are
equal to one another, and likewise all the n, equation (13) reduces to a finite geometric series that
can be summed explicitly:

P(E0) = ẞ[1 + γ + γ^2 + . . . γ^s] + γ^s+1P(Es+1) = ẞ[1 – γ^s+1/1 – γ] + γ^s+1P(Es+1).

(14)

Note that γ^s+1 is not the same as γs+1. The former means “γ raised to the power s + 1”, while
the latter is the Christensen-Joyce confirmation that Es+2 confers on Es+1—see equations (9)
and (10).

Now imagine a batch to be sampled after, shall we say, 150 generations. Then s = 149, and with
the above values of ẞ and γ, we obtain

P(E0) = 0.667[1 − (0.97)^150] + (0.97)^150P(E150) . (15)

The original great-great-grandmother, in generation 150 before the generation to be sampled, is

known to be a mutant, so P(E150) = 1, and therefore P(E0) is perfectly well defined: it works out
to be 0.670.

We will come back to the example of the bacteria in Section 4, but for the moment we may note
that P(E150) = 1 is not, as strong foundationalists would insist, a necessary requirement in order
that equation (15) determine P(E0). It may not be known if the original bacterium was mutated,
but it could have been obtained from a naturally occurring sample, in which the probability of
mutation had been determined by earlier observation of many samples. In this case P(E150)
would be set equal to this known probability: the situation formally resembles the Schrödinger
cat example, where the originating probability, associated with the decaying radium nucleus, was
known and unequal to unity.

A more extreme situation, and one that even more clearly flies in the face of strong
foundationalists like C. I. Lewis, is when P(E150) vanishes, which would be so if it were known
that the great-great-grandmother in the 150th generation was certainly not a mutant. It seems that
here a strong foundationalist has no choice but to deny any justification for the mutation in
generation 0. However, if P(E150) = 0, then P(E0) = 0.667[1 − (0.97)^150]= 0.660. Here is a
case that fits into the foundationalist’s requirement of linear finitude, but in which the ‘ultimate
ground’, in this case the mutation of the originating bacterium, is absent, although the probability
that is built up from the conditional probabilities is perfectly definite—and nonzero, of course.

Lest it be thought that the above considerations rest essentially on the requirement of uniformity,
let us briefly sketch our second example, that is, a generalization in which the conditional
probabilities are not the same from generation to generation. Suppose that an effect of an
increasing pollution of the nutrient, as a result of the growing mass of bacteria in it, is that the
probability of mutation of a daughter increases as time goes on, whether or not the mother is a
mutant. For example, if αn = P(En|En+1) = a +b^n+1 and ẞn = P(En|¬En+1) = b^n+1, where a
and b are positive numbers such that a+b < 1, then αn and ẞn are different from generation to
generation, although γn = a is constant. Equation (13) once more reduces to a finite geometric
series that can be summed:

P(E0) = b[1 + ab + (ab)^2 + · · · (ab)s] + a^s+1P(Es+1) =

b[1 − (ab)^s+1/1 – ab] + a^s+1P(Es+1). (16)

A more sophisticated abstract example, our third and last, is one in which n, as well as _n and _n,
depend nontrivially on n, and, moreover, in which the resulting infinite series is not geometric:

ẞn = 1/n+3 γn = n + 1/n+2 = 1 – 1/n+2, (17)

so that

αn = ẞn + γn = 1 – 1/n+2 + 1/n+3 .

Then

γ0γ1 . . . γs = 1/2 × 2/3 × · · · × s/s+1 × s+1/s+2 = 1/s+2

γ0γ1 . . . γs−1ẞs = 1/s+1 × 1/s+3 = 1/2(1/s+1 – 1/s+3),

so equation (13) reduces to

P(E0) = 1/3 + 1/2(1/2 – 1/4) + 1/2(1/3 – 1/5) + · · · 1/2(1/s+1 – 1/s+3 ) + 1/s+2P(Es+1) =

1/3 + 1/2(1/2 + 1/3 – 1/s+2 – 1/s+3) + 1/s+2 P(Es+1) . (18)
Like the geometric series, this series can thus be summed, and the result written down in explicit
form. In the next section we will investigate what happens to our three examples when the
number of terms, that is, the number of links in the epistemic chain, becomes infinite.

4 An Infinity of Conditional Probabilities Can Yield a Determinate Probability

During the past several years, Klein has repeatedly stressed that epistemic justification has the
character of infinite chains such as (5) or (12). He has defended infinitism as the one true faith
against attacks of foundationalists, who have traditionally argued that infinite chains do not make
sense. As Klein sees it, there is nothing problematic about an infinite linear chain, the reason
being that such a chain need not be completed. What is more, the requirement that an infinite
chain must be completed “would be tantamount to rejecting infinitism” ([10], p. 920). The only
thing that an infinitist à la Klein requires is that for every proposition En in the chain, there is a
proposition En+1 such that the conditional probability P(En|En+1) is known to be greater than
P(En|¬En+1), for that is the condition under which we can properly say that En+1
probabilistically justifies En.

We fully agree with Klein that an infinite linear chain need not trouble us, but our reasons are
somewhat different from his. For if such a chain is approached by way of a finite chain in the
limit as s tends to infinity, then we are able to show that the chain can be completed. In the
present section we first give an example of a probabilistic chain that, although infinite, can be
completed in the sense that its value can be computed. This example not only shows that infinite
chains of reasons can exist (contra Black [1]), but it also illustrates that such chains might
culminate in an unconditional justification that yields a number other than zero (contra Lewis
[13] and [12]). Last but not least, the example shows that an infinite regress of deferred
conditional justification can result in a determinate, unconditional justification for the target
proposition (contra Dancy [5] and Gillet [8]).

Consider again the bacterial colonies of Section 3, and particularly the first line of equation (14),
with its remainder term, the product γ^s+1P(Es+1). As s becomes larger, the number of terms
within the square braces increases, while at the same time γ^s+1 becomes smaller (recall that γ is
strictly smaller than one, and therefore that γ^s+1, the (s+1)st power of γ, is even smaller). In the
limit that s is taken to infinity, γ^s+1 disappears. Since γ^s+1 multiplies P(Es+1), which is a
probability and so can never be greater than one, we see that the remainder term vanishes in the
limit that s is taken to infinity. In this same limit the number of terms within the square braces
has become infinite, and so we obtain

P(E0) = ẞ(1 + γ + γ^2 + . . . ) . (19)

Although equation (19) is a geometric series with an infinite number of terms, it can nevertheless
be completed, in the sense that it can be computed exactly. An explicit form for its sum can be
read off from the second line of equation (14), if we use the fact that, in the infinite limit, γ^s+1
disappears. It is

P(E0) = ẞ/1 − γ. (20)

If we now take, as in the bacterium case, α = 0.99 and ẞ = 0.02, then γ = 0.97 and P(E0) = 2/3 .
In other words, even though the value of P(E0) is based on an infinite number of terms, we are
still able to calculate it (cf. Peijnenburg[15]).3 This suffices to show what happens when the
number of terms in the first example of Section 3 becomes infinite. The second example (16)
similarly has a well-defined limit as s tends to infinity, namely, b/(1 − ab).

At this juncture, a foundationalist objecting to infinite chains might argue that our story about the
bacterial colonies is not an example of infinitism at all. For no bacterium has an infinite number
of ancestor bacteria, if only because of the fact of evolution from more primitive algal slime,
which had evolved from earlier life forms, which sprang from inanimate matter, which
originated in a supernova explosion, and so on, back to . . . to what? To the Big Bang? But it
seems that the Big Bang may well not represent a beginning, in view of the deformation of
spacetime. The whole point here is precisely the question whether or not there was a starting
point. The foundationalist’s postulate that in the bacterial case there was a start begs the
question.

The fact that we can compute a probability from an infinite chain might be good news for the
infinitist, who may now relax the requirement that infinite chains cannot be completed. But could
not the same fact serve as grist to a foundationalist mill? After all, it seems that a foundationalist
might argue that, since the infinite chain converges to a number that we can compute, an ultimate
ground must exist. Could not the foundationalist claim that it is the limit itself that constitutes the
foundation? The answer is in the negative. For the limit of what exactly is supposed to provide
the ground? It cannot be the limit of the remainder term, γ0γ1 . . . γsP(Es+1), for as we have seen
this term goes to zero as s goes to infinity, and something that fades away into nothingness can
scarcely support an edifice. What about the limit of the entire infinite sequence? This cannot
serve the foundationalist purpose either. The limit of the sequence is the sum of an infinite
number of terms—in our example it is 2/3, being the value of P(E0). This value is, however, the
result of the concatenation of an unending sequence of terms describing probabilistic support; it
is not the initiator of it, for the purported initiator was precisely the remainder term γ0γ1 . . .
γsP(Es+1).4

The above considerations are not restricted to cases in which the infinite series is geometric, as it
was in our first two examples. They also apply to our third example. In the model (17), if one lets
s tend to infinity in equation (18), the terms involving 1/s+2 and 1/s+3 vanish, including of
course the last term, which carries the probability P(Es+1) as a factor. The result is P(E0) = 1/3 +
1/2(1/2 + 1/3) = 3/4. Here then is a case in which there is no uniformity, but in which the
unconditional probability of the target proposition can be calculated from an infinite, convergent
series of conditional probabilities.

Once you take seriously that justification admits degrees, then the bullet must be bitten: in
general there is no difficulty with infinite series of probabilistic support. The reason is that the
remainder term usually vanishes in the limit that s is taken to infinity. Our point is precisely that
a perfectly definite probability can be built up from an infinite number of conditional
probabilities, without any need, or room for an initiating proposition.

5 Afterword: The Bucket Brigade

Our conclusion that an infinite series of conditional probabilities can yield a definite,
unconditional probability value for the target proposition might seem counterintuitive, to say the
least. How can we ever justify a proposition if its justification is forever postponed?

The apparent oddity of our result can be illustrated by the saga of the bucket brigade.5 Suppose
there is a fire and Abby has to get her water from Boris, and Boris has to get it from Chris, and
Chris from Dan, and so on ad infinitum. It would seem that the fire will never be put out, since
there is no first member of the brigade who actually dips his bucket into the lake.

However, this problem would only arise if we drop the assumption that justification comes in
degrees. If instead we retain the assumption, and hold that justification is probabilistic, then the
matter is entirely different. Under this assumption, the proposition ‘Abby gets water’, A, is only
partially justified. In fact, we can calculate the exact probability value of A by applying the rule
of total probability that we cited earlier:

P(A) = P(A|B)P(B) + P(A|¬B)P(¬B) , (21)

where B reads ‘Boris gets water’. Of course, whether Boris gets water is also merely probable,
and its probability depends on whether Chris gets water, and so on. We face here an infinite
nesting of probability values calculated via the rule of total probability. As we have seen, we are
perfectly able to compute the outcome of this infinite series in a finite time: with the numbers
that we used in the bacterium example, the probability of A is 2/3. Note that this is completely
independent of whether we embrace an objective interpretation of probability (assuming, for
example, that the firefighters have propensities for handing over the water only now and then) or
a subjective interpretation (in which we specify our degree of belief in A). Both the objective and
the subjective interpretation are bound by the rule of total probability, and that is the thing that
counts here.

All four probabilities on the right-hand side of equation (21), the conditional as well as the
unconditional ones, are supposed to have values strictly between zero and one (in the interesting
cases). Now the traditional view, in which justification is not gradual, can be modeled by
restricting all four ‘probabilities’ to be 0 or 1. Within this nonprobabilistic approach, Abby either
gets water or she does not get water. The foundationalist moral of the saga about the bucket
brigade is precisely that she does not get water—if the number of brigadiers is infinite. Because
this is unacceptable, the foundationalist thinks that there must be a first firefighter who starts off
the whole rescue operation.

In the probabilistic scenario the existence of a primordial firefighter is not needed, since the
problem that it is supposed to solve does not arise in the first place. If we take seriously that
epistemic support comes in degrees, then the probability that Abby extinguishes the fire can have
a precise value that we are able to calculate, despite the infinite number of her teammates. As in
the examples that we considered above, this unconditional value is a function of all the
conditional probabilities.

Notes
1. Recently, Turri has argued that not only infinitists, but foundationalists, too, can make sense of
infinite chains of reasons (Turri, J., “On the Regress Argument for Infinitism,” forthcoming in
Synthese). In our view Turri’s argument is flawed. We will not go into the details here, but see
endnote 4.

2. Note that interpreting justification as probabilistic support is itself neutral with respect to the
normative-empirical debate. Defenders of the view that justification is normative will be inclined
to a subjective view of probability. Their more empirically-minded counterparts will tend to an
objective and even frequentistic interpretation. But both Atkinson and Peijnenburg factions are
bound by the Kolmogorovian axioms that underlie subjective and objective interpretations of
probability.

3. Equation (20) is the fixed point of a Markov process. Indeed, the stochastic matrix governing
this process is regular, and so the infinite iteration of

P(En) = ẞ + γP(En+1),

is guaranteed by Markov theory to converge to P* = ẞ + γP* = ẞ/(1−γ). This quick route to the
answer only works in examples employing uniform α and ẞ; in the general case Markov theory
does not help.

4. This is one of the reasons why Turri’s argument fails (see endnote 1).

5. We thank an anonymous referee for suggesting this illustration.

An Infinitist Account of Doxastic Justification

John Turri†
† Philosophy Department, Huron University College, 1349 Western Road, London,

Abstract

Any satisfactory epistemology must account for the distinction between propositional and
doxastic justification. Can infinitism account for it? Proposals to date have been unsatisfactory.
This paper advances a new infinitist account of the distinction. The discussion proceeds as
follows. Section 1 sets the stage. Section 2 presents Peter Klein’s account. Section 3 raises a
problem for Klein’s account and suggests an improvement. Section 4 raises a further challenge.
Sections 5 to 7 consider several unsuccessful attempts to meet the challenge. Section 8 presents
my new proposal, which can meet the challenge. Section 9 concludes the discussion.
1. Preliminaries

Epistemologists standardly distinguish propositional from doxastic justification. Some

proposition might be justified for you even though you don’t actually believe it, or believe it for
the wrong reasons, or believe it for the right reasons but in the wrong way. Call this type of
justification propositional justification. It attaches to propositions relative to individuals.
Doxastic justification attaches to concrete belief states (and to doxastic states more generally).
For your belief that Q to be doxastically justified, Q must not only be justified for you, but you
must believe Q for the right reasons and in the right way. Knowledge requires doxastic, not
merely propositional, justification.1

An adequate theory of justification must account for this distinction. Can infinitists account for
it?

Infinitism has been around since at least Aristotle’s time. But compared to its main non-skeptical
competitors, foundationalism and coherentism, its resources remain largely underdeveloped and
its potential benefits unappreciated. Through a series of papers over the last decade, Peter Klein
has done more than anybody to revive infinitism’s fortunes, thrusting it back onto the
philosophical scene (e.g. Klein 1999, 2005a, 2007a).2 So let’s begin with his account of the
distinction.

2. Klein’s account

We can encapsulate Klein’s account as follows. Here and throughout, understand each reason to
be unique, so that no two Rs with distinct subscripts are identical.

Infinitist propositional justification (IPJ): The proposition Q is propositionally justified

for you just in case there is available to you at least one infinite nonrepeating series of
propositions (or reasons) such that R1 is a good (and undefeated) reason to believe Q, R2
is a good (and undefeated) reason to believe R1, R3 is a good (and undefeated) reason to
believe R2, . . . ,Rm+1 is a good (and undefeated) reason to believe Rm, for any
arbitrarily high m. (Klein 2005a, 135–136; Klein 2007a, 8, 11; compare Klein 2005b,
166)

Infinitist doxastic justification (IDJ):Your belief that Q is doxastically justified just in

case Q is propositionally justified for you, and you have provided enough reasons along
at least one of the infinite non-repeating series of reasons, in virtue of which Q is
propositionally justified for you, to satisfy the contextually determined standards.3

A few more words about Klein’s positive view are in order. First, knowledge requires doxastic
justification (Klein 2007b, 158). This is uncontroversial. Second, doxastic justification requires
you actually to go through the process of justifying your belief. It must be earned, much as an
income or honest reputation is earned – at least for most of us (Klein 2005b, 158, 163, 170). This
is controversial but I will not question it here (see Leite 2004 for discussion). Third, contextual
standards determine what is admissible as a “bedrock” reason – a reason that, once you reach it,
permits you to stop (Klein 2007a, 10–12; Klein 2005b, 170–171). Crucially for the infinitist, it is
always possible to challenge properly ad infinitum the contextual standards and thereby the
erstwhile bedrock reason, although this never actually happens. Contextualism is very
controversial, but I will not question it here because I want to see whether Klein’s view granted
in its entirety can adequately account for the distinction between propositional and doxastic
justification.

3. An improvement

I will now suggest an improvement to IDJ. My motivation is not that we can dream up fanciful
counterexamples – as though that would be surprising. Rather, IDJ seems to omit something
fundamentally important.

Meet Samira. Samira knows that Q. Under our assumptions it follows that Samira’s belief is
doxastically justified, so Samira has produced enough reasons along at least one infinite non-
repeating series to satisfy the contextually determined standards. For simplicity suppose that Q is
propositionally justified for Samira in virtue of only one infinite non-repeating series of reasons,
and for convenience call this series ‘Z’. (Nothing that follows depends essentially on this
assumption; it merely simplifies greatly my presentation.) Again for simplicity suppose further
that the contextually determined standards require Samira to produce the first three reasons along
Z (again, nothing depends essentially on this assumption), which are:

R1, R2, R3.

So Samira has articulated the following justificatory structure:

R3→R2→R1→Q.

The arrow represents the ‘is a reason offered in favour of’ relation.

But here an issue arises. Given that Samira has produced R1, R2 and R3, Z is not the only
possible path she might be on, because Z is not the only path beginning with R1, R2, R3. An
infinite number of such paths exists. For Samira’s belief to be doxastically justified, it is not
enough for her to be on just any path beginning that way. If she were, it would be pure luck that
she produced the correct path’s first three steps. This sort of luck is perfectly compatible with
propositional justification, but obviously not with doxastic justification.

We could put the point this way. Simply producing R1, R2, R3 is insufficient. Samira could
accomplish that by lucky random guessing. She must properly produce the reasons.4
Accordingly we should amend IDJ.

(IDJ*): Your belief that Q is doxastically justified just in case Q is propositionally

justified for you, and you have properly provided enough reasons along at least one of the
infinite non-repeating series of reasons, in virtue of which Q is propositionally justified
for you, to satisfy the contextually determined standards.
So is Samira on the right path? Of course she is. By hypothesis she knows Q. And she would
know Q only if she were on the right path.

4. A question

But a further question arises. In virtue of what is Samira on the right path? Why is she on Z
rather than, say, on defective path A, whose first three steps are also R1, R2, R3?

Z: R1, R2, R3, R4Z, R5Z . . .

A: R1, R2, R3, R4A, R5A . . .

For the infinitist’s account to succeed, there of course must be a fact of the matter whether
Samira is on Z, given that she knows Q. For if there is no fact of the matter whether she is on Z,
then it is not true that she is on Z, whence it is not true that her belief is doxastically justified,
whence it is not true that she knows Q.5

The next three sections consider some proposals for answering our latest question.

5. Counterfactuals

a.

A tempting initial response is to go counterfactual: Samira is on Z because having reached R3, if

she were to consider the question ‘Why accept that?’, then she would offer R4Z rather than R4A.

But this does not solve the underlying problem. For there is also:

A*: R1, R2, R3, R4Z, R5A . . .

No matter the number of steps Samira takes, there will always be an infinite number of defective
paths that share with Z those steps, but then veer off into an epistemic dead-end.

b.

Consider next this proposal. Samira is on Z because for any step, RnZ, along Z, if Samira were to
reach RnZ and consider the question ‘Why accept that?’, then she would offer RnZ+1.

Any proposal featuring counterfactuals faces tricky counterexamples. Imagine a case for which
we stipulate: if Samira were to reach R40Z and was asked ‘Why accept that?’, then some
powerful agency would intervene and prevent her from providing R41Z. Or maybe at some point
Samira would have a nervous breakdown or get bored and just stop offering further reasons.
Indeed Samira is guaranteed to fall short of articulating the entire series because by stipulation:
she is mortal, the number of steps in Z is infinite, providing each reason takes some time, and
super-tasking is not an option.6 Such problems surely merit attention. But I think a more
alarming problem suggests itself, one that cannot be dismissed as a mere trick.
To illustrate this deeper worry, it will be convenient to stipulate that Q is the proposition Samira
would express by saying ‘I will never justify a belief beyond the 100th step’. Surely any normal
human could know this about him or herself. (If you’re doubtful, please adjust the example by
increasing it to whatever finite number step you wish.) At some point along the way – say the
1,000th step – Samira would have acquired such overwhelming evidence that Q is false that she
would simply stop producing further reasons in favour of Q because she would have stopped
believing Q.

This case is more alarming than your average tricky counterexample to a counterfactual analysis.
It is more alarming because Samira would rightly stop providing further reasons. She would have
acquired conclusive evidence that the belief she is ultimately defending is, in the counterfactual
situation, false.

The example is not an isolated fluke. We could plug in different clauses for ‘Q’ and create the
same problem. The truth-value of the proposition need only be sensitive to some threshold – be it
the passage of time or a limit on the repetitions of some performance – that the subject would
flagrantly violate if he continued indefinitely providing reasons.

For example consider the proposition I would express by saying ‘I will not live past my 200th
birthday’. Surely I know that. Yet since offering each reason takes at least some time, and super-
tasking is not an option, and any justifying series contains infinite steps, at some point I would
stop offering further reasons because I would recognize that I was older than 200 years. And
even if I did continue offering reasons, at some point it would become irrational because I would
obviously be older than 200. Imagine me at my 201st birthday party, still justifying my belief that
I will not live past my 200th birthday, even as I complain about how difficult it is to blow out all
201 candles and reminisce about how lucky I am to have reached the ripe old age of 201!

c.

Consider this variant of the counterfactual proposal. Samira is on Z because for any step, RnZ,
along Z, if she were to reach RnZ and consider the question ‘Why accept that?’ and offer a
reason in response, then she would offer RnZ+1.7 This proposal suffers the same fate. At some
point it would be irrational for Samira to offer any further reason in the service of justifying the
obviously false target belief that Q. For it would be painfully obvious to her that she long, long
ago transcended the 100-step threshold, many times over. And surely it would be surprising if a
patently irrational counterfactual performance could generate knowledge or doxastic
justification.

d.

Consider one final revision of the counterfactual strategy. Samira is on Z because for any step,
RnZ, along Z, if she were to reach RnZ and consider the question ‘Why accept that?’ and
appropriately offer a reason in response, then she would offer RnZ+1.
Notice two things about this proposal. First, it is puzzling because it is difficult to imagine
appropriately offering the 1,000th, 10,000th or 100,000,000,000th reason to support the claim
that you will never surpass the 100-step threshold. Samira would somehow have to be deprived
of virtually all the relevant evidence she had gathered. And it would be surprising if what Samira
would do if deprived of nearly all the relevant evidence somehow doxastically justified her actual
belief. Second, the revision is potentially trivial. In a straightforward sense it would be
inappropriate to offer any reason other than the correct next one in the series. For this proposal to
convince, then, the relevant sense of appropriateness would have to be explained.

6. Multiple paths

Infinitists might jettison the misbegotten counterfactual strategy and instead respond as follows.

Return to the point at which the problem originally seemed to arise. We observed that Samira’s
producing R1, R2, R3 is consistent with her being on either Z or A. We stipulated that she knows
Q, which requires her to be on Z, which in turn motivated us to ask what makes it the case that
she is on Z rather than A. But this overlooks an important possibility: perhaps she is on both
paths! And as long as she is on Z – regardless of whether she is also on A – her belief is
doxastically justified. Simply producing those three reasons makes it the case that she is on both
paths and so explains why her belief is doxastically justified.8

Grant that simply producing those three reasons makes it the case that she is on both paths. The
proposal is still inadequate. Simply being on Z in the way envisioned does not suffice for
doxastic justification. It fails to rule out even the most egregious form of epistemic luck:
randomly guessing R1, R2, R3 would suffice! The imagined response essentially reverts back to
IDJ, ignoring the lesson that drove us to adopt IDJ*. This type of luck is perfectly consistent with
propositional justification – guessing does not deprive you of the reasons or stop them from
propositionally justifying Q for you. But guessing obviously does not suffice for doxastic
justification. Samira must not only be on the right path. She must be on the right path in the right
way.

Abandoning the “path” metaphor, the point is that she must properly produce the relevant
reasons. The present proposal offers no insight into that important status, crucial to the
distinction between propositional and doxastic justification.

7. Tu quoque?

But is the infinitist being unfairly singled out for criticism here?9 Does anyone have a
satisfactory account of doxastic justification? You might suspect that infinitist and finitist alike
face the same problem, in which case we have located no special problem for infinitism.

My response is twofold. First, if the suspicion is correct, then the present discussion reveals an
urgent problem facing all theories of justification, considerably enhancing its importance. In a
certain sense, that would please me. Nevertheless, second, the suspicion is not correct.
Infinitism’s competitors face no such problem in principle.
To substantiate this last point, consider a simple foundationalist view, what I’ll call simple
empiricism. (I don’t here endorse simple empiricism, but rather deploy it to conveniently
illustrate a point.) Simple empiricism says (i) Q is propositionally justified for you just in case
you have an experience that Q; (ii) your belief that Q is doxastically justified just in case it is
properly based on an experience that Q; and (iii) your belief that Q is properly based on an
experience that Q just in case the experience’s causing your belief manifests your cognitive
disposition to take experience at face value.10 Now suppose that Samira has an experience that Q.
This renders Q propositionally justified for her. She knows Q, so her belief must be doxastically
justified, so she must be on the right path in the right way. According to simple empiricism, in
virtue of what is she on the right path in the right way? In virtue of the fact that her experience’s
causing her belief manifests her reliable cognitive disposition to take experience at face value.

Without ruining the basic point, the simple empiricist may add that for your belief to be properly
based on the experience, you must also cite the experience or verbally express your acceptance
of the experience as a good reason to accept Q. This is in keeping with Klein’s view that you
must do something to earn doxastic justification.

Consider also simple coherentism. (Again I do not here endorse simple coherentism.) Simple
coherentism says that (i) Q is propositionally justified for you just in case Q coheres with (some
proper subset of) your beliefs B1 . . . Bn, where n is finite; (ii) your belief that Q is doxastically
justified just in case it is properly based on B1 . . . Bn; and (iii) your belief that Q is properly
based on B1 . . . Bn just in case B1 . . . Bn cause you to believe Q through a manifestation of
your cognitive disposition towards coherence. Now suppose that Samira holds beliefs B1 . . . Bn.
This renders Q propositionally justified for her. She knows Q, so her belief must be doxastically
justified, so she must be on the right path in the right way. According to simple coherentism, in
virtue of what is she on the right path in the right way? In virtue of the fact that her beliefs
B1 . . . Bn caused her to believe Q through a manifestation of her cognitive disposition towards
coherence.11

Again without ruining the basic point, we may add that for your belief to be properly based on
the other beliefs, you must also cite them or verbally express your acceptance of them as a good
reason to accept Q.

8. Mimicry

Perhaps the infinitist can say that Samira’s belief is doxastically justified in virtue of the fact that
her production of R1, R2, R3 manifests her relevant cognitive dispositions.

The main task facing this proposal is to identify the relevant dispositions. We know what it is to
be disposed to take experience at face value – we do it all the time. We know what it is to trust
coherence – we do that every time we accept some conclusion because it “fits” with what we
already believe. What dispositions will the infinitist invoke?

But this task isn’t so difficult after all. The infinitist can help himself to whatever dispositions his
opponents help themselves to.12 Or he might “go contextualist” about the relevant dispositions:
along with contextual standards determining what is admissible as a bedrock reason, contextual
standards also determine what counts as an admissible cognitive disposition. Either way, a
penchant to just guess will almost certainly not show up on the list of relevant dispositions.

9. Conclusion

Where does this leave us? We asked, In virtue of what is Samira on the right path in the right
way?, or less metaphorically, In virtue of what has Samira properly produced the relevant
reasons? The infinitist can respond: in virtue of the fact that Samira’s production of R1, R2, R3
actually manifests her relevant cognitive dispositions, whatever those dispositions may be.

This proposal is not ad hoc. It does not violate infinitism’s other commitments. Invoking the
manifestation of cognitive traits is equally available to foundationalist, coherentist and infinitist
alike. At the very least, the infinitist appears to have just as plausible a story to tell as her
competitors. True, we might want a more complete story about what the manifestation of
dispositions requires. But infinitism suffers no relative disadvantage as a result. And it is
something we would have wanted in any case because the manifestation of dispositions is
pervasive throughout the physical and social world.

I conclude that we have identified a plausible infinitist account of doxastic justification,

including what distinguishes it from propositional justification. Of course this does not answer
all questions about infinitism. There remain serious questions about the main positive arguments
for infinitism.13 There also remain questions about the contextualist element of IDJ*. But those
are issues for another day. For now, infinitists can take heart in our present results.*

1 See Turri (forthcoming) for further discussion.

2 Fantl (2003) gives an importantly different argument for infinitism.

3 Says Klein (2007a, 10): “The infinitist will take the belief that p to be doxastically justified for
S just in case S has engaged in providing ‘enough’ reasons along an endless path of reasons.”
Notice that Klein says that providing the reasons is both necessary and sufficient for doxastic
justification. In response to Bergmann (2007), Klein (2007b, 26) indicates he might be willing to
add that doxastic justification requires S’s belief to be “based on” the justifying reasons. But he
also suggests that basing is tantamount to there being “an available reason” that you “cite . . . as a
reason” for your belief, so it remains unclear how this potential revision affects his theory.

4 Klein (2007, 6) might have this in mind when he says that doxastic justification requires
believing in “an epistemically responsible manner.” He doesn’t specify what believing
responsibly requires. My subsequent discussion can be read as an examination of what form an
infinitist account of this concept might take.

5 Philipp Keller questioned whether the reasoning here was careful enough. More fully spelled
out, here is why there must be a fact of the matter, given that Samira knows Q:

1. Samira knows that Q only if her belief is doxastically justified. (Premise)

 2. Samira’s belief is doxastically justified only if she is on path Z. (Premise)

3. She is on path Z only if it is true that she is on path Z. (Premise)

4. It is true that she is on path Z only if there is a fact of the matter whether she is on path
Z. (Premise)

5. Therefore she knows that Q only if there is a fact of the matter whether she is on path
Z. (From 1–4, Hypothetical Syllogism)

6 It is not an option for us humans as we are actually constituted, which is enough for my
purposes. I concede that there are possible beings for whom super-tasking is an option.

7 Peter Klein suggested this possibility in conversation. I do not know if he positively endorses
it.

8 Thanks to an anonymous referee for suggesting this.

9 Thanks to an anonymous referee for raising this question.

10 Turri (unpublished ms a) defends a related account of the epistemic basing relation.

11 An anonymous referee questioned whether the challenge I pose for infinitism stems from the
fact that infinitism disallows non-beliefs to play the role of reasons. The discussion of simple
coherentism puts this worry to rest because it disallows non-beliefs to play that role.

12 Klein (2005a, 136–137) makes a parallel suggestion regarding an infinitist account of what
makes something a (good) reason for you to begin with.

13 For discussion of Klein’s “regress argument” see Turri (2009); for discussion of Fantl’s
“features argument” see Turri (unpublished ms b).

* For superb feedback that helped improve this paper, I thank the editors of this journal
(along with any anonymous referees who helped them in the process). Special thanks also to
Peter Klein for helpful discussion early on.
How to be an infinitist about doxastic justification
Peter Klein
Philos Stud (2007) 134:25–29
Philosophy Department, Rutgers University

Michael Bergmann’s excellent comments contain some important insights which require some
clarifications and further refinements of my view. Nevertheless, I don’t believe that the three
issues he raises pose any fundamental difficulties for my defense of infinitism. Those three
purported difficulties are:

1. The proposed account isn’t really an infinitist account of doxastic justification. Indeed,
he says that my account is a form of foundationalism with unjustified propositions at the
base.

2. My account faces a troubling dilemma.

3. My objection to foundationalism fails.

Given Bergmann’s skepticism about the importance of propositional justification, before turning
to a discussion of the three objections, I want to make clear why I began with a discussion of
propositional justification and why that type of justification is interesting. (I will also return to it
in the concluding comments.) Recall that I said that a proposition, h, is propositionally justified
for S just in case there is an epistemically adequate basis for h that is available to S; and that a
belief that h is doxastically justified for S when and only when S is acting in an epistemically
responsible manner in believing that h. I said that:

Doxastic justification is parasitic on propositional justification. We have said that a belief

that p is doxastically justified for S iff S is acting in an epistemically responsible way in
believing that p. For a self-conscious epistemic agent—an agent who practices what she
preaches—what constitutes a responsibly held epistemic belief will depend upon what
that agent thinks is required for a proposition to be justified for her.

Hence, to understand the three concepts of doxastic justification employed by the

foundationalist, coherentist and infinitist we must first delineate the respective accounts of
propositional justification. For example, since the foundationalist holds that propositional
justification arises autonomously in basic propositions and is transferred by appropriate truth
conferring implication rules to non-basic propositions, a self conscious foundationalist would
seek to trace her beliefs back to an autonomously justified proposition by reconstructing the
inferences in reverse, so to speak. On the other hand, the infinitist does not think of propositional
justification as a property that is transferred from one proposition to another by such inference
rules. Rather, the infinitist, like the coherentist, takes propositional justification to be what I
called an emergent property that arises in sets of propositions. In particular, the infinitist holds
that propositional justification arises in sets of propositions with an infinite and non-repeating
structure such that each new member serves as a reason for the preceding one. Consequently, an
infinitist would seek to increase the doxastic justification for an initial belief—the belief
requiring reasons—by calling forth more and more reasons. The more imbedded the initial
belief, the greater its doxastic justification.

Bergmann introduces the discussion of his three objections by noting that I say that what ‘‘lies at
the heart of the infinitist’s requirement for doxastic justification’’ is not what causes our beliefs
but rather that we can cite a reason for our beliefs. That reason, of course, must be available or
we couldn’t cite it. What I meant, and perhaps didn’t say clearly, was that infinitism holds that a
particular belief is doxastically justified (at least to some degree) only if there is an available
reason and we cite that reason as a reason for our belief. Whether the belief with the reason as its
content is the cause of the belief is not central to infinitism, just as I claimed that foundationalists
need not be committed to the claim that reasons for beliefs are their causes. I said that because I
simply don’t know—nor do I think anyone else knows yet—whether doxastically justified
beliefs are caused by the beliefs with the reasons as their contents. So, since I thought
foundationalism had better not be wedded to the causal claim, I suggested that it should not be
considered to be central to infinitism either. Now Bergmann thinks that any plausible account of
the basing relationship will include a causal requirement.

But nothing significant depends upon remaining uncommitted about whether the beliefs with
reasons as their propositional contents are causes of beliefs. Perhaps they are. The important
issue is that infinitists hold that knowledge that p obtains only when our belief that p is
doxastically justified and in order for a belief to be doxastically justified it must be ‘‘based’’ (to
use Bergmann’s term) upon a belief whose propositional content is a reason, and that reason
must, of course, be available. Now, if basing is best understood as including a causal component,
so be it (for the sake of the discussion here).

Recall the capital-of-Montana example in my paper. The point was that infinitism could (but
need not) have a rather liberal view of what counts as an available reason. I suggested that the
infinitist could adopt a liberal notion of availability so that having easy access to an ordinary
encyclopedia and having epistemic practices that would result in consulting the encyclopedia
make its contents available to S as a reason. But the belief is not doxastically justified until the
encyclopedia is consulted and the reason is employed by S to render her belief doxastically
justified. The crucial point was that although propositional justification requires that there be an
infinite path of non-repeating reasons, in order for a belief to be (at least partially) doxastically
justified, it is not required that S possess that infinite set of reasons or that a belief be based upon
beliefs that have the infinite set of reasons as their propositional contents. However, it is required
that some of those reasons be available and that the belief be based upon those beliefs that have
the available reasons as their contents. In other words, by distinguishing between propositional
and doxastic justification the so-called ‘‘finite mind problem’’ would disappear.

One more comment before turning to the three main objections. Bergmann’s three objections
depend upon his view that my development and defense of infinitism contains, at least implicitly,
a denial of the following principle:

K2. A belief can be doxastically justified by being based upon some other belief only if that
other belief is itself doxastically justified.

So, just to be clear at the outset, let me say that I do think K2 is false, but I will suggest that a
closely related principle is true.

Is my view of doxastic justification an infinitist one?

Bergmann thinks mine is not an infinitist view because it denies K2. In other words, I hold that
our beliefs can be (at least partially) doxastically justified even if the beliefs from which they are
inferred are unjustified. Would some foundationalists (namely those who hold that the
foundational beliefs are arbitrary or unjustified) also deny K2? Yes. Is that sufficient to make my
view of doxastic justification a foundationalist one? No.

My view is not an arbitrary foundationalist view of doxastic justification because such a

foundationalist would also hold that with regard to some of the as-yet-unjustified beliefs (the so-
called basic beliefs) either they are such that a responsible epistemic agent need not or, perhaps
cannot, produce a reason for them. Whereas the infinitist would hold (1) that once any belief is
called into question by contextual requirements, a responsible epistemic agent would then have
to seek a reason for it and (2) that for any belief there are such contextual requirements and, most
importantly, that (3) once the as-yet-unjustified belief is called into question, there had better be
a further proposition that has not yet appeared in the reasons given, for if there weren’t, the
doxastic justification that the original belief had acquired through the process of reasoning would
be lost. (See my concluding comments for an important clarification.) One swallow does not a
summer make—and one point of agreement with a type of foundationalist does make one that
type of foundationalist.

But what about K2 itself? Bergmann gives no argument for it—probably because he thinks it is
so obviously true. K2 will strike those still under the spell of the foundationalist’s account of
propositional, and, hence, doxastic justification as obviously true. Foundationalists would
endorse the K2 principle because they think that if a belief is justified by another belief the latter
must be justified because they hold that justification is a property that is transmitted by inference
to non-basic propositions. Indeed, what Bergmann calls an ‘‘unjustified foundations view’’
strikes me as a non-starter since such a view has no way to account for the justification of non-
basic beliefs because there is no justification to be transmitted from the basic ones. On the other
hand, an infinitist can reject K2 because, just as propositional justification is not a property that
is transmitted from one proposition to another by truth preserving rules of inference, doxastic
justification in not inherited through the ‘‘basing’’ relation. A belief is doxastically justified (at
least to some degree) if we have located a good reason for holding it. And that is the case, even if
we have not yet located a good reason for holding the reason. But if we have a second reason for
the first reason, the first reason is now (at least partially) doxastically justified and the doxastic
justification of the original belief has further increased.

It’s like buying a car on payments. I am making progress in purchasing the car if I make a
payment. I am making more progress if I make more payments. I had made that progress even if
I run out of money and can’t make the next payment. But, if there is no money available to make
the next payment, I am in danger of losing all the value invested in the car.1

Just so with beliefs. I am making progress in increasing the doxastic justification of the original
belief when I locate more and more reasons for it. But, if I run out of reasons that are demanded
by the context, then I should either suspend belief in the original proposition or at least lower the
credence invested in the original belief.

Does my account face a troubling dilemma?

No. The putative dilemma is this: I must either accept or reject (K2). Never mind what the
supposed bad consequence of my accepting K2 is; I rejected K2. Now, what’s wrong with that?
Bergmann thinks there are two bad consequences. The first is that I then endorse an unjustified
(arbitrary) foundationalist view. But as argued above, rejecting K2 does not entail endorsing an
unjustified foundationalist view.

Nevertheless, Bergmann goes on to claim that there is even a deeper problem here, namely that I
seem to be committed to the view that ‘‘even apparently bad reasons will do [to make a belief
doxastically justified], since the reason can be an unjustified belief.’’ Now I agree that it is
epistemically better to have a justified belief than an unjustified one when the context requires it.
But it is not ‘‘bad’’ to have an as-yet-unjustified belief and employ it to justify a belief when the
as-yet-unjustified belief is not under scrutiny in the context. It’s not a ‘‘bad’’ reason simply
because it is not yet justified. (More about this in the concluding comments.)
Does my argument against foundationalism fail (for the reason Bergmann gives)?

No. If I understand Bergmann’s argument here, it rests on the claim that the examination of a so-
called basic belief, b, is satisfied either by basing b on an unjustified belief or the required
examination takes it for granted that K2. He says that taking K2 for granted leads to global
skepticism2 and an examination of one’s beliefs that uncovers an inference from an unjustified
belief is useless. Well, by now, I hope my response is predictable. If I locate a reason, r, for b,
the examination is useful since the reasoning has progressed, and, consequently, the belief that b
is (at least to some extent) doxastically justified. And if the context requires that r be doxastically
justified, the infinitist has an account of propositional justification that makes it possible to
satisfy that requirement.

Concluding comments

There is a refinement of my view that is required given Bergmann’s comments about K2. It
depends upon employing the notion of propositional justification and, consequently, once again
underscores the significance of that concept. In his footnote 8, he notes that in correspondence I
wrote that not only does a belief’s doxastic justification not require that the belief on which it is
based be doxastically justified, it also does not require that the doxastically justified belief be
propositionally justified. There is an ambiguity here that should be addressed.

It is a commonplace in epistemology to distinguish between subjective and objective

justification. I take it that distinction only applies to doxastic justification. But it depends
crucially on the concept of propositional justification. We can say that a belief is subjectively
doxastically justified iff the belief is inferred from another belief that is a reason for it regardless
of whether the reason is propositionally justified, and we can say that a belief is objectively
doxastically justified only if the reason is propositionally justified. Thus, if a belief is objectively
doxastically justified, its content is propositionally justified and the belief on which it is based is
potentially objectively justified (since there will be reason available for it); and the propositional
content of a belief that is merely subjectively doxastically justified is not propositionally justified
and the propositional content of the belief on which it is based is not even potentially objectively
doxastically justified. Now, when we say that doxastic justification is lost if there is no further
reason, we mean that the subjective doxastic justification is lost. If there is no further reason for
the belief in question, it was never objectively doxastically justified in the first place, so the
objective doxastic justification couldn’t be lost.

I will conclude with a comment about a possible rapprochement between Bergmann and me. I
think knowledge that p is best thought of as true, undefeated, objectively justified belief that p.
Now suppose that S has located a reason, r, for p, but has not yet located a reason for r. If the
propositional content of the as-yet-unjustified belief that r is not propositionally justified, then
the belief that r cannot become objectively doxastically justified and the belief that p can not be
objectively doxastically justified by the belief that r, since the series {p, r ...} is not infinite. So,
there is a K2-like principle endorsed by infinitism:
K2*: A belief can be objectively doxastically justified by being based upon some other belief
only if i) the propositional content of that other belief is propositionally justified and ii) the other
belief is objectively doxastically justifiable.

That’s pretty close to his K2.

1 The disanalogy, of course, is that (thank heavens!) cars have a finite cost.

2 He doesn’t say why it would lead to global skepticism, but I think he probably has this in mind:
If I were to accept K2, then the finite mind problem would reemerge and we could never have a
doxastically justified belief—which is required for knowledge.

I would like to thank Michael Bergmann for his many discussions and correspondence with me
on this issue. Although, regretfully, we still disagree about some of these issues, I have learned a
great deal from our exchanges.

ARBITRARY FOUNDATIONS?
MICHAEL HUEMER
THE PHILOSOPHICAL FORUM
Volume XXXIV, No. 2, Summer 2003

FOUNDATIONALISM AND THE OBJECTION FROM ARBITRARINESS

The central idea of foundationalism is that in some circumstances, we do not need reasons for
believing a proposition to be justified in believing it. More precisely:

Foundationalism: For some persons S and some propositions P, S is justified in believing

P, and some of S’s justification for believing P does not depend upon S’s having a reason
or reasons for believing P.

Thus, for example, most foundationalists would agree that, when I actually am in pain, I do not
require any reasons for believing that I am in pain to justifiedly believe that I am.

The objection to this doctrine that I consider here focuses on the following putative consequence
of foundationalism: It is possible for there to be a person who is justified in believing a
proposition but has no reasons for believing it.1 Hereafter, I shall call a justified belief for which
the believer has no reasons a “foundational belief.”

It is widely believed that foundationalism faces serious problems. Some of these alleged
problems turn on misconstruals that attribute needlessly strong assumptions to the
foundationalist—such as the assumption that (some or all) foundational beliefs are absolutely
certain, that they are indefeasible, or that it is impossible to have evidence for them.2 But there is
at least one objection to foundationalism, felt by many, that does not depend upon a
misunderstanding of the foundationalist’s commitments. Many suppose that foundationalism
faces a prima facie problem of showing why putatively foundational propositions are not
“arbitrary.” In the words of Peter Klein:

[F]oundationalism is unacceptable because it advocates accepting an arbitrary reason at

the base, that is, a reason for which there are no further reasons.3

I shall refer to this as the “Arbitrariness Objection” (AO). The most puzzling thing about this
objection is what is meant by “arbitrary.” In the following, I shall argue that the AO fails on any
of several obvious construals of the term.

OVERTLY QUESTION-BEGGING FORMULATIONS

“Arbitrary” seems to be a negative (epistemically) evaluative term. Surely no one would wish to
call any belief that he endorsed “arbitrary.” So perhaps “arbitrary” simply means “unjustified.”
On this interpretation, the AO objects that foundationalism advocates accepting unjustified
claims. But foundationalists, obviously, do not advocate accepting any claims that they, the
foundationalists, regard as unjustified. So the objection must be that foundationalists think
certain propositions are justified, whereas in fact those propositions are unjustified. But this is
evidently not a substantive objection to foundationalism. It is just a general assertion that the
foundationalist view of justification is wrong. For the alleged arbitrariness of uninferred beliefs
to furnish a reason for rejecting foundationalism, “arbitrariness” must denote some feature, other
than their alleged lack of justification itself, that shows why those beliefs cannot be justified.

An obvious alternative, then, is this: for a claim to be “arbitrary” (for S) is simply for there to be
no reasons (that S has) in its favor.4 On this interpretation, the AO charges that foundationalism
is unacceptable because it holds that one can be justified in accepting (it “advocates accepting”)
a claim for which there are no reasons (or: for which one has no reasons). But this is simply to
repeat the thesis of foundationalism, appending the assertion that the thesis is unacceptable.

Consider a third interpretation in the same vicinity. Perhaps “arbitrary” should be read as a
“thick” (epistemically) evaluative term—that is, a term whose meaning contains both evaluative
and nonevaluative components. The analogy in the moral realm would be to terms such as
“honest” and “courageous,” both of which are positively evaluative and also have a descriptive
meaning, as opposed to such terms as “good” and “bad,” which are purely evaluative. A rough
analysis of “S is courageous” might be, “S is able to overcome his fears (which is good).”5
Similarly, then, perhaps “P is arbitrary,” means something like, “There are no reasons for P
(which is epistemically bad).” Once again, the AO would transparently beg the question, merely
charging that foundationalists advocate accepting claims for which we lack reasons, and that this
is bad.

The situation, on any of these first three interpretations, is roughly this: one philosopher says it is
possible for some A’s to be F. A second philosopher objects that this cannot be, for such objects
would not be F. When pressed as to why they would not be F, he explains that they would not be
F because they were A’s (which cannot be F). If we dispute about whether there can be
uninferred but justified beliefs, the dispute cannot be resolved by pointing out that such beliefs
would be uninferred, nor by asserting that uninferred beliefs are not justified.

ARBITRARINESS AND PROBABILITY, PART 1

Perhaps a proposition is “arbitrary” when it is no more likely to be true than its negation. On this
interpretation, it seems, the AO alleges that propositions for which we lack reasons are, in
general, no more likely to be true than their negations—which is to say that they have
probabilities less than or equal to 1/2.

But this assertion is incompatible with the axioms of probability theory. No probability
distribution can assign a prior probability 1/2 to every proposition, nor even to every contingent
proposition in a given domain. For example, it is possible that one have no information about the
color of the sky on Venus, but there is no coherent way to assign a probability 1/2 to all of the
following five statements (given the incompatibility of red and blue): “The sky on Venus is red,”
“The sky on Venus is not red,” “The sky on Venus is blue,” “The sky on Venus is not blue,”
“The sky on Venus is red or blue.” One might wish to avoid this problem by positing that, when
one lacks reasons for or against a proposition, the proposition does not have a probability, rather
than its having a probability 1/2, and that propositions that have no probabilities also count as
“arbitrary.”6 So the antifoundationalist claim is that when one lacks reasons for believing a
proposition, then either it has no probability, or its probability is less than or equal to 1/2.

How should we understand the notion of probability at work here? We might understand it in any
of at least three ways: (i) in terms of a physical (e.g., a frequency or propensity) interpretation of
probability,7 (ii) in terms of a subjective interpretation,8 or (iii) in terms of a logical or
epistemological interpretation.9 Let us consider how the AO fares on each of these
interpretations.

(i) Physical probability: One problem with using a physical interpretation of probability is that
there seem to be some propositions that have no physical probabilities—for instance, necessary
truths, laws of nature, and perhaps even (on a frequentist view) propositions about particular
events (as opposed to event-types). If we say that all propositions that lack physical probabilities
are unjustified, we will have some uncomfortable results.10

A second problem is that it is obviously not true that, whenever we lack reasons for believing a
proposition, it either has no physical probability or has a physical probability less than or equal
to 1/2—at least, not if physical probabilities in general exist. For physical probabilities are by
definition objective and thus do not depend upon aspects of one’s epistemic position, such as
what reasons one has available, nor do they vary from one person to another as epistemic
position does.

Objection: According to the frequency interpretation of probability, a proposition’s probability is

relative to a reference class, and one might maintain that what reference class is relevant depends
upon one’s epistemic position. Thus, for example, reliabilists would hold that to be justified, a
belief must have a high probability of being true, relative to the reference class of “beliefs
formed by the same method.” However, this condition gives no comfort to the
antifoundationalist, as there is no ground for asserting that only beliefs based on reasons can be
probable in this sense (i.e., that no other belief-forming method can be reliable).11 Nor is it clear
how else the antifoundationalist might specify the relevant reference classes, and defend the
specification as correct, such that it would always turn out that beliefs formed in the absence of
reasons had a low probability.

(ii) Subjective probability: It is also false that, whenever a person lacks reasons for believing a
proposition, he either has no degree of belief in that proposition or has a degree of belief less
than or equal to 1/2. Indeed, securing high degrees of belief for putatively foundational
propositions would seem to be the least of the foundationalist’s worries. This is evidently not the
interpretation of probability that the antifoundationalist has in mind.

(iii) Logical/epistemic probability: According to the logical interpretation of probability,

probabilities are either logical properties of propositions (something like degrees of necessity) or
logical relations between propositions (something like degrees of entailment).12 The important
thing about this hypothesized logical property is that when a proposition has it to a high degree
(has a high probability), we are justified in believing it.13 Similarly, according to the epistemic
interpretation of probability, probabilities are degrees of justification, or degrees to which a
rational agent would believe a proposition.

On an epistemic/logical interpretation of probability, the AO transparently begs the question. It

collapses into the first interpretation considered in the section Overtly Question-Begging
Formulations above. For to say that a putatively foundational proposition lacks a high logical or
epistemic probability is just to say that the proposition is not justified, or that a rational person
would not believe it. And to say that propositions for which we lack reasons cannot, in general,
have a high logical or epistemic probability is just another way of saying that foundationalism is
false.

ARBITRARINESS AND PROBABILITY, PART 2

Perhaps a proposition is “arbitrary” when there are no reasons for believing (or: one has no
reasons for believing) that it is more likely to be true than its negation.14 Then the AO posits as a
necessary condition on the justification of a belief that one have a reason for thinking the belief
has a high probability. If we understand this as physical probability, then one of the problems
from the preceding section returns: The alleged necessary condition on justification would never
be satisfied in the case of laws of nature or necessary truths, since these propositions do not have
physical probabilities, nor do we have reasons for thinking that they do.
Moreover, there are three general problems for this interpretation of the AO, regardless of how
probability is understood. First, it appears to beg the question once again. The central thesis of
foundationalism is that we do not need reasons for certain propositions in order to be justified in
believing them. Certainly it would overtly beg the question to propose the following as a
necessary condition on justification:

a. In order for S to be justified in believing that P, S must have a reason for believing that
P.

The present version of the AO posits a just slightly different condition:

b. In order for S to be justified in believing that P, S must have a reason for believing that
(P is more likely than ~P).

But it seems that any reason for believing that P is more likely than ~P would also be a reason
for believing that P. If so, (b) obviously and directly entails (a), and it is unclear why anyone at
all tempted by foundationalism would be moved to accept condition (b). I submit that condition
(b) is too similar to the negation of foundationalism to be taken as a premise in a dispute directly
over foundationalism.15

Second: condition (b) seems to be motivated by a level confusion, particularly if we understand

“likely” in terms of subjective probability, logical probability, or epistemic probability.16 For on
these interpretations, (b) amounts to something like (c) or (d):

c. In order for S to be justified in believing that P, S must have a reason for believing that
S believes that P.

d. In order for S to be justified in believing that P, S must have a reason for believing that
S is justified in believing that P.

Why should one think that S must have a reason for a second-order belief in order to be justified
in a first-order belief? Let us focus on condition (d) as the more initially plausible of the two. It
seems that (d) is only plausible if one accepts:

e. In order for S to be justified in believing that P, S must be justified in believing that S

is justified in believing that P.

No clearer example of a level confusion could be desired.

Third and finally, condition (b) leads to a vicious infinite regress. For if (b) is true, then the
following series of conditions must also all be true:

b.’ In order for S to be justified in believing that (P is more likely than ~P), S must have a
reason for believing that [(P is more likely than ~P) is more likely than ~(P is more likely
than ~P)].
 b’’ In order for S to be justified in believing that [(P is more likely than ~P) is more likely
than ~(P is more likely than ~P)], S must have a reason for believing that {[(P is more
likely than ~P) is more likely than ~(P is more likely than ~P)] is more likely than ~[(P is
more likely than ~P) is more likely than ~(P is more likely than ~P)]}.

And so on. By the transitivity of the “in order for” relation, S’s justification for the initial, first
order belief turns out to require an infinite series of reasons for ever more complex, higher-order
beliefs. Even the infinitist would not countenance this sort of infinite series.17

THE SUPERVENIENCE OF EPISTEMIC PROPERTIES

A belief’s justificatory status depends upon its other, nonepistemic properties (including
relational properties). Which properties it depends on is a subject of great controversy, but
almost all epistemologists, including almost all foundationalists, will agree that it depends upon
some properties or other.18 Our final formulation of the Arbitrariness Objection charges that
foundationalism is incompatible with this principle. Let us call properties of a belief that would
count in favor of its being justified “positively epistemically relevant properties” (by analogy to
positively morally relevant properties), and those that would count against a belief’s being
justified “negatively epistemically relevant properties.” On the present formulation, a belief is
“arbitrary” just in case it lacks any positively epistemically relevant properties, and the AO
charges that foundationalism is committed to regarding some beliefs as justified even though
they have no positively epistemically relevant features.

Why might the objector think that putatively foundational propositions in general have no such
features? Let’s hear from Peter Klein:

Suppose some proposition, say F, is offered as a putative foundational one. . . . Either F

has some characteristic [ ] which makes it such that it is (highly likely to be) true or it
doesn’t. If it does, then the possession of [ ] can be and should be offered as a reason for
thinking that it is true—and the regress continues. If it doesn’t, then there is nothing that
distinguishes it from nonfoundational propositions and it becomes arbitrary to treat it as
foundational.19

Suppose S believes F and—grasping the first horn of Klein’s putative dilemma—that S’s belief
that F has a feature, Ф, which makes it justified. Pace Klein, the latter condition does not entail
that S has any reasons for believing F, not even the reason that the belief that F has Ф. First,
because S need not know or even believe that his belief that F has Ф. By hypothesis, it is the fact
that a belief has that makes it justified, not the fact that S believes or knows that it has Ф. Indeed,
in the case of some versions of foundationalism, S might not even have access to the information
that his belief has Ф (consider reliabilist and other externalist theories of justification). But one
need not be an externalist to appreciate the point: even an access internalist who holds that S
necessarily has access to (that is, is in a position to know) the conditions that make the belief that
F justified, need not think that S actually knows those conditions.20

Second, even if S knows that his belief that F has Ф, this need not—in fact, barring cases where
S has false beliefs about his own beliefs, it can not—be S’s reason for believing F. For S would
not have this putative reason available unless he already believed F (and that belief had Ф). It is
possible that his realization that his belief that F had Ф would increase his degree of justification
for believing F, and perhaps further strengthen his belief, but since by hypothesis the belief
already had Ф, the belief was already justified.21 Thus, it is not the case that S’s belief that F
depends for its justification on his believing that the belief that F has Ф. Any (nonskeptical)
theorist who accepts supervenience, not just the foundationalist, should accept this point. Given
the supervenience of epistemic properties, there must be some condition or set of conditions, Ф,
that suffices for a belief’s being justified. Ipso facto, it cannot be the case that the belief depends
for its justification on some condition additional to Ф, and so in particular, the belief’s
justification cannot depend upon the addition of the knowledge or awareness of Ф.

The presence of positively epistemically relevant features would indeed be a reason for
observers, such as epistemologists who are interested in questions of justification, to think that
S’s belief that F is justified. They can also be described as, in an agent-neutral sense, reasons
why S’s belief that F is justified—that is, factors explaining why S’s belief is justified. But this
should not fool us into thinking that they must constitute S’s reasons for believing F, or even
reasons that S has. To see the necessity of distinguishing the reasons why a person’s belief is
justified from his reasons for the belief, consider the case of inferential justification—if the
distinction is needed in that case, then it is surely also needed for the case of noninferential
justification. Suppose, then, that S is inferentially justified in believing that the streets are wet,
on the basis of his justified belief that it is raining (perhaps together with some background
beliefs that need not concern us). What is S’s reason for believing that the streets are wet? By
hypothesis, his reason is

A. That it is raining.

What is the reason why S’s belief that the streets are wet is justified? Not that it is raining. No
epistemologist will cite the mere state of the weather as either a necessary or a sufficient
condition on S’s belief being a justified one. Rather, an epistemologist will say the explanation
for why S’s belief is justified is something like the following:

B. That S justifiedly believes that it is raining, that the proposition that it is raining makes
it probable that the streets are wet, that S’s belief that the streets are wet is based on S’s
belief that it is raining, and that S lacks any defeaters for the belief that the streets are
wet.

The reason why S’s belief is justified is (B), but S’s reason for the belief is (A).22 The two are
logically independent propositions.23 And no one would argue that, because (B) must be true in
order for S’s belief that the streets are wet to be justified, S’s belief is really based on (B) rather
than (A), or that (A) is not an adequate justification for S’s belief. Nor, therefore, should one
argue that a putatively noninferentially justified belief is really justified by the analogous
condition (i.e., the condition that would explain why the noninferentially justified belief is
justified).

CONCLUSION
By their own lights, antifoundationalists are being arbitrary if they reject foundationalism
without presenting cogent arguments for this rejection. But antifoundationalist arguments prove
surprisingly few and weak on examination. An inarticulate charge of “arbitrariness” seems to
motivate many antifoundationalists at an intuitive level. Yet on the most obvious interpretations,
the charge amounts to no more than a bald denial of foundationalism. If “arbitrary” is taken as a
negative epistemically evaluative term (for example, as meaning “unjustified”), then the charge
of arbitrariness begs the question by simply asserting that putatively foundational propositions
are epistemically unacceptable. If “arbitrary” is taken to mean “not supported by reasons,” then
the foundationalist finds himself or herself merely charged with being a foundationalist.

The question-begging is slightly disguised when the antifoundationalist charges that putatively
foundational propositions are not “likely to be true,” or that we lack reasons for thinking they are
likely to be true. The most natural interpretation of the notion of “likeliness” in this context is
that of logical or epistemic probability, rather than, say, physical or purely subjective probability.
But on this interpretation, the objection begs the question once again, amounting to no more than
a paraphrase of the assertion that putatively foundational propositions lack a high degree of
justification. Furthermore, the objection seems to require the incoherent supposition of a default
state in which all propositions start out with probabilities equal to or less than 1/2.

The most sophisticated version of the objection from arbitrariness holds that foundational beliefs
cannot have any features that make them justified, since any such features would count as
reasons for believing them. Therefore, the objection goes, the foundationalist must deny the
principle of the supervenience of epistemic properties on nonepistemic properties. But we saw
that this objection involves a confusion between the reasons why a belief is justified and the
reasons that a believer has for the belief. Once we make this distinction, we see that the
foundationalist’s acceptance of the former sort of “reasons” does not commit him to the
existence of the latter sort of “reasons.”

Are there other possible interpretations of “arbitrariness” that might work in the argument?
Perhaps—I do not know how to prove that there are none. But any interpretation of
“arbitrariness” must respect the following constraints if it is to render the AO cogent: (i)
Arbitrariness should not be an epistemically evaluative property. (ii) It should not merely be one
of the defining characteristics of foundational beliefs, for example, the characteristic of not being
supported by reasons. (iii) It should be a property that all beliefs not supported by reasons have.
And (iv) it should be a property the possession of which prevents a belief from being justified.
We have seen that each of the most salient candidate interpretations of “arbitrariness” fails at
least one of these conditions.

1 In fact, this does not follow from foundationalism as defined, for two reasons. First, because a
foundationalist might hold only that some beliefs have a relatively low degree of noninferential
justification, insufficient by itself for justified belief. This would be plausible in conjunction with
the view that coherence among these beliefs ratchets up their justification. See Laurence
BonJour’s definition of “weak foundationalism,” The Structure of Empirical Knowledge
(Cambridge, MA: Harvard University Press, 1985) 28. Second, the claim that A does not depend
on B does not imply that it is possible for A to exist in the absence of B. For if B were an
inevitable consequence of A, or if B were necessary on its own, without reference to A, then it
would be true that one could not have A without B but false that A depends on B. One could
hold that there are some noninferentially justified propositions that we inevitably also have
reasons for. However, I shall pass over this point for now, because I do not consider it the most
interesting problem for Arbitrariness Objection to foundationalism.

2 None of these things follows from foundationalism. See William Alston, “Has
Foundationalism Been Refuted?” Epistemic Justification (Ithaca, NY: Cornell University Press,
1989) 39–56, for responses to several objections of this kind.

3 Peter Klein, “Human Knowledge and the Infinite Regress of Reasons,” Philosophical
Perspectives 13 (1999): 297. The AO is more commonly encountered in conversation than in
print.

4 See the quotation from Klein, “Human Knowledge,” above (footnote 3).

5 Here I pass over some nuances in the interpretation of thick evaluative terms. For instance,
does the user of the term “courageous” assert or merely presuppose that the ability to overcome
fears is morally good? Furthermore, the descriptive aspect of courage is almost certainly more
complex than “overcoming one’s fears” (consider the person who is said to have fought “a
courageous battle with cancer”). But these issues are not relevant to our purposes here. None of
these concerns serves to detract from the question-begging character of the AO when “arbitrary”
is interpreted along anything like these lines.

6 Alternately, of course, one could claim that the proposition’s probability is unknown—but this
sort of move is refuted by the considerations adduced in the section, Arbitrariness and
Probability, Part 2.

7 I have in mind the sort of interpretations offered by Richard von Mises, Probability, Statistics,
and Truth (London: Allen, 1957) and Karl Popper, “The Propensity Interpretation of
Probability,” British Journal for the Philosophy of Science 10 (1959): 25–42. “Physical
probability” here is slightly misleading, since frequentist and propensity interpretations may
apply equally well to probabilities of nonphysical events, if such there be, as to probabilities of
physical events. The term is intended merely as a contrast to “subjective probability,” “epistemic
probability,” and “logical probability.”

8 See Colin Howson and Peter Urbach, Scientific Reasoning: The Bayesian Approach (LaSalle,
IL: Open Court, 1989), ch. 3, for a subjective interpretation of probability. This interpretation is
closely related to the epistemic interpretation, since it is standardly stipulated that probabilities
are rational degrees of belief.

9 See John Maynard Keynes, ATreatise on Probability (London: Macmillan, 1921); Rudolph
Carnap, Logical Foundations of Probability (Chicago: University of Chicago Press, 1950). The
distinction between epistemic and logical interpretations is just this: the former identifies a
proposition’s probability with its level of justification, while the latter postulates a logical
property that grounds (and is equal to) a proposition’s level of justification. This distinction is
not important for our purposes.
10 I set aside the question of whether it is propositions or events that should be ascribed
probabilities. If one prefers to ascribe probabilities to events, as is perhaps more appropriate to a
physical interpretation of probability, one can rephrase the preceding statement in the text as
follows: “if we say that all propositions other than those asserting the occurrence of events that
have physical probabilities are unjustified, we will have some uncomfortable results.” Some
rephrasing of this kind is required in order to make out a version of the arbitrariness objection
referring to physical probability.

11 Thus, for example, Alvin Goldman’s reliabilism leads him to embrace a form of
foundationalism; see his “What Is Justified Belief?” Liasons: Philosophy Meets the Cognitive
and Social Sciences (Cambridge, MA: MIT Press, 1992) 117–18.

12 Keynes 5, 15–16. This cannot be taken literally, since neither entailment nor necessity comes
in degrees. Instead, the theory is that there is a logical property that does come in degrees, such
that necessity is one of the extreme values of that property (impossibility being the opposite
extreme); or that there is a logical relation that comes in degrees, such that entailment is one of
the extreme values of that relation (incompatibility being the opposite extreme).

13 Keynes 11–12.

14 This interpretation is suggested by the argument in BonJour, Empirical Knowledge 31–32.

15 Daniel Howard-Snyder has pressed this objection against BonJour (“BonJour’s ‘Basic
Antifoundationalist Argument’ and the Doctrine of the Given,” Southern Journal of Philosophy
36 [1998]: 163–77, esp. 165).

16 See William Alston on level confusions, “Level Confusions in Epistemology,” Epistemic

Justification (Ithaca, NY: Cornell University Press) 153–71. A level confusion is a confusion
between, for example, the conditions for knowing that P and the conditions for knowing that one
knows that P, or between the conditions for being justified in believing that P and the conditions
for being justified in believing that one is justified in believing that P, and so forth. Cf. Michael
Huemer’s objection to Fumerton’s Principle of Inferential Justification, “Fumerton’s Principle of
Inferential Justification,” Journal of Philosophical Research 27 (2002): 329–40.

17 Unlike “the foundationalist” and “the antifoundationalist,” “the infinitist” turns out to be a
definite description, referring to Peter Klein. Klein (“Human Knowledge” 309–10) rejects a
condition on justified belief very similar to (b) on the grounds that it “confuses having a justified
belief that p with having justified beliefs about p’s justificatory status.” He also seems to grant
the existence of a limit to the complexity of beliefs that a person can grasp.

18 See, for example, Ernest Sosa on “formal foundationalism,” Knowledge in Perspective

(Cambridge: Cambridge University Press, 1991) 179–80. For attempts to specify some
epistemically relevant features of beliefs, see Roderick M. Chisholm, The Foundations of
Knowing (Minneapolis: University of Minnesota Press, 1982) 12, 21–2; Robert Audi, The
Structure of Justification (Cambridge: Cambridge University Press, 1993) 308; Michael Huemer,
Skepticism and the Veil of Perception (Lanham, MD: Rowman, 2001) 99.

19 Peter Klein, “Foundationalism and the Infinite Regress of Reasons,” Philosophy and
Phenomenological Research 58 (1998): 919–25, esp. 924. See also BonJour 31, and Klein,
“Human Knowledge,” 303–04, for similar arguments. Note that Klein (“Human Knowledge,”
309, 319 n. 15) seems to accept an interpretation of probability in terms of justification for belief,
making “highly likely to be true” in the above passage plausibly interpreted as “justified.”

20 I here refer to what Richard Fumerton, Metaepistemology and Skepticism (Lanham, MD:
Rowman, 1995) 63, calls “weak access internalism.” “Strong access internalism,” on the other
hand, requires that S actually know or otherwise be aware of the conditions that make S’s belief
justified. The strong access internalist cannot make use of the reply I give here. However, strong
access internalism should be rejected on independent grounds, since it generates an infinite
regress of beliefs of ever-increasing complexity (Fumerton 64, 82).

21 Compare Fumerton’s argument against the form of access internalism that claims that access
to the conditions that constitute a belief’s justification is part of what constitutes the belief’s
justification: “The idea that X constitutes one’s justification for believing P only if one’s
awareness of X is added to X is equivalent to holding that X constitutes one’s justification for
believing P only if it does not really constitute one’s justification for believing P” (81).

22 It is not important for my present purposes that (B) reflect the actually complete and correct
theory of inferential justification, as the aim here is merely to illustrate the distinction, in general,
between S’s reason for believing P, and an epistemological account of S’s justification for
believing P. The addition of further clauses to (B), as might be required to turn it into a correct
account of inferential justification, would only serve to make this distinction more stark.

23 That is, (A) does not entail (B), nor does (B) entail (A).
Forthcoming in The Oxford Handbook of Skepticism, ed. by John Greco (Oxford: Oxford
University Press, 2008).
The Pyrrhonian Problematic
Markus Lammenranta

… 1. The Regress Problem

The core of the Pyrrhonian problematic is often taken to be an argument for the impossibility of
justified beliefs. This argument is found in the five modes of Agrippa and is called the regress
argument. It is an argument for a very strong form of global skepticism that denies the possibility
of any justified beliefs. Suspension of belief would then seem to be the appropriate attitude to
every proposition: if we are not justified in believing anything, we should not believe anything.

It was once common to interpret the Pyrrhonian skeptic as advocating this argument and basing
her suspension of belief on its negative epistemic conclusion. This raised the accusation of self-
refutation. If the conclusion of the argument is true, no one can be justified in believing anything,
not even the premises of the argument itself, and thus they cannot be used to justify the
conclusion. No one can thus effectively argue for the conclusion that there cannot be justified
beliefs. So why should anybody take the argument seriously?

Many recent scholars of Hellenistic philosophy take this to be a misinterpretation of the

dialectical strategy of ancient skeptics.3 The arguments of ancient skeptics are ad hominem
arguments. They are directed at us, the dogmatists, and use only premises that we accept. The
regress argument poses a problem for us because we are inclined to accept the premises and take
the argument to be valid, though we are not willing to accept the conclusion. As many
contemporary epistemologists see it, a genuine skeptical problem is a kind of paradox: we are
inclined to accept the premises and the denial of the conclusion, though they form an inconsistent
set of propositions.4

On the other hand, Sextus says that the skeptics welcome the self-refuting nature of their
arguments. Their arguments are like a ladder that we throw away once we have climbed up it, or
like fire that after consuming the fuel also destroys itself.5 However, as Hume noticed, our beliefs
tend to come back when we lose our reasons for giving them up. But then, the skeptic may point
out, we also return to the arguments that make us reject the beliefs once again. So the skeptical
arguments at least create instability in our beliefs. This may be enough to make them a serious
skeptical challenge to our beliefs.6

In order to pose a genuine paradox, the regress argument must have plausible premises. As
Michael Williams points out, they should not be based on theoretical ideas that we are not bound
to accept. They should be based on our ordinary intuitions about justification and knowledge.7
Does the regress argument have such intuitively plausible premises? This is the standard
formulation of the argument:

1. In order to be justified in believing something, one must believe it on the basis of good
reasons.

2. Good reasons must themselves be justified beliefs.

3. Therefore, in order to be justified in believing something, one must believe it on the

basis of an infinite number of good reasons.

4. No human being can have an infinite number of good reasons.

5. Therefore, it is humanly impossible to have justified beliefs.

Most epistemologists have not found all the premises plausible. There are several ways of
criticizing the argument. Indeed, epistemologists use the argument typically just to classify the
possible theories of justification. These are distinguished by virtue of which premise or step in
the argument they deny: Circular (linear) coherentists deny the step from premises 1 and 2 to 3
because they think that a circular chain of reasons can justify a belief. Wittgensteinian
contextualists think that the chain of good reasons can terminate in beliefs that are not
themselves justified, and thus they deny premise 2. Foundationalists think that the chain
terminates in basic beliefs that are justified but do not derive all their justification from their
inferential relations to other justified beliefs.8 They deny premise 1.

The skeptical conclusion is essentially based on the infinitist lemma (3) that only an infinite
chain of reasons can justify a belief. All other theories except dogmatic infinitism deny this.
Though Peter Klein, who has recently defended the position, argues that infinitism need not lead
to skepticism, the skeptic has a strong position if infinitism is true. Indeed, Klein acknowledges
that skepticism is a serious possibility for the infinitist.9

The regress argument thus provides a real paradox and a skeptical challenge only if infinitism is
an intuitively plausible account of justification. Initially, it does not seem to be plausible. Why
should we possess a concept of justification that is not applicable to any finite being? So the
skeptic must do something to convince us of the truth of infinitism.

In his reconstruction of skeptical reasoning, Michael Williams appeals to our practice of giving
and asking for reasons: Suppose I claim to know something. You can ask me how I know it. As a
reply, I give my reasons for believing what I claim to know. But then you can ask how I know
my reasons, and so on. Williams’s skeptic admits that in real life this process of justification has
an end. It ends when my interlocutor is satisfied with my reasons. In spite of this, Williams’s
skeptic insists that the regress goes on. Whether or not anybody really challenges my reasons,
they can reasonably be challenged, and so I must go on giving reasons. The skeptic concludes
that I do not know what I originally claimed unless I have first completed an infinite number of
prior justifications, which is, of course, impossible.10 But is this really the intuitive conclusion to
draw from our practice of justifying knowledge claims?

Klein imagines a similar dialogue between Fred, the foundationalist, and Sally, the skeptic. Sally
asks Fred his reasons for believing that p. Fred gives as his reason his belief that q. Then Sally
asks his reasons for q, and the regress continues until Fred gives a reason that he takes to be a
basic belief, which, according to foundationalism, is justified independently of reasons. Now
Sally asks why Fred thinks that the basic belief is true. In order to avoid arbitrariness, Fred says
that his basic belief has property P, and that all beliefs having P are likely to be true. But then
Sally insists that the regress does not stop, and asks Fred for his reasons for the metabeliefs that
his belief has P and that beliefs having P are probably true. Klein concludes that nobody is
completely justified in believing anything because this would require going through an infinite
number of reasons, but that people can still be provisionally justified in believing things because
such infinite chains are in a relevant sense available to them.11

Neither infinitist conclusion from our ordinary practices is intuitive. We see this more clearly if
we think about the point of our practice of giving and asking for reasons. The point seems to be
to evaluate each other as potential informants. When you claim to know something, I ask your
reasons because I want to decide whether I can trust you and learn from you. In this sort of case,
I already have many beliefs about the world and your situation. I use these beliefs in evaluating
your trustworthiness, and I do not expect you to justify them. So the regress may very well
terminate in basic beliefs because these are the kind of beliefs I take to be probably true in your
circumstances. You need not justify them to me. This also explains why we sometimes attribute
knowledge to small children and nonhuman animals though they cannot justify their beliefs at
all.

If we imagine that the practice of giving and asking for reasons is conducted between the skeptic
and us, the situation is different. The skeptic does not accept any beliefs as being justified or true.
There is no way to convince her because she does not accept any reasons that we could give, so
there is no way out of the regress. But why should this show that it is impossible for us to have
justified beliefs and knowledge? Our justificational practice does not aim at convincing the
skeptic. It aims at convincing someone who already has many beliefs and is able to use these
beliefs for evaluating the given reasons. The fact that we cannot justify our beliefs to the skeptic
is of no epistemological importance.

So the skeptical strategy of arguing ad hominem from the infinite regress of reasons has no
chance of succeeding. Our ordinary epistemic concepts and practices cannot be used to support
it. It is thus hard to see how it could have created the skeptical crisis in early modern philosophy
and played such an important role in shaping modern epistemology. Neither does it explain how
the Pyrrhonian skeptics themselves became skeptics. Sextus does have a story about this, and this
story seems to give us an independent skeptical problem.

… 4. The Modes of Agrippa

The considerations that Sextus relies on in supporting (4) are systematized in the five modes of
Agrippa. We can see that the dialectical interpretation also fits very well with how they work.
According to Sextus, every object of investigation can be brought under the following five
modes:

According to the mode deriving from dispute, we find that undecidable dissension about
the matter proposed has come about both in ordinary life and among philosophers.
Because of this we are not able either to choose or to rule out anything, and we end up
with suspension of judgement. In the mode deriving from infinite regress, we say that
what is brought forward as a source of conviction for the matter proposed itself needs
another such source, which itself needs another, and so on ad infinitum, so that we have
no point from which to begin to establish anything, and suspension of judgement follows.
In the mode deriving from relativity, as we said above, the existing object appears to be
such-and-such relative to the subject judging and to the things observed together with it,
but we suspend judgement on what it is like in its nature. We have the mode from
hypothesis when the Dogmatists, being thrown back ad infinitum, begin from something
which they do not establish but claim to assume simply and without proof in virtue of a
concession. The reciprocal mode occurs when what ought to be confirmatory of the
object under investigation needs to be made convincing by the object under investigation;
then being unable to take either in order to establish the other, we suspend judgement
about both.20

Sextus describes here five modes that are supposed to induce suspension of judgment about any
object of inquiry. We may call them the modes of (1) disagreement, (2) infinite regress, (3)
relativity, (4) hypothesis, and (5) circularity. It is not clear how they are to be understood.
Usually they are thought to work together: the modes of disagreement and relativity challenge us
to justify our beliefs by revealing that there are competing claims about the matter, and then the
rest of the modes show that the process of justification cannot be completed in a satisfactory
way. All attempts to justify a belief lead either to an infinite regress, an arbitrary assumption, or
circularity.
Often the challenging modes are seen to be unnecessary because the skeptic is thought to assume
implicitly that our concept of justification requires noncircular and nonarbitrary reasons. So the
three modes of infinite regress, hypothesis, and circularity—the so-called Agrippa’s trilemma—
alone form the problem that can be identified with the regress argument against the possibility of
justified beliefs.21

I have argued that this interpretation of Agrippa’s problem does not make it a serious skeptical
challenge to our beliefs. It also misses the dialectical nature of the problem. According to the
dialectical interpretation, it is the mode of disagreement that is the central one, and the others are
subordinate to it. Its structure is very much the same as in the ten modes, and we may regard it as
a generalization of them. Suppose we have a question to which there are just two possible
answers, p and ~p. Then the mode of disagreement works as follows:

1. S1 believes that p.

2. S2 believes that ~p.

3. At most one of them is right.

4. The disagreement between S1 and S2 is irresolvable.

5. We should suspend judgment about p.

So the mode of disagreement alone is supposed to induce suspension of judgment. The other
modes are used if the dogmatist wants to deny 4, that the disagreement is irresolvable. He is then
asked how the disagreement is to be resolved. If he gives a reason r for his belief that p, it is
pointed out that there is also an irresolvable disagreement about r. If he admits this, he also
admits the irresolvability of the original disagreement. If he denies it, he is asked how the
disagreement about r is resolved, and so on. By repeating the mode of disagreement, the
dogmatist is led into Agrippa’s trilemma. Modes 2, 4, and 5 cannot resolve the disagreement:
modes 4 and 5 are clearly question begging, and nobody can complete an infinite chain of reason
required by mode 2.22

The dogmatist may also try to appeal to a criterion of truth, but he is told that there is an
irresolvable disagreement about it. If he tries to resolve this disagreement by appealing to a new
criterion of truth, he is led into a regress. If he appeals to the same criterion again, he is in a
circle.

And if he just assumes a criterion, the mode of hypothesis applies to him.23 Sextus clearly
assumes that in order to avoid skepticism, the dogmatist must be able to resolve all
disagreements there are or could be about the object of inquiry and the reasons he appeals to.
This requirement may seem too strong. We will see, however, that it arises from presuppositions
that are quite plausible.

… 10. Relativism and Antirealism

It seems that there is no satisfactory epistemological resolution of the Pyrrhonian problematic.
Perhaps we should rather focus on its realistic presuppositions. A radical Protagorean relativist
avoids the problem by denying that there are genuine disagreements about reality. If I believe
that p and you believe that ~p, we are both right because p is true for me and ~p is true for you.
There is no disagreement between us. Indeed, there can be no disagreements because every belief
or appearance is true. If truth is relative in this way, everybody can retain his or her beliefs
without begging the question against others.

The traditional objection is that relativism is self-referentially incoherent: when the relativist says
that relativism is true, she may mean that relativism is absolutely true or that relativism is
relatively true. In the former case, she contradicts her own view that truth is relative. In the latter
case, she acknowledges that relativism is true only for the relativist. It is not true for the
absolutist. This is dialectically inefficient. She cannot argue the absolutist out of his position.

If we take relativism to be a response to the Pyrrhonian problem, the objection is misdirected.

The problem is composed of an ad hominem argument against the absolutist. Because the
absolutist accepts the premises of this argument, he cannot avoid the conclusion that he should
suspend belief. If, in spite of this, he continues to hold to his beliefs, it is his position that is
incoherent. The relativist avoids the skeptical conclusion and the possible incoherence that it
produces by denying the premise that there are genuine disagreements. There is no reason for her
to suspend belief.38

However, because it denies the possibility of error and disagreement, Protagorean relativism is
extremely implausible. A more plausible form of alethic relativism—relativism about truth
relativizes truth to something other than singular persons, such as social groups or cultures. If
truth is relative to something that can be shared by different persons, error and disagreement are
possible for them. This shared background also provides the neutral basis for resolving their
disagreements. When the background is not shared, there can be no genuine disagreements. So
alethic relativism gives some hope for resolving the Pyrrhonian problematic.

Alethic relativism denies alethic realism that many philosophers take to be a mere truism.
According to alethic realism, it is an obvious necessary truth that a proposition that p is true if
and only if p. It follows from this that one and the same proposition cannot be true for some
people and false for others. So the truth-value of a proposition cannot vary in the way that alethic
relativism suggests. For these philosophers, this is a sufficient reason for rejecting alethic
relativism.39

Another form of relativism denies metaphysical realism rather than alethic realism. Instead of
truth, it relativizes facts or reality. According to metaphysical relativism, there are no facts
simpliciter or absolute facts. There are facts only relative to conceptual scheme, culture or some
other parameter. So beliefs that appear to ascribe incompatible properties to the same object may
all be made true by different relative facts. Metaphysical relativism explains how people who
appear to disagree about the same absolute fact can all be right by interpreting their beliefs as
being about different relative facts.

Though alethic relativism and metaphysical relativism promise to resolve the Pyrrhonian
problematic, both are problematic if they are taken to be global in scope. The problem of global
relativism is not its self-referential incoherence or dialectical indefensibility but vicious regress.
In the case of metaphysical relativism, we get the regress as follows. Assuming that a is an
object, F is a property and P1,…,P∞ is a sequence of parameters, metaphysical relativism says
that there are no facts of the form:

1. a is F.

There are only facts of the form:

2. a is F relative to P1.

We can now ask whether this relational fact is itself an absolute fact or a relative fact. If it is an
absolute fact, global metaphysical relativism is false. But if it is a relative fact, we get a more
complex relational fact:

3. The fact that a is F relative to P1 is relative to P2.

If this is also a relative fact, there must be a more complex relational fact, and so on ad infinitum.
Global metaphysical relativism entails thus the absurd view that all facts are infinitely long.40
And global alethic relativism creates a similar regress.41

We can avoid the problem by restricting the scope of relativism to some particular subject
matter, such as aesthetic, moral or epistemic truths or facts. For example, epistemic relativism
says that there are no absolute epistemic facts. There are no epistemic facts of the form: S is
justified in believing that p. There are only facts of the form: S is justified in believing that p
relative to the system of epistemic rules that the ascriber A accepts. This relational fact is not
really an epistemic fact. It is a mental fact about S’s and A’s beliefs. And because this is an
absolute fact, there is no regress. So epistemic relativism avoids the regress by interpreting
epistemic judgment as being made true by absolute nonepistemic facts.

There is a current debate about the coherence of even this sort of local relativism. Paul
Boghossian argues that it cannot be coherently formulated, but there are others who disagree.42
We cannot and need not resolve this disagreement here. It is enough to note that these forms of
local relativism can at most resolve disagreements about those subject matters that they are
restricted to. As we have seen, the Pyrrhonian problem is general: it concerns any object of
inquiry. So the local forms of relativism cannot resolve it.

It may be objected that there is a sense in which epistemic relativism is quite general: about any
disagreement, we can ask what the participants are justified in believing. Epistemic relativism
allows that each participant can be justified in his or her belief even if he or she cannot rationally
convince the other participants. Each can be justified relative to different epistemic rules. So
epistemic relativism explains how people can justifiably or rationally disagree and persist in their
beliefs even though they cannot rationally resolve their disagreements. Does this not resolve the
Pyrrhonian problematic?
We saw earlier that the Pyrrhonian problematic presupposes that we take an impartial point of
view to our disagreements. The problem is that, from this point of view, we cannot rationally
decide who is right and who is wrong and are forced to suspend belief. The relativist comes to
our rescue by pointing out that the disputing parties can all be right relative to different
parameters.

So the problem of deciding between them disappears. However, the epistemic relativist cannot
say this. She can say at most that when the participants take themselves to be justified in their
beliefs, they are right relative to their respective epistemic rules. This does not help us, the
impartial observers, to say who is right in the original disagreement about the absolute facts
because, according to epistemic relativism, there are no facts by virtue of which one system of
epistemic rules is more correct than the others.

Relativism is thus unable to resolve the Pyrrhonian problematic. Global relativism has absurd
consequences, and local forms of relativism—assuming that they can even be coherently
formulated—can at most handle some disagreements. Disagreements about most matters we care
about are left unresolved.

…If there can be no justified beliefs, the suspension of belief should be global. But this demand
only concerns the infinitist. If we are not infinitists, there is no moral for us. We can keep all our
beliefs. The same is true of the skeptics who do not accept the infinitist dogma.

Notes

3. See, for example, Michael Frede, ”The Sceptic’s Two Kinds of Assent and the Question of the
Possibility of Knowledge,” in The Original Sceptics: A Controversy, ed. M. Burnyeat and M.
Frede (Indianapolis: Hackett, 1997), 127–131; Julia Annas and Jonathan Barnes, The Modes of
Scepticism (Cambridge: Cambridge University Press, 1985), 45. <<Comp: En dash in number
ranges.>>

4. See, for example, Stewart Cohen, “How to Be a Fallibilist,” Philosophical Perspectives 2

(1988): 93–94.

5. Sextus Empiricus, Against the Logicians, trans. R. G. Bury (Cambridge, Mass.: Harvard
University Press, 1935), 487–489 (M VIII, 480–481).

6. Michael Williams, Problems of Knowledge (Oxford: Oxford University Press, 2001), 66.

7. Michael Williams, “Skepticism,” in The Blackwell Guide to Epistemology, ed. J. Greco and E.
Sosa (Oxford: Blackwell, 1999), 35–36.

8. Because holistic coherentism rejects the idea that justification depends on a linear chain of
reasons, we may classify it as a special case of foundationalism that takes all justified beliefs to
be basic beliefs.
9. Peter Klein, “Human Knowledge and the Infinite Regress of Reasons,” Philosophical
Perspectives 13 (1999): 297-325.

10. Williams, “Skepticism,” 39.

11. Peter Klein, “The Failures of Dogmatism and a New Pyrrhonism,” Acta Analytica 15 (2000):
14–17; Klein, “Human Knowledge and the Infinite Regress of Reasons,” 312–316.

20. Ibid., 40–41 (I, 165–169).

21. This is the standard way of understanding the modes of Agrippa among analytical
epistemologists. See, for example, Robert Fogelin, Pyrrhonian Reflections on Knowledge and
Justification (Oxford: Oxford University Press, 1994), pt. 2; Greco, Putting Skeptics in Their
Place, chap. 5; Ernest Sosa, “How to Resolve the Pyrrhonian Problematic: A Lesson from
Descartes,” Philosophical Studies 85 (1997): 229–249; Williams, “Skepticism”; Williams,
Problems of Knowledge, chap. 5.

22. One mystery of Agrippa’s problem is the mode of relativity. It also appears as one of the ten
modes, and Sextus says that it is the most general of them. One possibility is that it just spells out
the conclusion of the mode of disagreement: We can say how things appear to the participants in
the dispute, but we cannot say how they are in themselves. The object of dispute is not how
things appear. It is how they are in their real nature. The other interpretation is that Sextus
confuses skepticism and relativism. Richard Bett argues that Aenesidemus was really a relativist.
If Sextus took Aenesidemus to be a skeptic, the relativistic phrases may have been left in the text
by mistake. See Bett, “What Does Pyrrhonism Have to Do with Pyrrho?” in Ancient Scepticism
and the Sceptical Tradition, ed. J. Sihvola (Helsinki: Acta Philosophica Fennica, 2000). In any
case, relativism is a possible response to Agrippa’s problem. It denies 3. See section 10.

23. Sextus Empiricus, Outlines of Scepticism, 72 (II, 20).

38. Because the relativist therefore has no need to defend relativism for the absolutist, she may
very well concede that relativism is only relatively true.

39. Alston, Realistic Conception of Truth, 180.

40. See Paul Boghossian, Fear of Knowledge: Against Relativism and Constructivism (Oxford:
Clarendon Press, 2006), 54-57.

41. See Frederick F. Schmitt, Truth: A Primer (Boulder: Westview Press, 1995), 68-71.

42. Boghossian, Fear of Knowledge, chap. 6. The coherence of epistemic relativism is defended
by Ram Neta in “In Defense of Epistemic Relativism,” Episteme 4 (2007): 30-48; and by Gideon
Rosen in “The Case against Epistemic Relativism: Reflections on Chapter 6 of Fear of
Knowledge,” Episteme 4 (2007): 10-29. See also Boghossian’s response in “The Case against
Epistemic Relativism: Replies to Rosen and Neta,” Episteme 4 (2007): 49-65.
Probabilistic Justification and the Regress Problem
Jeanne Peijnenburg and David Atkinson
Published in Studia Logica 89(3) (2008) 333-341.

Abstract We discuss two objections that foundationalists have raised against infinite chains of
probabilistic justification. We demonstrate that neither of the objections can be maintained.

Keywords Probabilistic justification, regress problem, foundationalism, infinitism.

Introduction

The definition of `knowledge' as `justified true belief' is part and parcel of many a textbook in
epistemology, even though Gettier-like counterexamples date back to Plato's Theaetetus. The
existence of counterexamples is however not the only problem that the definition faces. It also
runs into the Regress Problem: if my belief in a true proposition S0 is justified by my belief in
another true proposition, S1, which in turn is justified by my belief in still another true
proposition S2, and so on, ad infinitum, then I have no epistemic justification whatsoever for a
belief in any of the propositions S0, S1, S2, … etc.

One might try to take the sting out of the Regress Problem by toning down the definition to a
more modest version, one in which we content ourselves with probabilistic support rather than
full-blown justification and where we are satisfied with degree of belief rather than belief tout
court. But this move clearly does not help us, for a probabilistic variant of the Regress Problem
is readily made: if I partially believe S0 because S0 is probabilistically supported by S1, and
partially believe S1 because it is probabilistically supported by S2, and so on, then it seems that
we are not even partially justified in believing any of the propositions S0, S1, S2, … etc.

Although the Regress Problem is a threat to both foundationalists and coherentists, we will
concentrate on its significance for the former. As is well known, foundationalists hold that every
process of inferential or conditional justification must come to an end in a basic proposition that
is itself non-inferentially or immediately justified. Foundationalists of the old, non-probabilistic
school claim that this basic proposition must be absolutely certain, and some of them maintain
that it is even a priori. Contemporary foundationalists however mostly wear probabilistic colours,
and they argue that the basic proposition need not amount to absolute certainty. It is enough if its
unconditional probability can be established, and if it is such that it probabilistically supports
another proposition p, which means that the conditional probability of p given that the basic
proposition is true, is greater than the conditional probability of p given that the basic proposition
is false. But no matter what their different views may be on basic propositions and on the nature
of justification, both factions share the same horror infinitatis: the concept of an endless chain of
inferential justifications not only bewilders the old-style foundationalists, but remains anathema
for their probabilistic offspring as well.1

In this paper we focus on the probabilistic variant: we will deal with foundationalism and with
the Regress Problem only insofar as they are clothed in probabilistic raiment. In particular, we
will discuss two objections that probabilistic foundationalists have raised against infinite chains
of probabilistic justification. We shall first recall the reply of Peter Klein to these objections, but
then we will give a better reply, one which shows that neither of the objections hold water.

1. Two objections

When one looks at the plurality of objections that probabilistic foundationalists have raised
against infinite chains of probabilistic justification, two in particular stand out. Each of them can
in principle also be deployed by `ordinary' foundationalists in arguing against `ordinary' infinite
chains, but here we are especially interested in the way they are used within a probabilistic
setting. Both objections are explicitly discussed in the polemic between Carl Ginet and Peter
Klein [2, p. 131{155], but they can be found at many other places in the literature.

The first objection refers to our limited cognitive capacity and we will call it the finite mind
objection. The idea is that we, with our finite minds, are unable to complete an infinite
probabilistic chain:

“Even if I could have an infinite number of beliefs, how would I ever know anything if
knowledge required an infinite epistemic chain?” [3, p. 183].

“... finite minds cannot complete an infinitely long chain of reasoning and so, if all
justification were inferential, no-one would be justified in believing anything at all to any
extent whatsoever” [4, p. 40]; [5, p. 2]; [6, p. 150].
The second objection implies that (partial or probabilistic) justification can never be created by
inferences alone. We dub it the transfer objection, after its formulation by Carl Ginet:

“Inference cannot originate justification, it can only transfer it from premises to

conclusion. And so it cannot be that, if there actually occurs justification, it is all
inferential. .... [T]here can be no justification to be transferred unless ultimately
something else, something other than the inferential relation, does create justification” [7,
p. 148] (italics by the author).

Here the thought is that propositions in an endless chain of probabilistic justifications are not
really justified. The only way to justify them would be to end the chain with a proposition that is
itself noninferentially justified. Since the chain in question is supposed to be probabilistic, this
last proposition must be an unconditional probability statement of the form P(Sn+1), where it is
understood that Sn+1 probabilistically supports Sn, so that P(Sn|Sn+1) > P(Sn|~Sn+1). In an
attempt to lend force to this objection, Ginet cites Jonathan Dancy:

“Justification by inference is conditional justification only; [when we justify A by

inferring it from B and C] A's justification is conditional upon the justification of B and
C. But if all justification is conditional in this sense, then nothing can be shown to be
actually, non-conditionally justified.” [8, p. 55].

So in Dancy's view, a proposition is actually or really justified if and only if its justification is in
the end non-conditional. In the context of probabilistic justification, this view implies that a
proposition S0 is only really justified if in the end we can calculate its non-conditional
probability, P(S0). But if P(S0) were to depend on an infinite chain of conditional probabilities
of the form P(Sn|Sn+1), then, according to Dancy, there is no non-conditional probability to be
justified.

As a long standing supporter of epistemic infinitism, Peter Klein has repeatedly argued against
the foundationalists' dislike of infinite chains [9], [10], [11], [12], [13]. In his view there is
nothing troublesome with infinite regresses, the reason being that an infinite chain of reasoning
need not be completed. What is more, the requirement that an infinite chain must be completed
“would be tantamount to rejecting infinitism” [9, p. 920]. The only thing that an infinitist
requires, Klein argues, is that for every proposition Sn in the probabilistic chain, there is a
proposition Sn+1 such that the conditional probability P(Sn|Sn+1) is known to be greater than
P(Sn|~Sn+1), for that is the condition under which we can properly say that Sn+1
probabilistically supports Sn.

As Klein sees it, the Regress Problem is only a problem because it has been given a faulty
formulation. Once we phrase it correctly, the problem loses its force. Consider again our infinite
chain of propositions S0, S1, S2, … etc., where each Sn is probabilistically justified by Sn+1. If
we now were to ask how a belief in any of those propositions can be justified, we would be
asking too much. What we should ask is how a belief in each of those propositions can be
justified, and the answer to this question is simple: the justification comes from the proposition
one step up in the chain [11, p. 729]. Claiming that a particular proposition Sn is (partially or
probabilistically) justified means no more than being able to point to another proposition Sn+1
that bestows the support in question, and that in turn receives support from still another
proposition. This, and nothing more, is what infinitism claims.2

Given these views, it is clear what Klein's replies to the two objections are. His reply to the finite
mind objection is that we need not, and should not, attempt to complete an infinite chain: any
completion of such a chain would y in the face of infinitism. As to the transfer objection, Klein's
reply is plainly to deny that real justification implies unconditional justification, or that an
infinite chain of conditional probabilities can culminate in a particular value for an unconditional
probability. In Klein's view,

“Infinitism ... depicts justification as emerging when the set of propositions that are
appropriately adduced as reasons expands. Of course, were the foundationalist to insist on
thinking of warrant as originating in some propositions and then being transferred by
inference to other propositions, he or she would be begging the question at hand. For it is
this very concept of warrant that infinitism is challenging.” [13, p. 152].

Klein's replies to the two objections are interesting, but in our view he concedes too much. He
grants the foundationalists that finite minds cannot complete infinite chains and he also grants
them that, were `real warrant' the same as `unconditional warrant' or as unconditional probability,
justification on the basis of an infinite epistemic chain would be impossible. Neither of these
concessions is necessary, as will be explained in the next section.

2. The objections refuted

Consider once more our chain S0, S1, S2, … etc., where each Sn+1 probabilistically justifies Sn.
Are we able to justify S0, in the sense that we can compute P(S0)? As we have seen, Audi and
Fumerton deny that we can by referring to the finite mind objection; Ginet and Dancy deny the
same on the basis of the transfer objection. Below we propose to deal with both objections,
starting with the latter.

Dancy's text makes it clear that there is a striking similarity between the transfer objection and
C.I. Lewis's reasons for claiming that an endless epistemic chain does not make sense [8, p. 57].
It is true that Lewis is an old style foundationalist, but his ideas are easily adapted to
foundationalism of probabilistic stripe.3 What Dancy does not say, however, is that Lewis's
critique of infinite epistemic chains is ambiguous or even incoherent, and that the incoherence
may be present in one and the same article (for example in [14, p. 173]). On the one hand Lewis
suggests that, where any regress of probability values is involved, the value of P(S0) cannot be
computed. On the other hand he claims that, under these circumstances, the value of P(S0) will
always be zero.

It seems to us that the transfer objection suffers from a similar ambiguity. Consider Dancy's
phrase “if all justification is conditional, then nothing can be shown to be actually, non-
conditionally justified”. What does this phrase mean for the value of P(S0), if this value were to
be determined by an infinite chain of conditional probabilities? Does it mean that it is impossible
to say what exactly this value is? Or does it mean that the value becomes zero? Under the first
interpretation, the transfer objection seems to boil down to the finite mind objection; for the latter
assumes that P(S0) has a definite value, but that our tiny brains lack the equipment to compute it.
Under the second interpretation, on the other hand, the transfer objection once was supported by
no one less than Bertrand Russell. Russell argued that, in the case at hand, the value of P(S0) is
given by the product

P(S0) = P(S0|S1) x P(S1|S2) x P(S2|S3) x … and so on, ad infinitum:

Since all the factors in this product are less than one, Russell concludes that P(S0) “may be
expected to be zero” [15, p. 434]. However, as Hans Reichenbach first pointed out in a letter to
Russell on March 28, 1949, this argument is awed. For it is simply not true that the value of
P(S0) is given by the product above. Rather P(S0) is given by the rule of total probability

P(S0) = P(S0|S1)P(S1) + P(S0|~S1)P(~S1), (1)

where P(S1) is

P(S1) = P(S1|S2)P(S2) + P(S1|~S2)P(~S2), (2)

and P(~S1) is

P(~S1) = P(~S1|S2)P(S2) + P(~S1|~S2)P(~S2), (3)

and so on, ad infinitum. Since (1){(3) are sums rather than products, the fact that all the terms in
these equations are less than one does not imply that the outcome generally tends to zero. Two
weeks after having received Reichenbach's letter Russell replied, admitting his error.4

One may accept that P(S0) need not be zero, but still think that it cannot be computed. For how
should we compute an unconditional probability if the only thing we know is an infinite number
of conditional probabilities? This question takes us back to the finite mind objection. In the rest
of this paper, we nullify this objection by actually completing a particular infinite probabilistic
chain.

In our chain of propositions Sn, with n = 0; 1; 2; … etc., it is supposed that the conditional
probabilities αn and ẞn

αn = P(Sn|Sn+1)
ẞn = P(Sn|~Sn+1)

are known for all n. We may write

P(Sn) = αnP(Sn+1) + ẞnP(~Sn+1), (4)

where the unconditional probabilities P(Sn) and P(~Sn) are unknown. If we now write P(~Sn+1)
as 1 – P(Sn+1) and introduce the abbreviation
 γn = αn – ẞn = P(Sn|Sn+1) – P(Sn|~Sn+1),

then we can rewrite Eq.(4) as

P(Sn) = αnP(Sn+1) + ẞn – ẞnP(Sn+1) = ẞn + γnP(Sn+1). (5)

In order to calculate P(S0), we can now simply iterate (5), obtaining

P(S0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + … + γ0γ1 … γn–1ẞn + γ0γ1 … γnP(Sn+1). (6)

If P(Sn+1) were known, (6) could be used to calculate P(S0). But the difficulty is of course that
P(Sn+1) is not known. In contradistinction to the first n + 1 terms of (6), which are all
conditional probabilities and therefore known, the last term in (6), like the last term in (5),
contains the unknown unconditional probability P(Sn+1). So it would seem that we have come to
a dead end, and that old and new foundationalists have a point after all. The old ones claim that
the iteration can only be terminated by a supposed certainty, P(Sn+1) = 1; the new ones say that
a smaller value will serve, e.g. P(Sn+1) = 1/4 , but both insist that a definite value of P(Sn+1) is
necessary to enable one to calculate P(S0).

However, here is a counterexample that demonstrates the falsity of these foundationalist claims.
Elsewhere one of us has given a counterexample based on a geometrical series, and one based on
an exponential series [17]. The following example has the advantage that is much simpler than
either of these, in that the series reduces in the end to the sum of two terms only. Take

αn = 1 – 1/n+3 ẞn = 1/n+3, (7)

so that

γn = αn – ẞn = (1 – 1/n+3) – (1/n+3) = n+3/n+3 – 2/n+3 = n+1/n+3. (8)

In order to calculate (6), let us begin by evaluating parts of it, notably the last term and the
penultimate one. The last term is γ0γ1 … γnP(Sn+1) and the coefficient in this term can be
computed as

γ0γ1 … γn–1γn = 1/3 x 2/4 x 3/5 x 4/6 x … x n–1/n+1 x n/n+2 x n+1/n+3 = 2/(n+2)
(n+3). (9)

This is a very compact result, thanks to the cancellations between numerators and denominators
of the factors 3, 4, 5, … , n – 1; n; n + 1. The penultimate term in (6) can be written as

γ0γ1 … γn–1ẞn = 1/3 x 2/4 x .. x n–1/n+1 x n/n+2 x 1/n+3 = 2/(n+1)(n+2)(n+3) = 1/n+1

– 2/n+2 + 1/n+3. (10)

With the definition

δn = 1/n+1 – 1/n+2. (11)

one can check that

δn – δn+1 = 1/n+1 – 2/n+2 + 1/n+3. (12)

So from (10) and (12) it is clear that

γ0γ1 … γn–1ẞn = δn – δn+1. (13)

Plugging (9) and (13) into (6), we find

P(S0) = ẞ0 + (δ1–δ2) + (δ2–δ3) + … + ( δn – δn+1) + 2/(n+2)(n+3)P(Sn+1). (14)

Since δ2, δ3, … δn cancel out, (14) can be simplified to

P(S0) = ẞ0 + δ1 – δn+1 + 2P(Sn+1)/(n+2)(n+3). (15)

In the limit that n tends to infinity, (n+2)(n+3) in the denominator grows without bound.
Consequently the last term in (15) vanishes, since P(Sn+1), whatever it is, cannot be greater than
unity. Moreover, δn+1 disappears in the limit, as can be seen from (11). On taking the limit in
(15) we are left with

P(S0) = ẞ0 + δ1 = 1/3 + 1/6 = 1/2. (16)

This example undermines all claims that an infinite regress of probabilities cannot make sense,
or that it must always leads to zero, or that it can never be completed.

Acknowledgements. We thank two anonymous reviewers for their stimulating and useful
criticiscms.

1 The distinction between probabilistic and non-probabilistic foundationalism corresponds to

Bonjour's distinction between moderate and strong foundationalism [1, p. 26{
30]

2 Actually there is rather more. At several places Klein intimates that all justifications have this
structure, and that might be questionable.

3 For example, consider: “The supposition that the probability of anything whatever always
depends on something which is only probable itself, is flatly incompatible with the assignment of
any probability at all.” [14, p. 173]. This is easily probabilized to: “The supposition that the
unconditional probability of anything whatever always depends on something which is only
conditionally probable itself, is flatly incompatible with the assignment of any unconditional
probability at all.”

4 “... you are right as to the mathematical error that I committed ...” (Russell to Reichenbach in a
letter of April 22, 1949). We thank Carl Spadoni of the Mills Memorial Library, McMaster
University (Hamilton, Canada), for sending us a copy of the letter. C.I. Lewis remained
unconvinced, however. See [16] for a further analysis of the debate between Lewis, Reichenbach
and Russell.

On the Regress Argument for Infinitism

Synthese 166:1 (Jan 2009)
John Turri
Huron University College

Abstract: This paper critically evaluates the regress argument for infinitism. The dialectic is
essentially this. Peter Klein argues that only an infinitist can, without being dogmatic, enhance
the credibility of a questioned non-evident proposition. In response, I demonstrate that a
foundationalist can do this equally well. Furthermore, I explain how foundationalism can provide
for infinite chains of justification. I conclude that the regress argument for infinitism should not
convince us.

Infinitism offers a theory of justification as well as a solution to the epistemic regress problem.
Discussion of infinitism goes back at least to Aristotle.1 However, compared to its main non-
skeptical competitors, foundationalism and coherentism, its resources remain largely
underdeveloped, and its potential benefits, should there be any, unappreciated. This paper
critically evaluates one argument for infinitism: the regress argument. The dialectic is essentially
this. It has been argued that only an infinitist can, without being dogmatic, enhance the
credibility of a questioned non-evident proposition. I respond by demonstrating that a
foundationalist can do this equally well. Furthermore, I explain how foundationalism can provide
for infinite chains of justification.
Through a series of articles over the better part of the last decade, Peter Klein has done more
than anybody to revive the fortunes of infinitism, thrusting it back onto the philosophical scene.2
According to Klein, the regress problem “concerns the ability of reasoning to increase the
rational credibility of a questioned proposition.”

The Regress Problem can be put as follows: Which type of series of reasons and the
account of warrant associated with it, if any, can increase the credibility of a non-evident
proposition? Can a series with repeating propositions do so? Can one with a last member
do so? Can one that is non-repeating and has no last member do so?3

A non-evident proposition is one about which there could be “credible disagreement.” One sure
sign of a non-evident proposition is that our “epistemic peers could sincerely disagree with us
about [its] truth.” In such a case, the proposition lacks “the desired level of credibility” for us.
How might we, through reasoning, restore some of that credibility, once some measure of doubt
has set in? That is the key question for Klein, and what in this paper I shall refer to as the
regress problem.

Klein evaluates four main competing solutions: foundationalism, two forms of coherentism, and
infinitism. The test is whether a “self-conscious practitioner” of a theory can, once questioned by
an interlocutor, “employ” it to “increase the rational credibility” of a non-evident proposition. It
is important to note that, according to Klein, “it is not crucial what degree of credibility is at
stake.” The question is whether a self-conscious practitioner can enhance credibility at all.4 If a
self-conscious practitioner of a theory can do this, I shall say that the theory solves the regress
problem.

Klein’s thesis is that only infinitism can solve the regress problem. In order to establish his
thesis, Klein argues for two further theses: (1) neither coherentism nor foundationalism can solve
the regress problem, and (2) infinitism can solve the regress problem. Assuming that (3)
foundationalism, coherentism, and infinitism exhaust the alternatives, Klein's thesis follows. Call
the argument from 1 - 3 to Klein's thesis the regress argument for infinitism.

I will show that accepting what Klein says in defense of 2 should lead us to reject 1, and
consequently his overall argument. The reasons that Klein offers for thinking that infinitism can
solve the regress problem provide an equally adequate basis for concluding that foundationalism
can solve the regress problem, too, thereby casting doubt on 1.5

Klein distinguishes two sorts of coherentism, warrant-transfer and warrant-emergent. He rejects

warrant-transfer coherentism because it endorses question-begging circular reasoning as a way of
enhancing the credibility of a proposition, whereas surely any chain of reasoning that cites P as a
reason for believing P cannot increase the credibility of P.6 He rejects warrant-emergent
coherentism because it is just a “one-step” version of foundationalism, and foundationalism
cannot solve the regress problem. This places more weight on Klein's argument that
foundationalism cannot solve the regress problem.

According to Klein, the essence of foundationalism is the claim that warrant or justification
“aris[es] autonomously in so-called basic propositions and [is then] transferred to other
propositions through permissible forms of inference.”7 It should be noted that, in the course of
his discussion, Klein neither argues nor presupposes that foundationalism is incorrect in its
contention that some beliefs, properly basic beliefs, are “autonomously” warranted or justified,
where this positive epistemic status is achieved not by inference from further beliefs or reasons,
but instead accrues to the basic belief in virtue of its being F, where ‘F’ denotes the favored
foundationalist property that confers such status, whether it be indubitability, being caused by a
reliable process, etc.8

Klein tests foundationalism by means of an imaginary dialogue between a self-conscious

practicing foundationalist, Fred, and a doubting antagonist, Doris. I will now present a dialogue,
which is essentially modeled on Klein’s, but presented in a way most conducive to making my
main critical point shortly. In my dialogue, Fran is the foundationalist and Dan the doubter.

Before proceeding, I should note that many will object to Klein’s method here, on the grounds
that he has confused the property of being justified, or even that of enhancing justification, with
the process of showing that one is justified. Why think that explicit, reflective reasoning is the
only way to enhance one’s justification for a non-evident proposition, once it has been
questioned? Couldn’t the foundationalist reject this assumption, on the grounds that
foundationalism provides other ways of regaining credibility lost in the face of an aggressive
interlocutor? I am expressly setting aside all such worries here. For the sake of argument, I grant
that Klein is correct about the dialectical nature of justification.

Now for the dialogue.9 Fran says that the tree just off the ninth green is an elm. Let ‘P’ denote
the proposition that the tree just off the ninth green is an elm. Dan asks Fran something that
makes her believe that she had better have reasons for believing P. Fran responds by pointing out
that the tree has leaves characteristic of elms. Call this reason ‘R1’. Dan inquires about the status
of Fran’s acceptance of R1. Fran responds by citing the shape and texture of the leaves. Call this
reason ‘R2’. Dan inquires about the status of Fran’s acceptance of R2. This goes on until Fran, as
a self-conscious practicing foundationalist, reaches what she considers to be a properly basic
belief, B. Dan presses further. Fran resists at this point, claiming that B has foundational property
F, and so is properly basic, and thus “doesn’t need a reason.”10 Dan grants that B is F, and thus
properly basic, and not in need of a reason, but asks Fran whether properly basic beliefs are at
least likely to be true.

With respect to Dan’s last question, Fran has three options: affirm, deny, or withhold. If she
denies, then, although holding B is not arbitrary (remember, Klein is granting for the sake of
argument that B is a properly basic belief), using B as a reason for Rn is arbitrary. If she
withholds, then, from her own point of view, she ought not to use B as the basis for further
beliefs. If it is not good enough to affirm in and of itself, then how could it be proper to deploy it
as a basis for affirming something else? Neither of those options, then, allows Fran to regain any
of P’s lost credibility. What if she affirms? If she affirms, then B non-arbitrarily can serve as a
reason Rn, but only because the regress has continued. Fran has a very good reason for holding
B, namely, B has F and F-beliefs are likely to be true. It would now be legitimate to ask Fran to
offer a reason for thinking that B is F, and a reason for thinking that F-beliefs are likely to be
true. Therefore, “a practicing foundationalist cannot increase the rational credibility of a
questioned proposition through reasoning.”11
Before proceeding further, let us pause to take stock of the progress Fran made before answering
Dan’s last question. Fran had articulated the following justificatory structure:

B → Rn → … → R2 → R1 → P

in support of the original non-evident proposition P. It is consistent with the statement of

foundationalism that ‘n’ in ‘Rn’ be arbitrarily finitely large. (Which is not to say that it would
necessarily be inconsistent with foundationalism if ‘n’ were infinite; more on this below.)

We want a view that endorses neither circular reasoning nor arbitrariness. Yet, Klein maintains,
there is something to the foundationalist idea that some propositions are epistemically prior to
others, and something to the warrant-emergent coherentist idea that warrant emerges from a
proposition’s being embedded in an appropriate set of supporting propositions. Klein’s version
of infinitism (which he calls “warrant-emergent infinitism,” but which I shall continue to call
simply “infinitism”) purportedly speaks to all these needs. It avoids vicious circularity because it
forbids a proposition from appearing in its own evidential ancestry or series of supporting
reasons; it avoids arbitrariness because it says that, for each reason r in a series supporting a
belief, there is a further reason, available to the subject, supporting r.

But can a practicing infinitist increase the credibility of a questioned non-evident proposition?
Unless a practicing infinitist can do this, then, at least with respect to the regress problem,
infinitism is no better off than foundationalism. This brings me to my main point, which takes
shape in light of Klein’s response to an objection against infinitism raised by Sextus Empiricus. I
will argue that Klein’s response to Sextus reveals that, by Klein’s own lights, the practicing
foundationalist can, and in Fran’s case does, increase the credibility of a questioned non-evident
proposition.

I first quote the passage from Sextus and then Klein’s response at length:

[Sextus says:] The Mode based upon regress ad infinitum is that whereby we assert that
the thing adduced as proof of the matter proposed needs a further proof, and this again
another, and so on ad infinitum, so that the consequence is suspension [of judgment], as
we possess no starting-point for our argument.12

[Klein responds:] Now, if efficacious reasoning required that warrant originate in and be
transferred from a basic proposition, this criticism would be just. But for the reasons
given above, infinitism eschews such a view. The ‘starting point’ of reasoning is, as
Peirce says: doubt. A proposition becomes questionable and, consequently, it lacks the
desired rational credibility. Reasoning scratches the itch. The infinitist holds that finding
a reason for the questioned proposition, and then another for that reason, etc., places it at
the beginning of a series of propositions each of which gains warrant and rational
credibility by being part of the series. Warrant increases not because we are getting
closer to a basic proposition but rather because we are getting further from the
questioned proposition. But the Pyrrhonist is correct that the infinitist’s conception of
reasoning precludes assenting to a non-evident proposition. Dogmatism is incompatible
 with practicing infinitism. Warrant, and with it rational credibility, increases as the
series lengthens; but the matter is never completely settled.13

Warrant and rational credibility increase because we get further from the questioned proposition.
Rational credibility increases as the series of reasons offered in support lengthens. Recall Fran’s
progress. She started with P. Then she adduced R1. Then she adduced R2. And she kept on going
until she got to B. She lengthened the series supporting P, and thus increased the rational
credibility of P. Therefore, the self-practicing foundationalist (qua self-practicing
foundationalist) can enhance the credibility of a non-evident proposition. Moreover, consistent
with foundationalism, the length of the series can increase to any arbitrarily large finite
cardinality, so the prospects for increasing it to the level required for knowledge are at least
hopeful. As Klein puts it, “As the series lengthens, warrant and credibility increase. Nothing
prevents it increasing to the degree required for knowledge,” since it is implausible “that
knowledge requires the highest possible degree of warrant or absolutely credible belief,” which
only an infinitely long series could provide.14

Does the fact that Fran’s reasoning would eventually end with B, were she to continue, wreck the
supposed enhancement in justification? Nothing in Klein’s response to Sextus indicates that it
would. Klein does not take it as a premise of his reasoning that a chain of reasons must be
infinite to do its epistemological work; rather, that is his conclusion. And clearly it would be
unhelpful to assume the thesis of infinitism in order to argue for it.

Perhaps the fact that, as Klein puts it, practicing infinitism is incompatible with dogmatism is
enough to set it apart from foundationalism? Grant that falling into dogmatism is an intellectual
vice incompatible with enhancing the credibility of a questioned non-evident proposition. This
would aid Klein’s argument only if practicing foundationalism was incompatible with not falling
into dogmatism. But foundationalism need not lead to dogmatism: nothing prevents Fran from
practicing a version of foundationalism, according to which such matters are “never completely
settled.”15

I conclude that Klein’s regress argument for infinitism ought not to convince us. By the standard
in play, foundationalism provides just as satisfactory a solution to the regress problem as does
infinitism. Of course, some will object to Klein’s standards; that is, they will deny that merely
increasing the length of a series of reasons adduced to support the original belief can, all by
itself, enhance the justification (or warrant, or rational credibility) with which the original belief
is held. Let me be clear: I am not defending the standard in question. My point is that such a
standard provides no basis for preferring infinitism to foundationalism.

At this stage it would be natural to wonder, is warrant (or justification, or credibility) a matter of
having more and more reasons, as Klein insists, or a matter of basing your belief in non-evident
propositions on properly basic beliefs, as the foundationalist claims?16 But the two options are
not incompatible.17 A series of reasons, supporting a belief in a questioned non-evident
proposition, need not have a last member in order to have a foundational (properly basic)
member, as I shall presently demonstrate.
Foundationalism is consistent with there being available to the subject an infinite, non-repeating
series of reasons, of the sort the infinitist prizes. In other words, foundationalists needn’t limit
themselves to finite series of reasons.18 Suppose that Fran sees that it is 2:05. Suppose further that
she practices a version of foundationalism according to which such external-world beliefs can be
properly basic, and that in the present case her belief that it is 2:05 satisfies all the relevant
criteria. Now Fran asserts that it is past 2:00. Dan asks Fran something, which makes her think
that she had better have reasons for believing that it is past 2:00. She responds that she believes it
is past 2:00 because it is past 2:04. Dan presses further: why think that it is past 2:04? Because it
is past 2:04:30. Why think that? Because it is past 2:04:45. Why think that? Because it is past
2:04:52.5. Why think that? Because it is past 2:04:56.25. And so on. Each time Dan makes a
further request, Fran will respond by inching halfway closer to 2:05. But proceeding this way
ensures that she will approach the limit of, but never arrive at, 2:05. In other words, she has
available to her an infinite series of non-repeating reasons, each of which is entailed by its
successor. Moreover, the foundationalist has a principled story to tell about how each member of
this infinite series gets justified for her: namely, she can see that it is 2:05!19

of argument that dogmatism is a vice, non-dogmatic foundationalism remains a viable

alternative.

1 Posterior Analytics, 72b5 ff.

2 See, e.g., Klein (1998), (1999), and (2005a). See also Fantl (2003).

3 Klein (2005a, 132). Elsewhere Klein puts matters differently. “I will explain how the infinitist
can provide a solution to the epistemic regress problem[,] which is, as I see it, how to avoid
arbitrariness in belief without falling into dogmatism.” See Klein (2005b, 154).

4 Klein (2005a, 132).

5 It is worth noting that we need not completely answer the important question “just what makes
one proposition or belief a (good) reason for another?” in order to evaluate the regress argument
for infinitism. The argument purports to establish a certain formal constraint on good reasons: a
good reason must be supported by an infinite and non-repeating series of further reasons. Settling
the matter at hand thus goes some but not all the way toward a theory of good reasons.

A complete theory would settle whether there are any further substantive requirements on good
reasons. A complete theory far surpasses what I am capable of in the present paper, as it would
require, among other things, evaluating a vast range of arguments for and against
epistemological externalism, as well as addressing whether there are any higher-order
requirements on epistemically appropriate belief, e.g., whether and in what sense a subject must
believe that the reason for her belief is adequate in order for it to be adequate. (Thanks to an
anonymous referee for alerting me that this issue might be on readers’ minds.)

6 Klein (2005a, 134).

7 Klein (2005a, 132).

8 Klein (2005a, 133). Here we can see why Klein considers warrant-emergent coherentism to be
a thinly disguised version of foundationalism: “In this case, the foundational property, F, which
all warranted propositions have, is that each is a member of a set of coherent propositions,” Klein
(2005a, 135).

9 Much of this is very close paraphrase of Klein’s dialogue. However, some is also taken
verbatim from Klein’s dialogue, and, except where it seemed absolutely critical that the reader be
aware that I’m using Klein’s exact words, I don’t use quotation marks to indicate where I have
used Klein’s exact words. I’ve changed some names and variables in the dialogue, which
rendered the intermittent use of quotation marks and bracketed material both cumbersome and
distracting. Let this footnote stand as official attribution, to Klein, of all intellectual credit for the
structure, tone, and basic wording of the dialogue.

10 I doubt that foundationalism need be committed to the claim that properly basic beliefs don’t
stand in need of reasons. This is an artifact of Klein’s characterization of foundationalism, and I
will not stop to question its accuracy, since it is tangential to my project in this paper. Elsewhere,
Klein (2004, 168) remarks, “I take it that foundationalists hold, at a minimum, that some
propositions are prima facie justified in the absence of further reasons.”

11 Here ends the close paraphrase of Klein’s dialogue.

12 Quoted in Klein (2005a, 137).

13 Klein (2005a, 137 – 138), emphasis added.

14 Klein (2005a, 138).

15 Some might wonder what implications this has for versions of foundationalism that allow for
matters to be completely settled. Call such a view dogmatic foundationalism. (Not to be confused
with James Pryor’s (2000) foundationalist dogmatism.) I do not wish to rule out viable dogmatic
foundationalist views, and my argument here does not require that I do so. My point is that, even
granting for the sake

16 Thanks to an anonymous referee for putting this question to me.

17 It is worth remembering that we here focus on a dialectical conception of epistemic merit.

Many foundationalists concern themselves with a thoroughly non-dialectical brand of epistemic
merit; see, e.g., Pryor (2004).

18 Ben Fiedor and Ernest Sosa independently suggested in conversation examples similar to the
following one.

19 For feedback and discussion that helped improve this paper, I thank Ben Fiedor, Peter Klein,
Ernest Sosa, and two anonymous referees.
Synthese (2008) 163:175–185
Meta-epistemology and the varieties of epistemic infinitism
Scott F. Aikin

Abstract I will assume here the defenses of epistemic infinitism are adequate and inquire as to
the variety standpoints within the view. I will argue that infinitism has three varieties depending
on the strength of demandingness of the infinitist requirement and the purity of its conception of
epistemic justification, each of which I will term strong pure, strong impure, and weak impure
infinitisms. Further, I will argue that impure infinitisms have the dialectical advantage.
Keywords Epistemology · Infinitism · Foundationalism · Regress problem

Epistemic Infinitism is an underdeveloped view. On the old epistemic regress problem, infinitism
is the view that one can solve the problem by giving (or having) reasons on to infinity. Its prima
facie implausibility surely explains this lack of development—so few have been willing to go to
the mat for the thesis. Some have recently defended infinitism against both the initial
implausibility problem and actual arguments against it—notably, Peter Klein (1998, 1999, 2003),
Jeremy Fantl (2003), and myself (2005). I will proceed here as though these defenses are
adequate for the time being and inquire as to the variety standpoints within the view. I will argue
that infinitism has at least three varieties depending on the strength of demandingness of the
infinitist requirement and the purity of its conception of epistemic justification, each of which I
will term strong pure, strong impure, and weak impure infinitisms. Further, I will argue that
impure infinitisms have the dialectical advantage.
Before proceeding with infinitism, some theoretical apparatus must be assembled. Two
distinctions are necessary. The first is between pure and impure meta-epistemic theories. Pure
meta-epistemic theories are commitments to the exclusivity of one source or formal structure of
justification. So pure foundationalism, for example, is the commitment to basic beliefs being the
source of all epistemic justification.1 Pure coherentism is the view that coherence (of the right
kind) is the only source.2 Impure epistemic theories aremore ecumenical—there may be more
than one formal structure for justification. Some impure foundationalisms, for example, allow
both basic beliefs and perhaps systematic coherence to be sources for justification.3 Impure
coherentisms, too, may allow for basic beliefs to function alongside coherent sets of beliefs.4 In
fact, with some impure theories, the distinction between foundationalist versions and coherentist
versions seems more a matter of emphasis than a real theoretical difference.5

The dialectical factor counting in the favor of pure epistemic theories is that the regress problem
is a powerful tool to motivate a theory, since the classical version of the problem generally yields
at most five or six possible solutions (skepticism, foundationalism, coherentism, externalism,
basic unjustified beliefs, and infinitism). The purist then knocks off the relevant challenges and
then proceeds to elaborate her own view. The case for the purist’s view is made, then, on the
basis of the hexalemma (for six choices) where five of the choices are either refuted or
reasonably presumed out of bounds. It makes for neat scholarly work.

The dialectical situation for those defending an impure epistemic theory is that the history of
exchanges between purists has yielded some very good arguments against pretty much every
view. Some of these arguments proceed from prima facie cases of knowledge to showing how
competitors cannot countenance the cases, but the favored theory can. So foundationalists hold
out their hands, coherentists hold out interpretations and explanations. What impure theorists do
with the piecemeal of intuitive cases of knowledge is cobble together a systematic view of
knowledge that allows a variety of sources of justification. With some things, only basic beliefs
can do the work, and with other things, only epistemic relations between beliefs can do the work.
As such, the impure theorist has fewer bullets to bite dialectically, and she proceeds in a fashion
of systematically saving the appearances.

The second distinction is between the strong and weak demandingness of a meta-epistemic view.
Justification trees (J-trees) are an easy way to present this difference. A J-tree for some subject
S’s belief that p is a graphic representation of the justifying reasons S has for p. So if S believes
that p on the basis of q and r, and S believes q on the basis of s and r on t, S’s J-tree for p would
have two branches, with q and r as nodes on one level and s and t on nodes at a lower level:
p
/ \
q r
/ \
s t

We could ask further question about s and t, and depending on which epistemic theory is right, S
may in the end have a basic belief (one for which no more beliefs are necessary) or some
epistemic relation of coherence with the right set of beliefs as the justifiers. From the earlier
distinction between pure and impure theories, a pure foundationalism is the view that all J-trees
have only basic beliefs at their bottom nodes. Pure coherentism is that all final nodes bear
epistemic relations of coherence with the relevant beliefs.

Strong epistemic theories run that all J-trees must have at least one branch end with the relevant
epistemic type of belief. These beliefs are necessary for any J-tree. Weak epistemic theories hold
that some J-trees have at least one branch end with the relevant type of belief.

Strong epistemic foundationalism is the view that every J-tree has at least one basic belief. Weak
epistemic foundationalism is the view that there are J-trees with some basic beliefs. Basic beliefs
yield justification for weak theories, but they are not necessary for all J-trees. There are four
possibilities for combination for these meta-epistemic theories, then. Let us use foundationalism
as the exemplar:

• Strong pure foundationalism—basic beliefs are the only source of justification, and all J-trees
have only basic beliefs as their terminating nodes.

• Weak pure foundationalism—basic beliefs are the only source for justification, and some J-
trees have basic beliefs as their terminating nodes.

• Strong impure foundationalism—basic beliefs are not the only source of justification, but all J-
trees have at least one basic belief as a terminating node.

• Weak impure foundationalism — basic beliefs are not the only source of justification, but there
are at least some J-trees with basic beliefs as terminating nodes.

Clearly weak pure foundationalism is not a real possibility, since some of the J-trees with basic
beliefs (because of the exclusivity of basic beliefs yielding justification, on the purity
commitment) will in fact be all of the J-trees. This, then, yields a strong pure foundationalism.
So all pure meta-epistemic theories are strong. But impure meta-epistemic theories may either be
weak or strong. Strong impure foundationalism is the view that though there may be other
sources of justification, basic beliefs are still necessary for any J-tree. Basic beliefs are not
exclusive of justification, but are necessary. This distinction is particularly useful in light of the
structural arguments foundationalists make against coherentists. For example, Paul Moser (1984)
argues that coherentism cannot independently solve the regress problem, since assessments of
coherence cannot themselves be based on coherence. Intuitive or basic beliefs are necessary.
This argument may be correct, but it only is a case for strong foundationalism (as the
requirement that basic beliefs must play a role in all J-trees), but it does not necessarily count
against an impure coherentist commitment to coherence being a source of justification. As such,
Moser’s arguments clearly count against pure coherentisms, but they cannot yet clear the field.
Finally, weak impure foundationalism is the view that there are not only other sources of
justification, but there may even be J-trees without any basic beliefs as their terminating nodes.
Basic beliefs are productive of, but they are neither exclusive nor necessary for justification.

On these three combinations of strengths of demandingness and exclusivity of metaepistemic

theories, three correlate possibilities arise for epistemic infinitism:
• Strong pure epistemic infinitism—all branches on J-trees are infinitely extended. There are no
other sources of justification other than infinite chains of inference.

• Strong impure epistemic infitinitism—at least one branch on every J-tree must be infinitely
extended. There other sources of justification, but infinitely extended chains of inference are
necessary for any J-tree.

• Weak impure epistemic infinitism—Infinite chains of inference are productive of justification,

but they are not necessary for all J-trees.

Peter Klein’s infinitism, outlined in “Human Knowledge and the Infinite Regress of Reasons”
(1999), is a pure epistemic theory, as he holds infinitism’s “central thesis is the structure of
justificatory reasons is infinite and non-repeating” (1999, 297). Klein proceeds precicely from
the dialectical position of the hexalemma and eliminates foundationalism, coherentism,
unjustified foundations, and reliablism in order to motivate his view.6 Klein’s case against the
alternatives proceeds from two intuitive principles of good reasoning: the Principle of Avoiding
Circularity, and the Principle of Avoiding Arbitrariness. Together, they amount to the
requirement that subjects must have reasons as a whole that are non-circular and that subjects
must for each reason individually have reasons they can give in its favor. Klein’s argument is
that these requirements as conditions for rational believing stand as exclusive requirements in
favor of infinitism, because the alternatives clearly run afoul of one or two of them and infinitism
runs afoul of neither. So he concludes: “The infinitist is claiming that a normatively acceptable
set of reasons must be infinitely long and non-repeating if we are to avoid the pitfalls of
foundationalism (arbitrariness) and coherentism (begging the question)” (1998, 222).

Jeremy Fantl’s argument for infinitism in “Modest Infinitism” (2003) is also exclusivist, and is
as such, a strong pure theory. Fantl similarly introduces two requirements for a theory of
justification: that the theory account for justification coming in degrees and that complete
justification (or justification simpliciter, full justification, or justification of the highest degree)
makes sense. Fantl then proceeds to show that neither foundationalism nor coherentism can live
up to these requirements. Foundationalism fails because the traditional version of the view seems
to make all basic beliefs justified in virtue of their truth, which makes them all fully justified, and
as such, justification does not admit of degrees.7 Alternately, if foundationalism allows
incompletely justified beliefs to serve as basic, then foundationalism cannot explain the
possibility of complete justification. Coherentism fails because it is not, at least on internalist
requirements, capable of yielding complete justification (2003, 541).

That is, if the justification for some S’s belief that p was p’s coherence with the most coherent,
comprehensive, and explanatory group of beliefs, then surely the belief that p belongs
would improve the justification for p. But on the one hand, for a set of beliefs to satisfy these
requirements, it may turn out that there must be an infinite number of them. And on the other
hand, surely the belief that a proposition coheres with that set must itself be justified, and
that one also, for complete justification. As such, for coherentism to meet the completeness
requirement, it must be a tacit infinitism (2003, 546).
My own version of infinitism in “Who Is Afraid of Epistemology’s Regress Problem?” (2005) is
an impure version of the view, as I describe the infinitism I defend as “consistent with modest
epistemic foundationalism” (2005, 191). I introduce the foundationalist conception of
justification in order to answer the question as to how infinitism itself can eliminate arbitrary but
infinite chains of modus ponens inferences from the class of justifiers. The challenge is
formulated as what I call “the modus ponens reductio” (MPR) of infinitism. My strategy is to
concede that “beliefs based on (some acceptable set of) non-doxastic states are innocent until
proven guilty” (200), and show how infinite chains must be supplemented by non-doxastic
support. The non-doxastic support for these inferential chains counts as prima facie evidence that
they are correct (as opposed to the arbitrary ones). My view is clearly ecumenical enough to
qualify as impure, but it is unclear whether it is strong or weak, given the contents of the essay.
However, it is clear, from how I solve the arbitrariness problem with the MPR (by requiring
basic beliefs in every J-tree), my view is certainly a strong impure foundationalism in the
requirement of basic beliefs. With regard to infinitism, I lean toward strong infinitism, since even
for connecting appearances with beliefs about the external world, I require infinite inferential
series (201). But this requirement is one that only extends to fallible external world beliefs. It
still, on my account, is an open question as to whether there can be beliefs completely justified
by non-doxastic contents only, which would yield a weak impure infinitism.

It is clear, given the arguments both Klein and Fantl have against infinitism’s meta-epistemic
competitors in favor of pure infinitism, the dialectical burden is on the impure infinitist to answer
their challenges. That is, if the impure theory allows in sources of justification that Klein and
Fantl have shown inconsistent with infinitism, then it is incumbent on the impure infinitist to
show how the alternatives are in fact consistent with infinitism. What follows is a sketch of such
a defense.

Klein’s argument against foundationalism is that it either runs afoul of the arbitrariness
requirement or it is in fact infinitist. If a belief B is taken to be foundational, it must be so on the
basis of some property F of B. But what makes B’s having F a reason for it to be a basic belief?
How is F related, say, to B’s truth? Either the foundationalist answers the question (and the
regress continues, which means that foundationalism is on the road to infinitism) or no answers
are coming (and the basic belief, from the perspective of the reasoner, is an arbitrary foundation).
For responsible reasoners, “the meta-justification is invoked in order to avoid the appearance of
arbitrariness for it is designed to show why the ‘final’ beliefs are likely true” (1999, 304). The
point of Klein’s dilemma here is that foundationalism as a meta-epistemic theory cannot solve
the structural problem of the regress. Basic beliefs are held as basic with reasons or without, and
the former is a de facto failure to end the regress of reasons, and the latter is a de jure failure to
end the regress. It should be noted that Klein’s dilemma, as formulated, does not presume that a
subject must be justified in her assessments of her current justification, since the regress is not
regarding whether the subject is justified in taking B to be justified but regarding whether F is a
reason for holding B. That is, F (or B’s having F) stands as a reason for B, and so a further
question can be asked as to how F in fact justifies B. So the dilemma is not dependent on some
version of the JJ principle, but is a first order question about B’s justification.

A foundationalist response to Klein’s dilemma is that surely being able to explain the connection
between B’s property F and B’s truth is requisite for B’s justification, but that explanation itself
is not a continuation of the regress of jusitification. Take the old introspective belief held by
some subject, S:

(H) I have a headache.

I take it that some S would be justified in holding H if she did so on the basis of feeling
headachy. So F for H would be the basing relation between H and the subject’s awareness of her
non-doxastic state. It seems right that if S has got the conceptual tools to formulate H and for it
to be a belief held by a responsible human adult, our subject should be able to answer further
questions about H and F. But this ability is a condition for H being basic for S; it itself does not
comprise reasons for H. If asked, “how do you know that you have a headache?”, S would
probably respond that she feels it. Or, in classic Ducassian fashion, she would simply repeat H
(Ducasse 1944, 339). Klein’s default is to interpret this as either arbitrary or perhaps circular
reasoning. But on this reading of F, restating H is appropriate, because with the restatement S is
elliptically stating something along the lines of:

(H–F) I believe I have a headache on the basis of the fact that I have a headache.

On the Ducassian model, restating the belief on the challenge is a stand-in for articulating the
fact with which one is acquainted that comprises the belief. But more can be said, especially if
one is a good (or even passable) foundationalist. One strategy would be to turn the tables on the
Kleinian skeptic-infinitist and ask on the basis of what anyone would challenge H and perhaps
explain that these attitudes have default status, or explain that headaches are precisely the kind of
things that if you feel like you’ve got one, you’ve got one. But regardless of what S says in her
defense, these reasons may be indicative of how she is a responsive and capable reasoner, but
they are not justifiers for H. S can explain why H is a basic belief for her in showing that she is a
capable reasoner, but showing that she knows what she is doing when she avows H is not a
justification for H in the sense that it would serve as a set of reasons for H. Rather, they are
abilities S must be able to display for her to hold any kind of belief with the semantic contents of
H—she must know something about how headaches feel and how one might get them, have
some tacit familiarity with the default status of first person psychological statements, etc. Being
able to answer questions about H and F is clearly a condition for S to be justified in holding H
and answering them may in fact improve H’s epistemic status, but this does not mean that they
are justifiers for H. S’s independent justification for H should be just her being aware of her
headache.8 A necessary condition for her to be justified in holding H (perhaps conditions for her
to formulate H at all) would be that she is able to responsively regard her non-doxastic
psychological states.9 She can demonstrate that ability by answering questions correctly, but for
H to be justified for her, she must not only be able to answer these questions, she must have that
antecedent awareness of her headache. Showing that she can recognize a headache or that she
can defend some sense of priviledged access does not justify S in believing H, because without
the antecedent awareness of the headache S would not be justified in H. So Klein is right that the
meta-regress begins a new regress for justification, but it does not begin the original regress
anew.
As such, if this distinction between the doxastic conditions for being justified and the
nondoxastic justifiers is right, then Klein’s meta-regress argument may show the unavoidability
of infinitism, but it does not by necessity eliminate foundationalism.

Fantl’s argument against foundationalism is that the foundationalist takes basic beliefs to be
either self-justifying or in need of a meta-justification. He terms the first option ‘traditional
foundationalism’ and the second ‘meta-justificatory’ foundationalism. The traditional
foundationalist, on the one hand, faces a dilemma: either basic beliefs are fully justified or not. If
they are, then foundationalism cannot explain how justification admits of degrees. If basic beliefs
are not all fully justified (for example, the belief about the number of speckles on a hen. . . as
they reach 20 or so, our confidence would surely drop below complete confidence), then they
must be justified by something other than themselves. Otherwise, they would all have the same
justification (2003, 544).

Meta-justificatory foundationalists, on the other hand, cannot satisfy the completeness

requirement without becoming de facto infinitists (2003, 546). That is, if no belief B can be
completely justified without some feature F, then one will be in a better position to hold the
belief if one had a reason to hold that B had F. Of course, one would be in a much better position
if that belief itself were completely justified, which requires another reason, which itself should
be completely justified. . . and so on.

Fantl’s argument that infinitism is a consequence of meta-justificatory foundationalism is posited

on the same confusion behind Klein’s meta-justificatory argument—namely taking a subject’s
ability to answer challenges to and questions about a basic belief for a continuation of reasons
supporting the belief. Take, again, some subject’s belief that she has a headache (H). She could
not be in any better position to assess its truth than when she focuses her attention on how her
head feels. Her relative ability to answer questions about that privileged position does not
amount to an argument for H, but rather is evidence that she know what she is doing when she
avows it. Again, a subject’s ability to answer meta-challenges to basic beliefs is necessary for
their being justified for S as a reasonable and responsible believer, but they do not constitute
reasons for the belief—they are abilities that make up S’s knowing what she is doing when she
holds the belief. The consequence, then, is that though challenges to S’s belief that H as basic
and S’s defenses may go on to infinity, these do not undermine H’s status as basic, since they are
not arguments constitutive of the epistemic relation between S’s belief and her non-doxastic
state. So Fantl, like Klein, confuses the possibility of separate chains of argument being possible
with continuing the original chain.

Regarding cases like a subject believing that she has a headache, Fantl concedes that further
reasons are not necessary for such a proposition to be justified to a very high degree, nor are
further reasons necessary for the proposition to justified to a sufficient degree to count as
knowledge. However, these propositions cannot be fully justified for this subject because:

I also recognize that there are many difficulties with introspection. . ., including
difficulties that make fallibility in introspection a somewhat plausible consequence.
Given this, it’s hard to see why, if you can have as an adequate reason that, how and why
 introspection works, and in particular, that it worked well in this instance, the degree to
which ‘I have a headache’ is justified for you wouldn’t increase (2003, 561).

As such, Fantl holds, in the end, that in light of a reasonable fallibilism about introspection (and
other sources of information—rational insight and ordinary perception), no human can ever have
complete justification, because all of the worries about whatever source of information can never
be completely settled. Only infinite inquiry could do that. So foundationalism, if it allows basic
beliefs to be less than completely justified, cannot show how complete justification is possible.

First note that Fantl has conceded that there can be other sources of justification, though they do
not by themselves yield complete justification.10 It seems that this should have been his main
argument—fallible (or at least purportedly fallible) sources of justification cannot yield complete
justification. But putting things this way trivializes the argument, because it is precisely this
constraint on a theory of justification that most fallibilist theorists have been trying to do without.
Fallibilist theories of justification are designed precisely to make due without the concept of
complete justification, because of its either unattainability or its incoherence.11 Fantl, here,
argues that infinitism is still preferable even in if complete justification is incoherent, because it
still makes sense of the thought that for any degree of justification yielded by a fallible source,
there is a higher degree which is attainable with further reasons (perhaps eliminating pressing
defeaters). But, again, Fantl’s point holds fallibilists to standards they reject, since the point of
fallibilist theories is to find legitimate stopping places in these exchanges with the full awareness
that they can continue endlessly. If that is the case, Fantl is wrong that infinitism is preferable to
foundationalism, because his infinitism is only different in emphasis from modest
foundationalisms.

Further, fallibilism is not necessarily motivated by the thought that complete justification is
incoherent or unattainable, but rather it is motivated by the thought that complete justification is
not necessary for knowledge. The standard we should hold theories of justification to is that in
the service of being adequate to analyze knowledge. So if complete justification is not necessary
for knowledge, then it is an excessive requirement for theories of justification.

A final point is that Fantl’s argument that traditional foundationalism, if it requires that basic
beliefs are completely justified, cannot countenance degrees of justification clearly misses its
target. Fantl’s dilemma for basic beliefs is that either they are all fully or not all fully justified.
He holds that the consequence of holding that all basic beliefs are fully justified is that one
cannot then explain how justification comes in degrees. But surely this is not correct. Fantl
requires that foundationalists address the degree requirement only with basic beliefs: “This is on
the initial assumption that foundationalism can only satisfy the requirements by specifying the
way the foundational reasons should be treated” (2003, 545). But many traditional
foundationalists explain degrees of justification by appealing to weaker than deductive
inferential support yielded by basic beliefs. So, C.I. Lewis’s pragmatic foundationalism had
certainties as terminating judgments, but their epistemic support for the superstructure of the rest
of human knowledge was probabilistic (1946). Timothy McGrew, a modern-day Lewisian, has
subject’s basic beliefs justified by their indexical relation to the subject’s non-doxastic
psychological states, but beliefs about the world and so on are abductively justified (1995). One,
in fact, does not have to go any further than Descartes himself, the paradigmatic traditional
foundationalist, who held that one may nevertheless err in making non-deductive inferences from
clear and distinct perceptions, because we are not omniscient (Descartes 1984, CSM II 54–55;
AT VII 78–79). Traditional foundationalism can answer Fantl’s two requirements, and as such, it
is still a viable meta-epistemic theory.

I propose, then, that an updating of the dialectical situation within infinitism is necessary. Klein
and Fantl’s arguments for the exclusivity of infinitism with regard to the structure of reasons fail
with foundationalism. There may be other arguments against foundationalism that succeed on
infinitist grounds, but none are yet on the table. This, however, does not mean that infinitism is
wrong, but that the proponents of the pure versions of the view must either develop new
arguments or turn ecumenical with impure versions of the view.

There is a dialectical consideration that further distinguishes pure and impure infintitisms in what
I earlier (2005) termed the modus ponens reductio. This objection to infinitism amounts to the
argument that infinite chains of reasons are, when considered as mere relations between beliefs,
arbitrary. The thought was captured early by Max Deutscher:

Could it be one vast delusion system? Is a man reasonable in holding one belief merely
because he holds another whose propositional content is suitably related to the first, even
if he holds the second on account of a third which is suitably related to the second, and so
on? Might not a man just dream up a system and be ingenious enough to always extend
his story in logical fashion? How can the mere continuous extension of a belief system
guarantee the rationality of the members of the system? (1973, 6)

We can make good on this worry by formulating the following way of extending reason-giving
in Deutscher’s ‘purely logical fashion.’ Let the following beliefs stand in the support relation
such that B1 is supported by B2, and B2 by B3, and so on.

B1: p

B2: q &(q ⊃ p)

B3: r &(r ⊃ (q &(q ⊃ p)))

B4: . . . .

We can see clearly that if B2 is justified, it entails B1, so B1 is justified so long as B2 is. And B2
is justified by B3, and B3 by B4, and so on. So long as modus ponens is truth preservative and
epistemic justification follows what are intuitively truth indicative inferences (these are at least
plausible assumptions), p, on the infinitist thesis, is justified. But now note that if we change p to
not-p, we get the following set of beliefs:

B1’: ∼p

B2’: q &(q⊃∼p)
 B3’: r (r ⊃ (q &(q⊃∼p)

B4’: . . . .

The problem is that on the infinitist theory, we have no rational way of telling the difference
between one justifying set of beliefs that is conducive of truth and one that is not. Infinitist
justification, precisely because it is defined exclusively in terms of relations between beliefs,
looses its moorings with truth. Epistemic justification that does so loses the name.12

But here the issue can be resolved by a question as to whether the sets of beliefs (B1. . .Bn) and
(B1’. . . Bn’), if they are unadjudicable between their claims on the truth of p and ∼p, are the
only necessary components for a justificatory story. If they cannot be rationally adjudicated by
their own contents and inferential relations, then ex hypothesi, there must be some other feature
that is necessary. Insofar as any J-tree will be subject to the same structural issue (which is
necessary, given the syntax of J-trees), only beliefs not in need of J-trees can play the role of
rationally adjudicating the two set’s claims on truth. There, if there are infinitely iterating
branches of reasons, must also be terminating nodes on those branches, too. So, the difference
between the two sets can be captured between the cognitive situation for some subject between
B2 and B2’:

B2 : q &(q⊃p)

B2’: q &(q⊃∼p)

Both simplify to:

B2a: q⊃p

B2’a: q⊃∼p

Given that J-trees are supposed to be snapshots of a subject’s beliefs, and that a condition for
belief is understanding, any subject caught in this dilemma must be capable of understanding the
difference between these two conditionals. One may ask our subject, “When q, not p? Are you
sure?” A subject’s understanding of the propositions may, it seems, provide immediate
justification for accepting one conditional over another, say in cases of their relation being
relations like the following:

• Analytic (or logical) entailment: If Sam’s a bachelor, then he’s male. If something is blue, it’s
colored. If something has properties F and G, it has property F.

• Synthetic a priori support: If Sally is elderly, Sally is old. If something is a physical object, that
thing has spatial properties.

• Morally or practically intuitive connections: If you promised to take out the trash, you should.
If someone’s elderly, you shouldn’t torture her for fun. If someone gives you something, ceteris
paribus you should express gratitude.
• Experiential connections: If you hold your hand up in front of your face with your eyes open
and in good light, you’ll be appeared to your-hand-in-front-of-your-face-ly. If you drink a lot of
beer, you’ll feel light headed (at least).

Simply change all the above conditionals so that their consequents are negated, and we can tell
the difference epistemically between B2a and B2’a when they are formulated as such. We may
need further arguments to fix the consequent as justified, but we can tell the difference between
and even assign justificatory status to one or the other conditional depending on our
understanding and experience relevant to these propositions. If the foundationalist thesis is
broadly right that understanding, moral sense, and experience provide at least a prima facie
epistemic difference between propositions, we have a way of rationally adjudicating the two
competing sets. The conclusion, then, is that impure infinitism is dialectically preferable, since
the subjects considering these series of inferentially related beliefs have a means of deciding for
themselves which story is right. And that means has nothing to do with beliefs that are
inferentially justified, but those that are immediately so. As a consequence, then, if impure
foundationalism is true, then at least strong foundationalism is true, since no series of inferences
can have evidence it is not arbitrary unless it has non-inferential support from a subject’s
intuitions or experiential states.

The consequence, then, is that impure infinitism has a dialectical advantage over impure
infinitism, because it can address the challenge of the ‘modus ponens reductio.’ What seems to
follow, also, is that a strong impure foundationalism is necessary for the solution, but there may
be other features of the justification for choosing B2 over B2’ above that relies on some other
epistemic property of the beliefs I have not addressed here—perhaps its coherence with other
beliefs on the matter. However, regardless of this matter, which must be tabled for our purposes
here, it is clear that defensible versions of epistemic infinitism must be articulated as impure
theories of justification.

1 See, for example, Moser (1984, 1989) and Feldman (2003).

2 See, for example, Davidson (1986).

3 For example, Audi (2001).

4 BonJour, even as a coherentist about empirical knowledge, was still a foundationalist about a
priori knowledge (1985).

5 Which is solvable by Susan Haack’s term ‘foundherentism’ (1993).

6 Note that Klein does not eliminate skepticism for the argument. Klein’s take on the matter is
that the old Agrippan-Pyrrhonean skepticism of the five modes is consistent with epistemic
infinitism—namely that there are cases where we do not meet the infinitist requirement. See
Klein’s development of the view in 2003.

7 The argument is more detailed that this, and I will address it more fully below.
8 See Andrew Cling’s criticism of Klein on similar grounds: “Foundationalism implies that some
beliefs can be justified independently of propositional support, not that there are propositions for
which propositional support is in some absolute sense unavailable” (2004, 107).

9 Daniel Howard-Snyder makes a similar distinction between, on the one hand, a basic belief
being justified for S and S believing the premises of a meta-justificatory argument and, on the
other hand, the justification of S’s basic belief being derived from S’s meta-justificatory
argument (2005, 21).

10 In this respect, the classification of Fantl as a pure infinitist may be strained, since he is
willing to countenance other sources of justification. (Thanks to one of my reviewers for this
point.) However, given that Fantl’s requirements are still that infinite chains of inferential
support are the only sources of justification sufficient for complete justification, the view is
exclusive with regards to that element of justification. And again, Fantl’s concession here is
hypothetically dependent on his dialectical opponent allowing basic beliefs to be incompletely
justified. Fantl here does not categorically endorse the commitment.

11 Take, for example, Peirce’s argument that certainty is not something that can be rationally
attained, because reasons are always tentative and open for challenge (Peirce 1931–1935, CP
5.214).

12 Versions of this argument and a simpler form that I call ‘the modus ponens reductio’ and ‘the
simplification reductio’, respectively (2005) have been made by: Deutscher (1973, 6), Pollock
(1974, 28–29), Oakley (1976, 227–228), Cornman (1977, 290), Foley (1978, 313), Post (1980,
32–35; and 1987, 88–91), Moser (1985, 67), and Cling (2004, 110).
DON’T FEAR THE REGRESS: COGNITIVE VALUES AND EPISTEMIC INFINITISM
Scott Aikin
Think Autumn 209

We are rational creatures, in that we are beings on whom demands of rationality are appropriate.
But by our rationality it doesn’t follow that we always live up to those demands. In those cases,
we fail to be rational (and it is appropriate to use the term ‘irrational’), but it is in a way that is
different from how rocks, tadpoles, and gum fail to be rational. For them, we use the term
‘arational.’ They don’t have the demands, but we do. The demands of rationality bear on us
because we have minds that can move us to act, inspire us to create, and bring us to believe in
ways that are responsible and directed. My interests here are the demands rationality places on
our beliefs. Beliefs aim at the truth, and so one of the demands of being a rational creature with
beliefs is that we manage them in a way that is pursuant of the truth. Reasons and reasoning play
the primary role in that management – we ought to believe on the basis of good reasons. That is,
if you believe something, you think that you’re right about the world in some way or another.
You believe because you think that something (call it ‘p’) is true. Now, p’s truth is different from
all ways it could be false, and your being right about p isn’t just some arbitrary commitment, one
that could just as well have been its negation. This non-arbitrary specificity of beliefs is
constituted by the fact that they are held on the basis of reasons. Arguments are our model for
how these reasons go – we offer some premises and show how they support a conclusion. Of
course, arbitrary premises won’t do, so you’ve got to have some reason for holding them as
opposed to some others. Every premise, then, is a conclusion in need of an argument, and for
arguments to be acceptable, we’ve got to do due diligence on the premises. This, however, leads
to a disturbing pattern – for every premise we turn into a conclusion, we’ve got at least one other
premise in need of another argument. Pretty soon, even the simplest arguments are going to get
very, very complicated.

This problem is an old saw in philosophy, and it drives a number of classical works on
knowledge. In contemporary parlance, the challenge posed is termed ‘the regress problem.’
Traditionally, there have been a number of places where the story yielding the problem gets
interrupted. On the one hand, the argumentative model for reasoning can be called into question.
Perhaps argumentation requires more awareness and linguistic ability than what is required in
order to reason (e.g. babies don’t give arguments, but they seem to know things). On the other
hand, there have been special sorts of reasons posed, and the special properties of these reasons
make them so that they don’t have to be conclusions of arguments for them to serve their
purpose – they may be indubitable (you don’t have to argue for propositions nobody doubts),
they may cohere with other truths (sometimes it’s enough for a story to hang together), or the
premises may be yielded by some reliable source (who’s to argue with authority?). The thought
here is that some beliefs may end the regress of reasons by their having some special property
that makes them justified without having any further arguments that they are. The problem with
these solutions is that with all of them, you’ve still got to give arguments that one belief or other
not only has those properties, but also that those properties confer justification. Surely we need
an argument to stop with one sort of belief and not another. And so, we haven’t ended the
regress, have we? Call this the meta-regress problem – any time you propose a regress-ender,
you do so on the basis of an argument, which needs due diligence. And that puts us back on the
road to regress.

The regress problem is a consequence of a tension between our flatfooted intuitions about belief
and knowledge. The first is that ‘knowing’ and ‘rational belief’ are success terms. Knowledge
lays claims on us, and we may meet those demands or not. Second, knowing and rational
believing are reflective successes. We know and believe rationally by thinking hard, being
careful, doing our homework, getting our facts straight. And these are the applications of a
cognizer’s awareness of her responsibilities. Third, that reflective successes can be made explicit
and determinative. You can always show your work and explain why you arrived at one
conclusion instead of another. Let me call the collection of these first three intuitions epistemic
proceduralism – knowing and rational believing are consequences of proper belief-management.
The problem is that this model conflicts with some other intuitions that we have about knowing.
One is that human beings are good at knowing. Even dumb people still know lots of things,
despite the fact that they cannot put an argument together to save their lives. On top of that, even
smart folks, given the regress problem, won’t know much, since no matter how smart you are,
you can’t complete an infinite series of arguments. It looks like the standards are just too high
with prodecuralism to let in a good deal of our knock-about knowledge – what gets us to work on
time, what keeps us from stepping in front of buses or eating glass, what makes it so I can work
my TiVo, and so on. Let us call this perspective epistemic egalitarianism – knowing, though an
achievement, is something that is simple and widespread. The regress problem, then, is a case of
the clash between prodecuralism and egalitarianism. (The clash between the two perspectives is
not just limited to the regress problem. All the same intuitions clash in discussions of skepticism,
contextualism, the analysis of knowledge, religious epistemology, and so on.)
Epistemic infinitism is a thorough-going proceduralist view. The model runs that those who
know are those who have been maximally intellectually responsible. In essence, the thought
behind the view is that if you know, you can answer questions about what you know until there
just aren’t any more questions. But, as it turns out, there are in principle no final questions. So
knowers need to be able to keep coming with the answers.

This is a heavy task. And one reaction is that since knowledge is widespread and infinite reason-
giving is not, there must be something wrong with infinitism (namely, that it is false). This is an
illusion on two fronts. First, the fact that we allow people to say they know in cases were they
didn’t have infinite reasons, or, for that matter, had no reasons at all (maybe they were just
lucky) doesn’t mean that knowledge is so easy. For example, take Jerry, who bets on a 50-1 horse
to win, and the horse wins. Jerry proclaims, ‘There’s something about the name ‘Glue Factory
Bound’. . . I just knew he would win!’ We let Jerry get away with saying this not because it is
true that he knew, but because it doesn’t really matter whether or not he knew – what matters is
whether or not the horse won. But now change the situation. You are about to place a large bet on
a longshot horse based on whether or not you like her name. Do you know that ‘Pretty Pony,’
‘Firefly,’ or ‘Old Brown Shoe’ will win? Even if you made the bet and it paid, you may in a fit
of jubilation exclaim that you just knew it, but would you in cool reflection say you knew it? I
think not.

The point here is that knowledge-attribution is cheap. We allow people who do not know to
claim they do all the time, and we do so because we may waste time correcting them, it may be
rude, or it just doesn’t matter. Knowledge-attribution happens in contexts where there are many
other values on the table in addition to saying truly or not whether someone knows. But all you
have to do to burst the bubble is to ask the question, often in the appropriate tone of voice, ‘Yes,
but do you really know?’

The second illusion is the significance of the fact that infinite reason giving isn’t widespread.
The illusion, of course, is not that some cases of actual non-terminating reason-giving are being
overlooked; instead, the illusion is that when people stop giving reasons, they have satisfied the
demands of knowledge. A regular thought regarding arguments is that they are speech acts
addressed to an audience for the sake of either resolving a disagreement or settling an issue.
Once arguments accomplish these goals, there is no more social use for them – once we are in
agreement, we don’t argue any more. The fact that there are no infinitely long chains of
arguments is a social fact – we have a tendency to agree, and when arguments go on too long, we
give up on arguments and settle matters with our fists. But the epistemic question returns once
the issue is resolved – though we may persuade each other that p, does that mean that we now
know that p? A chasm yawns between the two thoughts, and it seems to demand we acknowledge
that we are often lucky we’ve never met a really smart person skeptical about the things we think
we know – else we suddenly would find that we don’t know.

Knowledge requires that you be able to give reasons you know are good reasons. It seems a
simple truism. Who would say someone knows that p, if asked why he believes it, he shrugged
his shoulders and uttered the an inarticulate ‘hmmm. . . idunno’? So what follows?
First, epistemic modesty. I have many beliefs, and I strive to know. But the task of holding these
beliefs properly and pursuing knowledge requires that I am constantly testing the reasons I have,
and that means I should always be open to the possibility that I am wrong. So I should seek out
the smartest people whom I disagree with and find out what they think, and I should thank
people who refute me. Fallibilism is the philosophical term of art for that collection of
intellectual virtues. The American philosopher Charles S. Peirce was a fallibilist, and he
famously claimed that knowledge is what constitutes the beliefs of inquirers at the end of infinite
inquiry. Fallibilism, for Peirce, is a natural infinitist outlook: since we are not at the end of
infinite inquiry, we don’t know yet if we have knowledge. So we have two duties – be open to
correction, and help move inquiry along. The question now is whether fallibilism is properly
held only on infinitist grounds.

If we were to have regress-ending beliefs, ones that settled the question of whether we know,
then if you thought that you had some of those kinds of reasons, you wouldn’t be open to
challenges from those who reject them. The matter is sealed for you, and those who disagree may
deserve engagement for the sake of correcting them, but they do not have an equal share in the
conversation. From the perspective of those who know, they are merely ignorant, stupid, or
confused. And they must be educated. This, of course, is not to say that one committed to
regress-ending reasons must always fail to charitably respond to those who question them, but
the question is what, exactly, does one say to one’s opponents when the commitments at issue are
those for which one thinks no more reasons have to be given? Non-infinitist (or finitist)
epistemology does not guarantee dogmatism and intellectual intolerance (contrary to what many
anti-foundationalists, for example, have claimed), but given the demands of resolving
disagreements, it is unclear what other options are available for the finitist except for adopting a
temporary infinitism. It is just that the infinitist is an infinitist all the time.

The second consequence of infinitism is that it is the natural intellectual home for the epistemic
proceduralist commitment to evidentialism, the view that one’s beliefs should be supported by
sufficient evidence. Given that the quality of the evidence is something always relevant to
assessing something as evidence (it’s good or bad evidence, strong or weak), we are always
facing a further set of questions when we proffer evidence. Any critical thinking textbook will
offer the same advice – always check your sources, make sure your sample is right, understand
your data, ensure that your experiment doesn’t yield vague or ambiguous results. Having
evidence isn’t enough. You, if you know, must know the quality of that evidence, which requires
that we know a whole lot more things.

A final concern looms: surely there are many things we know without having to give further
reasons: This is my hand. 2 + 2 = 4. If object x has properties P and Q, then x has property P.
Don’t torture innocent people for fun. Pain is bad. I am being appeared to red-bulgy-fruitly. But
with each of these, it seems that if someone weren’t convinced, or curious about how you knew
these things, you could (and should) still offer an argument. I know this is my hand, because I
am being appeared to my-handly and I’m having a kinaesthetic impression of holding my hand
in front of my face. From these, I’ve arrived at the belief that this is my hand. 2 + 2 = 4, because
if you take two of any thing and two of anything else, you’ll have four things. . . just try it! Each
argument here is a function of our concepts. . . but do we have the right concepts? For example,
couldn’t I add two things and two other things and have five things – four objects put together
and the collection of them? (Isn’t a collection a thing – baseball collections, coin collections,
aren’t they things? Why isn’t the collection also counted when we do addition?) There, I think,
are answers to these questions, but you see that it requires that we continue the reason-giving
even on the level of the concepts used. And the same with experiences – having the right
experiences is crucial for the empirical justification for many of our beliefs about the world. But
we, if we take those experiences to give us information about the world, we should be able to
give an argument that they are veridical and how they are relevant to the beliefs they support. If
you know, you should be able to answer questions with reasons instead of shoulder-shrugging or
the back of a hand. Now, we may say those who use those other means to answer questions
‘know,’ but this is out of our desires to be nice to them or save our skins. And those aren’t
reasons to say someone really knows.

Synthese (2010) 174:315–330

DOI 10.1007/s11229-008-9451-2
Elusive epistemological justification
Stephen Hetherington

Abstract What does it take for some epistemological thinking to be epistemically justified?
Indeed, is that outcome even possible? This paper argues that it is not possible: no
epistemological thinking can ever be epistemically justified. A vicious infinite regress of
epistemological reflection is the price that would have to be paid for having some such
justification. Clearly, that price would be too high.

Keywords Meta-epistemology · Scepticism · Justification · David Lewis · Externalism ·

Knowledge · Reflection · Regress

1 A meta-epistemological question

Epistemological practice includes, essentially, people making assessments of knowledge,

justification, or cognate epistemic features. It is routine to hear instances of thinking like this:
‘Imagine these circumstances:….Would the belief that p, formed in such circumstances, be
knowledge? I believe so [or: I do not believe so].’ Confidence is generally high at these
moments. Data thereby arise, fodder for supposedly fecund epistemological theorizing. Theories
are tested, advocated, or discarded. Epistemological life proceeds.
But it rarely includes much reflection upon the epistemic status of its own pronouncements.
What makes an epistemological claim justified or unjustified, knowledge or not knowledge?
When arguing on behalf of some theory of justification or knowledge, must our supporting
claims satisfy that theory—if any at all? Presumably so (we will say), if the theory claims to
describe all justification; otherwise, perhaps not. Yet is that all we can say about this issue?

Far from it. Long ago (1992), I developed some sceptical thoughts about it.1 Later, David Lewis
(1996) did likewise. My main goal was to reflect upon epistemological reflection; Lewis’s was
not. Still, he argued that when one engages in epistemology, one’s mind is opened to sceptical
possibilities; at which point, one loses whatever knowledge one previously possessed.2 On
different grounds, I had also argued for epistemological reflection’s having that sort of effect
upon itself.3 And this paper includes some related reasoning for the following meta-
epistemological conclusion:

No epistemological thoughts are ever epistemically justified. By having an

epistemological thought, one lacks epistemic justification for it. (For example, no one has
justification for an epistemological attribution of justification.)4

Naturally, this is not a thesis I hope is true; I welcome refutation of it. As we may say, adapting
some famous remarks by Hume (1978 [1739–1740], p. 269):

[even if] reason is incapable of dispelling these clouds, nature herself suffices to that
purpose, and cures me of this philosophical melancholy and delirium, either by relaxing
this bent of mind, or by some avocation [such as everyday epistemology], and lively
impression of my senses [and intuitions], which obliterate all these chimeras. I dine, I
play a game of back-gammon, I converse [with other epistemologists], and am merry
with my friends [including my epistemological ones]; and when, after three or four hour’s
amusement, I wou’d return to these speculations, they appear so cold, and strain’d, and
ridiculous, that I cannot find in my heart to enter into them any farther.

Whenever I turn away from this paper’s reasoning, even if to other philosophical pursuits, I find
myself blithely confident that some epistemological positions are justified, while others are not,
that some amount to knowledge, while others do not. Am I correct at those moments? Not if this
paper is correct.

2 An informal preview

Here is a preliminary sense, a brief and less formal one, of the paper’s main argument:

Having justification as an epistemological thinker requires too much epistemological

thought (either impossibly lengthy or impossibly complex epistemological reflection).
This is because no epistemological thinker can have justification, via an apt principle of
justification, without reflecting upon doing so—and because the need for this self-
reflectiveness never ends. (Whenever it seems to do so, this reflects a confusion between
being non-epistemological and being epistemological.)
Why is all of that so? What is epistemological reflection? And in what way, exactly, must it be
self-reflective if it is to be epistemically justified? How does it then become too lengthy or
complex ever to be epistemically justified? This paper will develop detailed answers to these
questions.

3 Being epistemological

What is it to think epistemologically? First, one must be theorizing. Epistemology is officially

‘theory of knowledge’, sometimes in effect ‘theory of justification’. We are not being
epistemological simply in describing Paula as knowing her name. We could be epistemological
in saying that she knows it through testimony and her senses.

I propose that someone is thinking epistemologically if and only if she is thinking about some
epistemic theory or principle—its truth and/or its epistemic status and/or its explanatory role in
accounting for her and/or someone else’s knowledge and/or justification, etc.5 I will say that the
person is being epistemological in relation to, or about, that principle. To be epistemological at a
time is to be thinking epistemologically. To do this is to be thinking about some given epistemic
principle. Hence, at a particular time, one is epistemological only in relation to the epistemic
principle about which one is being epistemological.

That relativisation to a specific principle matters. If a person is not thinking about a particular
epistemic principle, she is not being epistemological in relation to, or about, that principle. At
different times, she may be epistemological about different epistemic principles, by turning her
attention from one to the other.6 So, a person’s being epistemological simpliciter is her being
epistemological relationally—in relation to one or more epistemic principles or theories, about
which she is thinking.

What is an epistemic theory or principle? Here is a representative suggestion. Each such

principle articulates a condition C of an epistemic concept’s applicability. For simplicity, I will
talk only of principles concerning justification, and ones where C is a putatively sufficient
condition. (It could instead be a putatively necessary condition, or even a confessedly weaker
indication, of justification’s presence.) A putatively sufficient epistemic principle may take this
form:

J For any epistemic subject x and time t: If x satisfies condition C at t, then (other things
being equal) x is justified at t in believing that p (for any p whose content is appropriately
connected with the nature of C).7

(For instance, one principle might specify what kinds and amounts of data justify predictions
about the weather. Another could describe what is sufficient for having justified beliefs about
one’s physical surrounds.) If a given principle’s condition C is not putatively sufficient, that
principle says that justification is to only some extent, and in some way, constituted by condition
C.8

4 Two kinds of epistemic subject

What would it be to have justification as someone theorizing about what it is to have
justification? In general, to have justification is to satisfy a true and sufficient principle of it:9

JSat For any epistemic subject x and time t: x has justification at t if and only if at t there
is some true and sufficient epistemic principle P (one instantiating J, the generic form of
epistemic principle mentioned above, in Sect. 3), whose contained antecedent x satisfies,
and hence whose contained consequent x satisfies. (This constitutes her having
justification via P at t for her belief that p.)

JSat is a generic account of what it is to have justification. I will argue, however, that there is a
category of epistemic subjects who fail even JSat’s generic condition. Those epistemic subjects
are all and only epistemological thinkers.10

Let us distinguish, then, between epistemic subjects who are, and those who are not,
epistemological. Recall that to be epistemological at time t is to be epistemological in relation to
some epistemic principle P at t. This requires one to be thinking at t about P. Someone who is
epistemological in that way, in relation to P at t, would be an epistemological epistemic subject
of P at t.11 She is to be distinguished from what we may call a standard epistemic subject of P at
t—who is not thinking at t about P.12

Standard epistemic subjects are the more standard epistemological fare. Epistemologists enjoin
each other not to endorse epistemic principles that accord knowledge or justification only to
epistemic subjects who display marked intellectual sophistication (such as an epistemological
focus). But those same epistemologists tend to assume that the resulting epistemic principles are
true of all epistemic subjects if true of some. However, an epistemic subject is only ever an
epistemic subject of a given epistemic principle at a given time.13 Someone could be a standard
epistemic subject of a particular epistemic principle at one time, yet an epistemological epistemic
subject of that same principle at another time.

We may rewrite JSat with that distinction in mind:

JSat* For any epistemic subject x and time t: x has justification at t if and only if at t there is
some true and sufficient epistemic principle P (instantiating J) such that (1) x is either a standard,
or an epistemological, epistemic subject of P at t, and (2) x satisfies P’s contained antecedent
(hence P’s contained consequent) as the kind of epistemic subject of P at t which, by (1), she is.
That is, (2S) if x is a standard epistemic subject of P at t, she satisfies P’s contained antecedent
(hence P’s contained consequent) without thinking about P. And (2E ) if x is an epistemological
epistemic subject of P at t, she satisfies P’s contained antecedent (hence P’s contained
consequent) while thinking about P.

Now let us abstract from JSat* what will be needed, in Sect. 5, for the paper’s main argument.

To have epistemological justification is to have justification as an epistemological epistemic

subject. By JSat*, this is to satisfy a true and sufficient epistemic principle P’s contained
antecedent while also being epistemological in relation to P. An epistemic subject (call her Sub)
would thereby have that justification as an epistemological epistemic subject of P.
P’s contained antecedent requires Sub to be in some circumstance C; and we may assume that
she is indeed in C. But if Sub is not simultaneously thinking about P, then in relation to P at that
time she is not being epistemological. Nor, therefore, is she being an epistemological epistemic
subject of P. In which case, Sub is not in C as an epistemological epistemic subject of P. Thus,
she is not having justification via P as an epistemological epistemic subject at that time.

So, to have justification as an epistemological epistemic subject is to have it via the epistemic
principle in relation to which one is being that epistemological thinker. Correlatively, to be
epistemological in relation to P (by thinking about it), while having justification via some other
principle P* (about which one is not thinking), is to have justification only as a standard
epistemic subject—a standard epistemic subject in relation to P*. And to have justification as a
standard epistemic subject is not to do so as an epistemological epistemic subject. (I will return
to this distinction in Sect. 6.)

5 A meta-epistemological skepticism

From Sect. 4, a necessary condition of our epistemological epistemic subject Sub’s having
justification via an epistemic principle is that she be reflecting on that principle. But that
requirement entails her having no justification as an epistemological epistemic subject. This
section derives that meta-epistemologically sceptical result.

I begin by refining Sect. 4’s main point.

Suppose that Sub is thinking about epistemic principle P. (She might be reflecting on whether P
is true, or on whether it is justified.) But suppose that Sub has no views at that time on P’s
helping her to have justification. In that event, her relationship to P’s applying to her is not itself
epistemological. Therefore, P is not applying to her qua epistemological epistemic subject.
Accordingly, if Sub is to have epistemological justification via P, she must also be thinking
about the connection between P and her belief’s justificatory status at that time.

So, we have this thesis—an extension of JSat*(2E )—about epistemological justification:

EJ For any putatively sufficient epistemic principle P, time t, and epistemological

epistemic subject x: x has justification for her belief that p, via P at t, only if at t she is
thinking about P—including P’s applying to her at t.14

There are different ways to satisfy EJ’s contained consequent. Sub might think about whether P
is true of her. She could reflect upon how, in her view, P is true of her. But in each case, I will
now argue, she has no epistemological justification via P.

The basic reason is that no one can do enough epistemological thinking to have justification via a
given P. No matter what P’s contained antecedent asks Sub to do, EJ requires that she also be
thinking about her doing it. Again, this could be her considering whether doing what is required
by that contained antecedent suffices for having justification, and whether its doing so applies to
her own case. Or she may be thinking that doing what is required by P’s contained antecedent
does suffice for having justification, and that it is a condition which is satisfied in her own case.
In any event, if Sub is to have justification as an epistemologist via P, this will be because she
has it via a principle like P1:

P1 For any epistemological epistemic subject x and time t: If at tx satisfies P’s contained
antecedent, and at tx is thinking about how and/or whether she satisfies that contained
antecedent, then (other things being equal) she has justification at t for her belief that p.
(She does so via this putatively sufficient principle, P1.)

Why is that so? From where has P1 appeared?

It is needed in our account of Sub’s having justification via P, simply because her relationship to
P’s applying to her is to be epistemological. (1) In order to satisfy P’s contained antecedent as an
epistemological epistemic subject, Sub has to do more than that antecedent specifies. From EJ,
she must reflect, too, upon her satisfying P’s contained antecedent.15 (2)However, by engaging in
such reflection, she is doing what the further epistemic principle P1 asks her to do, in order that
she have justification. P1 specifies what Sub must be doing in addition to whatever P specifies—
so that, in doing whatever P specifies, she is doing it as an epistemological epistemic subject of
P.

Yet even that is not enough. We needed to call on P1 because Sub’s doing only whatever P asks
her to do is not enough to give her justification as an epistemologist. But that same problem
applies to P1 no less than to P: Sub is unable to have justification as an epistemological
epistemic subject by satisfying only P1’s contained antecedent. The reason for this is as follows.

If P1 is needed in order to account for Sub’s having justification via P as an epistemological

epistemic subject, this makes Sub an epistemological epistemic subject of P1 at that time. By EJ,
however, Sub has justification via P1 as an epistemological thinker only if P1 is true of her as
someone thinking about how and/or whether it is true of her. Yet if Sub is thinking about P1 in
that way, she is doing more than P1’s contained antecedent specifies as sufficing for justification.
A further putatively sufficient epistemic principle’s intervention is therefore needed. Sub would
have justification as that epistemological epistemic subject only because this further principle,
P2, is true of her—where P2 stands to Pl as P1 did to P. That is, P2 specifies what more is
needed if Sub is to satisfy P1 as an epistemological epistemic subject:

P2 For any epistemological epistemic subject x and time t: If, at t, x satisfies P1’s
contained antecedent, and at tx is thinking about how and/or whether she satisfies that
contained antecedent, then (other things being equal) she has some correlative
justification at t. (She does so via this putatively sufficient principle, P2.)

Nor does this sequence end with P2. Mutatis mutandis, the same reasoning as introduced P2 into
this tale now leads us to acknowledge a new putatively sufficient principle, P3—standing to P2
as P2 does to P1 and as P1 does to P. And this pattern continues without end, relying next upon
Sub’s thinking about P3’s applying to her, then upon her thinking about P4’s applying to her
(where P4 stands to P3 as P3 did to P2), then upon ... etc. The upshot is as follows.
 (1) Someone has justification via a true and sufficient epistemic principle P as an
epistemological thinker at time t, only if at t she is thinking about the applicability to her
of P1, of P2, of P3, and so on, ad infinitum. This is her thinking about the applicability to
her either of an infinite number of principles individually or of an infinitely large
principle.16

(2) She cannot accomplish this task, be it infinitely extended or be it infinitely complex.17

(3) Hence (by modus tollens, from 1 and 2), no one can have justification via a true and
sufficient epistemic principle P as an epistemological thinker. Epistemic success is
unattainable for anyone as an epistemological thinker.18

This result applies even to epistemic principles purporting to be about epistemological epistemic
subjects. Even for a principle that calls for an epistemological epistemic subject to be thinking
about its own applicability to her, there is a truth about how and/or whether it is applying to her.
By EJ, however, that truth cannot apply to her as an epistemological epistemic subject unless its
doing so is being thought about by her. (The fact that it is itself about epistemological-thinking-
about-itself does not change this point.) So, the epistemological thinking which that truth is about
must be nested within further epistemological thinking (then more, and more again, etc.). Our
sceptical conclusion thus remains intact: there is no epistemological justification.19

6 Disarming a non-reflectivist suggestion

It is a non-trivial inferential leap from A to B:

A There is no true and sufficient epistemic principle, via which some epistemologist has
justification while thinking about doing so;

B There is no true and sufficient epistemic principle, via which some epistemologist has
justification.

But Sect. 5’s regress argument made that leap. Should it have done so?

In endorsing EJ, I was denying what I have previously (Hetherington 1992, Chap. 6) called non-
reflectivism—epistemic externalism, directed at epistemological justification. For I was denying
that someone can (i) be thinking at t about the putative justificatory link between an epistemic
principle P* and herself at that time, (ii) fail to have justification via P* at t, (iii) possess
justification at t via a true and sufficient epistemic principle P about which she is not thinking,
yet (iv) possess the latter justification as the epistemological thinker about P* that she is at t.

Was that denial correct? Or should we reject EJ? Must epistemologists be as endlessly reflective
as I am requiring, if justification is to come their way? Non-reflectivism allows that an
epistemological thinker can have justification via a true and sufficient epistemic principle even
when not reflecting upon whether it is true of her. Non-reflectivism may allow, for example, that
an epistemological thinker’s development of a theory of empirical justification could be justified
a priori.20
That non-reflectivist objection to my argument might also be formulated in terms of epistemic
levels.21 Even for an epistemological epistemic subject, could there be different levels of
epistemic justification? Minimally, to have epistemological justification at a particular epistemic
level is to satisfy a true and sufficient principle P about justification at that level. But the regress
argument requires our epistemological Sub also to reflect on her having justification via P. And
surely (continues the objection) this further reflection is needed only if Sub is to have
justification for the higher-level belief that she has justification via P. This is so, no matter what
level of justification P describes. Hence (concludes the objection), EJ is wrong: an
epistemologist can have justification via a true and sufficient P without thinking about doing so
—even while being epistemological by thinking about a distinct P* instead. She could have
justification via P while lacking what would have been higher-level justification via P+, a
principle about P-justification.

That objection fails. Section 4 noted that someone is a standard—or is an epistemological—

epistemic subject, only relative to a given epistemic principle and a given time. Relative to a
given P and t, a person can be either an epistemological epistemic subject or a standard epistemic
subject—not both—depending on whether she is thinking at t about P’s applicability to her.
Now, the objection from epistemic levels posits someone, Sub, as thinking about P*—while
having justification via P, a true and sufficient principle about which she is not thinking. (For
example, she is pondering the nature of empirical justification for beliefs about the external
world. And these thoughts of hers are said to be justified a priori.) However, the objection does
not describe Sub’s having epistemological justification via P. At most, it describes what would
be epistemological justification via P, if (for a start) Sub was being epistemological in relation to
P. Relative to a true and sufficient principle P of a priori justification, say, Sub is being
epistemically successful—but not also epistemological. In contrast, relative to a true and
sufficient principle P* of empirical justification Sub is being epistemological—but not also
epistemically successful. So, the posited case does not describe someone’s attaining epistemic
success via a true and sufficient epistemic principle insofar as she is being an epistemological
epistemic subject of that principle.

Although it claims to describe her—a single entity—both as being an epistemological epistemic

subject and as having epistemic success, it does not show that a single epistemic subject is
satisfying both of these descriptions. The posited objection wondered whether my argument was
confusing epistemic levels; the objection confuses types of epistemic subject.

Moreover, it is possibly because the objection confuses types of epistemic subject that it presses
its point about epistemic levels.22 For an epistemological epistemic subject of a principle P,
justification does require higher-level justification via P (and, as it transpires, an infinitude either
of other principles or of complexity in the overall justification-licensing principle). It is only for
standard epistemic subjects of P that such a requirement should be rejected. Epistemologists are
right not to require higher-level justification of standard epistemic subjects—but wrong to
exempt themselves from that requirement.23

7 False epistemic principles

I am saying that no epistemological thinker can have justification via a true and sufficient
epistemic principle (and, given JSat*, there is no other way to have it). Yet this tells us that there
are no true and sufficient epistemic principles—unless we cease expecting them to quantify
univocally over all epistemic subjects.

For example, suppose that Sub is in a Chisholmian self-presenting state of seeming to see red24—
a circumstance which, if she were a standard epistemic subject, would suffice for her being
justified in believing that she sees something red. Suppose also that she is reflecting
epistemologically upon her having such justification in that way. Section 5’s regress argument
entails that she thereby fails to have such justification in that way. Her epistemological thinking
‘gets in the way’ of having justification as that epistemological thinker.

Accordingly, no epistemological epistemic subject can have justification by being in anything

like a self-presenting state. One would have to be reflecting at that time on being in that self-
presenting state and on thereby having that justification. But being in a self-presenting state
precludes one’s simultaneously reflecting on being in it. If we take seriously the distinction
between epistemological and standard epistemic subjects, therefore, we should accept that if
there is immediate justification, it is available only to standard epistemic subjects. Any
associated Chisholmian foundationalism is unavailable for epistemological epistemic subjects.

And that example may be generalized. Standardly, epistemologists seek both universality and
univocality when theorizing. Their epistemic principles quantify over all epistemic subjects at
once, offering a univocal account—of what it is for any epistemic subject to have a given kind of
justification. But if this paper is right, they should do that no more.

First, epistemological epistemic subjects and standard epistemic subjects must satisfy
respectively different conditions if they are to have justification. An epistemological epistemic
subject of a putatively sufficient epistemic principle needs to be thinking about its applying to
her (if it is to do this successfully); whereas no standard epistemic subject of that principle need
do so.25

Moreover, any epistemic principle is false insofar as it does include epistemological epistemic
subjects within its univocal scope; for none of these can have justification via it. No putatively
sufficient epistemic principle is a true account of how epistemological epistemic subjects can
have justification.26

Of course, we could decompose any putatively sufficient epistemic principle P into two sub-
principles, PS and PE. PS applies to standard epistemic subjects. PE is restricted to
epistemological epistemic subjects. I have argued neither for nor against any standard kinds of
scepticism—each being about standard epistemic subjects having some form of justification.27
But each PE has been found to be wholly false—false for all of its (epistemological) epistemic
subjects.

8 Persons
Our sceptical result obtains even for principles that seem as if they ought to be about all
epistemic subjects. Consider again a principle purporting to describe a simple sort of perceptual
justification. Seemingly, such a principle can be about everyone, because all of us employ our
senses in forming or supporting many beliefs. Even epistemological thinkers do so.

Well, that is not quite so. A principle about a simple kind of perceptual justification might indeed
be true of all persons—yet probably only because each person is at least sometimes a standard
epistemic subject. That epistemic principle is not thereby true of all epistemic subjects. Not every
epistemic subject has perceptual justification, since no epistemological epistemic subject does.
Although an epistemological epistemic subject may be using her senses, our regress argument
shows that this never suffices for her having justification as an epistemological epistemic subject.
Too much extra reflection is required of her, in addition to her sensing, if she is to have
perceptual justification as an epistemological epistemic subject.

Resistance to that conclusion might nonetheless linger:

Suppose that a person, Percy, is sensing in away described by a true and sufficient P’s
contained antecedent as sufficing for perceptual justification. Suppose also that Percy is
doing more besides, by thinking about his sensing. But the ‘more besides’, by coexisting
with the sensing, does not undo the sensing’s justificatory work. Percy is sensing in a
way that, were he not thinking epistemologically, would (according to P) give him
justification as a standard epistemic subject. Hence, he retains that justification even
when thinking epistemologically. He can have the perceptual justification when being
epistemological, even if not by being epistemological.

The flaw in this reasoning is that (as Sect. 5 explained) the ‘more besides’ does deprive an
epistemological thinker of justification. The ‘more besides’ coexists with the sensing within the
person, but not within any epistemic subject. Percy is using his senses epistemologically, as it
were. He therefore fails to use them in a way that gives him justification (an outcome
categorially available only to an epistemic subject). The regress to which the ‘more besides’
commits Percy leaves him too mired in epistemological thinking (about his sensing) for his
sensing to give him justification. He cannot have any standardly perceptual justification as an
epistemological thinker.

Standard forms of scepticism aside, maybe Percy has other kinds of justification at that time. Yet
he would do so only as a standard epistemic subject, by satisfying true and sufficient epistemic
principles about which he is not thinking. He might, for instance, be using his senses reliably and
non-epistemologically (that is, standardly), relative to another true and sufficient principle about
perceptual justification. But he would not thereby acquire perceptual justification in his capacity
as the epistemological thinker he is at that time, about a particular principle of perceptual
justification.

The temptation to think otherwise confuses significantly different epistemic modes of the one
person. The concepts of a person and an epistemic subject are not identical.28 Even if the one
person is perceiving reliably and thinking epistemologically about doing so, this does not entail
that the one epistemic subject is doing so. And epistemic principles quantify over epistemic
subjects, not (strictly speaking) over persons. Even when a principle claims to be talking about
people, it cannot quite succeed: a person qua object-of-an-epistemic-principle’s-focus is ipso
facto an epistemic subject. And a person qua epistemic subject need not be an epistemic subject
qua person. Epistemology is, first and foremost, about epistemic subjects. It talks about persons
only qua epistemic subjects.

9 Being impurely epistemological

Seemingly, this paper confronts us with a stark choice:

Either lack all epistemic justification as an epistemological thinker; or never be an

epistemological thinker.

But we may wonder whether that choice is needlessly stark.

In particular, is it possible to think epistemologically, and to gain epistemic justification in doing

so, by being less purely epistemological? This paper’s picture has been one of epistemological
inquiry (if it is to produce justified conclusions) as needing to possess, per impossibile, an
unending relentlessness, a crystalline purity—unceasingly pondering pertinent epistemic
principles. Maybe, however, epistemological thinking is not like that. Perhaps it is never so
purely or extremely epistemological. Might I have been analyzing the epistemic power only of
an unrealistically exaggerated conception of what it is to be epistemological?

Certainly in practice, that could seem so. It is undeniable that people, when thinking
epistemologically, will often not satisfy only my conception of epistemological inquiry. For
instance, even while reflecting upon an epistemic principle, you may also be evaluating related
observations or scientific data. You would be combining epistemological thinking with non-
epistemological thinking. Maybe epistemological thinking is always accompanied by some non-
epistemological thinking.

Yet even this would not allow it to evade our meta-epistemological scepticism. The strictly
epistemological part of your thinking would still be failing to provide you with epistemic
justification—irrespective of whether your non-epistemological thinking, when accompanying
that epistemological thinking, could be at all justified. We may put this point more strongly:
Within any quest for epistemically justified views, any strictly epistemological thinking is a
liability. No purely epistemological presence contributes justified views itself. So, if
epistemological thinking as we practise it is epistemological by having a strictly epistemological
core, then it is unjustified insofar as it is epistemological, strictly speaking.29

Still, might our sceptical regress argument show that there never is a purely epistemological core
within any attempts at thinking epistemologically? We may formulate that idea like this:
Epistemological thinking is only ever messily epistemological, never purely so. At any rate, it is
like that if it is ever to provide justification. In other words, we can regard epistemological
thinking non-sceptically only if we characterize it as being partly non-epistemological.
But care is needed here. The previous-but-one paragraph explained why we could not evade the
sceptical result by conceiving of our actual episodes of epistemological thinking as mere sums of
(i) epistemological thinking, independently characterisable as epistemological, plus (ii) non-
epistemological thinking, independently characterisable as non-epistemological. Instances of (i)-
plus-(ii) would not portray epistemological thinking as being present only messily (in the
intended sense of ‘messily’). That is because in principle the epistemological part of an instance
of (i)-plus-(ii) is separable from the non-epistemological part. The sum, (i)-plus-(ii), is therefore
a combination of two conceptually independent elements—the epistemological and the non-
epistemological. And, we have seen, the epistemological element provides only unjustified
views.

What the present suggestion needs, then, is a conception of our episodes of epistemological
thinking as including both epistemological and non-epistemological elements—but with these
not being characterisable, as respectively epistemological and non-epistemological,
independently of each other. In every relevant way, they would be ineliminably intertwined. No
epistemological thinking would be only epistemological. And this would be so, no matter how
finely we ‘slice’ the epistemological thinking: it would be partly non-epistemological, ‘all the
way down’. Not only ‘everyday’ epistemological thinking would be partly non-epistemological.
Even the most ‘conceptually pure’ epistemological thinking would remain partly non-
epistemological.30

We must wonder, however, whether that is really possible. Insofar as there would still be some
difference between epistemological thinking and non-epistemological thinking, something
constitutes that difference. What would do so? The criterion I have used in this paper for
characterizing epistemological thought—namely, reflection upon an epistemic principle—
remains the most apt. Maybe each ‘unit’, no matter how small, of epistemological thought also
includes some non-epistemological thought. Even so, some aspect of that unit includes reflection
upon an epistemic principle; else, it is not even a minimally epistemological aspect of some
thinking. We have seen, however, why that aspect of any such minimal unit of only-mixedly-
epistemological thought would contribute no justification. The point remains, therefore, that if
some justification is available, this is only because of the non-epistemological aspect of whatever
minimal mix is being hypothesized to be epistemological-(although only-mixedly-so). Even for
that minimal unit, then, unless we relinquish the category of the epistemological altogether we
continue to be susceptible to this paper’s meta-epistemological scepticism.31

An earlier version of this paper was presented at UNSW in 1992 (with the title, “Lacking
Knowledge and Justification By Theorizing About Them”). That earlier version was cited by
Lewis (1996, 550 n.). Since then, the paper has been presented at the 2008 Australasian
Association of Philosophy Annual Conference. I am grateful for the excellent comments made
on those occasions. Helpful suggestions by two anonymous referees were also very welcome.

1 For some recent thoughts, see Hetherington (2006a).

2 Lewis’s news was not wholly gloomy. He allowed that one might regain one’s former
knowledge, once the epistemological moment is past.
3 My argument was more general, in part by not relying upon the highly contestable assumption
—which Lewis clearly employed—that epistemological reflection includes attention to sceptical
possibilities.

4 My argument will also apply, mutatis mutandis, to epistemological theorizing about

knowledge. But this application will not be made explicit throughout the paper.

5 More than this can be involved in particular cases: the person could be reflecting upon more
than one epistemic theory, for example. But our generic description captures a core conception
of what it is to be thinking epistemologically. We may understand it as characterizing a ‘pure’
form of epistemological thinking—an idea which Sect. 9 will discuss more fully.

6 ‘What if she is thinking about two such principles over a sustained period, back and forth,
comparing them?’ Then she is epistemological during that extended period about both principles,
even if there are times within it when she is epistemological only about one of the two. For
simplicity (and because I believe that my argument is generalisable to these more complex
cases), I will continue talking just about the simpler case, of reflection upon a single principle.

7 For simplicity, I talk here of believing. But my discussion also applies, mutatis mutandis, to
acceptance and awareness, say.

8 Note 3 adverted to Lewis’s ‘highly contestable assumption … that epistemological reflection

includes attention to sceptical possibilities’. It may be objected that my characterization of
epistemological reflection—as involving attention to an epistemic principle—is no less
contestable. Notably, particularism might be cited—as an epistemological practice of assessing
the presence or absence of knowledge or justification without doing so by applying a principle or
theory of knowledge of justification. But particularism does this so as to test or support a
principle or theory of knowledge or justification. Consequently, even it accords with my criterion
of epistemological reflection upon some such principle or theory.

9 (i) I will use the phrase ‘true and sufficient’ when talking of a principle whose condition is both
putatively sufficient in form and actually sufficient in effect. Whenever I am not assuming a
principle’s truth, therefore, I speak only of its being ‘putatively sufficient’. (ii) As Sect. 3
allowed, some epistemic conditions are not even putatively sufficient. (They are putatively
necessary, or confessedly weaker still.) Nevertheless, one actually has justification by satisfying
a particular principle only if it is both putatively and actually sufficient in its description of that
kind of justification.

10 The terms ‘epistemologist’ and ‘epistemological thinker’ are being used to designate someone
only when she is being epistemological. In this sense, a professional epistemologist is not always
an epistemologist.

11 This is so, provided that she is an epistemic subject at all of P at t. What is required for this
prior condition to be true of her? We may treat ‘epistemic subject of P’ as short for ‘potential
subject of P’, meaning only to indicate a categorial aptness for being described more specifically
as satisfying P. (See also note 28 below.)
12 Here, too, see the previous note.

13 I have urged elsewhere (Hetherington 2006b, sec. 10) that we should focus epistemological
analysis upon epistemic agents, not epistemic subjects. This is especially apt when discussing
epistemological epistemic subjects/agents, because to be epistemological is to be thinking in
what is presumably an agentive way. I will continue writing in this paper of epistemic subjects,
including epistemological ones. But we may take it that they are agentive subjects. This point
will matter in note 20.

14 Would a similar constraint apply to epistemological beliefs as such—not only to justified

epistemological beliefs? It would, if a genuinely epistemological belief must either be (i)
accompanied by reflection upon some principle P that specifies a sufficient condition of the
belief’s being epistemological, or (ii) justified simply in order to be epistemological. The former
need would obtain if epistemological belief—justified or not—is always suitably self-reflective.
The latter need would obtain if epistemological belief is nothing—literally nonexistent—if not
justified. Here, I take no stand on either of these independently interesting ideas.

15 In arguing against what he calls metajustificatory foundationalism, Jeremy Fantl (2003) calls
upon similar reasoning to mine. In his terms: For any feature F that makes a foundational belief
justified, the epistemic agent’s having reason to believe that a belief has F makes the belief more
justified than the fact of its having F would make it.

16 ‘Surely no such regress could really develop. After…five? six? how many?…iterations, no
real sense could be made of them. (Would we lack the words even to formulate these iterations
after that while?) So, externalism is true of epistemological justification—if there is to be any.
And why, therefore, should there not be any—obtaining in some externalist way, bypassing this
section’s internalist strictures?’ But if internalism is ever to be aptly applied, epistemological
epistemic subjects in particular are those to whom it would be aptly applied. It is they (if anyone)
who are aware of, and who muse upon, their epistemic practice and epistemic principles. Notice,
also, how many recurrent aspects of epistemological focus reflect internalist ideas—for example,
that of ‘stepping back’ from an inquiry so as to think about it, even contemplating all of an
inquiry’s presumptions.

17 Fantl (2003) defends a moderate version of Peter Klein’s infinitism. (See Klein (2005) for a
more recent statement of his views.) Fantl’s aim (p. 560) is to avoid the implication ‘that humans
must have the ability to actually entertain an infinite series of reasons’. He does this by accepting
that propositions can be justified well enough for knowledge, even without being completely
justified (ibid.). But although I wish Fantl’s approach well, it does not save epistemological
thinkers from the present sceptical reasoning. For this reasoning commits such thinkers, if they
are to have (epistemological) justification at all, to perpetually continued reflection upon the
associated epistemic principle(s).

18 ‘Could there be propositional epistemological justification even if—courtesy of this section’s

regress argument—there is no doxastic epistemological justification?’ Propositional
epistemological justification would include propositions ‘reflecting’ upon other propositions in
epistemological ways. Yet then (by this section’s form of reasoning) there would be an infinity
of further propositions ‘reflecting’ in such theoretical ways. Indeed, it is even clearer for
propositional than for doxastic justification that an infinity of theoretical propositions would be
implicated in some such vicious regress.

19 A clarification: This sceptical conclusion is not saying that, even at best, epistemological
thinking can be only somewhat justified, or justified only fallibly. Our regress argument applies
to any amount, and any standard, of epistemological justification. Our sceptical worry is
structural. It pertains to the relationship of applicability between (i) any putatively true and
sufficient epistemic principle of any form and strength of justification, and (ii) any would-be
epistemological epistemic subject of that principle. Even to claim that some merely moderate
amount of merely fallibilist justification is present for an epistemological view is still to imply
that the support for the view in question satisfies a true and sufficient principle, one about that
lesser amount, and that fallibilist form, of justification. In other words, even when the amount
and form of justification are less impressive, some true and sufficient principle describes what is
required for possessing precisely that amount and form of justification.

20 Might non-reflectivism even allow that the epistemological thinker’s theory of empirical
justification is justified contextually? By this, I mean that it is justified through resting upon a
basis that is not itself justified—because it is taken for granted, simply not questioned—within
this context of inquiry. For such a view of justification, see Annis (1978) andWilliams (2001,
Chap. 14).). But that contextualist proposal is itself a principle of how justification is to be
present. Even a contextualist principle saying that justification can be contextual (by resting upon
some contextually unjustified and unquestioned beliefs) is itself an epistemic principle P telling
us what suffices for the presence of some justification. Thus, it does not escape the scope of this
paper’s argument: to end an epistemological inquiry (qua search for epistemological
justification) by pointing to some contextually unjustified beliefs is to cease being
epistemological at that time. It is to seek to ground what is epistemologically justified on what is
not only not justified (and not questioned) in that setting, but what is non-epistemological. For
once inquiry ends by taking some thesis for granted, epistemology is no longer being pursued, at
least in relation to that thesis. An epistemological moment includes inquiry, reflection upon
epistemic ideas—not a cessation of inquiry, a taking-for-granted. More precisely: although one
will take much for granted when being epistemological, one is not thereby being epistemological
about, or in relation to, that which is being taken for granted. Regardless of whether or not
justification can be grounded in non-justification (as contextualism claims), no epistemological
justification is grounded in non-epistemological justification. I return to this issue in Sect. 9.

21 On this notion, see Alston (1989, Chap. 6).

22 Or perhaps instead the confusion arises because the concept of a person is being substituted
for that of an epistemic subject. This could occur because seemingly a single person at the one
time is both (i) the epistemological epistemic subject (not gaining justification in that capacity) in
relation to P*, and (ii) the standard epistemic subject in relation to P (being claimed to have
justification by satisfying P). Section 8 will discuss this issue.
23 The inclination to accord an epistemological epistemic subject justification via a P about
which she is not thinking should also be allayed if we adopt note 13’s recommendation. We
would think of an epistemological x as first and foremost an epistemic agent. We would realize
that at t she is such an agent only in relation to some P. Then if we accept that she would have
justification only as an epistemic agent, we will infer that at t she has justification qua
epistemological epistemic agent only via P—the lone principle in relation to which, at t, she is
actively being epistemological.

24 See Chisholm (1989, pp. 18–20).

25 This complexity in how we must conceive of epistemic subjects is the reason why JSat
needed to be replaced by JSat*.

26 We might wonder whether this is so when the epistemic principle in question is about a form
of justification which (unlike Chisholmian self-presentation) has the potential to be
comparatively reflective and self-aware, such as coherence. Suppose that someone is reflecting
upon a coherentist principle CJ which her reflections at that time also seem to satisfy: ‘My
musings upon CJ and coherentist justification are themselves coherent. How apt.’ Yet even these
reflections are not thereby justified epistemologically, unless she is reflecting upon how her
reflections satisfy CJ: ‘These epistemological claims of mine about CJ and coherentist
justification are justified by satisfying CJ itself.’ In making this assessment, however, she would
be mistaken. She could have this epistemological justification only by satisfying, not CJ, but
instead a further principle. For simplicity and charity, let us assume that this new principle (call
it CJ+) would also be coherentist. It will describe meta-coherence—the coherence of (i) a belief
about one’s having coherentist justification for beliefs b1, …, bn via CJ, with (ii) those mutually
coherent-via-CJ beliefs b1, …, bn. But then this sort of need will continue ad infinitum. Now it
applies to CJ+—thereby generating a need for a new principle CJ++ to enter the story—as
previously it did to CJ. On and on, this pattern proceeds. Hence, even here, this thinker will not
be justified epistemologically. A vicious infinite regress sees to that. Although she may believe
that her beliefs b1, … bn at that moment are justified in a coherentist way, this meta-belief of
hers is not itself both justified and epistemological. Nor, therefore, are her beliefs b1, … bn—no
matter their content.

27 Equally, when arguing in Sect. 6 against non-reflectivism (externalism about epistemological

epistemic subjects), I was arguing neither for nor against standard epistemic externalism (about
standard epistemic subjects). For the standard distinction between epistemic externalism and
epistemic internalism, see Hetherington (1996, Chaps. 14, 15).

28 Although the full relationship between those two concepts is not discussed in this paper, here
is a brief suggestion. The concept of an epistemic subject is an epistemologist’s way of thinking
of something as being the subject of an epistemic principle. The ‘something’ might, or might not,
always be a person; I take no stand on whether animals are ever epistemic subjects. In any event,
the thing’s being an epistemic subject would be its having an epistemic property or mode of
being. To be an epistemological epistemic subject is to have a different property from whatever
makes one a standard epistemic subject. If this paper’s argument is sound, a belief’s instantiating
the property of being justified depends, in part, on which of these other properties the belief or
believer has at that time.

29 Moreover, it can distract you from pursuing alternative thinking which would provide some
justification (albeit standard—not-strictly-epistemological—justification).

30 Williamson (2008) argues that much philosophical thinking does not draw upon any special
faculty, such as of intuition. Instead, it applies a comparatively ordinary capacity to gain
knowledge of counterfactual truths (and thereby modal truths, in particular). Epistemological
thinking is clearly intended by Williamson to be subsumed under this characterization.

31 But see Hetherington (2008) for some further thoughts—non-sceptical ones—on the nature of
philosophical knowledge. A significant reconceptualisation of such knowledge might be needed
here.

The Epistemic Justification Puzzle

Christos Kyriacou
The University of Edinburgh

… I follow the mainstream view and consider the third option of Agrippa’s trilemma, infinitism,
to be entirely implausible and therefore ignore it. Infinitism is the view that the justification of a
belief requires the pursuit of justificatory reasons ad infinitum. One important reason counting
against infinitism is that it is cognitively too demanding for cognitively finite beings like us. As
Wittgenstein (1953:136) has said: ‘Justification comes by experience to an end. If it did not it
would not be justification’. Another reason is that it seems to be self-defeating, as it claims that
belief in infinitism is justified but this very belief does not pursue justificatory reasons to infinity.
But for a heroic defence of infinitism against all odds see P.Klein (2008).

… Coherentist positions can be broadly distinguished into linear and holistic. The linear
coherentist claims that there is a linear chain of individual beliefs that justify each one of our
justified beliefs. But linear coherentism has largely been out of favour and philosophers
sympathetic to coherentism don’t usually attempt to spell out an account of coherentism along
these tracks.

It has been rather unpopular because it seems to lead to an uncomfortable dilemma: either to an
unpalatable infinite regress of individual beliefs or to a narrow and, therefore, profoundly vicious
circularity. This is the case because we either have to push back the chain of individual justifying
beliefs ad infinitum or we have to break this infinite regress with a narrow justifying loop. That
is, we have to justify one of the supporting individual beliefs with one of the beliefs that were
initially in need for justification and, thus, come full circle.

But as with infinitism, an infinite regress of justifications is psychologically impossible for

cognitively finite beings like us and, therefore, unpalatable and a narrow justifying loop seems
entirely question-begging.

Truth is a One-Player Game: A Defense of Monaletheism and Classical Logic

Benjamin Burgis
University of Miami

… I would argue that the “what could be more basic or certain than logical truths?” question
illegitimately smuggles extremely dubious foundationalist epistemic assumptions into the
philosophy of logic.17 Why, after all, should we assume that our beliefs must be justified by
reference to more fundamental beliefs? From Descartes onwards, this epistemic program has
faced what so far look like insuperable difficulties. The narrower your class of foundational,
allegedly self-evident beliefs, the fewer interesting beliefs of the kind that we all ordinarily take
to be cases of knowledge can be derived from them. (Of course, you can always bite the bullets,
stamping your foot in each case and insisting that the various prima facie cases of knowledge
that turn out not to be knowledge on your theory really aren’t, because they aren’t derivable from
the class of foundational beliefs, but the more of that you do, the less convincing your account of
knowledge is going to be to anyone who doesn’t already agree with you.) Conversely, the
broader that class of allegedly foundational beliefs, the less plausible it becomes that they really
are all basic and self-evident.18

17 It could be objected that the claim that all beliefs must be justified on the basis of more
fundamental beliefs isn’t unique to foundationalism, but rather represents the overlapping
consensus of foundationalism and infinitism, but the objection I’m considering really does seem
to assume foundationalism rather than infinitism. After all, the idea in standard dismissals of the
debate about basic logical principles is not that we’re not justified in believing logical truths
because they can’t be justified by anything more basic, but rather that things don’t get more basic
than logical truth, and that as such, trying to have an argument about whether we should accept
them or not is a misguided enterprise. In other words, logical truths sit at the justificatory
foundation at the bottom of the stairs.

18 Some foundationalist views shy away from giving basic beliefs any label quite as emphatic as
“self-evident,” but it’s still the case that they take the basic beliefs to be (a) justified, but (b) not
on the basis of the requirements imposed on all other beliefs. Thus, I think, even for weaker
versions of foundationalism, we can say that the basic beliefs have a “special epistemic status
that makes it clearly right to believe them,” and the same criticism goes through—the broader the
class of beliefs for which one makes this claim, the less plausible the claim becomes.

INFINITE REGRESS ARGUMENTS

Raymond D. Bradley

Infinite regress arguments are used by philosophers as methods of refutation. A hypothesis is

defective if it generates an infinite series when either such a series does not exist or its supposed
existence would not serve the explanatory purpose for which it was postulated.

English philosopher John Locke criticized the view that an act is free only if we freely choose
whether to perform that action. Since freely choosing is itself an act, he argued, this commonly
held theory is absurd. For, that act of choice must then, according to the theory, be preceded by
still another act of choice that is free. It generates an infinite series of acts of free choice, one act
of choosing determining the acts of another, and so on ad infinitum. Since such a series does not
in fact occur, he argues, this account of what it is to act freely must be wrong. Infinite regress
arguments are frequently used to demonstrate that an explanatory hypothesis cannot in principle
explain what it is supposed to explain. It would be absurd to suppose that the only way we can
justify any of our beliefs is by appealing to other beliefs. Such an explanation is fatally flawed
since it generates either an endless chain of beliefs or a circular chain. And since we certainly
don't entertain an infinite number of beliefs, the chain must circle back on itself. But a circular
explanation doesn't explain anything, Hence an alternative account of belief-justification is
needed. The Foundationalist answer is to say that the foundation of some beliefs is experience
itself, not just beliefs about experience.
Some infinite regresses are linear, not circular. Suppose someone thinks the explanation as to
why anything at all exists is that God made things exist. Then the God-hypothesis, when thus
invoked, simply adds to the burden of explanation. On pain of infinite regress, we cannot explain
why anything at all exists by invoking the existence of still another entity (God). For then that
entity's existence calls for explanation.

Merely generating an infinite series is not in itself objectionable. The claim "Every natural
number has a successor" entails an infinite series of natural numbers. Likewise, the claim "For
every event there is a temporally precedent event that is its cause" entails both an infinite series
of events and an infinite series of moments of time at which they occur. Yet the existence of an
infinite series of natural numbers is mandated by logic and mathematics. And the concept of a
beginningless series of events and temporal moments is not self-contradictory. There is no
warrant in logic therefore for Aquinas's claim that these infinite series "cannot" go on forever.
This is a misuse of an infinite regress argument.

Throughout the history of philosophy, infinite regress arguments have sometimes been used to
demonstrate that ultimately some features of the universe cannot, on pain of infinite regress, be
explained at all. The brute fact that some things exist, is just one example.

Infinitism, Completability, and

Computability: Reply to Peijnenburg
Jeremy Gwiazda
The City University of New York Graduate Center
Mind, Vol. 119

In ‘Infinitism Regained’, Jeanne Peijnenburg argues for a version of infinitism wherein ‘beliefs
may be justified by an infinite chain of reasons that can be actually completed’. I argue that
Peijnenburg has not successfully argued for this claim, but rather has shown that certain infinite
series can be computed.

In ‘Infinitism Regained’, Jeanne Peijnenburg argues for a strong form of infinitism wherein
‘beliefs may be justified by an infinite chain of reasons that can be actually completed’
(Peijnenburg 2007, p. 598). Arguing against a more standard infinitist position Peijnenburg
writes:

Fumerton does not say why he believes that finite minds cannot complete an infinitely
long chain of reasoning. Presumably he thinks that such a task would be infinitely
complicated, or would take an infinite time to finish. (Peijnenburg 2007, p. 598)
Indeed, it seems that completing an infinite chain of reasoning would lie beyond the powers of a
finite mind and would take infinite time. One then expects Peijnenburg to argue that a finite
mind could complete this infinite task. Instead Peijnenburg demonstrates that P(E0) can be
computed in certain cases, where E0 is a proposition that is made probable by E1, which is in
turn made probable by E2, and so on indefinitely. After performing an example of such a
calculation Peijnenburg writes, ‘The justification [of E0] is, although infinite, perfectly
computable and completable’ (Peijnenburg 2007, p. 601). But Peijnenburg has merely shown
that P(E0) is computable; Peijnenburg has in no way shown that it is possible to complete such
an infinite chain of justification. The ‘and completable’ at the end of the above quote does not
follow from the argument. There is a wide gulf between demonstrating computability (in this
case showing that certain infinite series converge) and demonstrating completability (showing
that ‘beliefs may be justified by an infinite chain of reasons that can be actually completed’).
Peijnenburg claims to have done both, but has only done the former.

What’s So Bad About Infinite Regress?

Tony Roy (3/10/2010 !draft!)

Presumably I came from my parents, who came from theirs, etc. Many philosophers have
held that series of this sort, where one thing depends on the next, etc. cannot continue to
infinity. Or, at least, they have held that vicious infinite regress is impossible. But it’s
hardly clear what makes a regress vicious or benign, and so what justifies such
evaluations. For certain ancients, it may have seemed sufficient justification, that the very
idea of infinity is absurd. But in this day and age, with infinities routinely manipulated in
mathematics, it is more difficult to reject infinity as such. In this paper, I assume the
possibility of infinite series, and advance an account of their division into regresses,
vicious and benign. On this account, a vicious infinite regress involves straightforward
contradiction. Thus a theory which leads to vicious infinite regress is reduced to absurdity
in the usual way.

Regress arguments have been part of philosophy since the time of Plato, and Plato’s third man
represents a great moment in the history of philosophy (Parmenides 132a-b; much discussion
surrounds Vlastos 1954). As portrayed in the dialogue, Socrates offers a theory that explains, in
part, how distinct things can share a property or feature F and, against this theory, Parmenides
raises his regress objection. It is not my aim to engage in questions of exegesis. Perhaps, though,
Plato suggests a view on which (i) if some things are F, they are F only by participation in a
form distinct from them, and (ii) if some things are F by participation in a form, the form is itself
an F thing. On this view, there are forms, and things have their features by resembling them.
Now suppose there are some F things. By (i), the F things are F by participation in a form
distinct from them. By (ii), this form is an F thing. So the original things, and the form, are all F
things. So by (i), they are all F by participation in a form distinct from them. By (ii), the new
form is an F thing. So the original things, and the two forms, are all F things. Etc. It follows that
there are infinitely many F forms. In Parmenides’s hands, the argument takes the form of a
reductio: Plato’s theory has the consequence that there are infinitely many F forms; this
consequence is bad; so Plato’s theory is to be rejected.

But this reasoning raises the question, what’s so bad about infinite regress? why reject Plato’s
theory on the basis of this consequence? Perhaps certain ancients found the very idea of
infinitely many F things absurd. Supposing that there are some F things, (i) and (ii) imply the
denial of an additional premise, (iii) there are at most finitely many F forms. This may have been
trouble enough for Plato. However, it is not so easy for us to reject infinity as such. Thus, e.g.,
contemporary Cantorian mathematics accommodates infinite quantities as a matter of course.1

Even so, many have thought that there remains something wrong about (i) and (ii). The
consequence is not merely that there are infinitely many F things, but that there is a vicious
infinite regress. Infinite series are divided into those that are regresses and those (if any) that are
not, and regresses are divided into those that are benign and those that are not. A regress that is
not benign is vicious, and a vicious infinite regress is to be rejected. Like a donkey chasing a
carrot suspended before its nose, in a vicious infinite regress, every step toward a goal somehow
leaves the goal removed by another step.

But it is hardly clear what to make of this “eternal seeking,” and even the division of infinite
series into regresses, vicious and benign, is not well-understood. Given a regress, evaluations are
typically treated as obvious or on-the-surface. Thus, e.g., Russell claims a regress I describe on
p. 5 below (the fundamental relation regress), is “plainly vicious” (1911-12, 9). But other authors
hedge their bets. Armstrong says versions of the regress are “either vicious or at least viciously
uneconomical” (1989, 108), and J. Peterson that “while it may not be vicious, this regress... is
unbelievable” (1991, 154). Insofar as it has application against a range of competing theories,
Lewis treats the regress as a theoretical weakness, to be accommodated by one theory as much as
by another (1983, 353-4). This seems odd if, as Russell thinks, the objection is fatal — for there
is not much point contesting the viability of theories known to be dead. But these judgments are
all made apart from an explicit account of what viciousness amounts to. As Alex Oliver suggests,
“much of the trouble hinges on unclarities about the role of infinite regresses in metaphysics,
when they are vicious and when virtuous” (1996, 32). The problem is not merely when regresses
are vicious, but also what viciousness amounts to. We need to know what makes a regress
vicious, and why a vicious regress is to be rejected.

In this paper, I assume the possibility of infinite series as such, and advance an account of ones
that are regresses, both vicious and benign. On this account, a vicious infinite regress involves
straightforward contradiction. Thus a theory which leads to vicious infinite regress is reduced to
absurdity in the usual way. In the first section, I sketch some familiar regress arguments, along
with some initial attempts to understand them. In section two, I develop the account of infinite
regress arguments. The third section applies this account back to simple versions of the regress
arguments from section one. It is not my aim to dispute traditional evaluations of regress
arguments. By and large, I think such evaluations are correct. Rather, my aim is to expose the
way the regresses work. Given progress in other parts of logic, it is remarkable that regress
arguments remain so poorly understood.

I. Arguments and Evaluations

I begin this section by sketching some familiar regress arguments, and then turn to proposals for
their evaluation. I do not pretend to decide larger issues about truth, human origins, God, and
properties, or to give complete discussions of the proposals for evaluation. The larger questions
depend on much more than the evaluation of these particular arguments; and I develop only the
main outlines of proposals for evaluation. My aim is rather to put some issues on the table, to
indicate something of their breadth and significance, and so to set up the discussion that follows.

Let us begin with a simple truth series. Say a declarative sentence in corner quotes names the
proposition expressed by the sentence, and suppose we accept all expressions of the form,

(Tp) {p} is true iff p

where the same declarative sentence is substituted for both instances of ‘p’. Snow is white; so by
an instance of (Tp), {snow is white} is true; so by another instance of (Tp), {{snow is white} is
true} is true; etc. This may seem to be a paradigmatic example of a non-vicious series (see, e.g.
Armstrong 1989, 54, cf. Carruthers 1982, 19). However, the situation is not always so simple.

Consider a hereditary series. Suppose no person is human unless the biological offspring of
humans — that a person’s humanity depends on the humanity of his or her parents. Then if I am
human, I have human parents; if they are human, they have human parents; etc. One might
object, from a creationist perspective, that some human might have no parents at all, and thus no
human parents. Or one might object, from an evolutionary perspective, that a sorites problem
results from vagueness in the predicate ‘human’: perhaps a parent is the same kind as its child
only up to some “tolerance” — so that members of different kinds might emerge gradually
within a series. However, none of this matters. The question is rather, whether it is possible that
the humanity of each child depend on the humanity of its parent. So I simply assume the
condition. Romane Clark contends that this regress is vicious: evidence from biology to the side,
a person’s humanity cannot depend on the humanity of her ancestors this way (1988, 377).
Aquinas denies that this regress, or one very much like it, is vicious (Summa Theologiae
I.46.2ad7; cf. Sanford 1984, 113-115).

Of course, for Aquinas, the situation changes when it is the motion of a rock, which depends on
the motion of a stick, which depends on the motion of a hand, etc.

Whatever is moved must be moved by another. If that by which is moved be itself

moved, then this also must needs be moved by another, and that by another again. But
this cannot go on to infinity, because then there would be no first mover, and,
 consequently, no other mover, seeing that subsequent movers move only inasmuch as
they are moved by the first mover; as the staff moves only because it is moved by the
hand.

Since there cannot be an infinity of movers, Aquinas concludes that some mover is unmoved
(and, of course, such reasoning is not without precedent in Aristotle).2 I shall suppose the rock,
stick, etc. are something like a series of train cars, each of which is accelerated by the one in
front. Supposing a leftmost car accelerates, there is a series,

– – ––…→

where each car is accelerated. Since the leftmost car accelerates, it is accelerated by the car in
front of it; since the second car accelerates, it is accelerated by the one in front of it; etc. So the
leftmost car, and the entire train, accelerates, but without an engine.

Finally, consider the “fundamental relation” regress, familiar from Russell’s The Problems of
Philosophy. The third man regress is blocked by denying either that the form of F things is itself
F, or that the form of F things is distinct from them. While, depending on cases, one or the other
of these claims may be plausible against Plato, it is less easy to block a related difficulty
involving the relation between forms and things. The problem has application against multiple
solutions, both realist and nominalist, to the problem of universals. Russell’s version is directed
at resemblance nominalism.

If we wish to avoid the universals whiteness and triangularity, we shall choose some particular
patch of white or some particular triangle, and say that anything is white or a triangle if it has the
right sort of resemblance to our chosen particular. But then the resemblance required will have to
be a universal. Since there are many white things, the resemblance must hold between many
pairs of particular white things; and this is the characteristic of a universal (1959, 96).3

On the resemblance theory, there are no universals, and a thing is F only if it appropriately
resembles a standard exemplar. Russell claims that resemblance must therefore be universal.
However, as he suggests in another place, we might apply the same analysis again, and say a
relation is a resemblance just in case it resembles a standard exemplar of resemblance. “It is
obvious, however,” he says, “that such a process leads to an endless regress... and such a regress
is plainly vicious” (1911-12, 9 — note that his 1959, quoted above, was first published around
the same time, in 1912). Consider some color patch a; a is white only if it appropriately
resembles a standard white patch a*. So consider some relation r1 between a and a*; on the
current theory, r1 is a resemblance only if it appropriately resembles a standard resemblance r1*.
So consider some relation r2 between r1 and r1*; presumably, r2 is a resemblance only if it
appropriately resembles a standard resemblance r2*. So consider some relation r3 between r2
and r2*; etc. Maybe there is just one standard resemblance, so that r1*, r2*, etc. are the same.
Still, r1 is a resemblance only if r2 is a resemblance; r2 is a resemblance only if r3 is a
resemblance; etc. In this case, the situation might be pictured as follows:

r*
| r3/ /etc.
 | //
r2| /
|
| r1
a* a

Each relation requires another for it to be a resemblance. Thus this resemblance nominalism has
the consequence that there is an infinite series of resemblances. Russell, at least, thinks it is a
vicious regress, and that resemblance nominalism is therefore to be rejected. Similarly if, on a
realist view, things participate in forms, one might think that the relation between a form and
thing should itself participate in participation; etc. As I say above, not all agree that these series
are vicious.

In a recent paper, Oliver Black characterizes an infinite regress argument as a reduction with four
premises: three from which it follows that there is an infinite series of a certain sort, and a fourth
according to which there is no such thing (1996; cf. Sanford 1975, 520).4 For some property F,
begin by supposing existence — that there is some F thing.

(E) (∃x)Fx

In the relatively simple case of the hereditary series, the thing is a human. But also, for some
relation R, suppose generation — that any F thing stands in relation R to an F thing.

(G) (∀x)[Fx ⊃ (∃y)(Fy & Rxy)]

In the case of the hereditary series, each human is the biological offspring of others. For the third
premise, where some dots ● → ● → ● → ● are related by R when connected by an individual
arrow, say they are related by *R when there is a path along the arrows from one to the other. In
the case of the hereditary series, Rxy just in case y is a parent of x, and *Rxy just in case y is an
ancestor of x. *R is thus the ancestral of R.5 Given this, we require that nothing bears *R to itself
— that *R is irreflexive, and so a (strict) partial order.

(P) ~(∃x)*Rxx

To see how these conditions work, consider (a) - (e) below.

● ∩● O● ● ●
| |\ /\ |
| ●● ● ● ●
| | / \ |
∪● ● ● ●
… …
a b c d e

Where dots are F things, and related by R when they are connected by an individual arrow, they
are related by *R when there is a path along the arrows from one to the other. By (E), there is at
least one dot; (a) - (e) each satisfy (E). (G) requires that each dot bear R to some dot; (a) fails this
condition because the only dot does not bear R to any dot; (b) fails because the last dot on the
right-hand branch bears R to none; (c) meets this condition insofar as the only dot bears R to a
dot— namely itself. (P) rules out cases where (E) and (G) are satisfied in “R-circles”; the only
dot in (c) bears R to itself; so it bears *R to itself; so it violates (P); the dots in the main branch of
(b) all violate (P) insofar as there is a path along the arrows from any dot to itself. Thus (E), (G)
and (P) require that series continue “downward” as in, e.g., (d) or (e).

One might think we have just shown that (E), (G) and (P) guarantee the existence of infinite
descending paths, but this is not quite right. To see this — and that (E), (G) and (P) do in fact
guarantee the existence of such paths, say a path or sequence S = <s1, s2... > is a function whose
domain is an initial segment of the positive integers (an ordinal), with f(1) = s1, f(2) = s2, etc.
Where n’ is the successor of n, say a sequence S is an R-series just in case any sn,sn’∈S are such
that Rsnsn’; and say an R-series is an RF-series just in case each of its members is F. By (E),
there is an a such that Fa; so <a> is an RF-series. By (G), there is a b such that Fb and Rab, so
<a, b> is an RF-series. By (P), a and b are distinct. Etc. So there is a one-member series, a two-
member series — and there are infinitely many such series. But just as it is one thing to say there
are infinitely many integers, and another that some integer is infinite, so it is one thing to say
there are infinitely many such series, and another that some series is infinite; so to show that
there are infinitely many series (always continuing downward) is not itself to show that any
series continues infinitely. But suppose we are given linearity — that distinct objects on paths
from a are always connected by *R in one direction or another.

(L) (∀x)(∀y)[(*Rax & *Ray & x =/=y) ⊃ (*Rxy v *Ryx)]

Elements occurring on distinct branches are not connected by *R; (L) thus requires that there be
no distinct branches, so that the picture is as in (e) above. In this case, a “union” of the infinitely
many finite paths from a is a path of infinite length. Similarly, even without (L), it is possible to
“choose” a path of this sort. Supposing a restriction to some subclass of paths whose members do
not branch, again, a “union” of infinitely many finite paths from a is a path of infinite length.6
With or without (L), then, (E), (G) and (P) in fact require the existence of an RF-series with
infinitely many distinct members.

But (E), (G) and (P) are not therefore problematic. Or, at least, (E), (G) and (P) do not by
themselves lead to contradiction. To demonstrate their consistency, it is enough to find an
interpretation on which they are all true. But (E), (G) and (P) are, e.g., true on the integers, along
with (L), when R is successor and F is integer. So they are consistent. Black simply adds
finitude,

(F) There is no infinite series of Fs all related by R

as a fourth premise for reductio. Then a regress is vicious just when (E), (G) and (P) imply that
there is an infinite series of Fs related by R, and (F) that there is no such thing.7 Given this
premise, there is no problem about reaching a contradiction, but there is a problem about the
basis for the premise. (F) tells us that there is no regress, but leaves us wondering why. Maybe
something like (F) is true in every case where there is a vicious infinite regress. Even so, (F) is
question-begging insofar as the question is what’s so bad about infinite regress — for we want to
know why there is no such series. Thus Black does not answer the question we have asked.

For now, let me observe that the intuitive difficulty about vicious infinite regress arises out of
our very reasons for thinking there is a regress in the first place. Somehow, the elements of a
vicious infinite regress are introduced toward an end which remains forever unattained. The
problem is precisely that some supposed end remains unattained. If this is right, (F) or something
like it, is a consequence of reasons for regress, and so not independent of them. Insofar as (E),
(G) and (P) are consistent — and supposing that what’s bad about infinite regress exhibits itself
in contradiction — (E), (G) and (P) are therefore not the whole story about reasons for vicious
infinite regress.

Romane Clark suggests that a vicious infinite regress is characterized by “downward

dependence” (1988). From the standard truth table, the material conditional, g ⊃ Q, is true iff g is
false or Q is true; this condition does not require dependence between g and Q; so any such
conditional leaves it open whether Q depends on g, g depends on Q, or neither. Thus (G) leaves
it open whether one member of a series depends on another and, if there is dependence, what the
direction of dependence might be. On Clark’s account, what is missing from Black’s view
(though Clark writes before Black and so does not directly respond to him) is a dependence left
out by the material conditional in (G). Clark says a relation R is (upward) F-preserving if, for
any a, Rab and Fb guarantee Fa. Then,

Something is conditionally F just in case there is something to which it stands in an F-

preserving relation R which induces a partial order.... If this is the only way a thing comes
to be F, if something is only conditionally F, then, with respect to F, it is downward
dependent on its R-related heredity.... Something is categorically F just in case it is F but
not only conditionally so (1988, 173).

If the only way a thing can be F is to have R to an F thing, it is only conditionally F. A thing is
categorically F iff it is F but not only conditionally so. Given this, on Clark’s view, the typical
infinite regress argument is developed as a reductio. R induces a partial order, so (P) remains as
above. There is a premise according to which something is categorically F,

(E*) (∃x)(x is categorically F)

or, equivalently, (∃x)(x is F & x is not only conditionally F). Finally, for some upward F-
preserving relation R, it is sufficient that a target thesis implies that whatever is F is only
conditionally F. This condition may take the form,

(G*) (∀x)[Fx only if (∃y)(Fy & Rxy)]

If these premises collapse into (E), (G) and (P) then, as before, there is no contradiction.
However, the conditional in (G*) is not to be understood materially. (G*) goes beyond (G)
insofar as it constrains the direction of dependence, requiring that each member of a series
depends on the next. Clark leaves this notion at an intuitive level. Let us say we understand, and
do so as well.
Even so, it is not clear how to take Clark’s proposal. His initial idea seems to be that (E*) is
typically given and leads to contradiction with (G*). As developed above, (E*) does conflict with
G* — from (G*), whatever is F is only conditionally F, and from (E*) something is F but not
only conditionally F. Unfortunately, it is hard to see how (E*) is given in the ordinary case. It
may be given that someone (really!) is human; but it is not given that her humanity does not
depend on an infinite series. For this, we need to know what’s so bad about infinite regress. But
Clark also suggests (on the same page as the passages quoted above) that every F thing is
categorically F. If this is right, (E*) collapses into (E), and we get the contradiction for reduction
insofar as (E) is itself inconsistent with (G*) — insofar as something is supposed to be both F
and only conditionally F.

But (G*) does not contradict (E). For a relatively simple case, consider a series of cats stalking a
mouse, where any cat in the series bites the tail in front of it just in case its tail is bitten from
behind.

(In case of worry about cruelty to animals, substitute some mechanical device as below.) Now
suppose a first cat bites the mouse’s tail; then there is a second cat which bites the first cat’s tail;
so there is a third cat which bites the second cat’s tail; etc. So there are infinitely many cats.
Supposing, as we have, that infinity is not itself problematic, there should be no objection to an
infinite series of cats as such. And there should be no objection to cats that bite just in case they
are bitten. Insofar as a cat bites only if it is bitten, it is natural to think that biting is downward
dependent. But there is no contradiction supposing that the first cat bites the mouse’s tail. If the
cats are arranged in a circle, either each bites, or none bites. Similarly, in an infinite series, either
each bites or none bites. And the options seem equally plausible. God could create the series of
cats “all at once” in either state. All that is required for biting is that each cat be in proximity to
one that bites — where, seemingly, God could create them all that way. And similarly for not
biting. Given this, the downward dependence of F is not, in general, sufficient for the conclusion
that not-F. So (G*) does not imply the negation of (E). I return to this case below.

For now, notice that downward dependence (supposing we understand it) does distinguish the
truth series, snow is white, {snow is whit} is true, {{snow is white} is true} is true, etc. from
other, plausibly vicious, regresses. It is natural to say that this series is “upward,” rather than
“downward” dependent — and we might therefore call it a “progress” or “progression” rather
than a “regress” or “regression.” Perhaps downward dependence is necessary and sufficient for
regress. And there may be a kind of series which does not involve dependence at all, and so is
neither regress nor progress. But downward dependence does not distinguish, say, the motion of
a rock which depends on the motion of a stick, etc., from my humanity, which depends on my
parents’, etc. Each of these seems to involve downward dependence. As above, Clark uses his
condition to conclude that the hereditary series is vicious. So if the traditional evaluations, on
which not both are vicious, are correct, then downward dependence is not sufficient to
distinguish regresses that are vicious from those that are benign. And again, if there is no
problem about the regress of cats, Clark’s condition is not sufficient to distinguish benign from
vicious regresses. Having distinguished regresses from other series by means of Clark’s
condition, it remains to say which regresses are vicious, and which are benign.
The inadequacy of these proposals seems to cast us back on formulations, like one I use above,
according to which the elements of a vicious infinite regress are introduced toward an end which
remains forever unattained. Thus, e.g., Sanford, commenting on Passmore, and a remark by
Geach according to which the real trouble with vicious regress “arises already at the first step,”
considers that a theory which leads to regress makes some promise, and says,

The real trouble arising already at the first step is that of making no progress. We should
see this straight away. If we do not, we may see it after realizing that no number of steps,
not even an infinite number, makes any progress toward explaining, defining, analyzing,
or accounting for something. Drawing attention to an infinite regress can thus have a
function even though the real trouble is not due to the regress (1984, 96; cf. Passmore
1961, 19-37, and Geach 1979, 100-101).

Related points are sometimes made in terms of human capabilities: a series must fail to deliver
on some promise, because it is impossible for humans to complete an infinite series. No doubt, it
is impossible for humans to complete at least certain infinite series of tasks.8 But the cases we
have considered, at least, do not have to do with human capabilities; they have rather to do with
the existence of truths, ancestors, movers, and relations. And it is not necessary to cast the
current proposal in terms of human capabilities: a regress is vicious when even an infinite series
fails to deliver on some promise. Perhaps it is obscure how “lack of progress” is to be
distinguished from “downward dependence” — if one element of a series depends on the next
just because there is no progress, lack of progress may seem to go hand-in-hand with downward
dependence. However, I think these suggestions are on the right track. It is the task of the next
section to develop and defend this claim.

II. A Positive Theory

In this section, I propose a theory, exhibit it in application to (standard idealizations of) relatively
well-understood physical models, and comment on the result. Paradoxically, one advantage of
the cases is that they can be relatively complex while, at some level, philosophical examples may
be so simple as to obscure distinctions that matter. The examples have the advantage that devices
are subject to simple laws, and make possible different series from a small set of primitives. But
for infinite series, idealized laws are applied to current, and the like, of any finite magnitude —
though in reality such are discrete; so the models are just the idealized devices. To manipulate
cases, we need a little math and physics. But this should not be a problem. We return to
philosophical examples in the following section.

Recall that an R-series is a sequence with adjacent members related by R. Developing Sanford’s
suggestion that the real trouble with infinite regress “is that of making no progress,” I propose
that a valid infinite regress argument arises when premises imply that there is an R-series which
both is, and is not, adequate to some end. The reasoning involves considerations of three sorts:
(i) For some property F and relation R with irreflexive ancestral *R, there are adequacy premises
according to which there is an R-series whose first member is F. So far, it may be open whether
the series has just one member, or many. (ii) There are underlying premises which specify some
features of the members. Ordinarily, the features are “relatively intrinsic” insofar as they do not
depend on relations to other members of the series. And (iii) there are linking premises that fix
some functional relation between the underlying features and adequacy. As we shall see, a lot
hangs on the nature of this linking relation. An R-series is either finite or (countably) infinite. If
it follows from (ii) and (iii), by induction or whatever, that no finite Rseries is adequate, then the
series is a regress and infinite. If, in addition, it follows from (ii) and (iii) that neither does an
infinite series satisfy (i), then the original premises are inconsistent, the regress is vicious, and at
least one premise must be rejected. As we will see, however, not every series characterized by
such premises is a regress, and not every regress is vicious.

To put some flesh on these bones, let us begin with the series of cats turned around so that the
mouse bites the first cat, who bites the second, etc. Imagine that the force of each cat’s bite is in
proportion to the force with which it is bitten. Then the cats are like a series of electronic
amplifiers. An amplifier is a device with input voltage D, output voltage Q, and gain a, where Q
= a * D. Such devices may be connected in series as follows.

D1 |\a1 Q1 – D2 |\a2 Q2 – D3 |\a3 Q3 – D4 |\a4 Q4 – D5 |\a5 Q5 –

|/ |/ |/ |/ |/
… … … … …

As configured above, D1 is fixed at zero (ground), so Q1 - Q5 are zero as well. If there is a

power source so that D1 has some non-zero value V, we may reason from values at one amplifier
to values at the next. For any n, Qn = an * Dn and, since they are connected directly, Qn = DnØ.
So,

Q5 = a5 * D5
= a5 * Q4 = a5 * a4 * D4
= a5 * a4 * Q3 = a5 * a4 * a3 * D3
= a5 * a4 * a3 * Q2 = a5 * a4 * a3 * a2 * D2
…
= a5 * a4 * a3 * a2 * a1 * D1

where equalities in the horizontal direction are because Qn = an * Dn, and in the vertical because
Qn = Dn’. And, in general, D1 * a1 * a2 *...* an = Qn. Thus the state of each amplifier is fixed
once values are given for D1 and the gains. And similarly for the cats, where inputs and outputs
are like forces, and gains the proportions with which cats react to being bitten.

And similarly in the infinite case (if you like imagine that, through a miracle of miniaturization,
amplifiers get progressively smaller, so that the series fits into a finite space). To make the case
more specific, suppose the gains are arranged into a series 2√a, 4√a,... where for any n, an =
a^1/2^n. One reason for this choice is to keep the arithmetic relatively simple: For a series of n
members, the sum 1/2^1 + 1/2^2 +...+ 1/2^n = 1 - 1/2^n;9 and a product, a^1/2^1 * a^1/2^2 *...*
a^1/2^n is equal to a^S, where S is the sum of the exponents, so a^1/2^1 * a^1/2^2 *...* a^1/2^n
= a1-1/2^n. For an infinite series, the sum, 1/2^1 + 1/2^2 +... is the limit of finite partial sums, so
1/2^1 + 1/2^2 +... = 1, as 1 - 1/2^n approaches 1; and a product a^1/2^1 * a^1/2^2 *... is the limit
of finite partial products, so a^1/2^1 * a^1/2^2 *... = a1 = a, as the sum of exponents approaches
1.
Now suppose some amplifier system is such that: (i) There is an infinite R-series of amplifiers
with D1 = V. (ii) Individuals in the series are such that for any n in the series’s domain, an =
a^1/2n. And (iii) individuals in the series are linked so that, as above, with Π an extended
product function, for any n, Qn = Π[a1, a2 ... an] * D1. Then, in the infinite case, where Qω is
not a physical value in the series, but rather the limit of the outputs, Qω = Π[a1, a2...] * D1.
From (ii) and (iii), with the arithmetic from above, for any n, Qn = a^1-1/2^n * D1, and Qω = a *
D1. So with (i), Qn = a^1-1/2^n * V and Qω = a * V. Given a value for V, then, values for Qn
and Qω are fixed. And there is nothing inconsistent about this: (ii) and (iii) give us a functional
connection between the value of D1 and the values of Qn and Qω; but they do not thereby fix the
value of D1, and so do not force contradiction with (i). Rather, (ii) and (iii) make other values a
function of D1. This will be my standard example of an infinite progress. It is infinite because (i)
with (ii) and (iii) fix values “all the way out.” It is a progress insofar as values for later members
are determined by ones before, not the other way around. (With the details of this case under our
belts, details for others should be relatively straightforward.)

We get closer to the example of cats which simply bite or not by setting each gain equal to one,
and treating the amplifiers as two-state systems. For this, it is enough to let a voltage less than or
equal to some cutoff (say, ground) be the value low (L), and any voltage above it high (H) —
where amplifiers correspond to cats, output H to biting, and L to not biting. Then everything
works as before. With the gains equal to one, for any n, D1 = Qn = Qω. So if D1 = H, all the
values are H; if D1 = L all the values are L. And there is no contradiction.

For a regress, let us turn the cats back around so that they stalk the mouse. Again, we will model
the cats with amplifiers. As a first step, however, consider a bucket brigade in which members
pass their bucket, if they have one, to the next on command of a captain (who is like the
coxswain). This works like a series of electronic “flip-flops” arranged to form a “shift-register.”
Such systems “shift” values so that we will be able conveniently to isolate the effects of one
stage upon the next. The input D and output Q of a simple flip-flop take either of the states high
H or low L. The clock input cycles on and off. The flip-flop is such that the state of Q after c+1
clock cycles is equal to the state of D after c clock cycles. As pictured to the side, D is fixed at L
and, whatever its initial condition, Q is therefore L at every clock cycle after the first. Such flip-
flops may be arranged in series, to form a shift-register.

clock
|
––
Q D

clock
C1 C2 C3 C4 C5
| | | | |
Q1––D1 Q2––D2 Q3––D3 Q4––D4 Q5––D5 –
| | | | | |
… … … … … …
Supposing, e.g., that Q1 - Q5 are initially in the states LHLLH, after one clock cycle they are
HLLHL, after two LLHLL, after three LHLLL, after four HLLLL and after five, LLLLL. Thus the
initial pattern is “shifted” through the structure from right to left, with the “vacated” locations
taking the value from D5. If Q1 is connected around to D5, the pattern shifts round and round,
returning to its initial state every five clock cycles. Thus high values are like buckets passed from
one member to the next. (In my youth, I constructed a display which used shift registers to cycle
lights around like this in different patterns.) Say a pattern is stable just in case all output states
remain the same from one clock cycle to the next. Then LLLLL is the only stable pattern for the
above finite device. If Q1 is connected to D5, then LLLLL and HHHHH are stable. And, if the
series is infinite, all H, and all L, are stable.

This infinite case is related to “Hilbert’s Hotel” examples as applied by William Lane Craig
against the possibility of an actual infinite. His argument is by reductio: Thus, e.g., suppose a
hotel with infinitely many rooms and a guest in each room. Say a new guest arrives. The
proprieter vacates the first room by moving its occupant to room two, the occupant of room two
to room three, and so forth. So there is space for the new guest. Now suppose the first guest
checks out. The hotel remains full if the rest are returned to their previous places. Craig finds
these results “absurd”; if the rooms of an actually infinite hotel are full, there isn’t room for an
additional guest, and if rooms are empty, there are not enough guests to fill the rooms. He
concludes Hilbert’s infinite hotel is not actually possible.10 But there is a straightforward
response against this reasoning: An infinite series is “unbounded” in the sense that it has no last
member, and so can be put in correspondence with the integers. It is because the series is
unbounded that each room has a next, so that the new guest can be added. And there are ways of
removing individuals from such a series, some of which preserve this property, and some of
which do not. So long as unboundedness is preserved, as when just the initial guest (or every
other guest) is removed, the series remains infinite — and so capable of correspondence with the
integers (and rooms); if unboundedness is not preserved, as when, say, all the members after the
tenth are removed, the series is no longer infinite. Thus the nature of infinity explains why
addition or subtraction from an infinite series works the way it does. Craig is right to hold that
the possibility of Hilbert’s Hotel entails the possibility of shifting occupants. However, given an
assumption that the hotel is infinite, there is no mystery about shifting: we admit the
consequence, but deny that it is absurd. Having assumed the possibility of infinite series as such
— and thereby having rejected Craig’s conclusion against infinity, we already commit ourselves
to the possibility of “shifting” to and from infinity as above.

Removing the clock gets us closer to the series of cats. For this, let us return to a series of
amplifiers, now reversed to reflect the direction of the cats.

– Q1 a1/| D1 – Q2 a2/| D2 – Q3 a3/| D3 – Q4 a4/| D4 – Q5 a5/| D5 –

\| \| \| \| \|
… … … … …

Though we no longer “see” the pattern being shifted through the structure, this device works
very much like the shift register. For the above finite system, all L is the only stable state. If Q1
is connected around to D5, or the series is infinite, all H and all L are stable. Accounting for the
gains,
Q1 = a1 * D1
= a1 * Q2 = a1 * a2 * D2
= a1 * a2 * Q3 = a1 * a2 * a3 * D3
…
= a1 * a2 * a3 * a4 * a5 * D5

where equalities in the horizontal direction are by Qn = an * Dn, and in the vertical by Dn = Qn’.
As configured above, D5 = 0, so Q1 - Q5 are 0 as well.

Now suppose, (i) there is an R-series of amplifiers with Q1 = V for some V > 0. (ii) Individuals
in the series are such that for any n in the series’s domain, an = a^1/2^n. And (iii), as above,
individuals are linked so that for any n, Q1 = Π[a1, a2 ... an] * Dn; and for an infinite series, Q1
= Π[a1, a2 ... ] * Dω, where Dω is the limit of the inputs. As above, if there is a last member of
any such series, it has input 0. From (ii) and (iii), for any n, Q1 = a^1-1/2^n * Dn, and for an
infinite series, Q1 = a * Dω. So by (i), V = a^1-1/2^n * Dn, and if the series is infinite, V = a *
Dω. But for a series with just n members, Dn = 0 and, since V > 0, given V = a^1-1/2^n * Dn,
this is impossible. We can thus reason from the initial value V, to a positive value for Dn —
where this contradicts the assumption that the series has just n members, and so Dn = 0; since the
assumption that the series is finite leads to contradiction, the series is infinite. Or, put the other
way around, in finite cases, we can reason from Dn = 0 to a zero value for V, where this
contradicts adequacy. Thus (ii) and (iii) suffice to fix a value for Q1, and there is room for
contradiction. When the series is infinite, however, contradiction evaporates: we have V = a *
Dω, but without a value given for Dω. So again (ii) and (iii) do not determine a value for Q1, and
there is no room for contradiction. This will be my standard example of a benign infinite regress.
It is infinite because the value at every finite stage requires one at the next. It is a regress
because, in finite cases at least, later stages suffice to determine values at ones before. It is
benign since, in the infinite case, (ii) and (iii) are consistent with (i).

Again, we get closer to the original example of the cats by setting the gains equal to one, and
treating the amplifiers as two-state systems. With the gains equal to one, for any n, Q1 = Dn =
Dω. Suppose Q1 = H. Then, for an n member series, Dn = H; but in a finite n-member series, Dn
=/= H; this is impossible, so the series is not finite. However, there is no problem in the infinite
case, as nothing prevents Dω from taking the same value as Q1. So each of the cats can bite, and
the regress is benign.

For a vicious regress, begin with a series of tributaries flowing into a single river, where we are
interested in total flow from the tributaries. This is like a summation of electronic currents.11
Given some path to ground, in this case, an individual consists of a voltage source and resistor,
where the voltage is regulated so that potential across the resistor is constant and current through
the resistor is thus fixed at some constant value. Such individuals may be strung together as
follows.

---------v1------v2------v3------v4------v5----
| | | | |
v v v v v
 I1 I2 I3 I4 I5
| | | | |
<---------------------------------------------------
Q1 D1 Q2 D2 Q3 D3 Q4 D4 Q5 D5
|
R
…

For any n, the total current at Qn is the sum of In and Dn. So in this case,

Q1 = I1 + D1
= I1 + Q2 = I1 + I2 + D2
= I1 + I2 + Q3 = I1 + I2 + I3 + D3
…
= I1 + I2 + I3 + I4 + I5 + D5

where equalities in the horizontal direction are by Qn = In + Dn, and in the vertical by Dn = Qn’.
As here configured, D5 = 0, and Q1 is just the sum of I1 - I5.

Now suppose, (i) there is an R-series of current sources with Q1 = I0. (ii) Individuals in the
series are such that for any n, In = 1/2n. And (iii) individuals are linked so that for any n, Q1 is
the extended sum, Σ(I1, I2...In) + Dn. As above, if any such series has a last member, its input
Dn = 0. Similarly, in the infinite case, the only sources of current are from individual members of
the series, so Q1 = Σ(I1, I2,...). Now suppose I0 > 1. Then, in any finite case, I0 = 1/2^1 + 1/2^2
+ ... + 1/2^n = 1 - 1/2^n. But, 1 - 1/2^n < 1, and this is impossible. So the series is infinite. But
Σ(1/2^1, 1/2^2...) = 1.12 So again there is contradiction. This will be my standard example of a
vicious infinite regress. In both finite and infinite cases, later stages suffice to determine values
at ones before. It is vicious since in neither the finite nor infinite case are (ii) and (iii) consistent
with (i).

So, on this view, a typical regress argument has adequacy, underlying, and linking premises. In a
progress, underlying and linking premises do not fix relevant features for the first member.
Rather, underlying and linking premises, with claims about the first member, fix features for
ones that follow. In a regress, underlying and linking premises do fix relevant features for the
first, at least in finite cases. Thus there is grist for contradiction with adequacy. A regress is
vicious when underlying and linking premises are inconsistent with adequacy in both the finite
and infinite cases.13

Notice that the essential difference between benign and vicious cases is not that benign regresses
somehow involve infinite products, and vicious regresses summations. If, e.g., in a regress of
amplifiers (or cats), voltages were augmented by some value an, then for any n, Q1 = Σ(a1,
a2...an) + Dn, and in the infinite case, Q1 = Σ(a1, a2...) + Dω. With Dn = 0 for a last member,
there might be contradiction in every finite case, without contradiction when the series is infinite;
so the series related by summation would be benign. Similarly, series related by product
functions might be vicious.14 The important point is rather about the way linking functions
“collect” contributions of the members. In a vicious regress, the initial value is completely
accounted for by contributions of the members; contributions of the members are “collected” so
that the initial value is a direct function of them. For a benign regress, the initial value is not
entirely accounted for by features of the members. Rather, though finite series are sufficient to
force some initial value, an infinite series may have distinct stable states. In the vicious case, the
series has a determinate output value, incompatible with adequacy. But where a series has
multiple stable states, there may be no reason to deny that the series takes a state compatible with
adequacy.

One might desire some additional characterization of the dependencies and particular contexts
which result in one functional relation rather than the other. However that is wellbeyond the
scope of this paper. The current theory identifies a class of relations which much be in place for
vicious infinite regress; but that is not itself an account of the metaphysics to set up those
relations. Formally, the bottom line may be just that contexts are characterized by the different
functional relations.

To emphasize this point that vicious regresses may involve different functional relations,
consider a couple more cases. Instead of a voltage source and resistor, suppose an individual
consists of a voltage source and diode (where a diode lets current pass in only one direction).

---------v1------v2------v3------v4------v5---
| | | | |
v v v v v
| | | | |
---------------------------------------------------
Q1 D1 Q2 D2 Q3 D3 Q4 D4 Q5 D5
|
R
…

In this case, the voltage at Q1 is equal to the maximum of the individual voltages, Max(V1, V2,
V3, V4, V5). This works like pressures from columns of water, with check valves to prevent
flow back toward the sources; then the output rises just to the height of the highest input column.
Suppose we are given some such system with (i) Q1 = V; (ii) any Vn in the series < V; and (iii)
Q1 = Max(V1, V2...). Then a finite n member series has V = Max(V1, V2...Vn) < V. But this is
impossible, so no such series is finite. But similarly, in the infinite case, Max(V1, V2...) < V; so
the premises are inconsistent. Suppose an individual consists of a voltage source and switch, with
the voltage sources all fixed at some constant value V.

V>----------------------------------------------
| | | | |
o o o o o
S1 S2 S3 S4 S5
| | | | |
---------------------------------------------------
Q1 D1 Q2 D2 Q3 D3 Q4 D4 Q5 D5
|
R
…

Then the voltage at Q1 = V iff one or more switches is closed — iff the extended disjunction,
V[Closed(s1), Closed(s2), Closed(s3), Closed(s4), Closed(s5)] is true. Say we are given some
such system with (i) Q1 = V; (ii) no sn in the series such that Closed(sn); and (iii) Q1 = V ↔
V[Closed(s1), Closed(s2)... ]. Given (ii), for a finite n member series, V[Closed(s1), Closed(s2)...
Closed(sn)] is false; so by (iii), Q1 =/= V, which contradicts (i). But similarly, in the infinite
case, V[Closed(s1), Closed(s2)...] is false and the premises are inconsistent.

Perhaps these last cases are too obvious, and so motivate the intuition, expressed by
Geach, that there is something trivial or uninteresting about infinite regresses. Thus, e.g.,
V[Closed(s1), Closed(s2)... ] iff (sn)Closed(sn), and we are given ~(∃sn)Closed(sn); these
conflict directly, without the rigmarole of finite and infinite series. But note first, that there is
nothing wrong about reasoning with the finite and infinite cases — it is at least one way to
expose contradiction lurking in premises. Second, not all arguments of the proposed regress form
are so trivial. As we have seen, reasons which result in contradiction in finite cases may or may
not result in contradiction in the infinite. So the division into finite and infinite cases is not
superfluous. And the difficulty with infinite regress need not be that of making no progress.
There is progress in the vicious summation of currents case — only not enough; each member of
the series makes some positive contribution, though the sum of contributions remains inadequate.
And even in cases where there seems to be no progress from one step to the next, the situation
may be relatively complex. There seems to be no progress from one step to the next in the
regress of cats (or amplifiers), and similarly with the switches. Yet one is vicious and the other
not. In the regress of cats, we go from one finite stage to the next without biting; but in the
infinite case, all the cats may bite. In the regress of switches, we go from one finite stage to the
next with zero output voltage; and voltage remains at zero in the infinite case. In general, the
difficulty is making inadequate progress, where adequacy may be determined in relatively
complex ways.

I conclude this section with some brief remarks about what has been accomplished, and the
shape of argument to come. First, all our series — progress and regress, benign and vicious —
may be described by premises in the style of Black. We begin with premises according to which
there is an R-series whose first member is F; so we accept (E) and (P). And, insofar as we are in
a position to reason from one stage to the next, we accept something like (G) — though we have
seen cases where members do not share some constant property F, but rather have properties that
are indexed to position in the series, varying from one member to the next. In regress cases, there
is downward dependence as well. So far, then, the point is not that Black or Clark somehow go
wrong in their description of infinite regresses. Rather, it is that adequacy, underlying and
linking premises drive the relations described in (G). In vicious cases, these premises result also
in contradiction — something that (E), (G), and (P), with downward dependence, do not by
themselves do.

Second, on this account, a theory which results in vicious infinite regress is reduced to absurdity.
From this, it follows that some premise must be rejected. Of course, one might have reasons for
rejecting one premise rather than another. But the brute fact of inconsistency does not require
that one premise, rather than another, should go. In philosophical cases, there will be
philosophical reasons for rejecting specific premises. The current theory is an account just of
conditions for inconsistency as such.

Finally, one might object that series I say are not vicious are, nonetheless, impossible (Craig or
Clark might reason this way). Of course, I do not prove that series I count as progress and benign
regress are possible. However, I do try to motivate or ground claims about possibility by our
assumption that there is nothing the matter with infinite series as such, together with models for
the different series. I thus try to motivate the suggestion that consistency for our premises tracks
a larger sense of possibility. Suppose this is right, and valid regress arguments are generally
characterized by premises of the sort I describe. Then debate about regresses shifts to the
premises. If some situation is mistakenly described as a progress or benign regress, it is natural to
object by showing how the description is mistaken and exhibiting whatever strengthened
premises result in contradiction. Given that there are contradictions in the neighborhood to be
had, an infinite regress argument, if valid, should include whatever premises are required to
reach contradiction. However, as we will see, it remains possible to disagree about a theory’s
proper consequences, and so about whether a regress is vicious or benign.

III. Philosophical Applications

In this section, I merely scratch the surface of the arguments with which we began. In each case,
variant formulations, and a variety of objections and replies, go unexamined. The point is not so
much to obtain definitive results with respect to particular arguments, as to exhibit the overall
shape of the theory’s application to philosophical cases.

The simple truth series appears as non-vicious because it is not a regress at all; rather it is a
progress — or so it seems on a “backwards” or “upwards” looking account of truth along the
lines of the correspondence theory. As in the case of the mouse which bites the cats, a value at
one stage is simply propagated out to the next. Thus we are given a series of stages which take as
input a state or situation, and output a truth value to a proposition — where the proposition with
its truth value is the input to the next. Say p1 is the proposition that p and pi+1 the proposition
that pi is true. Then we are given a series of stages c1, c2,... which take as input a truth value for
pi and assign a value to pi+1, where by (Tp), pi+1 is true iff pi is true. Stages are linked so that
the output value from one stage is the input to the next. Thus, given an initial stage that makes p1
true, by (Tp), p2 is true; but this makes possible another application of (Tp); so p3 is true; and so
forth. Given initial situation p, we might picture the series as follows,

P – |\c0 T – D2 |\c1 T – D3 |\c2 T – D4 |\c3 T – D5 |\c4 T –

|/ P1 |/ P2 |/ P3 |/ P4 |/ P5

If the initial situation were not-p, each of the outputs would be false. So the stages are governed
by (Tp), and linked so that the output of one is the input to the next. Underlying premises about
the stages, together with the way they are linked, make other values a function of the first. Thus
it is a progress, not a regress. And there is no contradiction. The key to this series is that truth, on
the supposed account, is a “backwards” or “upwards” looking notion. Also, depending on our
view of propositions, there is nothing “uneconomical” or “unbelievable” about this series. What
would be odd, is if the series were somehow to end, or to have one member true, and the next
not. The situation might change on some other account of truth, but that is another story.

The hereditary series appears as a regress, but benign. In this case, we have a series of generative
events, arranged so that the input P (“parent”) to one is the output C (“child”) of the next. Then
we might see the series as follows,

– Q1 g1/| P1 – Q2 g2/| P2 – Q3 g3 /| P3 – Q4 g4 /| P4 – Q5 g5 /| P5 –
\| \| \| \| \|

The stages are linked so that for any n, Pn = Cn’. Say generative events do not preserve humanity
perfectly. Then, where the “humanity” for any output value falls between 0 and 1, and en is some
positive or negative error value, let us say that the stages have an underlying character so that Cn
= Pn + en. Then for a finite series,

C1 = e1 + P1
= e1 + C2 = e1 + e2 + P2
…
= e1 + e2 +... + en + Pn

where equalities in the horizontal direction are by Cn = Pn + en, and in the vertical by Pn = Cn’.
With the error values, there is the possibility that C1 = 1 and some Pn = 0 without contradiction.
But, if as in the original problem, we add that each ei = 0, then it is inconsistent to suppose that
C1 = 1 and some Pn =/= 1. Insofar as underlying and linking premises result in contradiction
with adequacy, it is a regress. But when the series is infinite, contradiction evaporates: for limit
value Pω, we have C1 = Pω, but without a value given for Pω. So in the infinite case, underlying
and linking premises do not determine a value for C1 and the regress is benign.

In contrast, the rock moved by the stick, etc. — construed as a series of train cars, is vicious. The
leftmost car is accelerated by the car in front of it, which is accelerated by the one in front of it;
etc. So far, then, it may seem as though acceleration is “shifted” from one car to the next and,
though there is a regress, it is benign. However, with a bit of (anachronistic, as applied to
Aristotle or Aquinas) physics, it is clear that the cars cannot accelerate. If the train is infinitely
massive, then no amount of force makes it move — or at least our ordinary notions from physics
do not apply. So suppose each car is half the size and mass of the one it pulls; then the total size
and mass of the train is finite, and we still have the question about how it moves. The
acceleration of the train is equal to the total force applied to it divided by its mass, a = F/m. The
cars are linked so that the force applied to an n-member train is the sum of the forces applied by
each of the cars, Fn = Σ(F1, F2...Fn). But if no car is an engine, we have the underlying fact that
each of the forces, and so their sum, is zero. And this is impossible if the train accelerates. So it
is a regress. But similarly, in the infinite case, the sum of the forces is zero, and the train does not
accelerate. So the regress is vicious. If we are tempted to see the regress as benign, I think it is
because we are tempted to see it under descriptions that do not properly sum the forces from the
cars.
The fundamental relation regress is more difficult — not so much because of problems with the
theory of regresses, but because the theory of resemblances is itself obscure. Russell thinks that,
on a strict resemblance nominalism, a relation is a resemblance only if it resembles some
standard. Thus each relation requires another in order for it to count as a resemblance. But there
are different ways to see this. On the one hand, we may think that a member of the series can be
a resemblance iff some member of the series is a resemblance of its own intrinsic nature — iff it
is a “mover” for the other members. Then stages are linked so that res(r1) ↔ V[mover(r1),
mover(r2)...]. Then, given the underlying premise that no member of the series is a resemblance
of its own intrinsic nature, in both the finite and infinite case, there is conflict with the adequacy
premise that the first member is a resemblance, so that the series is both a regress and vicious.
Something like this is Russell’s view.

But it may be that each relation is simple a “child” of the next. Suppose the underlying
premise that a relation is a resemblance iff it resembles the standard resemblance, res(ri) ↔ ri ≅
r*, where stages are linked so that ri ≅ r* ↔res(ri+1). In this case,

res(r1) ↔ r1 ≅ r*
↔ res(r2) ↔r2 ≅ r* etc.

So in the finite case, res(r1) ↔ res(r2) ↔ ... res(rn) ↔ rn ≅ r*. In this finite case, with no
resemblance after the last, rn ≅ r*; so there is conflict with adequacy, and the series is a regress.
But in the infinite case, res(r1) ↔res(r2) ↔ ... res(rω) and there is nothing to prevent the limit,
and each member, of the series from being a resemblance — so the regress, on this
interpretation, is benign.

In favor of Russell’s interpretation, it is natural to think that if the resemblance theory is to count
as a viable response to the problem of universals, a relation must be a resemblance purely as a
result of the (relatively) intrinsic natures of members of the series, together with the way they are
connected. If the series is benign precisely because the series has multiple stable states, it may
very well not be a successful account of universals. Perhaps, though, the reason such regress
arguments remain obscure is precisely because we have nowhere near so much clarity about
resemblances (justification, or the like) as about relations between parents and children or train
cars.

But there is yet another option. If you are six feet tall, and I am six feet tall, the way we are,
individually, guarantees that we are the same height. The fundamental facts seem to be our
individual heights, and the sameness arises because of the way we are individually. Similarly, a
resemblance theorist might say that the resemblance between color patches arises because of the
way they are individually. And there may be resemblances among relations but, again, because
of the way they are individually — so that resemblance is an “internal” relation in the sense of
Lewis (1986, 62ff). In this case, there may be a series of resemblances, but the series appears as a
progress, not a regress, with res(ri) ↔ri ≅ r*, and ri ≅ r* ↔res(ri+1) but no reason to deny that
rn ≅ r*. Quine pushes in this direction in a famous passage from his (1948). He allows that there
are red houses, roses, and sunsets, but denies that they have anything in common “except as a
popular and misleading manner of speaking.” That the houses and roses and sunsets are all of
them red is taken as “ultimate and irreducible.” Presumably, since they are all red, they resemble
one another in this respect. But they are not red because they resemble (and Quine would not
want to talk as though there are resemblances, except as a misleading manner of speech). Rather,
they resemble because they are all red. Given difficulties as above, I suspect that a viable
resemblance theory would need to be developed along these lines. And similarly for other
approaches to the problem of universals. But that is well beyond the scope of this paper. All I am
after is that we require clarification of the resemblance theory, before we can sensibly evaluate
related regress claims.

I take it as evidence for this approach to regress arguments, that its results coincide with
traditional evaluations — and even that results are indeterminate where tradition is less than
clear. One might object that the theory is therefore philosophically impotent. But this would be a
mistake. First, the account of regress arguments tells us what vicious infinite regress amounts to,
and so how it matters; insofar as a theory which results in vicious infinite regress is inconsistent,
it is reduced to absurdity in the usual way. Further, the theory tells us what to look for in the
evaluation of regress arguments, and so guides our approach to argument evaluation; their
evaluation depends on adequacy, underlying and linking premises.15 Even in cases where results
are not clear, the account points to ways in which a theory should be developed, before we can
even sensibly evaluate regress objections and so say whether a theory is coherent; and the
direction in which we ultimately develop some theory, may itself be constrained by regress
concerns. And this, I take it, is progress indeed.16

Notes

1 But see e.g. W. L. Craig (1979) and other places, who grants the point about mathematics, but
develops and defends traditional arguments against the actual infinite. I am inclined to think
these arguments fail. Pace Craig, difficulties raised by ancient and medieval philosophers seem
resolved with the rise of contemporary mathematics. See p. 16 below.

2 Aquinas, Summa Theologiae I.2.3, trans. Anton Pegis. See, e.g., Aristotle, Metaphysics 994a1-
19. In Aquinas, the distinction between this and the previous case is between accidental and per
se regresses. For discussion, see (Day 1987) and (Brown 1966).

3 Strangely, Russell seems not to have seen the possibility of application against his own view.
For general treatment, see (Armstrong 1974).

4 Even more recently, (Nolan 2001) argues that vicious infinite regresses are characterized by
various theoretical difficulties, including ontological extravagance. Of course, I do not deny that
even benign regresses and series that are not regresses at all may suffer from such difficulties.
My main response to Nolan is the account on which there is more than this to vicious infinite
regress.

5 The ancestral *R of a binary relation R, is the relation such that (i) for any x and y, if Rxy then
*Rxy; (ii) for any x, y and z, if *Rxy and *Ryz, then *Rxz; and (iii) for no other x and y is it the
case that *Rxy.
6 Let A be the set of all RF-series, and consider some B ⊆A such that for any C,D ∈B either C
⊆D or D ⊆C. For an application of the axiom of choice (in the form of Zorn’s lemma), we
require that ∪B ∈ A. If <n, c > and <n, d > are in ∪B, then <n, c >C∈B and <n, d >∈D ∈B; but C
⊆D or D ⊆C; either way, both <n, c > and <n, d > are members of some one function; so c = d;
so ∪B is a function. The domain of a union of functions is the union of their domains, and a
union of initial segments of the integers is an initial segment of the integers; so the domain of ∪B
is an initial segment of the integers; so ∪B is a sequence. If cn’ is in B then cn’∈C ∈B; since
the domain of C is an initial segment of the integers, cn∈C; since C is an R-series, Rcncn’; so
B is an R-series. Finally, since each member of ∪B is a member of an RF-series, each
member of B is F; so B is an RF-series. So ∪B ∈A. So by Zorn’s lemma, there is an RF
series M ∈A which is not a subset of any other member of A. Suppose M is finite; then for some
i, M = <m1, m2,...mi>. Since each member of M is F, mi is F; so by (G), there is an a such that
Fa and Rmia; so M ∪ {{i+1, a }} is an RF series; so there is an RF-series <m1, m2,...mi,
mi+1>of which M is a subset. This contradicts the maximality of M; reject the assumption: M is
infinite.

7 But on Black’s terms, ‘vicious’ in ‘vicious infinite regress’ is redundant. He sets things up so
that the notion of an infinite regress itself has ‘dyslogistic’ force (115-124).

8 In such cases we might have an independent premise like Black’s (F); see also Nolan’s
discussion of regresses with known finite domains (2001, 531-32). In these cases, the regresses
might or might not be vicious in the sense discussed below.

9 For the main argument above, it is sufficient simply to accept the claims about arithmetic.
However, this result should be familiar to philosophers from Zeno’s paradox of dichotomy. If
one goes half some distance, half the remaining, etc., after any n steps, 1/2n the original distance
remains. More formally, arguing by induction, Σ(1/2^1) = 1/2 = 1 - ½^1; suppose
Σ(1/2^1...1/2^k) = 1 – 1/2^k; then Σ(1/2^1...1/2^k, 1/2^k+1) = 1 – 1/2^k + 1/2^k+1 = 1- 2/2^k+1
+ 1/2k+1 = 1 - 1/2k+1; so for arbitrary n, Σ(1/2^1...1/2^n) = 1 - 1/2^n.

10 See, e.g., (Craig 1979, 83-87 or 1993, 12-16). Of course, there may be other reasons for
denying the possibility of an infinite hotel. But problems about infinite hotels as such are not
automatically problems about regress. Craig’s reasoning is important in this context insofar as it
may seem to apply against my example of “shifting from infinity.”

11 For another case with the same structure, consider a series of weights placed upon a scale.
Then the total force on the scale is equal to the sum of the forces from the weights, and the
system works like the summation of currents.

12 There is room for caution on this point: On the standard theory for real numbers, an infinite
sum is the series of its partial sums, where the series of sums is, or designates, a real number. So
it is not clear how or whether the sum counts as a straightforward total of the infinitely many
members. Consider, e.g., “conditionally” convergent series as, 1 - 1/2 + 1/3 - 1/4 + 1/5... and 1 +
1/3 - 1/2 + 1/5 + 1/7 - 1/4... which have all the same members, but different limits and so
different sums (see, e.g., Knopp 1928, 102-103 and 139ff). Such concerns are immaterial to the
overall view illustrated by this example, where the important point is just that the series has some
definite output value. Still, in the above case, it seems reasonable to think that the value is one.
Setting aside the point that current is discrete at the atomic level, the current at Q1 has some real
value; it can not be a value less than one since, at some stage, the current must exceed any such
value; it can not be a value more than one, since one is the limit of the series. So the value is one.

13 The above example of a progress takes the infinity of the series of amplifiers as a premise.
The progression is not from one amplifier to the next, but from the value of one output, to the
value of the next. If the series were finite, the values would progress through as many amplifiers
as there were and stop. In contrast, the regress of amplifiers does not take the infinity of the
series as a premise; the existence of members is driven by contradiction in every finite case. But
this is not so if Q1 = L. Since Q1 for every finite series = L, from Q1 = L with (ii) and (iii), there
is no requirement that there be any other amplifier. But, however many amplifiers there are, a
first value of L depends on a value of L at the next, all the way to the last, if there is a last. On the
account I offer, it is a regress insofar as the first value depends on the rest. It is benign insofar as
there is no contradiction in the infinite case. A vicious regress requires conflict between (i) and
the results of (ii) and (iii).

14 For this, we might imagine analog multiplication devices such that Qn = Vn * Dn, arranged in
parallel as for the summation of currents. Then, for the infinite case, Q1 = Π(V1 * V2 *...) and
with underlying premises about the values of V1, V2..., there is potential for conflict with
adequacy.

15 So, in a recent discussion, (Gillett 2003; Klein 2003) Carl Gillett argues that Peter Klein’s
“infinitism” as a theory of epistemic justification is subject to regress objections – on grounds
something like downward dependence of justification. Klein responds in part, that series
characterized by downward dependence are not vicious – using as his example a series like the
rock being moved by the stick, etc. But we have held that (i) downward dependence is not by
itself sufficient for vicious regress; and (ii) the series of train cars, at least, is in fact vicious. So
far, then, neither the attack nor the defense succeeds. (But in note 5 on p. 713, Gillett suggests
that downward dependence forces only the existence of an actually infinite series, so that “all the
problems surrounding the actual infinite consequently dog infinitism.” This suggests a premise
like Black’s – that there is no infinite series of the appropriate sort. But Gillett does not say what
the problems are supposed to be; and we have held that premises in the style of Black’s
themselves follow from reasons driving vicious regress.)

16 [Thanks to all! Note deleted.]

On an alleged completion of an infinite epistemic regress
Ronald Meester and Ren´e van Woudenberg

Abstract

This paper discusses the claim, issued by Peijnenburg and Atkinson, that some infinite epistemic
regresses can be completed. We argue (1) that Peijnenburg and Atkinson do not really address
the traditional epistemic Regress Problem, and (2) that the argument they offer does not qualify
as the completion of an infinite epistemic regress.

1 The traditional epistemic regress problem

This paper is a discussion of the claim, due to Peijnenburg and Atkinson, that, contrary to what is
traditionally assumed, there exist infinite epistemic regresses that can be completed. In addition
to being intrinsically interesting, this claim is also extrinsically important, for if it is correct, one
traditional argument for foundationalism, the Regress Argument, is undercut as this argument
involves the idea that infinite epistemic regresses cannot be completed. In this paper we present,
first, the traditional Regress Problem, or rather, various versions of it. Next, we present
Peijnenburg’s and Atkinson’s version of it, a version in which the notion ‘justified belief’ is
given pride of place. In the next section, we discuss the principle of justification that Peijnenburg
and Atkinson seem to adopt and argue that, with that principle, the epistemic Regress Problem
cannot even be addressed. Next we argue that the formal computations that Peijnenburg and
Atkinson perform do not qualify as “the completion of an infinite regress”. Finally, Peijnenburg
and Atkinson explicate what they mean by the ‘make probable relation’. We point to a problem
in their account, but this is not crucial as that relation doesn’t figure in their computations.

What is widely referred to as the epistemic regress problem has received rather different
formulations. In this section we consider various of such formulations in the interest of
explicating criteria that a regress must meet in order to qualify as an epistemic regress. That will
provide us with a firm basis for discussing the claim of Peijnenburg and Atkinson that there exist
infinite epistemic regresses that can actually be completed.

One locus classicus of the epistemic regress problem is Sextus Empiricus’ Outlines of
Phyrronism, where five “Modes” are distinguished, i.e. five lines of reasoning, that sceptics have
availed themselves of as a safeguard against dogmatism. The second of these modes has to do
with an infinite regress. Says Sextus:

The Mode based upon regress ad infinitum is that whereby we assert that the thing
adduced as a proof of the matter proposed needs a further proof, and this again another,
and so on ad infinitum, so that the consequence is suspension [of assent], as we posses no
starting-point for our argument.1

Now of what is the regress that Sextus refers to a regress? As the quote indicates, it is a regress
of proofs, or, in the Annas & Barnes translation, a regress of warrant. Beyond this point,
however, Sextus doesn’t provide any further details. He doesn’t tell what a proof is (or what
‘warrant’ comes to), nor what items are capable of being proved (or warranted). Others who have
filled in the details, have done so in various and different ways - and some have substituted
‘proof’ and ‘warrant’ for other notions, or, alternatively, have tried to explain them in other
terms.

Peter Klein for instance, commenting on Sextus, puts it like this: “The Modes were recipes for
avoiding dogmatism, i.e. the disposition to assent to non-evident propositions when it is not
settled whether they are true. One could locate such a non-evident proposition either by noting
that there was credible disagreement about it or by merely recognizing that there could be
credible disagreement. For in order to avoid epistemic hubris, the recognition that our epistemic
peers could sincerely disagree with us about the truth of some proposition forces us to regard it
as requiring reasons in order to rise to the desired level of credibility.”2 Klein, then, doesn’t talk
about ‘proofs’ but about ‘reasons’. He furthermore says that for Sextus reasons are required if
(and then he introduces another helpful notion not present in Sextus’ text) a proposition is to
have a desired level of credibility; he restricts this requirement to non-evident propositions - a
restriction not found in Sextus.

On Klein’s approach and in his terminology, then, an epistemic regress is a regress of reasons for
the credibility of a non-evident proposition. Accordingly, an infinite epistemic regress is an
infinite regress of reasons, so a series that contains a reason for a non-evident proposition, a
reason for that reason, a reason for that reason for that reason, a reason for that reason for that
reason for that reason, etc.

Michael Williams spells out Sextus’ Infinity Mode in a rather different way. “Suppose”, he says,
“that I make a claim - any claim. You are entitled to ask me whether what I have said is
something that I am just assuming to be true or whether I know it to be the case. If I reply that it
is something I know, you are further entitled to ask me how I know. In response, I will have to
cite something in support of my claim: my evidence, my credentials, whatever. But now the
question can be renewed: is what I cite in defence of my original claim something that I am just
assuming or something that I know? If the former, it will not do the job required of it: you can’t
base you knowledge on a mere assumption. But if the latter, it will in turn need to be backed up,
and so on.”3 We note that, other than Klein, Williams doesn’t talk about ‘propositions’ but about
‘claims’; not about ‘reasons’ but about ‘evidence’; not about ‘credibility’ but about ‘knowledge’.
He furthermore doesn’t restrict the requirement for evidential back up to what is non-evident.
Says Williams: “[W]hy [should] something striking us as self-evident [...] be a guarantee of its
correctness? If we ignore this question, the sceptic will say, we are just making a dogmatic
assumption. But if we try to find some further validation of intuitive self-evidence, we threaten
to open up the regress all over again.”4

On Williams’ approach, and in his terminology, then, an epistemic regress is a regress of claims
that are, or can be, backed up by evidence that the subject can cite, and the evidence cited
consists of what the subject knows to be true. And an infinite epistemic regress is an infinite
chain of such claims, each of which is backed up by evidence one can cite, while each piece of
evidence is something that is known by the subject to be true.

Richard Feldman explicates Sextus’ regress argument invoking yet other terminology. “The
argument”, he says, “begins with the observation that what makes a belief justified, at least in the
typical case, are other beliefs or reasons. [...] [I]f you think about this for a moment, you will
notice that a problem arises. If one belief is based on some reasons, but those reasons do not have
a basis themselves, then it looks as if what depends on those reasons is no better justified than a
belief for which one has no reason at all. [...] In short, you need reasons for your reasons, and
you need reasons for those reasons, and so on. But it does not seem as if any of us could ever
have this endless supply of reasons.”5 Two new terms appear in this explication, ‘belief’ and
‘justification’. Whereas Klein speaks about ‘the level of credibility of a non-evident proposition’,
Feldman speaks about ‘the justification of belief’. But both hold that what they have their eyes
on (‘the level of credibility’ and ‘the justification of belief’) is determined by reasons - and
Feldman seems to identify reasons with beliefs.

For Feldman, then, an epistemic regress is a regress of justified beliefs, and since justification
depends on the subject having reasons for her belief, we might also say that an epistemic regress
is a regress of reasons for a target belief. An infinite epistemic regress, accordingly, is a regress
of beliefs, each of which is justified by reasons that the subject has.

Richard Fumerton, not commenting on Sextus, has argued for a principle that is much in line
with Feldman’s explication of Sextus, viz. the Principle of Inferential Justification (PIJ):
 (PIJ) To have justification for believing P on the basis of E one must not only have (1)
justification for believing E, but (2) justification for believing that E makes probable P.6

Fumerton brings out the motivation for (PIJ) by means of the following case (that we have
slightly edited). Suppose you claim to be justified in believing that Fred will die shortly and that
you offer as your justification that a certain line across his palm (the infamous ‘lifeline’) is short.
Then, condition (1) of (PIJ) is satisfied, for you have justification for believing that Fred’s so-
called ‘lifeline’ is short - you have observed it yourself. Still, you are not justified in believing
that Fred will die shortly on the basis of that evidence, because condition (2) is not satisfied. For,
you have no justification for believing that Fred’s short ‘lifeline’ makes it probable that he will
die shortly. If one accepts (PIJ), Fumerton argues, an infinite regress will ensue: “If all
justification were inferential then for someone S to have justification for believing some
proposition P, S must be in a position to legitimately infer it from another proposition E1. But
E1 could justify S in believing P only if S were justified in believing E1, and if all justification
were inferential, the only way for S to be justified in believing E1 would be to infer it from some
other proposition E2 justifiably believed, a proposition which in turn would have to be inferred
from some other proposition E3, which is justifiably believed, and so on ad infinitum. But finite
beings cannot complete an infinitely long chain of reasoning, and so, if all justification were
inferential, no one would be justified in believing anything at all to any extent whatsoever.”7 In
this rendering of the regress argument, no use is made of such notions as ‘proof’, ‘warrant’,
‘claim’, ‘evidence’, ‘reason’, and ‘knowledge’. The only notions involved are ‘belief’,
‘justification’, ‘proposition’, and ‘making probable’. The latter notion is one we haven’t seen yet.
Still it might have been present just below the surface of other notions, esp. ‘proof’, ‘evidence’
and ‘reason’. After all, all these notions have been introduced without proper analysis, and the
analysis of each of them could easily be given in terms of ‘making probable’. So, ‘x proves y’, ‘x
is evidence for y’ and ‘x is a reason for y’ could all be analyzed as ‘x makes y probable’.

On Fumerton’s approach, then, an epistemic regress is a regress of justified beliefs - and a belief
is justified provided it satisfies (PIJ). This regress, as Fumerton has indicated, is infinite.

These, then, are strikingly different ways of rendering what is supposedly the same thing: the
epistemic regress argument. Nevertheless, there are a few constants in the arguments that we
must note:

[A] The arguments concern a subject that is trying to decide whether a certain proposition
is credible; or alternatively, they concern a subject that is trying to decide whether a
belief of hers is justified, whether she has (good, or sufficient) reasons for it.

[B] Whatever it is that makes a proposition credible, whatever it is that makes a belief
justified, the subject must be aware of it, she must be capable of citing it, she must be
capable to access8 it. Or: whatever it is that makes a proposition credible, if a subject
holds it is credible, she must be able to show that it is; if a subject’s belief is justified, she
must be able to show that it is.

As William Alston has pointed out, the notion of ‘justification’ is ambiguous.9 The justification
of a belief might be a certain process or activity - the process or activity of showing that the
attitude of believing has some positive status. The justification of a belief might on the other
hand also be a certain state one’s belief is in - the state namely of being justified. This is a real
and important distinction, obviously. For your belief that you are, for example, a native Dutch
person may be entirely justified, but this is very different from your being able to show that your
belief has that status; the latter might be a rather difficult thing to do. Standard modes of
expression in the epistemological literature such as that ‘S’s belief is justified’ and ‘S has a
justified belief’, are hence typically difficult to interpret. It should be clear however that the
notion of justification as used in [A] and [B] is the activity concept - the showing concept of
‘justification’.

There is a further constant, a principle that hasn’t explicitly surfaced so far, but that is implicit in
all of the above versions of the epistemic regress argument:

[C] In order to show that a certain proposition is credible it is insufficient to show that the
proposition is entailed by another; in order to show that belief in a certain proposition is
justified, it is insufficient to show that it is entailed by another proposition.

The reason for this is simple enough: if you see that ‘I have 10 million Euro’ entails ‘I have at
least 5 million Euro’ this doesn’t render the latter proposition credible, nor does it justify belief
in it.10 Entailment of a proposition by another is perhaps necessary but insufficient for credibility,
nor does it render the belief in it justified.

One point that has been made over and over again by those who have reflected on epistemic
regresses, is that if an epistemic regress is infinite, no proposition will be rendered credible by it,
no belief rendered justified by it.

It is this conditional claim that Peijnenburg and Atkinson are attacking. They don’t commit
themselves to the claim that all epistemic regresses are infinite; for all they say, some such
regresses may be finite. Nor is their claim that all epistemic regresses that are infinite can render
propositions credible, or beliefs justified. Theirs is the still weaker claim that some epistemic
regresses that are infinite, are nonetheless capable of rendering certain propositions credible, or
certain beliefs justified. Let us now turn to the argument advanced for it.

2 Peijnenburg’s and Atkinson’s version of the epistemic regress problem

As indicated, Peijnenburg and Atkinson have argued for the claim that “beliefs may be justified
by an infinite chain of reasons that can actually be completed” (Peijnenburg 2007: 598; also
Atkinson & Peijnenburg 2009) and they explicitly frame their discussion by reference to the
Regress Problem. They intend to provide a non-sceptical solution to it that differs from the
foundationalist, coherentist and infinitist solutions.

In order to be able to discuss their solution we first provide, and then discuss, some relevant
quotations in order to find out how they think about the notions involved in formulating their
version of the epistemic regress problem, viz. belief, propositions, justification and probability:
 (i) If a person S is epistemically (rather than prudentially) justified in believing
proposition E0, and if this justification is inferential (rather than noninferential or
‘immediate’), then typically S believes a proposition E1 which makes E0 probable.
(Peijnenburg 2007: 597)

(ii) Consider the following process of epistemic justification: proposition E0 is made

probable by E1, which in turn is made probable by E2, which is made probable by E3,
and so on. (Peijnenburg 2007: 597)

(iii) How to justify E1 epistemically? [...], if the justification is inferential, then there is a
proposition E2 that makes E1 probable. Imagine that E2 is in turn made probable by E3,
and that E3 is made probable by E4, and so on, ad infinitum. Is such a process possible?
Does the ‘ad infinitum’ make sense? (id.)

(iv) Against sceptics, foundationalists, and coherentists I will show that an infinite regress
can make sense; against infinitists I will show that beliefs may be justified by an infinite
chain of reasons that can actually be completed. (Peijnenburg 2007: 598)

(v) Can we continue this repetition, thus allowing for propositions made probable by
other propositions, made probable by still other propositions, and so on, ad infinitum?
(id.)

(vi) (A belief in) a target proposition E0 is justified by (a belief in) proposition E1, which
in turn is justified by E2, and so on. (Atkinson & Peijnenburg 2009: 1)

(vii) How close should the connection between En and En+1 be in order to say
legitimately that the one is justified by the other? [...] In this paper [...] we regard the
justification relation as a probability relation. (id.: 2)

A first thing to note is that the ambiguity of ‘justification’ is fully present in the quotes. In quote
(i) the concept seems to refer to a state, whereas in (ii) and (iii) it refers to an activity. We
construe the authors as holding that the activity or showing concept of ‘justification’ is leading.

A next thing to note is that whereas some quotations, notably (i) and (iv), identify beliefs as the
objects of justification, others, notably (ii), (iii), (vii), seem to identify propositions as the objects
of justification, while (vi) leaves open both possibilities. Of course, there is a connection
between beliefs and propositions: beliefs are attitudes towards propositions. Still, as we have
highlighted in the previous section under [A], one constant in the various formulations of the
traditional regress problem is that they concern a subject that is trying to decide whether a belief
of hers is justified (or whether a certain proposition is credible). Hence, in order to secure a
maximal interface with the traditional regress problem, we construe the authors as taking beliefs
to be the objects of justification, and interpret quotes (ii), (iii) and (vi) as abbreviated statements
of that idea. It is charitable to construe Peijnenburg and Atkinson in this way. For if justification
were to be construed as involving propositions only and not beliefs, this would have a most
unpalatable consequence. To see this we refer back to what we said in the previous section about
what motivated [C], and where we explained why it is insufficient for justification to show that
one proposition is entailed by another.

What do the authors say about how a subject should go about when she wants to decide to what
extent a belief of hers is justified? What does the process of epistemic justification (which is in
effect the process of showing that one of one’s beliefs has a positive standing) look like? Quotes
(v), (vi) and (vii) indicate the following: suppose subject S wants to decide to what extent her
belief in E0 is justified. Then what S must do is this: find another proposition E1 such that (a) S
believes E1, and (b) E1 makes E0 probable. Then find another proposition E2 such that (a) S
believes E2, and (b) E2 makes E1 probable, etc. By ‘makes probable’, our authors mean the
following: E1 makes E0 probable if P(E0|E1) > P(E0|¬E1), that is, if E0 is more likely given E1
than given the complement of E1. This notion will turn out to be somewhat problematic itself;
we will come back to this point in Section 4 and 5.

Given these explications, then, an infinite epistemic regress is an infinite series of beliefs, each
member of which has as its object a proposition that meets the two conditions specified in the
previous paragraph, so it is a proposition (a) that the subject believes, and (b) that is made
probable by yet another proposition.

The claim to be discussed is stated in (iv), viz. that there are regresses that satisfy these
conditions so there are infinite epistemic regresses that can actually be completed. ‘Completion’,
according to Atkinson and Peijnenburg, means that the unconditional probability P(E0) is well
defined, given a certain explicit infinite sequence of conditional probabilities P(En|En−1) and
P(En|¬En−1). By well defined is meant that any simultaneous assignment of probabilities to the
collection E0, E1, . . . which satisfies these conditional probabilites, necessarily have the same
probability P(E0).11

So we see that what Peijnenburg and Atkinson call the completion of an infinite epistemic
regress boils down to the computation of a probability, in line with quote (vii) which explicitly
identifies justification and probability.

We will argue that the claim of Peijnenburg and Atkinson is unconvincing. The next section
points out various problems with the notion of justification that is used. After that we argue that
even if these problems were non existent, the computations that the authors provide do not
qualify as the completion of an infinite epistemic regress.

3 Problems with justification

On the basis of our discussion in the previous section we can say that Peijnenburg and Atkinson
adopt the following principle of justification (PJ):

(PJ) If S believes that E0 and there is a proposition E1 such that (a) S believes it, and (b)
it makes E0 probable, then S’s belief is justified.
This principle, however, is not without problems. It doesn’t require that S knows or believes that
E1 makes E0 probable; it only requires that in fact E1 makes E0 probable. But this has unlovely
consequences, as the following example bears out.

Let E0 be the proposition that Dora will develop breast cancer during her life. Now there are
known genotypes which dramatically increase the probability of developing breast cancer.
However, it is strongly suspected that a number of yet unknown genotypes also increase the
probability of developing breast cancer. Suppose now that Dora has one of those yet unknown
genotypes - a genotype, let’s say that includes a certain allele A, which makes it highly probable
that she develops breast cancer. Suppose furthermore that E1 is the proposition that Dora’s
genotype includes allele A. Given these suppositions, we can easily see that E1 makes E0
probable. But now suppose also that S believes both E0 and E1, so believes that Dora will
develop breast cancer during her life, and also believes that Dora’s genotype includes allele A.
Then according to (PJ) S’s belief that E0 is justified. But that is unlovely. For how can S’s belief
be justified when S doesn’t even know, or believe, or have any inkling of the fact that E0 is made
probable by E1?

The problem with (PJ) is, in effect, that it doesn’t incorporate the idea that we identified under
[B] in Section 1 as a constant underlying the traditional regress argument, the idea namely that
whatever it is that renders a belief justified, it must be something the subject is aware of,
something she has cognitive contact with, something she can cite.

When we inject this idea into (PJ) we get:

(PJ*) If S believes that E0 and there is a proposition E1 such that (a) S believes it, (b) it
makes E0 probable, and (c) S believes that it makes E0 probable, then S’s belief that E0
is justified.

If this is to work, we must presumably require that the belief referred to in condition (c) must
itself be justified - for if S’s belief that E1 makes E0 probable is itself unjustified, S’s belief that
E0 will also remain unjustified. If we inject this requirement into (PJ*) we get:

(PJ**) If S believes that E0 and there is a proposition E1 such that (a) S believes it, (b) it
makes E0 probable, and (c) S justifiably believes that it makes E0 probable, then S’s
belief that E0 is justified.

With this principle it is as with Fumertons (PIJ): its application leads to an infinite regress. That
need not be a problem for Peijnenburg and Atkinson as they claim that some infinite epistemic
regresses can actually be completed.

The gain of the discussion so far is that the principle of justification (PJ) as it is used by our
authors has been shown to be unconvincing as it stands. And we have suggested a way to bold it
against an objection that we think Peijnenburg and Atkinson should take seriously - a suggestion
furthermore that is very much in the spirit of the traditional regress problem, and also in the spirit
of our authors.
Our suggestion, as incorporated in (PJ**), however, means that an ensuing infinite epistemic
regress (even the so-called simple linear ones) will have a more complex form than the regresses
that are referred to in quotes (ii), (iii), (vi), and (vii) in the previous section. An infinite epistemic
regress, contrary to what the quotations suggest, doesn’t have form I:

I: E0 is made probable by E1; and E1 is made probable by E2; and E2 is made probable
by E3; etc.

but the more complex form II12:

II: (a) S believes E0, E1, E2, E3, . . .

(b) E0 is made probable by E1, E1 is made probable by E2, E2 is made probable by E3,
etc.

(c) S is justified in believing that E0 is made probable by E1; that E1 is made probable by
E2; that E2 is made probable by E3; etc.

As we can see from the quotations, the regress that Peijnenburg and Atkinson claim can be
completed, is of form I, not of form II. This poses the problem that their claim in fact doesn’t
touch on what is traditionally conceived to be the regress problem. For, to repeat, an infinite
regress requires an infinite number of beliefs, while beliefs are absent in form I! It is at this point
that the problem of viewing the justification relation as a probability relation (quote (vii) above)
becomes apparent.

This conclusion, that the claim of Peijnenburg and Atkinson doesn’t touch the regress problem,
should already be enough to dismiss their conclusions. However, we have a few more things to
say about their setup and conclusions. In the next section we will argue that what our authors
offer, even if it isn’t the completion of an infinite epistemic regress, doesn’t even qualify as the
completion of a infinite regress full stop.

4 The main examples of Peijnenburg and Atkinson are not infinite regresses

We first recall the basic mathematical setup of Peijnenburg and Atkinson. They define two sets
of conditional probabilities, namely α as the probability of a proposition Em given that
proposition Em+1 is true, and ẞ as the probability of Em given that Em+1 is false. In formulas,
this reads

P(Em|Em+1) = α and P(Em|¬Em+1) = ẞ, (1)

for m = 0, 1, . . . For simplicity, they assume that neither of these two conditional probabilities
depend on m. Hence they deal with a situation in which the occurrence of E1 leads to a certain
conditional probability of E0, while E1 in its turn has a certain conditional probability given E2,
and so on. According to Peijnenburg and Atkinson, Em+1 makes Em probable if α > ẞ, but this
requirement (that, as we will argue in the next section, is in itself problematic) is not used in their
computations. After two pages of elementary algebra (including taking a limit), they conclude
(correctly) that this implies that P(E0) is well defined and equals

P(E0) = ẞ/1 − α + ẞ. (2)

This is the “completion” of the infinite regress that Peijnenburg claims makes her reject the
Regress Argument: despite the fact that there is an infinite chain of conditional probabilities, the
probability of E0 is still well defined, and hence an infinite chain of reasons can actually be
completed.

The computation of P(E0) above gives the correct answer (even though we have seen that it does
not mean that belief in E0 is justified) but we claim that it does not qualify as the completion of
an infinite regress. To see this, we use (1) to compute

P(Em) = P(Em|Em+1)P(Em+1) + P(Em|¬Em+1)P(¬Em+1)

= αP(Em+1 + ẞ(1 − P(Em+1))
= ẞ + (α − ẞ)P(Em+1). (3)

We claim that P(E0) follows from this equation immediately. Why? Well, note that the
assumptions (1) of Peijnenburg and Atkinson entails P(Em) to be the same for all m (this already
is a somewhat strange state of affairs of course) since instead of ‘starting’ at E0, we could also
start at E1, E2 or any Em for that matter. No matter where we start, the outcome must always be
the same since the probabilities α and ẞ do not depend on m. Hence in equation (3), we are
allowed to substitute x = P(Em) = P(Em+1), arriving at

x = ẞ + (α − ẞ)x,

which is immediately solved, without any infinite regress, to give

x = ẞ/(1 − α + ẞ),

and (2) follows without any infinite regress whatsoever. It is, therefore, simply not the case that
P(E0) is the result of a completion of an infinite regress.13

Perhaps the following analogy helps here: the fact that

2 = 1 + 1/2 + 1/4 + 1/8 + · · ·

does not imply that the number 2 should be seen, viewed or defined as the result of summing an
infinite series. Similarly, it is hardly reasonable to say that P(E0) is the result of completing an
infinite regress, if there is another, much better, way of computing it.

The computation of P(E0) without infinite regress above relies on the assumption that the
conditional probabilities P(Em|Em+1) and P(Em|¬Em+1) do not depend on m. At the end of her
paper, Peijnenburg (2007) remarks that this condition is in fact too strong. As an explicit
example she claims that it suffices to require that there is a constant c < 1 such that
 0 < αm − ẞm < c, (4)

for all m, where

P(Em|Em+1) = αm and P(Em|¬Em+1) = ẞm. (5)

Under these circumstances, a recurrence relation as in (3) cannot be written down. Nevertheless,
under condition (4), it is indeed the case that P(E0) is well defined.

Is this an example of a completed infinite epistemic regress? No. First of all, the aforementioned
problems with justification remain valid so the fact that P(E0) is well defined does not suffice for
any epistemic justification of E0 whatsoever, and neither does it mean that we can compute
P(E0).

Furthermore, condition (4) involves all possible values of m, of which there are - of course -
infinitely many. In fact, it is easy to see that any condition that would make P(E0) well defined
would either be essentially stationary, that is, with conditional probabilities independent of m, as
in (1), or necessarily include a condition for infinitely many values of m. So how can one check
whether or not (4), or any other sufficient condition, holds? This would involve a priori
knowledge of all propositions Em - note condition [B] stated above. But giving all propositions
and conditional probabilities in advance can hardly be called an infinite regress: one sets up the
full structure in one swap, and if we know all the relevant quantities in advance, then we simply
do not need a regress. In contrast, in a genuine infinite regress, one is interested in a proposition
E0. One then chooses another proposition E1 conditional on which E0 has a certain probability,
and one expresses the probability of E0 in terms of E1, etcetera. At no stage of this procedure,
unless one is allowed to look at the infinite future, does one know whether or not this leads to a
well defined probability of E0.

This is not to deny that regresses can play a positive role. We argued in the previous paragraph
that one can never know whether or not the infinite regress leads to a well defined P(E0), unless
we deal with a stationary situation, but we already concluded that a stationary situation is not an
infinite regress at all. However, we can use the conditional probabilities ẞm = P(Em|Em+1) and
αm = P(Em|¬Em+1) to give bounds for the probability of E0. Indeed, it is easy to compute (this
is equation (13) in Peijnenburg’s paper) that

P(E0) = ẞ0 + (α0 − ẞ0) ẞ1 + · · · +

+(α0 − ẞ0)( α1 − ẞ1) · · · (αm−1 − ẞm−1) ẞm +
+(α0 − ẞ0)( α1 − ẞ1) · · · (αm − ẞm)P(Em+1).

Since in the ‘future’ of the regression, P(Em+1) is always between 0 and 1, this means that P(E0)
will always be between the two values obtained from the right hand side by substituting
P(Em+1) = 0 and P(Em+1) = 1, even if P(E0) is not well defined in the infinite limit of the
regress.14

5 The ‘make probable’-relation

In this section we return to the notion of ‘making probable’. As Peijnenburg and Atkinson state
explicitly, this notion should be taken in the following way: proposition E0 is made probable by
E1, provided P(E0|E1) > P(E0|¬E1), or, in terms of the previous section, if α > ẞ. We already
mentioned that this requirement plays no role in the computation, but we would like to point out
nevertheless that this interpretation of ‘making probable’ has strange consequences.

To see this, suppose that P(E0|E1) = .0002 and P(E0|¬E1) = .0001. Then given the explication of
’making probable’ relation, we should say in this case that E0 is made probable by E1. And
given (PJ) we should furthermore say that if someone believes both E0 and E1, then his belief
that E0 is justified. But both of these things are, of course, unacceptable. It is unacceptable to say
that E0 is made probable by E1, when P(E0|E1) is as small as .0002. Why the second thing is
unacceptable can best be brought out by an example.

Suppose at the local fair a 10.000 ticket lottery is held in which buyers are allowed to purchase at
most two tickets. All tickets are sold. Suppose you know you are the only one who bought two
tickets and suppose furthermore you believe both E1 and E0 below:

E1: I bought two tickets;

E0: I will win the lottery.

Then given (JP) you are justified in believing that E1 (i.e. that you will win the lottery). For the
probability that you will win given that you bought two tickets is higher then the probability that
you will win given that you did not buy two tickets (that is, by buying one ticket). But this
cannot be right. Your belief that E0 isn’t justified even though you believe E1 and it is also the
case that P(E0|E1) > P(E0|¬E1). Hence, the ‘makes probable’- relation as interpreted by our
authors, renders (PJ) as well as its successors (PJ*) and (PJ**) unacceptable.

6 Conclusions

Peijnenburg and Atkinson argue that the Regress Argument for epistemic foundationalism is
wrongheaded because, contrary to what the argument assumes, an infinite regress is not, from an
epistemological point of view, lame or impotent because some infinite regresses can actually be
completed. We have argued that the justificatory principle they adopt suffers from serious
problems.

We have also argued that their set-up is not an epistemic regress (in the classical sense - the
sense in which it is meant in the Regress Argument) but a mere computation of probabilities, and
sidesteps the crucial issue of justification. Moreover, their main example is not even an example
of an infinite regress in the mathematical sense - the computation can be done with finitely many
operations, and the secondary example does not support their claim either.

So, we have argued that the Regress Argument for epistemic foundationalism stands unrefuted
by the arguments of Peijnenburg and Atkinson.
1 Quoted from Klein 131. A rather different translation of Sextus’ text reads: “In the mode
deriving from infinite regress, we say that what is brought forward as a warrant for the matter in
question needs another warrant, which itself needs another, and so ad infinitum, so that we have
no point from which to begin to establish anything, and suspension of judgement follows”.
(Annas & Barnes 1985: 182) Two differences stand out. Where the quotation from Klein has
‘proof’, the Annas & Barnes text has ‘warrant’, and where the quotation from Klein has ‘no
starting point for an argument’, the Annas & Barnes text has ‘no point from which to begin to
establish anything’.

2 Klein 2005: 132.

3 Williams 2001: 62.

4 Williams 2001: 63.

5 Feldman 2003: 49-50.

6 Fumerton 2002: 211.

7 Ibid.

8 See for this also Cling 2008: 402.

9 See for this Alston 1989: 43, 55, 70.

10 We borrow this example from Cling 2008: 402, who makes the same point as we in [C].

11 Having the same probability P(E0) does not imply that the assignment of simultaneous
probabilities is uniquely defined. For the current discussion, this is not important however.

12 At least when one adopts (PJ**). If one adopts (PJ) as Peijnenburg and Atkinson seem to do,
an infinite regress will have form II without (c).

13 There is much more to say about this computation and its consequences. Technically
speaking, the computation shows that any probability distribution on the (non)-occurrence of the
propositions involved which satisfies (1) has the same so called marginal probability P(Em).
This is not to say that the full probability distribution is now specified. Strictly speaking, this is
no problem for the issue under consideration here, but it shows that introducing probabilities in
this context is a very delicate issue.

14 In Atkinson and Peijnenburg 2009, a remark to - essentially - this effect is made by discussing
the consequences of substituting P(Em) = 0 and P(Em) = 1 for some m.
Pyrrhonism in Ancient, Modern, and Contemporary Philosophy
The New Synthese Historical Library
Texts and Studies in the History of Philosophy
Introduction
Dr. Diego E. Machuca

Preface

The bulk of the essays collected in this volume derive from papers presented at a conference on
Pyrrhonian skepticism held in Buenos Aires on August 6–8, 2008. I am grateful to all the
speakers who accepted the invitation to participate in the meeting. Other arrangements had been
made for the publication of several of the papers, so these are not included here. By contrast, four
additional essays by scholars who could not attend the conference have been incorporated. My
thanks to them for being willing to contribute to the volume. I wish as well to express my
gratitude to the Departamento de Humanidades y Artes of the Universidad Nacional de Lanús
and to the Agencia Nacional de Promoción Científica y Tecnológica of Argentina for the
financial support that made it possible to organize the conference. My thanks also go to Simo
Knuuttila for agreeing to publish the volume in The New Synthese Historical Library, and to
Willemijn Arts and Ingrid van Laarhoven from Springer for their help through the various stages
of the project. Finally, I would like to thank the referee for Springer for his helpful comments on
the manuscript.

Buenos Aires, Argentina Diego E. Machuca

April 2011

Introduction

Among scholars of ancient philosophy there is today considerable interest in the extant works of
the second-century physician Sextus Empiricus as our most important source for Pyrrhonian
skepticism. Until not long ago, though, they used to regard his writings exclusively as an
invaluable source of information about non-Pyrrhonian thinkers and schools whose positions
would otherwise be more obscure or even utterly unknown. A radical change in the way of
approaching Sextus’ oeuvre started to take place since the late 1970s and early 1980s, when a
group of scholars began to explore the complex history of ancient Pyrrhonism and the sui
generis character of this form of skepticism. Although some have questioned its coherence, it is
safe to say that most have recognized its philosophical import.

A fact that sometimes goes unnoticed even to this day regarding the Pyrrhonian tradition is that
its history did not end with Sextus and his immediate successors, since it has had a tremendous
impact on both modern and contemporary philosophy. As regards the modern period, a
considerable number of historians of ideas have argued that the Renaissance rediscovery of
Sextus’ works played a key role in the formation of modern thought.1 Richard Popkin in
particular maintained that the revival of Pyrrhonism triggered a “Pyrrhonian crisis.”2 According
to him, the history of modern philosophy should to a large extent be construed as the history of
the various strategies which modern thinkers devised to deal with that crisis.3 As for the
contemporary philosophical scene, an important number of epistemologists have vigorously
discussed the Pyrrhonian arguments against the rational justification of our beliefs – what they
now call “Agrippa’s trilemma.”4 Coherentists, foundationalists, and infinitists, as well as
externalists and contextualists have adopted different tactics to deal with that challenge.
Moreover, some present-day thinkers have characterized their own philosophical positions as
“(neo-)Pyrrhonian.”5 Thus, in parallel with the strong interest in Pyrrhonism aroused among
ancient philosophy scholars, Sextus’ writings have in recent years increasingly attracted the
attention of historians of modern philosophy and analytic philosophers.6

Collections of essays dealing with skepticism can in general be divided into two groups, each of
which has been the object of a criticism. Some collections are said to adopt an approach which is
exclusively historical and exegetical and to overlook the current discussion of skepticism in
systematic analytic philosophy, thus ignoring much of its philosophical import. By contrast,
others are said to take an approach which is solely systematic and to explore certain skeptical
arguments in abstracto, thus neglecting their origin and history, without which it is impossible to
fully appreciate them. The present volume intends to avoid these criticisms by integrating the
strengths and merits of both types of collections: it explores the history and significance of
Pyrrhonism in ancient and modern philosophy and examines the Pyrrhonian outlook in relation
to contemporary analytic philosophy. Moreover, going beyond the common distinction between
a historical and a systematic approach, some of the essays combine specialized historical
scholarship about the Pyrrhonian tradition with rigorous philosophical examination of the nature
of Pyrrhonism and the challenges it poses. The reason for combining historical exegesis with
systematic investigation is confidence in the viability and the desirability of exploring a
philosophical stance both in its historical context and in connection with contemporary concerns.

By analyzing various aspects of Pyrrhonian skepticism as it was conceived of in its original

Greek context and later on in the modern period, and as it is interpreted in the contemporary
philosophical scene, the essays collected in this volume will allow the reader to witness the
transformations undergone by the Pyrrhonian tradition. This tradition is complex and
multifaceted, since the Pyrrhonian arguments have been put into the service of different
enterprises or been approached in relation to interests which are quite distinct. It is thus
impossible to find an entirely homogenous or monolithic picture of the Pyrrhonian outlook from
the Hellenistic period to the present day. The diversity of uses and conceptions of Pyrrhonism
accounts for the diversity of the challenges it is deemed to pose and of the attempts to meet them.
This philosophical richness and adaptability should be borne in mind by anyone studying the
Pyrrhonian tradition.

Before introducing the twelve contributions that make up the present volume, it is perhaps
necessary to account for the absence of essays devoted to the presence of Pyrrhonism in Latin
medieval thought. The explanation is simple: although discussion of skeptical problems and
arguments is well attested in Western Europe during the Middle Ages,7 Pyrrhonism played a very
small part, since direct acquaintance with Pyrrhonian texts was rare, with the result that
knowledge of this brand of skepticism was extremely limited.8 First, three manuscripts of an
early-fourteenth century Latin translation of Sextus’ Outlines of Pyrrhonism probably by Niccolò
da Reggio (ca. 1280–1350) survive, but this translation seems to have exerted no influence
whatsoever.9 One of these manuscripts also contains a partial Latin translation of Sextus’
Adversus Mathematicos III–V, probably by the same author.10 Second, it has recently been
argued that, in Nicholas of Autrecourt’s treatment of the infinite regress argument, it is possible
to identify certain Pyrrhonian elements,11 which would show that in the fourteenth century there
was at least some knowledge of Sextus. Finally, a manuscript remains which contains a fifteenth-
century Latin translation of the first four books of Adversus Mathematicos by Giovanni Lorenzi,
but it does not seem to have had any circulation.12

In the Middle Ages, so-called Academic skepticism was better known than Pyrrhonism due
especially to Augustine’s Contra Academicos, and also to other works such as Cicero’s
Academica, De Natura Deorum, and Tusculanae.13 But there were also important skeptical
arguments peculiar to medieval philosophy, such as that which refers to an all-powerful God that
could deceive us.14 However, contrary to what happened before and after the Middle Ages,
during this period there were no skeptics in either the Christian, Jewish, or Islamic traditions,
with the only possible exception of John of Salisbury (ca. 1120–1180), who considered himself
an Academic skeptic.15 In any case, it is plain that there were no full-blown or radical skeptics
advocating universal suspension of judgment or denying the possibility of knowledge. In general,
although some thinkers adopted what might be deemed a skeptical position on certain specific
issues, when skepticism was the object of discussion, in most cases the aim was to refute it.16 In
addition, such refutations were not produced in the context of a serious skeptical crisis, unlike
what later occurred in the modern period. It must be noted, in this connection, that it has recently
been claimed that an important number of medieval authors made a methodological use of
skeptical arguments. This use consisted in employing them either to undermine a given
conception of knowledge which was then replaced by another one immune to the challenges
posed by those arguments, or to distinguish between types of knowledge and to determine the
kind of certainty proper to each of them.17 In the Middle Ages, skepticism seems to have been a
logical construction independent for the most part of the historical skeptical movement, which is
why it was understood as a reservoir of arguments to the effect that knowledge is impossible.18 In
this respect, the medieval conception of skepticism does not seem to differ much from the way
skepticism is generally conceived of in contemporary analytic epistemology.

In sum, the role played by Pyrrhonism in Western medieval thought was both minor in relation
to other forms of skepticism and insignificant in comparison with the part it has had in both
modern and contemporary philosophical discussions.

The present collection falls into three parts, the first focusing on ancient Pyrrhonism, the second
addressing its influence on modern philosophy, and the third dealing with the Pyrrhonian stance
in relation to contemporary analytic philosophy. While the approach taken in the third part is, as
expected, mainly systematic, some of the essays on ancient and modern Pyrrhonism combine
historical and exegetical analysis with an assessment of the philosophical merits of the
Pyrrhonian outlook. It should be noted that this collection does not aim to provide a
comprehensive discussion of Pyrrhonism in ancient, modern, and contemporary philosophy.
Rather, its goal is to open up stimulating new exegetical and philosophical perspectives on
Pyrrhonian skepticism and to motivate further examination of certain issues.

Richard Bett opens the first part by examining a facet of the practical nature of ancient
Pyrrhonism. As has frequently been noted, one of the aspects which differentiate it from the
various forms of contemporary skepticism is that the Pyrrhonist does not view his philosophy as
a merely theoretical stance with no implications for his life, but rather regards it as an α ’γωγ´η
or way of living. This practical character does not consist solely in the assumption that it is
possible to live one’s Pyrrhonism, but also in the thought that the Pyrrhonist, by virtue of his
skepticism, is better off than other people. Bett’s contribution focuses on an aspect of this further
claim. Granting for the sake of argument that Pyrrhonism is livable, he asks whether the type of
values the Pyrrhonian life includes and the adherence to them it allows make such a life ethically
acceptable and desirable for us. He thinks they do not, since the Pyrrhonist’s lack of commitment
to the moral values that form part of his society’s laws and customs and the passivity shown in
his practical decisions reveal that he is not an ethically engaged agent.

Next, Svavar Hrafn Svavarsson offers an original interpretation of the notion of α ’ταραξ´ια
(undisturbedness or tranquility) in Sextus’ account of Pyrrhonism. In the Outlines of Pyrrhonism,
Sextus tells us that the reason the proto-Pyrrhonist began to philosophize was his hope of
becoming undisturbed by resolving the disagreement among appearances, i.e., by determining
which are true and which are false. Being unable to resolve it, due to the equipollence of the
conflicting appearances, he suspended judgment. To his surprise, by doing so he became
undisturbed in matters of belief. Svavarsson argues that, although Sextus often suggests that the
skeptic’s undisturbedness comes about through his not having beliefs of any kind, in his explicit
explanation of skeptical undisturbedness he claims that it comes about only through not having
positive beliefs about natural values. Hence there are two notions of undisturbedness at play in
Sextus’ account of Pyrrhonism.

Katja Vogt’s essay examines a subject which has recently been a focus of much attention among
scholars, namely, the nature of the investigation conducted by the Pyrrhonist. One common
objection raised against his investigation is that it cannot be deemed genuine investigation, since
it does not aim at the discovery of the truth. Rather, what the Pyrrhonist searches for is the
attainment of the state of undisturbedness. Vogt argues that this objection is based on the
questionable view that the sole goal of investigation is the discovery of the truth. This overlooks
that there are other aims which philosophers strive for in their investigations, fails to distinguish
between the motivational basis of philosophical investigation and its goals, and ignores that
philosophical inquiry guided by the value of truth may not be immediately aimed at the
attainment of the truth but rather at the avoidance of falsehoods. These three points are to be
taken into account when trying to understand the Pyrrhonian sort of investigation.

For its part, my contribution explores whether the Sextan Pyrrhonist is committed to the law or
principle of non-contradiction, a topic which has not received much attention among students of
ancient Pyrrhonism. Although there are passages of Sextus’ oeuvre which seem to show that the
Pyrrhonist endorses that law, I argue that he actually suspends judgment about the truth of its
different dogmatic formulations. However, this does not preclude him from following, without
any conviction as to their correctness, certain qualified versions of that law when thinking and
acting, even though he would not present these as versions of a law or principle. I also claim that
the Pyrrhonist makes use of the dogmatic versions of the principle of non-contradiction only
dialectically. Finally, after showing that both the uncommitted observance of qualified versions
of that principle and the dialectical use of its dogmatic formulations are in agreement with other
aspects of the Pyrrhonist’s philosophy, I explore whether he is an anti-rationalist.

Peter Klein focuses on two of the so-called Five Modes of Agrippa, namely, the mode deriving
from regress ad infinitum and the mode based on reciprocity. He maintains that, in keeping with
his characteristic dialectical argumentation, the Pyrrhonist takes the premises employed in those
modes from Aristotle’s foundationalist conception of justification. This epistemological theory
was the predominant one in Sextus’ time and continues to play an important role in
contemporary foundationalism. However, the fact that the Pyrrhonist’s regress argument as well
as his reciprocal mode work with a particular conception of epistemic justification significantly
restricts both their generality and their power. For there are alternative theories – namely,
contemporary infinitism and coherentism – which conceive of epistemic warrant in such a way
that they reject some of the assumptions of the foundationalist conception of justification, and
which therefore accept that regress and reciprocal arguments can produce conclusions which are
epistemically justified. After some vacillation, Klein’s paper has been placed in the first part.
The reason is that, even though his purpose is to critically examine the epistemological
presuppositions of Agrippa’s trilemma using the tools of contemporary analytic philosophy, he
deals at great length with Sextus’ account of the Agrippan modes and Aristotle’s foundationalist
epistemology. His essay is a paradigmatic case of ancient texts being approached from the
perspective, not of a classical philosophy scholar, but of a contemporary epistemologist
concerned with assessing various theories of justification.

The second part of the volume, devoted to Pyrrhonism in modern philosophy, includes four
papers dealing with the outlooks that Francis Bacon, Pierre Bayle, and David Hume adopted
towards Pyrrhonism and skepticism in general. These essays take account of the specialized
literature of the past decades and offer new interpretations, thus advancing the study of the
presence of Pyrrhonian skepticism in modern thought.

Bacon’s relation to skepticism has not attracted much attention from specialists, and hence the
presence of Pyrrhonian elements in his philosophy has not been thoroughly studied. Accordingly,
Luiz Eva’s essay represents an important addition to the literature and will arouse further interest
in exploring the role played by skepticism in the thought of the author of the Novum Organum.
Bacon recognizes some important similarities between skepticism and his own philosophy,
especially the skeptic’s emphasis on the weakness of our cognitive faculties and his suspension
of judgment. But there is also a key difference: although both affirm the impossibility of
knowledge, the skeptic claims that nothing can be known tout court, whereas Bacon contends
that this is the case only as far as the traditional way of attaining knowledge is concerned. Eva
proposes to examine the affinities and differences between Bacon’s philosophy and ancient
skepticism, both Pyrrhonian and Academic, by focusing on a comparison between his doctrine of
the idols and the arguments and themes found in the ancient and contemporary skeptical sources
(in particular Montaigne’s Essais) which he presumably read.

It has sometimes been said that, for any given claim made by Pierre Bayle somewhere in his
oeuvre, it is possible to find a refutation of it somewhere else. Centering his analysis on Bayle’s
Commentaire philosophique, John Christian Laursen examines the presence in this work of both
arguments which purport to establish rationally the necessity of toleration and arguments which,
by contrast, undermine reason’s capacity to ground toleration. These latter are Bayle’s famous
arguments about the conscientious persecutor and the overwhelming force of custom and
ducation. Laursen analyzes in detail the tension between the two kinds of argument,
demonstrating that neither side wins and that Bayle finally turns to rhetoric to justify toleration.
He thus takes issue with recent interpretations of the Commentaire philosophique which,
ignoring or dismissing the arguments that undermine the reliability of reason in matters of
toleration, maintain that Bayle’s position in that work is wholly rationalist. Laursen’s paper is
therefore an important contribution to the more general debate about whether Bayle was a
rationalist or a Pyrrhonian skeptic.

Hume’s stance on Pyrrhonism, and on skepticism in general, is probably more complex than that
of any other modern thinker, with the possible exception of Bayle. Due to this complexity, two
essays are devoted to a careful examination of his skepticism, tackling both his overt disparaging
attitude towards (what he took to be) Pyrrhonian skepticism and the philosophical connection
between his own skepticism and the Pyrrhonist’s. Although Hume’s relation to skepticism has
received considerable attention from scholars, these two essays provide fresh insights. In the first
of them, Peter Fosl considers an at least apparent incoherence on the part of both the Pyrrhonist
and Hume, namely, their appeal to nature. For it seems that skeptics should refrain from claiming
to have understood nature, from making prescriptive or normative assertions based on the notion
of nature, and in general from espousing any form of naturalism. Fosl analyzes the Humean
conception of nature, arguing that it is characteristically skeptical and that, in Hume, skepticism
should be understood, not as a theory, but as a non-dogmatic way of addressing theory. This is
much in line with the ancient Pyrrhonist’s outlook, even if the similarity may be only accidental.
In this connection, Fosl thoroughly explores whether Sextus’ texts were available to Hume and,
if so, whether he read them and was influenced by them. He also considers in depth the
similarities and differences between Hume’s stance and the ancient forms of skepticism, and
examines to what extent one can characterize his skepticism as either Academic or Pyrrhonian.

For his part, Plínio Junqueira Smith explores in detail Hume’s treatment of skeptical arguments.
On the one hand, there is Hume’s distinction between two types of such arguments, namely,
popular or weak and philosophical or strong skeptical arguments. On the other, there is the
difference between two uses of these arguments, no matter whether they are weak or strong: the
Pyrrhonist utilizes them to suspend judgment, whereas the Academic employs them both to
restrict our inquiries to what is within the scope of our understanding and to show that we only
have probable knowledge and beliefs. Hume only accepts the latter use, because it is impossible
to abolish all belief as Pyrrhonists intend to do.

Present-day epistemological discussions of skepticism have mainly focused on so-called

Cartesian skepticism – i.e., the view that knowledge in general, or at least regarding a very large
area, is impossible. This view, which amounts to the one Sextus ascribes to the Academics
Carneades and Clitomachus, differs from Pyrrhonism in both its formulation and its scope. First,
the Pyrrhonist does not deny the possibility of ever attaining knowledge, but suspends judgment
about whether or not knowledge is impossible, restricting himself to saying that up till now he
has been unable to affirm that anyone knows anything. Second, he does not merely call into
question our knowledge-claims, but also casts doubt on whether we are in fact justified in
preferring any one of our beliefs to its opposite.19 This more radical and subtler form of
skepticism has more and more been attracting the attention of epistemologists, and the three
essays included in the third part of the volume are the result of careful study of the Pyrrhonian
outlook in connection with contemporary analytic philosophy. They deal with the relationship
between Wittgenstein’s thought and Pyrrhonism, the challenges posed by the Agrippan modes to
contemporary theories of knowledge and justification, and the question of whether Pyrrhonism is
livable. In the first essay, Duncan Pritchard explores Wittgenstein’s radical new conception of
the structure of reasons put forward in his posthumously published On Certainty. According to
such a conception, all belief-systems require the existence of fundamental “hinge” propositions
which are held to be most certain, but which can be neither rationally doubted nor rationally
supported. Pritchard claims that this account of the structure of reasons gives rise to a type of
restricted skepticism which bears significant similarities with the Pyrrhonian outlook. This is
why he thinks one can legitimately characterize that kind of skepticism as “Wittgensteinian
Pyrrhonism.”

In present-day epistemological discussions, the problem of disagreement has lately regained part
of the special relevance it has in Sextus’ account of Pyrrhonism. Thinking that the philosophical
import of this problem has not been fully appreciated yet, Markus Lammenranta proposes to
offer, on the basis of Sextus’ exposition of the modes of suspension of judgment and from the
perspective of analytic epistemology, a reconstruction of the argument based on disagreement.
He contends that this reconstruction provides us with a serious “skeptical paradox” which must
be taken into careful consideration by modern-day theories of knowledge and justification. He
reviews the epistemological theories of foundationalism, contextualism, coherentism, reliabilism
and evidentialism, and claims that they all fail to satisfactorily respond to the Pyrrhonian
argument from disagreement.

Probably the most vexed question regarding ancient Pyrrhonism concerns the scope of ε’πoχ´η,
namely, whether it is limited to philosophico-scientific beliefs or extends also to ordinary or
common-sense beliefs. The vigorous debate among specialists about whether the Pyrrhonist
disavows all, or only some, beliefs has been couched in different terms, e.g., whether his
skepticism is “rustic” or “urbane,” and whether or not all of his appearance-statements are “non-
epistemic,” “non-doxastic,” or “non-judgmental.”20 This issue is intimately related to the long-
standing charge that the Pyrrhonist is reduced to inactivity because action requires belief (the
famous α ’παξ´ια objection). If he advocates a moderate skepticism, then the charge seems easily
answerable, while if his skepticism is radical, he is required to explain how action is possible in
the absence of all beliefs. Juan Comesaña’s essay takes up this complex question, not from a
historical perspective, but from an exclusively philosophical one. His purpose is to determine
whether the urbane and the rustic Pyrrhonists’ replies to the inactivity charge are theoretically
acceptable. In his view, the rustic Pyrrhonist cannot successfully respond to the objection
because it is plain that action does require belief. The urbane Pyrrhonist therefore seems in a
better position. So Comesaña examines both whether this type of Pyrrhonist can respond to the
charge by using traditional resources and whether the contemporary theories of contextualism
and contrastivism can help him offer a philosophically satisfactory or plausible answer. His final
verdict is negative.

We hope this volume will contribute not only to further showing that Pyrrhonism has played a
crucial role in the history of philosophy since Antiquity, but also to highlighting the
philosophical import of this brand of skepticism.

1 The resurgence of Pyrrhonian skepticism was due especially to the publication of Henri
Estienne’s Latin translation of Sextus’ Outlines of Pyrrhonism in 1562 and of Gentian Hervet’s
Latin translation of Adversus Mathematicos and Adversus Dogmaticos in 1569.

2 Popkin (1960) is the pioneering work on the impact of ancient skepticism (especially in its
Pyrrhonian variety) on modern philosophy, covering the period from Erasmus to Descartes.
Popkin (1979) extends the analysis to Spinoza, and Popkin (2003) reaches back to Savonarola
and forward to Bayle. For the influence of Pyrrhonism, and ancient skepticism in general, on
modern thought, see also Popkin (1993), Maia Neto and Popkin (2004), Paganini (2008), Maia
Neto, Paganini, and Laursen (2009), Paganini and Maia Neto (2009), and Naya (2011).

3 Some scholars have argued that Popkin overstates the part played by Pyrrhonism in shaping
modern thought. See Ayers (2004) and Perler (2004).

4 See, e.g., Fogelin (1994), Sosa (1997), Williams (2004), Klein (2008), and Lammenranta
(2008). Agrippa’s trilemma is also known as “Münchhausen-Trilemma,” as Hans Albert (1985)
has called it.
5 See Fogelin (1994, 2004), Porchat Pereira (2006), and Sinnott-Armstrong (2006).

6 For an overview of Sextus’ legacy in modern and contemporary philosophy, see Machuca
(2008, 58–63).

7 It has recently been claimed that historians of skepticism have usually ignored the important
part played by both epistemological and external-world skepticism during the Middle Ages, and
that this has prevented them from realizing that medieval discussions of skepticism must be
taken into account in order to fully understand the history of modern skepticism. In this
connection, see the essays collected in Lagerlund (2010). For an overview of medieval
skepticism from the thirteenth century on, see Perler (2006).

8 See Schmitt (1983, 226–7), Porro (1994, 229–37), Floridi (2002, 13–25). Greek/Byzantine and
Arabic scholars, by contrast, continued to read Pyrrhonian texts during Late Antiquity and the
Middle Ages (see Porro 1994, 235 n. 16; Floridi 2002, 20–2, 24–5). For instance, in his
Myriobiblon or Library, Photius, the ninth-century Patriarch of Constantinople, offers a summary
of Aenesidemus’ lost Pyrrhonian Discourses, which is now our most important source for his
thought.

9 See Schmitt (1983, 227), Porro (1994, 230–5). For a detailed analysis of these manuscripts, see
Cavini (1977, 1–8), Floridi (2002, 63–9), and especially Wittwer (forthcoming).

10 See Cavini (1977, 4, 8–9), Floridi (2002, 79–80).

11 See Grellard (2007a).

12 See Schmitt (1976), Porro (1994, 236), Floridi (2002, 80–4).

13 See Schmitt (1983, 227), Porro (1994, 242–51), Grellard (2004, 114–5).

14 See Gregory (1974, 1984), Perler (2010a).

15 On the skepticism of John of Salisbury, see Grellard (2007b).

16 See, e.g., Grellard (2004) for a taxonomy of the thirteenth-century strategies against the
skeptical argument based on the illusion of the senses.

17 This view is fully advanced by Perler (2010b); cf. Grellard (2004, 128–9; 2007a, 342).
Whether an author felt the need to refute a skeptical argument or merely made a methodological
use of it seems to be, in principle, hard to determine with precision.

18 See Grellard (2004, 113; 2007a, 328).

19 The view, commonly accepted by scholars, that Pyrrhonism differs from Cartesian skepticism
in that it does not merely calls into question knowledge-claims but beliefs in general is rejected
by Brennan (1999).

20 The classic papers on this question are the five essays collected in Burnyeat and Frede (1997).
See also Glidden (1983), Stough (1984), Barney (1992), Brennan (1999), Fine (2000), Bailey
(2002, chapters 7–9, 11), Thorsrud (2009, chapter 9), and Perin (2010).

The Sceptic’s Tools: Circularity and Infinite Regress

Jan Willem Wieland*

Abstract
Important sceptical arguments by Sextus Empiricus, Hume and Boghossian (concerning disputes,
induction, and relativism respectively) are based on circularities and infinite regresses. Yet,
philosophers’ practice does not keep circularities and infinite regresses clearly apart. In this
metaphilosophical paper I show how circularity and infinite regress arguments can be made
explicit, and shed light on two powerful tools of the sceptic.

Keywords: scepticism, circularity, infinite regress, schema [Philosophical Papers 40: 359-69,
2011]

1. Three cases
Consider the following three well-known problems.

Problem of the Criterion. In order to decide a dispute, you need to obtain an agreed-upon
criterion by means of which you will decide the dispute. In order to obtain an agreed-upon
criterion, you need to decide a dispute about the criterion. Hence, there is a circularity or infinite
regress, and it is impossible to decide any dispute. Here is part of Sextus Empiricus’ initial text:

[…] if they wish to decide about the criterion by means of a criterion we force them into
infinite regress. Further, since proof requires a criterion that has been proved, while the
criterion has need of what has been determined to be a proof, they land in circularity.
(Outlines, Book 2, §4, 20)

Also compare Chisholm’s formulations from different pages:

[…] And so we are caught in a circle. (1973: 62)

[…] If we continue in this way, of course, we are led to an infinite regress. (ibid: 64)

Problem of Induction. In order to justify an inductive inference, you have to rely on the
assumption that the future resembles the past. In order to rely on the assumption that the future
resembles the past, you have to justify this assumption inductively. Hence, there is a circularity
or infinite regress, and it is impossible to justify any inductive inference. Compare part of
Hume’s initial words:1

[…] all our experimental conclusions proceed upon the supposition that the future will be
conformable to the past. To endeavour, therefore, the proof of this last supposition by
probable arguments, or arguments regarding existence, must be evidently going in a
circle, and taking that for granted, which is the very point in question. (Enquiry, §4; cf.
Treatise, Book 1, ch. 3, §6)

Problem of Relativism. In order to be entitled to your epistemic system, you have to justify its
rules (such as ‘if you believe p and believe q if p, then it is permitted to believe q, but not
permitted to believe not-q’). In order to justify epistemic rules, you have to be entitled to an
epistemic system. Hence, there is a circularity or infinite regress, and it is impossible to be
entitled to an epistemic system. Compare part of Boghossian’s own words:

Suppose that you doubt some claim C and I am trying to persuade you that it’s true. […]
Now suppose that the context in question is the special case where C is the proposition
that R is truthpreserving and my argument for C is rule-circular in that it employs R in
one of its steps. (2001: 11-2; cf. 2006: chs. 5-7)2

In all three cases, the problems have a sceptical conclusion and can take the form of either a
circularity or an infinite regress argument. Hence the question is: What is the difference? In the
following I first present an informal answer in terms of Sextus Empiricus’ case (§2), and then a
more precise and general one (§3). Finally, I argue that this metaphilosophical investigation is
not only relevant to get clear and valid arguments, but also to obtain general strategies for
resisting the sceptic (§4).

2. Informal distinction

Let us consider the Problem of the Criterion as construed by Amico (1993: 35-6). So you have to
decide whether a certain proposition p is true. You can do this critically, i.e. by a proof, or
uncritically. If you do it uncritically, then your decision is arbitrary and will be discredited. But if
you do it critically and use a criterion c1 to decide whether p is true, you first need to decide
whether c1 correctly rules what is true and what is not. Again, there are two options: you can do
this critically, or not. If the latter, your decision will be discredited. So you do it critically and
have two options.

Option 1: You prove that c1 correctly rules what is true and what is not by showing that it gives
the right results. In this case, you already know what is true and what is not (and hence whether p
is true or not). But this is impossible because we started from the situation where you still have
to decide whether proposition p is true. This is the circularity.

dispute p -> use c1

|
V
dispute c1 -> use p (circularity)
|
V
use c2
|
V
dispute c2 -> use c1 (circularity)
|
V
use c3
…
(regress)

Fig. 1: Circularity vs. regress

Option 2: You prove that c1 correctly rules what is true and not by appealing to a meta-criterion
c2 which correctly rules what criteria correctly rule what is true and not. But now you first need
to decide whether c2 correctly rules what are the correct criteria. Again, there are two options:
you can do this critically, or not. If the latter, your decision will be discredited after all. So you
do it critically and have two options. Either you prove that c2 correctly rules what is true and not
by showing that it gives the right results. This, again, is a circularity. Or you prove that c2
correctly rules what is true and not by appealing to a yet another meta-criterion c3. This is the
regress.
As both routes seem to lead to an impossibility, it follows that we cannot decide whether any
proposition is true (cf. Amico 1993: 36). The question is: How to make things precise, and
generalize this difference for other cases such as Hume’s and Boghossian’s?

3. Formal distinction

Hence the problem is to find something that the three sceptical arguments from §1 have in
common. I will be assuming three restrictions on, or desiderata for, my solution:

• What arguments can have in common is an argument schema of which they are
instances.

• Circularities and infinite regresses should be kept apart, but still support the same
sceptical conclusion.

• The argument schema should be as simple as possible, and the sceptical conclusion
should follow by classical rules of inference.

Here is my solution.

Sceptical Schema (SCEP)

(1) ∀x, you can φ x only if you can ψ x first.

(2) ∀x, you can ψ x only if (i) you can φ x first, or (ii) ∃y, you can φ y first.

(3) ∀x, you can φ x only if (i) you can φ x first, or (ii) you can φ an infinity of items first.
[from 1, 2]

(4) ∀x, you cannot φ x first, nor φ an infinity of items first.

(5) ∀x, you cannot φ x. [from 3, 4]

There are three premises, viz. lines (1), (2), (4), and two inferred lines, viz. (3) and (5). To get
instances of this argument schema, the relevant domain is to be specified, the Greek letters ‘φ’,
‘ψ’ to be replaced with a predicate which expresses an action involving the items in that domain.
For the three cases from §1, the instances of the letters would be the following:

domain φx ψx

disputes decide x obtain an agreed-upon criterion by which x can be decided

inductive inferences justify x assume that the future resembles the past such that x can be derived

epistemic rules be entitled to x justify x inferentially

By these filling instructions, the following sceptical conclusions can be obtained by SCEP:

• You cannot decide any dispute.

• You cannot justify any inductive inference.

• You cannot be entitled to any epistemic rule.

Here is for example the first case in full:

Problem of the Criterion (SCEP instance)

(1) You can decide a dispute x only if you can obtain an agreed-upon criterion by which
x can be decided first.

(2) You can obtain an agreed-upon criterion y by which a dispute x can be decided only if
(i) you can decide x first, or (ii) you can decide a dispute about y first.

(3) Hence: You can decide a dispute x only if (i) you can decide x first, or (ii) you can
decide an infinity of disputes first. [from 1, 2]

(4) You cannot (i) decide a dispute x first, nor (ii) decide an infinity of disputes first.

(5) Hence: You cannot decide any dispute. [from 3, 4]

It is worth noting that lines (1) and (2) in this case do not depart too much from Sextus
Empiricus’ initial premises:3

In order to decide the dispute that has arisen […], we have need of an agreed-upon
criterion by means of which we shall decide it; and in order to have an agreed-upon
criterion it is necessary first to have decided the dispute about the criterion. (Outlines,
Book 2, §4, 20)

The main rationale of SCEP is that two abilities (being able to φ and ψ the members of a certain
domain, e.g. being able to decide disputes and obtain agreed-upon criteria to decide them)
mutually depend one another such that they cannot get off the ground.4

Let me briefly explain the circularity/infinite regress difference in terms of SCEP. Most
importantly, the circularity is captured by the (i)-clauses of lines (2), (3) and (4), whereas the
infinite regress is captured by their (ii)-clauses.

In case of the infinite regress, the inference of clause (ii) of line (3) is important. It is to follow
from (1) and (2) by Transitivity in combination with Conjunction Introduction in the Implicatum.
Another way of putting this clause would be this: ‘∃y, you can φ y, and ∃z, you can φ z, and ∃v,
you can φ v, etc. first’ (where each item is new and distinct from the others). It may be
disputable whether you can reach infinity by Conjunction, but what is important for the argument
is that the number of items exceeds your capacity.

In case of the circularity, ‘first’ is indispensible. Without ‘first’ clause (i) of line (3) line would
read: ‘you can φ x only if you can φ x’, which is trivial. To say that you cannot φ x first is to say
that you cannot already φ x before φ-ing x (e.g. decide a dispute before deciding that very same
dispute). Yet, ‘first’ need not be read temporally. So ‘you can decide a dispute only if you can
obtain an agreed-upon criterion first’ does not necessarily mean that the criterion must be there
earlier in time. What matters is that your ability to decide a dispute depends upon your ability to
obtain an agreed-upon criterion, and not vice versa.5

Finally: for sure the Problems of the Criterion, Induction and Relativism do not exhaust
philosophy, and many more important and influential cases (such as the Justification Regress, the
Cartesian Circle, Carroll’s Tortoise, Wittgenstein’s Paradox, etc.) could be set out as instances of
SCEP.

4. Meeting the sceptic

It is often unclear whether, and sometimes even disputed that, a given circularity or infinite
regress is bad or vicious. Boghossian (discussing the problem of rule-circularity, as opposed to
premise-circularity), for example, asks: “But why should this be considered a problem?” (2001:
10) My general answer here is straightforward:

Viciousness. Infinite regresses and circularities are bad whenever they establish a
sceptical conclusion.

Three clarifications. First, sceptical conclusions are to be conclusions of the form ‘∀x, you
cannot φ x’. Second, how infinite regresses and circularities can establish such a conclusion I
discussed in the previous section (or at least, I presented one option). Last, to my knowledge this
take on viciousness has not been defended or even considered in the literature so far (for the
discussion on regress arguments, see Wieland 2012).

No matter which particular debate, if one regards the sceptic conclusion as absurd, or at least
unacceptable if alternatives are available, then a natural way to resist it would be to deny one of
relevant premises. On the basis of SCEP the options are easily available.6

Specifically, the idea is to add the anti-sceptical claim of the form ‘you can φ at least one item of
the domain’ as a premise to the argument (PREM), regard the others as hypotheses for reductio
from now on (HYP), and then conclude, by Reductio Ad Absurdum (RAA), that one of the
hypotheses has to go as they are jointly incompatible.

Anti-Sceptical Schema (ANTI-SCEP)

(1) ∀x, you can φ x only if you can ψ x first. [HYP]

(2) ∀x, you can ψ x only if (i) you can φ x first, or (ii) ∃y, you can φ y first. [HYP]
 (3) ∀x, you cannot (i) φ x first, nor (ii) φ an infinity of items first. [HYP]

(4) ∃x, you can φ x. [PREM]

(5) ~(1), ~(2) or ~(3). [from 1-4; RAA]

So, there are three main options, viz. you may reject (1), (2) or (3) of ANTISCEP (and all other
options are to be combinations of these). I shall not go through the options in terms of all three
debates mentioned in this paper, but limit myself to Chisholm’s anti-sceptical position. Chisholm
famously distinguished two views: particularism and methodism (1973: 66). Basically,
particularism is the view which takes particular instances of knowledge as primitive and criteria
for knowledge as derivative. Methodism has it the other way around: criteria for knowledge are
primitive and particular instances derivative.

In terms of the options just listed, particularism is an example of the first strategy (viz. ~(1)) and
resists the sceptical conclusion by rejecting ‘you can decide which are instances of knowledge
only if you can have criteria for knowledge’. This was Chisholm’s own choice. By contrast,
methodism is an example of the second strategy (viz. ~(2)). Specifically, it resists the sceptical
conclusion ‘you cannot decide which are instances of knowledge’ by rejecting ‘you can have a
criterion to decide whether proposition x is a piece of knowledge only if either (i) you can decide
whether x is a piece of knowledge first, or (ii) there is another criterion y and you can decide
whether y is a piece of knowledge first’. Chisholm did not discuss the third strategy. Still, that
position could be called circularism or infinitism (depending on what part of (3) is rejected).

All in all, all anti-sceptical responses to a circularity or infinite regress argument can nicely be
framed in terms of ANTI-SCEP. Hence, the argument schema presented in this paper is not only
useful to get clear and valid arguments, but also to map the logical space for meeting the sceptic.

5. Global scepticism

I would like to conclude with a note on global scepticism. Usually the global/local distinction is
taken along the following lines. Local skepticism denies that knowledge regarding a specific
subject matter is possible. For example, one could be a local sceptic about religious matters and
deny that knowledge about such matters is possible. Whereas there can be many different local
sceptics, it is furthermore thought, there can be only one form of global scepticism. Namely,
global scepticism would be the view that knowledge regarding each and every subject matter is
impossible.

Yet, given what I have said in the foregoing, such a form of global scepticism is not even global
enough. To say that knowledge in general is impossible is merely to say that one instance of
SCEP is sound and has a true conclusion (i.e. ‘you cannot know any proposition’). Fully global
scepticism, by contrast, would be the position that all instances of SCEP are sound. Hence, it
accepts that it is impossible to decide disputes, to justify inductive inferences, to be entitled to
epistemic rules, and so on for all filling instructions of SCEP. Indeed: a view that I will defend at
another occasion.
* Thanks to: Mark Nelson, Maarten Van Dyck, Erik Weber, and audiences in Amsterdam and
Milan. The author is a PhD fellow of the Research Foundation Flanders at Ghent University.
Email: Jan.Wieland@UGent.be.

1 For the regress version of this argument, cf. Popper (1934: 29).

2 Rule-circular arguments are distinct from premise-circular ones. Whereas the latter purport to
prove a proposition by already assuming its truth, the former purport to prove the validity of a
rule by already assuming its validity (cf. Boghossian 2001: 11).

3 Assuming that: ∀x, you can φ x only if you can ψ x first = in order to φ x, you need to ψ x.

4 This notion of circular abilities is to be distinct from other common circularities: circular
arguments, definitions and explanations (yet perhaps the latter may be discussed in terms of the
former).

5 However this asymmetric dependence relation is spelled out. For a similar take on ‘first’ see
Van Cleve (2003: 50, fn. 12). Also cf. his reconstruction of the Cartesian Circle (1979: 55-6).

6 Cling (1994) makes a similar point for the Problem of the Criterion. Still, his reconstruction of
the problem is substantially different from an instance of SCEP. The main difference is that line
(4) of SCEP has no parallel in Cling’s case.

THE PYRRHONIAN PROBLEMATICAND GENERALITY IN ARGUMENTS

CONCERNING THE STRUCTURE OF JUSTIFICATION BY
RYAN MICHAEL HEBERT
WASHINGTON STATE UNIVERSITY

Abstract

Some of the most interesting criticisms of foundationalism, coherentism, and infinitism have
been shown to be rather general: some criticisms of a particular theory can be shown to apply to
the competitor theories as well. This opens the door for a number of important questions. How
many of the prominent arguments for or against the various structural theory theories can
actually be generalized in the sense of applying to all competing theories? What does it mean if
an argument or criticism can be generalized? If the property of generality is pervasive among
prominent arguments, why is it pervasive? If the property of generality is pervasive among
prominent arguments, what should we make of it? What would the consequences be for the
debate about the structure of justification if generality is pervasive?

We prepare the groundwork necessary to answer these questions. We explore foundationalism,

coherentism, and infinitism so that we may assess a sample of the most iv prominent arguments
concerning these theories. In particular, we study the regress argument for foundationalism, the
alternate systems objection to coherentism, and the finite minds objection to infinitism. In the
course of our study we find a peculiar result. For any of the three arguments, if we press it, the
core permute into a general argument. In other words, the reasons offered in support of a given
theory or to criticize a particular theory can, if pressed, be generalized such that they apply to all
structural theory theories of justification. What should we make of this peculiar result? It is hard
to say. While we cannot begin to explain this peculiarity in the present essay, we sketch a
number of possible interpretations before closing.

TABLE OF CONTENTS
Page
ABSTRACT……………………………………………………………………………….. iii
SECTION
1. PRECIS OF THE THESIS…………………………………………………….. 1

1.1 Agrippa’s Trilemma……………………………………………………..3

1.2 Three Justificatory Principles……………………………………………5
1.3 Propositional and Doxastic Justification……………………………….. 8

2. FOUNDATIONALISM………………………………………………………... 12

2.1 Basic Beliefs…………………………………………………………… 13

2.2 Varieties of Foundationalism…………………………………………... 15

3. COHERENTISM………………………………………………………………. 17

3.1 Coherence Relation……………………………………………………. 22

3.2 Coherence and Entailment……………………………………………... 23
3.3 Coherence and Consistency……………………………………………. 27
3.4 Coherence and Explanation……………………………………………. 35

4. INFINITISM…………………………………………………………………… 41

4.1 Infinitism and Reasons………………………………………………… 44

4.2 Infinitism and Availability…………………………………………….. 46

5. DIALECTICAL PECULIARITY……………………………………………… 49

5.1 Regress Argument for Foundationalism……………………………….. 49

5.2 Alternate Systems Objection to Coherentism………………………….. 56
5.3 Finite Minds Objection to Infinitism…………………………………....61
5.4 Interpreting the Peculiar…………………………………………………68

BIBLIOGRAPHY………………………………………………………………………….. 76 1

1. Précis of the Thesis

The Pyrrhonian problematic is the perennial challenge to the possibility of a structural theory of
justification or knowledge. A structural theory of knowledge describes the structural features of
knowledge while remaining silent as to the question whether anyone, in fact, knows anything. A
structural theory of justification does the same for epistemic justification. Such a theory is
hypothetical: if someone possesses an item of knowledge (or justification), that item of has such-
and-such features. Thus a structural theory of knowledge (or justification) is, in principle,
logically consistent with the thesis of global ignorance (Klein 2007a: 4). It might well be the case
that no one, in fact, knows anything but that if anyone did know anything, whatever was known
would have those features described by the correct structural theory.

The conclusion of the trilemma is a skeptical one suggesting that no structural theory is possible.
Theorists have generally taken the skeptical conclusion as a reductio ad absurdum for at least
one of the premises of the ancient trilemma, leading to the development of the three competing
structural theories of justification: foundationalism, coherentism, and infinitism. In light of
dialectical competition among these options, theorists have adduced arguments – some quite
famous – thought to be specific to one structural theory in an attempt to support the theorist‟s pet
theory or refute competitor theories.

As it turns out, some of the most interesting criticisms of foundationalism (whether from
coherentists or infinitists), coherentism (whether from foundationalists or infinitists), or
infinitism (whether from foundationalists or coherentists) have been shown to be rather general:
some criticisms of a particular theory can be shown to apply to the competitor theories as well.
Lawrence BonJour‟s argument from doxastic ascent stands as a recent example. We contend that
similar results can be achieved, if only on a limited basis, when arguing for a particular theory.
This opens the door for a number of important questions. How many of the prominent arguments
for or against the various structural theory theories can actually be generalized in the sense of
applying to all competing theories? What does it mean if an argument or criticism can be
generalized? If the property of generality is pervasive among prominent arguments, why is it
pervasive? If the property of generality is pervasive among prominent arguments, what should
we make of it? What would the consequences be for the debate about the structure of
justification if generality is pervasive?

While we cannot begin to claim to answer any of these questions, we prepare the groundwork
necessary to answer them. It is chiefly the purpose of the present essay to explore
foundationalism, coherentism, and infinitism as the consequences of the ancient trilemma. We
explore these structural theories so that we may assess a sample of the most prominent
arguments concerning these theories. In particular, we study the regress argument for
foundationalism, the alternate systems objection to coherentism, and the finite minds objection to
infinitism. Admittedly, the cogency among these three arguments vary (the former two being
much better arguments than the latter), we choose these arguments because of their traditional
prominence in the literature.

After we complete our exploration of foundationalism, coherentism, and infinitism, we study the
three aforementioned arguments and find a peculiar result. For any of the three arguments, if we
press the argument, its core elements permute into a general argument. In other words, the
reasons offered in support of a given theory or to criticize a particular theory can, if pressed, be
generalized such that they apply to all structural theory theories of justification. What should we
make of this peculiar result? It is hard to say. While we cannot begin to explain this peculiarity in
the present essay, we sketch a number of possible interpretations before closing

1.1 Agrippa’s Trilemma

Agrippa‟s Trilemma (also called “the ancient trilemma” and “the Pyrrhonian problematic”) is
perhaps the most famous of ancient skeptics‟ arguments. It is a special argument for skepticism.
It is the perennial argument against the possibility of a structural theory of knowledge.

The trilemma appears in a terse passage of Sextus Empiricus‟s Outlines of Scepticism:

In the mode deriving from infinite regress, we say that what is brought forward as a
source of conviction for the matter proposed itself needs another such source, which itself
needs another, and so ad infinitum, so that we have no point from which to begin to
establish anything, and the suspension of judgment follows. … We have the mode from
hypothesis when the Dogmatists, being thrown back ad infinitum, begin from something
from which they do not establish but claim to assume simply and without proof in virtue
of a concession. The reciprocal mode occurs when out to be confirmatory of the object
under investigation needs to be made convincing by the object under investigation; then,
being unable to take either in order to establish the other, we suspend judgment about
both (Sextus 2000: 41).

A similar passage exists in the Lives of Eminent Philosophers, chronicled by Diogenes Laertius:

But Agrippa and his school added to them five other modes, resulting respectively from
disagreement, extension ad infinitum, relativity, hypothesis, and reciprocal inference. The
mode arising from disagreement proves, with regard to any inquiry whether in
philosophy or in everyday life, that it is full of the utmost contentiousness and confusion.
The mode which involves extension ad infinitum refuses to admit that what is sought to
be proved is firmly established, because one thing furnishes the ground for belief in
another, and so on ad infinitum. The mode derived from relativity declares that a thing
can never be apprehended in and by itself, but only in connection with something else.
Hence all things are unknowable. The mode resulting from hypothesis arises when people
suppose that you must take the most elementary things as of themselves entitled to
credence, instead of postulating them: which is useless, because some one else will adopt
the contrary hypothesis. The mode arising from reciprocal inference is found whenever
that which should be confirmatory of the thing requiring to be proved itself has to borrow
credit from the latter, as, for example, if anyone seeking to establish the existence of
 pores on the ground that emanations take place should take this (the existence of pores)
as proof that they are emanations (Diogenes 2005: 501).

The argument proceeds roughly as follows. Take any arbitrary proposition p that an epistemic
agent S affirms. The trilemma outlines three possible models S can employ in the attempt to
justify the belief that p.

The first is a model of an endless regress of reasons. When asked why S believes that p, S offers
and affirms q. When asked why S affirms q, S offers and affirms r. When asked why S affirms
r…, so on ad infinitum. The trilemma suggests that the problem with this model is that there is no
way to motivate the justification of a given belief. Moreover, finite minds cannot have an infinite
number of beliefs. Ergo, the first model cannot be the correct model for the structure of
justification for S’s belief that p.

The second is a model of terminal reasons. When asked why S believes that p, S offers and
affirms q. When asked why S affirms q, S ceases and terminates the chain of reasons at q. The
trilemma suggests that the problem with this model is that the terminus is arbitrary. Without any
good reason to accept q, it is puzzling how it could appropriately justify S’s belief that p. Ergo,
the second model cannot be correct the correct model for structure of justification for S’s belief
that p.

The third is a model of circular reasons. When asked why S believes that p, S offers and affirms
q. When asked why S affirms q, S offers and affirms r. When asked why S affirms r, S offers and
affirms p. The trilemma suggests that the problem with this model is that it endorses circular
reasoning, which is fallacious. Ergo, the third model cannot be the correct model for the structure
of justification for S’s belief that p.

In short, the Pyrrhonian problematic contends that epistemic justification “must either proceed in
a circle, lead to a regress, or be ended by some mere assumption” (Lehrer 1989b: 141). Given
that none may be a satisfactory structural theory of epistemic justification for S‟s belief that p at
t, skepticism obtains.

1.2 Three Justificatory Principles

We may crystallize these considerations with the following three principles. By “justifier” we
refer to a proposition or belief that supports a proposition or belief. By “justifier ancestry” we
refer to the links in the chains of justifiers, sometimes branching, that support beliefs (Klein
1999: 298):

Principle of Avoiding Regress (PAR): for all propositions, x, if x is warranted for a

person S, at t, then there is some justifier, j1, available to S for x at t; and there is some
justifier, j2, available to S for j1 at t, etc., and there is some justifier, jn, available to S for
jn-1 at t such that no further justifier, jn+1, is needed for jn for S at t.

Principle of Avoiding Arbitrariness (PAA): for all propositions, x, if x is warranted for

a person S, at t, then there is some justifier, j1, available to S for x at t; and there is some
 justifier j2, available to S for j1 at t, etc., and there is no last reason in the series (Klein
1999: 299; Klein 2005a: 136).

Principle of Avoiding Circularity (PAC): for all propositions, x, if x is warranted for a

person S, at t, then for all y, if y is in the justifier-ancestry of x for S at t, then x is not in
the justifier-ancestry of y for S at t (Klein 1999: 298; Klein 2005a: 136).

These principles correspond to the three structural models of Agrippa‟s Trilemma. Either
accepting or rejecting all three principles has the consequence of skepticism. The various
possibilities for a structural theory of knowledge lie between the two extremes. There are a few
things to be said about the principles.

First, the principles do not jointly exhaust the necessary conditions for an adequate theory of
justification, whether foundationalist, coherentist, or infinitist. As Klein notes, at least non-
defeater and non-overrider principles would need to be introduced (Klein 2005a: 139).1 Non-
defeater and non-overrider principles are not necessary for the entailment of any of the structural
theories, but any theory without the principles would not be satisfactory. Thus when we say that
PAC and PAR jointly entail foundationalism, or that PAA and PAR jointly entail coherentism, or
that PAA and PAC jointly entail infinitism, we do not pretend that these substantive principles
are sufficient for an adequate theory. We ask the reader to generously interpret these statements
elliptically; if the reader so desires, the reader may insert a clause excluding defeaters and
overriders where appropriate. We suppress all such clauses throughout.

Second, as we have noted above, all structural theories of knowledge or justification are
consistent with certain varieties of skepticism. Insofar as skepticism is characterized by its
commitment to theses of ignorance, ignorance is diverse in its type and degree. On the one hand,
skeptics of a certain sort could hold that we, in fact, possess no justified beliefs. This might be
because some global agniology, a thesis to the effect that it is not possible to acquire any justified
beliefs, is true. Sextus Empiricus (2000) and Peter Klein (2003) accuse Academics skeptics of
dogmatically accepting such an agniology. Ironically, David Hume (2007) and A.C. Graying
(2003) accuse the Pyrrhonians of the same dogmatism. Regardless, it is important to note that
such a thesis is blind to the structural question: what are the features of justified beliefs if anyone
were to possess any? Skepticism of this first sort is consistent with actuality of a correct
structural theory.

On the other hand, skeptics of another sort could hold that it is impossible to develop an adequate
theory about the features of justified beliefs. This is the kind of skepticism featured by the
Pyrrhonian problematic. Barry Stroud (1989; 2000) interprets this Pyrrhonian argument as a
challenge to the possibility of providing a general explanation of the structure of justification.
Strictly speaking, such a thesis is blind to the questions of actuality: do we, in fact, have any
justified beliefs? Is possible to acquire any justified beliefs? PAR, PAA, and PAC jointly entail
that there can be no satisfactory structural theory of knowledge. Skepticism of this second sort
obtains.

It should not come to the surprise that Agrippa‟s Trilemma has not secured the skeptical
conclusion in the minds of many philosophers. Nonetheless it has been the source of long-
standing debate about the correct structural theory of epistemic justification. The dominant
structural theory has been and continues to be foundationalism. Foundationalism rejects PAA;
PAC and PAR jointly entail the structural theory of foundationalism. Another view, coherentism,
attempts to offer a structural theory by rejecting PAC; PAA and PAR jointly entail the structural
theory of coherentism. The classical debate about the correct structural theory has taken place
between the foundationalists and coherentists to the exclusion of the infinitist option. This is
because the structural theory of infinitism, the view that rejects PAR and is jointly entailed by
PAA and PAC, has received little scholarly attention until very recently.

1.3 Propositional and Doxastic Justification

Before proceeding to the explication of foundationalism, coherentism, or infinitism, we must

prepare the underpinning of the analysis. The underpinning consists in the definition of the
relationship between two types of epistemic justification, propositional and doxastic.
Propositional justification is potential justification of sorts for an epistemic agent. As John Turri
expresses, propositional justification “attaches to propositions relative to individuals” or, as
Klein expresses, “such justification is an epistemic property of propositions rather than a
property of belief states” because a belief may be “justified for you even though you don’t
actually believe it, or believe it for the wrong reasons, or believe it for the right reasons but in the
wrong way” (Klein 2007: 8; Turri 2009: 209). Propositions for which S is propositionally
justified in accepting are those which are merely “justifiable” or if S is in “a position to
justifiedly believe” (Turri 2010a: 312). We might say that S is entitled to believe that p if p is
proposition justified for S.

Doxastic justification is actual justification of sorts for an epistemic agent, attaching “to concrete
belief states (and to doxastic states more generally)” or, put another way, “belief states are the
bearers of doxastic justification” (Klein 2007: 8; Turri 2009: 209). If the belief that p is
doxastically justified, it is not merely justifiable but “justified”, where p is something for which
S is “justifiably believing” (Turri 2010a: 312). Without pausing to survey2, the view that doxastic
justification is little more than propositional justification plus the basing principle is pervasive in
the literature. The basing principle is a condition that specifies that if the belief that p is
propositionally justified for S because some reason r and S believes that p on the basis of r, then
the belief that p is doxastically justified for S (Turri 2010a: 314).

Turri (2010a) has produced compelling counterexamples to the basing principle. Roughly and
very briefly, the counterexamples proceed as follows. Suppose that reasons r1, r2, r3, …, rn,
propositionally justify belief that p for S and that S believes that p on the basis of r1, r2, r3, …,
rn. What is left to be explained the manner in which S accepts r1, r2, r3, …, rn as reasons for p.
Presumably the only epistemically admissible ground for accepting r1, r2, r3, …, rn as reasons
for p is that they make p sufficiently likely to be true. Yet it is perfectly possible that S accepts
those reasons on epistemically inadmissible grounds. For instance, S might accept r1, r2, r3, …,
rn as reasons for p on a whim, or by luck, or because of the advice of an astrologist, or because
of the reading of tea leaves, or because a hallucinated interlocutor coerced S to assent, etc. Since
S can satisfy the conditions of the basing principle and yet fail to achieve doxastic justification,
the principle is false.
Turri argues that we need a principle wherein “the subject’s intellectual abilities explain why she
is in a position to justifiedly believe or know p” (Turri 2010a: 320). Put another way, if S has
properly produced the relevant reasons for the belief that p such that S is doxastically justified in
the belief that p, then S has manifested the relevant cognitive dispositions in the production of
those reasons, whatever the relevant cognitive dispositions may be (Turri 2009: 217). He
therefore endorses the following principle, especially as an explanation of the relationship
between propositional and doxastic justification:

(PJ) Necessarily, for all S, p, and t, if p is propositionally justified for S at t, then p is

propositionally justified for S at t because S currently possesses at least one means of
believing p such that, were S to believe p in one of those ways, S’s belief would thereby
be doxastically justified (bold mine; Turri 2010a: 320).

In light of Turri’s remarks, we formulate propositional and doxastic justification as follows:

A proposition p is propositionally justified for S at time t if and only if there is at least

one justifier, j1, of kind k such that j1 is a good and undefeated reason to believe p.

A proposition p is doxastically justified for S at time t if and only if p is propositionally justified

for S at t and S‟s production of justifiers, j1, j2, j3, …, jn, for p actually manifests S’s relevant
cognitive dispositions.3

There are two important elements in the formula: justifiers of kind k and the relevant cognitive
dispositions. Let us begin with a word about the latter.

It is in the interest of every theorist for an account about what exactly constitutes those
epistemically relevant cognitive dispositions, whether they are perceptual faculties, the
deliverances of memory, the reliable testimony of others, causal efficacy, the production of
conscious reasons, etc. None of these dispositions excludes, at least from the outset, any
structural theory. Arguably, since no well-developed theory exists about those relevant cognitive
dispositions, the formulation of propositional and doxastic justification will have to be
sufficiently vague to accommodate the development of an adequate theory. If a theorist so
desires, they may insert whatever dispositions they find preferable (within reason, of course).

Similar remarks apply to justifiers of kind k. The precise nature of justifiers of kind k is a matter
of philosophical dispute. Neither the foundationalist, nor the coherentist, nor the infinitist should
reject either formulation because theorists of any stripe are free to specify what qualifies as kind
k for a justifier. Foundationalists might say that a justifier of kind k is either a belief that
possesses or stands in the appropriate justificatory relation to foundational property Φ (the
property constitutive of basic beliefs). Coherentists might say that a justifier of kind k is a belief
that coheres with S’s other beliefs better than its competitors. Infinitists might say that a justifier
of kind k is a good and undefeated reason that is a member of an infinite, nonrepeating series.
We examine all these proposals in detail below.

2. Foundationalism
Foundationalism is, broadly, defined by three theses. First, foundationalism is a terminal or, as
Carl Ginet puts it, a “finitist” theory of justification (Ginet 2005: 141). It describes the ramifying
chain of justification as one that finds its terminus at some belief or class of beliefs. This is a
consequence of foundationalism’s endorsement of PAR.

Second, foundationalism distinguishes between two types of doxastically justified beliefs:

justified beliefs are members of either the “superstructure” or the “foundations” (Alston 1993:
144). Member beliefs of the superstructure are nonbasic and member beliefs of the foundations
are basic. This is a consequence of foundationalism’s rejection of PAA.

Nonbasic beliefs derive their justification by standing in the right kind of justificatory relation –
inferentially – to epistemically prior justified beliefs, whether these prior beliefs are nonbasic or
basic. However, no nonbasic belief is autonomously justified. All prima facie warrant enjoyed by
nonbasic beliefs can be traced back to foundational beliefs.

The defining quality of a basic belief is the privilege basic beliefs enjoy. The virtue of such
epistemic privilege allows them, at least in principle, to transmit justification to nonbasic beliefs.
Unlike nonbasic beliefs, “the source of the justification of foundationally justified beliefs is not
other beliefs” (Audi 1993: 8). The source of a belief’s prima facie justification is criterial for its
taxonomical identification as either basic or nonbasic.

Third, foundationalism is a linear theory of justification. This is a consequence of

foundationalism’s endorsement of PAC. Linearity is a poorly-explicated concept, but I will
utilize two related traits: the direction of justification and the privilege of certain beliefs with
respect to other beliefs. In reality these traits are two faces of the same coin. Foundationalism
depicts the direction of justification in a justificatory chain one-way, usually from the bottom up.
We saw this, above, with foundationalist’s second thesis. Those “bottom beliefs” are
epistemically prior; they serve as premises for the conclusion of later beliefs and are the
derivative source of justificatory merit possessed by these conclusions. Foundationalism affords
a strong privilege to certain classes of belief – basic beliefs – in that basic beliefs only serve as
premises.

2.1 Basic Beliefs

There are, broadly, two formulae by which one may typify basic beliefs. Call these two models
F1 and F2. F1 and F2 are disparate formulae of basic beliefs, but the trait they share is Φ. The
property Φ is denotative of basic beliefs since it is the source of not only the autonomous warrant
enjoyed by basic beliefs but the justificatory status of the totality of the foundationalist’s beliefs,
basic and nonbasic alike. A belief is autonomously warranted just in case the belief possesses at
least prima facie epistemic justification and the source of the justification is not other beliefs.
This justification is, in some sense, an intrinsic feature of the basic belief. It is noteworthy that
the mere possession of Φ is sufficient for a basic belief to be propositionally justified.

The formulae of F1 and F2 foundationalisms are therefore as follows:

F1 Basic beliefs are autonomously warranted in virtue of some feature Φ. Beliefs

possessing property Φ are properly basic.
 F2 Basic beliefs are not autonomously warranted and possess some feature Φ. Beliefs
possessing property Φ are properly basic but are neither epistemically justified nor
unjustified.

Consider a parallel analysis. For what Sosa calls “formal foundationalism”, the epistemic
justification possessed by a given belief supervenes on the relevant property possessed by that
belief (Sosa 1980: 229 – 231). Thus when a basic belief is justified, its possession of the property
of being propositionally justified, supervenes on property Φ. When a nonbasic belief is justified,
its property of being propositionally justified supervenes on two properties: first, the property Φ
possessed by basic beliefs and, second, the property of standing in the right kind of justificatory
relation to basic beliefs.

There are many candidates for the defining property Φ of a properly basic belief: rationalists like
Descartes (1998) and BonJour (1998, 2003) roughly hold that innate or rationally intuited ideas
are properly basic; empiricists like Hume (2007, 2008) and Russell (1948) roughly hold that
experiential beliefs are properly basic; a tracking theorist like Nozick (1981) roughly holds that
sensitive beliefs are properly basic; virtue epistemologists like Greco (2000, 2010) and Sosa
(2003, 2007, 2009) roughly hold that apt or responsible beliefs are properly basic; reliabilists like
Goldman (1986, 1999) and Plantinga (1993) roughly hold that beliefs reliably formed are
properly basic; a Wittgensteinian roughly holds that a hinge proposition is properly basic
(Pritchard 2004; Strawson 1985; Wittgenstein 1969). Any number of different formulations of Φ
are, at least one the face of it, up for grabs. We primary speak of the abstract property Φ because
it generalizes over all foundationalist theories.

2.2 Varieties of Foundationalism

Positism (not to be confused with positivism), as Van Cleve calls it, adopts F2. Roughly, positist
foundationalism holds that some beliefs, though they are not themselves (propositionally or
doxastically) justified, can serve to justify other beliefs (Van Cleve 2005: 168). The details of the
interpretation diverge, but it does not seem unfair to attribute the makings of such a view to
Wittgenstein in his On Certainty (McGinn 2008; Wittgenstein 1969; Wright 2004).

With very few exceptions, F1 has been the principal model for basic beliefs. F1 foundationalism
goes as least as far back as Aristotle. It has been so commonly proposed since Aristotle that
foundationalist theses can be subdivided among three classes. Lawrence BonJour has outlined
the characteristic features of these classes, calling them strong, moderate, and weak
foundationalisms:

According to moderate foundationalism, the noninferential warrant possessed by basic

beliefs is sufficient by itself to satisfy the adequate-justification condition for knowledge.
Thus on this view a basic belief, if true, is automatically an instance of knowledge
(assuming that Gettier problems do not arise) and hence fully acceptable as a premise for
the justification of further… beliefs. By virtue of their complete justificatory
independence from other… beliefs, such basic beliefs are eminently suitable for a
foundational role (BonJour 1985: 26).
Notice that moderate foundationalism is a fallibilist model. Infallibilism in a nutshell is the thesis
that a belief can be epistemically justified but false. Put another way, justification does not
purchase the guarantee of truth for justified beliefs; they are only likely to be true. Infallibilism is
the denial of fallibilism. Infallibilism holds that epistemic justification purchases a guarantee of
truth for those beliefs that are epistemically justified. Justified beliefs are not merely likely to be
true, they must be true.

Strong foundationalism is an infallibilist model. It adopts the core tenets of moderate

foundationalism, taking exception only with the traits of basic beliefs. The basic beliefs of strong
foundationalism are “infallible, certain, indubitable, or incorrigible” (BonJour 1985: 26).
Weak foundationalism, being weaker than moderate foundationalism, is a fallibilist model but
does not characterize its basic beliefs as possessing sufficient autonomous warrant to qualify
them either as knowledge or “as acceptable justifying premises for further beliefs. Such beliefs
are only „initially credible,‟ rather than fully justified” (BonJour 1985: 28).

If basic beliefs are insufficient to meet the adequate-justification condition for knowledge, then
how can the weak foundationalist account for instances of knowledge (assuming that we have
any)? The problem seems troublesome when it is remembered that basic beliefs are the
epistemically most prior in a justification chain. The answer is an appeal to coherence (Poston
2010). “The weak foundationalist solution to this problem is to attempt to augment the
justification of both basic and nonbasic beliefs by appealing to the concept of coherence”
(BonJour 1985: 28).

3. Coherentism

Coherentism arose to a short-lived supremacy among British idealists in the beginning of the
20th century. Its early proponents were Harold H. Joachim (1906), F.H. Bradley (1914), and
Bernard Bosanquet (1911; 1920). Brand Blanshard (1939; 1962) developed his distinctive
variety of coherentism after spending time in Oxford, where he met F.H. Bradley. Despite all
this, coherentism was overshadowed by the rise of the Vienna Circle and its philosophy of
analysis. Since then it been tenuously held by only a handful of philosophers. Most famously this
includes contemporary thinkers like Lawrence BonJour (1985) and Keith Lehrer (1974; 2000),
though BonJour has since formally disavowed coherentism (BonJour 1999). Susan Haack (2009)
and Catherine Elgin (1996; 2005) also offer important contributions to the theory of coherence,
though their status as coherentists is less than incontrovertible. Haack (1997) has defended her
position as essentially foundationalist despite argument to the contrary from BonJour (1997).
Elgin (2005), who identifies as a coherentist, has received criticism from Van Cleve (2005) to
the effect that her theory is actually a form of weak foundationalism.

Coherentism was historically a reaction to foundationalism and the correspondence theory of

truth, though the coherentists were sometimes confused about the latter score. We speak only of
coherentist theories of justification. Because coherentism was reactionary, it is easy to elucidate
it in terms its qualms with foundationalism.
Coherentism is, broadly, defined by three theses. Like foundationalism before it, coherentism is,
first, a terminal theory of justification. Foundationalism and coherentism agree that the regress of
justifying reasons for the belief that p comes to an end. This is a consequence, as we have seen,
of foundationalism’s and coherentism’s endorsement of PAR.

Second, coherentism rejects the distinction between basic and nonbasic beliefs. This is a
consequence of coherentism’s endorsement of PAA. Whereas foundationalism conceives of the
ultimate origin of doxastic justification to be some feature Φ of a basic belief rather than the
belief itself, one way of thinking about coherentism is that coherentism emphasizes the primacy
of other beliefs in the doxastic justification of any given belief. In a sense, all beliefs, justified or
not, are nonbasic in the sense that they lack property Φ of basic beliefs – no belief for the
coherentist is autonomously warranted. In another sense, all justified beliefs are basic in sense of
a specific property that all justified beliefs possess; namely, coherence. Coherence can be
distributed among two models of coherence-bearing.

The first model of coherence-bearing, the warrant-transfer model, posits that members of a
coherent set have the primary function of communicating propositional justification. This means
that member beliefs are the primary bearers of the coherence relation and would therefore qualify
as being propositionally justified.

The second model of coherence-bearing, the warrant-emergence model, makes use of two
separate properties when talking about coherence. There is the property of membership in a
coherent set possessed by individual beliefs and there is the property of coherence possessed by a
set. Coherence is not a reducible property of any particular belief, but instead a property of a set
of beliefs whereby the set is the primary bearer of justification (Klein 2007a: 8). For a belief to
be propositionally justified is for it to possess the relational property of coherence – belief p is
coherent with beliefs q, r, etc. – and coherence is intelligible only in lieu of other beliefs, just as
the concept of validity applies only to arguments or the designation of marriage applies only to
those individuals with at least one spouse.

Third, coherentists tend to reject the linear model of epistemic justification in favor of a
nonlinear model. This is a consequence of coherentism’s rejection of PAC. As we saw, above,
linearity is a function of the direction of justification and the privilege of certain beliefs vis-à-vis
other beliefs. Consider three argument arrangements for the set of beliefs, {p, q, r}, and assume
the arguments are valid:

(i) (ii) (iii)

p q r
q r p
r p q

On a linear model, the belief r in (i) would be derivatively justified via its relationship to p and q,
as conclusions are analogously derived from premises such that p and q are epistemically prior to
r; p and q are privileged with respect to r. If, say, p is basic, then it is always privileged in the
sense that it its source of justification is never derived from its position as a conclusion in any
argument. Therefore foundationalist models cannot endorse (i), (ii), and (iii) provided that p, q,
or r is basic. If, as in our example, a foundationalist endorses (i), the foundationalist has excluded
their ability to endorse (ii); whether the foundationalist is permitted to endorse (iii) depends on
additional factors. Weak foundationalists might endorse both (i) and (iii), for example. By
contrast, a nonlinear model of justification is free to endorse (i), (ii), and (iii). There are two
possible ways to construe the coherentist’s endorsement of (i), (ii), and (iii). They are, as we
have already seen, the warrant-transference model and the warrant-emergence model (BonJour
1985; Klein 2005a: 135; Klein 2007a: 8; Sosa 1980).

The warrant-transfer model is roughly a linear model. The property of being propositionally
justified is transmitted via one belief to another whereby all recipients of justification are
circularly arranged. Thus propositional justification, as it is for foundationalism, is transmitted
when beliefs stand in the right kind of relation – an inferential relation – to one another. The
source of the property of being propositionally justified is a fair question, and it seems rather
doubtful that the circular arrangement of beliefs can spontaneously generate warrant. Klein
adopts the following analogy for the warrant-transfer model to express the problem:

…Think of basketball players standing in a circle passing the ball round and round. It is
important to note that this view is parasitic on the foundationalist account of
propositional justification. For once again, somehow justification arises in the circle of
propositions and is transferred via inference from one proposition to another. Thus, it is
not surprising that Aristotle employed the foundationalist concept of a fixed epistemic
priority among propositions in developing his criticism of this transference form of
coherentism mentioned earlier. To exploit the analogy a bit: Transferring the basketball
seems easy to understand, but how the basketball got there in the first place is mysterious
(Klein 2007a: 8).

A related stumbling block for the warrant-transfer model is that it seems to endorse fallacious
reasoning. When a given belief, say r, is called into question, S will offer some reasons that p and
that q; as in (i). But, on the warrant-transfer model, S ultimately endorses that p and that q on the
basis of that r; as in (ii) and (iii). In short, that r provides its own justificatory basis by virtue of
justificatory transitivity on the warrant-transfer view.4 If that r is called in question such that
some other reasons, that p and that q, is supplied to support that r, S would have no rational
footing for the belief that r. It would simply be question-begging to assert that r, the thing in
question, as grounds for that p and that q (Klein 2005a: 134). Perhaps even worse for the
coherentist, if r provides its own justificatory basis on the warrant-transfer view, then r – to the
extent that it possesses any warrant at all – is autonomously warranted. In other words, it would
seem that r would be more like a foundational belief than a candidate for coherentism.

The second form is the warrant-emergence model. According to the warrant-emergence model,
the property of a belief’s being propositionally justified supervenes on the property of
membership in a coherent set. Warrant-emergence models do not “think of propositional
justification as a property attached to a proposition that can be transferred to another one. Rather,
it views justification as an emergent property such that when sets of propositions have a certain
arrangement – a coherent structure – all members of the set of propositions are justified” (Klein
2007a: 8). The model posits that a member belief, p, is justified in virtue of the fact that it both
serves as a reason for other beliefs, q and r, and that other beliefs, perhaps even q and r, serve as
reasons for p. There is no privileged belief in the sense that any belief may serve as premises in
some arguments and be derived as a conclusion from other, though distinct, arguments. If p and
q are members of a coherent set, p will sometimes be a reason for q; other times, q the reason for
p. It depends on what is being granted in the course of the argument. This is what we understand
to be the essence of the Neurath metaphor.5 What matters is that the sense in which these beliefs
may be either premises or conclusions is that they do so at different reasoning sessions. If p is 22
challenged in an argument and q is offered in support of that p, then p cannot be available as a
reason for q. “That would be to fall into circular, question-begging reasoning” (Klein 2005a:
136).

The three theses of coherentism intertwine when the regress terminates at a sufficiently coherent
set of beliefs wherein the member beliefs are justified as a result either because of their circular
arrangement or because of their membership in the coherent set. However, it is unlikely that any
person has endorsed warrant-transference coherentism. BonJour and Lehrer, among others, have
endorsed warrant-emergence coherentisms.

3.1 Coherence Relation

Lawrence BonJour has offered an excellent intuitive description of coherence, saying that
“coherence is a matter of how well a body of beliefs „hangs together‟: how well its component
beliefs fit together, agree or dovetail with each other, so as to produce an organized, tightly
structured system of beliefs, rather than either a helter-skelter collection or a set of conflicting
subsystems” (BonJour 1985: 93). However appealing the intuitive description may be, one of the
traditional bugbears for coherentists is the lack of precision or consensus about what exactly
constitutes the coherence relation. What does it mean for beliefs to “hang together”? What are
the necessary and sufficient conditions that a collection of beliefs needs to satisfy to qualify as a
minimally coherent set?

Pollock and Cruz, for example, opt to divide theories of coherence among positive and negative
accounts. Positive coherence theories require positive support in the form of reasons for the
acquisition of justified beliefs and negative coherence theories construes reasons negatively,
affording grounds for the rejection of beliefs, but do not require justification for acquired beliefs
because all beliefs are prima facie justified; beliefs are epistemically innocent until defeated
(Pollock and Cruz 1999: 70). Keith Lehrer (op. cit.) is exemplary of positive coherence and
Gilbert Harman (1984) is exemplary of negative coherence.

Peter Murphy (2006) utilizes an interesting taxonomy for proposed elements of the coherence
relation. The classification is distributed among propositional and psychological relations, where
propositional relations include deductive, probabilistic, and explanatory subcategories. Though
my organizational structure has been inspired by Murphy, we consider only three candidate
conditions for the coherence relation. I have opted to trade an economy of breadth for one of
rigor. In the following sections, we examine entailment, consistency, and explanatory
requirements for coherence.

3.2 Coherence and Entailment

Early coherentists conceived of an important element of coherence as one involving implication
or, what amounts to the same, necessary connection. Contemporary coherentists have construed
early formulations of the coherence relation as being constituted by entailment and criticized it
on that basis (BonJour 1985: 96 – 98, 214 – 217; Lehrer 1974: 157 – 159; Lehrer 2000: 100 –
101). Lehrer offers the following formulation of the entailment condition for coherence:

Thus, a belief that p coheres with other beliefs of a system C if and only if p either
necessarily implies or is necessarily implied by every other belief in C (Lehrer 1974: 157
– 158; Lehrer 2000: 100).

Compare Blanshard:

Fully coherent knowledge would be knowledge in which every judgment entailed, and
was entailed by, the rest of the system (Blanshard 1939: 264).

Blanshard’s formulation is at once stronger and weaker than Lehrer’s. It is stronger in that it
requires a proposition to both entail and be entailed by all other propositions in a fully coherent
set whereas Lehrer’s formulation requires the satisfaction of only one of these conjuncts. Let us
consider two objections to the strong formulation of the entailment requirement of coherence.
The first objection comes from Lehrer (1974: 165; 2000: 100). It shows that we can construct
two fully coherent, though mutually exclusive, systems that are formally identical. For any set,
C, of logically consistent propositions, if some members of C are contingent and every member
of C necessarily implies or is implied by every other member of C, we can construct another set,
C*, which is made up of all the members of C except that those contingent member propositions
are negated. Given their mutually exclusivity, we must choose between C and C*, and coherence
is supposed to furnish the means to do so but fails. On the entailment conception of coherence,
the two sets are equally satisfactorily coherent. Ergo, entailment is not a sufficient condition for
coherence.

The second of Lehrer’s criticisms attempts to demonstrate that entailment is not a necessary
condition for coherence. The principle of the argument appears to be correct but the example
Lehrer adduces is problematic. We turn to Murphy for a support of Lehrer’s contention. Consider
the following four beliefs:

the belief that Moe is wincing, the belief that Moe is squealing, the belief that Moe is
yelling “that hurts”, and the belief that Moe is in pain. None of these beliefs logically
implies any of the others. Nor does the conjunction of any three of them imply the fourth.
Despite the lack of entailments, though, the beliefs together seem to constitute a system
of beliefs that is intuitively quite coherent. So coherence can be earned by relations
weaker than entailment (Murphy 2006).

We accept these two criticisms as decisive against the strong formulation of the entailment
requirement of coherence.

Let us take a moment to examine the weak formulation, which seems to be what Blanshard has
in mind. The weak formulation concedes that entailment is not a necessary condition for
coherence. The saving qualification for Blanshard is that it is a fully or maximally coherent set
that displays the conjunctive requirement of entailing and being entailed by every other member
belief. Blanshard says that we probably will “never in fact find a system where there is so much
interdependence” (Blanshard 1939: 265). The maximally coherent set, as we call it, is an ideal
(Blanshard 1939: 264, 265, 266). Despite our imperfect approximations of the maximally
coherent set, knowledge is still possible and it is justified in lieu of the coherence among beliefs
(Blanshard 1930: 265). This is because coherence is comes in degrees. The ideal set has other
qualities: member propositions are consistent, no member proposition is arbitrary, no member
proposition is independent or atomic, and every proposition entails and is entailed the other
member propositions jointly and singly (Blanshard 1939: 264, 265 – 266).

Blanshard offers an analogy of types of sets which display different degrees of coherence and, it
would seem, degrees of epistemic justification. The set that is not coherent, presumably 26
one in which member beliefs of the set are not epistemically justified, “would be a junk-heap,
where we could know every item but one and still be without any clue as to what the remaining
item was. Above this would come a stone-pile,” which we take to be the minimally coherent set
satisfactory for justification, “since here you could at least infer that what you would find next
would be a stone” (Blanshard 1939: 265). If indeed Blanshard is suggesting that the stone pile is
analogous to a minimally coherent set, then, strictly speaking, he is not appealing to deductive
inference to justify the belief that the next would be a stone. Rather, the mode of inference is
induction:

All items i1, i2, i3, …, in from the heap have been stones.
Item in+1 is from the heap.
Item in+1 is a stone.

If Blanshard sanctions inductive inference as satisfactory for minimal coherence, then it would
seem that there is agreement from early coherentists that entailment is not a necessary condition
for the coherence relation. Another way to think of the analogy via its extension is to consider
explanation for a phenomenon or set of phenomena. If, for instance, one knows something about
members of the heap, namely that they are stones, and finds a sample of similar enough stones
nearby, one might infer that the stones came from the heap. The inference would be legitimated
by appeal to Peircean abduction (Burch 2010):

All items i1, i2, i3, …, in from the heap have been stones.
Item in+1 is a stone.
Item in+1 is from the heap.

Regardless of whether agreement exists between early and contemporary coherentists over the
role of entailment in the coherence relation, the foregoing arguments demonstrate the difficulty
for entailment requirements of the coherence relation.

3.3 Coherence and Consistency

Blanshard posits that consistency is a desirable quality of coherence. We take him to mean
logical consistency. Some contemporary coherentists posit that logical consistency is a necessary
condition for knowledge. For example, BonJour says that logical “consistency is one requirement
for coherence, [and] that inconsistency is obviously a very serious sort of incoherence” (BonJour
1985: 95). However, BonJour notes that logical consistency is inadequate and adds a
probabilistic consistency requirement:

Suppose that my system of beliefs contains both the belief P and also the belief that it is
extremely improbable that P. Clearly such a system of beliefs may perfectly well be
logically consistent. But it is equally clear from an intuitive standpoint that a system
which contains two such beliefs is significantly less coherent than it would be without
them and thus that probabilistic consistency is a second factor determining coherence
(BonJour 1985: 95).

An important distinguishing factor of probabilistic consistency is that it depends on how many

probabilistic conflicts exist within the belief system and that probabilistic consistency is a matter
of degree (BonJour 1985: 95). Thus BonJour expresses two criteria necessary for the satisfaction
of the coherence relation:

(1) A system of beliefs is coherent only if it is logically consistent.

(2) A system of beliefs is coherent in proportion to its degree of probabilistic consistency.

Lehrer too takes issue with a logical consistency requirement for the coherence relation, saying
that consistency, “even if necessary for coherence, as it is here assumed to be, is not sufficient
for coherence” (Lehrer 1974: 163; Lehrer 2000: 104). Lehrer sees consistency as an issue
wrapped up in the larger project of explanatory coherence:

We must require that a belief cohering with a system either explain or be explained in
relation to the system better than anything that contradicts it. Contradiction must be made
relative to the system. Two mutually contradictory statements may be such that a system
of beliefs entails that they cannot both be true. We shall speak of such beliefs
contradicting each other and thus employ a relativized concept of contradiction. […] A
belief coheres with a system of beliefs if and only if the belief is consistent with the
system and either explains something in relation to the system not explained better by any
belief that contradicts it, or the belief is better explained by something in relation to the
system, and nothing that contradicts it is explained better (Lehrer 2000: 105).

Setting aside explanatory coherence, Lehrer’s use of the word “contradiction” is not equivalent
to its formal meaning. For Lehrer’s purposes, a candidate belief q “contradicts” candidate p if q
is either a contrary or contradictory of p. Contraries and contradictories of p compete with p for a
place in S‟s acceptance system because if q is admitted into S‟s acceptance system, whether q is
either a contrary or a contradictory, entails that p is inconsistent with S‟s acceptance system.
Whether p or its competitor q is accepted depends on their complementarity with S‟s acceptance
system. Complementarity, for explanatory coherence, is assessed in terms of whether and how
well a candidate acceptance explains or is explained by other members of S‟s acceptance system.
It is not difficult to see the importance of consistency. Without consistency, it is difficult to see
how p and q compete for admission into S‟s acceptance system. For any acceptance p, the
acceptance that p must be consistent with the rest of the system. Is this type of consistency
tantamount to logical consistency? It is difficult to say. Perhaps Lehrer has in mind a type of
inconsistency that a responsible epistemic agent must avoid that is weaker than its logical
formulation. If so, we should naturally ask: what does this type of egregious inconsistency look
like and how is it different than an acceptable inconsistency?

Peter Klein (1985) distinguishes between what he calls strong inconsistency and weak
inconsistency. A set of propositions is strongly inconsistent when two member beliefs, p and q,
are inconsistent; or, alternatively, when the conjunction of two member beliefs, p and q, is
necessarily false. A set of propositions is weakly inconsistent when, strictly speaking, it is
inconsistent but not strongly inconsistent. Klein appeals to the Lottery Paradox as an example of
a weakly inconsistent set (Klein 1985: 108). Another example of a weakly inconsistent set, γ, is
the following: {p ⊃ q, p, ~q}. The set γ is inconsistent, but the conjunction of any two members
of γ is not necessarily false. Obviously the truth of any two member propositions of γ would
entail the falsity of the remaining member proposition. Perhaps Lehrer wishes to say that no
coherent set can be strongly inconsistent. If he wished also to reject the permissibility of weak
inconsistency, he would endorse logical consistency as a necessary condition for justified beliefs.

It is clear that logical consistency is not a sufficient condition for coherence. The example from
Murphy, above, is telling against the sufficiency of consistency. However, this does not answer
the question whether logical consistency is a necessary condition for coherence and, by
extension, justification and knowledge. The answer is highly contentious. Foley (1978), Klein
(1985), and Kornblith (1989) have argued that logical consistency is improbable or impossible to
attain, though by different means and for different dialectical purposes; Wunderlich (2010) notes
that Goldman (1999) and Cherniak (1986) also argue against the attainability of consistency.
Kornblith and Wunderlich quote a notable passage from Cherniak:

Given the difficulties in individuating beliefs, it is not easy to estimate the number of
logically independent atom propositions in a typical human belief system, but 138 seems
much too low – too „small minded‟. Yet suppose each line of the truth table for the
conjunction of all these beliefs can be checked in the time a light ray takes to traverse the
diameter of a proton, an appropriate cycle of time for an ideal computer. At this
maximum speed, a consistency test of this very modest belief system would require more
time than the estimated twenty billion years from the dawn of the universe to the present
(Cherniak 1986: 211).

The contention, in view of its relation to coherentist theories of justification, is that “it is
impossible for human beings, or any information process device, to accept beliefs in virtue of
their coherence with others, so long as the field of beliefs over which coherence must operate is
tolerably large” (Kornblith 1989: 211). We may express the argument thus:

P1 Consistency is a necessary condition for coherence.

P2 Consistency is unattainable.
C Coherence is unattainable.

There are four possible assessments of the argument. The first is to agree with Kornblith and
reject coherentism as a plausible structural theory. The second is to reject the first premise by
eliminating logical consistency as a necessary condition for coherence. Third, coherentists could
reject the second premise and attempt to argue that consistency is attainable for typical epistemic
agents. Fourth, coherentists could stick to their guns and bite the skeptical bullet. By insisting
that they have the correct structural theory account of justification, coherentists could take
Kornblith’s argument to be an argument for skepticism: if coherence is necessary for justification
and we cannot attain coherence, then we could just as easily give up the unstated premise that we
possess justified beliefs (and, by implication, that it is possible to obtain justified beliefs).

Wunderlich (2010) adopts the third assessment. He proposes a three-step algorithm for checking
consistency in our set of beliefs:

First, one identifies propositions that must be understood as logically complex. Second,
one divides those into groups of logically complex propositions that contain overlapping
sets of atomic propositions. It is only within such a group that one could find a logical
inconsistency. Finally, one checks each of the groups for consistency (Wunderlich 2010:
4).

One must formalize and exhibit the logical complexity of the various sentences because
formalizing them sententially, e.g. the set {A, B, C, …}, precludes the possibility of
inconsistency. Inconsistency is possible only with sufficient logical complexity. However, one
need not translate all possible complexity of implicit in a given sentence. Instead, one need only
translate “as much of the logical complexity of the first sentence as is necessary to exhibit
relationships with other sentences” (Wunderlich 2010: 4).

After one has formalized the propositions under review, they are arranged according to
relevance. Those propositions which are overlapping are placed within their own groups.
Intuitively, logically related propositions about the history of the Roman Empire are presumably
grouped into a set distinct from those logically related beliefs about ice cream. Within these
distinct groups, one performs consistency checks.

Kornblith uses the number of sentences in a short novel as a proxy for a belief set. Wunderlich
references Jane Austin’s Pride and Prejudice as an illustrative example. He puts his method to
the test by performing a consistency check on the opening lines of the novel:

It is easy to check that nowhere else in the novel is the truth discussed (that a man in
possession of good fortune must be in want of a wife). However the first and second
sentences are formalized, one need not worry that they will conflict with the rest of the
novel…

One must then check that the first and second are compatible with each other. Is the
universal acknowledgement of a truth logically consistent with that truth being
particularly fixed in the minds of a group of people? Yes. There are therefore no worries
about the logical consistency of Pride and Prejudice regarding the first two sentences
(Wunderlich 2010: 5).

“The lesson,” Wunderlich writes, “is that sensible consistency checks proceed by dividing up the
task into manageable subtasks, and that the division itself is not too difficult” (Wunderlich 2010:
5). Though a belief corpus could be very large indeed, all that matters is that “sizes of the subsets
of atomic propositions that are linked by the subject‟s beliefs. Kornblith has given us no reason
to think that these subsets are large” (Wunderlich 2010: 5).

If “Kornblith has done nothing to show that any difficulties one might have in discovering such
inconsistencies have anything to do with the cosmically unmanageable size of a truth table,” then
we seem to have an attractive, elegant method to check for consistency (Wunderlich 2010: 6).
Perhaps consistency can be retained as a necessary condition for coherence after all.

There are two potential worries that coherentists may not be able to adopt Wunderlich’s
algorithm. The following considerations may be framed in the form of a dilemma: either it is
possible for epistemic agent S to check for consistency but not satisfy other coherence conditions
or it is possible for S satisfy other coherence conditions but not possible for S to check for
consistency. If consistency is posited as a necessary condition for coherence, then the resulting
conclusion would be that coherence is unattainable. To be clear, I am not arguing that
Wunderlich’s algorithm is faulty. It strikes me as plausible that a foundationalist could adopt it.
Rather I am worried that the algorithm presents an obstacle for the coherentist.

A subset is, as we have seen, populated only by those propositions which are logically related,
e.g. by those beliefs which are not logically relevant to other beliefs. If they were, the subset
would be populated by those relevant beliefs. For consistency checks to be possible, the
population must be sufficiently small. This is where the first worry emerges. If the population is
too small – perhaps as it must be to perform a consistency check – then it likely fails to be robust
enough to satisfy other necessary conditions for coherence. For example, a collection too small
likely cannot satisfy Lehrer’s explanatory characterization of coherence (whereby member
beliefs either explain or are explained by other member beliefs), above, in such a way that could
satisfy the adequate-justification condition of knowledge.

The natural response to this worry is to emphasize that coherence ranges over the whole web of
belief and is not limited to operation within the subsets. However, there is a potential partitioning
problem. This is the second worry. Presumably there is a rough inverse relationship between the
size of the subsets and the number of subsets in S‟s overall belief set. Given how small a subset
must be so that a consistency check is possible, there is likely large numbers of these subsets. If
each subset comprises only those propositions that are logically related to one another, then it
should conversely hold that subsets are not logically related to other subsets. If not, then it
becomes difficult to see just how S’s beliefs, once divided among these subsets, are able to work
together in order to achieve coherence. Given enough partitions in the web of belief, the
justificatory power of coherence is diminutive.

To illustrate these worries, consider the opening lines of Pride and Prejudice concerning the
truth discussed (that a man in possession of good fortune must be in want of a wife). Though
easy to check for consistency, it is difficult to see how, for instance, they adequately explain or
are explained by components in the rest of the novel. If they are so distinct that mention of them
never appears anywhere else in the novel, it would seem that they fail to sufficiently cohere with
the rest of the novel such that they are not epistemically justified beliefs for S. It is clear that the
two sentences are not jointly incoherent; but not being incoherent is not sufficient for coherence.
Put another way, incoherence and coherence are contraries, not contradictories (Murphy 2006).

In short, we might worry that there is a tension between Wunderlich’s algorithm and other
necessary conditions of coherence. If, on the one hand, a subset small enough to perform
consistency checks, then it may not be to satisfy other necessary conditions such that S never
satisfactorily justify member belief of those subsets. If, on the other, a subset is large, then it may
not be possible to perform consistency checks. Moreover, if the subsets are too distinct, then it is
questionable whether one has formed partitions in one’s belief set and, therefore, whether these
subsets can cohere.

Before closing, I wish to gesture in the direction of one possible reply to these worries.
Wunderlich could insist that logical relevance does not exhaust the categories of relevancy over
which coherence ranges. Perhaps the example from Murphy, above, is instructional. For
example, coherentists who adopt an explanatory requirement for the coherence relation could
maintain that the explanation is not reducible to logical terms. If not, then the mere fact that
subsets are logically irrelevant to one another would not imply that they are, for instance,
explanatorily irrelevant to one another. If explanatory coherence can range over the subsets
despite their formal distinctness, then troubles arising from the smallness of each subset are
dispelled.

3.4 Coherence and Explanation

Coherentists have noted the shortcomings of consistency, entailment, and other proposed
conditions for the coherence relation. Some have turned toward explanation to supplement their
accounts of coherence. An immediate shortcoming with these attempts is that a theory of
explanation is no less a vexing philosophical problem than a theory of coherence. Moreover, in
hanging their hopes on explanation, coherentists are obliged to adopt an adequate theory of
explanation. One condition for adequacy is that the terms “justification” or “knowledge” are not
part of the analysis of explanation. Given failure of this condition, the analysis of explanation
would be in terms of justification and knowledge which would in turn be analyzed in terms of
explanation, leading the coherentist in a small circle (Lehrer 2000: 106).

Harman sees explanation as analogous to “because” statements: “On the revised account, we
infer not just statements of the form X causes Y but, more generally, statements of the form Y
because X or X explains Y” (Harman 1973: 130). Harman notes that explanations are of the form,
“In circumstances C, X’s tend to be Y’s” (Harman 1973: 131). Explanatory inferences are
justified “if the explanatory claim that X’s tend to be Y’s will explain why the next X will be Y is
sufficiently more plausible than competitors such as interfering factor Q will prevent the next X
from being a Y” (Harman: 132).

Compare with Lehrer who similarly regards explanatory coherence, as we have seen, as
dependent on the comparative reasonableness of a claim given its competitors. A “system C1 has
greater explanatory coherence than C2 if and only if C1 is logically consistent and C2 is not, or
both are consistent but more is explained in C1 than C2, or both explain the same things but
some things are explained better in C1 than C2” (Lehrer 2000: 105). Harman and Lehrer seem to
share the thought that the best explanation is an undefeated explanation.

We can characterize explanatory defeat through the analogue of justificatory defeat. Lehrer’s
account of justificatory defeat is too subtle and sophisticated to adequately exegete for present
purposes, but the relevant feature can be explicated through the heuristic of a justification game
(Lehrer 1989a: 253 – 256; Lehrer 1989b: 148 – 152). Consider S’s acceptance system where
member acceptances are personally justified for S if and only if they cohere with S‟s acceptance
system. In lieu of explanatory coherence, this might mean, among other things, that member
acceptances explain or are explained by other member acceptances.

A system X is an acceptance system of S if and only if X contains just statements of the form – S
accepts that p – attributing to S just those things that S accepts with the objective of accepting
that p if and only if p (Lehrer 1989a: 254). Imagine that S engages in the justification game with
an omniscient critic. This omniscient critic is allowed to make revisions to S’s acceptance system
when a target acceptance – S accepts that p – is false, replacing one or more such statements with
statements of the form – S accepts that not-p. In addition, the omniscient critic replaces all
statements that are closed by such revisions (Lehrer 1989b: 148 – 149). Statement q is closed,
roughly, when S accepts that p and p entails q. Call the resulting system of revised statements M:

If the critic can form a system M with the result that I am not justified in accepting p on
the basis of M at t, then the critic wins the game, and my justification is defeated. If, on
the other hand, my acceptance is such that no such system M has the result that I am not
justified in accepting p on the basis of M at t, then I win, and my justification is
undefeated (Lehrer 1989b: 149).

Analogously an undefeated explanation is an explanation which, like justification for Lehrer’s

justification game, is not eliminated by correction of content in S‟s web of belief. An
explanation is initially warranted if it is not beaten by competitor explanations. A competitor
explanation, c, is one in which it is more reasonable for S to accept the explanation p because q
on the assumption that c is false than on the assumption that c is true, on the basis of the
acceptance system X at t (Lehrer 1989a: 255). Explanation p because q beats competitor
explanation q because c for S at t if and only if it is more reasonable, given S’s acceptance
system X, to accept q than c on the basis of X at t. Undefeated explanations are like acceptances
of system M for which S retains personal justification. If, given M, an acceptance is justified,
then S‟s personal justification for the acceptance that p is vindicated. Likewise, explanations are
undefeated when, given corrections in S’s acceptance system X, the explanation is not eliminated
or undermined.

BonJour appeals to explanatory relations as a means to enhance the coherence of a belief set.
Construing the central insight of explanation negatively, he suggests that the web of belief
suffers from a loss of coherence if the set contains anomalous members; that is, if the set
contains member beliefs that cannot be explained by reference to other beliefs in the web. The
“distinctive significance of anomalies lies rather in the fact that they undermine the claim of the
allegedly basic explanatory principles to be genuinely basic, and thus threaten the overall
coherence of the system” (BonJour 1985: 99). He offers the following desideratum for an
account of coherence:

The coherence of a system of beliefs is decreased in proportion to the presence of

unexplained anomalies in the believed content of the system (BonJour 1985: 99).6

BonJour’s desideratum, though intuitively appealing, is deficient. If, upon reflection, we discover
unexplained anomalies in our belief set, we can repair our web in one of two ways. Either we can
increase the explanatory coherence of a system either “by adding statements that explain or by
subtracting statements to be explained” (Lehrer 2000: 109). From a formal point of view, both
are equally good. From an epistemic point of view, adding explainers tends to be more preferable
to ignoring evidence contrary to one’s world view. Yet to “reduce what is unexplained, one may
refuse to concede the truth of those statements that need explanation” (Lehrer 2000: 108). For
example, the belief that humankind literally descended from Adam and Eve seems to be at odds
with our best-established evolutionary theories and their data. But one could, at least in principle,
maintain a coherent belief set that retains the literal belief in the common ancestral linkage to
Adam and Eve if one jettisons the corpus of evolutionary theory, the data, and its corollaries.
Similarly for the Flat Earth Society, members can, at least in principle, coherently maintain the
belief that Earth is flat by eliminating evidence to the contrary. These examples smack of old
objections to coherentism. While we do not take them to be good arguments against coherence,
we refer to them as a way to suggest that explanatory coherence is needs to be supplemented
with a condition that eliminates cherry picking as a legitimate means of achieving coherence.

BonJour ultimately agrees with this assessment. The central lesson is that the “epistemologically
significant concept of coherence is bound up with the idea of justification, and thus any sort of
inference relation which could yield some degree of justification also enhances coherence,
whether or not such a relation has any explanatory force” (BonJour 1985: 100). Justification is
explicated, though not exhaustively, by explanation. Explanation is a condition of justification,
but since “coherence is also enhanced by inferential connections of a nonexplanatory sort,”
explanation is therefore not a sufficient condition of the coherence relation (BonJour 1985: 100).

A number of reasons can be adduced to supplement the conclusion of the insufficiency of

explanation for coherence. We will only mention two from Lehrer. First, there is the opacity
problem (Lehrer 2000: 116). If the fact that S’s belief that p explains or is explained by some of
other S’s beliefs is opaque to S, then the presence of opacity will fail to justify S in their belief
that p. An important element in explanation is transparency, meaning that S must recognize that
their justified beliefs are somehow related or hang together (Murphy 2006). This is to introduce a
psychological requirement for coherence that we will not further explore. Second, there are cases
in which one may arrive at the best explanation, given the available information, for something
but that even the best explanation is inadequate. “This suggests that we must require that a
hypothesis not only explain better than any alternative we can conceive but also that it must be a
comparatively good explanation, good enough so that we are justified in accepting it” (Lehrer
2000: 117).

4. Infinitism
Infinitism has not been endorsed at all in the history of philosophy until very recently. The name
“infinitism” did not appear in the literature, Klein notes, until Paul Moser (1984) and John Post
(1987) (Klein 1998: 919). The earliest known infinitist is Charles Sanders Pierce who has at least
the makings of the view (Aikin 2009: 72; Klein 1999: 306, 320 – 321). Though remaining
thoroughly underdeveloped, infinitism has benefitted from the erudition of Scott Aikin (2005,
2008), Jeremy Fantl (2003), Peter Klein (1998, 1999, 2005a, 2005b, 2007a, 2007b), and John
Turri (2009).

Infinitism is, broadly, defined by three theses. First, infinitism is a nonterminal theory of
justification. Unlike foundationalism and coherentism, infinitism posits that no “final” reason
can be offered such that the regress of reasons ends. Whatever the matter in question, the matter
is never completely settled – not even whether reasoning can completely settle a matter (Klein
2003; Klein 2005a). Roughly this is because for each answer, one can ask at least one question;
for each question, one can offer at least one further answer, etc. This is a consequence of
infinitism’s rejection of PAR.

Second, infinitism joins coherentism in its rejection of basic beliefs. The consequence of
infinitism’s endorsement of PAA leads to some parallels with coherentism. For both coherentism
and infinitism, there is a sense available in which all justified beliefs are nonbasic because that
they lack the denotative property Φ of basic beliefs. No belief for the infinitist or coherentist is
autonomously warranted. There is another sense available for coherentism and infinitism such
that all justified beliefs are basic in sense of a specific property that all justified beliefs possess.
This property can be distributed among two models of infinitist relations.

The first model of infinitist relations, the warrant-transfer model, posits that members of an
infinite series are justified only if each member of the series has communicated propositional
justification inferentially to each member sequentially. This means that member beliefs are the
primary bearers of both the property of being propositionally justified and of the property of
standing in the right kind of inferential relation to epistemically prior beliefs.

The second model of infinitist relations, the warrant-emergent model, makes use of two separate
properties when talking about warrant. There is the property of membership in an infinite series
possessed by individual beliefs and there is the property of warrant possessed by the series. The
latter property, like coherence, is not reducible property of any particular belief. The property of
being propositionally justified is emergent for both coherentism and infinitism – it arises, but
supervenes upon, the property of membership in the right kind of set, whether a coherent set or
infinite series (Klein 2005a: 135 – 138; Klein 2007a: 8).

Third, infinitism is, like foundationalism, a linear model of justification. This is a consequence of
infinitism‟s endorsement of PAC. As we saw, linearity is a function of two traits: the direction of
justification and the privilege of certain beliefs with respect to other beliefs. Infinitism is like
foundationalism in that justification moves only one-way, from those beliefs which appear
earlier in an infinite series to those which appear later. No belief may appear in its own evidential
ancestry. Moreover, infinitists posit that some beliefs are privileged with respect to other beliefs.
Consider the three argument arrangements for the set of beliefs, {p, q, r}, which we saw above,
maintaining the assumption that each argument is valid:
 (i) (ii) (iii)
p q r
q r p
r p q

As in the case of any of these arguments, the premises are both epistemically prior and privileged
with respect to the conclusions. Like foundationalism, infinitism rejects that all beliefs are on par
in the sense that we could justifiedly endorse (i), (ii), and (iii). If, for example, (i) is endorsed as
epistemologically cogent support for the belief that r, then the infinitist has rejected the
possibility of (ii) and (iii). Infinitism has something common with moderate foundationalism in
this respect.

That is where the similarities end however. Since infinitism rejects the doctrine of basic beliefs,
there can be no justified belief that is privileged in the sense of never having a role as both a
premise in one argumentative arrangement and the conclusion in another – though distinct –
arrangement. What distinguishes infinitism from coherentism is that those beliefs which serve as
premises for certain beliefs will always be privileged with respect to those beliefs. Thus if the
infinitist endorses (i), p and q are always footing for r and r can never be a member of the
evidential ancestries for either p or q. But there will be reasons that support p and q for which p
and q cannot, in turn, serve as support; and reasons for those reasons; etc.

The three theses of infinitism intertwine when S‟s belief that p is justified because an infinite,
nonrepeating series of reasons are available to S as a result of either its inferential arrangement in
an infinite, nonrepeating series or because of its membership in an infinite, nonrepeating series.
Thus in returning to PAA and PAC, justifiers take the form of good, undefeated reasons and kind
k takes the form of available reasons. To gain a clearer understanding of infinitism, we must
therefore answer two questions:

(1) What is it for S to have a reason available for the belief that p?

(2) What is it for S to have a reason available for the belief that p?

In section 4.1, we explain the relevant features of reasons so that we may answer (1). In section
4.2, we explicate the concept of availability so that we may answer (2). With both concepts at
our disposal, infinitist theories of epistemic justification can be appreciated for their depth and
sophistication.

4.1 Infinitism and Reasons

Like the use of “justifiers” and “kind k” for formulating propositional and doxastic justification,
the use of “reasons” by infinitists is such that any adequate theory of reasons may be inserted.
The theory of reasons at the infinitist‟s disposal is equally well-disposed to infinitism’s
competitors. It is in the interest of foundationalists, coherentists, and infinitists alike, Klein notes,
to develop an account of reasons because theorists of all stripes “will have to employ [“reasons”]
because each view holds that there are reasons for at least some of our beliefs” (Klein 2005:
137).

Some examples of candidate theories of reasons are the following (Klein 1999: 299; Klein 2005:
137; Klein 2007: 12). For all such theories, p is a reason for q if and only if:

1. p is true and it renders q probable; or

2. if p is probable, then q is probable, and if p is not probable, then q is not probable; or

3. if p were true, q would be true, and if p were not true, q would not be true; or

4. p would be acceptable as a reason for q in the long run by the appropriate epistemic
community; or

5. an impartial, uninformed observer would accept p as a reason for q; or

6. an omniscient critic would not revise the manner in which p is a reason for q; or

7. p would be offered as a reason for q by an epistemically virtuous individual; or

8. believing that q on the basis of p is in accord with one‟s most basic epistemic
commitments; or

9. there is at least one cognitive process available to S which reliably takes true beliefs
that p as input for the output true beliefs that q; or

10. p is evident to S and p makes q evident to S; or

11. p meets the appropriate conversational presuppositions.

The list is by no means meant to be exhaustive, but it should be amply clear that infinitism can
get behind any of these theories and “opt for whatever turns out to be the best account since each
of them is compatible with what the infinitist is committed to” (Klein 2007: 12). This result
should not be terribly surprising since foundationalists and coherentists can equally well get
behind any of these theories and opt for whatever turns out to be the best account. It is important
to therefore note that for any theory of reasons, it will be the case that “not just any proposition
will function as a reason for other beliefs” (Klein 2009: 300). Reasons are special in the sense
that they must behave like justifiers: they make the proposition that p sufficiently probable for S
at t.

Given the array and generality of possible varieties of reasons, it perhaps best simply to recall
our definition of doxastic justification:
 A proposition p is doxastically justified for S at time t if and only if p is propositionally
justified for S at t and S‟s production of justifiers, j1, j2, j3, …, jn, for p actually
manifests S’s relevant cognitive dispositions.

The manifestation of S‟s relevant cognitive dispositions in concrete instances is what we may
call reasoning, i.e. the production of reasons. The reasons thus produced are justifiers, j1, j2, j3,
…, jn, for S at t for the belief that p.

4.2 Infinitism and Availability

Klein (1999, 2005, 2007) has used examples, analogies, and ostensive definition to clarify the
components of availability to which infinitism is committed. If a proposition to be available to S,
it “must be appropriately ‘hooked up’ to S’s beliefs and other mental contents at t. In order for a
proposition to be available in this sense it need not be occurrently available or endorsed by S at
t” (Klein 2005: 136). Propositions available to S at t “are like money in S’s bank account that is
available to S if S has some legal way of withdrawing it even if S is unaware that the money is
there or takes no steps to withdraw it” (Klein 2007: 13). We formulate availability accordingly:

A proposition p is available to S at t if and only if S can produce p at t via the

manifestation of the relevant cognitive dispositions. 47

Consider the example of the mathematical proposition <366 + 77 = 437> for the relevant
mathematical cognitive dispositions. We can to come to accept that <366 + 77 = 437> by
utilizing our cognitive machinery despite the fact that we have never entertained the proposition.
We are, as Klein puts it, “disposed to think that 366 + 71 = 437 after a bit of adding given our
belief that 6 + 1 = 7, that 7 + 6 = 13, etc… Thus, there is clearly a sense in which we believe that
366 + 71 = 437 is subjectively available [to us] because it is correctly hooked up to already
formed beliefs” (Klein 1999: 308).

The lesson is threefold. First, occurrent beliefs do not exhaust the categories of belief that human
beings possess. It is in this wider sense of belief, one which includes dispositional beliefs,
whether construed as first- or second-order dispositions, that counts for the availability of p for S
at t. Second, as should come to no surprise to the reader, the phrase “relevant cognitive
dispositions” is a placeholder for an adequate theory of such dispositions. They probably include,
as we have seen, mathematical beliefs, but most definitely include others. Theorists are free to
insert whatever dispositions they require for their purposes. Third, nothing in the present
discussion of availability precludes foundationalism or coherentism. It may well be remarked
that availability, like reasons, is sometimes as essential to the foundationalist or to the coherentist
as it is to the infinitist. What separates infinitism from foundationalism and coherentism is a
peculiar specification of availability: for any doxastically justified proposition p, there is at least
one infinite, nonrepeating series of reasons available to S at time t for p. We might hasten to add,
however, that each theory has, with respect to its competitors, a unique specification of
availability.

One may ask whether it is possible for S to have an infinite number of available reasons for p at
t. Our prima facie answer is in the affirmative. Reconsider the example of the proposition <366 +
77 = 437>. The proposition was selected arbitrarily from an infinite set of such propositions. It
seems, at least on the face of it, that S could similarly have available the corpus of all such
arithmetic propositions. Nothing stops S from forming the belief that <984 + 593 = 1157>, or
that <4938 + 384 + 1465 = 6787>, etc.

Consider a supplementary example from Klein (2005). Suppose S knows only the artificial
language l consisting of no more than the primitive vocabulary {x is F, red, indexical “that”}
such that S can believe of an object: that is red. Suppose further that there are an infinite number
of red objects. S could believe of each object: that is red. Since the truth conditions for each of
the propositions affirmed is different, the beliefs are distinct. We should mention that infinitism,
at least as far as a structural theory of epistemic justification is concerned, is not committed to
the existence of an infinite number of red objects. That is not the point. Klein‟s example
admirably demonstrates how far – indeed infinitely far – our access to new propositions can
reach. Klein has shown that finite minds have available to them an infinite number of beliefs
(Klein 2005: 138).

Mere possibility is not enough. One may ask whether it is plausible for S to have an infinite
number of available reasons for p at t. While the answer to that question cannot be properly
addressed presently, it is worth suggesting that human understanding is a dynamic enterprise. We
are constantly inventing and refining new concepts to understand what was hitherto unknown.
Recent additions to the collective corpus of human beliefs include atoms, dark matter, evolution,
germs, gravitation, greenhouse gases, magnetism, nanotechnology, quarks, strings, etc., each
ushering huge expanses of theoretical domains, instrumentation, and methods.

They invite inquiry, questions, conundrums, etc., to be squared in light of our already existing
corpus of beliefs. Problems invite further expansions of our vocabulary and understanding.
In short, these terms are not mere neologisms. They are expansions in the availability of potential
reasons. How far can this enterprise go? Is it endless? Is there an end to inquiry? It hardly seems
worth staking a claim as to whether there is an end to inquiry so that one can summon up an
objection to the plausibility of S’s possession of an infinite number of available reasons for p at t.

5. Dialectical Peculiarity

We have completed our exegesis of foundationalism, coherentism, and infinitism. In the

following subsections we examine three of the most prominent arguments concerning each
structural theory respectively. In section 5.1, we inspect the regress argument for
foundationalism. In section 5.2, we scrutinize the alternative systems objection to coherentism.
In section 5.3, we look at the finite minds objection to infinitism. In 5.4, we conclude by
sketching a number of possible analyses of our peculiar findings.

5.1 Regress Argument for Foundationalism

One of the oldest arguments adduced for foundationalism is the regress argument. Aristotle
produced one of the first:
 Now some think that because one must understand the primitives there is no
understanding at all; others that there is, but that there are demonstrations of everything.
Neither of these views is either true or necessary.

For the one party, supposing that one cannot understand in another way, claim that we are
led back ad infinitum on the grounds that we would not understand what is posterior
because of what is prior if there are no primitives; and they argue correctly, for it is
impossible to go through infinitely many things. And if it comes to a stop and there are
principles, they say that these are unknowable since there is no demonstration of them,
which alone they say is understanding; but if one cannot know the primitives, neither can
what depends on them be understood simpliciter or properly, but only on the supposition
that they are the case.

The other party agrees about understanding; for it, they say, occurs only through
demonstration. But they argue that nothing prevents there being demonstration of
everything; for it is possible for the demonstration to come about in a circle and
reciprocally.

But we say that neither is all understanding demonstrative, but in the case of immediate it
is non-demonstrable – and that this is necessary is evident; for if it is necessary to
understand the things which are prior and on which the demonstration depends, and it
comes to a stop at some time, it is necessary for these immediate to be non-demonstrable.
So as to that we argue thus; and we also say that there is not only understanding but also
some principle of understanding by which we become familiar with the definitions.

And that it is impossible to demonstrate simpliciter in a circle is clear, if the

demonstration must depend on what is prior and more familiar; for it is impossible for the
same things at the same time to be prior and posterior to the same things – unless one is
so in another way (i.e. one in relation to use, the other simpliciter), which induction
makes familiar. But if so, knowing simpliciter will not have been properly defined, but
will be twofold. Or is the other demonstration not demonstration simpliciter in that it
comes from about what is more familiar to us (Aristotle 1995: 117 – 118)?

The argument is interesting because two distinct regress arguments can be generated from the
schema of the regress argument. The first regress argument is epistemic: if some epistemic
justification is inferential, some must not be inferential (Fumerton 2010; Klein 1999 310 – 311).
The second regress argument is conceptual: fundamental to our understanding of inferential
justification is an understanding of noninferential justification because we need to “introduce a
concept of noninferential justification in terms of which we can then recursively define
inferential justification” (Fumerton 2010). We consider only the epistemic regress argument
(ER).

Given Agrippa‟s Trilemma, Aristotle’s ER is but a familiar variation. To see how they compare,
consider the following formulation of the ancient trilemma:
 (1) For any belief, p, if p is justified for S, then p is justified either by an endless regress
of reasons, terminal reasons, or circular reasons.

(2) If S‟s belief that p is justified by an endless regress of reasons, then S must have a
good and undefeated reason, q, for p; and S must have a good and undefeated reason, r,
for q; and S must have a good and undefeated reason…

(3) But S cannot have an infinite number of good and undefeated reasons. For example, S
cannot because finite minds cannot have an infinite number of beliefs; or because S has
no point from which to begin establishing anything.

(4) So S’s belief that p cannot be justified by an endless regress of reasons.

(5) If S’s belief that p is justified by terminal reasons, then S must have a good and
undefeated reason, q, for p such that q needs no further reason.

(6) But S cannot have a good and undefeated reason, q, for p such that q needs no further
reason. For example, S cannot because there is no reason to think that that q is a good and
undefeated reason for p; and if there is a reason, r, to think that q is a good and
undefeated reason for p, then the regress has not terminated.

(7) So S’s belief that p cannot be justified by terminal reasons.

(8) If S’s belief that p is justified by circular reasons, then S must have a good and
undefeated reason, q, for p and S has a good and undefeated reason for q, namely p.

(9) But cannot have S must have a good and undefeated reason, q, for p and S has a good
and undefeated reason for q, namely p. S cannot because this is an instance of fallacious
reasoning, whether circular reasoning or begging the question.

(10) So S’s belief that p cannot be justified by circular reasons.

(11) It is not the case that S’s belief that p is justified by an endless regress of reasons,
terminal reasons, or circular reasons.

(12) Therefore, S’s belief that p is not justified.

Compare ER:

(1a) S has some justified beliefs.

(2a) For any belief, p, if p is justified for S, then p is justified either by an endless regress
of reasons, circular reasons, or noninferentially.

(3a) Assume for reductio that none of S’s beliefs can be justified noninferentially.
 (4a) If S‟s belief that p is justified by an endless regress of reasons, then S must have a
good and undefeated reason, q, for p; and S must have a good and undefeated reason, r,
for q; and S must have a good and undefeated reason…

(5a) But S cannot have an infinite number of good and undefeated reasons. For example,
S cannot because finite minds cannot have an infinite number of beliefs; or because S has
no point from which to begin establishing anything.

(6a) So S’s belief that p cannot be justified by an endless regress of reasons.

(7a) If S’s belief that p is justified by circular reasons, then S must have a good and
undefeated reason, q, for p and S has a good and undefeated reason for q, namely p.

(8a) But cannot have S must have a good and undefeated reason, q, for p and S has a
good and undefeated reason for q, namely p. S cannot because this is an instance of
fallacious reasoning, whether circular reasoning or begging the question.

(9a) So S’s belief that p cannot be justified by circular reasons.

(10a) It is not the case that any of S’s belief can be justified by an endless regress of
reasons or circular reasons.

(11a) S’s belief that p cannot be justified.

(12a) But (11a) contradicts (1a).

(13a) Therefore, some of S’s beliefs are justified noninferentially.

Unfortunately, Aristotle‟s argument does not offer good grounds to accept foundationalism. It
does not offer good grounds for two reasons.

First, ER does not qualify as a suitable response to skepticism. Premise (1a) is indispensible for
ER because it is required for the reductio to generate the contradiction that ultimately drives the
proof of noninferential justification. Given that the skeptic calls the possession of knowledge into
question, (1a) will not be so readily granted. ER is question-begging argument in the face of
skepticism. The foundationalist will need to find independent grounds to suppose that S has any
justified beliefs. But if the foundationalist can show that S has any justified beliefs, then S has
foundationally, e.g. noninferentially, justified beliefs and ER is superfluous.

Second, ER does not qualify as a suitable argument for foundationalism vis-à-vis coherentism or
infinitism. ER is something of a straw man. Foundationalism is as much a model of fallaciously
arbitrary reasoning as coherentism is a model of vicious circular reasoning or as infinitism is a
model of a vicious regress in reasoning. Foundationalism does not posit that basic beliefs are
arbitrary, yet that is what one horn of the trilemma, specifically premises (5) – (7). If the
foundationalist may reject (5) – (7), the coherentist should be able to reject (8) – (10) and the
infinitist (2) – (4). If the foundationalist is allowed to deny the anti-foundationalist premises of
the ancient trilemma and press the others, it would seem like an equally legitimate rhetorical
maneuver for the coherentist or the infinitist. Consider an analogue to ER for coherentism, ERC:

(1c) S has some justified beliefs.

(2c) For any belief, p, if p is justified for S, then p is justified either by an endless regress
of reasons, terminally, or by coherence.

(3c) Assume for reductio that none of S’s beliefs can be justified by appeal to coherence.

(4c) If S‟s belief that p is justified by an endless regress of reasons, then S must have a
good and undefeated reason, q, for p; and S must have a good and undefeated reason, r,
for q; and S must have a good and undefeated reason…

(5c) But S cannot have an infinite number of good and undefeated reasons. For example,
S cannot because finite minds cannot have an infinite number of beliefs; or because S has
no point from which to begin establishing anything.

(6c) So S’s belief that p cannot be justified by an endless regress of reasons.

(7c) If S’s belief that p is justified by terminal reasons, then S must have a good and
undefeated reason, q, for p such that q needs no further reason.

(8c) But S cannot have a good and undefeated reason, q, for p such that q needs no
further reason. For example, S cannot because there is no reason to think that that q is a
good and undefeated reason for p; and if there is a reason, r, to think that q is a good and
undefeated reason for p, then the regress has not terminated.

(9c) So S’s belief that p cannot be justified by terminal reasons.

(10c) It is not the case that any of S’s belief can be justified by an endless regress of
reasons or terminal reasons.

(11c) S’s belief that p cannot be justified.

(12c) But (11c) contradicts (1c).

(13c) Therefore, some of S‟s beliefs are justified by coherence.

It would not be difficult to make an analogue of ER for infinitism. The principal lesson is that
there is little reason, at least from the logic of the trilemma alone, to not instead privilege PAR
and PAA to conclude with the denial of PAC and the affirmation of coherentism; or not to
privilege PAA and PAC to conclude with the denial of PAR and the affirmation of infinitism.
One needs more than formal principles to argue for any of the three structural theories. One
needs some additional dialectical resource that legitimates the privilege Aristotle gives to PAR
and PAC.
Suppose for a moment that the foundationalist has recourse to some additional dialectical
resource that provides grounds to prefer PAR and PAC to conclude with the denial of PAA and
the affirmation of foundationalism. What shall we make of Aristotle’s argument then? In such a
case, the epistemic regress argument is at best doing the least amount of work in the overall
argument for foundationalism and is at worst completely superfluous in the overall argument for
foundationalism. The failure exhibited by the regress argument is that Aristotle offers the
epistemic regress argument precisely as the grounds to prefer PAR and PAC to conclude with the
denial of PAA and the affirmation of foundationalism.

5.2 Alternate Systems Objection to Coherentism

One of the classical objections to coherentism is, as Feldman (2003) calls it, the Alternate
Systems Objection (AS). It was one of the original arguments developed by classical
foundationalists in their efforts to undermine the case for coherence. BonJour expressed the
objection thus:

According to a coherence theory of … justification … the system of beliefs which

constitutes … knowledge is epistemically justified solely by virtue of its internal
coherence. But such an appeal to coherence will never even begin to pick out one
uniquely justified system of beliefs, since on any plausible conception of coherence, there
will always be many, probably infinitely many, different and incompatible systems of
belief which are equally coherent. No nonarbitrary choice between such systems can be
made solely on the basis of coherence, and thus all such systems, and the beliefs they
contain, will be equally justified (BonJour 1985: 107).

By way of summary, BonJour says that “no matter how high the standard of coherence is set, it
seems clear that there will be very many, probably infinitely many, systems of belief which will
satisfy it and between which such a coherence theory is will be unable to choose in an
epistemically nonarbitrary way” (BonJour 1986: 25). In a nutshell, the objection purports to
show that coherence is an inadequate theory of justification because we can construct several
mutually inconsistent belief sets that robustly satisfy all the conditions of achieving coherence.
Given that the beliefs are mutually inconsistent, they cannot all be true. Since the possession of
justification is a testament to a belief or belief set’s likelihood of being true and the coherence
theory of justification posits that beliefs are justified just in case they cohere with a belief set,
coherence alone cannot decide between these mutually inconsistent belief sets. If coherence
cannot so decide, coherence is an inadequate theory of justification.

We have two modes of general reply to AS. The first is an aggressive response. The aggressive
response notes that if AS grounds enough to reject coherentism, it would lead to skepticism. This
is because there is nothing about foundationalism per se or infinitism per se that eliminates the
possibility of alternate, inconsistent belief sets. Given disparate enough foundations between
different belief sets, resulting belief sets will likely be mutually inconsistent with one another. It
is doubtful that the foundationalist could combat this possibility by including all noted basic
beliefs of mutually incompatible belief sets for two reasons. First, there is probably an infinite
number of such belief sets. Second, completeness is likely unobtainable, especially where some
beliefs pertaining to complex formal systems are concerned. The latter reason alludes to Gödel.
The issue is equally well pronounced for the infinitist. For any segment of any length of a
propositionally justified series of reasons, there might be an infinite number of series that
replicate that segment.

It might be objected that the aggressive reply cannot be generalized so easily. Consider classical,
i.e. strong, foundationalism wherein the basic beliefs are justified and therefore true.
Foundationally justified beliefs are true because strong foundationalism maintains the thesis of
infallibilism. The infallibilist thesis is that justification has a guaranteed purchase on truth.
Consequently, if one were to compare any two strongly foundationally justified belief sets, one
might hold that either they must actually be the same belief set or that they can, at least in
principle, be subsumed under a single belief set. In either case, there cannot be any genuinely
incompatible strongly foundational belief sets.

We have a tempered concession. Strictly speaking, the objection is correct in that one cannot
generate any strongly foundationalist belief sets that are genuinely incompatible with one
another. But what is doing the work for the guarantee of consistency among belief sets is not
foundationalism but rather infallibilism. Our thesis is that it is that there is nothing about
foundationalism per se that precludes the possibility of the formation of a general AS that applies
to all structural theories. Our thesis obtains even in the face of such an objection because 59
infallibilism is not a thesis of foundationalism per se. Given that infallibilism is not an essential
thesis of foundationalism, our original contention is correct.

Of course, the thesis of infallibilism is essential to strong foundationalism. This leads us to our
second point. The move to infallibilism is not one that precludes adoption by a coherentist or an
infinitist. If one were to couple the thesis of infallibilism with coherentism or infinitism, one
would get the same result for coherentism or infinitism. The overarching thesis of the present
argument, namely that the essential dialectical features of AS, good or ill, can be generalized to
apply to all theories, would command additional support. This would be – and is – just another
instance in which the peculiar result of generality rears its head.

Additionally, the thesis of the aggressive response obtains because neither infallibilism nor the
possession of infallibly justified beliefs is inconsistent with the aggressive reply. It could just as
well be that the antecedent – that AS grounds enough to reject coherentism – is false and that
therefore the conditional – if AS grounds enough to reject coherentism, then skepticism – is true.
This could be the case, for example, if strong foundationalism was true and that AS was an
inadequate criticism of coherentism.

The second mode of reply is defensive. We should remark that AS is problematic only where
both of the two incompatible belief sets, λ and η, are well-justified. The attainment of sufficient
justification with respect to λ and η is theory-neutral. To the extent that λ or η are only weakly
justified, our commitments to either are analogously weak, and it seems like we would be free to
give up either – or both – in lieu of recognizing that they are incompatible. If, however, λ and η
individually exhibit high levels of justification, it seems like we could adopt both belief sets as
subsets of our overall belief set in an attempt to dissolve the apparent conflict.

The attempt to dissolve the apparent conflict between that λ and η would come via the formation
of the metabelief that λ and η are not, in actuality, incompatible. The content of that metabelief
would need to specify how the apparent incompatibility points not to a failure of theory, but an
incompleteness of theory. We should take the commitment to such a metabelief in an analogous
spirit to the commitment to theoretical constructs, like certain unobservable particles in
theoretical physics, or abstract objects in certain types of metaphysics. The metabelief would
therefore attempt dissolve the apparent conflict between λ and η by pointing to a failure of
imagination.

The high levels of justification for λ and η intimate to a believing subject that they are likely
correct. Of course, it may turn out that neither λ nor η are correct. The metabelief cannot
guarantee that λ and η are not genuinely incompatible with one another. It may turn out that they
are. The metabelief solution is admittedly hazardous if adopted without care. Precaution dictates
that this metabelief come with specifications concerning how to alleviate the apparent tension
between λ and η. Such specifications might include the direction of a research program,
reflective equilibrium, and belief revision in the face of genuinely incompatible or
propositionally unjustified sets of beliefs.

Concerns aside, it is hard, without further principle, to rule the metabelief solution out. It is not
obviously doxastically irresponsible. Those individuals prone to abuse the solution are likely
doxastically irresponsible agents for whom the solution is just another means to manifest that
irresponsibility. We should recognize that the fact would remain that doxastically irresponsible
agents are irresponsible with or without recourse to the solution. The solution would not be the
source of the problem. Furthermore, since, ex hypothesi, λ and η are well-justified, it appears that
the introduction of the metabelief is the theoretically most elegant solution available. It simply
seems more plausible, i.e. a better explanation, to suggest that the state of theory vis-à-vis λ and
η is systematically incomplete rather than systematically incorrect.

5.3 Finite Minds Objection to Infinitism

The Finite Minds Argument has been one of the most commonly cited reasons for the rejection
of the infinitist option. It has been articulated by many. Lawrence BonJour, for example, is one
such scholar who has appealed to it:

For if construed as a claim about actual human knowers, the infinite regress view clearly
entails the dubious thesis that any person who has any… knowledge at all possesses an
infinite number of… beliefs. And surely, the argument continues, this is impossible for a
creature with only a finite mental capacity and a finite brain. Though it is difficult to state
in a really airtight fashion, this argument seems to me an adequate reason for rejecting
[infinitism] (BonJour 1985: 24).

John Williams has also given the Finite Minds Argument clear expression:

The [proposed] regress of justification of S‟s belief that p would certainly require that he
hold an infinite number of beliefs. This is psychologically, if not logically, impossible. If
a man can believe an infinite number of things, then there seems to be no reason why he
cannot know an infinite number of things. Both possibilities contradict the common
 intuition that the human mind is finite. Only God could entertain an infinite number of
beliefs. But surely God is not the only justified believer (Williams 1981: 85).

For our purposes the Finite Minds Argument (FM) is formulated accordingly:

(1) If infinitism is true and we have any doxastically justified beliefs, then we must have
an infinite number of beliefs.

(2) But we cannot have an infinite number of beliefs.

(3) So either infinitism is false or we do not have any doxastically justified beliefs.

(4) We have some doxastically justified beliefs.

(5) Infinitism is false.

FM seems initially plausible. Unfortunately it rests upon deep misunderstandings about the
theoretical commitments of infinitism and is consequentially subject to a number of criticisms
that undermines its effectiveness. Peter Klein has spent much ink rebutting FM, especially
objecting to (1) and (2) (Klein op. cit.). We consider some of the rebuttals briefly.

We should reject (1) because infinitism does not require an infinite number of beliefs. As we
have seen, infinitism posits that for any doxastically justified belief that p, if S is justified in
believing that p at t, S must have an infinite, nonrepeating series of reasons available for p at t.
Propositions that are available for S, i.e. potential beliefs, are not tantamount to actual beliefs S
holds.

We should reject (2) because it rests on a questionable specification of belief. It is obvious

enough that finite minds cannot possess an infinite quantity of occurrent beliefs. If infinitism
required that we need to possess an infinite number of occurrent beliefs, it would be a bad
theory. But it does not. If the classification of belief is wider to include those relevant cognitive
dispositions, as discussed in section 4.2 above, then it is less than obvious that a subject cannot
have an infinite number of beliefs. In fact, it would seem that there are cases in which an infinite
number of beliefs can be had by a typical, finite epistemic agent.

It would be inappropriate to conclude that the advocates of FM ever intended to attack a straw
man. Infinitism is a newcomer in the debate about the correct structural theory of justification
and most iterations of FM predate any earnest developments of infinitist theory. Nevertheless,
FM pretty clearly misunderstands infinitism’s theoretical commitments and fails on that basis.
Some might suggest that though FM fails, it can be reformulated in a way that preserves FM’s
core elements but does not attack a straw man. Consider the following reformulation of FM
which we call the Finite Availability Argument (FA):

(1*) If infinitism is true and we have any doxastically justified beliefs, then we must have
an infinite number of beliefs.

(2*) But we cannot have an infinite number of beliefs.

 (3*) So either infinitism is false or we do not have any doxastically justified beliefs.

(4*) We have some doxastically justified beliefs.

(5*) Infinitism is false.

FA merits two immediate remarks. First, advocates of FM could get behind FA insofar as it
fulfills its promise of salvaging the essence of FM in such a way that does not suffer from FM‟s
immediate defects. Moreover FM and FA are formally identical and possess very little difference
in content. Second, the alterations in the content from FM to FA render Klein‟s responses to (1)
and (2) ineffective to either (1*) or (2*). FA forces the infinitist to seek a new line of defense.
Alas, whatever virtues FA preserves, it inherits serious defects from FM.

Before we analyze the familial defect shared by FM and FA, a word on one motivation to exploit
the familial defect. One option universally available to the infinitist is the rejection of the
premise that we possess any doxastically justified beliefs. If done so, infinitism would be
purchased at the cost of skepticism. We need not pause long to mention the obvious
disadvantages of such a response. Even worse than the theoretical commitment to skepticism is
the prospect that competitor theories can be defended without such a burdening commitment. If
so, the demands of parsimony might be sufficient to rule infinitism out as a viable competitor.
Infinitists prepared to defend their theories by appeal to skepticism will want to level the playing
field by arguing that FM and FA obliges foundationalists and coherentists to appeal to skepticism
as well.

The question now is whether there is a familial defect in FM and FA that can be exploited in
such a way that generates skeptical implications for all structural theories. It seems that there is
such a defect. By way of introduction, let us say that the most interesting feature of this defect is
that has a great number of different – though related – strains. We begin with the readiest strain
of the defect, namely that both (2) and (2*) are question-begging, in order to make the defect as
clear as possible. Consider a rough analogue of FA (call it FAf) for moderate foundationalism:

(1f) If foundationalism is true and we have any doxastically justified beliefs, then we
must have at least one basic belief available to us.

(2f) But we cannot have any basic beliefs available to us.

(3f) So either foundationalism is false or we do not have any doxastically justified

beliefs.

(4f) We have some doxastically justified beliefs.

(5f) Foundationalism is false.

More sophisticated iterations of FAf could be generated to include all foundationalisms. We

could similarly generate variants of FAf for strong foundationalism, weak foundationalism, and
even coherentism. Whatever the variant of FA, the shared error among the second premise of all
iterations of the argument is that it would all too be easy to refute any target structural theory by
simply asserting that one of the target theory’s core theses is false. In any argumentative context,
theorists are obligated to adduce good reasons to accept their preferred view and, what amounts
to the same, to reject competing views. In the debate about the structure of justification, there are
three competing views: foundationalism, coherentism, and infinitism. Given that there are
competitors, to offer that the competitor view is false as a reason for the belief that the
competitor view is false is a textbook case of begging the question. The seed of error existed in
FM, but it is difficult to detect in FM because (2) leans upon the intuitive force that we cannot
have an infinite number of occurrent beliefs. That seems true.7 But, as we have seen, infinitism
does not have such a commitment. When the second premise of FM is reformulated in FA so as
to avoid attacking a straw man, (2*) has no analogous intuitive support. We are left in want of
grounds to accept any variant of the second premise.

We should not be mistaken to say that no such grounds can be adduced. Perhaps they can. But,
as a matter of fact, no such grounds have been adduced for either FM or FA. Even if such
grounds were adduced such that either FM or FA is not question-begging, FM and FA would be
superfluous. We would not need either argument to make the case against the targeted structural
theory. The case would have already been made.

Consider two additional iterations of the familial defect. These iterations of the familial defect
are subtle. Their nature differs depending on how the elements of FM or FA are emphasized. The
diverse iterations are most clearly seen when considering infinitist responses to FM. This point is
perhaps best illustrated by example. Consider two cases.

First iteration. An objector might complain that a lot of to do is made of “availability” yet the
account is, admittedly, no more than a stand in for a more complete account. An objector may
well find the role of background beliefs – taken even in the wider sense of relevant cognitive
dispositions – somewhat mysterious about how these background beliefs make an infinite
number of reasons available to us. The objector might press that it is too difficult to see how, to
use a mathematical set of dispositions, a series of reasons, like that 6 + 1 = 7, that 7 + 6 = 13,
etc., form the justificatory basis for the belief that 366 + 71 = 437. The objector might be
dissatisfied with the lack of specificity about the nature of the justification chain for the belief
that 366 + 71 = 437, let alone more complex beliefs. Given these difficulties, the objector might
conclude, availability is an inadequate basis to defend infinitism.

Second iteration. An objector might notice that it is unclear how, given the warrant-transfer
model of infinitism, S could ever possibly be doxastically justified in their belief that p.
There is a serious problem lurking below the surface, one which Klein (cf. 2005) has discussed.8
Presumably there are both deductive and nondeductive inferences that are epistemically
acceptable means of transmitting warrant from beliefs with it to beliefs without it. The objector
might note that the beneficial properties – being erosion-proof and truth-preserving – that make
deduction epistemically desirable, assuming the relevant closure principles hold, do not exist for
nondeductive inferential principles. Over the course of warrant transmission, the credibility of
deductive inferences is not lost, but the same is not likely true for nondeductive inferences.
Granting that only a modicum of credibility is lost through nondeductive inferential warrant
transmission, then it might be objected that it would seem that an infinitely long series of reasons
connected inferentially would most certainly fail to provide warrant to satisfy the adequate
justification condition of doxastic justification for posterior beliefs in the series. Thus for any
transfer t, if t occurs after n, then, for any p inferred at n, p will not be justified for S; formally,
∀n [(n < t) ⊃ ∀p ~Jsp]. More plainly stated, given enough noninferential warrant transmissions
(those which occur at or after t) all epistemically posterior members would fail to possess
adequate warrant for justification.

The infinitist may exploit the defect in the first case by pointing out that specificity about
“availability” exists for foundationalism or coherentism either. If it is the case that
foundationalists and coherentists make use of the notion of availability, even if they do not call it
such, the objection cuts both ways. It is equally mysterious how a set of basic beliefs, whatever
they are, can ramify to justify the belief that 366 + 71 = 437, let alone more complex beliefs
much further away from the foundations. Parallel remarks apply to coherentism. There simply is
no developed theory of any type that provides such an explanation. If the lack of development of
an adequate theory of availability is good grounds to reject infinitism, it should be equally good
grounds to reject foundationalism and coherentism.

The infinitist may exploit the defect in the second case by pointing out that the problem is with
warrant-transfer models of justification, not with infinitism per se. It would not matter whether
the ultimate support schema specified by the structural theory was foundationalist, coherentist, or
infinitist. It is likely that, for any theory of justification, the number of possible nondeductive
inferences that retain an acceptable level of credibility is too small to support more than a
handful of even our occurrent beliefs, let alone our dispositions. This is not to say that no options
are available to salvage warrant-transfer models of justification. But it is to say that the problem
is one that applies generally – as would likely be the means to salvage warrant-transfer models.

Thus if it is granted that FA is a good objection to infinitism, it should be an equally good

objection to foundationalism and coherentism. If infinitism can be saved by the appeal to
skepticism, then so should foundationalism and coherentism. The kind of skepticism that
emerges from such a commitment is consistent with the possibility of a correct structural theory
of justification (or knowledge). While neither FA nor any of its variants should be granted as a
good objection, the universal commitment to skepticism is not completely moot. Though perhaps
useless as a reason to adopt infinitism, the skeptical implications of FM and FA is an additional
instance in which the peculiar result of generality appears.

5.4 Interpreting the Peculiar

Up to this point we have examined our sample of prominent arguments concerning

foundationalism, coherentism, and infinitism respectively and have found a very peculiar result –
generality. We have found that the core elements of our sampled arguments can be reformulated
into more general arguments that apply to all competing structural theories of justification. It is
now our mission to comment on what to make of such a result.

We recognize that we have excluded from our analysis – for the sake of space – several other
prominent arguments that deserve to be examined. We recognize that our data is so far from
satisfaction that it cannot hope to lend adequate support for any of the interpretations sketched
below. That is a task designated for future work. The data of such work may help provide
evidence for or against any of the possible interpretations. It may even propose new ones.
For now we must content ourselves with a few competing suggestions for what we are to make
of our peculiar findings. In the following subsections we draft five possible interpretations of our
data. We know not which is the most preferable or even whether, in the end, they constitute
substantively different interpretations. One more or more interpretations may be consistent with
one another. We leave these puzzles for the reader.

Refusal. The extent and strength of refusal comes in degrees. The weakest degrees of refusal are
the most sober, reminding us that our sample of three arguments – four if one counts the barest
allusions to BonJour’s argument from doxastic ascent – is scant evidence for conclusions of any
kind. At the lowest level of refusal, one waits for more information.

The next level of refusal takes the results as instructive about the business of defending and
criticizing any structural theory. Moderate refusal holds that the phenomenon of generality is real
but does not see it as necessarily problematic. It may posit that the peculiar result is a
consequence of a kind of confusion (whatever that confusion may be) that permeates the way
business as usual is conducted. Moderate refusal lends itself to the rejection of the three
arguments we analyzed, but it is open to the possibility that they – or other arguments – may
return with whatever confusion that originally infected them cleared up.

Refusal need not stop there. Stronger degrees of refusal represent a positive attitude about the
preservation of the literature’s assumption that there is a fact of the matter; namely that there is a
correct structural theory of justification. The height of this attitude may broach dogmatic levels.
Strong degrees of refusal maintain that there is a fact of the matter about which structural theory
of justification is the correct one even in the face of mounting contrary evidence. For those
sympathetic to strong refusal, even if it turned out to be the case that all the prominent arguments
concerning the structure of justification were infected with the peculiarity of generality, it would
at best be inductive evidence for the wrong conclusion. The proponent of strong refusal would
maintain that arguments will be adduced which do not possess the defective property of
generality.

Strong Particularism. Chisholm remarked that epistemology is confronted with two questions:
“„What do we know?‟ and „How are we to decide, in any particular case, whether we know?”
(Chisholm 1966: 56). The famous problem of the criterion is that if we have no answer for the
first question, we cannot answer the second and that if we have no answer to the second
question, we cannot answer the first (Chisholm 1966: 56). Methodists try to solve the problem of
the criterion by beginning with an answer to the second question while particularists try to solve
the problem of the criterion by beginning with an answer to the first question. Particularists
emphasize the priority of epistemological facts in theory construction. In the context of the
debate concerning the structure of justification, the facts would be all items of knowledge and the
competing methods would be the competing structural theories.

Let us just state the relevance of Chisholm’s remark. The strong particularist could maintain that
the data, i.e. all our items of knowledge, including the peculiar result of generality, are prior to
theory and that one should simply stop at – or cannot go beyond – the data. Because the data is
the totality of data supporting competitive structural theories of justification, the strong
particularist takes the feature of generality as evidence for the claim that the data does not – and
perhaps cannot – choose between the theories. Depending on the pervasiveness the feature of
generality, the strong particular posits that the data underdetermines – and may always
underdetermine – which structural theory is the correct one. The strong particularist is therefore
officially agnostic about whether there is a fact of the matter and, if there is a fact of the matter,
which structural theory is the correct one.

Strong particularism is therefore a concession to iterative skepticism of a very special kind; that
we cannot know that we know which of the structural theories is correct. But its concession does
not end there. It refrains from comment about even first order structural-theoretic knowledge. It
must so refrain precisely because it is officially agnostic about whether there is a fact of the
matter. And there must be a fact of the matter in order for there to be a theory that correctly
apprehends the structure our knowledge. Strong particularism is content with the whole of our
epistemological data without the umbrella of a structural theory.

Theoretic Relativism. The theoretic relativist contrasts the particularist in holding that there is no
fact of the matter concerning the correct structural theory. There may be various forms of
theoretic relativism and it is, as such, difficult to articulate. But theoretic relativists might
univocally hold that any structural theory can be adopted or rejected depending on one’s
background assumptions and commitments. The background need not even be epistemic.
Pragmatic concerns may require the adoption of a structure of justification in some instance and
other, incompatible structures at other times. All would be equally epistemologically legitimate.

This is why skepticism, given a certain set of epistemic and pragmatic background commitments,
the theoretic relativist might argue, can sometimes be as theoretically legitimate as it is
theoretically illegitimate given other backgrounds. In those cases it is not simply that skepticism
(or any structural doctrine) appears to legitimate in some cases and illegitimate in others, it is.
The theoretic relativist might also posit that relativism explains why no structural theory has
been adequately supported for the duration of the debate about the structure of justification;
namely because the data cannot pick out what does not exist, just as one cannot offer genuine
reasons in support for one’s tastes (in matters of taste).

Theoretic Unity. Proponents of theoretic unity hold that there is a fact of the matter and the fact
is that all structural theories of epistemic justification are correct. Prima facie, the unity thesis is
incoherent because the three structural theories are defined as contraries. The unity theorist
attempts to resolve the apparent conflict by positing that they are all correct because they are
instances of a single, more general theory. For the unity theorist, foundationalism, coherentism,
and infinitism amount to no more than different ways of expressing the same theory. If so, then
our results should not be surprising and, in fact, one should expect that all cogent arguments
concerning the structure of justification to be arguments susceptible to generalization.

The thesis of theoretic unity is not a new one. Sosa (1980; 2004) has argued for the unity of
foundationalist and coherentist theories of justification. If one finds Sosa persuasive and
infinitism as a viable option, one has adopted the unity thesis by default. Bergmann (2007) has
argued that Klein’s solution to the regress problem9 presented by the ancient trilemma, which
ultimately leads him to infinitism, isn‟t a genuine infinitism but rather a kind of foundationalism.
Turri (2010a) has made the converse argument that Klein’s solution to the regress problem can
be adopted by foundationalists – and therefore coherentists since Klein accepts Sosa’s claim that
coherentism is a type of formal foundationalism.

Consider one possible specification of the unity thesis that all structural theories of justification
are iterations of formal foundationalism.10 As we saw above, the epistemic justification
possessed by a given belief supervenes on the relevant property possessed by that belief (Sosa
1980: 229 – 231). When a basic belief is justified, its property of being propositionally justified
supervenes on property Φ. When a nonbasic belief is justified, its property of being
propositionally justified supervenes on two properties: first, the property Φ possessed by basic
beliefs and, second, the property of standing in the right kind of justificatory relation to basic
beliefs.

The unity theorist might contend that coherentism is a formal foundationalism because the
property of being propositionally justified supervenes on a very special specification of property
Φ, namely the property of either a circular arrangement or, more likely, membership in a
coherent set. In a sense all beliefs are properly basic. Similarly for infinitism, the property of
being propositionally justified supervenes on the very special specification of property Φ,
namely the property of membership in an infinitely long, nonrepeating series of reasons. If Turri
(2010a) is correct in his contention that foundationalism can accommodate Klein’s solution to
the regress argument, then it seems that there are the makings of the case for theoretic unity.

Pyrrhonianism. Ancient Pyrrhonianism was a type of skepticism that aimed at the achievement
of suspension of judgment for any given matter of dispute by demonstrating the equipollence of
two opposing accounts, theories, etc. (Sextus 2000: 4). Equipollence is a kind of evidentiary or
argumentative deadlock, one in which the available data, reasons, justification, etc.,
underdetermines the competing accounts such that one cannot choose between competitors. Thus
a Pyrrhonian suggests that the exhibition of generality in our sampled arguments is but a small
data point in a well-documented history of philosophy.

The Pyrrhonian will hold that generality is likely a pervasive feature that infects all arguments
such that for any argument concerning any structural theory, it can be generalized so as to
neutralize grounds to support or reject the theory in question, thus moving us to equipollence.
One possible explanation for this result, a Pyrrhonian might merely suggest with the faintest
shrug, is because skepticism is true. Perhaps the Pyrrhonian does not advocate the thesis in
earnest, but that is another matter entirely.

If generality is extremely pervasive, one might find it good evidence for skepticism. But this
leads to a paradox – perhaps even the implosion of skepticism. This is the argument of Cruz
(2010). But Pyrrhonians are unconcerned with paradox or implosion. Sextus Empiricus, for
example, seems perfectly cognizant that skepticism implodes when he writes in his Outlines of
Skepticism, “’In no way more’ says that it too, along with everything else, is no more so than not
so, and hence it cancels itself along with everything else” (2000: 6). Later he notes that “skeptics
will perhaps be found to determine nothing – not even „I determine nothing‟ itself” (2000: 49).
Sextus remarks that this is a characteristic trait of all the skeptic’s arguments because

you should understand that we do not affirm definitely that they are true – after all, we
say that they can be destroyed by themselves, being cancelled with what they are applied
to, just as purgative drugs do not merely drain the humors from the body but drive
themselves out too along with the humors (2000: 52).

Pyrrhonianism is skepticism par excellence. It is the skepticism of Agrippa, the skepticism of the
perennial challenge to the possibility of a structural theory of knowledge. If one opts for
Pyrrhonianism, one will have gone full circle: begin with the ancient trilemma; grapple with each
horn of the trilemma and its consequences; consider foundationalism, coherentism, and
infinitism; stumble upon argumentative generality; end with equipollence. Perhaps the ancient
trilemma is little more than a clever trap that is no more than a distraction, allowing the
Pyrrhonians to corner their prey before the fatal strike. If so, then the Pyrrhonian might tout the
emergence of generality as the death knell for foundationalism, coherentism, and infinitism.

1 Roughly, y is a defeater for S‟s the justification of x at t, “if and only if, when y is conjoined
with whatever justified x for S, the resulting conjunction no longer justifies x for S” (Klein 1981:
24). Overriders are basically a complicated species of defeaters. For more detail, see Klein
(1981).

2 See Turri (2010a).

3 The production of justifiers does not require that S have an epistemological theory nor does it
preclude the Wittgensteinian analysis of knowledge. If one considers, for example, the
production of justifiers for the belief <I have a hand>, the relevant cognitive disposition might
just be that S sees that S has a hand. Nothing about the definition requires that S be able to say
more than that. The question about which of the cognitive dispositions are epistemically relevant
and why is up to the theorist (within reason, of course). We have no assumptions about the
dispositions in this essay.

4 If r provides its own justificatory basis on the warrant-transfer view, then r – to the extent that
it possesses any warrant at all – is autonomously warranted. It would seem that r would be more
like a foundational belief than a candidate for coherentism. For more on this point, see Audi
(1993).

5 For discussion of the Neurath metaphor, see Neurath (1959), Sosa (1980), Quine (1980), and
Devitt (2005).

6 BonJour (1985) offers five desiderata for a theory of coherence. In section 3.3, we saw the first
two desiderata. This is the fifth.

7 I am suggesting that it is possible that the intuitive support for FM can be Gettierized and that
this might be the explanation for the appearance of (2)‟s justified status. Since whether (2) can
be Gettierized does not affect my argument, I do not explore this possibility.
8 Klein actually endorses warrant-emergence infinitism on the basis of this difficulty (Klein
2005: 134 – 135).

9 Cf. Klein (2005) and Klein (2007).

10 It should be made clear that unity theorists are not committed to such a specification. They
might hold that all theories are as much iterations of formal foundationalism as they are “formal
coherentism” or “formal infinitism”.

Logique & Analyse 216 (2011), 589.597

FILLING A TYPICAL GAP IN A REGRESS ARGUMENT
JAN WILLEM WIELAND
Abstract

In the following we fix a typical regress argument, locate a typical gap in the argument, and try
to supply a number of gap-filling readings of its first premise.

1. Locating The Gap

As Armstrong said about someone who gets stuck in an infinite regress:

“He is like a man who presses down the bulge in a carpet only to have it reappear
elsewhere.” (1978: 21)

I take it that Armstrong appeals to a typical regress argument:

(1) There is a problem P.

(2) There is a theory S of how to solve P.
(3) S generates a regress R consisting of problem/solution pairs similar to P/S.1
(4) R is unacceptable.
(5) Therefore: S is to be rejected.

Filling out the details of Armstrong's example:

(1) There is a bulge in the carpet, and we want to get rid of it.
(2) This problem can be solved by pressing it down.
(3) If we press down the bulge in the carpet, another bulge appears elsewhere in the carpet,
which in turn can be eliminated by pressing it down, so that yet another bulge appears in the
carpet, etc.
(4) The regress is unacceptable.
(5) Therefore: the solution of pressing down the bulge is to be rejected.

Impressive as it seems, the regress argument can be resisted simply by denying (4). Yes, there is
a regress, but so what? We need a justification of why the Carpet Regress (say) is bad. In
particular the question is: why does the regress in which we keep on pressing down bulges lead
to the conclusion that we cannot eliminate the bulge by pressing it down, and that we have to
search for another solution for our problem (e.g. buying a new, smaller carpet)? Put in general:
how to get from regress R to the rejection of theory S? Hence, there is a gap in the argument, and
it must be filled.

Two disclaimers. First, there are no doubt more gaps in the argument, but they will be left for
future research. Second, examples of regress arguments can easily be multiplied, in both
philosophical and everyday contexts. As my aims are metaphilosophical, we shall stick to the
carpet case in the following.

2. Filling It

There are a number of familiar ways to get from a regress R to the rejection of theory S. For
instance, if R is to be plainly unbelievable or absurd (the classic way), then S which generates R
is committed to an absurdity, and to be rejected for this very reason. Likewise, if it can be shown
that R is logically impossible, uneconomical, or similarly disadvantageous for some principled
reason, then S can be rejected on that basis.

In the following we set such straightforward gap-filling strategies aside, and assume that the
unacceptability (or viciousness, as it is often said) of regresses has somehow to do with failures
of problem solving. More specifically, to get from R to the rejection of S, we shall take it, is to
show how R prevents S from solving a relevant problem P. The unacceptability of R can be
defined accordingly:2

R is unacceptable (for S with respect to P)$R prevents S from solving P

By this, the Carpet Regress is bad for our theory (i.e. that the bulge can be eliminated by pressing
it down) just in case it prevents our theory from solving the problem of the bulge in the carpet.
By this definition, regresses are not unacceptable full-stop; they can only be unacceptable
relative to a certain relevant problem P. Thus without P there can be no unacceptable regress
(and, indeed, no motivation to start the regress in the first place).

In order to see how a regressive solution S fails to meet P (if so), we should get clear on what P
actually consists in. In the literature this usually remains unclear; at best one may find hints (cf.
some of the upcoming footnotes). In the following we show that the problem solving task is
rather ambiguous, and distinguish several readings of it. As it turns out, not all of them guarantee
that the Carpet Regress is unacceptable, and hence only some of them are able to fill the gap in
the regress argument.

Consider the first, most obvious, reading on the problem solving task:

(P1) Get rid of that particular bulge in the carpet.

Now run the argument on the basis of P1. There is one particular bulge in the carpet. Call it b1.
b1 can be eliminated by pressing it down. If we press down b1, b1 is completely gone. Another
bulge b2, distinct from b1 (assuming some intuitive identity conditions of bulges), however
appears elsewhere in the carpet, which in turn can be pressed down, and so on. In this case, the
regress is irrelevant and does not lead to the conclusion that the problem cannot be solved by
pressing down the bulge, for the problem solving task is already met. It is true that there is
another bulge, but so what? b1 is eliminated, and this was all we had to do.

But one might think P1 is too weak. If we want to get rid of the bulge in the carpet, do we really
want to eliminate one particular bulge, as P1 says? What we really want, presumably, is to get rid
of any bulge in the carpet whatever. So here is a revised take on the problem solving task:

(P2) Get rid of all the bulges in the carpet.

This task is clearly distinct from P1: it is stronger, i.e. harder to be met. Run the argument again.
There is a bulge in the carpet, b1. As the theory as stated in line (2) goes, b1 can be eliminated by
pressing it down. If we press down b1, another bulge b2 appears elsewhere in the carpet. P2 says
that we have to eliminate all bulges, so b2 is to be eliminated too. As the theory goes, b2 can be
eliminated by pressing it down. If we press down b2, another bulge b3 appears elsewhere in the
carpet. P2 says that we have to eliminate all bulges, so b3 is to be eliminated too. And so on: it's
bulges all the way down. In this case, the regress seems relevant concerning the issue whether
the solution at issue is a good one, or not, for the problem solving task is not already met by
managing the first bulge. So the question is: does the regress lead to the conclusion that our
solution of pressing down the bulge is a bad one? Surprisingly, perhaps, the answer is negative.
The problem solving task P2 motivates the regress, rather than conflicts with it. Time and again
P2 says that we are to eliminate the new bulge, and time and again the theory says we are to
press it down.3

Hence we need a still stronger reading of the problem solving task to obtain a conflict with the
theory. Consider the following:
(P3) Get rid of all the bulges in the carpet, and complete this task.

P3 is stronger than P2: it puts a constraint on how to solve the problem. In case of P2 we are
allowed to solve the problem in whatever way we like, specifically: by completing it, or by never
completing it. But in case of P3 it is not allowed to eliminate bulges by never completing the
task. This conflicts with the regress. At no point of the regress it is possible to complete the task
of eliminating all bulges in the carpet.4 This is not just because we are too slow, or too tired after
a couple of months of working on the problem, nor because we simply die at a certain point.
Rather, it is structurally impossible to do it: if we have solved a problem, there is always yet
another problem to be solved. We can never say that we are ready with pressing down bulges,
and that the job is done.5 Even immortality would not help us here. Hence, at no point we can
meet the task P3, which warrants the conclusion that our solution is a bad one.

This may fill the gap in the regress argument, but there is at least a second possibility. To wit:

(P4) Get rid of all the bulges in the carpet at once.

P4 too is stronger than P2. In case of P2 we are allowed to eliminate bulges in carpets in
whatever way we like, specifically: successively, or all at once. But in case of P4 it is not
allowed to eliminate bulges successively. This conflicts with the regress. At no point of the
regress it is possible to eliminate all bulges in the carpet at once. This, again, is not because we
are too restricted in our capacities to fulfil the task. It is not that if there are, say, 16 bulges in an
enormous carpet, that we lack hands to press all down at the same time. It is structurally
impossible to do it: if we have eliminated bn, there is always yet another bulge bn+1 which is to
be eliminated (and the other way around: bn+1 does not exist unless bn has been pressed down).
This means that at no point our solution of eliminating bulges by pressing them down can meet
the task P4, which supports the conclusion that our solution is a bad one.

In sum, both P3 and P4 can fill the gap in the regress argument. Specifically, they provide the
following two gap-filling modifications of premise (1):

(1*) There is a bulge in the carpet, we want to get rid of it, we want to get rid of any other bulge
in the carpet, and we want to complete this task.

(1**) There is a bulge in the carpet, and we want to get rid of all the bulges in the carpet at once.

Both modifications warrant the conclusion that the problem of the bulges (as specified in the first
premise) cannot be met by pressing them down. Furthermore, the Carpet Regress (line 3) is
unacceptable (line 4) exactly because it prevents our theory (line 2) from meeting the relevant
problems (lines 1* and 1**).

So far so good.

3. Emptying It Again
In the previous section we singled out two strategies which succeed in filling the gap in the
typical regress argument. In this section I nevertheless suggest that they are unmotivated, so that
they are of no help to the typical regress argument.

Consider the generalizations of the problem solving tasks P1.4:

(P1*) Solve problem X of type Y.

(P2*) Solve all problems of type Y.
(P3*) Solve all problems of type Y, and complete this task.
(P4*) Solve all problems of type Y at once.

As we have seen at their instances, P2* is a strengthening of P1*, and both P3* and P4* are
strengthenings of P2* (and hence of P1*). In other words: it is easier to meet P1* than to meet
P2*, and it is easier to meet P2* than to meet P3* or P4*. Furthermore, P1* and P2* are too
weak to fill the gap in the typical regress argument, but P3* and P4* are powerful enough for the
job.

The worry about the latter two tasks naturally stems from this. The point is: surely it is possible
to strengthen the problem solving tasks in such a way that they cannot be met. But why accept
the strengthenings? Why accept that P3* and P4* rather than P1* or P2* are to be met? Why, in
all generality, are we to complete the task of solving all problems of the same type at some point,
or are we to solve all of them at once? Or again: where do these constraints on problem solving
acquire their normative force?

True, in the carpet case one could suggest that the problem solving needs completion (and hence
that P3 is justified), because we obtain a fiat carpet in that case only. Yet presumably this
motivation cannot be generalized. As Klein says on the Justification Regress:

“The infinitist cannot accept the Completion Requirement because it is clearly

incompatible with infinitism. Justifications are never finished.” (1999: 314)6

In brief, if P3* and P4* are unmotivated, they can be resisted.

4. Filling It Again?

At this point there is still a gap in typical regress arguments, and this consequently undermines
them. Is there something more promising available?

Consider the following, fifth take on the problem solving task:

(P5) Get rid of the bulge in the carpet, and do not shift this problem.

P5 is rather different from the other tasks P2, P3 and P4, because it is, like P1, about one
particular bulge only. But unlike P1, it puts a constraint on how to deal with the problem, i.e. that
we should better not shift it. Generally the constraint is this: do not solve a problem P by a
solution S where S gives rise to a problem which is similar to P (there are several ways in which
solutions can `give rise' to similar problems, but we shall not go into that here). Our theory
violates this constraint P5, because it does give rise to a problem which is similar to the initial
problem. Pressing down a bulge makes it the case that another bulge appears elsewhere in the
carpet, and hence the problem of getting rid of the first bulge has shifted to the problem of
getting rid of the second bulge.

P5 involves the following modification of premise (1):

(1***)There is a bulge in the carpet, and we want to get rid of it in such a way that the problem
does not get shifted.

This too will fill the gap in the regress arguments, and is what Armstrong must have had in mind
when he invented the Carpet Regress:

“What the [Z] regresses bring out is that the [Z-theorist] does not in fact solve his
problem, he simply shifts it.” (1978: 21)7

P5 is however not disconnected from the other tasks. P5 relates to P4 and P3 in an interesting
way. First, a violation of P5 entails a violation of P4, moreover: it is because time and again our
theory shifts the problem of getting rid of a particular bulge that we cannot eliminate all bulges
in the carpet non-successively (i.e. meet P4). Second, it is because our theory shifts the problem
of getting rid of one particular bulge, and keeps on doing this, that we can never complete the
task of getting rid of any bulge in the carpet whatever (i.e. meet P3). Here, a violation of P5
entails a violation of P3; however only under the condition that the problem shifting does not
terminate at some point. This condition can be guaranteed by adopting yet another task:

(P6) Get rid of all the bulges in the carpet in the same way.

P6 is an instance of the general principle of uniformity according to which similar problems

require similar solutions. It establishes that we keep on pressing down bulges, and refrain from
trying out other solutions at some point (e.g. cutting out a piece of the carpet, buying a smaller
carpet after all, etc.). Furthermore, as long as we keep on pressing down bulges, the problem
shifting will not terminate. So by P6 a violation of P5 smoothly entails a violation of P3.

All this can be generalized. Recall the generalizations of P3 and P4, and let us add the
generalizations of P5 and P6:

(P3*) Solve all problems of type Y, and complete this task.

(P4*) Solve all problems of type Y at once.
(P5*) Solve, but do not shift, problem X of type Y.
(P6*) Solve all problems of type Y in the same way.

How about their interconnections? As said, P3* and P4* are strengthenings of P2*. P6* also is a
strengthening of P2*. P5* is however a strengthening of P1*. Furthermore, a violation of P5*
entails a violation of P4*, and a violation of P5* coupled with P6* entails a violation of P3*. The
entailments in the opposite direction however fail: (i) a violation of P5* is not entailed by a
violation of P4*, and (ii) a violation of P5* is not entailed by a violation of P3* (with or without
supplementation of P6*).

Here are the counterexamples. First, consider a special carpet which is such that (for no special
reason) each five seconds a bulge pops up in it. It is impossible here to press all down at the
same time and to solve all problems of the same type at once, but this has nothing to do with
problem shifting. This example blocks the entailment mentioned at (i). Second, consider an in-
finite carpet with an infinity of bulges. It is impossible here to press all down and finish the
problem solving at some point, but this has, again, nothing to do with problem shifting. This
example blocks the entailment mentioned at (ii).

This is important. That a violation of P5* (coupled with P6*) entails but is not entailed by a
violation of P3* or a violation of P4* means that P5* can be taken as more basic than P3* and
P4*. I conjecture that violating P5* captures the essence or distinguishing characteristic of a
regress (and the role of P6* is nothing but guaranteeing that P5* will be violated to eternity).8

The question follows: can P5* and the corresponding premise (1***), then, fill the gap in the
typical regress argument? It still depends. (1***) is justified only if it is bad to violate P5*, i.e. to
commit to problem shifting. Perhaps in some cases it is, and in others not. But then we need a
demarcation criterion. To my knowledge this has nowhere been put forward in the literature, but
the results of the present section suggest that if the typical regress argument is to be saved, it is to
be saved here.9

Be that as it may, for now we may still, among other things, eliminate bulges in carpets by
pressing them down.10

*Thanks to Jo Van Cauter, Anna-Sofia Maurin, Benjamin Schnieder, Maarten Van Dyck, Erik
Weber, and audiences in Tilburg, Ghent and Manchester. I am PhD fellow of the Research
Foundation Flanders.

1 This problem/solution pattern of regresses derives from Schlesinger (1983: 221). Cf. Gratton
(1994) for discussion.

2 Or weaker (leaving open the possibility that there are other sufficient conditions under which
regresses can turn out unacceptable): R prevents S from solving P->R is unacceptable (for S with
respect to P).

3 Cf. “There is also an obvious way of looking at matters optimistically, asserting that when the
question is raised why a certain evil is permitted, we can at once come up with a solution.”
(Schlesinger 1983: 225)

4 Cf. “Thus, if the regress is vicious, it is vicious because it prevents Resemblance Nominalism
from accomplishing its explanatory project of accounting for all properties in terms of
resembling particulars: such a project remains forever incomplete.” (Rodriguez-Pereyra
2002: 108)
5 At least, if it is denied that the notion of completed/actual infinities is applicable here.

6 For an important qualification of this, see Peijnenburg (2007).

7 This sentence precedes, without further clarification, the carpet sentence cited at the very
beginning of this paper.

8 Cf. “It is the first step in the regress that counts, [. . . ] that if there was any difficulty in the
original situation, it breaks out in exactly the same form in the alleged explanation.” (Passmore
1961: 31)

9 This and other queries raised in this paper (which was written in 2009) are extensively
discussed in my dissertation entitled `And so on. Two theories of regress arguments in
philosophy.'

10 On Sundays, of course, as Philipp Keller suggested.

Externalism/Internalism
Hamid Vahid
Routledge Companion to Epistemology

A) A Sellarsian Dilemma

… Let us recall that according to SAI an agent’s belief p is justified only if he justifiably believes
that (i) the ground (e) of his belief obtains, and that (ii) e adequately supports p. However, as
Alston (1986) and others have noted, the second requirement of SAI engenders an infinite
regress. For if in order to be justified in believing that p, I must be justified in believing that my
evidence e adequately supports p, the justification of this latter (higher-order) belief would, in
turn, require that it is based on further evidence e1 and that I justifiably believe that that e1
adequately supports my belief that e adequately supports p and so on ad infinitum. The ensuing
regress not only involves an infinite number of beliefs, but an infinite number of beliefs of ever-
increasing complexity (Fumerton 1995: 89).
Foundationalism
Michael DePaul
Routledge Companion to Epistemology

… Why don’t foundationalists think the chain can go on forever? We are finite beings, so we
clearly do not have infinitely many beliefs. Hence, we simply could not go on indefinitely
justifying beliefs with different beliefs we already hold. Therefore the chain of reasons cannot be
infinitely long.
INFINITISM
Peter D. Klein
Routledge Companion to Epistemology

Introduction

Infinitism, along with coherentism and foundationalism, is a view about the structure of reasons
and reasoning that is designed to provide a solution to the epistemic regress problem. The regress
problem can be put this way: Suppose we give a reason, r1, for holding one of our beliefs, b.
Then, we are asked for our reason for holding r1, and we provide the reason, r2. Then, we are
asked for our reason for r2, and we give r3. Now, either this process could go on indefinitely,
which seems to suggest that nothing has been gained by providing a reason because there is
always another one needed; or, if some reason repeats, it seems that we have argued in a circle
and that no such argument could provide a good basis for accepting b; or, if at some point there
is no further reason, it seems that the stopping point is arbitrarily held because there is no
reasonable basis for holding it. The problem is that, contrary to strong pre-theoretical intuitions,
there seems to be no point in giving reasons for our beliefs.

Infinitism holds that there is no reason that can be given for any belief which is so privileged that
it is immune to further interrogation. There are circumstances in which even the most
commonplace reasons require further reasons. Even so, knowledge based upon such reasoning is
possible, and giving reasons does increase the warrant for our beliefs.

The primary purpose of this chapter is to sketch the case for infinitism. It has three main steps.

First, I will discuss the way in which the regress problem was originally conceived by the
Pyrrhonians and Aristotelians. The upshot will be that given two presuppositions that underlie
the regress problem as originally conceived, the Pyrrhonian response, namely that reasoning is
unable to resolve disputes, is highly plausible.

Second, I will discuss three challenges to the Pyrrhonian response. The first challenge arises
from various forms of foundationalism including what I call ‘austere reliabilism’ and
‘embellished reliabilism.’ I will argue that these forms of foundationalism fail to adequately
address the normative basis motivating the regress argument. The second challenge originates
with contemporary coherentism. I will argue that contemporary coherentism is not a viable
response because it is subject to the same objections that apply to foundationalism. That leaves
infinitism, the third challenge, as the only viable, non-skeptical response.

Third, I will sketch infinitism, point to some of its advantages, and try to show that the primary
objections to it miss the mark.

1. The Traditional Problem

The traditional regress problem was known to Aristotle, who wrote this in the Metaphysics:

There are . . . some who raise a difficulty by asking, who is to be the judge of the healthy
man, and in general who is likely to judge rightly on each class of questions. But such
inquiries are like puzzling over the question whether we are now asleep or awake. And all
such questions have the same meaning. These people demand that a reason shall be given
for everything; for they seek a starting point, and they seek to get this by demonstration,
while it is obvious from their actions that they have no such conviction. But their mistake
is what we have stated it to be; they seek a reason for things for which no reason can be
given; for the starting point of demonstration is not demonstration. (Aristotle 1941b:
1011a2–14)

Even though Aristotle is speaking about “demonstration,” and there is a special meaning that he
would sometimes attach to that concept involving syllogistic reasoning from intuited first
principles, his point here is that reasoning, in general, reaches an end because there are some
privileged starting points “for which no reason can be given” because “the starting point of
demonstration is not demonstration.” No reason can be given because reasoning presupposes
something not inferred—namely the premises that provide the basis for the reasoning.

This argument still motivates foundationalism. Here is a redacted paragraph from William
Alston’s Epistemic Justification that faithfully renders his general point:
 The argument [for foundationalism] is that the original belief [the one that requires
justification] will be mediately justified only if every branch [of the justificatory tree] . . .
terminates in an immediately justified belief. Positively, it is argued that on this condition
the necessary conditions for the original belief’s being mediately justified are satisfied,
and negatively it is argued that if any branch assumes any other form, they are not.
(Alston 1989: 54)

Alston goes on to say that this argument “gives stronger support to foundationalism than any
other regress argument” (Alston 1989: 55).

The foundationalists’ response is an answer to the skeptics’ use of the regress argument whose
classical formulation is due to Sextus Empiricus:

The later Skeptics hand down Five Modes leading to suspension, namely these: the first
based on discrepancy, the second on the regress ad infinitum, the third on relativity, the
fourth on hypothesis, the fifth on circular reasoning. That based on discrepancy leads us
to find that with regard to the object presented there has arisen both amongst ordinary
people and amongst the philosophers an interminable conflict because of which we are
unable either to choose a thing or reject it, and so fall back on suspension. The Mode
based upon regress ad infinitum is that whereby we assert that the thing adduced as a
proof of the matter proposed needs a further proof, and this again another, and so on ad
infinitum, so that the consequence is suspension [of assent], as we possess no starting-
point for our argument. The Mode based upon relativity . . . is that whereby the object has
such or such an appearance in relation to the subject judging and to the concomitant
percepts, but as to its real nature we suspend judgment. We have the Mode based upon
hypothesis when the Dogmatists, being forced to recede ad infinitum, take as their
starting-point something which they do not establish but claim to assume as granted
simply and without demonstration. The Mode of circular reasoning is the form used when
the proof itself which ought to establish the matter of inquiry requires confirmation
derived from the matter; in this case, being unable to assume either in order to establish
the other, we suspend judgement about both. (Empiricus 1976: I: 166–69)

There are five modes mentioned in the passage from Sextus Empiricus. The modes of relativity
and discrepancy are crucial to understanding the reductio put forth by Sextus because those
modes are designed to show that neither a judgment based on how things appear nor a judgment
based upon what we collectively believe (either qua “philosophers” or qua “ordinary” persons) is
so privileged that it does not need to be supported by further reasoning. As we will see,
considerations similar to those motivating the modes of relativity and discrepancy form part of
the motivation for infinitism.

The foundationalists’ answer to the skeptical conclusion is that there must be some beliefs that
cannot be justified by further reasoning because, as they see it, reasoning cannot create epistemic
warrant, so warrant must be present in some basic beliefs. From the foundationalists’ perspective,
the problem is typically not whether there is sufficient warrant for knowledge, it is, rather, how
sufficient warrant arises and is transferred.
This is clear, for example, from Aristotle’s rather dismissive attitude towards skepticism
manifested in the citation above, and even in the carefully constructed answer in the Posterior
Analytics designed to show that if some knowledge is the result of demonstration, then some
knowledge must not be the result of demonstration. There he argues that either the series of
demonstrations is finite or infinite. It must be finite because “one cannot traverse an infinite
series” (Aristotle 1941a: 72b10). But if it terminates, it cannot terminate in another belief that
requires a demonstration because the conclusion would not be “properly” known and “rests on
the mere supposition that the premisses are true” (Aristotle 1941a: 72b14). It cannot be finite and
circular because the premisses in a demonstration must be “prior to and better known than the
conclusion” and “the same things cannot be simultaneously both prior and posterior to one
another” (Aristotle 1941a: 72b25–28). Thus, if there is demonstrative knowledge, then there
must be non-demonstrative knowledge.

Near the end of the Posterior Analytics Aristotle does provide a sketch of how such non-
demonstrative knowledge reliably originates with sensation and ends with rational insight. The
details of Aristotle’s proto-reliabilist sketch are not important at this point, although I will return
to it and a general discussion of reliabilism in section 2. What is important here is to understand
how the foundationalists use the regress argument.

The Regress Argument as Used by Foundationalism

1. Reasoning has only three possible structures: it is finite and has a beginning point, it is
circular, or it is infinite.

2. Circular reasoning is not acceptable because a belief would have to be epistemically prior to
itself.

3. Reasoning infinite in length could not be carried out by humans.

4. Thus, if there is knowledge that results from reasoning, the reasoning must be finite in length.

5. The beginning points of the reasoning must be known (otherwise it would be mere
supposition).

6. Thus, if there is reasoning that results in knowledge, there must be some beliefs (the beginning
points) that are known by some process other than reasoning.

The conclusion is the basic claim made by the foundationalist, namely, if there is some
knowledge that is the result of reasoning, some knowledge is not the result of reasoning. Note the
hypothetical nature of the conclusion. As mentioned above, although almost all foundationalists
eschew skepticism, a foundationalist need not hold that there is knowledge in any specific area,
or even in general. There can be and have been skeptical foundationalists: Hume, for example.
There can be non-skeptical foundationalists: Locke, and of course Aristotle, for example.

I think it is fair to say that there are two core presuppositions underlying the regress argument as
put forth by foundationalists without which the argument could not succeed:
 Non-Originating Principle: Reasoning, alone, cannot produce epistemic warrant.

Inheritance Principle: Reasoning can transmit the requisite epistemic warrant for
knowledge from other beliefs.

For the sake of the discussion, the Pyrrhonians can accept the hypothetical in step 6 (above) as
well as the two principles, but they would invoke the modes of relativity and discrepancy in
order to show that there are no legitimate firm beginning points. Aristotle might be right that in
practice we do not push for reasons beyond those that are taken for granted by all of the
participants in a discussion, but skeptics would argue that such contextually based agreements do
not indicate the presence of a belief that has the requisite epistemic warrant because at other
times and in other circumstances, different agreements are, or can be, made. In addition, skeptics
would point out that beliefs based upon perception are person and circumstance relative. That’s
not to say that reasons for holding such beliefs can’t be located; rather, it is to say that they are
not privileged in the way required by foundationalism.

The skeptics would point out that the inheritance and non-originating principles are telling
against infinitism and coherentism because if reasoning cannot originate epistemic warrant, then
neither view can explain how warrant arises in the first place. Each belief in the potentially
infinite reasoning process is warranted on the condition that the previous belief is warranted, but
that previous belief is warranted only if the previous one is, etc. So, how does warrant originate?
(see Dancy 1985: 55). Similarly, even if the beliefs in a set of coherent beliefs are mutually
warranting—each increasing the warrant of the other—the question of how the beliefs obtain
warrant to begin with still remains.

The upshot, from the Pyrrhonian point of view, is withholding beliefs. To them, what initially
looked like a good argument for foundationalism, when examined more care fully, actually
provides a basis for a skeptical attitude towards beliefs because the origin of warrant remains
mysterious.

2. Responses to the Skeptics’ Use of the Regress Argument

Aristotle was not content with his response to skepticism quoted in the previous section
(Aristotle 1941b: 1011a2–14). As mentioned earlier, in the Posterior Analytics he provides the
sketch of another type of response, namely, one designed to provide a basis for explaining the
origin of warrant. Here is a somewhat redacted and interpolated quotation that remains true to the
basic Aristotelian view. (I have indicated exact quotes with double quotation marks):

In order for us to acquire the basic beliefs “we must possess a capacity of some sort”
which is “a characteristic of all animals, for they all possess a congenital discriminative
capacity which is called sense perception. But though sense perception is innate in all
animals, in some the sense-impression comes to persist, in others, it does not.” In those
animals in which sense perception persists, there “comes to be what we call memory, and
out of frequently repeated memories of the same thing develop experience . . . [and] from
 experience . . . originate the skill of the craftsman and the knowledge of the man of
science”. (Aristotle 1941a: 99b33–100a8)

The essence of this form of foundationalism is what I call ‘austere reliabilism’ with regard to
basic beliefs that acquire their warrant simply in virtue of having the right kind of causal history.
What makes this form of reliabilism “austere” is that although reasoning can produce new
knowledge, reasoning neither creates new types of epistemic warrant nor augments the amount
of warrant, it merely transmits the warrant inherent in basic beliefs. (See Goldman 1979 for a
contemporary form of austere reliabilism.)

‘Embellished reliabilism’ does not adhere strictly to the two principles mentioned in the previous
section because it allows that reasoning can produce either a new type of epistemic warrant or
augment the amount of epistemic warrant inherent in basic beliefs. Nevertheless, embellished
reliabilism, like austere reliabilism, holds that some beliefs have a type of epistemic warrant that
obtains because of the way in which such “basic” beliefs arise. But once the basic beliefs, or
those inferred from them, become members of a set of beliefs that have been subjected to careful
self-reflection—including reflection about the reliability of our (or, in a Cartesian mode, my)
epistemic capacities—a different type of (or at least more) warrant can arise. Here is a passage
from Ernest Sosa that makes that very point:

Admittedly, there is a sense in which even a supermarket door “knows” when someone
approaches, and in which a heating system “knows” when the temperature in a room rises
above a certain setting. Such is “servo-mechanic” knowledge. And there is an immense
variety of animal knowledge, instinctive or learned, which facilitates survival and
flourishing in an astonishingly rich diversity of modes and environments. Human
knowledge is on a higher plane of sophistication, however, precisely because of its
enhanced coherence and comprehensiveness and its capacity to satisfy self-reflective
curiosity. Pure reliabilism is questionable as an adequate epistemology for such
knowledge. (Sosa 1991: 95)

It is not my purpose here to examine either austere (“pure”) or embellished reliabilism in detail
and I grant that this taxonomy might be difficult to apply in some cases. Nevertheless, it should
be clear that although the embellished form of reliabilism does recognize the normative
imperative to provide reasons for some of our beliefs, both forms fail to fully recognize the
fundamental intuition informing the regress—namely, that any belief for which one can produce
reasons is better or differently warranted than a belief for which one cannot produce reasons.

If good reasoning cannot be circular, and if being able to provide reasons for our beliefs is
importantly epistemically better than not being able to do so, then infinitism is the only solution
to the regress argument—other than skepticism. To see that, take any proposed “basic” belief in
the regress. Call it “E.” One can ask the following question: In virtue of what is E a proper
ending point? If no answer is forthcoming, then it clearly appears arbitrary to believe E without a
reason because up to that point reasons were needed. Why should the regress end at E rather than
at some earlier step or at some possible later step?
Suppose that the answer is that E is the appropriate ending belief in virtue of E’s having some
foundational property, F. Then, the next question becomes obvious: Does E’s possessing F make
it more likely that E is true than it would be if E did not possess F?

The imperative to produce an answer strikes me as obvious. Consider what I have called
elsewhere a “Wednesday Foundationalist” who holds that a belief formed by any person on
Wednesday has the austere form of warrant (Klein 2007a: 15). No one is such a foundationalist
because there is absolutely no reason to believe that Wednesday beliefs are any better than, say,
Friday- or Sunday-beliefs. What foundationalists typically put forth as the F-property in virtue of
which E beliefs are foundational is such that E’s possessing F readily provides a basis for
believing that E is likely to be true.

Once the question is asked about whether E’s possessing F is truth conducive, there are four
possible responses: It can be ignored, or “yes,” “no,” or “I don’t know.” I take it that ignoring the
question is to fail to grasp the normative imperative underlying the regress argument, and the
“no” and the “I don’t know” answers place S’s acceptance of E in jeopardy. Once the question is
asked and understood, the only answer that at least preserves all of E’s warrant is “yes.” But,
then, a reason for believing E has been given and the regress has continued.

Let me note in passing that this argument against foundationalism, if sound, works against the
current forms of emergent coherentism as well—and emergent coherentism strikes me as the
only plausible form. The other form—what I call transference coherentism—was probably never
held since it takes individual beliefs to be the primary bearers of warrant and leads to circular
reasoning. That logically possible but completely unsatisfying view was well disposed of by
Aristotle and the Pyrrhonians.

Emergent coherentism is best exemplified by BonJour (BonJour 1985: 87–110). In this view, it is
sets of beliefs that are the primary bearers of warrant. All beliefs in the appropriate type of
coherent set are warranted simply in virtue of being members of that set. Thus, warrant is not
transferred from one belief to another—rather, warrant emerges as a result of the mutual support
provided by the beliefs in the set.

As Ernest Sosa has pointed out, this form of coherentism shares a formal structure with
foundationalism (Sosa 1980). Using the terminology I am employing, the emergent coherentist
takes the foundational property F to be E’s being a member of a set of beliefs that is coherent
(and perhaps has other features as well). In other words, emergent coherentism can be seen as
one-step foundationalism because all beliefs are foundational. (Perhaps some are relatively
“more foundational” than others because they are more important to the coherence of the set. But
they all gain their initial warrant because they are members of the appropriate type of set.) But
once the foundational property, F, is identified as “being a member of a set of coherent beliefs,”
the question arises about whether E’s being a member of such a set is truth-conducive. Without a
positive answer to that question, acceptance of the coherent set seems arbitrary. The regress has
continued.

Now, it could be objected (1) that this very general argument against foundationalism (and
emergent coherentism) conflates an important distinction between a belief itself being justified
with the meta-belief that the belief is justified and (2) that knowledge only requires that the belief
be justified (see Alston 1976).

To assess the force of the objection, it is important to distinguish two senses of belief and the
concomitant two senses in which a belief is justified. In one sense, “belief” refers to the
propositional content of a belief as in “that belief is true” or “her belief was implied by what she
said earlier.” In the other sense, “belief” can refer to the belief-state as in “she had that belief for
many years” or “her belief was caused by a reliable process.” The concomitant distinction
regarding “justified belief” is between the proposition being justified for someone, that is,
propositional justification, and the believing (i.e., the state of believing) being justified, that is,
doxastic justification (see Firth 1978).

The objection mentioned above would be valid only with regard to propositional justification.
There is a clear distinction between a proposition, say p, being justified and the meta-proposition
‘p is justified’ being justified. Any argument that conflated the distinction is built upon a pun. I
grant that in order for p to be justified for a person, it is not required that ‘p is justified’ is
justified for that person.

Nevertheless, it is crucial to note that what is required for knowledge is that S’s believing that p
be justified (even if the believing is only a dispositional state). For even if p is true, believed, and
propositionally justified for S, S could fail to know that p because either S believed p for the
wrong reasons or no reasons whatsoever (as in a guess). The regress argument and any possible
responses are concerned with whether the belief that p is doxastically justified sufficiently for the
belief to rise to the level of knowledge.

Once the question is raised concerning whether E’s possessing F makes it more likely that E is
true than it would be if E did not possess F, it is S’s entitlement to continue to believe that p that
is being questioned. If S is not able to defend the “yes” answer to the question, some adjustment
of S’s entitlement to believe E and every belief that depends upon E is called for. It might not be
required that S give up E because E (as opposed to any of the contraries of E) might possess the
kind or amount of epistemic warrant that austere reliabilism would attribute to it, but those views
that recognize the importance of having reasons for our beliefs when their epistemic credentials
are challenged (i.e., embellished reliabilism, coherentism, and infinitism) would require some
recalibration of S’s entitlement to believe that p.

In other words, the “meta-question” concerning whether E’s possessing F makes it more likely
that E is true is directly relevant to determining whether S’s believing that E is warranted. It is
only austere reliabilists who will not grant this point. For them, the belief that E is fully
epistemically warranted just in case it is produced by an appropriate process. As mentioned
above, the normative force behind the regress argument is simply that having reasons for
believing a proposition adds a type of epistemic warrant. Lacking a reason is problematic only
when seen from the standpoint of normative epistemology in which knowledge is taken to be the
most highly prized form of true belief—where, of course, it is the believing that is prized, not the
propositional content (see Plato 1980: 97a–98b).

3. Infinitism
Brief sketch

The upshot of the argument up to this point is that either we have to reduce what it takes to be
the most highly prized form of true belief to something akin to austere reliabilism or it appears
that there is no privileged belief which is immune to interrogation. The first alternative simply
ignores the normative intuition underlying the regress. But a major obstacle to accepting
infinitism remains. Recall the two principles that motivated foundationalism: the Inheritance
Principle and the Non-Originating Principle.

Together they rule out infinitism. For even if we had infinite time to produce reasons, it still
seems mysterious, if not downright impossible, that some belief could ever be warranted because
reasoning alone cannot warrant a belief. Coupled with the fact that compared to an infinitely
enduring being, we live but a nanosecond, the upshot seems to be that the Pyrrhonians were right
after all. Suspension of belief is the only apt attitude.

The answer to these worries and the key to understanding infinitism is that neither of the
principles, though they motivate and imply foundationalism, is required by all accounts of
epistemic warrant. Having reasons for a belief does add a type of warrant for holding it. Indeed,
having reasons for a belief is required for it to be the most highly prized form of true belief. In
other words, although there is some type of epistemic warrant that a belief acquires in virtue of
its etiology, having a reason for the belief provides a different type of warrant for believing it. I
say “different type” of warrant rather than just “more warrant” because no matter how reliable
the process is that produced the belief, the belief does not rise to the status of the most highly
prized form of true belief unless there are good reasons for holding it. So, although there is one
form of warrant that does not originate with reasoning, another form of warrant does. Thus, the
Non-Originating Principle is false. Having reasons for a belief provides it with a new type of
warrant. In addition, the inheritance principle is, at best, misleading since it seems to imply that
the warrant required for knowledge is transmitted by reasoning. But the reason, r, for a belief, b,
can provide b with a type of warrant that r, as yet, does not possess because no reason for r has
yet been given or located. So b could be known without r being known.

The infinitist will take the belief that p to be doxastically justified for S only if S has engaged in
providing “enough” reasons along the path of reasons. S would be completely doxastically
justified if every reason in the path were provided. But since it takes some time to discover and
offer reasons, even though a proposition might be completely justified (if there is a suitable
endless path of reasons), no belief could ever be completely doxastically justified. Nothing is
ever completely settled in the sense that it is beyond interrogation, but as S engages in the
process of providing more reasons for her beliefs they become better justified—not because S is
getting closer to completing the task but, rather, because S has added some warrant for her belief.
How far forward in providing reasons S needs to go in order to acquire knowledge seems to me
to be a matter of the pragmatic features of the epistemic context—just as which beliefs are being
questioned and which can be taken as reasons is at least partially contextually determined (see
Fantl 2003; Klein 2005a, 2005b, 2007a, 2007b).

Responses to Some Objections to Infinitism

Infinitism has not been taken as a serious contender among the answers to the regress problem
because there seem to be obvious, clear objections. But I think these objections to infinitism miss
the mark. Let us consider five of them:

1. The Finite Mind Objection

Aristotle correctly observed that beings with a finite mind cannot traverse an infinitely long
inference path because each inference takes some time. But infinitism—or at least the kind that
makes proper use of the distinction between propositional and doxastic justification—does not
require that an infinite set of reasons be produced or located in order for a belief to rise to the
level of the most highly prized form of true belief. Knowledge requires being able to provide
enough reasons for our believing. It does not require completing a task with an infinite number
of steps.

What constitutes “enough” reasons requires careful elaboration and I have not done that here.
Such an elaboration would include a discussion of the role of the contextual considerations that
make further questioning either necessary because a legitimate question has been raised and
understood or frivolous because the amount of added warrant that further investigation would
produce is minuscule. Those issues are beyond the scope of this essay.

2. The No-Starting Point Objection

The Pyrrhonians said that the process of reasoning endorsed by infinitism could not succeed in
justifying a belief sufficiently for us to adopt it because “we possess no starting-point for our
argument.” That objection has an intuitive tug only if we thought that knowledge could be
produced by reasoning only if all of the positive epistemic properties required for belief rising to
the level of knowledge had to be present in the reasons for the belief. But I hope I have dispelled
their intuitive appeal by showing how reasoning can produce a new type of warrant that is not
inherited from the offered reason.

3. Skepticism

Some philosophers have argued that knowledge entails certainty, where certainty includes at
least having finally settled the matter. And they would point out that infinitism makes that kind
of certainty impossible and, thus, infinitism leads to skepticism. There are two replies to this
objection.

First, as I mentioned earlier, there are both skeptical and non-skeptical forms of foundationalism.
There would be skeptical forms of coherentism if no belief set held by creatures like us could be
sufficiently coherent to satisfy the requirements of knowledge. In a similar vein, there certainly
could be skeptical forms of infinitism that held that the normative requirements of justification
simply cannot be fulfilled. The fact that a theory of justification leads to skepticism might
provide a basis for looking more carefully at whether the theory is correct, but that, alone, does
not strike me as a sufficient reason for rejecting the theory.
Second, the form of infinitism that I am defending does not lead to skepticism. It is a form of
fallibilism that eschews certainty as a requirement for knowledge, where certainty is construed as
requiring that the degree of epistemic warrant necessary for knowledge makes the belief immune
to further interrogation. Indeed, I think this form of infinitism can explain why certainty is taken
to be both a relative notion as when we say that one belief is more certain than another, and an
absolute notion as when we say that a belief is certain only if there is no belief that is more
certain. It can also explain why absolute certainty cannot be obtained because any belief can
always be made a little more certain by producing more reasons along the path of reasons while
at the same time it can explain how a belief can be certain enough to rise to the level required by
knowledge (see Klein 2005c).

4. Infinitism Really Endorses a Form of Arbitrary Foundationalism

It has been claimed that (1) infinitism is really a form of an unjustified (arbitrary) foundationalist
view, and (2) that a “bad” reason, r, could justify a belief, b. (See Bergmann 2007 for the
objection and Klein 2007b for a full response.) That infinitism is not a form of foundationalism
should be clear because it eschews the fundamental claim endorsed by foundationalists, namely,
that there are some beliefs immune to further interrogation. The answer to (2) is more complex.
There are several distinct factors that could make a reason, r, “bad” for believing b:

(i) A reason, r, could be “bad” because it was not formed in a reliable manner. Such a bad
reason could not transfer the kind of warrant required by the austere reliabilist to b by
reasoning, and consequently, neither b nor r would be knowledge—even according to the
infinitist. In other words, the infinitist can embrace the reliabilists’ basic insight that a
belief must be properly caused in order to be knowledge. So, in this sense r could not be
“bad” and lead to knowledge.

(ii) A reason, r, could be “bad” because there is no further reason for it. But note that in
such a case, r couldn’t have been formed reliably because the belief that b was reliably
formed is a good reason for thinking b is true. Hence, what was said with regard to (i)
applies here as well.

(iii) A reason, r, could be “bad” because S does not have available an answer to the
question as to why she believes that r is likely to be true. In such a case, although b has
gained some warrant because r was produced as a reason for believing b, b’s degree of
warrant would diminish. That strikes me as just what a theory of justification should
dictate. We are a bit better off by possessing r as a reason for b than we would be if we
had no reason for believing b, but we are not completely in the epistemic clear.

(iv) A reason, r, could be “bad” because it is false or there is a defeater of the reason for r.
If it is false, there is a defeater of the inference from the “bad” reason (namely, ~r).
Infinitism, per se, is an account of only the justification condition in knowledge; an
infinitist can include a no-defeater condition in the necessary conditions for knowledge.
So, such a “bad” reason could not lead to knowledge. (I should add parenthetically that I
think on some occasions a false belief can lead to knowledge and, hence, such useful
falsehoods are not “bad” reasons, but those considerations are irrelevant here because
 those false beliefs could appear in chains of reasons endorsed by foundationalists (see
Klein 2008).)

5. The Something from Nothing Objection

An anonymous reviewer of this chapter poses this question:

Q: Can a belief B be warranted (to at least some degree) by being based on a belief in
reason R1 if both of the following are true: (i) the belief in reason R1 is not reliably
formed and (ii) the believer has no reason for thinking the belief in reason R1 is likely to
be true?

The reviewer writes that a ‘yes’ answer “seems completely implausible” and that I seem
committed to a “no” answer. I suppose that a ‘yes’ answer seems so implausible because if B can
be warranted (at least partially) on the basis of R1, when R1 isn’t warranted at all, it seems that
some warrant is originating from nothing. The reviewer’s point is that if the correct answer to Q
is ‘no,’ then the Non-Originating Principle is true.

He/she writes:

Klein doesn’t directly answer Q in the paper, though he says that under these
circumstances, a belief in R1 couldn’t transfer the kind of warrant required by the austere
reliabilist and so B couldn’t amount to knowledge. So I think we should assume that
Klein thinks that B couldn’t be warranted to any degree at all by being so based and that
R1 couldn’t transfer any degree of warrant at all under conditions (i) and (ii).

I agree that R1 couldn’t transfer any degree of warrant under conditions (i) and (ii) because R1
has no warrant to transfer. But the reviewer is wrong in thinking that “B isn’t warranted at all by
being so based.” To repeat, a basic claim of infinitism is that reasoning can originate warrant.
When we locate a reason for a belief, we have provided that belief with some warrant which the
reason might not (yet) possess. Warrant hasn’t originated from nothing. It has originated through
the process of locating and citing the reason. Of course, B falls short of being knowledge because
“R1 was not reliably formed” is a defeater of R1’s justification for B, and, as mentioned above,
B lacks the kind of warrant that the reliabilists require of a belief.

Conclusion

I conclude that neither foundationalism nor coherentism provides an adequate non-skeptical

response to the epistemic regress problem. Only infinitism does.
The Solvability of Probabilistic Regresses. A Reply to Frederik Herzberg
David Atkinson and Jeanne Peijnenburg
Published in Studia Logica, 94 (2010) 347-353.

Abstract

We have earlier shown by construction that a proposition can have a well-defined nonzero
probability, even if it is justified by an infinite probabilistic regress. We thought this to be an
adequate rebuttal of foundationalist claims that probabilistic regresses must lead either to an
indeterminate, or to a determinate but zero probability. In a comment, Frederik Herzberg has
argued that our counterexamples are of a special kind, being what he calls `solvable'. In the
present reaction we investigate what Herzberg means by solvability. We discuss the advantages
and disadvantages of making solvability a sine qua non, and we ventilate our misgivings about
Herzberg's suggestion that the notion of solvability might help the foundationalist.

We further show that the canonical series arising from an infinite chain of conditional
probabilities always converges, and also that the sum is equal to the required unconditional
probability if a certain infinite product of conditional probabilities vanishes.

Keywords Probabilistic justification, regress problem, foundationalism, infinitism.

According to classical foundationalism, an infinite probabilistic regress either yields zero or
remains indeterminate. In several papers we have given counterexamples to this claim. That is,
we have demonstrated that a proposition can have a well-defined nonzero probability, even
though its probabilistic justification is forever postponed [1], [2], [3], [4].

In an interesting comment involving a pretty use of nonstandard analysis, Frederik Herzberg [5]
has proved that our counterexamples are consistent, in the sense that they have a model.
However, he also argues that they are of a very special kind. For our counterexamples are what
he calls `solvable' and most consistent probabilistic regresses lack that property. Our
counterexamples are thus exceptional, and this leads Herzberg to conclude that there is a way in
which a foundationalist might defend herself. For she might now argue that generically we are
unable to calculate or estimate an infinite probabilistic regress explicitly. In other words, she
might concede that some propositions which are justified by a probabilistic regress can have a
well-defined nonzero probability, but still deny that this is generally the case. In the present reply
to Herzberg we will focus on the notion of solvability.

(i) We will first investigate what Herzberg means by `solvable'. (ii) Then we discuss the pros and
cons of the requirement that a probabilistic regress can justify a proposition only if that regress is
in fact solvable. (iii) Finally, we consider two alternative ways in which a foundationalist might
defend herself.

(i) What does it mean to say that a probabilistic regress is solvable? Herzberg writes: “A
probabilistic regress < α, ẞ> is solvable if for any model <P; S > of <α; ẞ > one can derive a
closed-form expression for P(S0).” [5, p. 3; for the notation, see Herzberg's Definition 2].
However, he also notes that “[i]t is difficult to provide a precise definition of solvability, since
the notion of a closed-form expression is not a well-defined mathematical concept.” (ibid.) In the
first footnote of his paper Herzberg suggests that closed-form expressions include rational
numbers, but that it is unclear whether some irrational, algebraic or transcendental numbers also
fall under this heading. “All one can say”, he writes, “is that in all of these cases, any candidate
for [a] set of closed-form expressions will be a countable set C of real numbers.” (ibid.)
Acknowledging that there is no agreement on what this set C looks like, he concludes with a
counterfactual: “if there were universal agreement on the set C of closed-form expressions of real
numbers, a candidate for a definition of closed-form solvability would be as follows” (ibid., pp.
3-4) – after which he defines closed-form solvability in terms of C.

In order to gain a better understanding of solvability, let us take a closer look at the structure of
our counterexamples. The typical form of these counterexamples is:

P(S0) = A + B; (1)

where P(S0) is the probability associated with the target proposition S0, A is an infinite series of
conditional probabilities only, and B is a remainder term, containing precisely one unconditional
probability as a factor:

A = lim s→ {ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + … + γ0γ1 … γs-1ẞs}

 B = lim s→ {γ0γ1 … γs P(Ss+1)} (2)

(for the meaning of the symbols ẞ and γ see Herzberg's [5] and our [2]). Following Herzberg, we
shall say that Eq.(1) is solvable if and only if we can write down a closed-form expression for the
sum A + B, for example an expression in terms of a finite number of rational terms. Now it can
be proved that the infinite series of conditional probabilities A always converges, so A certainly
is equal to a unique and well-defined number (for a proof, see Appendix A). Moreover, in the
usual cases B will tend to zero (see Appendix B). This means that, as a rule, the infinite regress
of conditional probabilities produces a convergent series that is equal to the unconditional
probability of the target proposition S0. In other words: as a rule the unique and well-defined
value of A coincides with P(S0).

From this it does not follow, however, that this unique value can actually be written down as a
closed-form expression consisting only of known functions. It might after all be an expression
outside Herzberg's set C. In general the structure P(S0) = A + B is solvable if we can write the
unique value of A in an acceptable closed form. Since mostly we are not able to do this, we agree
with Herzberg that most convergent series are not solvable in his sense. The generic case is
indeed one in which an infinite probabilistic regress yields a convergent series, the sum of which
cannot be written down in a closed form based on familiar functions.

(ii) In Herzberg's view, this might help the foundationalist. For she can now argue that most
probabilistic regresses are unsolvable and thus that our solvable counterexamples form an
exceptional class. At the end of his paper Herzberg concludes that the generic, unsolvable
probabilistic regresses could produce a very small number: “for a given regress, P(S0) might be
very close to zero” [5, last sentence]. He thus suggests that, while the solvable regresses yield a
non-zero value for P(S0), the unsolvable ones might confer on this probability uncontrollably
tiny values. But this suggestion could easily mislead the reader. As we see from Eqs. (1)-(2), the
value of P(S0) cannot be smaller than ẞ0. Whenever the latter is not very close to zero, P(S0)
will not be close to zero either. Since this goes for solvable and unsolvable regresses alike, most
of the generic, unsolvable regresses will thus not end up close to zero, much as in the case of
nongeneric, solvable regresses.

A foundationalist might react to this by banning unsolvable regresses altogether. Indeed, if she is
a rigid constructivist, she would doubt the very existence of such regresses, so she will certainly
not acknowledge them as potential counterexamples in our sense.

How reasonable is the position of this constructive foundationalist? In order to answer this
question, we should bear in mind that invoking solvability carries with it two moments of choice.
First, we must choose a definition of solvability, which in this case amounts to specifying
Herzberg's set C. Which expressions are we supposed to include and which are we supposed to
leave out? As Herzberg notes, there is no clear answer to this question. Second, and
independently of what solvability looks like, we must decide whether or not we require
counterexamples to be solvable. A rigid constructivist, as we have seen, will indeed require
solvability as a sine qua non. For her, a probabilistic regress justifies a proposition only if its sum
can be written as a closed-form expression. We are however inclined to take a laxer stance. It is
true that most of our counterexamples yield rational numbers and thus are solvable in Herzberg's
sense. But this is only an accidental feature, introduced for reasons of simplicity and immediate
intelligibility. We never meant solvability to be an essential characteristic of our
counterexamples.

Of course, demanding that only solvable series can be counterexamples has the advantage of
ensuring that the latter are easily calculable. But the demand has obvious drawbacks as well. If
interpreted narrowly, it would exclude the use of many known and unknown higher functions.
Isn't it standard practice to define transcendental numbers and functions by means of convergent
series, by integral transforms, or by other unsolvable, but calculable expressions? And are there
not infinitely many other new transforms that can be computed? Personally we see no good
reason to ban any of them; but this is after all more a matter of taste than a matter of fact.

(iii) Assuming that a constructivist approach does not really help our hypothetical
foundationalist, what other lifelines are available to her? We can think of two. The infinite
product γ0γ1γ2 … normally diverges to zero (see Appendix B), so that the term B in Eq.(1)
vanishes. However – and this is the first lifeline – a foundationalist might wish to claim that a
`prime mover' (Ss+1 in the limit s → , thus infinitely far away) is still required to bring the
whole epistemic chain into existence, even though the stochastic support that it gives to S0 has
shrivelled away to nothing. If she were to grasp this lifeline, the foundationalist would commit
herself to a position that we might call `epistemological deism'. Seventeenth and eighteenth
century deists preached that God created the world, including the laws of nature, but waived any
interference after that. In a similar vein, our counterexamples postulate an infinitely remote
beginning of the justificatory regress, but deny that this beginning has any inuence whatsoever
on the probability value of the target proposition. All justification for the target proposition stems
from the conditional probabilities that make up the infinite series A.

Any foundationalist who feels uncomfortable with this meagre option might choose the second
lifeline. She could seek solace in the atypical case where the infinite product γ0γ1γ2 … (and thus
B) converges to a value that is not zero. The general expression (1) for the required probability
leads to

P(S0) = ẞ0 + ∑ n=1 ẞn IIn-1 s=0 γs + P* II n=0 γn.

where

P* = lim n 1 P(Sn) ;

on condition that this limit exists. In order to calculate P(S0) we need now not only the
conditional probabilities, but also P*; the unconditional probability associated with an infinitely
distant prime mover; for the latter still exerts probabilistic influence on the value of P(S0).

However, the second lifeline comes with a price tag. For not only is this class atypical, it is also
characterized by a sequence, {γn}, that tends very rapidly to unity as n tends to infinity. This
means that ẞn tends very rapidly to zero, and αn = ẞn + γn tends very rapidly to one. If it were
the case that αn = 1 and ẞn = 0 exactly for some n, then the relation between Sn and Sn+1 would
be one of bi-implication, Sn Sn+1, thus guaranteeing that Sn and Sn+1 have the same truth
values. Stochastically speaking, the link between the two propositions has been short-circuited,
and the same would apply to any link for which the corresponding γn is precisely equal to one. If
however γn is not precisely equal to one, but tends asymptotically very quickly to that value, we
might dub the relation between successive propositions in the epistemic chain, at any rate for
sufficiently large values of n, one of quasi-bi-implication. For small values of n the values of γn
might depart radically from unity, so we have the finite beginnings of a normal epistemic chain.
The infinite tail, however, would be one of quasi-bi-implication. In other words, all but a finite
part of the chain is approximately short-circuited through to the ground. The probabilistic chain
might be called quasi-finite, and it could perhaps serve as a generalization of the standard
foundationalist requirement that the series of justification be finite. However, this option seems
little more than a simple reformulation of the foundationalist stance and thus appears to come
close to begging the question of whether justification by probabilistic regress is possible.

Appendix A. Proof that the series converges

The unconditional probabilities P(Sn) and P(Sn+1) are related by the rule of total probability,

P(Sn) = ẞn + γnP(Sn+1), (3)

where

ẞn = P(Sn|~Sn+1) and γn = P(Sn|Sn+1) - P(Sn|~Sn+1), (4)

with γn > 0 for all n, which is the condition of probabilistic support. Eq.(3) can be iterated from
n = 0 up to n = s, with the result1

P(S0) = ẞ0 + ∑s n=1 γ0γ1 … γn-1ẞn + γ0γ1 … γsP(Ss+1), (5)

In the limit that one lets s go to infinity, the resulting infinite series is in fact convergent, as we
will now prove. Since ẞn + γn = P(Sn|Sn+1) ≤ 1 ; it follows that ẞn ≤ 1 - γn, and so

∑s n=1 γ0γ1 … γn-1ẞn ≤ ∑s n=1 γ0γ1 … γn-1(1 - γn)

= (γ0 - γ0γ1) + (γ0γ1 - γ0γ1γ2) + … + (γ0γ1 … γs-1 - γ0γ1 … γs)
= γ0 – γ0γ1 … γs ≤ γ0. (6)

which is finite and independent of s, and therefore also valid in the limit that s is taken to
infinity. Since all the ẞn are nonnegative and all the γn are positive, it follows that the infinite
series of positive terms

A = ẞ0 + ∑ n=1 γ0γ1 … γn-1ẞn

is convergent, for the partial sums (6) are bounded by γ0.

Appendix B. When does the remainder term vanish?

Consider the remainder term in Eq.(5), namely γ0 γ 1 … γsP(Ss+1). We will show that the
coefficient in front of P(Ss+1), i.e. γ0γ1 … γs, usually tends to zero in the infinite s limit. Since
γn = exp[log γn] = exp[- |log γn|],

γ0γ1γ2γ3 … = II n=0 γn = exp [-∑ n=0 |log γn|]. (7)

So γ0γ1 … γs has the limit zero as s goes to infinity if and only if the sum

∑ n=0 |log γn| (8)

diverges. Since all the terms in the series (8) are positive, the series can only converge, or
diverge to + . Unless γn tends very quickly to 1, the series will indeed diverge; hence an
infinite product of terms, each of which is smaller than one, is usually equal to zero. In this case

B = lim s→ {γ0γ1 … γsP(Ss+1)}

is also zero, since P(Ss+1) is bounded by unity.

Exceptionally, if γn tends rapidly to 1, this will not be the case. For example, if

1 - γn γn~n^-a as n → ,

then the sum (8) converges if a > 1, but it diverges if 0 < a ≤ 1. If γn tends less quickly still to
one, or does not tend to unity at all, which may be regarded as `usual', the series is divergent, and
by convention one says that the corresponding infinite product (7) diverges to zero.

Acknowledgement

We should like to thank Frederik Herzberg for his comments on an earlier version of this paper,
which led to a more nuanced discussion of solvability.

1 Formula (5) can be formally proved by mathematical induction as follows:

1. Since P(S0) = ẞ0 + ∑γ0P(S1), Eq.(5) is clearly true when s = 0.

2. Suppose P(S0) = ẞ0+ s-1 n=1 γ0γ1 … γn-1ẞn+γ0γ1 … γs-1P(Ss) for some s. Then, on
substituting ẞs +γsP(Ss+1) for P(Ss) in this formula, we obtain the right-hand side of Eq.
(5).

Therefore, if (5) is true for some s, it is true for any larger value of s. But it is true for s = 0, so it
is true for s = 1, 2, 3 ...
Notre Dame Journal of Formal Logic
Volume 51, Number 4, 2010
Justification by Infinite Loops
David Atkinson and Jeanne Peijnenburg

Abstract In an earlier paper we have shown that a proposition can have a well-defined
probability value, even if its justification consists of an infinite linear chain. In the present paper
we demonstrate that the same holds if the justification takes the form of a closed loop. Moreover,
in the limit that the size of the loop tends to infinity, the probability value of the justified
proposition is always well-defined, whereas this is not always so for the infinite linear chain.
This suggests that infinitism sits more comfortably with a coherentist view of justification than
with an approach in which justification is portrayed as a linear process.

Keywords: probabilistic justification, coherentism, infinitism

1 Introduction

Present-day epistemologists routinely assume that justification is probabilistic in character: a

proposition En may be justified by another proposition En+1, even if the latter only partially
supports the former. Recently we have shown that this apparently innocent assumption flies in
the face of classical foundationalism [1]. For if we take seriously that justification comes in
degrees, then there is in general no need for a foundation from which the justification springs.
The target proposition En may have a perfectly well-defined probability value, even though its
support consists of an infinite linear chain in which En is probabilistically justified by En+1,
which in turn is probabilistically justified by En+2, and so on, ad infinitum.1

In the present paper we extend this research on infinite linear chains by exploring the viability of
infinite epistemic loops. We show that, once justification is interpreted probabilistically, the
prospects for infinite loops are even brighter than those for infinite linear chains. If a proposition
is justified by an infinite loop, it always has a well-defined unconditional probability, whereas
this is not always so for infinite linear chains. An infinite linear chain normally confers upon the
target proposition an unconditional probability that is well-defined, but there are exceptional
cases in which it fails to do so.

Here is how we plan to make our case. In Section 2 we start with a discussion of finite and
infinite linear probabilistic chains, briefly summarizing the results that we derived in [1]. We
explain that infinite linear probabilistic chains always converge and that they normally yield a
well-defined probability value for the target proposition. In addition to the results in [1], we also
delineate a class of exceptional cases, in which the target proposition lacks a unique value. After
this analysis of justification by probabilistic chains, we turn our attention to justification by
probabilistic loops. We first discuss in Section 3 loops that are finite, describing and analyzing
several examples. We show that justification by a finite loop is often nontrivial, yielding a
definite value for the target proposition where we would not immediately expect it. In Section 4
we extend our study to loops of infinite size. We demonstrate that justification by an infinite
linear chain is usually indistinguishable from justification by an infinite loop. The only cases in
which an infinite chain and an infinite loop differ are the exceptional situations that we had
already identified in Section 2. In those situations, the infinite loop does, whereas the infinite
chain does not yield a well-defined unconditional probability for the target proposition. Finally,
in Section 5, we sum up our results.

2 Finite and Infinite Linear Chains

Let E0, E1, E2, . . . be a sequence of propositions, finite or infinite in number. We say that En is
probabilistically justified by En+1 if and only if the conditional probability of En, given that
En+1 is true, is greater than the conditional probability of En, given that En+1 is false:

P(En|En+1) > P(En|¬En+1) . (1)

The unconditional probabilities P(En) and P(En+1) are related by the rule of total probability,

P(En) = P(En|En+1)P(En+1) + P(En|¬En+1)[1 − P(En+1)] . (2)

With the abbreviations

αn = P(En|En+1),
ẞn = P(En|¬En+1),
 γn = αn − ẞn,

rule (2) becomes

P(En) = ẞn + γnP(En+1) . (3)

Clearly, γn > 0 is equivalent to the condition of probabilistic support as expressed in (1).

If each member of the sequence E0, E1, E2, . . . , Es+1, except Es+1, is probabilistically justified
by its successor, we speak of a finite linear chain of probabilistic support. The linear chain is then
grounded in the ultimate link Es+1, which is unsupported by any of the other links. We can
consider a finite chain in which the number of links, s+1, is fixed, for example when there are
just three of them, or we might be interested in a situation in which the length of the finite chain
is allowed to vary, thus turning s into a variable. In both cases we must find some reason for the
veridicality or plausibility of Es+1, on pain of leaving the entire chain hanging in the air.

Equation (3) can be iterated from n = 0 up to n = s, with the result

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + · · · + γ0γ1 . . . γs−1ẞs + 01 . . . γsP(Es+1) . (4)

Equation (4) is the most general formulation of a finite linear chain.2 It applies not only to cases
where both αn and ẞn are uniform, that is, where they remain constant from one link to the next
throughout the chain, but also to cases where both αn and ẞn are nonuniform, taking on different
values that depend on n. In [1] three examples were given of a finite linear chain with variable s:
one in which αn and ẞn do not depend on n, one in which they do depend on n although γn does
not, and finally one in which αn, ẞn, and γn all depend on n.

The structure of (4) can be represented as

P(E0) = X + YZ, (5)

where X is a finite sum of conditional probabilities only and YZ is a remainder term. The factor Y
is the finite product of γ factors, while Z is the unconditional probability of the alleged ground of
the chain, Es+1:

X = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + · · · + γ0γ1 . . . γs−1ẞs

Y = γ0γ1 . . . γs
Z = P(Es+1). (6)

In the limit that s goes to infinity, all the members of the sequence E0, E1, E2, . . . are
probabilistically justified by their successors. In this case the chain of support is infinite and its
structure becomes

P(E0) = X’ + Y’Z’, (7)

where X’, Y’, and Z’ are

X’ = lim s → {ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + · · · + γ0γ1 . . . γs−1ẞs}
Y’ = lim s → {γ0γ1 . . . γs}
Z’ = lim s → P(Es+1). (8)

We have proved that X’ always converges to a unique and well-defined number ([2], Appendix
A). In addition, we showed that Y’ usually tends to zero ([2], Appendix B). This means that, in
general, the product Y’Z’ vanishes and thus that, as a rule, P(E0) in (7) takes on the well-defined
value X’. It is only in very exceptional cases that Y’ does not tend to zero, and then the value of
P(E0) cannot be determined (see the Appendix of the present paper for the condition under
which this happens). Surprisingly enough, even in these exceptional cases the value of P(E0) can
still be determined if the probabilistic justification has the form of an infinite loop rather than an
infinite linear chain. We will come back to this point in Section 4, where we discuss justification
by infinite loops. First, in Section 3, we consider loops of finite size. We show that justification
by a finite probabilistic loop is nontrivial in most cases. For usually the conditional probabilities
on the loop vary, and then the value of P(E0) is a nontrivial function of the length of the loop.

3 Finite Loops

We have seen that Equation (4) is the general formulation of a finite linear chain. The general
formulation of a finite loop has a similar form, but for some finite s it is so that Es+1 = E0.
Mathematically, there is no problem whatsoever if we insert Es+1 = E0 into Equation (4):

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + · · · + γ0γ1 . . . γs−1ẞs + γ0γ1 . . . γsP(E0), (9)

for this yields

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + · · · + γ0γ1 . . . γs−1ẞs/1 − γ0γ1 . . . γs, (10)

which is well-defined, on condition that γ0γ1 . . . s is not equal to unity.3 With that proviso, the
solution furnishes a generic justification of the viability of the coherentist scenario in its simplest
form, that of a finite one-dimensional ring.

So mathematically speaking a self-supporting finite loop or ring is certainly possible. But the fact
that something makes good mathematical sense is, of course, not enough. Can a loop that closes
upon itself really occur? A temporal example of such a loop is difficult to come by in the real
world, but the science fiction of time travel can provide one. E0 could be the event that young
Bif decides in 1958 to use the 2018 edition of the sports almanac, E1 the event that he continues
his successful career as bettor until 2018, and E2 the event that old Bif succeeds in stealing Doc
Brown’s time machine in 2018, returning to 1958 in order to give the almanac to his younger
self. E3 = E0 could be the event that young Bif decides in 1958 to use the 2018 edition of the
sports almanac. . . and so on.

However, the events need not follow one another in time. For example, consider the following
three propositions:
C: “Peter read parts of the Critique of Pure Reason.”
P: “Peter is a philosopher.”
S: “Peter knows that Kant defended the synthetic a priori.”

Assuming that all philosophers read at least parts of the Critique of Pure Reason as
undergraduates, if Peter is a philosopher, then he read parts of the Critique. Of course, even if he
is not a philosopher, he may still have read Kant’s magnum opus. If Peter knows that Kant
defended the synthetic a priori, he very likely is a philosopher, whereas if he does not, he is
probably not a philosopher, although of course he might be an exceptionally incompetent one,
not having understood anything of Kant or the Critique. Finally, if he read the Critique, he quite
likely knows that Kant defended the synthetic a priori, whereas this is rather less likely if he
never opened the book.

Here then is a simple finite loop, consisting of a fixed number of links, namely three:

C <− P <− S <− C, (11)

where the arrow indicates that the proposition at the right-hand side probabilistically justifies the
one at the left.

We can make loop (11) nonuniform by investing the three propositions C, P, and S with, for
example, the following dissimilar values for the the conditional probabilities:

C: α0 = P(C|P) = 1 ; ẞ0 = P(C|¬P) = 1/10 ; γ0 = α0 − ẞ0 = 9/10

P: α1 = P(P|S) = 9/10 ; ẞ1 = P(P |¬S) = 1/5 ; γ1 = α1 − ẞ1 = 7/10

S: α2 = P(S|C) = 4/5 ; ẞ2 = P(S|¬C) = 2/5 ; γ2 = α2 − ẞ2 = 2/5.

Then the unconditional probabilities4 are

P(C) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2/1 – γ0γ1γ2 = 0.711

P(P) = ẞ1 + γ1ẞ2 + γ1γ2ẞ0/1 – γ0γ1γ2 = 0.679

P(S) = ẞ2 + γ2ẞ0 + 20ẞ1/1 – γ0γ1γ2 = 0.684.

The number of links in the above nonuniform loop is fixed—there are exactly three propositions.
Often, however, we are dealing with cases in which there is a variable s. Let us therefore look at
an example of a nonuniform loop in which the number of links varies. We will see that, in this
example, the value of P(E0) depends on the length of the loop in a nontrivial way. Consider

ẞn = 1/n + 3 γn = n + 1/n + 2 = 1 – 1/n + 2. (12)

Then
 γ0γ1 . . . γs = 1/2 × 2/3 × · · · × s/s+1 × s+1/s+2 = 1/s+2

γ0γ1 . . . γs−1ẞs = 1/s+1 × 1/s+3 = 1/2(1/s+1 – 1/s+3),

so Equation (10) reduces to

P(E0) = 1/3 + 1/2(1/2 – 1/4) + 1/2 (1/3 – 1/5) + · · · 1/2(1/s+1 – 1/s+3 )/1 – 1/s+2
= s + 2/s + 1(1/3 + 1/2 (1/2 + 1/3 – 1/s + 2 – 1/s + 3)] = 3s + 8/4(s + 3). (13)

Here P(E0) does depend on the number of sites on the finite loop (and, moreover, there is a
definite limit as the number of sites tends to infinity, namely 3/4). What is more, since in this
example the conditional probabilities, P(En|En+1) and P(En|¬En+1), are not the same for
different n, the unconditional probabilities, P(En), also are not the same for different n. In fact,
one finds

P(En) = 1 – 1/2(1/n + 2) – 1/(n + 1/s + 3), (14)

which indeed depends nontrivially on n, as well as on s, the length of the loop.5

Cases like (13), in which the value of P(E0) varies with the number of links, form in fact the
generic situation. There also exist special cases, where the value of P(E0) does not depend on s.
Such nongeneric cases arise when the conditional probabilities are uniform (that is, all the ẞn are
the same, and all the γn are the same, independent of n). Intuitively, it is clear that loops for
which the conditional probabilities are uniform will yield unconditional probabilities that are
independent of s. Here is a formal proof. In the uniform case, Equation (10) becomes

P(E0) = ẞ(1 + γ + γ^2 + · · · γ^s)/1 – γ^s+1. (15)

The finite geometrical series 1 + γ + γ^2 + · · · γ^s is equal to (1 – γ^s+1)/(1 − γ), and on
substituting this we find

P(E0) = ẞ/1 − γ. (16)

Indeed, this does not depend on s at all. The value of P(E0) is the same, irrespective of s, that is,
however long or short the loop may be. Moreover, in this uniform case P(E0), P(E1), P(E2), and
so on, are all equal, since they are all determined by the same cyclic expression (15). The number
of links is completely irrelevant to the value of the unconditional probabilities in the uniform
case; moreover, this holds whether s is finite or infinite.

4 Infinite Loops

We consider now an infinite loop of probabilistic support, that is, one where s in (9) and (10)
goes to infinity. We first look at what happens when the product γ0γ1 . . . γs tends to zero as s
goes to infinity, and then what happens when it doesn’t. The former represents the typical case,
the latter the atypical one.
If γ0γ1 . . . γs tends to zero, (10) yields the infinite series

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + γ0γ1γ2ẞ3 . . . (17)

and we have seen that such an infinite series is always convergent. In fact, Equation (17) is what
one obtains by letting s in Equation (4) go to infinity. So in the typical case there is no difference
between the infinite chain and the infinite loop.

The above equation covers both the uniform and the nonuniform cases. In the uniform infinite
situation, in which ẞn and γn are constant, that is, independent of n, Equation (17) reduces to

P(E0) = ẞ(1 + γ + γ^2 + . . . ) = ẞ/1 − γ. (18)

This infinite uniform loop yields the same unconditional probability as does the finite uniform
loop, a fact that is intuitively easy to grasp. After all, in the latter case, propositions are
uniformly connected round and round ad infinitum.

The nonuniform case is more interesting. Here (17) can take many different forms, dependent on
the values of _n and n. If we choose for these values the ones that were given in (12), then we
obtain the expression (14) in the limit that s is taken to infinity, namely,

P(En) = 1 – 1/2(1/n + 2) = 2n + 3/2(n + 2). (19)

In particular, P(E0) = 3/4, as we have already noted—see (13) and the lines following that
equation.

So much for the typical case. What of the atypical situation in which the infinite product of the γs
is not zero? Here the linear chain fails, in the infinite limit, to produce a definite answer for the
probability, whereas the infinite loop gives a unique value. To illustrate this, consider the
specific example

ẞn = 1/(n + 2)(n + 3) γn = (n + 1)(n + 3)/(n + 2)^2 = 1 – 1/(n + 2)^2.

The crucial difference is that here 1 − γn tends to zero as fast as 1/n^2, whereas this difference
had the slower asymptotic behavior 1/n in Equation (12). We find now

γ0γ1 . . . γs = [1/2 · 3/2 ] × [2/3 · 4/3] × · · · × [s+1/s+2 · s+3/s+2] = 1/2(s+3/s+2)

γ0γ1 . . . γs−1ẞs = 1/2(s+2/s+1) × 1/(s+2)(s+3) = 1/4(1/s+1 – 1/s+3),

so Equation (4) becomes

P(E0) = 1/6 – 1/4 (1/2 – 1/4 + 1/3 – 1/5 + 1/4 – 1/6 + · · · + 1/s – 1/s+2 + 1/s+1 – 1/s+3)
+ 1/2[(s+3)(s+2)] P(Es+1)

= 3/8 – 1/4 [(1/s + 2) + 1/(s + 3)] + 1/2[(s + 3)/s + 2)] P(Es+1). (20)
Note that the coefficient in front of P(Es+1) does not vanish in the limit that s tends to infinity.
Indeed, in this limit we find formally

P(E0) = 3/8 + 1/2 P(E ),

where P(E1) is an indeterminate number in the interval [0, 1]. The infinite linear chain has in this
case failed to produce a definite value for the probability. However, for the infinite loop we can
set P(Es+1) = P(E0) in Equation (20) and then we can solve the linear equation for P(E0) to
obtain

P(E0) = [3/8 – 1/4(1/s+2 + 1/s+3)] / [1 – 1/2(s+3)(s+2)] = 3/4 – 1/4(1/s + 3),

which has the perfectly definite limit 3/4. Thus, the infinite chain and the infinite ring only differ
when γs tends to unity with sufficient rapidity. In the Appendix we formulate a necessary and
sufficient condition under which this happens, thus delineating the entire class of cases in which
the infinite chain fails to give a definite answer, but the infinite loop does so.

5 A Pluralistic Picture

In 1956 Sellars diagnosed the malaise of epistemology as an unpalatable either/or: “One seems
forced to choose between the picture of an elephant which rests on a tortoise (What supports the
tortoise?) and the picture of a great Hegelian serpent of knowledge with its tail in its mouth
(Where does it begin?). Neither will do” ([3], p. 300). Sellars was presumably thinking of chains
and loops, both involving entailment relations only, and then indeed neither is adequate.
However, if support is interpreted probabilistically, then we are not confined to these two
possibilities. For then a picture of justification emerges that is distinctly pluralistic. A target
proposition, En, can be probabilistically justified by a finite or an infinite chain, or it can be
justified by a finite or an infinite loop. In each of these four cases the conditional probabilities
might be uniform or they might be nonuniform. These three parameters (finite versus infinite
chain, finite versus infinite loop, uniform versus nonuniform) thus yield eight different varieties
of probabilistic support.

The main result of our paper pertains to probabilistic support that is nonuniform and infinite. At
first sight one might think that a nonuniform loop of infinite length cannot really be called a
loop, since there is no end of the tail that the Hegelian serpent can swallow. After all, is it not the
case that a Wiederkehr des Gleichen, to quote another German philosopher, must require that the
loop be finite? The loop may be long, indeed more than cosmologically long, but it seems that it
may not be infinite, on pain of having no Wiederkehr at all. Moreover, even Poincaré, when he
formulated his recurrence theorem, had to assume that the universe is finite in spatial extent and
of finite energy: those are necessary conditions for a recurrence in finite time.

So it seems that a real loop differs from an infinite “loop.” However, from this it does not follow
that, therefore, an infinite loop is in fact an infinite chain. Our investigation shows that such a
conclusion would be unwarranted. It is true that an infinite uniform loop cannot be distinguished
from an infinite uniform chain: both yield the same trivial result. It is also true that, usually, the
infinite nonuniform loop produces the same value as does the infinite nonuniform chain.
However, there are exceptional situations in which infinite nonuniform loops and infinite
nonuniform chains yield different results. These exceptions consist in cases where the infinitely
far away “end” of the chain can still exert some influence on the probability of the target
proposition. As we have shown, an infinite loop has an even wider domain than does an infinite
linear chain: an infinitely long serpent succeeds even when an infinite stack of tortoises fails.

Appendix

What is the general condition under which the infinite linear chain and the infinite loop fail to
agree? Clearly if γ0γ1 . . . γs does not vanish in the limit of infinite s. Every n lies in the open
interval (0, 1), the extreme values 0 and 1 being excluded by fiat. Since γn = exp[log γn] = exp[−|
log γn|], we have

γ0γ1γ2γ3 · · · = II s=0 γs = exp [ − Σ s=0 |log γs|]

.
So γ0γ1 . . . γs has a nonzero limit as s goes to infinity if and only if the sum

σ=Σ s=0 |log γs| (21)

is convergent. Clearly, convergence can occur only if γ s tends sufficiently quickly to 1. For
example, if

γs ~ 1 – s^−a

for large s, then Equation (21) converges if a > 1, and in that case the coherentist loop (10) yields

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + . . ./1 − e−^a,

but the linear infinitist chain instead gives

P(E0) = ẞ0 + γ0ẞ1 + γ0γ1ẞ2 + · · · + e^−a P(E ).

Whereas the infinite loop gives a definite answer in terms of all the conditional probabilities, the
linear chain is problematic in the infinite limit, for P(E1) must be construed as an indeterminate
number between 0 and 1.

Notes

1. For a more extensive discussion of how patterns of probabilistic dependence are relevant to an
understanding of epistemic justification, see our earlier paper [1]. The present article may be
regarded as a sequel to that work.

2. Equation (4) is the same as Equation (13) in [1].

3. If γ0γ1 . . . γs = 1, it follows that all the γs are separately equal to one. But then all the αs are
equal to one also, and all the ẞs are equal to zero, which is the condition of biimplication. It is
clear logically that the propositions could all be true, or all be false, for these two extreme
possibilities are obviously consistent with the bi-implication. However, the lack of uniqueness
goes further, for any probability between 0 and 1, if it is shared by all the propositions, is
consistent with bi-implication. This is a direct consequence of the rule of total probability when
α = 1 and ẞ = 0. So if the product of the γs is unity, the indeterminacy is maximal, and if it is not
unity, then the unconditional probabilities are determined uniquely.

4. As they must, these numbers satisfy

P(C) = ẞ0 + γ0P(P) P(P) = ẞ1 + γ1P(S) P(S) = ẞ2 + γ2P(C).

Incidentally, there is a good reason for considering a loop of at least three propositions. For in a
“loop” of just two links, there are only three independent unconditional probabilities, for
example P(E0), P(E1), and P(E0 ^ E1), whereas there are four conditional probabilities around
the loop, P(E0|E1), P(E0|¬E1), P(E1|E0), and P(E1|¬E0). So there must be a linear relation
between them, which means that all four may not be chosen independently. This difficulty does
not arise for a loop of three links, for in this case there are seven independent unconditional
probabilities and only six conditional probabilities around the loop, so the latter may be chosen
arbitrarily. With more than three links on the loop there is even more freedom, so the conditional
probabilities may again be chosen freely.

5. For a general value of m between 0 and s, we see by iteration of Equation (3) from n = m to n
= s that P(Em) = ẞm + γmẞm+1 + · · · + γm . . . γs−1ẞs + γm . . . γsP(Es+1). Much as in the
case m = 0, we find γm . . . γs = m+1/s+2 and

γm . . . γs−1ẞs = m+1/2(1/s+1 – 1/s+3).

With these expressions in hand, we work through steps entirely analogous to those given in (13),
ending with formula (14).

Acknowledgments

At a conference in Amsterdam in May 2007, David Makinson challenged us to make sense of a

coherent loop of probabilities. This paper is the direct result, and we thank Makinson for the
fruitful stimulus.
Foundations of Deduction’s Pedigree: A Non-Inferential Account
Jeremy Seitz
A thesis presented to the University of Waterloo in fulfillment of the thesis requirement for the
degree of Master of Arts in Philosophy

I hereby declare that I am the sole author of this thesis. This is a true copy of the thesis, including
any required final revisions, as accepted by my examiners. I understand that my thesis may be
made electronically available to the public.

Abstract

In this thesis I discuss the problems associated with the epistemological task of arriving at basic
logical knowledge. This is knowledge that the primitive rules of inference we use in deductive
reasoning are correct. Knowledge of correctness, like all knowledge, is available to us either as
the product of inference, or it is available non-inferentially. Success in the campaign to justify
the correctness of these rules is mired by opposing views on how to do this properly. Inferential
justifications of rules of inference, which are based on reasons, lead to regressive or circular
results. Non-inferential justifications, based on something other than reasons, at first do not seem
to fare any better: without a basis for these justifications, they appear arbitrary and unfounded.
The works of Boghossian and Dummett who argue for an inferentialist approach, and Hale who
supports non-inferentialism are carefully examined in this thesis. I conclude by finding
superiority in Hale’s suggestion that a particular set of basic logical constants are indispensable
to deductive reasoning. I suggest that we endorse a principle which states that rules are not
premises, and are therefore to be excluded from expression as statements in a deductive
argument. I argue that the quality of being indispensable is sufficient for a basic rule of
deduction to be countenanced as default-justified, and therefore need not be expressed in
argument. By a rule’s evading expression in argument, it avoids circular reasoning in deductive
arguments about its own correctness.

Another important outcome that emerges from my research is the finding that non inferential
knowledge is ontologically prior to the inferential sort. This is because plausible inferential
knowledge of basic logical constants shall always be justified by circular reasoning that already
assumes the correctness of the rule to be vindicated. This initial assumption is tantamount to non-
inferential knowledge, and therefore this latter is more primitive—in fact the only primitive—
species of basic logical knowledge.

Acknowledgements

I have my supervisor, Dave DeVidi, to thank for guiding me through this thesis experience.
Among the many helpful aspects of his supervision, foremost was his practice of assuring that I
said what I meant, and meant what I said—no more and no less. I am also thankful for Dave’s
patience and diligence when reading my research and providing feedback. I could not have
learned as much as I did about the epistemology of logic without this high level of dedication
Dave maintained in his role as supervisor.

I would also like to acknowledge and express my gratitude to the University of Waterloo, its
Faculty of Arts, and Philosophy Department for the generous scholarships and teaching
assistantships I received during my wonderful Master’s degree experience at UW.

Contents
1 Basic Logical Knowledge and its Problems 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Rules and Regress Explained . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Usefulness of Deduction and a Slice of its History . . . . . . . . 8
1.4 The Problems Facing BLK . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Carroll . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.6 Agenda for Resolution of the Problem of Warrant . . . . . . . . . . 15
2 Treatment of the Problems 19
2.1 Tonk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.2 Tonk gets Belnapped . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.3 Carroll, Regress, and Scepticism . . . . . . . . . . . . . . . . . . . . 26
2.4 Circularity, Consistency, Harmony, Scepticism . . . . . . . . . . . . 31
2.5 Boghossian and Conceptual Role Semantics . . . . . . . . . . . . . 39
2.6 Hale’s Non-Inferential Route . . . . . . . . . . . . . . . . . . . . . . 44
2.7 Summary of Findings in this Chapter . . . . . . . . . . . . . . . . . 49
3 The Primacy of Non-Inferential Justification 53
3.1 Distinguishing Types of Non-Inferential Knowledge . . . . . . . . . 53
3.2 Indispensability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.3 The Primacy of Non-Inferential Knowledge of Logical Laws . . . . . 57
References 59

Chapter 1

Basic Logical Knowledge and its Problems

1.1 Introduction

Every thing we know is based on some other thing that we know, or so the saying goes. If this
were true, then there should be some knowledge with origins that are inaccessible to us, and so
this adage might never be confirmed or disproven. The issue of finding a basis for knowledge is
of primary concern to the logician and epistemologist, because knowing how we possess the
knowledge we do, if such sort of knowing is possible, gives us confidence to pursue new
knowledge that we may regard as soundly derived. It also provides us with some reassurance that
our ‘webs of belief’ are consistent, viz., that the corporate body of individual beliefs we each
possess more-or-less agree with each other as a conglomerate whole. In the case of logic, my
concern in this project, finding a basis for our knowledge that a given logical language permits
the inference only of sentences that are consistent with each other and valid in that system
involves demonstrating that its most fundamental elements are correct. If we may be persuaded
of this, then warrant exists to reason according to the principles of a particular logic with
confidence.

To establish the correctness of these principles, we shall likely have to proceed in a fashion that
is at odds with the supposition of the adage above, or else we shall never arrive at what is first,
basic, and certain. The following example from fictional literature suggests a strategy we might
apply to overcome the obstacle presented by the adage. It suggests an epistemological attitude
towards knowledge that I shall favour and defend in this thesis: Baron Mu¨nchausen recounts the
story of his falling into quicksand. To avoid certain drowning, he hoists himself to safety by a
firm tug of his own ponytail. This amusing tale is analogous to the problem that concerns me
where, in the case of knowledge of logic, some approaches to justification like the one
championed by the adage succumb to a regress and fail to arrive at a basis—at what lies first in a
chain of reasoning. But Munchausen’s tale portends an alternative; the ‘bootstrapping’ that he
performs stands as a place to start that begins in itself. The solution to his problem comes from
nothing external to his situation—such as a low-lying vine or wooden plank to facilitate his
escape.

In similar spirit, this thesis shall defend the idea that knowledge of the correctness of the rules
governing basic logical constants is knowledge that is non-inferential in status. Furthermore, I
shall argue that any plausible inferential strategy proposed to justify that correctness shall depend
on non-inferential knowledge. The latter therefore stands prior to what has been derived
inferentially, and this renders non-basic any inferentially-derived knowledge of correctness. I
shall presently deliver a characteristic of basic logical knowledge, followed by an account of the
two ways available to us for arriving at knowledge: by inference or by non-inferential means.

A rule of inference in deductive logic is a metalinguistic expression. It governs operations on—

or defines the function of—a logical operator, which is a syntactic element of the formal system
of deductive logic. Rules of inference that define an operator come in pairs; one rule articulating
how to introduce the operator in a chain of reasoning, another telling us how to eliminate that
operator from the chain. Rules of inference act as functions taking us from premises to
conclusions, and the rules of inference admitted in deductive logic are universally valid. Basic
logical knowledge (or BLK) is knowledge we purportedly have about rules of inference;
knowledge that a rule is correct. Important paradoxes emerge from attempts to arrive at this
knowledge; the substance of this thesis is largely a reckoning of whether this knowledge is
possible and if so, how.

An inferential justification of a claim about a rule of inference, for instance the claim that the
basic rule modus ponens (hereafter MPP) is correct achieves its objective by means of inference;
it uses deductive reasoning to corroborate that claim. The virtue of an inferential justification in
the eyes of a reasoner who assesses an argument for validity is that an inferential argument
presents us with some sort of substantiation in support of our conclusion. I anticipate one point
of view an inferentialist might hold about deductive arguments such that in an important sense it
seems like the preferred approach to justification—for it appears to answer the question how are
we justified? in a rather effective way, viz., via deductively valid argument. In this thesis
deductively valid arguments shall be regarded as rigourous ones because deductive logic allows
us to establish, for any given conclusion, whether it follows from premises by means of a
systematic proof-checking procedure based on accepted rules of inference or axioms. Similar
rigour is attributable to the domain of mathematics. Kreisel for instance describes the sort of
rigour that commonly serves as the touchstone of mathematical proofs: rigour may be thought of
in the ‘Turing sense’ with respect to decidability.1

The non-inferential account of knowledge takes a different approach by grounding a justification

of our knowledge that a rule of inference is correct in a stipulation, a proffered definition, a
stated intuition, or something along those lines. Many philosophical views exist about these
forms of non-inferential knowledge, and what is common to all of them is the view that they are
basic in themselves. Some (radical) non-inferentialists hold that these rules are not in need of any
justification, or perhaps that they are default-justified. Other non-inferentialists such as Hale
argue more reasonably that non-inferential knowledge of a specific kind is indispensable to the
practice of deduction, and therefore that our non-inferential knowledge about basic rules is
justified because of this indispensability. Hale’s view is the one I defend in this thesis. The
inferentialist repudiates the basis on which rests non-inferential knowledge, and this creates an
epistemological tension between deductive and rigourous knowledge on one hand, and intuitive
notions on the other.

This project fundamentally involves a rejection of the attitude expressed by the adage at the
beginning of this section. It argues for the primacy and superiority of non-inferential justification
of the correctness of the most basic logical constants on the basis that (a.) we can effectively
bestow basic logical rules and operators with meaning when their introduction and elimination
rules are in harmony. (b.) That a particular subset of basic rules or operators with fixed meaning
should not be doubted because those rules are indispensable to the deductive reasoning process.
(c.) That those rules that are impossible to doubt participate in a minimal kit of rules (viz., basic
rules) that we cannot help but use in reasoning. (d.) Because we cannot help but use them in
reasoning they are essential to logic and deductive reasoning, and are correct. (e.) Finally,
because the basic rules enjoy this status as indispensable, essential, and implicit in the reasoning
process they need not, and should not figure as premises in deductive reasoning. Items (a.)
through (d.) are ideas belonging to Hale which I expound and endorse. (e.) is one of my
hypotheses, and the following is my own contribution to this enquiry: The upshot of a defense of
non-inferentialism is that in the case of an argument that seeks to justify a basic rule of inference,
we shall have license to suppose the correctness of that rule without a deductively valid proof of
its correctness. A further result emerging from the status of a basic rule’s implicit nature is that
the reasoner is absolved from obligation to have that rule figure assertorically in reasoning. The
theory establishes a purported distinction between rules and premises in deductive reasoning. By
omitting explicit expression of a particular rule in an argument, we avoid circularity when
reasoning about the correctness of that rule and thereby arrive at a plausible justification for our
application of the rule—viz., for its correctness. The most original argument I advance in my
thesis is one that I have not come across anywhere in the literature: it says that non-inferential
knowledge of logical rules can be shown to be ontologically prior to knowledge arrived at
inferentially. In fact, inferential knowledge depends on non-inferential knowledge, and this in
turn renders inferential knowledge non-basic.

Most people familiar with deductive logic understand notions such as validity and soundness of
arguments. While these notions are discussed at great length in this thesis, it is important to
acknowledge my principal concern here—for the correctness of the rules of inference. Validity
and soundness are not called into question - I am not concerned to know, for example, whether
validity is the right sort of criterion to respect in order for our conclusions to follow necessarily
from premises. We enter into an enquiry about the correctness of basic rules because, as we will
see, their origin and epistemological status is still undecided, evidenced by the opposing views
held by inferentialists and non-inferentialists on the matter.

This thesis discusses problems nested in the epistemology of justification— the study of how
belief is justified. Its philosophical literature generally countenances three main underlying
theories: foundationalism, coherentism, and infinitism. Infinitism holds that the structure of
justificatory reasons is infinite and non-repeating. It is an appealing theory because it can provide
an acceptable account of rational belief, viz., beliefs held on the basis of adequate reasons.2 But
most theorists reject infinitism as absurd because under this theory, the evidential ancestry of any
statement is infinite. This is particularly troublesome to a deductive justification of a conclusion.
Foundationalism is typified by Descartes’ famous statement ‘I think therefore I am’. He believed
that he could build his edifice of knowledge on that fact; a fact about which he was undoubtedly
certain; a fact that is ‘default-justified’, or ‘self-justifying’. The fact served as his foundation—as
a place to begin. Opponents of a foundationalist theory of justification argue that it is
unacceptable because it countenances an arbitrary reason as its base—arbitrary in a similar way
to the status of non-inferential knowledge I described above. In an important sense we cannot
call a foundationalist belief rational because no further reasons exist as justification for it.
Coherentism sees justification as holisitic. Instead of there being a chain of reasons justifying a
statement, a statement is justified because it is included in and coheres with a set of statements
that are accepted. Minimally, the statement must be consistent with the other statements of the
set. Objections to coherentism include the argument that it leads to circularity. Furthermore, if
we examine the set of accepted statements just described, no justification of the set of statements
necessarily exists, and may have been accepted in foundationalist spirit. In this sense
coherentism is disguised as foundationalism. Just as I have shown the correlation between
foundationalism and non-inferentialism several sentences above, so too do I loosely suggest a
synergy between coherentism and inferentialism.

A defense of my thesis shall unfold in the following manner. Chapter 1 introduces the concept of
basic logical knowledge and the practice of deductive inference. It presents a sceptical case from
Lewis Carroll challenging the possibility of justifying our knowledge that the rule MPP is
correct. Chapter 2 surveys responses to this challenge offered by Belnap, Boghossian, Dummett,
Engels, Hale, and Prior. In all cases these responses include profound and insightful discussion
about how the basic rules figure in logic and reasoning. Chapter 3 will conclude the thesis by
defending the claims (a.) through (e.), as well as by defending my own findings about the
primacy of non-inferential knowledge, in support of my preference for a non-inferential strategy
in any justification of the correctness of a select number of basic logical constants.

1.2 Rules and Regress Explained

What qualifies as a rule of inference? Since we are concerned with logic in this enquiry, we
discuss forms of reasoning rather than particular arguments, forms such as {α → β, α, ∴ β}. The
form expressed, which instantiates conditionalelimination or MPP, is universally valid: it
exemplifies the standard or universal form of a deductive argument that applies MPP in order to
arrive at its conclusion, and because any contentful argument, viz., one that contains at least one
premise leading to a conclusion about something in a particular domain of discourse, can be
translated into this (or some similar) logical representation, we call it universal. But this is not to
make an argument by enumerative induction for the validity of MPP. Enumerative induction fails
to give a rigourous justification because it is impossible to test every (past, present, and future)
instantiation of MPP for validity. Must we qualify the universality of MPP’s logical applicability
such that it is universal in the context of deductive or classical logic but perhaps not in some
other contexts? Well, we don’t want to.3 This is one of several important issues on the block in
any attempt to justify basic rules of inference. The universality, necessity, and a priority of MPP
that we seek to demonstrate should tell us that MPP is truth-preserving, in all instances, in any
model, and showing that MPP has these properties constitutes some way towards success to the
epistemologist attempting to justify basic rules of inference.

Rules of inference must be necessary. Necessity emerges from one of the characteristics of
validity: it obtains when the premises of an argument are true, and the conclusion is validly
inferred from the premises by means of a correct rule of inference. It’s impossible for that
conclusion to be false in any case where the premises are true. That was a brief characteristic of
necessity as it pertains to deductive arguments. But we are also concerned with what it means for
a rule of inference to be necessary, and necessity comes in two forms, absolute and relative. The
sort of necessity we wish to assign to rules of inference in order for them to be warranted in
deduction is absolute necessity. The difference between the two is that relative necessities are
only applicable in specific or restricted contexts. A useful example might be mathematical
induction; while Frege took it to be a logical principle, it is nowadays not reckoned as such
because while it holds for the natural numbers and the many domains for which suitable
isomorphisms to the natural numbers can be defined, there are many domains in which it does
not hold, including, for instance, the reals but also even some weak arithmetics. Hence the
principle of mathematical induction is merely a relative necessity. In the present case, absolutely
necessary rules of inference are those for which there is no possibility that (a.) they are
inapplicable or not warranted in the domain of discourse, and (b.) do not preserve the validity of
arguments. We may cautiously call certain rules of inference such as MPP absolutely necessary
and accept the impossibility that they are doubtable. We also accept the indefeasibility of certain
other types of knowledge such as knowledge of the axioms of arithmetic. However we must
acknowledge the case of Euclidean geometry, knowledge and understanding of those axioms
having been held to be tantamount to logical knowledge and therefore by extension necessary for
a very long time. This geometry showed us that even seemingly necessary knowledege is
defeasible. So acknowledging this possible obstacle, we nevertheless want to talk about MPP as
a rule of inference that is absolutely necessary, universal as described above, and finally a priori.
A priority characterises the way in which a reasoner arrives at a conclusion, or arrives at
knowledge. To be realised a priori means that a reasoner surpasses the character of experience to
deduce his conclusion. We then want to present the right sort of reasons that demonstrate basic
rules of inference as being in possession of each of these properties.

Without giving much thought to the matter, the rule of inference MPP seems like one of those
rules we need not question. The obvious nature of a deductive argument that eliminates its
conditional by applying MPP thus; {α → β, α, ∴ β}, seems like an argument that employs a rule
that may be unreflectively countenanced as correct. However, as we unpack this simple rule of
inference and attempt to justify our warrant for applying it, we run into the obstacles of regress
and circularity. The work of the philosophers I discuss in Chapter 2 is concerned with finding
solutions to and ways-around these obstacles.

The background to these problems circulate around epistemological and metaphysical stances
towards the right method of justification—of basic rules of inference. Lewis Carroll’s puzzle
presented in his paper What the Tortoise Said to Achilles4 demonstrates that knowledge of logic
is insufficient to warrant our use of basic rules of inference. Warrant must therefore be grounded
in something other than our knowledge of logical laws if it is groundable at all. Otherwise, a
warrant grounded in such knowledge may lead to regress. Patrice Philie presents this regress in
an extremely clear way in his paper Carroll’s Regress and the Epistemology of Logic:5

First we consider an argument employing MPP:

(1) If Socrates is a man, then Socrates is mortal.

(2) Socrates is a man.

Therefore,

(C) Socrates is Mortal.

We must know two things about the argument; first, that the premises themselves are warranted,
and second, that the argument possesses a valid logical form. We must therefore be in possession
of (viz., we must know) the following principle;

(W) If an argument displays the logical form of MPP, then it is valid.

Principle (W) is therefore the warrant that moves us from premises to conclusion. But how does
it do this? A reasoner in this case must know two things; first he must know (W), and second that
the argument above is an instance of MPP.

So what one must know can be summarized as follows:

(a) (W): If an argument displays the logical form of MPP, then it is valid.

(b) The argument above ((1), (2), (C)) is an instance of MPP.

(c) Therefore, the conclusion (C)—that Socrates is mortal—can be drawn.

The problem is immediate: the step from (a) and (b) to (c) is itself inferential and possesses the
MPP form. In order to have warrant for drawing conclusion (c), the reasoner would need to
possess an additional piece of knowledge, namely:

(i) (W): If an argument displays the logical form of MPP, then it is valid.

(ii) The argument above ((a), (b), (c)) is an instance of MPP.

(iii) Therefore, the conclusion (c)—that the conclusion (C) can be drawn—can be drawn.

This can go on indefinitely and constitutes an infinite regress of warrants. The challenge to the
epistemologist and philosopher of logic, then, is to establish a cogent apparatus for justifying our
entitlement to reason according to (this) basic rule of inference. I want to ask about an argument
that contains as conclusion the very claim that MPP is itself a correct rule of inference.
Intuitively, we seem to know that MPP preserves truth, yet the way we know this is not so
obvious, as evidenced by the regress argument above. I seek the knowledge that MPP is a correct
rule of inference and is truth-preserving in all cases, delivered via rigourous justification if
possible. Alternatively, I seek some sort of vindication for the intuition that MPP is correct. We
know that an argument with true premises that instantiates MPP is a valid one. We now seek an
answer to a how? question that stands one step prior to the application of a rule of inference, a
question that asks after justification of MPP itself which, if answered, can give us the knowledge
about MPP that we seek, viz., the substantiated knowledge that it is unexceptionally truth-
preserving, delivered via the force of an argument in a deductive proof, or via some other
convincing sanction of our intuition. The sort of knowledge sought is knowledge about logic
rather than knowledge gained by using it. Paradoxically however, it seems as though we might
be required to use logic in order to gain knowledge about it.
The problem of justifying our knowledge that MPP is a truth-preserving rule of inference seems
to lie in the fact that MPP is construed as basic. A principle of inference is basic logical
knowledge according to Hale if ‘...the explanation of how we have it makes no appeal to the
soundness of any principle of inference.’6 In other words, our knowledge that MPP is truth
preserving does not rest on our knowledge of any other rules of inference that allow us to arrive
at this knowledge. Under Hale’s definition, we therefore might be inclined to believe that the
justification of our knowledge that MPP is truth-preserving cannot be the product of some piece
of reasoning, for reasoning employs rules of inference, thereby violating Hale’s requirement.
What other approaches to justification of knowledge exist then that might allow us to justify the
knowledge about MPP that we seek to vindicate?

This is basic logical knowledge (BLK)—knowledge not only of correct forms of inference, but
also knowledge of the warrant that blesses these forms as correct. The apparatus we endorse for
the task must naturally avoid the problem of regress just explained, and it should not come to a
sceptical conclusion about this sort of knowledge. If it succeeds in its task, then the account we
ultimately adopt shall come about as the product of either an inferential piece of reasoning or a
determination of non-inferential composition. These are the two epistemological approaches we
employ to arrive at knowledge, and any proffered solution to my enquiry will be grounded in one
of these two strategies.

1.3 The Usefulness of Deduction and a Slice of its History

Philosophers have disputed the usefulness of deductive inference in the past, and my interest for
this section is the difference in attitudes maintained by Kant and Frege towards deduction. The
brunt of the difference pertains to disagreement over the fruitfulness of analytic judgments.
Briefly, an analytic judgment occurs according to Kant, when some predicate B is judged to be
covertly contained in the concept A.7 The statement ‘All bodies are extended’ is an analytic one
for Kant because the concept ‘body’ possesses extension as an essential property.8 Conversely,
the statement ‘All bodies are heavy’ is not analytic, but rather synthetic, for ‘weight’, according
to Kant, is not essential to (and thus not covertly contained in) the concept ‘body’. A synthetic
judgement is one such that predicate B lies outside the concept A. For Kant therefore, analytic
judgments clarify and synthetic ones amplify. Though both sorts of judgments can be inferred a
priori at least for some propositions. Deductive inferences that arrive at conclusions of an
analytic nature are less interesting for Kant because he held that inferred conclusions are simply
reiterations of that already contained in the premises.

Frege’s understanding of analyticity differed from Kant’s. An analytic statement for Frege is one
that follows from the fundamental laws of logic. He holds that we can establish these
fundamental laws in the following manner: He distinguishes the content of a judgement (such as
the concept ‘extension’ being contained in the concept ‘body’) from the justification for making
the judgement.9 The latter is ‘...a judgement about the ultimate ground upon which rests the
justification for holding it to be true.’10 We attain the ground by commencing from deductive
inferences and working backward.11 This process demands that a proof for each justificatory step
be demonstrated, and if it is possible to arrive at primitive truths (which are general logical laws
and definitions according to him) in this reverse-forensic manner, then the truth is analytic. So
we have just seen how Frege thought ‘the ground’ could be established, viz., by deducing the
fundamental laws of logic and nominating accompanying definitions. Now to arrive at analytic
truths, we do not use the reverse-forensic process just described, rather, we apply the basic rules
and definitions to arrive at particular analytic judgments. The following quote from The
Foundations supports this method and sheds some light on how analyticity is to be preserved:

Often we need several definitions for the proof of some proposition, which
consequently is not contained in any one of them alone, yet does follow purely
logically from all of them together.12

With an understanding of what it means for a statement to be analytic according to Frege we can
now explain why he held some analytic propositions to be fruitful—that is, why they extend our
knowledge rather than merely clarify it as Kant believed. Frege distinguishes between the terms
‘function’ and ‘argument’. A function is a statement containing a constant component ‘...which
represents the totality of the relations.’13 An argument signifies the object that stands in those
relations. In other words, the argument is replaceable by a variable. Therefore, in the proposition
‘David has reached the age of majority,’ ‘David’ is the argument, and ‘reaching the age of
majority’ is the function. Dummett14 observes that functions arise in one of two ways according
to Frege’s ontology. The first is by the identification of a pattern (what I have called ‘form’ in my
discussion of arguments in the previous sections) that stays constant regardless of what argument
is inserted. The other way in which a function can be distilled, applicable to such concepts as
shape, number, and direction, is by abstraction of concepts from complex predicates which are
then transformable into definitions by way of simple predicates. The example below consisting
of (1.), (2.), and (3.) will clarify what this means. Frege believed that the sense of a complex
predicate (viz., what the complex means) could not be entirely derived from its components,
which is why he endorsed a strategy that dissects a proposition into components rather than build
it up from them. This strategy agrees with the ‘backwards method’ of Frege’s later writings
briefly mentioned in the preceding paragraph. This is also our first clue that provides insight into
Frege’s ontology, where, by supposing that a complex designates something ‘richer’ than the
sum of its constituent components is capable of expressing, we see that the complex extends the
content articulated by its own constituent components. By means of example Dummett explains
Frege’s reasons for preferring dissection over the ‘building up’ of a concept;15

(1.) Consider the proposition ‘Either Jupiter is larger than Neptune and Neptune is
larger than Mars, or Mars is larger than Neptune and Neptune is larger than Jupiter.’

(2.) We extract the predicate ‘Either Jupiter is larger than x and x is larger than Mars, or
Mars is larger than x and x is larger than Jupiter’

(3.) We thus attain the concept ‘intermediate in size between Jupiter and Mars.’

But if we are to regard the logical connectives ‘and’ and ‘or’ as primitive (or ‘basic’ in the
terminology of this project), then they cannot be explained in terms of how they function in a
complex predicate. They must be explained rather in terms of how they function when acting as
main connectives. Complex predicates are therefore only to be understood such that they are
extractable from sentences such as (1.). Such an extraction is exemplified by (2.), and (3.) is a
new simple predicate by definition.
 ‘...its [viz., the complex predicate’s] sense may be seen as being given as a function
carrying the sense of the name Neptune [or Venus, or some other body] on the thought
expressed by [(1.)] ...We can regard it as such a function only because we already
understand the complete propositions; it is in grasping their contents that we directly
advert to the meanings of the connectives ‘or’ and ‘and.”16

Dissection of a proposition is therefore counter to the process of ‘building one up,’ where in the
latter, we incite its constituent components. We don’t need to acknowledge the pattern of the
complex predicate (exemplified by (2.)), instead, all that is required is to be able to understand
the subsentences of the complex and the meaning of ‘and’ in order to arrive at our definition (3.).

Dissection is therefore justly described as a process of concept-formation: it reveals something

new, one pattern among many discernible in the proposition and shared with it by others, but not,
in general, intrinsic to a grasp of its content.17

The ‘richness’ I adverted to earlier is what our understandings bestow upon a complex predicate.
It is an aspect of the content over and above that offered by its constituent components. When we
acknowledge a complex predicate, we exceed the content offered by the components of that
predicate (or the content offered by the premises in the case of a whole argument), and in this
way what has been derived extends our knowledge.

Why discuss analyticity in a study of deduction? Having surveyed two opposing views about the
fabric of deduction, one being Frege’s claim that deductions leading to analytic judgments
advance our knowledge, the other view belonging to Kant who claims that analytic judgments do
not, we should want to confront Kant’s charge. One way in which Frege interprets Kant’s
ontology is by means of analogy. Frege describes the Kantian view as seeing deductive
inferences follow from premises ‘as the beams are contained in a house.’ Whereas Frege sees
them as ‘plants are contained in their seeds.’18 My view is that complex predicate (1.) above is
not the only one that can be dissected to arrive at definition (3.). There are probably an infinite
number of complex predicates that can be so dissected to arrive at that exact definition, and
conversely, we should be able to dissect (1.) and arrive at some definition other than (3.) as well.
In the present case, the choice to dissect (1.) in the particular way expressed in order to arrive at
(3.) constitutes something extra— something over and above the content expressed in (1.); ‘here,
we are not simply taking out of the box again what we have just put into it.’19 Frege therefore
regards both form/pattern recognition and dissection as fruitful, and the consequences so derived
advance our knowledge.

On this much Kant and Frege were in agreement: deductively valid arguments deliver a new item
of knowledge—an unobserved conclusion— that was purportedly unobvious until a rule of
inference was applied. They also agreed that a priority was a quality that a deductive argument
must possess. A conclusion must follow necessarily from the premise(s) of an argument, and the
argument must be universal. The condition of universality as it relates to the a priori is as
follows: where the conclusion of an argument is deduced beyond the character of our experience,
the universality of the way we come about that deductive inference takes the form of argument
(3.) above. We are permitted to countenance it as universal because in talking about its form
(once again, exemplified by (3.)), we have established that every instantiation of its form is valid,
rather than pointing to a subset of particular instances as evidence for its validity. Divorcing the
notion of particular instantiation (viz., experience) from any circumscribed argument form is
essential to claiming the right to countenance that argument form as universal.

I conclude this section with emphasis on the notion that deduction is useful. Frege’s arguments,
accompanied by Dummett’s analysis give strong reasons for endorsing the fruitfulness of
deduction. If we were to go with Kant on this matter and suppose that only the synthetic is
capable of bearing fruit, then we would have to possess an unreasonable amount of background
knowledge. For consider: If the statement ‘all bachelors are unmarried men’ is an analytic one
for Kant and tells us nothing new, then we would have to know ALL the synonyms for
‘unmarried man’ that also make the statement analytic. For Kant’s story to work, the thesauruses
of our minds would have to be unabridged. We would then be able to say that synonymies are
unfruitful because we would already possess knowledge of all concepts. To put the case in other
terms, we would have to know the meanings of all words in our language in order for us to
already possess all the knowledge that is being deduced. Knowledge of such synonymies is
something we just don’t have. In light of persuasive arguments for the fruitfulness of deduction,
we should therefore acknowledge that it is a valuable practice and that every effort to understand
its foundations, most notably a defense of the correctness of its basic rules of inference, should
be pursued.

1.4 The Problems Facing BLK

I have thus far defined BLK as the sort of knowledge we might possess about rules of inference.
The knowledge that interests me is about the correctness of the basic rule MPP. In order for us to
have this knowledge, we either (a.) suppose it to be knowledge in one of the non-inferential
forms previously mentioned, or (b.) we attempt to demonstrate via inferential justification that
the inference rule is correct. Both of these accounts are problematic however, and in this section
I explain why.

If we attempt to provide an inferential justification for our entitlement to reason according to

MPP, then we must offer up an argument to that effect. An argument, as we have seen, employs
inference to arrive at its conclusion. But an argument that defends the validity of a basic rule of
inference must make use of some rule of inference in order to arrive at that conclusion. The basic
rule we are attempting to defend would cease to be basic by virtue of being justified by
something (viz., another rule) more basic than it. This leads to an infinitistic view of justification.
Carroll’s message presented in the introduction is similar but not identical. Carroll does not
argue for a regressive chain of increasingly basic rules, rather, he argues that a regressive series
of applications of MPP ensue from any attempt to justify it. This will be discussed to greater
depth in sections §1.5 and §2.1. As I will show in those sections, there is an additional
complication in the case of a justification for our warrant to reason according to MPP, because
the only rule available to us for that justification is MPP itself. This will highlight the problem of
circularity: the use of a rule of inference in an argument that defends the validity of that very
same rule.

A non-inferential account of a justification for our entitlement to reason according to MPP does
not rely on inference. Instead, it finds its grounding in stipulations that are informed by, inter
alia, beliefs or intuitions. This account, some argue, can lead directly to regress; if our
entitlement to reason according to MPP is grounded in a stipulation (for example), what
justification exists then for our entitlement to countenance that stipulation as correct? The next
level in a regress never answers this question in a satisfactory way, always requiring some sort of
‘reinforcement’ one level beyond. But that is not even the most important shortcoming of the
non-inferential account. Without a deductive demonstration showing how we come to adopt the
stipulation as correct, our endorsement of that stipulation is less than rigourous; its correctness is
not demonstrable by some formal method governed by rules. A stipulation tells us that
something is the case, but not how it comes to be so.

The final problem in need of attention pertains to the possibility of justification altogether. Upon
evaluation of the problems facing inferential and non-inferential accounts of our entitlement to
reason according to a rule of inference, a stalwart sceptic would dismiss both citing their
problematic natures just discussed as his reason. Instead he would defend the thesis that no
satisfactory justification exists. Is this a reasonable position? How would we have to change our
attitudes about deductive logic if we were to accept the position that the basic rules of inference
we endorse are unjustifiable? The sceptic introduces a legitimate challenge to the possibility of
the sort of justification we seek. However, accepting the sceptical thesis is tantamount to defeat:
it is diffcult to accept that no proof exists for the correctness of our logics and basic
epistemology, for, if we cannot justify basic logical rules, how may we declare with certainty
that our science of deductive reasoning is sound? Scepticism is a thesis that we should refuse to
accept. Once again, intuitively, we feel as though we cannot be wrong about the soundness of
deductive logic. The sceptic’s thesis is self-undermining and therefore indeterminate, for, in
order to support his thesis that MPP is unjustifiable, an argument employing MPP must be
presented, which commits the sceptic to its correctness. In this paradoxical case, the sceptic’s
endorsement of correctness is inevitable because it would be unreasonable for him to apply a rule
of inference in an argument that he deemed unjustifiable or even worse, incorrect. If the sceptical
thesis is vindicated, then so will the thesis that logical norms (such as the necessity of MPP) do
not exist. Without them we no longer have a science of deductive reasoning—of logic.
Assessment of the validity of arguments thus becomes a bankrupt practice.

1.5 Carroll

Lewis Carroll’s 1895 paper What the Tortoise Said to Achilles shows that any application of
MPP is inferentially justified only by appeal to an additional application of that same rule. The
result not only leads to infinite regress, but also highlights a diffculty present in any attempt to
justify a rule of logic using that same logical framework. What follows is a truer iteration of the
Tortoisean regress argument. Despite presenting a variation of it in §1.2, a rendition of its
original form is appropriate here because of its historical importance, and also because I will
refer to specific aspects of the argument throughout this thesis. Carroll’s story begins with
Achilles and the Tortoise discussing their race. The Tortoise piques Achille’s interest by boasting
of being able to demonstrate a sort of race that ultimately has no end. The proverbial race, of
course, is explained in such a way as to mimic the path a logician might take in her justification
of our entitlement to apply a rule of inference. The Tortoise presents a Euclidean-inspired
argument about a triangle that applies MPP as a rule of inference:
 (A) Things that are equal to the same are equal to each other.

(B) The two sides of this triangle are things that are equal to the same.

Therefore,

(Z) The two sides of this triangle are equal to each other.

Achilles initially believes that the argument arrives at its conclusion after only two steps (two
premises), but the Tortoise holds otherwise; she distinguishes between thinkers who might hold
one of several attitudes towards an argument taking that form:

1. The thinker who believes in the logical necessity of the truth of Z following from true
premises A and B in a valid argument.

2. The thinker who regards the hypothetical sequence as valid, viz., someone who
endorses the validity of an argument employing the MPP rule of inference, but who does
not yet accept the truth of A or B.

3. The thinker who accepts the truth of the premises A and B, but does not accept the
hypothetical {(A ∧ B) → Z}.

Carroll wants to make an argument for the case of thinker (3.), and his Tortoise defies Achilles to
make her accept Z. Achilles responds by entreating the Tortoise to accept statement (C):

(C) If A and B are true, Z must be true.

But (C) is another hypothetical; we are therefore left with a new argument:

(A) Things that are equal to the same are equal to each other.

(B) The two sides of this triangle are things that are equal to the same.

(C) If A and B are true, Z must be true.

Therefore,

(Z) The two sides of this triangle are equal to each other.

But the Tortoise steadfastly resists acceptance of Z—she refuses to accept the consequent of the
hypothetical, whereby Achilles argues that if the Tortoise accepts A and B and C, then she must,
by logical necessity, accept Z. However this last argument by Achilles involves another
hypothetical of the form,

(D) If A and B and C are true, Z must be true.

So long as the Tortoise fails or refuses to acknowledge the truth of Z which, being unconvinced
of the correctness of MPP, she is under no logical obligation to accept, justification for Z must
appeal to a regressive series of hypotheticals. The next, and obvious step in this series would be:

(E) If A and B and C and D are true, Z must be true.

This argument persists ad infinitum. Pascal Engel20 proposes several morals that we might draw
from Carroll’s paper. One moral that is relevant to this enquiry is that a paradox arises when we
countenance a rule of inference, in this case MPP, as a premise. If we fail to see the distinction
between premises and rules (and subsequently treat the latter as having the same status or force
as the former), then it is conceivable that someone—an inferentialist—might never be able to
grasp the necessity of conclusion Z following from premises A and B. The same applies for the
necessity existing between A, B, C, and conclusion Z, and so on. If we always preserve the
distinction between rules, premises, and conditionals, then the regress cannot start because the
regress of hypotheticals simply does not arise. In this case a non-inferential justification for our
warrant to use MPP would be required, a justification that does not call upon an argument in
order to demonstrate validity. The idea behind the infinitely regressive nature of an inferential
justification for MPP is therefore that any axiom, rule, or principle we adopt and hold to be valid
is always subject to the sceptical challenge that asks us to justify the decision we have made to
regard the rule as correct. This challenge will always ask after a level of justification higher than
the one at which we are presently working. In this way the regress conceivably never ends, or
results in circularity.

1.6 Agenda for Resolution of the Problem of Warrant

Several problems confront attempts to justify our knowledge of basic rules. The Carroll story
shows how infinite regress ensues from an inferential justification of MPP. Regress is
unsatisfactory because—roughly—we never arrive at some grounded principle whose
correctness we can be certain of. That principle is important; upon it we could demonstrate our
entitlement to reason according to a given rule of inference, and subsequently effect valid
deductive inferences culminating in conclusions that follow necessarily from their premises. This
last sentence describes the objective of the inferentialist approach.

Further scrutiny of inferential justification also reveals that circularity ensues from any attempted
justification of MPP by deductive argument. This is because a deductive argument that concludes
that MPP is correct must use MPP in its reasoning. If we wish to vindicate an inferential
approach to the justification of our entitlement to use basic rules of inference in deductive
reasoning, we shall have to find a strategy that does not succumb to regress or circularity.
Michael Dummett proposes a solution available to the inferentialist in the form of pragmatic
circularities to be discussed in detail in the following chapter. This proposal improves the
prospects for the inferentialist campaign, and figures in numerous solutions proposed by other
philosophers, including that of Boghossian also to be discussed.

Non-inferentialism countenances justifications for rules of inference that are not themselves
arguments. These sorts of justifications, as we have seen, include stipulative definition, sense
experience or even an intuition. They seem less satisfactory as grounding principles however
because of the lack of support available to corroborate them as correct, support such as a formal
argument. Sense experience is not inter-subjective, and therefore it is impossible to codify a
justification for it in universal fashion. Intuition is the very practice we seek to explain (viz., our
intuition to countenance certain rules of inference as correct), and therefore appealing to intuition
in order to justify it leads to circularity. And stipulative definition without justification is
arbitrary, regardless of how effective and plausible the definition is. How then do we satisfy
someone seeking greater explanatory power in a justification than the solution proffered by the
non-inferentialist? It might seem as though an inferential justification, one that is a product of
reasoning, better satisfies the criteria for justificatory rigour. Consider the non-inferential
stipulation: ‘MPP is a correct and truth-preserving rule of inference,’ and let us attempt to
compose an argument in its defense;

P1. If a rule of inference is correct and truth-preserving, then it facilitates the

inference of a valid argument.

P2. MPP is a correct and truth-preserving rule of inference.

Therefore,

C. MPP yields valid arguments

This simple argument seems to show that the non-inferentialist’s supposition that MPP is correct
may be even more deeply justified by argument—by reason— in an inferential context. The
ultimate goal of any justificatory account of rules of inference would be the presentation of a
justification in a chain of reasoning that terminates. We see that the non-inferentialist is satisfied
with justification that begins and terminates at once in definition, intuition, or sense-experience,
but bythe most recent argument above, the inferentialist account seems to be capable of
justifying MPP more ‘deeply’. Now what would a terminated justification look like? Should it be
presented in inferentialist or non-inferentialist fashion, and what are its properties? Intuitively,
we should seek a justification for which complete justificatory rigour exists, if possible.
Ultimately, this ends up being the yet unachieved objective of all philosophers mentioned in this
thesis, not to mention my own.

Rules of inference lie, more-or-less, at the beginning of an evidential chain. Again, what strategy
do we have at our disposal that guarantees that MPP is a correct rule of inference? What, in other
words, would an argument consist in, that has as conclusion the very claim that MPP is a correct
rule of inference? The next chapter will present and evaluate several proposed solutions to the
challenge of justifying our entitlement to reason according to basic rules of inference.

Chapter 2

Treatment of the Problems

2.1 Tonk

Carroll’s Tortoise lays a large obstacle in the inferentialist’s path. Achilles’ attempt to justify a
rule of inference by argument is hampered by the fact that he cannot do so without the
consequence of regress in that argument. The alternatives to justification by argument, described
above, may seem less palatable in light of our interest in justificatory rigour, but this non-
inferentialist strategy of justification does carry with it results that are worthy of consideration.
By regarding the rules governing deductive validity as analytically correct, and at the same time
holding that those rules are conventions that emerge based on the language(s) we use, we
improve the prospect and plausibility of logical rules being arbitrary; ‘arbitrariness’ is effectively
what a stipulation connotes.

A statement is analytic when its truth can be determined solely by the meanings of the terms
contained in that statement. Quine’s example No unmarried men are married is the archetypical
analytic statement in philosophical literature. It is analytic because no matter what the non-
logical elements of the sentence represent, we can determine the truth of a sentence by what its
logical structure forces upon us. Quine’s example can be converted into the universal statement
No non-xs are x, which is a tautology—a logical truth. An analytically valid inference is one that
arrives at its conclusion by respecting prescribed meanings of the rules of inference and
operators involved. These meanings are expressible in terms of introduction and elimination
rules which explain when and how we are permitted to apply the logical term in deductive
inference. An analytically valid inference is simply an inference that has arrived at its conclusion
by respecting the meanings of the logical terms, that is, by conducting inference in a manner
sanctioned by rules telling us how a logical term may be used in inference. In The Runabout
Inference Ticket, Arthur Prior asks whether the notion of an analytically valid inference is
corrupt, based on his demonstration that ‘...any statement whatever may be inferred, in an
analytically valid way, from any other.’1

He observes that in inference involving conjunction, the validity of that inference is buttressed
‘solely [in] the meaning of the word and’. He demonstrates by applying its elimination rule to
the following inference:

Grass is green and the sky is blue;

Therefore,

Grass is green.

Furthermore, Prior holds that we may deliver a complete (or all-encompassing) account of the
meaning of the word ‘and’ by enumerating that operator’s introduction and elimination rules,
thereby specifying the role this operator plays in our inferences:

i. from any pair of statements P and Q, we can infer the statement P-and-Q.

ii. from any conjunctive statement P-and-Q we can infer P.

iii. from any conjunctive statement P-and-Q we can infer Q.

Prior then points out that based on this account of analytic validity, we can generate all sorts of
absurdities in an analytically valid way; for example, the statement ‘2 and 2 are 4, therefore 2
and 2 are 5.’ He accomplishes this with the following argument:
 (P1.) 2 and 2 are 4.

(P2.) 2 and 2 are 4 tonk 2 and 2 are 5.

Therefore,

(C.) 2 and 2 are 5.

Tonk, being a hitherto unknown operator, takes its complete definition in similar fashion to the
definition of ‘and’ above using the following introduction and elimination rules:

i. from any statement P we can infer any statement formed by joining P to any
statement Q by tonk (which compound statement we hereafter describe as ‘the
statement P-tonk-Q’).

ii. from any ‘contonktive’ statement P-tonk-Q we can infer the contained statement Q.2

Prior sarcastically characterises this new logical construction as convenient, promising ‘to banish
falsche Spitzfindigkeit from Logic for ever.’ It is convenient and banishes false niceties from
logic because this operator as defined admits the inference of any statement from any another in
an analytically valid way. Naturally, we should not take this message to mean that Prior
subscribes to the thesis that analytically valid statements are corrupt. He acknowledges that as
response to his introduction of tonk, ‘more enlightened views will surely prevail at last...’3 We
find such an enlightened view in Belnap’s discussion of conservative extensions in the next
section.

So it would seem as though some analytically valid arguments are semantically legitimate, this
last term being characterised by arguments employing operators actually used in English and
extra-logical contexts such as the connective ‘and’. Some other operators however are
illegitimate, tonk being our prime example, where we find no such term in everyday parlance. If
we are to preserve and vindicate the notion of analytic validity, we should have to (a.) explain it
in such a way that it permits of all inferences we hold to be legitimate (such as those inferred
from conjunction-elimination), while at the same time excludes all inferences that have no
semantic relevance or value (such as the inference of ‘2 and 2 are 5’ from ‘2 and 2 are 4’); and
(b.) present analytic validity in a way that does not enable it to generate inconsistencies.
Furthermore, analytic validity in the more formal sense ensures that all statements expressible in
a system of logic are, programmatically, permitted by the rules and axioms, and are valid. There
is something about the nature of analytic validity that is not captured in the account articulated
by Prior, something that erroneously vindicates illegitimate operators such as tonk. If we can
identify what that is and incorporate it into a more restrictive account of analytic validity,
perhaps the notion may be preserved. Analytic validity is a goal worth pursuing because if it can
be made stable, viz., if it can satisfy our requirements for semantic legitimacy and avoid
inconsistency as mentioned in the previous paragraph, then the inferentialist can point to the
quality of being analytically valid as warrant for his entitlement to reason according to a rule of
inference. The introduction and elimination rules could conceivably lie at the onset of the
inference pedigree, and a possible Carroll-style regress might not arise in an analytically valid
scenario.

2.2 Tonk gets Belnapped

Let us recharacterise the most salient point of Prior’s argument. Tonk’s introduction and
elimination rules are expressible thus;

A |- A-tonk-B (tonk-introduction);

A-tonk-B |- A (tonk-elimination);

A-tonk-B |- B (tonk-elimination).4

By the transitivity of deducibility in a logical framework that recognises tonk as an operator, the
logical form A |- B follows. However A |- B does not express a universal truth, for if A is true
and B is false, then B does not follow from A. The tonk framework allows us to make this
contingent inference, and when B is actually false, we have a counterexample derived in an
analytically valid way. The crux of Prior’s point therefore argues that operators characterised in
an analytically valid way permit the valid derivation of falsehoods from true premises. But
Belnap shows his point to be mistaken, as I shall presently explain.

Belnap observes two options for defining connectives. The first option is to define them in terms
of the role they play in inference and with respect to deducibility. The second option is to deliver
some conception of what the connective means ‘independently of the role it plays as premiss and
as conclusion’.5 This second option might see, for example, the creation of simple truth tables to
establish the independent meaning of the given connective. Yet despite the problems caused to
the first option by tonk, Belnap nevertheless wishes to preserve that approach to defining
connectives. But how is this to be achieved without generating tonktive results? ‘How are we to
make good the claim that there is no connective such as tonk though there is a connective such as
and...?’6 What is required is a characterisation of a connective in terms of the role it plays in
inference such that when that connective is used in argument, only analytically valid inferences
that lead to sound conclusions are the result. Furthermore, when the connective is used in
deductive inference according to the rules governing its meaning, the conclusion inferred must
follow necessarily from the premise or premises of that argument.

The solution, explains Belnap, is to acknowledge that connectives that are defined using the first
option, viz., in terms of the role they play in inference, are not defined or conjured up from
nothing—ab initio, to use his words—but rather are defined in terms of an antecedently given
context of deducibility. In the recharacterisation of Prior’s argument at the beginning of this
section, we see that Prior has helped himself to one of the properties of deducibility—that of
transitivity. That is how he is able to validly deduce the argument form A|-B using tonk.7 But
nowhere in Prior’s characterisation of tonk are the rules for deducibility characterised; they are
presupposed, and constitute the antecedent context just mentioned.

When we deliver a formal characterisation of deducibility however, we see that introduction of

the tonk rules to that characterisation (viz., adding tonk as an extension to deducibility) renders
the newly conjoined system inconsistent. Belnap chooses Gentzen’s characterisation8 of
deducibility (hereafter D) to demonstrate his point. This characterisation is a sequent calculus,
one among several used to demonstrate proofs that apply certain styles of formal inference. In
this case, D was conceived (by Gentzen, who named it ‘LK’ for Logischer Kalkul) to study
natural deduction. All and only sequents that correspond to universally valid arguments are
provable in D, and it, in Belnap’s words, ‘completely determine[s] the context’ of deducibility.
The difference between sequents and statements is that sequents are better suited to express
forms of reasoning, whereas statements, as have been used in Chapter 1, are employed to express
particular premises and conclusions, viz., non-universal content.9 A sequent such as Γ |- Σ in D
expresses the intuitive notion that if every formula contained in Γ is true, then at least one
formula in Σ shall be true as well. This also means that Σ is provable in D. There were numerous
attempts to characterise natural deduction in the 1900s. Common to all these attempts was the
aim to minimise or eliminate the number of axioms used to characterise deduction, attempting
instead to formally model reasoning as it ‘naturally’ occurs.

Table of Belnap’s Presentation of G. Gentzen’s Characterisation of Deducibility (D)

Axiom. A ` A

Rules. Weakening: from A1,..., An |- C to infer A1,..., An, B |- C.

Permutation: from A1,..., Ai, Ai+1,..., An |- B to infer A1,..., Ai+1, Ai,..., An |- B.

Contraction: from A1,..., An, An |- B to infer A1,..., An |- B.

Transitivity: from A1,..., Am |- B and C1,..., Cn, B |- D to infer A1,..., Am, C1,..., Cn |-
D.

While D permits the sequent A |- A, it does not enable us to derive A |- B. However, as we have
seen, the addition of the tonk rules to D makes that sequent derivable. Since D already outputs all
and only the universally valid sequents expressible in that language, and since A |- B is not one
of those sequents, A |- B is inconsistent with D. This problem is the result of the tonk rules not
being a conservative extension of D.

An extension of a system (or language) is characterised by the introduction of operators or new

rules and/or axioms to that system, comprising an articulation of how the sequents deduced by
way of those extensions may be proved.10 A system might be extended so as to allow it to deduce
sequents that it was unable to previously, or to allow the system to deduce more thriftily that
which it was already able to in more cumbersome fashion. A conservative extension has the
following essential property: every new sequent provable by means of the extension must
include instances of the new operator introduced by that extension. When these conditions are
transported to contexts involving particular arguments, as in Chapter 1, we realise the result that
a conclusion follows from premises not including the new operator only if the conclusion
contains the new operator. The reason we require a sequent to contain a newly introduced
operator is that if it was already possible to prove some sequent without it, then it was never
necessary to introduce the new operator in the first place. It is illuminating to note as well that
were the characterisation of deducibility to augment the tonk rules instead of vice-versa, then
that would be a conservative extension because standard deducibility is (much) more restrictive
in terms of what can be proved than the tonk rules, in fact, anything follows from a rule that
permits the sequent A |- B. When the tonk rules extend the formal characterisation of
deducibility, the augmented system makes the sequent A |- B provable, violating the
conservativeness requirement: the sequent A |- B, deduced by means of tonk-elimination, does
not contain tonk, and was not expressible by the original characterisation of deducibility alone.

Belnap realised several items of importance in his reply to Prior. He describes a concise criterion
for deciding whether or not a newly proposed operator shall be problematic in the framework of
natural deduction. One aspect of that criterion requires that a newly introduced operator be
verified against an antecedent context of deducibility exemplified by (but not limited to) the
Gentzen characterisation above. The other aspect involves the conservativeness requirement: it is
impermissible to admit an operator to a system of natural deduction if that operator makes
possible the deduction of sequents that do not contain that operator and which are not provable in
the unextended system. Belnap showed how tonk violates this criterion by permitting the
deduction of a non-conservative, inconsistent sequent through the application of the rule of
transitivity. Belnap’s suggestion is that taking these points into account rescues the supposition
that correct inferences such as MPP are analytically valid and truth-preserving. He has shown
that tonk does not corrupt the notion of analytic validity where logical rules and operators are
defined in terms of the role they play in inference.

So might we regard analytic validity as capable of adequately characterising BLK in a way that
satisfies all of Hale’s criteria mentioned in §1.2? Not yet, one problem still plagues analytic
validity: If a logical operator is implicitly characterised by prescribing its introduction and
elimination rules, thereby delivering its meaning, once those operators are part of a deductive
context then there is nothing stopping a reasoner from using the very defining rule to vindicate
the defined operator. Suppose we were to adopt the introduction and elimination rules for tonk
presented at the beginning of this subsection. We could then put forward the following argument:

(P1.) 9 X 7 = 63,

(P2.) 9 X 7 = 63 tonk ‘tonk’ is a valid operator, (by tonk-introduction)

therefore,

(C.) ‘Tonk’ is a valid operator. (by tonk-elimination)

This argument is circular and certainly appears to lead to a problematic conclusion. It argues for
the validity of tonk while using that very operator in its reasoning. We know that tonk is an
unsound rule of inference, and that this argument is unsound, but it is no thanks to our test for
validity that we know this. For in a situation in which tonk is part of the deductive context, this
argument is valid. Were a similar argument put forth to demonstrate the correctness of MPP, it
would, unlike tonk, result in a conclusion of which we approve and regard as sound. But what,
then, distinguishes these two valid cases, each of which involves its own justification via a rule
(namely, itself) which is supposed to be part of the pre-established deductive context? The lesson
suggests that we cannot rely solely on analytic validity—mere acceptance of introduction and
elimination rules—as the strategy for establishing the correctness of those rules. There are two
reasons for this; (a) because analytic validity does not prevent the vindication of a rule by means
of a circular argument. And (b), it allows us to find validity in some patently bad rules of
inference.

As reasoners who admit MPP to deductive logic, we occlude tonk from deduction because of the
unfavourable results that emerge when we apply it. But again this is no thanks to our test for
validity. Furthermore, the disapproval we have for tonk is informed by our acceptance of MPP—
the latter is already part of our established logical context, and so when we evaluate those
unfavourable results that ensue from an application of tonk, we find the rationale behind an
adoption of tonk to be broken-backed. But suppose some reasoner accepts tonk in the same way
MPP is accepted. He might view the choice to accept MPP just as erroneous because when MPP
is applied in his ‘tonk-context,’ a whole bunch of statements that were previously expressible
(via tonk) in the system become invalid. And since the tonk-reasoner previously endorsed the
system that contains tonk, he would find any acceptance of MPP to unfavourably alter the
expressive power of his logical system. As we have seen, tonk enables us to infer any statement
from any other, while MPP is much more restrictive. So why do we accept MPP and reject tonk?
Analytic validity is an inadequate criterion for distinguishing bad from good rules of inference.
Furthermore, we are no better ahead in vindicating one rule of inference over another in the
framework of deduction solely on the basis of analytic validity. Other criteria shall therefore be
required.

2.3 Carroll, Regress, and Scepticism

It will be useful at this point to elaborate on the lessons drawn by Pascal Engel from his
examination of the Lewis Carroll problem discussed in Chapter 1. He identifies at least four
morals to be drawn from the Carroll paper:

1. The need to preserve the distinction between rules and premises;

2. The need to distinguish between propositional and practical knowledge;

3. The need to recognise and deal with an unavoidable circularity in any attempt to justify
a basic rule of inference.

4. Scepticism about the force of logical reasons.

Per 1. Engel expresses the view exhorted by many in response to the Carroll paper that regress
does not start if we prohibit rules from being countenanced as premises. Referring back to the
structure of the Euclidean argument, it is apparent that A, B, and C are treated equally as
premises leading to the conclusion Z. ‘Equally’ here means that they all have the same force in
their status as premises playing a specific (assertoric) role in a deductive argument. But observe
that A and B, whether countenanced as statements or sequents, represent assertions (in this case
pertaining to geometry), while C articulates a rule expressing the validity of the sequence. Engel
suggests that in a conditional argument, the major and minor premises, as well as the rule, are
each distinct in status. If we countenance the rule as something distinct, and subsequently
disallow rules from figuring in arguments in that way, then the regress cannot begin.
Let us flesh out this possibility proposed in moral 1. It suggests the prohibition of logical rules
from figuring as premises, thereby sidestepping the regress problem. In a typical valid deductive
inference, if the premises are true, the conclusion necessarily follows. Were the rule of inference
to figure in statement form (cf. statement C) and held to be true, then this would be an inferential
justification of our right to move from premises A and B to conclusion Z, thereby expressing an
argument that uses (and explicitly expresses) MPP as its rule of inference. C would
propositionally express how the deduction is valid. In other words, it would solidify the
relationship existing between A and B, forcing the conclusion Z. So what can we do when the
rule C remains unexpressed? What then could justify the procedure of this deduction? Dummett
explains his stance on the matter of the justification of deduction in the following quote.

This is the philosopher’s question. In everyday life, we do not wait upon a justification, or ask
for one. It is not rational to entertain any serious doubt about the matter; both thought and
discourse would break down if we attempted to eschew all deductive inference until a
justification of the practice was forthcoming, which it would never be, because any such
justification must involve some deductive argument.11

We can distill at least two messages from this quote. First, Dummett points out that in common
discourse, the justification of simple inferences just does not occur. If it were a requirement in
such a circumstance, communication would be difficult and awkward. The second message is
that by acceding that a justification for a basic rule of inference not be required, one must be
fairly certain and convinced of the correctness of the rule; for if not, why would its application be
permitted unchecked? This ties fittingly into another of his theses, to be discussed at length in the
next section, that in satisfying the philosopher’s curiosity about the correctness of the rule, a
curiosity that demands the ‘checking’ just mentioned, we must demonstrate that correctness by
argument, and that argument will plainly be circular. In the case presently considered, we are
examining the possibility of justifying a rule of inference. We may do so by applying the very
rule we seek to justify, but we may also apply some other rule to arrive at our proof. While the
justification of a rule of inference by means of the application of some other rule may not
initially seem to be circular, we are reminded that our concern is for basic rules of inference such
as MPP, and so any non-basic rule should conceivably be justifiable by a basic rule after the
requisite number of line-items in a proof. Once again, the next section will examine Dummett’s
theory explaining exactly when we might be permitted to champion a circular argument as a
justification for a rule.

Leaving Dummett’s work aside now, what other possibilities exist, aside from circular or petitio
principii arguments, that might serve to justify the correctness of rules of inference? We recall
that we are presently attempting to do this in accordance with moral 1, that is, without allowing
the rule to figure as a premise. The thesis that rules of inference are ‘self-justified’ is therefore
also worthy of consideration, and countenancing a rule as such has a few advantages. It avoids
circularity because no argument is required to vindicate correctness. It avoids regress because it
is held to be a bottom-certitude of sorts. Engel recounts Dummett’s suggestion that a rule is self-
justified when it satisfies the requirement of harmony12 and conservative extension, and that such
a proposal is among the most forceful responses to the Tortoisean challenge.13 Dummett’s
solution is challenged however by the responsibility to correlate a rule that is self-justified with a
rule that is conservative and in harmony. Our entitlement to countenance a rule as self-justified
will depend on a reliable method to verify correctness without using an argument containing an
expression of the rule’s correctness among its premises.

Per 2. The Tortoise’s main argument is that no thinker of any persuasion is under obligation to
accept the necessity of the claim labeled C of the Carroll argument: if the premises of an instance
of MPP are true, so is the conclusion. The problem resulting from her puzzle shows that the sort
of knowledge required to carry us from premises to conclusion is not the sort of knowledge that
can be expressed in a proposition—it is not a form of knowledge that or propositional
knowledge, but rather a form of knowledge how, or practical knowledge. So far, it does not seem
that practical knowledge is the sort that can be expressed in a proposition. Some philosophers
attribute the sort of knowledge required to move from premises to conclusion to a faculty of the
understanding. Comprehension of how MPP works and that it facilitates valid inferences is a
function of that faculty— it is possession of the practical knowledge required. To know ‘if P
then Q’ is to know that when we know P we are able to ‘see’ the consequence Q. If someone
were to know that P, yet be unable to see Q, then that would be what it is not to know that ‘if P
then Q’. This moral therefore sees the Tortoise arguing that a particular type of understanding is
required to apply MPP in inference, and that application of it is not the sort of move that can be
articulated by a proposition. Practical knowledge of MPP’s necessity is therefore something very
different from an item of propositional knowledge which attempts, yet fails, to express MPP’s
necessity.

Per 3. We have already seen indications of the likelihood that circular arguments are going to be
involved in the justification of BLK. Engel presents an argument that this is unavoidable. Given
the infinitude of instances of MPP and our ability to acknowledge only a finite number of them,
our ground for logical truths cannot simply lie in a finite number of conventions because then we
would be using a rule of inference to derive truths from only finitely many conventions; For
grasp of the rule involves grasp of its applicability to the infinitely many instances. Recognising
that a particular argument is an instance will involve reasoning akin to:

P1. This argument has the form A → B, A, therefore B;

P2. Any argument of the form ‘A → B, A, therefore B’ is an instance of MPP;

therefore,

C. this is an instance of MPP.

Of course, if we recast this argument in another form, say as a disjunctive syllogism, the problem
simply becomes one of the circularity of the justification of that rule, which would then be basic
in place of MPP. This would mean that we would be using logic to derive logic from convention.
In Engel’s words, ‘...if we want to explain the nature of logical truths or rules from the existence
of conventions, we must presuppose these logical truths and rules to derive them from
conventions, and our derivation is thus circular.’14

Per 4. Unlike the other cited morals where the Tortoise has a grievance with premise C, viz.,
MPP in statement form, Engel considers the possibility that the Tortoise endorses C, but does not
accept the conclusion Z, or more specifically, does not accept the logical force or necessity that
takes us to the conclusion Z. Some cited reasons for this refusal are ignorance, stubborness,
recalcitrance, akrasia, and outright scepticism. If the reason for the Tortoise’s refusal to accept
the logical truth C is actually that she does not understand C, then this brings us back to moral 2
which discusses the epistemology of understanding. If she actually does understand C but still
refuses to accept its logical force, then we can label her as a sceptic about the force of logical
reasons.

Scepticism can give rise to some philosophically damaging results. Moral 4. seems to show that
the position of the stalwart sceptic leaves us at an impasse in the task of justifying our warrant to
apply MPP in either inferential or noninferential fashion. The sceptic argues that there does not
seem to be any attitude, disposition, principle, or rule that allows a reasoner to bridge the gap
existing between premises and conclusion inferred by means of conditional elimination. In other
words, practical knowledge for our warrant to reason according to MPP is unavailable or non-
existent, and warrants expressed propositionally engender regressive results. Our discussion of
Dummett’s treatment of this problem beginning in the next section will figure integrally in
opposing this sceptical claim.

The final comment I would like to make about MPP in this section pertains to the integral way in
which it figures into any logical system. Above I referred to paraconsistent logics that exclude
MPP as one of its admissible rules of inference, and observed that for any valid rule actually
admitted by a logical framework, we can offer up a deductive argument that employs MPP in
order to assert that rule’s validity. It’s not much of a feat:

(P1.) If rule x is one of the rules of inference of language y, then it is valid,

(P2.) Rule x is part of y,

therefore,

(C.) Rule x is valid.

The paraconsistent logician, who does not admit MPP as a rule in her system, cannot deny this
piece of reasoning which employs MPP to arrive at its conclusion. It has figured into her system
whether she likes it or not. In this way MPP plays a role similar to transitivity in Gentzen’s
characterisation of deduction; Prior overlooked transitivity in his characterisation of tonk.
However it was implicit or unavoidable as a constituent rule of deduction. MPP seems to possess
that same status: we know it belongs in any deductive system, and it seems as though it belongs
to any consistent logic as some sort of an antecedent or implicit context, even in cases where
MPP is programatically excluded as a permissible rule as in the case of paraconsistency. So
whence comes this universality or unavoidable applicability of MPP? In the section focusing on
Hale’s work below we will see that he argues that use of MPP in a consistent system is
unavoidable. He takes this finding and uses it as a foundation for a non-inferential account of our
knowledge of the correctness of basic rules of inference.

Since any argument about the truth, meaning, or validity of a particular rule can be framed in
terms of MPP, and any argument that attempts to explain the validity of MPP has the problems
named by the morals in this subsection, then perhaps a special status needs to be assigned to
MPP that considers our unreflective intuition that it is valid, despite our inability to offer up
proof of its necessity. One argument for the special status of MPP might be that we are
predisposed to reason in that way. This could mean one of at least two things. First, that our
reasoning faculties are ‘hard-coded’ to behave in such a way that propels us to arrive at
conclusions using a strategy such as MPP. This argument is weak though, because we are at a
loss to conclusively explain what this ‘hard-codedness’ would consist in. Even more problematic
is that were we actually to characterise this hard-codedness, there is no guarantee that the hard-
coded rule is actually correct—we should only be able to point to its correctness in an inductive
way, that is, by pointing to as many instances of our application of this rule that have helped us
deduce sound and consistent results. Presently, as we come closer to understanding the nature of
quantum science, the hypothesis that the principle of bivalence, for example, is not a universally
applicable principle becomes plausible. If ever the mind was hard coded with a rule, the principle
of bivalence should qualify as one, yet quantum science shows that this principle should be
called into question. The second argument might be that given the way our languages and
histories of thought have unfolded, the status of MPP as the underived rule of inference might
have evolved as their product. To this we might respond with an analogy to Neurath’s ship, such
that the way we reason could have turned out to be otherwise had our ascent into language and
epistemology a few millenia ago unfolded differently. This might shed some light on the
necessity of MPP, such that if a language system could be developed by a sufficiently large
group of people who know nothing about language or theory of knowledge, people who have
essentially been subjected to the conditions described in Plato’s cave in book IX of The
Republic,15 then we might gain some insight into the necessity of MPP insofar as that rule of
inference would either emerge as a natural output of their linguistic and reasoning practice or it
would not—and this would be useful knowledge gained about the necessity of MPP. However, to
begin with a veritable tabula rasa in some linguistic community in an attempt to see if MPP
would figure integrally in some newly-contrived system in the same way that it has in our current
system(s) would be impossible. I doubt such an attempt could even be conducted via thought-
experiment because implicit and unavoidable usage of concepts that we already possess will be
involved in, and thus inform the outcome of our enquiry. The subjects of such an experiment
would truly have to be non-lingual for such an experiment to be possible.

MPP figures integrally in our reasoning processes and seems to enjoy a special status as a rule of
inference. We have run into difficulty trying to come up with a justification for our entitlement to
reason according to MPP, and appeals to the make-up of our minds or of our linguistic practice
fall short of providing the justification we seek. Now does moral 1 propose a potential solution to
the Carroll problem? Did he intend that the readers of his paper distill the important distinction
between rules and premises from Achilles’ and the Tortoise’s banter? The rules-are-not-premises
will have an important place in my concluding findings presented in Chapter 3.

2.4 Circularity, Consistency, Harmony, Scepticism

The way in which the Carroll story is laid out shows how premise C is required to justify {(A ∧
B) → Z}, and then how premise D is required to justifty {(A ∧ B ∧ C) → Z}, and ultimately how
this leads to an infinite regress of justification that seeks to, but never succeeds in vindicating the
conclusion, Z. A regress of this sort is fatal to the logician’s task of offering up a justification for
our entitlement to reason according to this rule where he seeks to demonstrate the rule’s
correctness. Perhaps then, a circular approach to justifying MPP should be reexamined to see if a
plausible justification for this rule of inference can be exacted from a strategy, viz., circularity,
which we were quick to dismiss earlier on.

Dummett begins his enquiry into the possibility of circular justifications from the important
observation that there really is no reasoning with or proof available as response to the stalwart
sceptic. Appealing to the Carroll morals (3.) and (4.) of the previous section, the sceptic argues
that there is no satisfactory justification to be had for MPP as a rule that preserves truth. But this
denial of the possibility of justifying MPP is a categorical statement that can be turned into an
argument framed in terms of MPP. The sceptic would have to acknowledge our rule in order to
proffer the conclusion he bolsters, which renders the sceptic’s view self-defeating.

The crux of scepticism is relevant to Dummett’s work in metaphysics because there16 he is

concerned ultimately to defend intuitionistic instead of classical logic as the logic. He champions
such a defense by arguing that classical logic is informed by metaphysical presuppositions. For
instance, when we consider an object, and suppose that there is a fact-of-the-matter about any
statement that mentions it, we are taking the metaphysical stance of realism. In this case classical
logic is the accepted framework we use to reason about the members of this domain. But an anti-
realist stance about a particular domain emerges from an adherence to intuitionistic logic for that
domain; if it is regarded as the logic of choice to discuss the objects in some domain, then we do
not, for example, admit the principle of bivalence to figure in our reasoning —for any
proposition P, either P is true or P is false, for the nature of this principle implies a metaphysical
stance. Realism is supposed to mirror a correspondence to reality exclusive or regardless of our
access to a domain. Realism also entails belief in a fact-of-the-matter about any proposition of
the domain in question. It is a presupposition of which intuitionism in mathematics arguably
dispenses, and Dummett extends the lesson to non-mathematical domains. A sceptic
(categorically) denies the possibility of justification thereby suggesting a fact-of-the-matter in
itself and thus proffers a realist thesis. Intuitionism repudiates the fact-of-the-matter way of
reasoning and is therefore not amenable to the realist metaphysical stance (of bivalence for one)
presupposed by classical logic. The categorical statement made by the sceptic agrees with a
realist attitude about justification, and so in the broader context of Dummett’s entire book (and
life’s work for that matter), which seeks to defend intuitionistic logic as the logic of choice, one
that attempts to circumvent metaphysical presuppositions, the reader evolves to understand how
the sceptic’s thesis is dissatisfying and untenable according to Dummett.

Continuing with his enquiry into the possibility of the existence of a plausible circular
justification for our entitlement to reason according to MPP, Dummett distinguishes circularities
of two sorts:

Gross Circularity. A grossly circular argument contains among its premises the very claim
sought to be proven in the conclusion. In the case of justifying a basic rule of inference such as
MPP, one of the premises in a deductive inference will be the expression that MPP is valid. Such
an argument might resemble the following:

(P1.) MPP is a correct rule of inference. (Premise)

(P2.) MPP is a correct rule of inference or the sun sets at 18h03.17 (Or-introduction)
 Therefore,

(C.) MPP is a correct rule of inference. (Or-elimination)

The principal issues facing gross circularities however touch on the fact that a grossly circular
argument assumes what it seeks to prove, and subsequently, by assuming what it seeks to prove,
such an argument can prove anything, even falsehoods or inconsistencies. The gross circularity
manifests in the fact that the conclusion is identical to one of the stated premises, and, in this
particular case, the argument arrives at a conclusion about the rule MPP itself. A gross or
viciously circular argument is immune to failure, and by extension, is empty: in reasoning about
a rule, if we state its correctness in a premise and then conclude that that rule is correct, then the
argument cannot fail. But the conclusion doesn’t tell us anything more than what we already
knew in our premises/assumptions, and so nothing has been gained—this is what I mean by
calling a grossly circular argument an empty one.

These two aspects of grossly circular arguments described in the first sentence of the paragraph
above are strong enough reasons to impel the inferentialist to seek out a more satisfactory
account of justification than gross circularity is capable of providing.

Pragmatic Circularity. A pragmatically circular argument that justifies a basic rule of inference is
one that reasons according to that rule but does not contain an expression of its validity among
the premises of the argument. In this way, the argument is not guaranteed to succeed, and
because it is not guaranteed to succeed we thereby actually prove something.

I believe that this distinction between vicious and pragmatic circularities can be read as part of
Dummett’s response to the Tortoise and sceptic. Consider the remark Dummett makes about the
stalwart sceptic:

Since a justification of a logical law will take the form of a deductive argument, there can be
no justification that appeals to no other laws whatever; but that does not matter, since there is no
sceptic who denies the validity of all principles of deductive reasoning, and, if there was, there
would obviously be no reasoning with him.18

This quote is almost a reductio, highlighting the largest problem confronting the sceptic because
it leads to the conclusion that the sceptic’s attitude is unreasonable. The Tortoise systematically
refuses to accept any conclusion arrived at by inference, and Dummett’s quote reiterates that
there simply isn’t any reasoning with an attitude of that ilk. The Tortoise is trying to point out the
futility of grounding any warrant for reasoning according to MPP because regress ensues, while
Dummett hones in on the fact that if we are to accept that sceptical thesis, then our entire science
of deductive reasoning cannot rest on firm foundations. But this is unreasonable. Dummett’s
suggestion is that our inferential practice can be justified via pragmatically circular reasoning
that has the potential to avoid regress, and at least be enlightening in some way. Dummett’s task
then, should be to build on his defense of pragmatic circularities and develop a theory that does
not admit defeat in the face of susceptibility to inconsistency and unsoundness.
Let us begin by looking more closely at Dummett’s claim that, unlike viciously circular
arguments, pragmatically circular ones are not guaranteed to succeed. Our earlier discussion of
tonk highlights this case. Consider the following pragmatically circular argument.

(P1.) Stephen Harper is current Prime Minister of Canada;

(Premise)

(P2.) Stephen Harper is current Prime Minister of Canada, tonk, C´eline Dion was
his predecessor;
(tonk-introduction)

Therefore,

(C.) C´eline Dion is former Prime Minister of Canada. (tonk-elimination)

In this case, The introduction and elimination rules for tonk allow us to infer, in an analytically
valid way, a conclusion that is false. The inference shows us that it is possible to reason
according to tonk (viz., it is possible for such a sequence of reasoning to be valid) and still
generate an argument that fails. A pragmatically circular argument is therefore enlightening in
the sense that some such valid arguments will yield true conclusions, and others will be false. If
some arguments fail because they are unsound, yet some others obtain, then we can legitimately
countenance pragmatically circular reasoning as fruitful and contingent on the meanings of
operators and rules involved. Tonk was shown to be meaningless above, and therefore the
meaning of any sentence employing it should be difficult to ascertain.

Now exactly what aspect of tonk renders it incorrect and causes some arguments that apply it to
fail? The examples under consideration are arguments such that a true premise is followed by
two correct applications of the tonk rules (viz., its introduction and elimination rules) which lead
to a false conclusion. Because of the false conclusion, we know that one of the inferential steps
taken is incorrect, despite being validly applied. This tells us that there is a problem with tonk—
we’ve used it properly, only to infer a false conclusion. The problem is with tonk’s rules. Taken
separately, both rules are correct, for, as we have seen, the tonk-introduction rule is identical to
the rule for and-introduction and tonk-elimination is identical to or-elimination. So both rules are
correct when applied in other contexts, yet when used together to characterise the meaning of
tonk, they define a bogus operator.

As expressed in the metalanguage, it is not immediately apparent that tonk is contingent.19

Tonk’s failure emerges through exemplification in the object language. First consider its rules in
the metalanguage as expressed in §2.1; We then observe the expressions in the object language
below. They demonstrate that arguments applying tonk result in sentences that are inconsistent
with the system of natural deduction, namely,

P |- Q and Q |- P.

Whereas a sound and consistent characterisation of natural deduction allows P|-P, it does not
permit P |- Q, for the latter does not express a universal truth. It is thus that we reason according
to tonk in a pragmatically circular way, and find that the rules governing this operator fail a
conservativeness test, which must be passed if we are to adopt the rule in natural deduction. Our
justification of tonk by pragmatic circularity has therefore failed . A different operator that is
characterised in such a way that it is conservative will be a better candidate for admission into a
system of natural deduction.

Now in their status as circular patterns of reasoning, are pragmatically circular arguments as
useless as the grossly circular ones in their attempts to justify logical rules? It depends, Dummett
holds, on the purpose of the argument put forth. If its purpose is explanatory, whereby the
argument serves to clarify the reasons an explainer has for holding that rule to be valid, and
whereby the explainee endorses the correctness of that rule as well, but merely seeks a
demonstration of that correctness, then pragmatically circular arguments will have achieved their
intended purpose. However if the purpose of a pragmatically circular argument is suasive,
whereby the explainer intends to use a pragmatically circular argument to convince or persuade a
doubtful explainee (such as the Tortoise) of the correctness of that rule, then the argument will
fail because the explainee shall not allow the explainer to employ that rule of inference to
generate a justification for that very rule—it would be question-begging. The explanatory
campaign takes for granted that the rule in question is correct, and subsequently delivers a
circular vindication of this belief. But the intuition just mentioned is the very state that the
sceptic, and the Tortoise for that matter, take issue with. For as we have seen in the Carroll story,
the only sort of inferential justification available is an infinitely regressive one.

An inferentialist endeavour with explanatory rather than suasive objectives takes a leap of faith
in assuming the validity of MPP prior to delivering his explanation. It’s a small leap, because
even the (natural) language from which the rule derives its characterisation and use20
countenances that rule as correct.21 The difference between a thinker who accepts pragmatic
circularities and the sceptic who denies them lies in the former’s willingness to allow certain
types of faithful leaps by accepting that a rule be applicable in inference ab initio. A proponent
of pragmatically circular arguments will then use that supposition to formulate an argument that
justifies our entitlement to use that particular inferential rule. By contrast, the tortoisean (or
sceptic) prohibits those same faithful leaps. This is an example of a pragmatically circular
argument arguing for the correctness of MPP:

(P1.) A rule of inference that is conservative and truth-preseving is a correct rule of

inference.

(P2.) MPP is conservative and truth-preserving

Therefore,

(C.) MPP is a correct rule of inference

So we have come some way in establishing what sort of circularities might be permissible in
arriving at legitimate justifications for logical laws and operators according to Dummett, namely
pragmatic ones. But we also saw that some validly characterised operators such as tonk were
problematic insofar as they generated inconsistencies. Since illegitimate operators can still be
characterised in a pragmatically circular way, how might we then circumscribe those operators
that preserve the soundness and consistency of any argument containing true premises? We have
yet to bridge the leap of faith I mentioned above, and must consider whether a conventionalist
attitude about the meanings of operators might secure a justification where circularities fall short.
Under conventionalism, we might hold that the logical constants or rules are valid simply in
virtue of being the ones we treat as valid, thereby replacing the leap of faith at the beginning of
the inference pedigree (viz., where the justification for the correctness of a rule of inference is
delivered prior to the application of that rule of inference in argument) by accepting those rules
as basic, regarding them as default-reasonable, or countenancing them as unjustifiable. But any
rule system can be conjured up, such as the system that admits of tonk—if we are to speak of
systems that we simply adopt as valid. Dummett likens conventionalism and its shortcomings to
following the rules of a game. If all the rules of a particular game are understood by some player,
and the overall objective of the game was also understood by that player, then he would have a
complete understanding of the game. But the game could also be played by a player who knew
the rules perfectly well, but didn’t understand its purpose, objective or strategy. It is questionable
then whether the person has a full grasp of the game despite her strict adherence to the rules.
This is analogous to a context where, hypothetically, a reasoner understands the rules governing
tonk, and adopts them conventionally. The reasoner could apply the tonk rules but he would
generate inconsistencies. And because the reasoner endorses a system (in the case of logical
expression) that generates inconsistencies, we can say that perhaps this reasoner does not really
know what he is doing—that is, does not seem to understand the purpose of reasoning. When we
use a logical operator, we do not always use it in the context of deductive inference—sometimes
our usage is non-assertoric, and our conception of the operator is therefore richer than a purely
formal conception is capable of supporting. The understanding of the game’s purpose and
strategy figures integrally in helping us know which are the appropriate rules to the game, and
this understanding seems to be the very thing we have yet to describe. This understanding is
intuitive or inchoate. We barely know we have it, but intuition does do its work when we come
to understand that we have been playing a game lacking of objective or strategy in the case of a
game, or reasoning according to an inconsistent rule of inference in the case of an argument.

The important question in the preceding paragraph therefore asks after a method of
circumscribing those pragmatically circular justifications of logical laws that are sound and valid
from those that are not, such as tonk. We saw that metalinguistic characterisation was
insuffcient, for it was possible to characterise tonk in analytically valid fashion, despite our
awareness of its shortcomings. Dummett’s response to this problem begins by acknowledging
that the ends of natural language are internal. We presumably need natural language in order to
express our thoughts. The meanings of our words and linguistic devices are determined by the
rules governing their use. Failures such as inconsistency arise in our linguistic tools when the
multiplicity of principles governing our linguistic practice are discordant. ‘For the language to
function as intended, these principles must be in harmony with each other.’22 We shall first
identify Dummett’s distinction of the two categories of meaning-constituting rules of use, and
then we shall investigate exactly what he means for rules to be in harmony:

Verificationist Principles. Are those that have to do with the circumstances that warrant
assertion, indicating when we are entitled or required to assert a particular statement.
This is established both by inference (in the case say, of mathematics) or by observation,
and often our entitlement is based on a hybrid of the two. The verificationist component
of the meaning of the logical operators are particularly easy to specify: they are the
 introduction rules. And if we know the introduction rules, it can be said that we possess
the complete meaning of the operator in question because we may verify the meaning
against the rules—something akin to analytic validity.

However if we were to rely solely on what is required to verify a particular statement in

order to be entitled to assert it, then ‘...we should be skilled at making assertions but
incapable of responding to the assertions of others.’23 This is because we do not all know
the same things—we are not all in possession of the same fact and meaning base, and
what a thinker knows about any one particular thing ultimately depends on everything
else he knows. And therefore given knowledge of the rules that characterise a logical
operator (admittedly it gets trickier with the non-logical terms of natural language), we
can form statements and verify their conformity to the rules, however this principle is
incapable of considering the consequences of our choice and application of rules.

We therefore counterbalance verificationist principles with;

Pragmatic Principles. These principles are the ones that determine what we can
warrantedly infer from a claim. They prescribe that the content of a statement should be
regarded as determined by its consequences for one who accepts it as true. This principle
asks what the difference made to a thinker consists of should she choose to accept the
statement. In the case of a logical operator, the pragmatist component of the meaning is
once again easily specified: it is given by the elimination rules. The pragmatic approach
to governing our linguistic practice therefore assesses the logical consequence of an
application of an elimination rule.

Suppose we interpret formalism generally as the principle that allows us to assemble the rules for
any language, and that a language so created is nothing but a series of symbols and rules telling
us what may be expressed in the system. It is nothing but such a series because it does not
express statements about anything in particular. In order for such a system to have meaning, we
must assign it a semantics, and this is where pragmatist principles figure: once the semantics is
assigned, we can decide whether the statements inferred in that system agree with what we hold
to be a correct view of the domain of discourse. It decides about the consequences of being
allowed to assert a particular statement.

The principles of verificationism and pragmatism account for essential aspects of meaning, and
neither of them is singly sufficient for a complete meaning theory. Dummett held that a complete
meaning theory must be informed by both principles, and that if a defined operator is going to be
correct in a logical system, its introduction and elimination rules must be in harmony. The
introduction rules prescribe which statements involving the logical vocabulary are expressible in
the language, and the elimination rules prescribe what inferences we are permitted to draw from
those statements. The introduction and elimination rules for a particular logical operator are in
harmony when the elimination rules allow us to derive from a sentence with that operator as its
main connective all and only those statements that the introduction rules require that we know in
order to warrantedly infer that statement in the first place. Harmony guarantees the absence of
inconsistency when we consider the addition of any operator to a system that is consistent to start
with, because it guarantees conservativeness.
The logical conception of a conservative extension best reflects the semantic, syntactic, and
meta-theoretical requirements prescribed by the attribute of harmony. Conservative extensions
were discussed in detail in §2.2 above, and what we must carry forward from that discussion
involves the requirement that any rules we wish to introduce into a system that characterise an
operator be a conservative extension of that system in order to preserve validity and consistency,
and that harmony preserves this conservativeness. When we add a new constant to a natural
deduction system, we threaten the consistency of the newly anointed system. If we can prove that
the addition of the new constant does not generate inconsistencies in the augmented system (i.e.
the natural deduction system augmented by the new constant), then the newly anointed constant
will be correct in the new system. The anointed system will not yet be a conservative extension
of the original one however, for conservativeness carries with it the additional requirement that
we should not be able to express conclusions in the new system that are unexpressable in the old
system, save for conclusions that contain the new operator themselves. Conservativeness is
therefore a stricter criterion than harmony—its stringency surpasses that of consistency’s with its
additional requirements pertaining to new expressions expressible in the anointed system: that
they do not facilitate new expressions unless the new operator is included in that expression.

Dummett believes that harmony implies conservativeness because of the sharp correlate existing
between the metalanguage and the object language in logic. Suppose some consistent system of
natural deduction. We then augment that system by an operator whose rules are in harmony. We
can then see, by observing stepwise proofs in a corresponding object language, that the use of
that operator is superfluous. We never required the operator to be added to the system in the first
place. A step of a proof that utilises such an operator introduces it by its rule. An inference
immediately ensues by applying that operator’s elimination rule. Dummett calls such an
occurrence in a proof a ‘peak’. The superfluousness just mentioned is apparent when we consider
that the propositions (represented by variables) that we admit into a deductive proof are already
ones that we accept. What remains unknown as yet are the properties of the resultant inference -
the conclusion. But we were already able to articulate the premises and effect all the possible
valid operations on them using the original system; we were able to do away with the new
operator; to level those peaks, much in the same way as we eliminate steps of a proof using the
cut-rule in the predicate calculus. Harmony therefore implies conservativeness because no new
statements are inferable via the conservatively augmented system that we couldn’t infer in the
original one.

What lesson might we draw about BLK from Dummett’s enquiry into harmony? Acknowledging
that the basic rules must be introduced first into any system, prior to the more complex rules that
will likely employ BLK in their own characterisations, it is obvious that basic logical rules will
comprise a consistent natural deduction system. This is because, given nothing but basic
operators, there are no other components of the system with which they may be inconsistent.

Which rules are basic? In other words, when do we know that we have enumerated them all? The
concept of harmony answers this question with its requirements of consistency and parsimony. If
we can raze a logical system of an operator without sacrificing explanatory power then the item
razed is not basic. The criterion of conservativeness is therefore useful and effective to the end of
circumscribing which elements of a system are to be basic.
2.5 Boghossian and Conceptual Role Semantics

Boghossian’s treatment of the problem of characterising the logical operators takes an

epistemological approach, which is to say that he examines our practice of belief in order to
ascertain what epistemological stances we should hold, or are permitted to hold, in the case of
the basic rule of inference MPP. He promotes a theory of justification for BLK that argues in
favour of our entitlement to do so, provided that the meanings of the logical operators may be
fixed, and once again, provided we are not attempting to demonstrate this entitlement to a
stalwart sceptic.

As we have seen, gross circularities give rise to two familiar fallacies in argumentative
reasoning, namely, the fallacy of begging the question, which takes the strategy of assuming in
the premises of an argument that which we seek to prove in its conclusion. The other fallacy that
arises from a grossly circular argument is the problem of bad company, whereby in assuming
what we seek to prove, we become capable of proving anything.

Boghossian agrees with Dummett’s view on the issue of circularity by distinguishing the
pragmatic from the gross, and sets up his endorsement of the former as a justificatory strategy
with promise:

...it is not immediately clear that we should say that an argument relies on its
implicated rule of inference in the same way as we say that it relies on its premises.24

The problem of begging the question gives rise to two principal (epistemological) challenges
according to Boghossian.

1. ‘...is to say in what the entitlement to use a rule of inference consists, if not in one’s
justified belief that that rule is truth-preserving;’

2. ‘...is to say how a rule-circular argument can confer warrant on its conclusion even if it
is powerless to move the relevant sceptic.’25

If we are obliged to justify our belief that a rule is correct before we can legitimately employ it in
argument, then the sceptic’s challenge will indeed block our entitlement to reason according to
that rule. For according to him, our belief in its validity is not justified or worse, unjustifiable.
However, rather than focus on our beliefs about validity, an alternative response to the sceptic’s
prohibitions (who, as we know already, repudiates pragmatically circular justifications) consists
in the delivery of a correct account of what determines the meanings of our logical constants.
With the meanings of those constants fixed, Boghossian argues that we are entitled to employ the
rule without first being justified. It is proposed that meanings be delivered by way of a
conceptual role semantics (CRS), which seeks to demonstrate the meaning of an operator or rule
by enumerating the role it plays in inference. That role is thereby meaning-constitutive of the
logical operator in question. To wit, of the entire set of statements employing the ‘if-then’ rule in
a thinker’s repertoire, a specific subset of those statements will constitute the meaning of that
operator. We have referred to these throughout this text as introduction and elimination rules.
At least two challenges can be mounted against a CRS. A semantic indeterminist might oppose
this conceptual-role model on the basis that only an infinite (and therefore unquantifiable)
number of demonstrations can fix the meaning of an operator. But a CRS implies that a specific
subset of inferences involving some operator fix its meaning. If that specific subset is
quantifiable, and we have reason to believe that it is in consideration of characterisations of
logical constants through specifically designated introduction and elimination rules, then the
implications of the challenge of semantic indeterminacy change because the indeterminist is no
longer capable of challenging the practice of prosecuting an infinite set as determinative of
meaning. Another angle to these opposing views might be understood along the following lines;
a semantic indeterminist would argue that an infinite number of logical truths represent the
meaning of an operator, while the proponent of a CRS believes a finite subset is sufficient for the
job. The upshot of these opposing views is that the latter shall be able to ostensively generalise
the meanings of logical laws or operators by pointing to a subset of rules that make up a
meaning, while the indeterminist is unable to express the same generalisations if he maintains
that the total set of logical truths make up that meaning. As we shall see in this section,
Boghossian will have to make a strong case for a CRS in order to persuade an enquirer that
meaning indeterminacy can be overcome.

The second challenge is that Boghossian’s proposal of employing a CRS is threatened by the
potentially lax results permitted by the acceptance of introduction and elimination rules as
constitutive of meaning. As we have already seen in the case of tonk, nothing blocks us from
nominating any collection of introduction and elimination rules to characterise meaning. When
used in a rule circular justification however, we saw how tonk enabled us to legitimately derive
any statement—even inconsistencies and falsehoods. But Boghossian wants to say that not every
collection of introduction and elimination rules determines a meaning, and in his story tonk
would be an operator that fails to do so. Certain criteria must be satisfied in a CRS in order for an
operator to have meaning by characterisation through introduction and elimination rules—truth-
preservation is one such criterion. Some sets of rules are truth-preserving, such as Belnap’s
presentation of the rules for the operator ‘and’, while other sets of rules are not, again tonk. To
avoid the lax results facilitated by tonk and similarly characterised operators, Boghossian
explains that for an operator to have meaning, the given subset of (introduction and elimination)
rules ‘...means that unique logical concept, if any whose semantic value makes the inferences in
that subset truth-preserving.’26 Tonk and other operators devoid of meaning do not satisfy this
last requirement.

Supposing that a CRS succeeds in delivering the meanings we seek, Boghossian suggests the
following epistemological principle:

(L) If M is a genuinely meaning-constituting rule for S, then S is entitled to infer

according to M, independently of having supplied an explicit justification for M.27

He then remarks that this principle,

...does not require that S know that M is meaning-constituting for S if S is to be entitled to

infer according to M but only that M be meaning-constituting for S.28
(L) he argues, is intuitively plausible. That we have the disposition to reason according to some
rule should stimulate an entitlement to reason according to it that is prior to, and independent of,
an explicit justification for that rule. Without the disposition or intuition to reason as such in the
first place, there would be no rule of inference for which we would need to seek out a
justification. Boghossian seems to be saying something along the following lines; ‘we’ve used
some rule in our inferential practice and it is indeed as yet unjustified, but something compelled
us to conceive of and use that rule, rather than some other. We should recognise this fact as
significant, and give this rule that we’ve been disposed to use some credit as being a rule that is
correct.’ If we didn’t allow these dispositions to stimulate inferences, then the only thing we
would be capable of doing in a sceptic’s world is assert ‘contentful’ items (and not transitions
between contentful items), which in my view amounts to a license to only make statements about
what is directly observable in the case of physicalia, or knowable in the case of memory, belief,
or intuition.

Entitlement to use a rule of inference ensues from a CRS in Boghossian’s story. A CRS is a
theory whereby introduction and elimination rules determine the meanings of operators. But he
also holds that not every rule set determines a meaning. His claim is that only operators with
genuine meanings may be used to justify rules of inference in an argument. The fundamental
issue then is the problem of determining what makes a class of rules genuinely meaning-
constituting. The sceptic may continue to deny beliefs (namely those beliefs hanging on to the
correct nature of some rule) from figuring into any justification of a rule of inference, but
Boghossian holds that there are forms of reasoning that emerge a fortiori from our propensity to
infer, that is, we are disposed to reason in these ways. Provided that characterisation in terms of
introduction and elimination rules in the case of the logical constants can be delivered and can
establish an operator as genuinely meaning-constituting, then the CRS delivers the entitlement to
reason according to a rule. Our entitlement therefore reduces to the problem of determining what
makes a rule meaning-constituting. Boghossian’s first epistemological challenge listed above,
(1.), is thus resolved. In order for Boghossian’s story to work, he would need to give us a
worked-out account explaining exactly how he links the conceptual role an operator plays in
inference with its meaning. That is the very large proviso mentioned five sentences above, and
this Boghossian attempts in his paper Analyticity Reconsidered.29

His second challenge, (2.) at the beginning of this section, treats of the problem of persuading the
sceptic. Supposing the sceptic were stalwart, is there anything we could do to persuade him? He
blocks all explicit categorical statements to the effect that a particular rule of inference is correct.
By extension, our entitlement to reason according to a rule of inference is nullified as well, on
pain of the lack of justification for the belief that the rule is correct. We would also have
difficulty persuading him that a rule is meaning-constituting because that would involve an
argument, and that argument would have to employ MPP. So Boghossian asks the important
question, what is the epistemological significance of the fact that we are unable to persuade the
sceptic about MPP?

Boghossian evaluates Dummett’s response to the problem of replying to the sceptic where
Dummett distinguishes between the distinct projects of quelling the sceptic’s doubts versus
explaining to a non-sceptic why MPP is valid. The key to Dummett’s response is, as we have
seen, the distinction he makes between two sorts of arguments, which he baptised with the terms
suasive and explanatory. But Boghossian argues that he does not understand what it would be to
explain why a given law is true, short of adopting some (unfavourable) conventionalism about it.
Instead;

The question that we need to be asking, I think, is rather this: Can we say that something
is a real warrant for believing that p if it cannot be used to answer a sceptic about p? Is it
criterial for my having a genuine warrant for believing that p that I be able to use it to
persuade someone who doubts whether p?30

This question travels with the thought that if we are genuinely justified in believing p, then we
should be able to convince others as well—to ‘see as we do’, so to speak. Boghossian believes
that it is this propensity, one that supposes the possibility of intersubjective and unanimous
assent, that informs our theory of knowledge, and that lies behind the persistence of the feeling
that we must satisfy the sceptic—that having a genuine warrant includes the sceptic’s blessing.
But Boghossian holds that this thought is false, and the following vein of reasoning which I have
distilled from page 253 of his paper elucidates exactly why:

P1. We deny that warrants do not exist for BLK;

P2. We don’t know what that warrant might consist in if not rule-circular inference.

P3. The sceptic denies rule-circular accounts of BLK.

Therefore,

C. There must be genuine warrants that do not carry any sway with the sceptic.

Although this argument expresses Boghossian’s view on the status of warrants for BLK, it is an
argument that would not satisfy sceptical scrutiny. P1. is a premise that opposes the sceptical
thesis, and essentially names that thesis as incorrect. It is only on the basis of acceptance of P1.
that we may arrive at the stated conclusion. Boghossian presumes—he begs the question—
concerning the existence of such a warrant via P1., and this presumption conspires to deliver an
argument in favour of the possibility of justifying a warrant for BLK that is as effective as an
explanatory argument, but not a suasive one. On this issue then, Dummett and Boghossian seem
to arrive at the same result.

Once we accept the concessions an explanatory argument forces us to make, we may enjoy an
entitlement to reason according to MPP because we are able to deliver the meanings of the
logical constants via introduction and elimination rules that preserve truth. Truth preservation is
the criterion, according to Boghossian, that demarcates operators with genuine meaning from
those devoid of it. Finally, we do not need to satisfy the sceptic if we agree to allow just one
premise citing the validity of MPP into an argument that justifies our warrant for applying it in
inference; with that one assumption we may present a deductively valid justification that even the
most stalwart sceptic would have diffculty denying.

2.6 Hale’s Non-Inferential Route

In his paper Basic Logical Knowledge,31 Hale argues that there must be some noninferential
knowledge underlying knowledge derived inferentially. This constitutes his greatest departure
from the views of Dummett, Boghossian, and other inferentialists. It is possible to define
conditions delivering a characterisation of what an inferential justification for a basic principle of
logic consists in, but Hale does not see any way for these conditions to be met. Recalling from
§1.2, those conditions are:

• That an argument to the effect that some basic rule of inference is sound does not use
that rule, nor any other rule to corroborate that argument.

• That an argument not assume that some inference rule is sound in order to corroborate
an argument that some other rule is sound.

Obstructing the non-inferential route from arriving at a plausible account of BLK is its lack of a
satisfactory answer to the question: what is to constitute the fabric of non-inferential knowledge?
Again, §1.2 discusses the problems associated with grounding this sort of knowledge in sense-
perception, introspection, self-evidence, and other similar ideas. In response to this challenging
question, Hale argues that we can answer it by directing enquiry towards two distinctly different
projects:

Project A Explaining how we can come to know that basic rules such as modus ponens are
sound.

Project B Explaining why it is not possible intelligently (i.e. clear-headedly and

coherently) to doubt the soundness of basic rules...32

Dismissing rule-circular attempts to arrive at BLK, Hale sees promise in a strategy that explains
BLK as grounded in the conditions for understanding the logical constants. In order to
understand the functioning of a particular operator, we must at the very least accept certain basic
patterns of inference as criterial and constitutive of the meaning of that operator. Of course, this
by itself does not seem sufficient to ground a claim that we have logical knowledge; Hale cites
Paul Horwich with noting that acceptance of a rule is distinct from its truth or correctness. But
Horwich’s point only threatens project A.

Project B is the more modest of the two, and Hale believes that success in that project will permit
advancement to the more ambitious project of explaining BLK identified in Project A. By
accepting certain simple patterns of inference such as MPP as constitutive of understanding, Hale
suggests we commit ourselves to the claim that it is impossible to countenance an (accepted)
inference as unsound, and also commit ourselves to the claim that no counter-examples to that
rule exist.

But acceptance as arbiter of our understanding of meaning fails to be a sufficient condition for
that understanding, for it is possible to accept rules that are, unbeknownst to us, defective in
some way. Acceptance can be at most constitutively necessary for understanding. A sceptic can
oppose our present train of thought about the practice of acceptance; she might charge that if
acceptance of certain basic patterns of inference is constitutive of (a.) understanding the meaning
of a constant, and (b.) being sound and immune to counterexample (viz., impossible to doubt),
then in the same way we accept MPP, so too must an operator such as tonk enjoy these qualities
in a case where somebody erroneously accepts that operator as well. But we know that there is
something wrong with the operator tonk, and therefore something amiss with the sceptic’s
reductio argument. There must be a codifiable difference between accepting MPP and accepting
tonk.

The sceptic’s charge is correct in that we should be able to entertain doubts about an operator
such as tonk. It is also correct that if acceptance entails immunity to counterexample and doubt,
then that immunity must extend to all rules of inference we accept: both to sound and unsound
ones. Yet the sceptic is wrong to suppose that a sound rule of inference (such as MPP) should be
doubtable through counterexample. He would only be able to make this claim if it were possible
to find a counterexample, say, in the case of MPP, where the premises {A, A → B} are true but
the conclusion B is not. But such an instance of counterexample does not exist. In the case of
tonk, the abundance of counterexamples we are able to infer from it, as well as the proof above
that shows we can infer A |- B by tonk, amounts to the possibility of inferring any statement
whatever, or of inferring contradictory statements. Tonk therefore has no fixed or coherent
meaning because its rules are non-conservative and inconsistent. Hale draws two morals here,
the first is that:

‘...whilst there is undoubtedly something wrong with the tonk rules, their fault is not most
happily or illuminatingly characterized in terms of failure to preserve truth—neither can
properly be convicted separately of this failing,...’33

What goes wrong, he suggests, might be that tonk fails a conservativeness constraint on rules of
deduction. The second moral is that examples such as tonk do not corrupt the idea that it is
impossible to infer counterexamples or inconsistencies using logical operators whose
characterisations we accept as constitutive of their meanings. Hale shares this belief with
Boghossian, and both would agree that conservativeness (according to the former), and truth-
preservation (per the latter) distinguish the meaningful operators from the meaningless.

Hale maintains that the sceptic’s challenge, which ultimately says that we can doubt the
correctness of basic logical rules, does not disturb the prospect of succeeding at project (B).
Suppose we wish to verify that an operator satisfies a conservativeness requirement. Tonk fails,
and we must demonstrate this in an argument. If tonk is not conservative, then we certainly
cannot use tonk as a rule of inference in argument, for tonk is unsound. But we can reason about
tonk, which is different from reasoning using it. And reasoning about does not require that we
reason according to in an inference. Consider the following argument charging tonk of non-
conservativeness:34

The introduction and elimination rules for tonk are not conservative: take any pair
of tonkless propositions, say, ‘7 + 5 = 12’ and ‘Hilary Putnam is a brain in a vat.’

(P1.) 7 + 5 = 12 (P2.) ‘7 + 5 = 12’ tonk ‘Hilary Putnam is a brain in a vat.’

(TonkIntroduction).
 (C1.) ‘Hilary Putnam is a brain in a vat.’ (Tonk-Elimination). Since the undischarged
premise and final conclusion are tonkless, and the latter cannot be inferred from the
former without using the tonk rules, those rules are non-conservative. 35

The above argument does not use tonk to reach its conclusion that that operator is bunk, it merely
reasons about the operator, and concludes something important: that it is non-conservative. This
is a metatheoretical statement about our knowledge of tonk - specifically about its non-
conservativeness, and by extension, its meaninglessness in the context of deduction. This
argument about Putnam cannot be considered a counterexample to tonk because we do not give
an argument applying tonk any clout: if a reasoner is to accept a given argument, she must accept
the applied rule of inference, but she doesn’t because tonk is inconsistent and therefore
meaningless. Hale has shown that any vindication of a doubt about conservativeness must use
rules other than those whose conservativeness is in question—rules whose reliability is assumed
during the reasoning.

It does not, of course, follow from this that there must be some rules whose reliability
must, and may properly, be assumed in any demonstration we can give of the
conservativeness or non-conservativeness (more generally, soundness or unsoundness) of
any (other) rules. It does not follow, but it is—or so I believe—true. (Hale’s emphases.)36

Hale proceeds by discussing this explicit proof of the non-conservativeness of tonk at length,
arguing that in giving any such proof, he cannot help but use MPP, universal quantifier
introduction and elimination. Part of the proof involves asserting that some particular inference is
an instantiation of its general form (viz., ‘p, so p tonk q’ is of the general form ‘A, so A tonk B’).
In a page-long footnote37 regarding this last inference he observes that despite the obviousness of
the content it asserts, we should still ask after some sort of verification of its correctness. But we
would fall to an infinite regress of the Tortoisean sort to ask for an inferential justification, and so
the correctness we seek is, he suggests, the result of a seeing—a species of non-inferential
intellectual recognition. Hale reminds us that this insight cannot provide us with a view into
something more complex such as validity, but it does allow us to mediate recognition of some
particular inference as exemplifying some general rule. If we are to evaluate a rule of inference
for soundness, that rule must be both general and conditional: general in the sense that some
conclusion of some specified general form may be drawn from premises of some specified
general form; conditional in the sense that given premises of the specifed form, a conclusion of
the specified form may be drawn. There must exist a minimal kit of inference rules that consists,
according to Hale, of the above-mentioned MPP, universal quantifier introduction, and
elimination.

Hale has therefore pointed out two ways in which it might be thought possible to doubt the
soundness of an inference rule:

1. We might envisage a counterexample to it.

2. We might doubt its possession of other properties required for soundness such as
conservativeness.
He has shown per 1., that because we initially accept the rule of inference we have no right to
doubt it by counterexample. But acceptance is insufficient to guarantee correctness because if we
were to hold the same attitude about an operator such as tonk, we would have reason not to doubt
this rule either. But this clearly isn’t the right way to view the situation. We do indeed need to be
able to doubt the soundness of tonk, but not through supposing that counterexamples to it exist.
Rather it should be per 2., through a demonstration showing how it lacks other properties such as
conservativeness. The reasoning we employ to prove non-conservativeness avoids circularity by
not using the rule of inference in question, and furthermore employs a minimal kit of inference
rules, consisting of MPP for one, that is required to reason about the soundness of any inference
rule. 2. is inapplicable to sound rules of inference, but perfectly applicable to unsound ones. If a
minimal kit exists, then it shall enable us to articulate doubts about unsound rules while
preserving the charge that the sound ones are immune to doubt.

To complete this survey of Hale’s views, we now address the objective mentioned in project A
above, the explanation of how we can come to know that basic rules are sound. We recall Hale’s
suggestion that we might build upon the success realised in project B in order to realise success
in A. A legitimate concern exists in this regard about the prospect of mitigating the connection
between these two projects without the use of those items contained in the minimal kit, for in
utilising the kit we shall have employed circular reasoning to achieve the goal, and Hale has
avoided circularity at all costs thus far. The answer to this dilemma Hale suggests, is that rather
than assuming that some theory renders project B successful, and then, armed with the resultant
knowledge that the operators of the minimal kit are impossible to doubt, we proceed to leverage
this knowledge by using those operators to argue in defense of a positive conclusion in support
of project A. Instead we here obviate the need to express their indubitability in a premise.

Hale is, unsurprisingly, less comfortable with this suggestion but speculates,

Perhaps it can be maintained that there is, simply, no inferential gap to be traversed—that
is, that knowledge can, much as Descartes thought, consist in believing truly something
which it is impossible rationally to doubt.38

It is interesting to note that treating basic rules as a foundationalist agrees with the first moral
mentioned by Engel in §2.3, where he draws the distinction between rules and premises, and
suggests that we might consider prohibiting the former from being expressed assertorically in a
deductive inference. Without them explicitly expressed, there doesn’t seem to be much else to do
other than accept them implicitly, which would suggest that we have license to reason according
to them without justification—that they are default reasonable.

If this is the case, we should be able to countenance basic logical knowledge as non-inferential,
different from the problematic sorts of non-inferential knowledge discussed in §1.2 insofar as it
is, in this way, general and necessary, rather than subjective and contingent in the way memories,
intuitions, and beliefs can be.

2.7 Summary of Findings in this Chapter

In §1.6 I enumerated a series of questions and problems that an account of our entitlement to
reason according to BLK must address. In this section I shall assess whether or not those
questions have been answered, or rather, to what degree they have been answered by our survey
of philosophical views presented in this chapter. Recalling from Chapter 1, the principal
problems are described thus:

(i.) Carroll shows us that paradox ensues from an inferential argument offered to justify
the correctness of a rule. In the case of MPP, such an argument makes use of that rule to
arrive at its conclusion. A subsequent argument is then required to defend our warrant to
reason according to the rule, and this repeats itself ad infinitum.

(ii.) There are problems associated with all manner of non-inferential support for the
basic logical constants: such support is subjective and arbitrary rather than universal and
necessary as we need it to be.

Also in §1.6 it was stated that any treatment of these problems would have to address the
following questions and issues:

(a.) That inferentialism and non-inferentialism constitute the two strategies we have at
our disposal in order to arrive at knowledge. In the scope of this thesis, that knowledge is
about the correctness of basic logical rules.

(b.) Inferentialism seems like a more rigourous strategy for characterising the logical
constants because it gives us reasons in the form of valid arguments that conclude that a
logical constant is correct. Non-inferentialism is only as rigourous as the intuitions it
codifies, yet intuition about logical constants seems like the very sort of knowledge we
seek to explain, and therefore noninferentialism has an aire of circularity to it.

(c.) That a strategy for demonstrating the correctness of a rule exists and does not
succumb to regress or vicious circularity. We extrapolate from this that the
Sceptic’s position on this matter is to be rejected.

Upon completion of our survey, it is obvious that (b.) is wrong-headed. It seems as though (a.) is
strengthened, and we have also improved the likelihood that (c.) is correct, with qualification. As
for (b.), both reason-based and non-inferential approaches to justification can be interpreted as
yielding circular results: Carroll showed, as we saw, that given an inferential argument about the
correctness of MPP, an infinite regress of justifications followed. If instead we hold that
justification terminates (by accepting rule-circular, or pragmatically circular sequences of
reasoning), then the terminating inference will be the circular supposition that the rule in
question is correct. A non-inferential justification of rules can be equally interpreted as circular,
for such an explanation is grounded in, as we saw, intuition, belief, definition, etc. The
correctness of a logical constant under a non-inferential characterisation is itself an intuition - a
hunch we seek to vindicate, and so in the case of the basic logical constants, a non-inferential
justification also appeals to the very principles sought to be justified. Therefore the illusion that
presented itself to me in Chapter 1 suggesting that an inferential justification is more rigourous
than a non-inferential one because it offers reasons is just not right because the reasons are
couched in circular justifications, in the same way as an intuition.

I do not think that (a.) can be proven as a cold hard fact just yet, if ever at all, but intuitively it
does seem as though these are the only two strategies for arriving at knowledge. I exclude the
thesis of scepticism from these strategies because it rejects the possibility of justifying
knowledge. I think many of us would like to admit or begin with the initial assumption in
epistemological enquiry that knowledge exists and that much of it—if not all of it—is capturable
via one of our justificatory strategies. Regardless of the domain and whether we call the strategy
inferentialist or non-inferentialist, proof-theoretic or model-theoretic, or some other account that
amounts to the same fundamental principles, I can think of no other positive accounts of
justification than the ones described by inferentialism and noninferentialism.

The qualification that needs to be made regarding (c.) is exactly addressed by Dummett’s
suasive/explanatory distinction. He is correct about suasive justifications; there is no convincing
a stalwart sceptic by argument, because that argument will ultimately use the rule of inference
under question in a pragmatically circular way. Nor will the sceptic accept stipulations, decrees,
definitions, or others discussed throughout because no manner of verification for their
correctness will satisfy her. Therefore, so long as our purpose for demonstrating correctness is
explanatory in nature, then we should be allowed to admit pragmatically circular justifications in
a defense of the correctness of a rule of inference. But Hale’s point needs to be considered here.
For if we are going to admit a pragmatically circular justification, then we must, for at least a
fundamental collection of rules, admit of their correctness. If we do, and assume their correctness
ab initio—which is essentially what the admission of pragmatically circular arguments consists
in, then perhaps we can countenance them as default-justified. A viable non-inferential defense
of the correctness of basic rules may then be launched from these assumptions. I will have more
to say about this in Chapter 3.

How have the proposed solutions to problems (i.) and (ii.) unfolded? In the case of (i.) we are
confronted with the difficulties suggested by both regress and circularity. Someone who
addresses these would have to present an account that circumvents them. Focus on meaning has
been thought to be an effective strategy: if we can fix the meaning of an operator, then perhaps
we shall be able to stave off regress and circularity. Dummett’s and Boghossian’s accounts may
both be interpreted as being grounded in meaning by virtue of developing a theory couched in
analytic validity, although their strategies explaining how to arrive at meaning differ
significantly.

We have yet to discuss problem (ii.) pertaining to non-inferentialism, mentioned at the beginning
of this section. Hale’s piece is the one that makes a significant contribution to the progress of
enquiry into that problem. To avoid repetition, I shall save discussion of Hale for the next
Chapter, for it is his contribution that figures chiefly in my concluding findings.

Chapter 3

The Primacy of Non-Inferential Justification

Obvious questions to ask at this point, subsequent to my enquiry into the problem of justifying
basic logical rules and thereby arriving at BLK, is the question that asks, which justificatory
strategy? Should inferentialism or non-inferentialism be our epistemic attitude when justifying
warrant? Which is better capable of delivering the correct characterisation of BLK, if such a
thing exists? In this concluding chapter, I answer these questions with a defense of non-
inferentialist approaches to justifying rules of inference. I shall argue that non-inferential
knowledge of basic logic should be distinguished from other forms of non-inferential knowledge,
and that while non-inferential knowledge in general is typically dubious, non-inferential
knowledge of basic logic should be accepted. The thesis ends with my most important argument
in favour of non-inferentialism: that this sort of knowledge is more deeply grounded in a chain of
reasoning than knowledge arrived at inferentially. In light of this, non-inferential knowledge of
basic logic remains possible, while inferential knowledge of basic logic is not.

3.1 Distinguishing Types of Non-Inferential Knowledge

Let us take stock of the kinds of non-inferential knowledge previously discussed: beliefs,
memories, intuitions, stipulations and a few others fall into one category. They are basic in the
sense of being the starting point for an item of knowledge—and in the cases under consideration,
the items didn’t exist until thought-up or recalled by someone. We generally regard this sort of
knowledge as dubious because of its subjective and oftentimes non-transmittable1 nature. If an
item of knowledge depends on a single individual in order to be completely understood, or if it is
simply non-transmittable, then the item of knowledge does not rest on the firm foundations we
aspire to lay down for logic, nor can they serve as foundations themselves.

But non-inferential knowledge of logic—the second category—fares differently for two reasons.
First, I have demonstrated in this thesis how we may countenance our knowledge of basic logical
rules as universal, absolute, general and necessary. Furthermore, knowledge that they are
conservative and in harmony lends strength to our knowledge of their correctness. We cannot,
and do not, unanimously assign these logical properties to the first sort of non-inferential
knowledge. The second reason that non-infererential knowledge of logic is exceptional is
because of its indispensability to deduction and reasoning in general. This is one of the principal
messages in Hale’s paper. In this chapter I take the indispensability thesis further. I demonstrate,
by extrapolating from Hale’s paper, exactly how and where in the reasoning process we use basic
knowledge, why that knowledge should be countenanced as logical knowledge, and why this
species of non-inferential knowledge of logic is distinct from other non-inferential knowledge.
We should conclude with Hale, that non-inferential knowledge of logic is essential to the
reasoning process. I shall then expand on this in the final section by arguing that not only is non-
inferential knowledge essential, but non-inferentialism is also the only method available to us
suitable for knowing the correctness of basic rules of logic.

3.2 Indispensability2

We cannot, as we have seen, simply suppose that an operator is impossible to doubt—it should
be demonstrated in an argument. In§2.6, I presented my adapted argument from Hale which
showed, via a chain of reasoning, that tonk is nonconservative. That argument is not circular if
we acknowledge the difference between an argument that uses or reasons according to an
operator versus one that reasons about that operator. When we reason about an operator, we do
not use it in the same way Achilles proposes we do with his premise (C): we may therefore
conclude something about the operator without reasoning with it viciously. The following is the
chain of reasoning that Hale uses to demonstrate the nonconservativeness of tonk. It is a more
explicit version of the argument I presented in §2.6, and it is important and useful to present here
because it highlights the application of basic rules in their most primitive forms:

(1) Tonk-introduction allows you to make inferences of the form ‘A, so A tonk B’.
so: (2) If the inference: ‘p so p tonk q’ is of the form ‘A, so A tonk B’, then tonk-
introduction allows you to make it.
(3) The inference: ‘p, so p tonk q’ is of the form ‘A, so A tonk B’.
Hence: (4) Tonk-introduction allows you to infer ‘p tonk q’ from ‘p’.
Further: (5) Tonk-elimination allows you to make any inference of the form ‘A tonk B, so
B’.
So: (6) If the inference: ‘p tonk q, so q’ is of the form ‘A tonk B, so B’, then tonk-
elimination allows you to make it.
(7) The inference: ‘p tonk q, so q’ is of the form ‘A tonk B, so B’.
Hence: (8) Tonk-elimination allows you to infer ‘q’ from ‘p tonk q’.
Hence: (9) Tonk-introduction and tonk-elimination allow you to infer ‘q’ from ‘p’.
Hence: (10) The tonk rules together allow you to derive any conclusion from any
premise.3

Among the observations we might make about this chain of reasoning, of note for our purposes
is the fact that (2) and (6) are the result of an application of universal quantifier elimination such
that we recognise a particular statement to be an instance of its universal form. Furthermore,
MPP is used to derive the conclusions found in (4) and (8) from the conditionals expressed in (2)
and (6), via the minor premises expressed in (3) and (7).

We have already discussed how Hale seeks to avoid inferential reasoning that uses basic logical
operators because of an inability to offer-up a satisfactory inferential explanation of their
correctness. Another aspect of Hale’s non-inferentialism can be clearly understood by examining
his view about what is actually going on in premises (3) and (7); despite the obviousness of these
two premises, we should still enquire into their justification, that is, a justification of our
entitlement to mediate recognition of a particular statement as being an instance of some general
rule. That would essentially constitute a justification of UQI&E. But what sort of premises
should we put forth in such a defense? There do not seem to be any. Additionally problematic is
the fact that an inferential justification would lead to the usual regress or circularity. In a footnote
Hale therefore argues:

Our recognition of their correctness must, it seems, be a non-inferential matter. That is,
the right answer is just the one we should naturally give, viz. that we can just see that the
particular inference is of the displayed general form...what is involved here is a species of
non-inferential intellectual recognition—which we may call rational insight, and which
has an indispensable role to play whenever we operate with rules of inference.4
The demonstration contained in (1) - (10), as well as this last quote reveal the significant aspects
of Hale’s epistemology that I wish to highlight: correctness of a small number of items of logical
knowledge is a non-inferential matter. It is natural to make a few of the basic inferences that we
do, and we just ‘see’ that these inferences are correct. This amounts to a species of ‘non-
inferential intellectual recognition’. As I mentioned in §2.6, Hale does not think that the
recognition of something more complex such as validity may be realised in this way, yet simple
inferences such as those inferred by UQIorE may be countenanced as the products of rational
insight.

If we subscribe to this point of view thus far, then it is not a large leap to conclude that the sort of
inference pattern just discussed plays an indispensable role in an argument for the non-
correctness of tonk. It does not prove that these basic inferences must be used in reasoning, but I
cannot think of a non-circular way of expressing tonk’s non-correctness without using them, and
that takes us a good way towards the qualification of being indispensable.

Now I must answer my own question: how does the above figure into a justification for MPP? It
may be justified in the same manner as UQIorE. §2.6 discussed the stipulation that a sound rule
of inference will be general in the sense that a conclusion of some specified general form may be
drawn from premises of some specified form. A sound rule shall also be conditional such that
given premises of a particular form, we may draw conclusions of a corresponding form. If we are
going to analyse the pedigree of an argument to its most primitive level as demonstrated in (1) -
(10), then Hale points out that we are going to have to reason from explicit formulations of those
rules (exemplified in (1) - (10) by lines (1) and (5)). Mediating particular instances of UQI&E or
MPP from their explicit formulations will, once again, be an indispensable step in the reasoning
process. MPP and UQI&E therefore constitute a minimal kit of inference rules required to reason
about the soundness of rules of inference.

And how does (1) - (10) and the preceding vein of reasoning show that the items in the minimal
kit are impossible to doubt? Well, we showed how they are impossible to doubt by
counterexample because they were the rules that we accepted. Furthermore, we cannot accept a
bogus rule of inference if we can demonstrate its inconsistency and/or non-conservativeness.
With this achieved, we have a viable way to distinguish the rules of inference that yield sound
conclusions from those that don’t. The former shall be the correct rules. In consideration of an
argument presented to refute the correctness of the rules contained in the minimal kit, we shall
have had to use those rules to argue for their non-correctness, and this is something that we are
not permitted to do. It therefore seems as though a viable defense of the indispensability of MPP
and UQI&E has been presented.

The final task in this project involves linking indispensability with correctness. If Hale is right, it
would be the most significant achievement of his paper. How do we move from the former to the
latter? This question motivates Hale’s thesis about the possibility of leveraging the findings of
Project B to successfully achieve the goal set out in Project A. But it seems as though we would
have to present an argument in order to support this claim, an argument something similar to the
following: ‘because certain basic rules are immune to doubt or are indispensable, claims about
their correctness will be true’. This argument cannot be supported, as we have seen, without
applying some sort of pragmatically-circular reasoning. To avoid this problem, Hale suggests a
solution that has already been aptly described in this thesis by the first moral presented in§1.5,
the one that says that rules are not premises. Something that is impossible to doubt need not be
taken as a premise. Knowledge of the correctness of simple logical constants is impossible to
doubt, and instead of proffering this fact as a premise in a argument, we take a non-inferential
stance on this sort of knowledge by claiming that there is no inferential gap to be traversed. I
reiterate one of Hale’s final comments in his paper to conclude my summary of his arguments:

Perhaps it can be maintained that there is, simply, no inferential gap to be traversed—that
is, that knowledge can, much as Descartes thought, consist in believing truly something
which it is impossible rationally to doubt.5

I have shown that non-inferential knowledge can be bifurcated into two distinct types: the
logical, and the non-logical. I also explained why knowledge of non-inferential, non-logical
statements is dubious: because none of it is universal, necessary, and completely general. Basic
logical rules do have these qualities however. And when we tie this into the thesis that they are
indispensable in reasoning, we come away with a viable non-inferential grounding of basic
logical rules.

3.3 The Primacy of Non-Inferential Knowledge of Logical Laws

In a chain of reasoning, if something comes first then anything that follows from it cannot be first
as well. In this sense, whatever comes first is basic. The strength of my thesis—that non-
inferential knowledge of logic is the only sort we can have about the basic rules of logic—
comes from (a.) the weakness I find in the argument for inferentialism with respect to its
inability to deliver anything but circular or regressive results, and (b.) the strength of Hale’s
account.

Boghossian and Dummett share an endorsement for the suasive versus explanatory distinction
and both believe that the latter notion advances their cause. We recall from Chapter 2 that
suasive justifications convince a doubtful explainee, in this case of correctness. Explanatory
justifications demonstrate correctness to an explainee who already endorses that correctness, but
simply wants to see the justification. The doubtful explainee, like the Tortoise, refuses to accept
circular or regressive demonstrations. Both Dummett and Boghossian agree that there is no
persuading this sort of sceptical attitude. But explanatory justifications facilitate inferential
justifications. Dummett says of the non-suasive explainee:

He does not need to be persuaded of the truth of the conclusion; what he is seeking is an
explanation of its being true. An explanation frequently takes the form of a deductive
argument, in which the conclusion is the fact to be explained. There is therefore no
uncertainty about the conclusion, which we already know; and often the best reason for
believing the premisses is that they offer an explanation for the conclusion’s being true.6

Dummett goes on to argue that an explanatory justification, which is a pragmatically circular

justification, has some value. In fact, he says that it’s the best we can do, or, to use other words,
it’s the most basic sort of justification for basic logic that is possible. Whether he believes an
explanatory and pragmatic justification is the most basic of all justifications or merely the most
basic sort of inferential justification Dummett does not say. But there is little if anything in his
paper Circularity, Consistency, and Harmony that endorses non-inferential justification. So I take
it that Dummett’s position on justification for basic knowledge is anchored by his explanatory
stance about justification described in this most recent quote. Boghossian’s story leads him to a
near-identical conclusion:

To put this point another way: we must recognize a distinction between two different
sorts of reason—suasive and non-suasive reasons. And we have to reconcile ourselves to
the fact that in certain areas of knowledge, logic featuring prominently among them, our
warrant can be at most non-suasive, powerless to quell sceptical doubts.7

Boghossian’s and Dummett’s stories are right-on to somebody who accepts (pragmatically)
circular reasoning. One reason for this might be because pragmatically circular reasoning that
uses the basic rules endorsed by classical logic yields valid results. But I do not believe that
circular reasoning has any place in a justification of basic logic because the justification falls
short of rigour. Rigour requires that proof-checking can be effected through application of rules
of inference or axioms. But the justification sought in this case is one for the very rule of
inference accepted. There are numerous paradoxes associated with the quality of being self-
referential, and a pragmatically circular justification of a basic rule of logic is just that. My
conviction is that this denies the reasoner the right to accept pragmatic circularities as
justification for basic laws of logic. The moment we prohibit circularity, Boghossian’s and
Dummett’s arguments fizzle.

I shall now describe the flaw in their reasoning, and I believe what follows here to be the most
important contribution this thesis makes to enquiry into basic logical knowledge. The simplest
way to explain the flaw is that the supposition that MPP is correct, held by an explainee under
explanatory—but not suasive— circumstances is tantamount to non-inferential knowledge of
correctness. This is why: if a reasoner is to ask for an inferential justification of the correctness of
an item of knowledge, an item about which she is already convinced of correctness, then her
conviction is prior to any argument that demonstrates that correctness. Believing or knowing that
something is correct prior to a demonstration of the correctness is non-inferential knowledge.
Knowledge arrived at by reason is, as we know, inferential knowledge, but knowledge we
possess prior to reasoning must be non-inferential knowledge. Non-inferential knowledge about
basic logical rules is therefore basic because it lies prior to the inferential knowledge that
depends on it.

It is possible to establish basic knowledge as correct in a non-inferential way without depending

on any inferential steps or attitude in the process. Yet it is impossible to achieve the opposite,
that is, to justify basic knowledge in inferential fashion without the application of some non-
inferential knowledge. For consider: Hale showed us how we may countenance the basic rules as
correct by showing that they are impossible to doubt and subsequently that their correctness need
not be enumerated as premises in reasoning. No inferential step figures here. If we examine
Dummett’s inferential strategy on the other hand, explanatory rather than suasively circular
explanations are the best outcomes we can hope for. Furthermore, the explanatory account, being
pragmatically circular, presupposes the correctness of the rule.
For these reasons, an inferential determination of the correctness of basic rules of inference
cannot be basic because that determination depends on non-inferential knowledge. Hale has
described non-inferential knowledge as a ‘minimal kit’ of default-justified rules. This kit, which
we have shown to be indispensable to deduction is therefore the only sort of logical knowledge
we may countenance as basic with any confidence.

1 See his paper Informal Rigour and Completeness Proofs, International Colloquium in the
Philosophy of Science. Amsterdam: North-Holland Press, 1967. Pg. 138.

2 This explanation of infinitism comes from Klein in his (1999).

3 Rules of inference also exist that are valid in some contexts but not in others, for example the
inference rule {¬¬ α → α} (viz., the rule of double-negation) is valid in classical logic, but not in
intuitionistic logic.

4 Carroll, (1895).

5 Philie, (2006), 185.

6 Hale (2002), pg. 280.

7 Kant (1998), B10.

8 The apriority of this statement is also apparent because we do not need to appeal to intuition or
experience in order to establish the truth of the statement (on Kant’s account).

9 Frege (1953), §3.

10 ibid.

11 This process is discussed in Frege (1997), pg. 309

12 Frege (1953), §88. 13Frege (1967), §9.

14 Dummett (1991), pg. 39 - 40.

15 ibid, pg. 40 - 41.

16 ibid, pg. 40. Dummett’s emphases, my clarification and adaptations in square braces.

17 ibid, pg. 41.

18 Frege (1953), §88.

19 ibid.
20 Pascal Engel (2007), Dummett, Achilles, and the Tortoise, Open Court, 728 - 745.

1 Prior (1960), pg. 38.

2 Prior (1960), pg. 39.

3 ibid.

4 Note that tonk’s introduction rule is coextensive with disjunction-introduction, and its
elimination rule with conjunction-elimination.

5 Belnap (1962), pg. 130.

6 Belnap (1962), pg. 131.

7 Transitivity functions largely like the cut rule in propositional logic, where the cut rule states
that If Φ |-0 δi for each i = 1, 2,..., k and Ψ ∪ δ1, δ2,..., δk |-0 α then Φ ∪ Ψ |-0 α.

8 summarized by Belnap in his (1962), pg. 132.

9 Other important differences between statements and sequents exist as well. For instance, the
order in which formulas occur in a sequent may have bearing on the resultant sequent inferred.
Additionally, the possibility of a formula being repeated in a sequent exists as well. This is in
contrast to sets where the order of set members is insignificant, and multiple occurrences of
members is prohibited.

10 In the case of tonk for example, the rule ‘for any P, P-tonk-Q’ , explains how to use/introduce
the rules geverning the tonk operator.

11 Dummett (1994), pg. 194.

12 A metalinguistic cognate of conservatism explained in the next section.

13 Engel, pg. 730 - 731.

14 Engel (2004), 727.

15 And the conditions for us as observers, and for the subjects, as the objects of our enquiry
would not be the same still because our language and theory of knowledge emerged as a result of
groups of people developing them in this world, rather than off in some Platonic cave.

16 in his book The Logical Basis of Metaphysics, Harvard University Press, (1991).

18 Dummett (1994), 203.

19 The contingent nature of tonk might be apparent to a seasoned logician or somebody with
sharp analytic skills. Yet it is possible to characterise some other operator in the metalanguage
possessing introduction and elimination rules that are too complex for a reasoner to immediately
intuit its status as valid.

20 I am referring here to the way introduction and elimination rules are delivered in English, as
presented for the case of tonk in §2.1.

21 In his paper A Counterexample to Modus Ponens. The Journal of Philosophy, Vol. 82, No. 9.
1985. pp. 462 - 471. Hartry Field argues that counterexamples to MPP exist. I do not discuss his
observations here because I do not think it will help this project to study a counterexample to a
rule, viz., MPP, that is uncontestably valid to the extent that it is. Additionally, the
counterexample Fields uses contains another conditional embedded into its antecedent, and I
think that I would be straying too far afield to enter into an enquiry that enumerates the
differences existing between the example Fields proposes compared to a more conventional
exemplar of MPP.

22 Dummett (1991), 210.

23 Dummett (1991), 211.

24 Boghossian (2001) 245.

25 Boghossian, (2001), 246.

26 Boghossian (2001), 249.

27 ibid.

28 ibid.

29 Paul Boghossian, Analyticity Reconsidered, Nouˆs, Vol. 30, No.3, 373 - 387.

30 Boghossian (2001), 252.

31 Hale (2002).

32 Hale (2002), 289.

33 Hale (2002), 293. By ‘neither’, Hale must undoubtedly be referring to the introduction and
elimination rules that characterise tonk.

34 My adaption of Hale (2002), 295.

35 This argument can be interpreted as making the following point(s): ‘(P1.) For a rule to be
conservative, the following qualities must obtain..., (P2.) Here is an argument that uses the
operator in question. (P3.) The operator’s rule(s) do not satisfy the conditions for
conservativeness in this argument. (C) Therefore, This operator is bogus.’ Framed correctly, the
argument contained in this footnote could be turned into a deductive one that reasons according
to (i.e. uses) MPP.

36 Hale (2002), 297.

37 Hale(2002), 298.

38 Hale (2002), 302.

1 Examples of non-transmittable knowledge might be knowledge of one’s own pain or pleasure.

A somewhat more tangible example, which I credit to Tim Kenyon, is a person’s knowledge of
their ‘lost sock circle’, viz., the collection of socks a person has lost in their lifetime. The circle
has a distinct and detailed status to its owner, but only represents a vaguely defined and
undifferentiated collection to anyone else.

2 In this section I introduce the abbreviations UQI, UQE, and UQI&E, UQIorE which shall stand
mutatis mutandis for universal quantification introduction and/or elimination.

4 ibid. Hale’s emphases.

5 Hale (2002), 302.

6 Dummett (1994), 202, his emphasis.

7 Boghossian, 253, his emphasis.

BOOK REVIEW
Sci & Educ (2013) 22:1257–1263
Matthias Steup and Ernest Sosa (eds): Contemporary Debates in Epistemology
Kevin McCain

... 3 Foundational Knowledge

This section begins with a debate between Peter Klein and Carl Ginet over the merits of
infinitism as a response to the regress problem. The regress problem arises if all justification is
inferential. If all justification is inferential, then in order for S’s belief that p to be justified it
must be inferred from S’s belief that q. However, in order for S’s inferring that p from her belief
that q to justify p her belief that q must itself be justified. S’s justification for believing q will
come from its being inferred from another belief that r, which itself must be justified and so on.
Peter Klein attempts to support the infinitist response, which claims that this sort of infinite
regress is required for justified belief and that it actually does produce justification. Klein’s
argument in support of infinitism has three main steps. First, he argues that neither
foundationalism, the view that the regress terminates with justified beliefs whose justifiers do not
themselves require justification, nor coherentism, the view that there is not an infinite regress
because there is self-support among one’s justified beliefs, can adequately respond to the regress
problem. Second, Klein presents and motivates infinitism. Third, Klein responds to two common
objections to infinitism. Carl Ginet, however, argues that infinitism is implausible. Ginet begins
by first offering examples, which he claims demonstrate that foundationalism can solve the
regress problem. After this, Ginet argues that infinitism faces two serious problems. First,
infinitism faces the problem of making sense of how an infinite regress of justified beliefs is
available to the believer. Ginet doubts that the relevant notion of availability can be secured.
Second, Ginet argues that infinitism entails that inferences can actually generate justification for
their conclusions which the premises of the inference lack. Ginet maintains that it is clearly false
that inferences can generate justification.
Infínitism Regained
JEANNE PEIJNENBURG
Mind, Vol. 116.
July 2007

Consider the following process of epistemic justification: proposition E0 is made

probable by E1, which in turn is made probable by E2, which is made probable by E3,
and so on. Can this process go on indefinitely? Foundationalists, coherentists, and
sceptics claim that it cannot. I argue that it can: there are many infinite regresses of
probabilistic reasoning that can be completed. This leads to a new form of epistemic
infinitism.

Today epistemologists are usually probabilists: they hold that epistemic justification is mostly
probabilistic in nature. If a person S is epistemically (rather than prudentially) justified in
believing a proposition E0, and if this justification is inferential (rather than noninferential or
'immediate'), then typically S believes a proposition E1 which makes E0 probable.1

How to justify E1 epistemically? Again, if the justification is inferential, then there is a

proposition E2 that makes E¡ probable. Imagine that E2 is in turn made probable by E3, and that
E3 is made probable by E4, and so on, ad infinitum. Is such a process possible? Does the 'ad
infinitum' makes sense? The question is known as the Regress Problem and the reactions to it are
fourfold. Sceptics have hailed it as another indication of the fact that we can never be justified in
believing anything. Foundationalists famously argued that the process must come to an end in a
proposition that is itself noninferentially justified. Coherentists, too, maintain that the infinite
regress can be blocked, but unlike foundationalists they hold that the inferential justification need
not be linear and may not terminate at a unique stopping point. Finally, infinitists have claimed
that there is nothing troublesome about infinite regresses, the reason being that an infinite chain
of reasoning need not be actually completed for a proposition or belief to be justified. One of the
leading infinitists, Peter Klein, has even stated that the requirement that an infinite chain must be
completed 'would be tantamount to rejecting infinitism' (Klein 1998, p. 920).

In this paper I will defend a view that is different from all four. Against sceptics,
foundationalists, and coherentists I will show that an infinite regress can make sense; against
infinitists I will show that beliefs may be justified by an infinite chain of reasons that can be
actually completed.

Suppose that E0 is made probable by E1. The probability of E0, P(EO), can be calculated by
means of the rule of total probability:

(1) P(E0) = P(E0|E1) P(E0) + P(E0|~E1) P(~E1)

where P(E0|E1) is the probability of E0 given E1 and P(E0|~E1) is the probability of E0 given
not-E1. Since E0 is made probable by E1, E0 is more probable if E1 is true than if E1 is false, so
we have

P(E0|E1) > P(E0|~E1)

If E1 is in turn made probable by E2, we must of course repeat the rule:

(2) P(E1) = P(E1|~E2) P(~E2) + P(E1|~E2) P(~E2)

where again it is assumed that P(E1|E2).

Can we continue this repetition, thus allowing for propositions made probable by other
propositions, made probable by still other propositions, and so on, ad infinitum? Supporting
foundationalism, Richard Fumerton has claimed that we cannot, since 'finite minds cannot
complete an infinitely long chain of reasoning and so, if all justification were inferential, no-one
would be justified in believing anything at all to any extent whatsoever' (Fumerton 2006a, p. 40;
Fumerton 2006b, p. 2; Fumerton 2004, p. 150).

Fumerton does not say why he believes that finite minds cannot complete an infinitely long chain
of reasoning. Presumably he thinks that such a task would be infinitely complicated, or would
take an infinite time to finish. Such worries are understandable. If P(E0) is the outcome of an
infinite regression, the calculation of P(E0) seems at first sight too lengthy and too complex for
us to complete. After all, insertion of Equation (2), together with

(3) P(~E1) = P(~E1|E2) P(E2) + P(~E1|E2) P(~E2)

into the right-hand side of Equation (1) leads to an expression with four terms, namely

(4) P(E0) = P(E0|E1) P(E1|E2)P(E2) + P(E0|~E1) P(~E1|E2)P(E2) + P(E0|E1) P(E1|

~E2)P(~E2) + P(E0|~E1) P(~E1|~E2)P(~E2)

A repetition of this manoeuvre to express P(E2) and P(~E2) in terms of P(E3) and P(~E3) would
produce no less than eight terms. After n + 1 steps, the number of terms is 2^n+1, which yields
an ungainly expression that seems hard to calculate in a simple, closed form.

There is however an easy way to reduce this complication of the rapidly increasing number of
terms. Replace P(~E1) in Equation (1) by 1 - P(E1), and then rewrite this equation as

(5) P(E0) = P(E0|~E1) + [P(E0|E1) - P(E0|~E1)]P(E1)

A similar treatment can be applied to Equation (2), which then becomes

(6) P(E1) = P(E1|~E2) + [P(E1|E2) - P(E1|~E2)] P(E2)

and so on. Although these changes may seem minimal, their advantages are significant. For they
enable us to obtain a closed and completable expression for P(E0), not only when the number of
steps is finite, but also when it is infinite. This can be further explained as follows.

Clearly we can only use Equation (5) to compute the value of P(E0) if we know the value of
P(E1). Similarly, we can only use (6) to compute P(E1) if we know P(E2). So knowing P(E1) is
necessary for knowing P(E0), knowing P(E2) is necessary for knowing P(E1), knowing P(E3) is
necessary for knowing P(E2), and so on. If we generalize (5)-(6) to

(7) P(Em) = P(Em|~Em) + [P(Em|Em+1) - P(Em)|~Em+1) P(Em+1)

which gives the probability of Em, the conclusion remains unaltered: we need to know the value
of P(Em+1) in order to be able to compute the value of P(E^). Now let us call the probability of
Em α if Em+1 is true, and ß if Em+1 is false:

P(Em|Em+1) = α and P(Em|~Em+1) = ß

For simplicity we assume that neither of these two conditional probabilities depends on m, in
other words, α and ß are the same for any m. For example, we might have α = 0.9 and ß = 0.3, or
any other pair of numbers that satisfies 1 > α > ß > 0. With α and ß in place. Equation (7)
becomes

(8) P(Em) = ß + (α – ß)P(Em+1)

Being special cases of (7), Equations (5) and (6) can also be written

(9) P(E0) = ß + (α - ß)P(E1)

(10) P(E1) = ß + (α - ß)P(E2)

Let us now apply the rule expressed in these equations to the finite case, beginning with m = 0,1,
2. This move gives us a finite series, consisting of two steps:

(11) P(E0) = ß + (α - ß)P(E1)

= ß + (α - ß)[ß + (α - ß)P(E2)]

We can continue this process for any finite m = 0,1, 2, 3, ... , n. The result is still a finite series,
and moreover one that can be summed explicitly, yielding

(12) P(E0) = ß + (α - ß) [ß + (α - ß)[ß + (α - ß)[...P(En+1)]...]]

= ß[1 + (α - ß) + (α - ß)^2 + ... (α - ß)^n] + (α - ß)^n+1P(En+1)
= ß/1- α - ß + (α - ß)^n+1[P(En+1) - /1 – α + ß]

Here the value of P(E(1) is ultimately derived from one single term, the remainder term (α -
ß)^n+1P(En+1), containing the probability of En+1 (see the second line in Equation (12)).
Clearly, the value of this remainder term cannot be computed unless we know this probability.
Does this mean that Fumerton and other foundationalists would be right if they were to claim
that Equation (12) can only be solved if we assume that the value of P(En+1) is known and
hence that En+1 is noninferentially justified?

The answer is negative. To see this, let us consider the infinite case. The standard way to
investigate the convergence of an infinite series is first to look at a finite series of, say, n+1 terms
only, with a remainder term, and then to investigate what happens as n tends to infinity.
Applying this procedure to Equation (12), we observe that, since 0 < α - ß < 1, the factor (α -
ß)^n+1 becomes smaller and smaller as n becomes larger and larger. In the formal limit that n
goes to infinity, we find that the series has an infinite number of terms, and that the terms in the
second and third lines of Equation (12) that contain the unknown P(En+1) tend to zero, and
hence disappear completely.

In the limit of an infinite number of terms in the series, corresponding to an indefinite iteration of
Equation (7), we find

P(E0) = ß/1- α – ß = 0.3/1 – 0.9 + 0.3 = 0.75

with the values given above for α and ß, 0.9 and 0.3. Thus, even after an infinite number of steps
in the inferential justification, the value of P(E0) can be exactly calculated: it is 0.75. The
justification is, although infinite, perfectly computable and completable.

One might object that the argument developed above binges on a very special case. For in
demonstrating that an infinite regress can make sense, and that justification by an infinite chain
of reasoning can indeed be carried out, I have assumed that the conditional probabilities a and ß
remain the same throughout the entire process. Such an assumption is rarely fulfilled. It is very
unusual that the degree with which a proposition E0 is made probable by E1 is identical to the
degree with which E1 is made probable by E2, and E2 is made probable by E3, and so on, ad
infinitum. The fact that, in addition, the probabilities P(E0|~E1), P(E1|~E2), P(E2|~E3), etc. are
also identical only underlines the special nature of this case.

My answer to such an objection would be twofold. First I would point out that one
counterexample is enough to refute the foundationalist's claim that all inferential chains of
reasoning must come to an end. Similarly, one counterexample is enough to confute the claim of
the infinitists that no infinite chain can be actually completed. The fact that this counterexample
presupposes special situations is not really relevant.

Second, it is relatively easy to construct counterexamples without making the assumption that α
and ß are the same for each step. For suppose that the conditional probabilities P(Em|Em+1) and
P(Em|~Em+1) do change as the index m varies, which we indicate by adding the index m to α
and ß:

P(Em|Em+1) = αm and P(Em|~ Em+1) = ßm

The second line of Equation (12), where a and ß are the same, is a special case of

(13) P(E0)= ß0 + (α0 – ß0)ß1 + (α0 – ß0)(α1 – ß1)ß1 + … + (α0 – ß0)(α1 – ß1) … (αn-1
 – ßn-1)ßn + (α0 – ß0)(α1 – ß1) … (αn - ßn)P(En+1)

where αm and ßm vary with m. Equation (13) can be proved by mathematical induction, but we
will not stop to do that here. Now, since 1 > α - ß > 0, we might erroneously think that

(14) 1 > αm – ßm > 0

would be enough to make the remainder term in the last line of (13) vanish in the limit that n
goes to infinity, and hence might provide us with counterexamples of the sort that we are looking
for. However, condition (14) does not guarantee the vanishing of the remainder term, let alone
does it ensure summability of the entire series. We need something stronger, for example the
constraint that am – ßm be uniformly bounded from above by a constant, c, that is strictly less
than 1:

(15) 1 > c > αm – ßm > 0

Under constraint (15), Equation (13) does the trick. For now not only does its remainder term
tend to zero as n tends to infinity, it is also completable: we can find instances of ßn and αn – ßn
such that the sum that (13) expresses can be explicitly calculated. These instances are our
counterexamples, and there are infinitely many

1 This relation of making probable might be conceived externalistically or internalistically. In the

first case the emphasis will be on an objective interpretation of probability, in the second case a
subjective interpretation seems more appropriate. For the present argument, it does not matter
which stance one takes.

2 Here is one: if αn - ßn = α/(n + i) and ßn = b/(n+1), where a and b are positive constants, then
P(E0) = b(e^a-1)/a.

3 I thank an anonymous referee and Tom Stoneham, in his role as associate editor of Mind, for
stimulating comments. Thanks also to David Atkinson for help with mathematical details.
Ineffectual Foundations: Reply to Gwiazda
Jeanne Peijnenburg
Mind, Vol. 11

In an earlier paper I argued that there are cases in which an infinite probabilistic chain can be
completed (Peijnenburg 2007). According to Jeremy Gwiazda, however, I have merely shown
that the chain in question can be computed, not that it can be completed (Gwiazda 2010).
Gwiazda thereby discriminates between two terms that I used as synonyms. In the present paper I
discuss to what extent computability and completability can be meaningfully distinguished.

1. Introduction

Suppose a proposition or belief E0 is probabilistically supported by E1, which in turn is

probabilistically supported by E2, and so on ad infinitum. Can E0 have a definite, nonzero
probability value? Foundationalists like Bertrand Russell and C. I. Lewis thought it cannot. They
argued that the probability of E0 in this case always tends to zero, since it is an infinite product
of conditional probabilities, all of which lie strictly between 1 and 0:

(1) P(E0) = P(E0|E1) x P(E1|E2) x P(E2|E3) x …

= 0.

It was Hans Reichenbach who first pointed out the mistake in this argument. He explained that
the probability of E0 is not given by the infinite product (1), but by the infinite recursion (2):

(2) P(E0) = P(E0|E1)P(E1) + P(E0|~E1)P(~E1)

P(E1) = P(E1|E2)P(E2) + P(E1|~E2)P(~E2)
P(E2) = P(E2|E3)P(E3) + P(E2|~E3)P(~E3)
…

Both Russell and Lewis accepted Reichenbach’s correction. Lewis, however, stressed that this in
no way affected the conclusion that the probability of E0 tends to zero, since according to him
(2) likewise converges to nought.1 In Peijnenburg 2007 I showed that (2) can yield a value other
than zero, thereby countering Lewis’s original argument, as well as modern versions of it. For an
example of the latter, we may look at Peter Klein’s comment on Richard Fumerton’s book
Metaepistemology and Skepticism. Klein notes:

There is one other objection to infinitism that I did not find in Fumerton’s book but which
seems to jump from the pages.…Consider the expression ‘e makes probable p’. If the
relevant notion of ‘making probable’ is some sort of statistical one then the longer the
chains, the less likely the conclusion. If the chains were infinitely long, the conclusion
would be so improbable that it would not be reasonable to accept it. (Klein 1998, p. 925)
If we interpret ‘would be so improbable’ as ‘would tend to zero’, then we have exactly Lewis’s
argument.

By demonstrating that P(E0) can have a nonzero probability even if its supporting chain is
infinitely long, I took it I had shown also that (2) can actually be completed. Jeremy Gwiazda
however claims that I merely demonstrated that (2) can be computed (Gwiazda 2010). He thus
draws a distinction between two terms that I used as synonyms. To what extent can we
meaningfully distinguish between computability and completability? ‘Computability ’ is a
reasonably well-defined term, although it leaves open whether the actually computed conclusion
P(E0) must be a rational number or may perhaps be an irrational or even a transcendental
number (cf. Herzberg 2010). ‘Completability ’, on the other hand, seems to be more vague.
Gwiazda’s definition of it—‘demonstrating completability ’ means ‘showing that…an infinite
chain of reasons…can be actually completed’ (Gwiazda 2010, pp. 1123–24)—is not very
helpful, since the definiendum is contained in the definiens. Let us therefore take a closer look at
the works of the people who are central to the present discussion and from whom I borrowed the
term ‘completable’, namely Fumerton and Klein. By observing how they use the term, we might
form an opinion as to what Gwiazda could have had in mind.

Fumerton writes:

[For a foundationalist] having justification can only be constituted by a chain of

reasoning that one actually completes, as opposed to [what would satisfy an infinitist like
Klein — JP] one each link of which could be completed. (Fumerton 2004, p. 150)

Here Fumerton distinguishes between ‘completing a chain’ and ‘completing each link’. In order
to see what this might mean, imagine a finite chain C, consisting of three links:

C = ½ + ¼ + 1/8

What can it mean to complete each link in this chain? Completing the first link, I would say,
means that we substitute 1/2 for C. Completing the second link then means adding 1/4 to 1/2, and
completing the third adding 1/8 to 3/4. What does it mean to complete the entire chain? Nothing
more nor less, it seems to me, than concluding that the value of C is 7/8. But concluding that C
equals 7/8 is no different from computing the chain or computing the value of C; so in this finite
case, computing and completing seem entirely equivalent. The matter appears to be similar when
the chain is infinite:

C’ = ½ + ¼ + 1/8 + …

Here, too, completing the nth link means adding (1/2)^n to what we already have, whereas
completing the entire chain means concluding that the value of C’ is 1. And the latter is precisely
to say that the chain C’ has been computed.

So what exactly is Gwiazda aiming at when he intimates that an infinite epistemic chain might be
computed, but can never be completed, and hence can never have any supporting or justificatory
role? I am able to think of three possible answers. In the next three sections I consider them in
succession, arguing that each of them has its inadequacies.

2. An infinite number of beliefs

The first possible answer is that an infinite chain cannot be completed (although it can
sometimes be computed) because our mental make-up is such that we cannot have an infinite
number of beliefs. This answer would bring Gwiazda’s position close to that of, for example,
Noah Lemos, who argued that a belief can never be justified by an infinite chain, because this
would have the impossible consequence that the number of our beliefs is infinite:

One difficulty with [an infinite chain] is that it seems psychologically impossible for us to
have an infinite number of beliefs. If it is psychologically impossible for us to have an
infinite number of beliefs, then none of our beliefs can be supported by an infinite
evidential chain. (Lemos 2007, p. 148)

However, Fumerton (whose position Gwiazda seems to share) repeatedly stressed that we can
have an infinite number of beliefs:

There is nothing absurd in the supposition that people have an infinite number of justified
beliefs. (Fumerton 2006, p. 49)

Klein is right that we do have an infinite number of beliefs. (Fumerton 2001, p. 7)

…there probably is no difficulty in supposing that people can have an infinite number of
beliefs. (Fumerton 1995, p. 140)

So the assumption that we have only a finite number of beliefs cannot be the reason why
Fumerton, or Gwiazda for that matter, denies that infinite chains are completable.

Perhaps Gwiazda’s idea is that, although we might have an infinite number of beliefs, we will
never be able to actually spell them out one by one. But this idea would make ‘completability ’
unduly restrictive. If ‘completing a chain’ were to mean ‘spelling out each of its links’, then
hardly any chain could be completed, long finite ones included. A fortiori, hardly any belief or
proposition could be justified. As Peter Klein, pondering what he calls the Completion
Requirement, puts it:

It is not required that [a person] S actually have carried out that process by rehearsing the
justification out loud or sotto voce. That’s a good thing, too. For…the Completion
Requirement would rule out most, if not all, of my beliefs as justified—even on
foundationalist grounds. (Klein 1998, p. 921)

The question whether we can or cannot have an infinite number of beliefs is a tricky one, and it
seems ill-posed. For it presupposes that we can have a method for individuating and counting
beliefs, and this is at present highly disputable. My preference is to remain neutral on this matter.
This seems perfectly possible: you might claim that someone can complete an infinite series
without committing yourself to the claim that this person must have an infinite number of beliefs.

3. The limit cannot be reached

A second possible way of differentiating between computing and completing hinges on the
notion of a limit. If an infinite series is suitably convergent, we can compute it in the sense that
we can calculate its limit, for example the number 1 in the case of chain C’ and the value of
P(E0) in the case of chain (2). However, it is impossible to actually reach this limit in a finite
number of steps, although we come increasingly close to it the longer we make the chain.
Perhaps Gwiazda is thinking of a parallel with Zeno’s paradox of Achilles and the tortoise: we
cannot reach the limit of C’ or of (2), in much the same way that Achilles cannot reach the point
where he draws level with the tortoise. But we can compute that point, and Achilles can come
closer and closer to it; in the same vein, we can come closer to 1, or to the actual value of P(E0).
If this is indeed what Gwiazda has in mind, then two observations must be made.

First, under this interpretation Gwiazda’s argument resembles that of James Thomson in his
paper on super-tasks (Thomson 1954). Imagine that C’ represents the race track on which
Achilles and the tortoise compete: the so-called ‘Zeno points’ are 1/2, 3/4, 7/8, and so on. Since
Achilles has generously granted his rival a head-start, the tortoise will be at Zeno point 1 –
(1/2)^n+1 whenever Achilles is at Zeno point 1 – (1/2)^n. Thomson would heartily agree that we
can compute the point where the fleet-footed Greek catches up with the slowest of all animals:
this occurs at the limit point 1. However, he would deny that Achilles can complete the series C’,
if ‘completing’ is understood as ‘performing an infinite number of tasks’. In Thomson’s view,
not even a hero like Achilles can carry out such a super-task, for super-tasks are logically
impossible.

Paul Benacerraf showed that Thomson’s argument is invalid (Benacerraf 1962). The assumption
that Achilles occupies all the Zeno points one after another implies neither that Achilles reaches
nor that he fails to reach point 1—additional assumptions are needed to uphold one of these
implications. For example, with the extra assumption that Achilles runs at a constant speed, it
follows that Achilles indeed overtakes the tortoise. But if we additionally assume that Zeus, in a
capricious mood, transports Achilles back to the starting point in such a way that 1 is the first
point that Achilles does not reach, then our hero will never overtake the reptile. Conclusion:
Zeno failed to show that Achilles cannot reach the limit point 1, and Thomson failed to show that
completing an infinite number of tasks is logically impossible. The question of whether super-
tasks are logically possible or impossible is still open, and the answer depends in part on the way
in which a particular super-task is described.

Having said this, I come to my second observation. The image of Achilles struggling on the track
triggers associations that are not called forth by justification through infinite chains such as (2).
While it is not too odd to distinguish between computing a point in a race and actually reaching
that point in vivo, it does seem awkward to distinguish between computing a probability and
actually reaching it. Suppose I have calculated the probability of E0 on the basis of the series (2).
Then I have clearly computed the limit of this series, but have I not thereby also reached the
limit? What else is there to reaching the outcome of an infinite sum than calculating the value of
that sum? What more can we do to justify E0 probabilistically than to compute the (finite or
infinite) series of the probabilities in question?

Here is another way to phrase the difference between the Achilles paradox and our chain (2). In
the case of the Achilles, two questions have to be answered:

(a) Can we compute the point where Achilles draws level with the tortoise?

(b) Does Achilles actually reach that point?

The answer to (a) is ‘yes, we can’, while the answer to (b), as we learned from Benacerraf, can
only be given if extra information is provided. In the case of our probabilistic chain (2), the
question corresponding to (a) is:

(a’) Can we compute the probability that E0 is true?

Again the answer is ‘yes, we can’. But what question could correspond to (b)? There seems to be
none. While there is a fact of the matter over and above computing the point where Achilles
draws level with the tortoise, there seems to be no fact of the matter over and above computing
the probability that E0 is true. For what fact could that be? That E0 is indeed probabilistically
justified? But as I intimated above, probabilistic justification is computing the series of the
relevant probabilities. Is the fact perhaps that E0 is actually true? Or that we know that E0 is
actually true? Either option would betray a poor understanding of what it means to establish how
probable a proposition is. The very process of finding out the probability of a proposition is
orthogonal to settling its truth value; once we have computed the probability of E0, it is
inappropriate to then ask what the truth value of E0 actually is. In a probabilistic context, the
significant question is the value of the probability, not the question of whether the proposition is
in fact true or false.

4. The structural objection

But perhaps Gwiazda’s worries about completing an infinite chain find their breeding-ground in
an entirely different domain. Infinite regresses have often been the target of what Carl Gillett
calls the ‘Structural Objection’ (Gillett 2003, p. 713). Applied to our probabilistic regress (where
E0 is probabilistically justified by E1, which is probabilistically justified by E2, and so on), the
Structural Objection entails that the probabilistic justification of E0 is blocked by the very
structure of the regress itself. For if justification is deferred infinitely, then there is no source
available for E0 to be justified. In the words of Gillett (where we might read ‘probabilistically
justified’ or ‘probabilistically supported’ for ‘property H’):

Let us call an arbitrary member of the regress ‘En’ and consider what effect its addition
would have to the regress. Regardless of the nature of this entity, and its place in the
chain, all the members, E0…En–1, prior in the regress to En only instantiate H in virtue
of En being H. And, of course, this will be the situation regardless of how many more
entities are added to the regress. The question consequently arises how it could ever come
 to pass that any member of the chain has the property H? (Gillett 2003, pp. 712–13;
Gillet’s text has sn instead of En)

This question of where the property H comes from is a serious one, and it has no counterpart in
the case of the Achilles or other paradoxes dealing with infinity. This notwithstanding, the
question can be answered, and the answer follows directly from the structure of the probabilistic
regress as presented in Peijnenburg 2007. Suppose that our probabilistic chain were finite,
consisting of three links only:

(3) P(E0) = P(E0|~E1) + [P(E0|E1) - P(E0|~E1)]{P(E1|~E2) + [P(E1|E2) - P(E1|

~E2)]P(E2)}

In (3), the target proposition E0 receives much of its probabilistic justification from E2, the
source or foundation of this chain, which itself has a certain probability, P(E2). However, some
of the justification for E0 comes from the conditional probabilities. For in chains like (3), the
conditional probabilities are not just neutral conduits that merely transport the probabilistic
support from one place to another. Rather, they can actively contribute to the probabilistic
justification of the target proposition. This may be seen when we realize that P(E0) in (3) could
still be positive even if P(E2) were zero.

In general it is the case that, as the chain becomes longer, the justification provided by the
conditional probabilities increases, while the justification given by the source of the chain
decreases. And as the n of En grows larger and larger, a law of diminishing returns comes into
force: the influence of P(En) on P(E0) diminishes with each link, until it finally vanishes
entirely. In the limit that n tends to infinity, all the probabilistic support for E0 will come from
the conditional probabilities, and none from the source or foundation.

So Russell and Lewis were right that, in a probabilistic regress, something goes to zero if n goes
to infinity. But this ‘something’ is not as they thought P(E0), the outcome of the probabilistic
regress. Rather it is the justification given to the target proposition E0 by the foundational
proposition En. This is not to say that En itself has to be highly improbable. En may have any
probability value ranging from 0 to 1. It is only to say that, in the limit, the would-be ground En
has become wholly ineffectual: the support it gives to E0 is nil.

A foundationalist may be convinced by our argument that En in the limit is impotent, but still
maintain that the existence of such an En is necessary for the chain to come into being. En would
then become an infinitely far-off ‘prime mover’, as it were, who kicked off the whole process,
but whose influence has now withered away to nothingness. Foundationalists taking this route
would resemble in epistemology what deists are in theology. The latter picture God as a
nonintervening and nonsupervising creator, who permits the universe to run itself according to
the natural laws that he himself engendered. In a similar vein, so it could be argued, our
demonstration of how P(E0) in (2) can have a nonzero value shows that the infinitely remote
origin of the justificatory regress has no influence whatsoever on the exact value of P(E0).

This version of foundationalism is slimmer than the standard variants, but some might deem it
healthier. Unfortunately it has a weak side as well. Our infinite probabilistic chain may have
many concatenated chains, going in different directions, each of them infinitely long. This would
imply the existence of many remote ‘prime movers’, all engaged in a network that has P(E0) as
its output, and none having any effect on the precise value of P(E0). It would then be difficult to
say whether we are still dealing with a version of foundationalism or whether we have already
crossed the bridge to an infinitist variant of coherentism.

1 More details on the debate between Lewis, Russell, and Reichenbach can be found in
Peijnenburg and Atkinson forthcoming.
Foundationalism for Modest Infinitists
The Canadian Journal of Philosophy
John Turri

Abstract: Infinitists argue that their view outshines foundationalism because infinitism can,
whereas foundationalism cannot, explain two of epistemic justification’s crucial features: it
comes in degrees and it can be complete. I present four different ways that foundationalists could
make sense of those two features of justification, thereby undermining the case for infinitism.

1. Introduction

We find two main contemporary arguments for the infinitist theory of epistemic justification
(‘infinitism’ for short): the regress argument (Klein 1999, 2005) and the features argument
(Fantl 2003). I’ve addressed the former elsewhere (Turri 2009a). Here I address the latter.

Jeremy Fantl argues that infinitism outshines foundationalism because infinitism alone can
explain two of epistemic justification’s crucial features, namely, that it comes in degrees and can
be complete. This paper demonstrates foundationalism’s ample resources for explaining both
features.

Section 2 clarifies the debate’s key terms. Section 3 recounts how infinitism explains the two
crucial features. Section 4 presents Fantl’s argument that foundationalism cannot explain the two
crucial features. Section 5 explains how foundationalism can explain the two crucial features.
Section 6 sums up.

2. Terms and Requirements

Infinitism is the view that a proposition Q is epistemically justified for you just in case there is
available to you an infinite series of non-repeating reasons that favors believing Q (Fantl 2003,
539).1 Foundationalism is the view that Q is epistemically justified for you just in case you have
a series of non-repeating reasons that favors believing Q, terminating in a properly basic
foundational reason “that needs no further reason.”2

I cannot here fully characterize epistemic justification, partly because doing so would beg
important questions in the present context, but I may say this much. Epistemic justification is
the positive normative status needed for knowledge, closely associated with having evidence in
favor of the truth of some claim, and typically contrasted with the practical justification, whether
moral or prudential, involved in action.3 I shall refer to it simply as ‘justification’.
Doubtless justification comes in degrees. You can obviously be more or less justified in
accepting some claim. An adequate theory of justification must respect this, and “explain why or
show how” justification comes in degrees. Call this the degree requirement.

Complete justification is “justification for which there is no higher degree” (Fantl 2003, 538),
or otherwise put, “that degree of justification that cannot be increased further” (Fantl 2003,
547). This contrasts with adequate justification, which is the minimal degree of justification
required for knowledge. It is not plausible to identify adequate justification with complete
justification. That justification can be complete is less obvious than that it comes in degrees. For
the sake of argument, I grant that justification can be complete. As such, an adequate theory of
justification must likewise explain why or show how justification can be complete. Call this the
completeness requirement.

3. How Infinitism Proposes to Meet the Requirements

Infinitists satisfy the degree requirement by pointing out that length comes in degrees, which
justification may mirror. “All else being equal, the longer your series of reasons for a
proposition, the more justified it is for you” (Fantl 2003, 554).

Fantl offers the following analysis of complete justification: Q is completely justified for you just
in case you have an infinite array of adequate reasons for Q (Fantl 2003, 558). Having an infinite
array involves infinitely more than merely having an infinite series. To have an infinite array of
reasons favoring Q, for each potential challenge to Q, or to any of the infinite reasons in the
chain supporting Q, or to any of the inferences involved in traversing any link in the chain, you
must have available a further infinite series of reasons. In a word, it requires having an infinite
number of infinite chains.

This analysis of complete justification ensures that no proposition is ever completely justified for
any of us. Fantl does not view this as a problem, because he intuits that although many
propositions are adequately justified for us, none is completely justified.

There is an alternative view, however. It seems that we are justified in being absolutely certain of
some claims. For example, I know for absolute certain that I exist, and that something exists.
Furthermore, it is natural to suppose that we are completely justified when we know for absolute
certain. So at least some claims would seem to be completely justified for us. If correct, this
confounds Fantl’s infinitist analysis of complete justification. However, the following discussion
does not presuppose that it is correct.

4. How Foundationalism Supposedly Fails to Meet the Requirements

Foundationalists divide over how to understand foundational reasons. Traditional

foundationalists contend that foundational reasons are “self-justifying” because their mere truth
suffices to justify them. The claims <I am thinking> and <There is at least one proposition that is
not both true and false> are plausible candidates for self-justifying reasons. Metajustificatory
foundationalists deny that the mere truth of a foundational reason ensures its foundational
status. Instead, they say, foundational reasons must have property F. Different theorists adopt
different values for ‘F’. Some say it is ‘is reliably caused’, others say it is ‘coheres with the
subjects other beliefs’, others say it is ‘is clearly and distinctly perceived’, and yet others might
say it is ‘is approved by society’ or ‘is approved by God’. Other values are possible. Importantly,
metajustificatory foundationalism “cannot require that a believer have access to the
metajustificatory feature as a reason for the foundational reason,” because that would rob the
putative foundational reason of its status as foundational (Fantl 2003, 541). It would effectively
require a further reason for the reason that supposedly stood in no need of it.

Fantl’s division may not capture every important distinction among varieties of foundationalism.
But it does divide all foundationalists into two neat groups, and this suffices for his purposes. We
may represent his basic argument as follows:

1. All foundationalist theories are either traditional or metajustificatory.4 (Premise)

2. Traditional foundationalism cannot satisfy the degree requirement. (Premise)

3. Metajustificatory foundationalism cannot satisfy the completeness requirement.

(Premise)

4. Therefore no foundationalist theory can satisfy both the degree and completeness
requirements. (From 1 – 3)

5. An adequate theory of justification must satisfy both the degree and completeness
requirements. (Premise)

6. Therefore no foundationalist theory of justification is adequate. (From 4 – 5)

The argument is valid, so it remains to ask whether the premises are true. I am granting premises
1 and 5 for the sake of argument. This leaves 2 and 3. The remainder of this section presents
Fantl’s case for each.

We begin with Fantl’s case for 2, i.e. against traditional foundationalism. Traditional
foundationalism has insufficient resources to satisfy the degree requirement. All self-justifying
reasons are by definition true, and their truth justifies them. Yet “truth per se cannot determine
which self-justifying reasons are more or less self-justifying” (Fantl 2003, 544). Appealing to
properties other than the truth of foundational beliefs is inconsistent with traditional
foundationalism; it would effectively transform it into a form of metajustificatory
foundationalism. So truth alone won’t suffice, and no other property is eligible. Traditional
foundationalism cannot satisfy the degree requirement.

We now move on to Fantl’s case for 3, i.e. against metajustificatory foundationalism. To satisfy
the completeness requirement, the metajustificatory foundationalist will have to say something
like this:
 Q is completely justified for you iff you have a non-repeating series of reasons for Q,
ultimately founded on a reason that exemplifies the metajustificatory feature [F] to the
highest possible degree. (Fantl 2003, 546)

But the proposal fails. No matter what value ‘F’ takes, if you gain a reason to think that the
foundational reason completely exemplifies F, and that exemplifying F is epistemically
important, then Q will thereby become better justified for you. Consider, for example, a
reliabilist version of metajustificatory foundationalism, which says that Q is completely justified
for you just in case you have a non-repeating series of reasons for Q, ultimately founded on a
100% reliable reason. If you gain a reason to believe that your foundational reason was 100%
reliably caused, and that reliability is epistemically important, then Q will thereby become better
justified for you.5 But then metajustificatory foundationalism has not satisfied the completeness
requirement, for it will be possible to increase your justification for Q beyond what maximal
exemplification of F would allow.

5. Foundationalist Solutions

I begin by responding to 2. The simplest response is to endorse the view, familiar from fuzzy
logic, that truth comes in degrees (see Priest 2001, chapter 11). On this view, we may represent a
proposition’s degree of truth by assigning it a real number in the interval between 0 and 1,
inclusive (represented by ‘[0, 1]’). The degree of justification could then covary with the
foundational reason’s degree of truth. Complete justification would correspond to a foundational
reason true to degree 1. A traditional foundationalist may, if she likes, treat 1 as an ideal limit
that is never actually reached.

A second response suggests itself. Retain a standard non-degreed theory of truth. A foundational
reason’s truth suffices to render it adequately justified for you; it is in this sense that a
foundational reason’s truth suffices to justify it. However, any degree of justification beyond
adequate requires that you be aware of the foundational reason’s truth. Awareness comes in
degrees. Complete justification requires perfect awareness of the foundational reason’s truth.
Should we desire, we may represent the degree of awareness by assigning it a real number in the
interval [0, 1]. Again, the traditional foundationalist may treat perfect awareness as an ideal limit
that is never actually reached.

I shall now respond to 3. One response is to adopt a two-dimensional model of belief. The first
dimension is credence, or how strongly you believe. Credence comes in degrees, which we may
represent by real numbers in the interval [0, 1]. 1 represents full belief, 0 full disbelief, and .5
perfect suspension of judgment. The second dimension is fixation. At any point in time, you can
be more or less fixated on a certain degree of credence, i.e. it could more or less easily turn out
that your credence shifts from where it is to some other point in the interval. Fixation likewise
comes in degrees, which we may also represent by real numbers in the interval [0, 1], with 0
indicating the maximum level of volatility consistent with the state in question being a belief,
and 1 the maximum level of imperviousness to change consistent with the state in question being
a belief. We can plot these two dimensions on a graph, the y-axis representing credence and the
x-axis representing fixation.
With this framework in place, a metajustificatory foundationalist could say that having a non-
repeating series of reasons for Q, ultimately founded on a reason that fully exemplifies the
metajustificatory property F, justifies full belief in Q. That is, it justifies you in being here
(represented by the star) on your belief-graph for Q:

Credence 1*|
|
|
0 | 1
Fixation

Gaining a reason to believe the proposition <Full belief in Q is justified for me> increases the
degree to which you are justified in being fixated in your full belief in Q. In other words, it
justifies for you a greater degree of fixation, moving you further to the right along the x-axis.

Credence 1 |*
|
|
0 | 1
Fixation

Gaining a further reason to believe the proposition <Full belief in <full belief in Q is justified for
me> is justified for me> justifies an even greater degree of fixation in your full belief in Q.
Further iterations at higher levels are handled similarly. There is no limit in principle to the
number of levels.

We are now positioned to offer the metajustificatory foundationalist proposal:

Q is completely justified for you just in case you are justified in being maximally fixated
at full belief in Q.

The metajustificatory foundationalist may treat maximal fixation as the ideal limit, which can be
approached to varying degrees but never actually reached.

A second response suggests itself. Set aside the two-dimensional model of belief, and adopt
instead a qualitative commonsense framework for measuring strength of belief, with full belief
being the strongest. Q is adequately justified for you when full belief in Q is justified. Adequate
justification could be overdetermined. That is, you could have more than one reason, or chain of
reasons, in virtue of which Q is adequately justified for you. One such reason could be the
proposition <Full belief in Q is justified for me>.6 There is no limit in principle to the degree of
overdetermination. Having recognized all this, the metajustificatory foundationalist could
propose:

Q is completely justified for you just in case it is infinitely overdetermined that full belief
in Q is justified for you.7
6. Conclusion

That brings my discussion to a close. We have seen four responses to Fantl’s argument, two on
behalf of both the traditional foundationalist and the metajustificatory foundationalist. If any of
the four works, then foundationalism can indeed satisfy the twin requirements of degree and
completeness, thereby undermining Fantl’s case for infinitism.

1 Fantl (2003, 539 – 540) indicates that he is concerned with propositional, rather than doxastic,
justification. The latter requires that the belief be properly held on the basis of the good reasons
you possess; the former does not. Stated more fully, the infinitist theory of propositional
justification is: The proposition Q is propositionally justified for S just in case there is available
to S at least one infinite series of propositions (or reasons) such that R1 is a good (and
undefeated) reason to believe Q, R2 is a good (and undefeated) reason to believe R1, R3 is a
good (and undefeated) reason to believe R2, …, Rm + 1 is a good (and undefeated) reason to
believe Rm, for an arbitrarily high m. See Turri 2009b.

2 I doubt that this fully satisfactorily characterizes foundationalism, but I won’t stop to argue the
point here. It isn’t fully satisfactory because foundationalism needn’t require that the chain of
reasons terminate. See Turri 2009a, 162 – 3.

3 Though see Fantl and McGrath 2002 and 2007 for more on the complex relationship between
epistemic and practical matters.

4 The ‘or’ here should be understood exclusively.

5 Fantl (2003, 540 n.7) says that reasons are propositions. Yet it makes no sense to talk about a
proposition being reliably caused. We must, I believe, understand this to mean a reliably caused
token belief state, which takes the foundational proposition as its object.

6 Here I rely on a principle that Fantl (2003: 549) himself relies on when arguing against
metajustificatory foundationalism: “Your epistemic position can be improved in one of two
ways: 1) the degree to which p is justified for you can increase and 2) the degree to which it is
justified for you that p is justified for you can increase.”

7 Note that accepting this doesn’t require the metajustificatory foundationalist to reject the
earlier claim, “if you gain a reason to think that the foundational reason completely exemplifies
F, and that exemplifying F is epistemically important, then Q will thereby become better justified
for you.” For we can understand ‘better justified’ to include not only an increase in the degree of
justification, but an increase in the security or stability of the degree of justification, which
overdetermination promotes. This remains true whether justification is determined by reliability
or any other plausible candidate the metajustificatory foundationalist suggests.
Synthese (2011) 178:515–527
Infinitism and epistemic normativity
Adam C. Podlaskowski · Joshua A. Smith

Abstract Klein’s account of epistemic justification, infinitism, supplies a novel solution to the
regress problem. We argue that concentrating on the normative aspect of justification exposes a
number of unpalatable consequences for infinitism, all of which warrant rejecting the position.
As an intermediary step, we develop a stronger version of the ‘finite minds’ objection.
Keywords Infinitism · Epistemic responsibility · Normativity · Regress problem ·
Klein

Epistemically responsible agents face the difficult task of accepting only justified beliefs. As
Peter Klein points out, taking seriously our epistemic responsibilities quickly leads to an infinite
regress, since any belief is justified only if it is based on good reasons, and the beliefs serving as
reasons also stand in need of justification, as do those beliefs serving as reasons, and so on
(2007a, p. 5). This difficulty, the regress problem, threatens the possibility of an agent meeting
any of her epistemic responsibilities.

Klein’s solution to the problem, infinitism, is that one is provisionally justified when there is an
infinite, non-repeating chain of reasons for one’s belief. His solution is not to stop the regress in
its tracks (as foundationalists insist), nor it is to suggest that each belief is justified by its place
within a network of beliefs (as coherentists insist). Rather, the infinitist embraces the infinite
regress by suggesting that it does not pose a genuine problem.

While there is much to be admired in Klein’s view, we are concerned that infinitism cannot
adequately account for one’s epistemic responsibilities. More specifically, we shall argue that
infinitism cannot account for the normative feature of epistemic responsibility. Our case against
infinitism proceeds from the assumption that epistemically responsible agents ought only
maintain justified beliefs. (Since Klein himself makes this assumption, there is no need for
providing a full-fledged theory of epistemic responsibility here. In order to be as charitable as
possible to Klein, we shall give a weak reading of this idea.) Along with this assumption, we
appeal to a weakened version of the commonly held view that one ought to act (or think) thus-
and-so implies that one can, at least sometimes, do so. We should expect, then, a similar
relationship between one’s epistemic commitments (the ‘ought’) and capabilities (the ‘can’). But
there is good reason to think, as we shall argue, that on the infinitist view, one cannot ever meet
one’s responsibilities. This denial of the epistemic ‘can’ comes in the form of a strengthened
version of the finite minds objection.
The typical formulation of the objection is that a person cannot have infinitely many reasons at
one time, given that humans have finite minds.1 We suggest that, though Klein has adequately
addressed this worry as originally stated, a stronger version of the objection lurks in the
background (what shall be called the finite and less-than-ideally-ordered minds objection).2 By
developing this objection, we argue that Klein’s view commits him to one of a number of
unpalatable consequences, depending on how beliefs are based on reasons. However Klein
proposes to cash out the epistemic ‘can’, his account suffers from debilitating problems. Since
Klein cannot furnish this ‘can’, our epistemic responsibilities fall to the wayside—an
unacceptable result. Therefore, infinitism should be rejected.

1 Infinitism and epistemic responsibility

It is widely held that justification is a normative notion.3 It is also commonly maintained that the
normative feature of justification serves as a constraint on any acceptable theory of the notion
(e.g., Chisholm 1977; Sellars 1956). Though there is less agreement about the nature of this
normativity, one important sense stands out: namely, that an epistemically responsible agent
ought to form only justified beliefs (e.g., Bonjour 1980;Kornblith 1983). The role of
responsibility suggests the presence of action-guiding norms, such as obligations and
permissions. Admittedly, though, there are a number of different respects in which justification
might be normative without implying an ‘ought’. For instance, Pollock (1985) outlines one
respect in which justification is action-guiding by focusing on epistemic permissions.
Nevertheless, since the ordinary sense accorded to ‘responsibility’ (on which the epistemic
notion relies) is tied to respect for how one ought to act, it is quite plausible to assume that
justification implies an ‘ought’ insofar as it is a kind of epistemic responsibility.

For any view on which only justified beliefs can provide justificatory support (for other beliefs),
an epistemically responsible agent faces the regress problem in epistemology. One’s belief that
p, to avoid being arbitrarily held, must be justified by some reason r1. But in order for r1 to be
able to justify p, it must be justified by some reason r2, and so on. It appears that we face an
infinite regress, and unless there is some acceptable way to address this regress, it appears that no
belief within that chain of reasons is left justified.4 In short, taking seriously our epistemic
responsibilities requires justifying our beliefs, but doing so requires an infinite regress. As such,
it is unclear how one could meet one’s epistemic responsibilities. As Klein suggests:

A key notion here is, of course, ‘epistemic responsibility.’ It is an unabashed normative

notion. And that is as it should be since the regress problem is about what kind of
reasoning can satisfy the norms of epistemic responsibility. (2007a, p. 5)

The debate about epistemic justification and the regress problem has, until very recently, been a
debate between foundationalists, who believe that there are noninferentially justified basic
beliefs whose justification transmits to other beliefs, and coherentists, who maintain that no
beliefs are justified non-inferentially, and that a belief is justified when it is a member of a
coherent system. Klein’s own alternative, infinitism, is that one is provisionally justified when
there is an infinite, non-repeating chain of reasons for one’s belief. Instead of stopping the
regress in its tracks (as foundationalists insist), or distributing the regress phenomena over a
network of beliefs (as coherentists insist), the infinitist embraces the infinite regress by
suggesting that it does not pose a genuine problem and insisting that, properly understood,
justification actually requires infinite justificatory regresses.

The plausibility of Klein’s solution to the regress problem depends, in part, on a shift away from
the typical conception of reasons and how they relate to justification. Instead of holding that one
can, if one accumulates sufficient reasons, be fully justified, Klein maintains that the best one
can hope to do is to be provisionally justified. The degree of provisional justification increases
with each reason back in the infinite, non-repeating chain one actually obtains. As such, one
need not have in hand an infinite number of reasons (Klein 2007a, p. 10). Rather, Klein requires
only that those reasons be available to the epistemic agent.

Klein has employed two different (though related) ways of articulating the notion of availability.
In his (1999), Klein says that a reason is available when one has either a disposition to believe
the reason, or has a second-order disposition to form a disposition to believe the reason. And in
his (2007a), Klein says that a reason, p, is available to an epistemic agent, S, just in case:

[T]here is an epistemically credible way of S’s coming to believe that p given S’s current
epistemic practices. Available propositions to S are like money in S’s bank account that is
available to S if S has some legal way of withdrawing it even if S is unaware that the
money is there or takes no steps to withdraw it. (2007a, p. 13)

This way of understanding a reason’s availability seems to be a way of better characterizing the
dispositional view. For there being a way of coming to believe something seems just to mean
that one has a disposition to believe something in certain epistemically credible circumstances or
a (second-order) disposition to form that belief. If one did not have a (first or second-order)
disposition to believe something in any circumstances, it is difficult to see how there could be a
way for one to believe it. As such, we shall understand Klein’s view about the availability of
reasons in dispositional terms.

One benefit of Klein’s conception of the availability of reasons is that it does not require a
particular view on what makes one proposition a reason for another. But Klein does seem to take
a stand on an important issue related to one’s having reasons: what it is for one’s belief to be
based on a reason. The epistemic basing relation is the connection between one’s beliefs and
one’s reasons. Put roughly, the reasons on which a belief is based are the reasons for which one
holds the belief. In his (2005b), Klein says that “infinitism holds that a particular belief is
doxastically justified (at least to some degree) only if there is an available reason and we cite that
reason as a reason for our belief” (p. 26). In response to a challenge from Bergmann (2007),
however, Klein is willing to allow that a belief be based on a reason when the reason causes (in
an appropriate way) the belief.

In the sections that follow, we shall raise objections to infinitism which turn, to some extent, on
the relationship between Klein’s view of available reasons and the basing relation (especially in
Sect. 3). So it is worth spending some time getting clear about what is at stake over the basing
relation. Perhaps the most well developed account of the basing relation is due to Korcz (2000).
According to Korcz, one’s belief is based on some reason just in case either the belief was
caused by the reason, or one has a meta-belief to the effect that the reason is a good reason for
holding the belief.5 If we understand Klein’s requirement that an agent cite his reason (as a
reason) in terms of the agent’s having a meta-belief that the reason is a good one (as suggested in
his (2007a)), then Korcz’s account of the basing relation seems to capture what Klein is willing
to allow for basing (given his concession to Bergmann).6 So for the purposes of this discussion,
the infinitist will be taken to understand basing as Korcz articulates it.

One might worry that we are unfairly saddling Klein with a particular view of the basing
relation. However, this worry would be misplaced for two reasons. The first is that we have
motivated Klein’s tacit acceptance of Korcz’s account. Klein seems to think that meta-beliefs can
establish basing, and he is willing to allow (in response to Bergmann) that a causal connection
can establish basing. This just seems to be Korcz’s view. The second reason the worry is
misplaced is that Korcz’s view is the most permissive account of basing in the literature. Other
accounts of basing allow one or the other of Korcz’s disjuncts, but not both. At the very least, if
infinitism requires some other account of basing, it is incumbent on the infinitist to say what that
account is.

One final clarification is in order. With the notion of the basing relation in place, we may now
consider the distinction between propositional and doxastic justification. A proposition may be
justified for a person even if the person does not believe the proposition. This type of
justification concerns how one’s evidence relates to some proposition (whether the person
believes that proposition or not). Doxastic justification concerns whether a belief the person
actually holds is justified by the reasons on which it is based. So one way to think about the
relationship between these two types of justification is that doxastic justification results from
adding the basing relation to propositional justification. This distinction is important for
understanding Klein’s view. He suggests that:

Infinitism is committed to an account of propositional justification such that a

proposition, p, is justified for S iff there is an endless series of non-repeating propositions
available to S such that beginning with p, each succeeding member is a reason for the
immediately preceding one. It is committed to an account of doxastic justification such
that a belief is doxastically justified for S iff S has engaged in tracing the reason in virtue
of which the proposition p is justified far forward enough to satisfy the contextually
determined requirements. (2007a, p. 11, emphasis in original).

In the sections that follow, we take issue with Klein’s account: our principal worry is that, even
with the impressive apparatus available to him, infinitism is inadequate to accommodate the
normative feature of justification.

2 Problematizing infinitism

We shall argue that a tension arises for infinitism, between what one ought to do and what one
can do, so far as justification is concerned. The basic shape of our argument follows. Since
justification is a normative notion, it implies an ‘ought’. And that one ought to act (or think)
thus-and-so implies that one can do so.7 Unless we can make sense of how one can, as a matter
of principle, engage in any given instance of this process, we lose sight of our epistemic
responsibilities. In Sect. 2.1, we argue that a strengthened version of an objection already raised
against infinitism (the finite minds objection), aimed at propositional justification and
availability, eliminates the relevant ‘can’.8 And in Sect. 2.2, we argue that the success of this
objection (from Sect. 2.1) renders infinitism incapable of accommodating the normative feature
of justification qua epistemic responsibility. In short, we shall use a strengthened version of the
finite minds objection to serve as the basis for our deeper normativity objection. Finally, in Sect.
3, we address the infinitist’s suggestion that, even if propositional justification is problematic for
capturing the epistemic ‘can’, doxastic justification might serve the infinitist’s needs. In that
section, we argue that, even if Klein’s emphasis on doxastic justification avoids the finite minds
objection, it proves too restrictive as an account of justification. Thus, Klein’s infinitism faces
problems at every turn.

2.1 The finite and less-than-ideally-ordered minds objection

At first blush, it appears that anyone attempting to justify the belief that p must cite the infinite
sequence of reasons issuing from p. Yet human minds are finite so that one could not possibly
hold an infinite number of beliefs; and having a finite life span precludes citing each reason in
the sequence. As such, one could not ever be fully justified in holding a belief. This objection,
typically referred to as the finite minds objection, has been discussed by Klein a number of times
(e.g., Klein 1999, 2005a, 2007a).

How we have articulated Klein’s view suggests his response. First, Klein agrees that on his view
one cannot be fully justified, and he embraces this result. All that we can hope for, according to
Klein, is provisional doxastic justification sufficient to meet the demands set by a particular
conversational context. For any such case, we ought to cite a series of reasons until we reach a
reasonable place for ending an inquiry supplied by the epistemic context (akin to reaching what
Wittgenstein (1969) calls ‘hinge propositions’).

Secondly, Klein maintains that the notion of availability (which is built into the notion of
propositional justification) allows that, at every step in the chain of reasons there is a reason to be
cited and, in some sense, one is disposed to do so. Though one might not presently possess the
(first-order) disposition to believe that which may be cited as a reason, one nevertheless
possesses the second-order disposition to form that belief. This view of available reasons, with
its emphasis on epistemically credible first and second-order dispositions, aids in avoiding the
finite minds objection by making clear that the objection is based on the (mistaken) presumption
that justification requires possession of an infinite number of beliefs. Moreover, Klein’s position
keeps its distinctively infinitist flavor by maintaining that there is an infinite non-repeating chain
of reasons, any given member of which is, in some respect, available to an agent. In short, Klein
insists that, “We don’t have to traverse infinitely many steps on the endless path. There just must
be such a path and we have to traverse as many as contextually required” (2007a, p. 13).

Though we are sympathetic to Klein’s suggestion that justification is an ongoing process and that
there is, “always a further step that can be taken should we become dissatisfied with the point at
which we stopped the progress of inquiry,” (Klein 2007a) there is reason to doubt that the
position, as Klein puts it, can be maintained easily. For as we shall argue, though Klein’s
response is adequate against the traditional form of the finite minds objection, it is not similarly
effective against the version that follows.
The problem we point out is not that a finite agent cannot hold an infinite number of beliefs, nor
is it that a finite agent cannot hold a particular belief about the order of an infinite number of
beliefs. Rather, our concern is that, for many cases, a finite agent cannot hold a belief that stands
in the proper place within a sequence of reasons, so as to serve as an available reason in the
sense that Klein requires. That is, Klein’s infinitism requires that every step in the infinite
sequence of reasons must be available to an agent. Our strategy is to undermine Klein’s reliance
on second-order dispositions (to form beliefs as reasons) by arguing that there is a new finitude
objection that applies to second-order dispositions as well. (This version of the objection shall be
called the finite and less-than-ideally-ordered minds objection.) This objection is meant to show
that not every reason in a particular infinite series is available to a given subject.

For the sake of argument, we grant Klein that, for a belief that p, that there might exist the right
sorts of justificatory links between p and an infinite, non-repeating chain of reasons r1 . . . rn,
even if that sequence is not entirely easily accessible to S. To use Klein’s (2007a) bank account
metaphor, there might exist a bank account containing an infinite amount of money. However,
this does not guarantee the possibility that we can supply the right sort of legal means for
withdrawing funds from it, at least not to the extent required in order to satisfy S’s needs. After
all, our interest in any given belief here is to act as a reason for some other belief. Though S
might have the second-order disposition to form the belief that p which serves as a reason ri,
citing ri as a reason for another belief depends on being already epistemically situated in the
right place within the sequence of r1 . . . rn. This requires sensitivity to the order within the
sequence of beliefs that the cited reason resides, not just that one has a given belief. If each
transaction were a token of the same type, we might make sense of the same (second-order)
disposition to form beliefs being manifested in a variety of justificatory contexts. But
transactions from this account do not share such a resemblance. Unlike many financial
transactions, each epistemic transaction (i.e., citing a reason) depends on making another one,
and that transaction depends on making another one, and so on. One cannot make an arbitrary
transaction, breaking into the sequence without having made the transactions leading up to the
entry point, and expect to benefit from the justificatory work (i.e., epistemic transactions) that led
to that point. Instead, each transaction depends on its place within the sequence.

That any given epistemic transaction depends on its place within a chain of reasons throws doubt
on S possessing second-order dispositions to form beliefs to cite as reasons for every given place
in that chain. For many such epistemic transactions ultimately outstrip S’s finite capabilities. It is
not just that S does not have the opportunity to manifest second-order dispositions past some
distant place in the chain of reasons. Most links in an infinite chain of reasons could not be cited
within the course of S’s lifetime. Hypotheses about the immortality of the soul aside, S does not
possess any dispositions to act beyond her lifetime.9 And there are plenty of cases, at the borders
of one’s finite capabilities, where one possesses the wrong dispositions. Faced with an infinite
chain of reasons to cite, it is more likely that, at some point along the chain, S has the disposition
offer a guess or become bored with the whole enterprise (instead of having the epistemically
credible disposition to continue citing reasons). There is good reason to think, then, that for a
great many cases, S does not possess the relevant second-order dispositions whatsoever.10 So the
finite minds problem originally aimed at S and her first-order dispositions to hold any given
belief can also be raised for the relevant second-order dispositions.
2.2 The normativity objection

The finite and less-than-ideally-ordered minds objection serves as the basis for our deeper worry:
that infinitism cannot sustain the normative feature of justification. To establish this conclusion,
we take seriously Klein’s (2007a) insistence that justification, as a kind of epistemic
responsibility, is a normative notion. As such, agents ought to form beliefs by epistemically
respectable means. This requires that, at least sometimes, one is able to act in a certain way. If
one cannot ever act in a way so as to meet one’s epistemic responsibilities, it is very difficult to
see in what way the notion associated with one’s behavior is normative. So we maintain that if
one ought to form beliefs in a particular way, then it must in principle be possible for such a
being to form beliefs in that way. (This weakened version of the slogan ‘ought implies can’
should suffice to secure our conclusion while avoiding most of the objections to stronger
versions of the principle.)

Given this understanding of normativity, when one ought to justify the belief that p, it has to be
the case that, at least sometimes, one can do so. And as already argued, for most of the members
of any infinite chain of reasons, one’s finite nature precludes citing most of those reasons, or
even possessing the second-order disposition to do so. Therefore, one is not responsible for
justifying any given belief. Without the right capabilities, epistemic responsibilities fall to the
wayside. Clearly, this is an unacceptable result. While Ginet (rightly or wrongly) argues that, if
infinitism is true, then no beliefs are justified (Ginet 2005, p. 148), we argue that, if the thesis is
true, then we are always epistemically blameless.

In response to this objection, one might maintain that what has been illustrated so far is that the
way Klein thinks of normativity is the cuprit, not infinitism. Suppose that Klein acknowledged
that his view had the consequence that we are always epistemically blameless. In order to avoid
this consequence, suppose he endorsed some other understanding of epistemic normativity—one
that does not involve citing reasons. Infinitism, per se, would then be immune to the normativity
objection developed above.

While it would be immune to the normativity objection, it would still fall to the finite and less-
than-ideally-ordered minds objection. In order to make sense of how an infinite string of reasons
could justify a proposition for us, Klein needs for it to be possible for us to be able to follow such
a string. Regardless of how we understand epistemic normativity, we have been urging that this
is impossible given our cognitive resources. Klein has failed to appreciate that it is not only the
finitude of our minds that is at issue. It is also how well-ordered they are. And while this led to
the normativity objection, it is an independent issue for the infinitist.

3 The doxastic fix?

So much for the notion of an available reason in Klein’s account, and the difficulty it poses for
capturing the normative feature of justification. But one might still wonder whether Klein’s
emphasis on doxastic justification aids in avoiding our concerns. Consider, for instance, Klein’s
response to Bergmann on the (original) finite minds objection:
 The crucial point was that although propositional justification requires that there be an
infinite path of non-repeating reasons, in order for a belief to be (at least partially)
doxastically justified, it is not required that S possess that infinite set of reasons or that a
belief be based upon beliefs that have the infinite set of reasons as their propositional
contents. However, it is required that some of those reasons be available and that the
belief be based upon those beliefs that have the available reasons as their contents. In
other words, by distinguishing between propositional and doxastic justification the so-
called ‘finite mind problem’ would disappear. (2007b, pp. 26–27)

If this response is adequate for the original finite minds objection, is it adequate for the finite and
less-than-ideally-ordered minds objection? Recall that our objection has it that even if we could
have (second-order) dispositions to form beliefs that serve as reasons, the infinitist needs those
dispositions to be properly ordered, which is implausible for an infinite chain of reasons. But
surely, the infinitist might respond: it is not at all implausible for one to have properly ordered
dispositions to meet the finite and often meager demands of doxastic justification.

In what follows, we shall argue that there are two problems with Klein’s supposed doxastic fix to
the finite minds problem, both stemming from how the infinitist might understand the epistemic
basing relation. Recall that, in Sect. 1, we suggested that Klein seems to endorse something like
Korcz’ account of the basing relation. According to that view, a belief can be based on a reason
if one has a meta-belief that the reason is a good reason for holding the belief, or if there is an
appropriate causal connection between the reason and the belief. As we shall argue, on either
account of the basing relation, reasons cannot play their proper role. Klein’s account, when
understanding epistemic basing as holding the right meta-beliefs, rules out unsophisticated
epistemic agents’ justified beliefs; and Klein’s account, when taking basing to be a causal matter,
fails to allow causally based reasons to do any justificatory work. In short, Klein’s account of
justification is too restrictive. (Note, though, that these problems do not arise for the infinitist
qua infinitist. Rather, they arise from Klein’s requirement on doxastic justification that one cite
one’s reasons.)

To see the problem causal basing raises for Klein, consider the following case.11 Suppose that
Nick and David have been in their offices all day, and David goes to Nick’s office to ask whether
Nick would like to get a beer after a long day at work. Suppose that Nick’s answer depends on
the weather. Also suppose that earlier, David looked at the weather report online, which said that
there would be no rain that evening. David believes that it is not going to rain because he saw the
weather forecast online. Suppose that, instead of going to talk to Nick, David went outside to
check the weather for himself. We may imagine that David is generally reliable about whether
there is going to be rain, even though he does not exactly appreciate why he is reliable. Now,
suppose that both his seeing what the weather is like and the online forecast causally sustain his
belief that there is not going to be rain. When he goes in to ask Nick about going out, suppose
Nick asks whether it is going to rain. David replies, “No, the online weather report called for
clear skies.” Suppose Nick then asks, “Which weather report did you use? Is it reliable?” To
which David responds, “Yes, I used the one from the local news, which tends to be very
reliable.” Nick is thereby satisfied, and they go have their drink.
Recall that Klein says, “infinitism holds that a particular belief is doxastically justified (at least to
some degree) only if there is an available reason and we cite that reason as a reason for our
belief” (2007b, p. 26). David’s belief that there will not be rain seems to be justified on the basis
of two distinct bodies of reason: his checking the weather online and his experiencing the
weather first-hand (and we may assume that the evidence David has is, in fact, part of an infinite,
non-repeating chain of reasons that justify the proposition David believes). But recall that David
does not understand how his experiencing the weather constitutes a reason for thinking it will not
rain. We may suppose that he has a meta-belief to the extent that his experiencing the weather is
why he believes that it will not rain, but that he lacks a meta-belief about whether this is a good
reason for thinking it will rain.12 He understands, roughly, that he has oftentimes been right in
the past, and that there is a good inductive inference for his being right this time, but he is simply
too lazy to work through all of that. So suppose he avoids tracing that line of reasons. Then,
according to Klein’s infinitism, David is not justified on the basis of experiencing the weather.
There is a reason available to David, but he would not cite it. Thus, the infinitist must say he
gains no provisional justification from those reasons. But this seems incorrect. Surely those
reasons play a role in how well justified David’s belief is.

More generally, any time S has more than one set of reasons (R1 and R2, say), any of which
would justify S’s belief that p, Klein is committed to maintaining that only those S would cite can
justify S’s belief. So if S would not cite the members of R2 because one is too lazy, or due to any
other psychological issue, then R2 cannot play a justificatory role for S. This seems to be the
wrong result, since features of S’s situation unrelated to S’s justificatory status ought not
preclude R2 from playing that role for S. Thus, adding a causal basing requirement renders
Klein’s account too restrictive in that it does not allow all of one’s reasons to play a justificatory
role.

One might respond by suggesting that this example fails to appreciate the contextualist aspect of
Klein’s account. In a context in which the standards are low, David’s other reasons will play no
role in determining whether he is justified, since his evidence from the weather report was
sufficient for his belief to be justified. So the fact that his having experienced the weather plays
no role should not be cause for alarm.

The problem with this response is that it misses the point of the example. Even in a context with
low standards, one can have reasons beyond those that satisfy the standards. Klein’s account of
doxastic justification fails to allow those reasons that causally sustain one’s belief but one would
not cite as reasons to play any role. Thus, his account is too restrictive when a causal connection
is allowed to establish the basing relation.

Perhaps, then, Klein should simply require meta-beliefs for basing. Unfortunately, this is
restrictive in another way. Meta-beliefs that establish the basing relation are beliefs that a reason
is a good reason for holding a particular belief. But there are epistemic agents who have justified
beliefs despite lacking the concepts required for such meta-beliefs. For instance, many adults
have justified beliefs, yet lack an understanding of what a “good reason” is.13 Giving up on
unsophisticated believers having justified beliefs is too high a price to pay to save infinitism.
Thus, if only meta-beliefs could establish basing, Klein’s account would be too restrictive.14
Klein suggests that his distinction between propositional and doxastic justification helps to avoid
the finite minds objection. But once we appreciate the need for the basing relation in his account,
we realize that his view is overly restrictive. Either it fails to allow causally based reasons to do
any justificatory work, or it rules out unsophisticated epistemic agents’ justified beliefs. And
either way, Klein’s attempt to avoid the finite and less-than-ideally-ordered minds problem fails.

4 Conclusion

Peter Klein’s infinitism, we have argued, suffers from a number of defects. Understanding his
account of propositional justification leads to the finite and less-than ideally-ordered minds
objection, which motivates the normativity objection. We noted that Klein might adopt a
different way of thinking about epistemic normativity (in terms other than epistemic
responsibility) to avoid the normativity objection, but he would still have to face our
strengthened finite minds objection. Moreover, his attempt to avoid that objection by
distinguishing between propositional and doxastic justification as he does leads to further
problems, given his concession to Bergmann on the basing relation. We conclude that infinitism
should be rejected.

Acknowledgements Thanks to George Pappas, Bob Stecker, Gary Fuller, Matt Katz, Rob
Noggle, Mark Shelton, and Nicholaos Jones for helpful comments on an earlier version of this
article. Thanks to Bill Melanson and Aaron Cotnoir for helpful discussions. Thanks also to two
anonymous reviewers for their helpful suggestions.

1 Infinitism has already been met with a number of other objections (Bergmann 2007; Cling
2004; Ginet 2005; Turri 2007), which Klein has admirably addressed. For present purposes, we
do not address any of them.

2 See also Aikin’s (2005) treatment of versions of the finite minds objection.

3 For a dissenting view, see Fumerton’s (2001).

4 Separate versions of the regress problem may be raised for being justified and being able to
identify justifications (e.g., Simson 1986, pp. 180–181). It is not entirely clear how Klein’s own
conception of the issue relates to this distinction, though it oftentimes appears that he focuses on
the former version of the problem. Nevertheless, the criticisms we raise of Klein’s position do
not depend on settling this exegetical matter.

5 This is a rough characterization of Korcz’s admirably intricate account. The details of Korcz’s
account are not relevant to this discussion.

6 Well worth noting is that a number of people maintain that an adequate account of the basing
relation is to be understood only in causal terms. Bergmann seems to suggest this, Mittag (2002)
suggests this, and Swain (1981) defends a purely causal account of basing.

7 The principle that ought implies can is not uncontroversial. See, for instance, Ryan’s (2003).
The version of the principle we employ, however, is quite weak. It is weaker than any of the
versions for which Ryan provides counter-examples. Moreover, only the principal objection
raised depends on the assumption. The other objections raised in Sects. 2.1 and 3 stand even if
one rejects our weakened principle that ought implies can.

8 Ginet offers another means of eliminating this epistemic ‘can’, though he does so quite
separately of normative considerations. He entertains this worry, directed at propositional
justification, that we may only make sense of an infinite chain of available reasons on the
condition that we have in hand a generalized manner for establishing the justificatory links
between those reasons, such as an algorithm. Moreover, it is unlikely that such a thing can be
given (Ginet 2005, p. 147).

9 Note that these claims about one’s finite dispositions are partially inspired by a similar
suggestion made by Kripke (1982) while arguing against dispositional accounts of meaning. In
his famous interpretation of Wittgenstein, Kripke suggests that agents like us do not possess
many of the dispositions needed in order to grounds the rules one follows. But whereas our
objection to infinitism appeals to dispositions to form beliefs, Kripke focuses on dispositions of a
non-intentional variety. From this, he argues that a standard dispositional account cannot ground
the indefinite applications of rules (such as the one governing the use of ‘+’) because the rules
one follows (on a dispositional account) are to be read off one’s dispositions and, for cases
outstripping one’s actual capabilities, one has no dispositions whatsoever to act. Therefore, there
is no fact of the matter which rules one follows in extreme cases.

10 In response to the claim that S does not possess whatsoever the dispositions required by
Klein’s account, one might be tempted to appeal to the dispositions possessed of idealized
counterparts to ourselves. Were S under sufficiently ideal conditions (e.g., having infinite time,
processing power, memory), then S would be disposed in the manner that Klein’s account
requires. Despite the initial plausibility of this suggestion, though, this is an unwarranted appeal
to idealizations. For the regress problem is peculiar to finite agents like us. As such, suggesting
an account of justification for our idealized counterparts does not serve as a solution to the
regress problem for us, since the problem is distinctly raised for agents with our limited
epistemic capabilities.

11 Note that the way in which the beliefs are causally based on reasons is not going to be that the
reasons play a causal role in producing the beliefs. Very often, on Klein’s view, one must have
the belief before one possesses the reasons. So it seems that the reasons must causally sustain the
belief, and this is how the belief is causally based on them.

12 Notice, this is not to say that David has a meta-belief that experiencing the weather is not a
good reason for thinking it will rain.

13 This worry is raised by Korcz (2000, p. 536). See also Schmitt (2001, pp. 184–185).

14 Klein says that the kind of knowledge of interest to him is “knowledge that results from
carefully examining our beliefs in order to determine which, if any, deserve to be maintained”
(2007a, p. 4). Even if we adopt this view of knowledge/justification, Klein’s account is too
restrictive. Imagine a case in which one clearly has a justified belief, in virtue of having all of the
right evidence, even though one has not carefully worked through all of it. For instance, imagine
that one sees roughly how their evidence will provide an argument, without seeing each step in
the argument yet. Such a person seems to be justified in Klein’s sense, yet lacks the requisite
meta-beliefs. Cases like this show that Klein’s account of doxastic justification is too restrictive
if it requires meta-beliefs for basing.

Epistemic Principles and Epistemic Circularity

Byeong D. Lee
Philosophia (2014) 42:413–432

Abstract Can we show that our senses are reliable sources of information about the world? To
show this, we need to establish that most of our perceptual judgments have been true. But we
cannot determine these inductive instances without relying upon sense perception. Thus, it
seems, we cannot establish the reliability of sense perception by means of an argument without
falling into epistemic circularity. In this paper, I argue that this consequence is not an
epistemological disaster. For this purpose, I defend a normative claim that it is reasonable to
accept the general reliability of our perceptual judgments, instead of a factual claim that our
perceptual judgments are generally reliable. More specifically, I offer a normative practical
argument which explains why it is reasonable to accept the general reliability of our perceptual
judgments, even though we cannot establish the general reliability of our perceptual judgments
by means of theoretical reasoning.

Keywords Epistemic circularity . Epistemic first principle . Instrumental reasoning . Epistemic

presumption . Wilfrid Sellars . Crispin Wright

Preliminary Remarks

Our ultimate sources of information about the world are perception, memory and introspection.
Thus, it seems, we cannot have justified beliefs about the world unless these are reliable sources
of information about the world. Let us consider the following epistemic principles:

EP1: Our perceptual judgments are generally reliable.

EP2: Our introspective judgments are generally reliable.

EP3: Our memory judgments are generally reliable.

Let us focus on EP1. If it is reasonable for us to accept this epistemic principle, and if we
ostensibly perceive (without ground for doubt) something to be Φ, then we may say that we are
justified in believing that the perceived object is Φ. Thus, if it is reasonable for us to accept such
an epistemic principle, we can form justified beliefs about the world by relying on sense
perception. But the big question is: can we justify such an epistemic principle as EP1?1 Note that
EP1 is a fallible claim with substantial empirical content. In order to establish the correctness of
EP1, we need to establish that most of our perceptual judgments have been true. But, without
relying on sense perception, we cannot determine that most of our perceptual judgments have
been true. To put it another way, we need to apply the very principle, EP1, to our previous
perceptual judgments in order to use these judgments as inductive evidence. Thus, it seems, we
cannot inductively justify EP1 without presupposing the general reliability of our perceptual
abilities. An argument is called epistemically circular if it supports the reliability of a source of
belief by relying on a premise that is itself based on that source. Therefore, we face the problem
of epistemic circularity in justifying epistemic principles such as EP1. Due to this sort of reason,
epistemic foundationalists such as C. I. Lewis and Roderick Firth claim that such a fundamental
epistemic principle cannot be justified by virtue of other beliefs (cf. Lewis 1946, 1952; and Firth
1964).

If epistemic principles such as EP1 cannot be justified by other beliefs, why is it reasonable for
us to accept such principles? One possible proposal is that they are first principles such that they
must be accepted without further justification. Quite a few past philosophers took such first
principles to be self-evident.2 Another possible proposal is that we have no option but to accept
such fundamental principles unless we are willing to deny the possibility of empirical
knowledge. Wilfrid Sellars is critical about these proposals, however. On his view, the first
proposal, which usually appeals to self-evidence, is too atomistic (cf. Sellars 1979, p. 176). As
noted, EP1 is a fallible claim with substantial empirical content, and so it is implausible to take
such a fallible claim to be self-evident without any adequate ground. Besides, it is desirable to
justify such a substantial claim in a naturalistic setting (cf. Sellars 1975, p. 345). The second
proposal amounts to the old slogan ‘This or nothing’. Sellars claims that this old slogan is too
weak, because it does not give us any illuminating insight as to why such principles as EP1 are
correct (cf. Sellars 1979, p. 176).

The above problem of epistemic circularity is one of fundamental problems in epistemology.

Most notably, William Alston has addressed this sort of circularity problem in a number of
studies (cf. Alston 1986, 1993, 1996, and 2005). His view is that we cannot establish the general
reliability of sense perception without falling into epistemic circularity. In a similar vein, Ernest
Sosa argues that a fully general theory of knowledge is impossible without epistemic circularity
(cf. Sosa 2009, pp. 175–176).

Why then do we consider epistemic circularity as a bad thing? One important reason is that a
circular argument is dialectically ineffective. For us to convince someone who doubts a claim,
we must offer reasons that she can take to be less questionable than the claim in question. Thus a
circular argument cannot rationally convince someone who doubts the conclusion because the
premises are as questionable as the conclusion.

As noted above, we cannot inductively justify EP1 without presupposing that sense perception is
a reliable source of information. In addition, EP1 is an empirical claim and so, it seems, not
something that can be justified a priori. If we cannot justify the general reliability of sense
perception without epistemic circularity due to the above reasons, we cannot defend our
epistemological endeavor against skeptical challenges. The purpose of this paper is to suggest a
new way of dealing with this recalcitrant problem. In this paper, I shall begin by admitting that
we cannot establish the general reliability of sense perception by means of theoretical reasoning,
without falling into epistemic circularity. Nonetheless, I shall argue that this consequence is not
an epistemological disaster. I shall make this point by defending something virtually equivalent
to EP1, namely EP1*:

EP1*: It is reasonable for us to accept the general reliability of our perceptual judgments.

If we can defend the above normative claim, then we can defend pursuing the epistemic goal by
relying on our perceptual judgments. As a consequence, even though we cannot establish EP1 by
theoretical reasoning, its analogue EP1* would suffice to get us out of a philosophically
embarrassing situation where we cannot rationally defend our ordinary perceptual beliefs against
skeptical challenges. For the above purpose, I shall argue that it is a minimum presumption for
our epistemic discourse that the epistemic goal is reasonable. I shall also argue that accepting the
general reliability of our perceptual judgments is ultimately the best means of bringing about the
epistemic goal. And then I shall offer a normative practical argument about a means-goal relation
which has EP1* as its conclusion. Thus, it is not the aim of this paper to offer a theoretical
argument for an epistemic principle such that without relying on premises that are themselves
based on sense perception it shows that the beliefs based on sense perception are likely to be
true. Instead, I shall offer a normative practical argument which explains why it is reasonable for
us to accept the general reliability of our perceptual judgments, even though we cannot establish
the general reliability of our perceptual judgments by means of theoretical reasoning.

Instrumental Reasoning and Reasonable Goals

As mentioned in the previous section, I shall defend the claim that it is reasonable to accept the
general reliability of our perceptual judgments. And I shall use, in part, instrumental reasoning in
order to defend this claim. Thus, let me briefly discuss the form of instrumental reasoning I shall
use.

(α) (All things considered) it is reasonable for us to pursue goal G. (Relevant things
considered) the best means of bringing about G requires taking M. Therefore, it is
reasonable for us to take M.

The above argument is not a theoretical argument, but a normative practical argument about
means-goal relations. Note that the first premise and the conclusion are normative claims rather
than factual claims. Let me briefly explain why the above argument is valid. Suppose that we can
defend the first premise of (α). Then we can say that it is reasonable for us to pursue G. To say
that it is reasonable to pursue G is to say that it is reasonable to take an appropriate means for G.
Suppose also that we can defend the second premise of (α). Under such conditions, we can
defend taking M on the grounds that the best means of achieving our reasonable goal requires
taking M. For example, if (all things considered) it is reasonable for a person to quench her thirst,
and if (relevant things considered) the best means of quenching her thirst requires drinking
water, then it is reasonable for her to drink water. To put the same point another way, she can
defend drinking water on the grounds that the best means for her reasonable goal, namely,
quenching her thirst, requires drinking water. In this sense, the above instrumental reasoning is
valid.3
One point that should be noted here is that instrumental reasoning by itself does not justify goals,
because it is about means-goal relations. As Nicholas Rescher aptly puts it, “A voyage to a
foolish destination–no matter how efficiently conducted–is a foolish enterprise” (Rescher 1988,
p. 96). Thus, even if M is the most effective means to G, it may be unreasonable to take M in
case G is an unreasonable goal. Therefore, we need to distinguish two kinds of rationality. One is
instrumental rationality. It tells us what to do to effectively achieve one’s goal, but it does not tell
us what to adopt regarding goals. Thus, we must reason not only about matters of efficiency of
goal attainment, but also about the appropriateness of our goals. For now, an important point for
the purpose of this paper is that we can defend taking a means M for a goal G just in case we can
defend not only the reasonableness of G we are pursuing but also the appropriateness of M with
regard to G.

What then are reasonable goals? As Rescher also points out, “the crucial question for rationality
is not that of what we prefer, but that of what is in our best interests–not simply what we happen
to desire, but what is good for us in the sense of contributing to the realization of our real
interests” (ibid., p. 99). Our survival and welfare are vital to our real interests, and so we may say
that our survival and welfare are reasonable goals for us. In other words, it is reasonable for us to
act so as to promote our survival and welfare. I shall discuss the reasonableness of the epistemic
goal in the next section.

Our Epistemic Presumption and Shifting the Burden of Proof

In this section, I shall argue that the reasonableness of our epistemic goal is a minimum
presumption for our epistemic discourse, and that we can shift the burden of proof for this
fundamental epistemic presumption to the skeptic.

As mentioned in the previous section, our survival and welfare are vital to our real interests, and
so we may say that our survival and welfare are reasonable goals for us. Our survival and welfare
can be in jeopardy in case we do not know what to avoid and what to obtain in the world. Thus
correct information about the world is of the utmost importance in order to make rational
decisions for the sake of our real interests at least in the long run. In other words, correct
information about the world is vital to our long-range welfare. Thus we may say that the
epistemic goal is to acquire correct information about the world in such a way that reduces the
risk of getting incorrect information. In this regard, it is worth noting that Sellars argues for the
reasonableness of the epistemic goal on the basis of the reasonableness of the common good.

[T]he prime mover of the practical reasoning involved in probabilistic thinking, indeed
all logically oriented thinking, ‘deductive’ as well as ‘inductive’, ‘practical’ as well as
‘theoretical’, is not an idiosyncratic wish to promote truth, but the intention as a member
of a community to promote the total welfare of that community. This intention is implicit,
and, when it becomes explicit, can be overcome by impulse and self-interest. But it is
because truth is a necessary condition of securing the common good that the search for it
presents itself to us, on reflection, as categorically reasonable—in the truest sense a moral
obligation. (Sellars 1974b, pp. 437–438)
Thus, on Sellars’s view, it is categorically reasonable for us to promote our general welfare, and
truth is a necessary condition for securing the common good, and hence it is categorically
reasonable for us to promote our epistemic goal as well.4

Let us now consider whether the epistemic goal is indeed reasonable. First of all, our concept of
epistemic justification is a normative concept. In our epistemic discourse, we evaluate beliefs as
justified or not on the basis of epistemic norms such as the one that one ought to believe
propositions on the basis of adequate grounds. And we ought to accept justified beliefs, whereas
we ought not to accept unjustified beliefs. Here it is worth distinguishing the deontological
‘ought’ from the propriety ‘ought’. If one’s failure to act in accord with the ‘ought’ implies one’s
being culpable or blameworthy for it, then the ‘ought’ is the deontological ‘ought’. The
representative example is the moral ‘ought’. But there is another kind of ‘ought’ which does not
imply such culpability or blameworthiness. For example, if a botanist says about a trillium that it
ought to have three petals, then he does not mean that the trillium is culpable or blameworthy for
not having three petals; instead, what he means is just that it is appropriate or proper for the
trillium to have three petals. In such a case, the ‘ought’ is the propriety ‘ought’ (cf. Wolterstorff
2005, p. 330). On my view, the epistemic ‘ought’ is the deontological ‘ought’. As mentioned
before, truth is a necessary condition for securing the common good, and so it is categorically
reasonable for us to pursue the epistemic goal. Recall that incorrect information about the world
might put our general welfare in jeopardy. Consequently, if one says to you that p is unjustified
then one’s point is not merely to describe the fact that it fails to meet certain epistemic norms.
The point of saying that p is unjustified involves one’s recommendation or insistence that you
ought not to accept it for the sake of the epistemic goal. Thus, if you hold on to an unjustified
belief despite this kind of recommendation or insistence (without providing any adequate
defense), then you can be blamed for violating the epistemic norms (or being utterly
unreasonable). To put it another way, unjustified beliefs can be subject to rational criticism.5

Second, our concept of epistemic justification is a goal-dependent concept.6 To say that a belief
is justified is to say that it has a favorable status vis-à-vis the epistemic goal. By contrast, to say
that a belief is unjustified is to say that it has an unfavorable status vis-à-vis the epistemic goal.
In addition, as argued in the second section, it is not reasonable for us to take a means M for a
goal G unless G is a reasonable goal. Thus, the normative status of a belief depends on the
reasonableness of the epistemic goal; in other words, if the epistemic goal were unreasonable, the
alleged justified beliefs would lose their normative status.

Third, the reasonableness of the epistemic goal depends on the possibility of bringing about the
goal at least to some extent. As mentioned before, we may say that our epistemic goal is to
acquire correct information about the world in such a way that reduces the risk of getting
incorrect information. But it is the key coherentist insight that we cannot step outside our minds
to judge that our own conceptual states agree with something external to them. As Otto Neurath
(1959, p. 201) aptly puts it, “We are like sailors who must rebuild their ship on the open sea,
never able to dismantle it in dry-dock and to reconstruct it there out of the best materials”. Thus,
the best we can do for the epistemic goal is to gradually improve our conceptual framework from
within. Therefore, on Sellars’s view, our epistemic goal is to gradually improve our conceptual
framework so as to maximize its explanatory coherence, especially to the effect that unexplained
events are minimized and successful predictions are maximized.7 If, however, our conceptual
framework (or our world-view) were widely off the mark, we could not achieve the goal, and so
our epistemic endeavors would be futile. By contrast, if our conceptual framework is not widely
off the mark, we could improve it little by little through try and error. Admittedly, it is possible
that our conceptual framework might be widely off the mark to the effect that the epistemic goal
is unrealizable. But, as argued above, the normative status of the alleged justified beliefs depends
on the reasonableness of the epistemic goal. Thus, if our epistemic goal were unreasonable, then
the alleged justified beliefs would lose their normative status. Then we could not defend or
criticize any claim, and so the end of rational discourse. To avoid this kind of epistemological
disaster, therefore, we should not give up the epistemic goal unless we are given some positive
reasons for thinking that it is indeed unreasonable.

Our survival and welfare are reasonable goals, which we should not give up, unless we are given
some positive reasons to think otherwise. It is surely unreasonable to give up our real chance to
promote our real interests unless we are given some positive reasons for thinking that we have no
such chance. Likewise, we should not give up the epistemic goal unless we are given some
positive reasons for thinking that it is indeed unreasonable. If there were a game which no one
can possibly win, it would be futile for us to try to win the game. Therefore, we try to win a
game on the presumption that it is at least in principle possible for us to win the game. In a
similar vein, if the epistemic goal were unreasonable, it would be futile for us to try to defend or
criticize claims, and so the end of rational discourse. Accordingly, we engage in the practice of
defending or criticizing claims on the presumption that the epistemic goal is reasonable.
Therefore, our epistemic discourse must begin with the presumption that the epistemic goal is
reasonable. Along these lines, we may argue that our epistemic discourse does not begin without
any presumptions, and that it is a minimum presumption for our epistemic discourse that the
epistemic goal is reasonable. Hence, we may say that until and unless the skeptic somehow
provides some positive reasons for thinking that the epistemic goal is indeed unreasonable, it is
reasonable for us to pursue the goal. To put it another way, the burden of proof for the
reasonableness of the epistemic goal lies on the shoulders of the skeptic. Let me elaborate on this
point a bit further.

We have no other way but to rely on our concept of justification in order to justify some claim,
and even the skeptic is no exception insofar as she engages in epistemic discourse with us. In this
regard, notice that the skeptic and the non-skeptic cannot engage in genuine debates unless they
assume (at least for the sake of argument) that one side or the other side can defend their
position. In addition, the infinite regress of justification is impossible. Suppose that a claimer
defends his claim by offering a ground, p. A challenger can call this ground into question by
saying, ‘Why p?’ To meet this challenge, the claimer might provide another ground, q. The
challenger can, in turn, call this ground into question by saying, ‘Why q?’ Here it should be
noted that if the challenger were allowed to keep raising a question, ‘Why is that?’ to any reply
of the claimer, there would be no claim that the claimer can ultimately justify.8 In this regard, it
is worth noting that genuine doubt requires grounds, which are not doubted at the same time.
According to Charles S. Peirce, the mere thought that the Cartesian demon hypothesis might be
true is not a genuine doubt, but an artificial doubt. Genuine doubts require grounds. He says:

Some philosophers have imagined that to start an inquiry it was only necessary to utter a
question whether orally or by setting it down upon paper, and have even recommended us
 to begin our studies with questioning everything! But the mere putting of a proposition
into the interrogative form does not stimulate the mind to any struggle after belief. There
must be a real and living doubt, and without this all discussion is idle. (Peirce 1955, p.
11)

Like Peirce, Ludwig Wittgenstein claims that genuine doubt takes place against a background of
beliefs which are not doubted at the same time. He says:

That is to say, the questions that we raise and our doubts depend on the fact that some
propositions are exempt from doubt, as it were like hinges on which those turn.

That is to say, it belongs to the logic of our investigations that certain things are indeed
not doubted.

But it isn’t that the situation is like this: We just can’t investigate everything, and for that
reason we are forced to rest content with assumption. If I want the door to turn, the
hinges must stay put. (Wittgenstein 1969, §§341–343)

To put the same point another way, there must be a conceptual framework within which doubts
and settlement of doubt take place. Such a framework provides the norms, criteria, or rules for
defending (or criticizing) particular claims under consideration.

Due to the above reasons, the status of the aforementioned epistemic presumption is so
fundamental that it forms the basis of rational discourse. In other words, the epistemic
presumption is part of a conceptual framework which makes our epistemic discourse possible.
Consequently, if the skeptic denies the epistemic presumption, she thereby denies the conceptual
framework within which giving reasons, grounds, and justification takes place. Hence, even the
skeptic must accept our epistemic presumption (at least for the sake of argument) to engage in
epistemic discourse with us. Along these lines, we may say that we can hold the fundamental
epistemic presumption until and unless the skeptic somehow gives us some positive reasons for
thinking that it is false. This is why the burden of proof for the fundamental epistemic
presumption can be shifted to the skeptic.

Let us consider one possible objection here. When the skeptic says that our beliefs are
unjustified, this should be understood as a conditional claim: ‘if justification is what we say it is,
then none of our beliefs is justified.’ Thus the skeptic is not committed to saying that our
understanding of justification is correct. However, even in that case, the skeptic needs to defend
her position by arguing for the conditional claim. In other words, the skeptic needs to assume our
concept of justification (at least for the sake of argument) in order to show us that none of our
beliefs is justified. Second, it is not the purpose of this paper to refute all kinds of skepticism.
Rather, it is to show that we can ward off skeptical challenges to the effect that we can
reasonably continue to pursue our epistemic goal as usual.9

The Normative Practical Argument for an Epistemic Principle

As mentioned in the first section, epistemic principle EP1* says that it is reasonable for us to
accept the general reliability of our perceptual judgments. In this section, I shall offer a
normative practical argument about a means-goal relation which has EP1* as its conclusion.
More specifically, I shall argue that it is reasonable for us to accept the general reliability of our
perceptual judgments on the basis of the following normative argument:

(β) (All things considered) it is reasonable for us to pursue the epistemic goal. (Relevant
things considered) the best means of bringing about this goal requires accepting the
general reliability of our perceptual judgments. Therefore, it is reasonable for us to accept
the general reliability of our perceptual judgments.

This argument is valid, because it is an instance of the normative argument about means-goal
relations, (α), mentioned in the second section. Thus, we can defend the conclusion of (β) by
defending its two premises.

Let us begin with the first premise. I have already argued in the previous section that this premise
is a minimum presumption for our epistemic discourse. Thus, until and unless the skeptic
somehow provides some positive reasons to think otherwise, we can hold this premise. In other
words, as for the first premise of (β), we can shift the burden of proof to the skeptic.

Let us turn to the second premise. We ultimately obtain information about the world through the
senses. In other words, our senses are the ultimate source of information about the world. Thus,
if we could not use our sense organs, then we could not realize our epistemic goal of achieving a
world picture with a maximum of explanatory coherence, especially to the effect that
unexplained events are minimized and successful predictions are maximized. Especially, without
reliable perceptual judgments, we could not bring about the epistemic goal. For this reason, if we
do not accept the general reliability of our perceptual judgments, we cannot get our epistemic
pursuit off the ground. For denying the general reliability of our perceptual abilities is
tantamount to closing our ultimate epistemic door to the world. By contrast, if we accept the
general reliability of our perceptual judgments then, as it seems, we could pursue the epistemic
goal, and thereby have a real chance for bringing about it unless there is no such chance. Along
these lines, we may argue that the best means of bringing about the epistemic goal ultimately
requires accepting the general reliability of our perceptual judgments.

As mentioned above, my defense of the second premise depends on the claim that our senses are
the ultimate source of information about the world. But one may object that this view has been,
historically, a controversial one. Philosophers like Plato and Descartes would reject it, and there
are undoubtedly religious people now who would take the ultimate source of information about
the world to be their preferred scriptures or religious leaders. Moreover, consider a variant of
Berkeley’s view: science allows us to make useful predictions about which experiences we will
have, but it does not capture very well the true account of what produces those experiences. But
these objections do not pose a serious problem for my view.

To begin with, as pointed out in the previous section, we have no other way but to rely on our
concept of justification in order to justify any claim. And it is our conceptual framework that
provides the norms, criteria, or rules for defending (or criticizing) particular claims under
consideration. In addition, I agree with Sellars that it is desirable to justify our epistemic
principles in a naturalistic setting. On his view, our epistemic goal is to gradually improve our
conceptual framework so as to maximize its explanatory coherence, especially to the effect that
unexplained events are minimized and successful predictions are maximized. Granted that we are
in our current conceptual framework, thus, we may argue that the naturalistic world view is such
that it maximizes the explanatory coherence of our conceptual framework. Along this line, we
may take the view that our senses are the ultimate source of information about the world as the
one that maximizes the explanatory coherence of our conceptual framework. I shall say more on
this in the next section.

In addition, in my previous papers (Lee 2008, 2013a, b), I have criticized a foundationalist
picture such that empirical justification has a foundation in beliefs about sensory experiences,
and beliefs about physical objects and their properties are non-foundational beliefs. Thus, I shall
not repeat my arguments here. Instead, let me just mention two things. In the first place, the main
problem with the above picture is that the correct world-view we seek to acquire cannot be built
up merely on the basis of one’s very limited sensory experiences. As far as our present world-
view is concerned, there are more facts in the world than the entire human race have found so
far. And the correct world-view must also cohere with a lot of people’s various observations
about the world. In the second place, as mentioned before, our epistemic goal is to gradually
improve our conceptual framework so as to maximize its explanatory coherence. On this view, if
science allows us not only to explain the past and present facts of the physical world but also to
predict the future facts to the effect that it helps secure the common good, then we can say that it
fully serves its intended purpose. To put it another way, insofar as information about the world is
such that it maximizes the explanatory coherence of our conceptual framework, it can be taken as
correct.10

The Genetic Question Versus the Justification Question

In this section, I shall discuss two important objections to my defense of EP1*.

The main topic of this paper is the problem of epistemic circularity, and so the most crucial
objection to my argument would be that it is not completely free from the problem of epistemic
circularity. In other words, one might argue that I cannot defend a premise (or premises) of my
argument (β) discussed in the previous section unless I assume that sense perception is generally
reliable. For example, if sense perception is not generally reliable, how can I know that the best
means of bringing about the epistemic goal requires accepting the general reliability of our
perceptual judgments?

As noted before, it is inevitable for us to rely on our conceptual framework in order to judge that
our senses are the ultimate source of information about the world. Admittedly, we got into such a
conceptual framework partly by relying on our perceptual abilities. Nevertheless, as Sellars
points out, we should distinguish between the genetic question and the justification question with
regard to our conceptual framework (cf. Sellars 1979, pp. 179–180).

(1) How did we get into our conceptual framework?

 (2) Granted that we are in our conceptual framework, how can we defend accepting the
general reliability of our perceptual judgments?

According to Sellars, the first question addresses the genetic question, which requires “a causal
answer, a special application of evolutionary theory to the emergence of beings capable of
conceptually representing the world of which they have come to be a part” (ibid., p. 180). By
contrast, the second question addresses the justification question. As has been emphasized, we
have no other way but to rely on our conceptual framework in order to justify something. And it
is our conceptual framework that provides the norms, criteria, or rules for defending (or
criticizing) our claims. As a consequence, it is inevitable for us to address any justification
question on the basis of our conceptual framework. This is also the case for the issue of
defending an epistemic principle. This fundamental fact about the justification question does not
necessarily show that any defense of an epistemic principle is bound to be circular.

Keeping the above point in mind, let us consider why my proposal offers a non-circular defense
of the general reliability of our perceptual judgments. The first thing to note is that my argument
is not a theoretical argument but a normative practical argument about means-goal relations.
Thus, the conclusion of my argument is not that our perceptual judgments are generally reliable,
but that it is reasonable for us to accept the general reliability of our perceptual judgments
(unless we are willing to give up our epistemic pursuit). The second thing to note is that my
argument is based on a conceptual analysis of the very concept of epistemic justification.
Admittedly, this sort of conceptual analysis is based on our conceptual framework, which we
have acquired partly by relying on sense perception. But, as pointed out above, we should
distinguish between the genetic question and the justification question. Regardless of how we got
into our conceptual framework, we have no way but to justify the general reliability of our
perceptual judgments on the basis of our conceptual framework, within which my normative
practical argument (β) is valid. In addition, within our conceptual framework, our senses are the
ultimate source of information about the world, without which we could not bring about the
epistemic goal of acquiring correct information about the world. Thus, without reliable
perceptual judgments, we could not bring about the epistemic goal, and so our epistemic pursuit
would be futile. Due to this sort of reason, the reasonableness of the epistemic goal is a minimum
presumption for our epistemic discourse. And, for this reason, unless the skeptic somehow
provides some positive reasons for thinking that our epistemic presumption is false, we can keep
holding it. Therefore, granted that we have no way but to justify the general reliability of our
perceptual judgments on the basis of our conceptual framework, and also that my argument in
question is based on a conceptual analysis, I don’t have to establish the general reliability of our
perceptual judgments in advance before I defend the claim that the best means of bringing about
the epistemic goal ultimately requires accepting the general reliability of our perceptual
judgments. For this reason, even though we do not know in advance that our perceptual
judgments are generally reliable we can still argue that the burden of proof for our fundamental
epistemic presumption lies on the shoulders of the skeptic. To put the same point another way,
insofar as we are not given some positive reasons to give up our epistemic pursuit, it is
reasonable for us to accept the general reliability of our perceptual judgments. In this sense, my
argument is not epistemically circular.
Now let me briefly mention what this paper contributes to the issues of epistemic circularity. As
mentioned in the first section, one important proposal about fundamental epistemic principles
such as EP1 is that we have no option but to accept them unless we are willing to deny the very
possibility of empirical knowledge. This proposal amounts to the old slogan ‘This or nothing’.
As I also mentioned in the first section, this old slogan does not give us any illuminating insight
as to why such epistemic principles are true. My proposal adequately addresses this problem. My
proposal is not just that we have no option but to accept such epistemic principles unless we are
willing to deny the very possibility of empirical knowledge. The most important feature of my
proposal is that it defends such epistemic principles by way of a conceptual analysis of the
concept of epistemic justification. On my proposal, epistemic justification is a goal-dependent
concept in such a way that our epistemic obligation to accept justified beliefs depends on the
epistemic presumption that the epistemic goal is reasonable. Thus, the reasonableness of the
epistemic goal is part of a conceptual framework which makes our epistemic discourse possible.
Under this conceptual analysis, I have provided a normative practical argument which shows
why it is reasonable to accept the general reliability of our perceptual judgments. Therefore, on
my proposal, the reasonableness of accepting the general reliability of our perceptual judgments
is, in an important sense, conceptually connected to the very concept of epistemic justification.
In addition, we have no way but to justify the general reliability of our perceptual judgments on
the basis of our concept of justification. Moreover, it is unreasonable to give up our real chance
to achieve our presumably reasonable goal unless we are given some positive reasons for
thinking that we have no such chance. Hence, on my proposal, we can say that it is reasonable to
accept the general reliability of our perceptual judgments until and unless the skeptic somehow
manages to offer some positive reasons to think otherwise. As has been emphasized, insofar as
the skeptic engages in epistemic discourse with us, the burden of proof for the fundamental
epistemic presumption lies on the shoulders of the skeptic. Due to this reason, an important merit
of my proposal is that I can use the strategy of shifting the burden of proof for our fundamental
epistemic presumption to the skeptic.

Let us consider another important objection. The general form of my main argument is as
follows:

(α) (All things considered) it is reasonable for us to pursue goal G. (Relevant things
considered) the best means of bringing about G requires taking M. Therefore, it is
reasonable for us to take M.

But this argument form permits an instantiation like the following:

(γ) (All things considered) it is reasonable for me to pursue the goal of paying off my ten-
million-dollar debt to Fat Tony (because he’ll kill me if I don’t). (Relevant things
considered) the best means of paying him requires my buying a lottery ticket and hoping
for the best (because I have absolutely no other way of gaining even a chance at ten
million dollars).

Therefore, it is reasonable for me to buy a lottery ticket.

In the above case, buying a lottery ticket may turn out to be the reasonable thing for me to do,
but I have virtually no chance of successfully paying back Fat Tony. In the same way, it may be
reasonable for me to accept that my perceptual judgments are reliable. But that still seems to be
quite a ways from knowing that they are reliable. Along these lines, one might object that my
main argument (β) discussed in the previous section fails to dispel the epistemic worry posed by
the skeptic.11

However, there is an important disanalogy between the above practical argument (γ) and my
practical argument (β). Recall my practical argument:

(β) (All things considered) it is reasonable for us to pursue the epistemic goal. (Relevant
things considered) the best means of bringing about this goal requires accepting the
general reliability of our perceptual judgments. Therefore, it is reasonable for us to accept
the general reliability of our perceptual judgments.

My practical argument (β) is concerned with a means-goal relation, and the relevant goal
involved in this case is the epistemic goal, so that the relevant reasonableness is epistemic
reasonableness (cf. Sellars 1974a). This epistemic reasonableness allows me to claim that it is
epistemically reasonable for us to accept the general reliability of sense perception. By contrast,
the relevant goal involved in (γ) is not the epistemic goal, but a practical goal, and so the relevant
reasonableness is not epistemic reasonableness but practical reasonableness. It is certainly
practically reasonable for me to avoid being killed. In addition, as we shall see shortly, it is
extremely improbable that my buying a lottery ticket really helps to achieve my goal of paying
off my ten-million-dollar debt to Fat Tony in the case of (γ), whereas it is probable in an
important sense that accepting the general reliability of our perceptual judgments really helps to
pursue our epistemic goal.

Sellars distinguishes the non-metrical sense of ‘probable’ from the metrical sense of ‘probable’,
which is found in statistical contexts. In the basic non-metrical sense of ‘probable’ (in relation to
which all other senses are to be understood), to say of a statement or proposition that it is
probable is, in first approximation, to say that it is worthy of credence, that it is acceptable in the
sense of being worthy of credence; that is, to put it in a way which points toward a finder grained
analysis, it is to say that relevant things considered there is good reason to accept it. (Sellars
1974a, p. 368)

On his view, probability (or truth-aptness) in the non-metrical sense is more fundamental than
probability in the metrical sense.12 In the case of (γ), although the best means of paying off my
debt to Fat Tony might require my buying a lottery ticket, we have good reason to think that it is
extremely improbable that my buying a lottery ticket really helps to achieve my goal. This is
because the mathematical probability of winning the lottery is extremely low. In other words,
winning the lottery in this case is extremely improbable in the metrical sense. But the case for (β)
is quite different.

First, as has been argued, it is a minimum presumption for our epistemic discourse that it is
reasonable for us to pursue our epistemic goal. Second, as noted before, we may take the view
that our senses are the ultimate source of information about the world as the one that maximizes
the explanatory coherence of our conceptual framework. Third, it is unreasonable to give up our
real chance to pursue a presumably reasonable goal unless we are given some positive reasons
for thinking that it is very unlikely for us to achieve the goal. Therefore, (relevant things
considered) there is good reason to think that accepting the general reliability of our perceptual
judgments really helps to pursue our epistemic goal. Accordingly, we can say that it is probable
or truth-apt (in the basic non-metrical sense) that accepting the general reliability of our
perceptual judgments really helps to pursue our epistemic goal. In this case again, the skeptic
bears the burden of proof to show that accepting the general reliability of our perceptual
judgments does not really help to pursue our epistemic goal. Due to the above reasons, (γ) is not
a counterexample to show that it is not epistemically reasonable for us to accept the general
reliability of our perceptual judgments.

Sellars and Wright on Epistemic Principles

In this final section, let me further clarify my proposal by making comparisons with Wilfrid
Sellars’s and Crispin Wright’s views, respectively.

To begin with, my defense for the general reliability of our perceptual judgments is somewhat
similar to Sellars’s defense for epistemic principles. Very roughly, his argument runs as follows.
We are effective agents in the world. To be effective agents, we need reliable cognitive maps of
our environment. If epistemic principles such as EP1 did not hold, we could not have reliable
cognitive maps of our environment at all. Thus, epistemic principles such as EP1 must be
elements in our conceptual framework in order for us to be effective agents in the world. Along
this line of reasoning, Sellars claims that fundamental epistemic principles such as EP1 are
justified not by virtue of empirical evidence, but as necessary conditions for a hardly disputable
assumption, namely that we are effective agents in the world. In other words, he defends
epistemic principles by using this kind of transcendental argument.13

As we shall see shortly, there are some important differences between my proposal and Sellars’s
proposal. In the first place, my argument relies on a conceptual analysis of the concept of
epistemic justification. On this analysis, epistemic justification is a goal-dependent concept in
such a way that our epistemic obligation to accept justified beliefs depends on the epistemic
presumption that the epistemic goal is reasonable. Thus, the reasonableness of the epistemic goal
is part of a conceptual framework which makes our epistemic discourse possible. Accordingly,
my proposal does not rely on Sellars’s assumption that we are effective agents in the world. On
my view, his assumption is a bit more vulnerable to objections. For example, one might argue
that other higher animals can be effective agents who have reliable cognitive maps of the world,
but these animals are presumably incapable of making epistemic appraisals by using an
epistemic principle such as EP1. Unlike Sellars’s proposal, my proposal can use the strategy of
shifting the burden of proof for our fundamental epistemic presumption to the skeptic, because
even the skeptic must accept the concept of justification with us (at least for the sake of
argument) in order to engage in epistemic discourse with us (or she needs to show that this
presumption implies an absurd consequence). In the second place, the status of a normative
practical argument about means-goal relations like (β) is less controversial than the status of a
Sellarsian transcendental argument.14 My proposal has these advantages over Sellars’s proposal,
although my view is heavily influenced by his views.
Let us now turn to Wight’s view. My response to the problem of epistemic circularity is
somewhat similar to the response he made in his 2004 paper ‘Warrant for Nothing (and
Foundations for Free?)’. For both are partly based on Wittgenstein’s insight that we cannot but
take certain things for granted. Nonetheless, there are still a few important differences between
the two.

First of all, the problem of epistemic circularity my paper addresses is the one that arises for the
justification of fundamental epistemic principles such as EP1, whereas the problem of circularity
Wright’s paper addresses is a circularity involved in a Moorean anti-skeptical argument.
Admittedly, these problems are related, but nonetheless they are not the exactly same problem.

Second, and more importantly, Wright’s approach to entitlement is a ‘skeptical solution’ to the
challenge of skepticism. On his view, there is a special class of propositions, which he calls
‘cornerstones’. These cornerstone propositions are such as ‘there is an external world,’ ‘there are
other minds,’ and ‘the world has an extended history’. He claims that these propositions are
necessary presuppositions for the significance of our reasoning practice. Thus, to doubt such
propositions is tantamount to jeopardizing our reasoning practice itself. Therefore, our reasoning
practice, which is an indispensable feature of our effective agency, requires us to take
cornerstone propositions on trust. In other words, such propositions play an indispensable role in
our reasoning practice, and so it is rational for us to trust them, despite the fact that we have no
evidence for them. One important feature of Wright’s proposal is that he does not try to show
that such cornerstone propositions are true. Instead, he seeks to show that we are rationally
entitled to accept them, despite the fact that we have no adequate evidence for them. To put it
another way, his strategy starts by granting that we do not possess a proper epistemic
justification for cornerstone propositions, and yet tries to alleviate the uncomfortable
consequences of this skeptical conclusion by way of defending the idea that we may nonetheless
be rationally entitled to accept them. In this sense, Wright’s strategy amounts to a partial
surrender to skepticism, because it starts by conceding the thrust of the skeptical arguments.

However, my approach is not a skeptical solution to the challenge of skepticism. As I

emphasized in the previous section, my proposal uses a conceptual analysis of the concept of
epistemic justification. On this conceptual analysis, it is a minimum presumption for our
epistemic discourse that the epistemic goal is reasonable, and the status of this presumption is so
fundamental that it forms the basis of rational discourse. This strategy is more effective against
the skeptic because even the skeptic must accept the concept of epistemic justification with us in
order to engage in epistemic discourse with us, and also because to give up our fundamental
epistemic presumption is tantamount to giving up rational discourse. Another thing worth noting
is that the skeptic usually does not dispute the legitimacy of our deductive reasoning. Her usual
point is that almost all of our beliefs do not meet her demand for justification. In this regard, it
should be noted that my normative practical argument such as (β) is a kind of deductive
argument. It is so in the sense that its conclusion always holds under the condition that all of its
premises hold.15 Due to these features of my proposal, I can use the strategy of shifting the
burden of proof to the skeptic not only for our fundamental epistemic presumption but also for
the validity of a practical argument such as (β). Due to these reasons, contra the skeptic, I can
argue that it is reasonable for us to accept the general reliability of sense perception until and
unless the skeptic provides some positive reasons to think otherwise. As a consequence, we don’t
need to admit even a partial defeat in our debate with the skeptic. The ball is in the skeptic’s
court.

Third, Wright’s strategy can best be construed to have the structure of a transcendental argument.
On his view, our reasoning practice requires us to take cornerstone propositions on trust, because
to doubt them is tantamount to jeopardizing our reasoning practice. For this reason, it is rational
for us to trust cornerstone propositions without evidence. But this kind of transcendental
argument does not allow us to go beyond the indispensability of our reasoning practice.16 Thus
Wright’s approach is vulnerable to the following line of objection. What the skeptic is targeting
is not the thought that trust in cornerstone propositions is practically rational, but rather the
thought that it is epistemically rational. For any proposition p, one cannot be epistemically
rational in accepting p unless one’s acceptance of p is brought about in such a way that is
appropriately related to the epistemic goal. However, as noted, Wright does not seek to show that
it is epistemically rational for us to accept cornerstone propositions. His point is rather that such
propositions play an indispensable role in our reasoning practice, so that we are entitled to trust
them. For this reason, his appeal to the notion of entitlement is vulnerable to the charge that it
fails to dispel the epistemic worry posed by the skeptic.17 By contrast, my approach is not subject
to this line of objection. I defend epistemic principles such as EP1 by a particular type of
practical argument. This type of practical argument has a certain desirable feature for our issue.
As noted in the previous section, my practical argument is concerned with a means-goal relation,
and the relevant goal involved in it is the epistemic goal, so that the relevant reasonableness is
epistemic reasonableness. This epistemic reasonableness allows me to argue that it is
epistemically reasonable for us to accept the general reliability of sense perception.18

Fourth, my approach appeals to a normative practical argument about means-goal relations rather
than a transcendental argument and the status of such a practical argument is less controversial
than the status of a transcendental argument. Thus, an additional advantage of my proposal is
that it can avoid unnecessary controversies regarding the status of a transcendental argument.

Concluding Remarks

The reasonableness of the epistemic goal is a minimum presumption of our epistemic discourse.
And the status of this presumption is so fundamental that it forms the basis of rational discourse.
Consequently, even the skeptic must accept this presumption in order to engage in epistemic
discourse with us. Hence, by relying on our fundamental epistemic presumption and also by
using a certain practical argument I have defended in this paper, we can defend accepting the
general reliability of our perceptual judgments without falling into epistemic circularity, until
and unless the skeptic provides some positive reasons to think otherwise. That is, the burden of
proof for the fundamental epistemic presumption can be shifted to the skeptic.

In addition, perceptual, introspective and memory judgments have similar epistemic status in our
conceptual framework. Thus, we can defend accepting the general reliability of our introspective
and memory judgments in a similar way as we can defend accepting the general reliability of our
perceptual judgments. Therefore, the fact that we cannot establish the general reliability of our
perceptual, introspective, and memory judgments only by means of theoretical arguments
(without falling into epistemic circularity) is not an epistemological disaster.19

1 In this paper I assume that it is desirable to show that it is reasonable for us to accept epistemic
principles such as EP1. But some philosophers do not think it necessary to show this. Notably,
James Van Cleve (1979) argues that knowledge of an epistemic principle is not necessary for
knowledge to arise in accordance with it, so that it suffices for an epistemic principle to be just
true. However, there is an important reason why we should not ignore the demand for justifying
an epistemic principle. If we cannot meet this demand for justification, we will lose an important
ground on which to rightly reject beliefs based on unreliable principles.

To see the above point more clearly, consider a crystal ball gazer who holds beliefs on the basis
of his crystal ball gazing, because he believes in its reliability. Suppose that we cannot show that
it is reasonable to accept an epistemic principle such as EP1. Then we cannot rationally criticize
the crystal ball gazer’s beliefs despite the fact that he fails to meet the demand for justification,
and so we cannot help but allow his beliefs. This is because we cannot rationally defend our own
epistemic principle such as EP1, and so our epistemic situation is basically equivalent to his
situation. The only difference between the two cases is that crystal ball gazing happens to be
unreliable, whereas our perceptual abilities happen to be reliable. But the big problem with the
above supposition is that we might not know the difference. Consequently, if we cannot help but
allow beliefs based on unreliable epistemic principles, we cannot pursue our epistemic goal in
such a way that we improve our chances of success at achieving the goal by engaging in rational
debate.

In addition, if we cannot epistemically evaluate epistemic principles so as to reliably distinguish

between reliable principles from unreliable ones, having knowledge in accordance with an
epistemic principle will be a matter of luck. This is because having knowledge will depend on a
contingent fact that an epistemic principle on the basis of which one holds beliefs happens to be
reliable. But our epistemic endeavors should not depend on such a sheer luck. Due to these
reasons, it is the core intuition of epistemic internalism that we should not ignore the demand for
justification even regarding epistemic principles. In this respect, it is worth noting that Van Cleve
is an epistemic externalist.

2 If any reason or principle needs to be defended by another reason or principle, our knowledge
is bound to be threatened by an infinite regress of justification. Thus, the Rationalists, such as
Descartes, Spinoza, and Leibniz thought that self-evident first principles, which do not need to
be proven, are required as a solution to the regress problem. For example, Descartes writes:
“[These first principles] should be so clear and evident that the mind of man cannot doubt their
truth when it attentively applies itself to consider them” (Descartes 1911, p. 204). In addition,
according to Roderick Chisholm, if a subject is in what he calls ‘a self-presenting state’, then it is
evident to him that he is in the state in question. For instance: “If seeming to have a headache is a
state of affairs that is self-presenting for S at the present moment, then S does now seem to have
a headache and, moreover, it is evident to him that he seems to have a headache” (Chisholm
1977, p. 22).

3 For a detailed defense of this type of instrumental reasoning, see Lee 2011.
4 For a detailed discussion of this view, see Lee 2011.

5 My claim that justification is a normative concept is a widely accepted view. But there are
admittedly some challenges against it. Most notably, Richard Fumerton makes the following
objection:

Consider, for example, a patient who is told that if she can get herself to believe that she
will recover from the devastating cancer ravaging her body, that will at least increase the
probability that she will recover. That might give the patient a strong reason to try to
acquire the belief even if it is not epistemically rational to believe that she will get well.
… But let us suppose that by forming the relevant belief, the patient produces for herself
a long life devoted to scientific and philosophical investigation, investigation that results
in an enormous number of true beliefs. Despite accomplishing the goal of believing what
is true through believing that shewill get well, our patient (by hypothesis) had no
epistemic reason to believe that she would get well. (Fumerton 2006, pp. 34–35)

On my view, we can handle the above kind of objection in a similar way as utilitarians in ethics
handle the so-called problem of justice by adopting rule utilitarianism instead of act
utilitarianism. On the problem of justice, the overall happiness of a society could sometimes be
maximized by taking an innocent person as a scapegoat, but it is against justice to sacrifice an
innocent person even in such a case. To avoid this problem, some utilitarians uphold rule
utilitarianism, according to which the correct ethical standards should be constructed on an
analogy with the way the utilitarian principle applies to juridical laws. Thus, on this proposal, an
ethical rule is correct if and only if having the rule in effect tends to maximize the overall
happiness of those to whom it applies. I agree with Sellars that the concept of being justified in
holding a belief has been developed on the basis of our social practice of demanding justification
and responding to such demands. On this view, in an analogous way to rule utilitarianism, we
can say that an epistemic norm is correct just in case having the norm in effect is conducive to
our public epistemic goal, and that it is our epistemic norm that one ought to hold a belief only
when it is based on appropriate epistemic grounds. As a consequence, whether one’s belief is
justified or not should be evaluated in terms of whether it conforms to epistemic norms rather
than whether it produces more true beliefs (see Lee 2008 for a detailed defense of this view).
Along these lines, we can argue that the patient in Fumerton’s example violates an epistemic
norm, because her belief that she will recover from the cancer is not based on appropriate
epistemic grounds. Another thing to note is that if a groundless claim is allowed without being
punished or criticized for it, the survival and general welfare of our community might become
more vulnerable to risks due to wrong information at least in the long run.

6 For a detailed defense of the goal-dependent view of justification, see Lee 2011.

7 I have provided a detailed discussion of this view in B. D. Lee, The truth-conduciveness

problem of coherentism and deflationalism, unpublished.

8 The infinitists such as Peter Klein, Jeremy Fantl, and Scott Aikin challenge the claim that the
infinite regress of justification is impossible. It is beyond the scope of this paper to fully address
their challenges. Thus, let me just mention why I do not accept infinitism. On infinitism,
inferential justifications can ramify without end and must do so for a belief that is truly justified.
The most serious problem with this view is, as Carl Ginet (2005, p. 155) points out, that “it is
committed to the thesis that inference alone can create justification.” According to Klein (2005,
p. 136), “[infinitism] can solve the regress problem because it endorses a warrant-emergent form
of reasoning in which warrant increases as the series of reasons lengthens.” Likewise, Fantl
(2003, p. 554) says: “All else being equal, the longer your series of adequate reasons for a
proposition, the more justified it is for you”. Even on Aikin’s (2011) ‘impure’ form of infinitism,
a long chain of inferential relation can create (partial) warrant. This is because he claims that
inferential relations are necessary for any justification, although he denies that they are
sufficient.

To begin with, I agree with Ginet (2005, pp. 148–49) that a formally good inference can only
transfer justification from premises to conclusion so that it cannot create justification even if its
inferential chain is long enough. Suppose that there is an infinite series of justification such that

B1←B2←B3←B4←…

If the justification of B2 depends on B3, we cannot say that B1 is justified by B2 until we know
that B3 has a positive justificatory status. Likewise, if the justification of B3 depends on B4, we
cannot say that B2 is justified by B3 until we know that B4 has a positive justificatory status.
Clearly, to determine whether B1 is justified, we cannot continue this process ad infinitum.
Therefore, unless at a certain finite stage of this process we have a belief which can be accepted
by default without further justification, we cannot determine whether B1 is justified. Along this
line of thought, we can argue that if, as the infinitists claim, justification is a matter of having an
infinite series of non-repeating reasons, we cannot evaluate whether a given belief is justified or
not. If, on the other hand, certain beliefs have default justification such that the burden of proof
for them is shifted to the challengers, then we can evaluate whether or not beliefs are justified on
the basis of such default beliefs, and so we don’t need infinitism. In this regard, it is worth
comparing the following two claims. (i) You always have to answer questions about what you
know until there are not any more questions. And there are in principle no final questions. (ii) As
for beliefs which have default justification, you have to answer a challenge only when it offers a
positive reason for thinking that a given belief is false. We can still be intellectually responsible
by embracing (ii) rather than (i), and so we are not forced to accept infinitism.

9 I have claimed that the skeptic and the non-skeptic cannot engage in genuine debates unless
they assume that one side or the other side can defend their position. One may object, however,
that the ancient skeptics would have disagreed with this claim. According to Casey Perin (2010),
the Pyrrhonists took themselves to be searching for the truth but without assuming that they
could find it. Admittedly, it could be that Pyrrhonists took themselves to be searching for the
truth but without assuming that they could find it. But, as noted, the purpose of this paper is to
show that we can ward off skeptical challenges to the effect that we can continue to engage in
our epistemic endeavors as usual. Thus Pyrrhonists’ view above poses a serious problem for my
view only if they can provide us with a positive reason for thinking that our epistemic
presumption is false. As noted, we have no other way but to rely on our concept of justification
in order to justify some claim, and so we should not give up our epistemic presumption unless
the skeptics show that it is false. And the skeptics can successfully do so only by engaging in
genuine debates with us. Moreover, as also noted, we can engage in genuine debates with the
skeptics only when they assume (at least for the sake of argument) that one side or the other side
can defend their position. It is due to these reasons that Pyrrhonists’ view above does not pose a
serious problem for my view.

10 I have provided a detailed defense of this claim in B. D. Lee, The truth-conduciveness

problem of coherentism and deflationalism, unpublished.

11 This objection is raised by an anonymous reviewer.

12 For a detailed discussion of this distinction, see Lee 2011. See also footnote 18.

13 For this transcendental argument, see Sellars 1979; and for the interpretation of his argument
as a transcendental argument, see deVries 2005, pp. 64–66; O'Shea 2007, pp. 132–34; and
Williams 2009, p. 174.

14 For standard objections against transcendental arguments, see Stroud 1968, pp. 241–256;
Körner 1979; and Stern 2000.

15 An anonymous reviewer raises the following objection: “The author says that the skeptic
usually does not dispute the legitimacy of our deductive reasoning. But some skeptics would, in
fact, dispute the legitimacy of our deductive reasoning. On some interpretations of Descartes, for
example, that is what he is doing in calling clear and distinct perception into question. And,
arguably, Carneades (the Academic skeptic) thought he was engaging in rational discourse even
though he thought we could at most have a limited assent to what is plausible.”

As mentioned before, it is not the purpose of this paper to refute all kinds of skepticism. Rather it
is to show that we can ward off skeptical challenges to the effect that we can continue to engage
in our epistemic pursuit as usual. And we should not give up our concept of justification unless
the skeptic shows us that it is indeed incoherent. Moreover, we can engage in genuine debates
with the skeptic only when the skeptic agrees with us (at least for the sake of argument) that one
side or the other side can defend their position. In this regard, notice that if the skeptic refuses to
accept even the legitimacy of our deductive reasoning, the skeptic and the non-skeptic share
virtually no rule of reasoning on the basis of which the skeptic can refute the non-skeptic. For
this reason, even if the skeptic claims that she can only give a limited assent to what is plausible
the legitimacy of our deductive reasoning must be the kind of thing to which she must give at
least a limited assent in order to engage in rational debates with the non-skeptic. If the skeptic
even refuses to give such a limited assent, the skeptic cannot succeed in refuting our view by
means of a correct argument. For a further discussion of this claim, see Lee 2011.

16 Philie (2009) argues against Wright along this line.

17 For this line of objection, see Pritchard 2007 and Jenkins 2007.
18 Let me discuss one more objection raised by an anonymous reviewer: “Here, the author is
trading on there being two senses of ‘epistemic’. It might mean truth-apt, or it might mean only
relevant to the pursuit of truth. The practical argument is clearly insufficient to establish that it is
reasonable in the sense of being truth-apt (or likely to be true) that our perceptual judgments are
reliable. But if it means only that it is reasonable, in the sense that it is related to our pursuit of
truth (a pursuit that might be totally unsuccessful), then the author has not provided a reply to the
skeptic.”

However, on my view, the above-mentioned two senses of ‘epistemic’ are closely interrelated
with each other at the fundamental level. I endorse the deflationary conception of truth, which is
nowadays very influential among truth theorists in the philosophy of logic. Consider the
following equivalence schema: ‘p’ is true if and only if p. On the deflationary conception of
truth, there is complete cognitive equivalence between the left-hand side of the biconditional and
its right-hand side, and there is nothing else to say about truth other than what the truth predicate
does. And what the truth predicate does is to serve as a vehicle of generalization, semantic
ascent, and certain other logical or expressive function. Due to this reason, truth is not a
substantial concept, and hence it cannot mark a substantial norm of belief independent of the
norms of justification. In addition, and more importantly, I agree with Sellars that probability or
truth-aptness in the non-metrical sense is more fundamental than probability in the metrical
sense. On this view, to say that p is truth-apt in the basic non-metrical sense is to say that
(relevant things considered) there is good reason to accept p vis-à-vis the epistemic goal.
Consequently, if we can defend that (relevant things considered) there is good reason to accept
the general reliability of our perceptual judgments vis-à-vis the epistemic goal then this is
tantamount to defending that the general reliability of our perceptual judgments is truth-apt (in
the basic non-metrical sense). For a more detailed discussion and defense of this view, see Lee
2011, and B. D. Lee, The truth-conduciveness problem of coherentism and deflationalism,
unpublished.

19 I’d like to thank an anonymous reviewer for his or her very helpful comments on an early
version of this paper.
Viciousness and the structure of reality
Ricki Leigh Bliss
Philos Stud (2013) 166:399–418

Abstract Given the centrality of arguments from vicious infinite regress to our philosophical
reasoning, it is little wonder that they should also appear on the catalogue of arguments offered
in defense of theses that pertain to the fundamental structure of reality. In particular, the
metaphysical foundationalist will argue that, on pain of vicious infinite regress, there must be
something fundamental. But why think that infinite regresses of grounds are vicious? I explore
existing proposed accounts of viciousness cast in terms of contradictions, dependence, failed
reductive theories and parsimony. I argue that no one of these accounts adequately captures the
conditions under which an infinite regress—any infinite regress—is vicious as opposed to
benign. In their place, I suggest an account of viciousness in terms of explanatory failure. If this
account is correct, infinite grounding regresses are not necessarily vicious; and we must be much
more careful employing such arguments to the conclusion that there has to be something
fundamental.

Keywords Vicious infinite regress _ Metaphysical foundationalism _ Grounding _ Metaphysical

explanation

Arguments from vicious infinite regress have been a mainstay in the philosophical tradition
throughout its history. They are employed in a rich variety of contexts to demonstrate the
unacceptability of the position under consideration. Given the centrality of arguments from
vicious infinite regress to our philosophical reasoning, it is little wonder that they should also
feature on the catalogue of arguments offered in defense of theses that pertain to the fundamental
structure of reality. In particular, proponents of a position we can call metaphysical
foundationalism will claim that the infinite regress(es) of grounds that ensues failing the
existence of something fundamental is vicious, and thus, unacceptable.1 The foundationalist
concludes that, on pain of vicious infinite regress, there must be something fundamental.

Not all infinite regresses are vicious. Benign infinite regresses, or what some, perhaps, feel more
comfortable referring to as infinite series, are commonplace. The set of natural numbers forms an
infinite series, for example; and the truth regress is commonly held to be benign. The
foundationalist, then, needs to provide us with a reason for taking infinite regresses of grounds to
be vicious instead of benign. One possible reason to suppose that regresses of grounds are
vicious is that to be a vicious infinite regress just is to be a grounding regress: all vicious infinite
regresses are grounding, or in virtue of, or dependence regresses.2

But claiming that the infinite regresses of grounds that would ensue failing the existence of
something fundamental are vicious because to be a vicious infinite regress is to be a grounding
regress is different to telling us why infinite regresses of grounds are vicious. Although it may
well be true that all vicious infinite regresses are grounding regresses, as a premise employed in
an argument to the conclusion that grounding regresses are necessarily vicious, this is hopelessly
circular. And to move from this assumption to conclusions about the fundamental structure of
reality would seem wholly unacceptable. Again, it may well be true that infinite grounding
regresses are vicious, and thus unacceptable, but what cannot be true is that we are supposed to
believe this because to be a vicious infinite regress just is to be a grounding regress. What the
foundationalist requires, then, is an answer to the question of why infinite regresses of grounds
are vicious as opposed to benign.

In the existing literature on viciousness, there is limited agreement on what makes for the
viciousness of a vicious infinite regress. In fact, there is limited agreement on which regresses
we can consider vicious at all. Of the regresses that are taken to be genuinely vicious, there are a
variety of prospective analyses available.

Some philosophers are of the view that vicious infinite regresses are involved with contradiction:
theories that generate vicious infinite regresses also harbor contradictions.3 Mention is also made
of underlying dependence relations: vicious infinite regresses are ordered by a dependence
relation that transmits a property, the origins of which we seek for.4 Another suggestion available
in the literature is that viciousness is involved with reduction: theories that aim at being reductive
but fail generate vicious infinite regresses.5 It has even been suggested that viciousness is
involved with theoretical virtue: infinite regresses can be deemed vicious because of their
ontological extravagance.6

In this piece, I explore the more general question of what the viciousness of a variety of vicious
infinite regress consists in. Although independently interesting, not to mention philosophically
significant, I explore these issues, however, with an aim to understanding the more specific
question of why, or indeed whether, infinite grounding regresses are vicious. An answer to this
question, I hope to have impressed upon the reader, is vital to any defense of the foundationalist
picture that employs an argument from vicious infinite regress. Without a clear account of what
the viciousness of the relevant purported vicious infinite regresses consists in, this argument in
defense of the necessity of the existence of something fundamental doesn’t even get going.

In Sect. 1, I introduce some examples of vicious and benign regresses. In Sect. 2, I discuss the
most common accounts of viciousness available in the literature: contradictions, dependence
relations, reductionism and parsimony respectively. I argue that, although each may only be
proposed as an analysis of only a subset of vicious infinite regresses, each fails to explain the
viciousness of even the subset at which they are aimed. In Sect. 3, I propose an alternative
account of viciousness in terms of a specific kind of explanatory failure. According to the
proposed account, an infinite regress is considered vicious when the regress is generated because
we are seeking an explanation for a proposed phenomenon but fail. Whilst this account does not
seem to be true of all genuinely vicious infinite regresses, I propose it as an account of a subset
of them; amongst which are grounding regresses. In Sect. 4, I come finally to a discussion of
metaphysical foundationalism and grounding regresses. Here I argue that if the proposed account
is correct then it is only under certain conditions that regresses of grounds are vicious.

1 Infinite regresses: vicious and benign

Take the fact, or the state of affairs, that a is an F - [Fa]. If Fa is a contingent truth, it seems to be
the case that both F and a could just as well have existed without being united in a particular
state of affairs at all. a could have been a G instead of an F; and b could have been an F instead
of a. For anyone who posits the existence of states of affairs, there is an additional metaphysical
story that needs to be told regarding their unity.

Accounts of structured facts (or propositions) are notoriously associated with what is commonly
taken to be a version of a Bradley regress. In answer to the question of how universals and
particulars come together to form unified entities—states of affairs—one might be inclined to
posit the existence of an instantiation relation. We could say that a instantiates by relation R the
property F, giving us something like [aRF]. But this very quickly leads us off on a regress, for
we then need to say how this relation, R, relates a to F—giving us something like [R1[aRF]];
how R1 does its relating and so on ad infinitum. It seems that with every structured state of
affairs (or proposition) comes an infinitude of further states of affairs or propositions.7

The problem of unity, as it pertains to states of affairs, meets with one (or many) of several
responses. Many consider the regress to be vicious, and its appearance the death knell for states
of affairs. Others admit that the regress is engendered, but claim that it is benign;8 and others yet
deny its appearance at all.9 I will not discuss accounts which deny the appearance of a regress,
for they are not relevant to the present discussion. Let’s begin by considering the benign reading
of the regress.

The benign reading of the regress has it that the initial state of affairs acts as a truthmaker for the
ensuing regress of truths that pertain to the unity of the state of affairs under question. A state of
affairs will engender an infinite number of propositions that pertain to the unity of that state of
affairs. It will be true that <aRF> and true that <R1<aRF>>, and so on ad infinitum. The regress
is benign in the same way that the Tarski truth regress is taken to be benign: every truth entails
an infinite number of truths that pertain to its truth, where the truth of the initial truth is in no
way dependent upon any of the truths (pertaining to its truth) that it entails. If ‘snow is white’ is
true if and only if snow is white, then ‘‘snow is white’ is true’ is true if and only if snow is white,
and so on ad infinitum.

How does the vicious reading differ from the benign reading? Let’s say that in order to explain
how the unity of the initial state of affairs is achieved, we need some relation of instantiation, R.
Let’s also say that in order to account for how R relates a to F we need some further relation R1
—giving us some further fact [R1[aRF]]. But of course, in order to account for how R1 relates R
to a and F, we need some further relation R2—giving us some further fact—and so on ad
infinitum. At each stage, in order to explain the unity of the state of affairs we need yet another
instantiation relation. And in needing yet another relation, we are never able to account for how
the unity of the state of affairs is achieved. This vicious reading of the regress is often cast in a
dependence, or in virtue of, idiom. It is claimed that the regress is vicious where each fact of
instantiation obtains in virtue of some further fact of instantiation.10

I do not purport to adjudicate on the question of the unity of states of affairs. Whether or not a
regress is even engendered is itself debatable, and an issue for elsewhere. All I hope to have
demonstrated is that if one is of the view that a regress is generated, there are accounts that take
the regress to be benign and accounts that take it to be vicious.

2 Some prospective analyses

There are a variety of prospective analyses of viciousness available in the literature. Not only do
philosophers disagree on which infinite regresses are vicious as opposed to benign, and if
vicious, the reasons for which they are vicious. Philosophers also disagree on whether there is
but a single explanation of what the viciousness of a vicious infinite regress consists in. For
present purposes it matters not whether we take the prospective analyses to be analyses of
vicious infinite regresses simpliciter, or merely analyses of a subset of regresses. I will argue that
each of the prospective accounts of viciousness is flawed. In which case, they cannot even be
analyses of a sub-set of regresses.

The first suggestion commonly made in the literature is that there is a tight connection between
viciousness and contradiction: theories that engender vicious infinite regresses also harbor
contradictions.11 An example of a theory that generates a vicious infinite regress, and also
contains a contradiction, is the Theory of The Forms. To see this, we can look to the Third Man
Argument of the Parmenides as presented by Vlastos and discussed by Nolan—who follows
Vlastos. The first assumption of the argument is as follows:

(A1) If a number of things, a, b, c are all F, there must be a single form, F-ness, in virtue
of which we apprehend a, b, c as all F.12

F here is ‘largeness’, but the same argument could be run for any discernible property. We
therefore have four things that fall under the form F-ness—a, b, c and F. This assumption is
followed by a second, together which, essentially give us the Third Man Argument.

(A2) If a, b, c and F-ness are all F, there must be another form, F1-ness, in virtue of
which we apprehend a, b, c, and F-ness as all F.13

It is clear that the theory of the forms leads us off on an infinite regress. In order to apprehend F
it must fall under a form—F-ness. But in order to apprehend a, b, c and F-ness as F, there must
be another form in virtue of which F-ness is apprehended as F, F1-ness, and so on ad infinitum.
The regress also appears to be vicious because in order to apprehend any form, we require yet
another form.
To better understand this argument, we must introduce two additional and suppressed
assumptions to which Vlastos believes Socrates must be committed for the argument to work:

(A3) Any form can be predicated of itself. Largeness is itself large. F-ness is itself F.14

And the final premise:

(A4) If anything has a certain character, it cannot be identical with the form in virtue of
which we apprehend that character. If x is F, x cannot be identical with F-ness.15

In order to all be F, a, b, and c fall under the form F-ness. By (A3), this form—F-ness—is itself
F. Yet, by (A4), we are told that in order to be apprehended, no character by way of which we
apprehend anything can be identical to itself. Therefore, if we are to apprehend anything as being
F, if we are to apprehend F-ness, F-ness must itself fall under a form to which it is not identical.
F-ness (plus a, b, etc) must fall under the form F1-ness and so on ad infinitum. The theory of the
forms generates a vicious infinite regress. But what of the contradiction?

Vlastos is explicit that having made the two suppressed assumptions explicit, we need not go by
way of the infinite regress to see that the theory harbors a contradiction. Taken together, (A3)
and (A4) lead to the conclusion that every form both is and isn’t identical to itself. Contradiction.

It is difficult to establish the priority ordering between the vicious infinite regress and the
contradiction in theories that contain both. Although it is agreed that there is a tight connection
between some cases of viciousness and contradiction, it is not clear which is the symptom and
which is the disease. Nolan at one place states ‘the first pathology of which an infinite regress
may be a symptom is that of a lurking contradiction’.16 Whilst he abandons this suggestion, he
entertains the thought that if we scratch around internal to a subset of theories that generate
vicious infinite regresses, we will likely find a contradiction; and it is the contradiction that may
be the real source of the problem. If this view is correct, then we have not only an account of a
correlation between viciousness and contradiction, but also an account of which is responsible
for the generation of the other: the contradictions in contradictory theories generate vicious
infinite regresses. All that remains is to develop an account of how.

In the absence of an account which explains the mechanism involved whereby the contradiction
yields the infinite regress, I would like to suggest that there seems to be at least one prima facie
plausible picture on which the vicious infinite regress could be taken to be responsible, albeit
indirectly, for generating the contradiction. I would like to suggest that it seems plausible that the
vicious infinite regress is the underlying problem with the contradiction surfacing as its
symptom.17

One way in which a vicious infinite regress might deliver a contradiction is if one of the
premises that generates the contradiction is inserted as a means of halting the looming infinite
regress. To see this point, let’s return briefly to Vlastos’ rendering of the Third Man Argument.
As already stated, taken together, (A3) and (A4) lead to the generation of a contradiction. But the
theory also harbors a second contradiction. In (A1) we can see the stipulation that there ‘must be
a single form’. But as we know, the theory leads to the generation of multiple forms. In
particular, the demand in (A4) that ‘it [a character] cannot be identical with the form...’ drives us
to this conclusion. The theory harbors two contradictions: the first contradiction arises as we
have both one and many forms of a single form at the one time, namely F-ness, F1-ness, F2-ness,
and so on ad infinitum. And the second arises because we have F-ness being both identical to and
different from itself ((A3)+(A4)). It is the first of these contradictions that is of particular
interest; the insistence that there ‘must be a single form’, (A1), could be viewed as aimed at
halting a looming infinite regress of forms delivered by (A4). And it is the inclusion of this
assumption—(A1)—that, in turn, pushes the theory to contradiction.

If even a subset of vicious infinite regresses are generated by theories that also harbor
contradictions, then it seems plausible that one is involved in delivering the other. One
suggestion is that contradictions are the pathology of which the infinite regresses are but a
symptom. If this proposal is correct, what the appearance of a vicious infinite regress ought to
alert us to, then, is that the theory that has generated it is contradictory. But I have suggested,
absent any clear articulation of how the relevant contradiction generates an infinite regress, it
seems equally as plausible that the contradiction is the symptom rather than the disease. And I
have suggested the possible mechanism by way of which this might take place. But in this latter
case, where we may have uncovered something interesting about viciousness, namely that
vicious infinite regresses harbored within theories are (in some cases) involved in rendering them
contradictory, we have not yet uncovered what the viciousness of a vicious infinite regress
consists in.

I’m not entirely sure how we are supposed to adjudicate between these two possible accounts of
the relationship between viciousness and contradiction.18 And the possibility of the one seems to
mitigate against the certainty of the other. Of course, there remains a third possibility: that both
the contradiction and the vicious infinite regress are symptoms of a deeper problem. This seems
to be the position that

Nolan settles with.19 If this is correct, which I think it probably is, we are still wanting for an
account of what the viciousness of a vicious infinite regress consists in. Again, at most we can
note that there is an interesting and important connection between some cases of viciousness, and
contradiction, without yet having an account of what that viciousness consist in.

We turn now to the relationship between viciousness and dependence. Another suggestion
commonly made in the literature is that vicious infinite regresses involve an underlying
dependence relation.20 What is significant about this dependence relation is that it involves the
transfer of a certain significant property—being true, being, being justified, for example. The
regress gets going because the phenomena invoked at each stage depend upon something else for
said property; and the regress is deemed vicious because although it doesn’t terminate, the
thought is that it should.

Many of the vicious infinite regresses with which we are familiar are of this structure. One
powerful reason to be an epistemic foundationalist, for example, is the threat of looming vicious
infinite regress. If each one of our beliefs depends upon some further belief for being justified, it
is not clear, says the foundationalist, how any belief is justified at all. The regress must terminate
for any of our beliefs to count as properly justified. To this end, epistemic foundationalists posit
the existence of foundational, non-inferentially justified beliefs, which serve to terminate the
regress and ensure that the beliefs that depend upon them are indeed justified.

Some construals of the metaphysical foundationalists’ version of the grounding regress also fit
this structure. If everything depends upon something else for its being or existence, says the
foundationalist, then it is not clear how anything has any being or existence at all. In this same
vein, Leibniz, for example, was of the view that there needs to be simples—terminus points of
the regress—if things are to have reality:

Where there are only beings by aggregation, [composite] objects, there are no real
beings. For every being by aggregation presupposes beings endowed with real unity
[simples], because every being derives its reality only from the reality of those beings of
which it is composed, so that it will not have any reality at all if each being of which it is
composed is itself a being by aggregation, a being for which we must still seek further
grounds for its reality, grounds which can never be found in this way, if we must always
continue to seek for them.21

What is wrong with these kinds of regresses is often expressed in terms of a variety of different
metaphors. Philosophers will say that if the regress does not terminate, it is not clear how things
got started, or how things got off the ground.22 They will worry that if everything is borrowing a
property from something else, then it is not clear where this property is actually coming from.23
The employment of different metaphors to express the concern evidenced by the charge of
viciousness will no doubt influence what we take to be at issue. Worse, they will likely bring
with them conceptual baggage that muddies our ability to think clearly about what the actual
matter to hand is. In any case, these arguments all seem to fit the structure set out above.

What gets us from the appearance of the regress to the charge of viciousness is the thought that
the regress must terminate, which of course, is exactly what it doesn’t do. But why suppose the
regress must terminate? One suggestion is that it must terminate because we are dealing with a
known finite domain.24 But this cannot possibly be a reason to think that the regress must
terminate. After all, isn’t what we are trying to do to establish whether or not the regress can be
infinite? To understand what is problematic about this argument, we need first to consider the
variety of reasons we could have for taking a domain to be finite: sometimes our reasons will be
pragmatic or based on theory preference, for example. The appearance of an infinite regress in
such a domain ought to force us to seriously question the reasons we have for supposing the
domain to be finite in the first place. More to the point, this claim, as it stands, is question-
begging. Exactly what we are seeking for is a reason to think that the regress cannot be infinite,
to claim that it must be finite is patently circular.

Another common reason to suppose that the regress must terminate is that it must be completed
by some categorical property possessor.25 If this argument is simply a disguised version of the
question-begging formulation of the argument just mentioned, it is unacceptable.26 The
assumption that the regress must be completed cannot serve as an assumption in an argument to
the conclusion that the series cannot be infinite.
A more sophisticated rendering of this formulation of the argument targets the idea that unless
there is something in the chain that possesses the property under consideration categorically,
nothing within the chain can be said to possess that property at all. Take the worry over the
property being real expressed by Leibniz above. If every composite object has its reality in
dependence upon its parts, then it is not clear where all this reality is actually coming from. So
too with the justification regress. If all of our beliefs are justified, in virtue of their dependence
upon other beliefs, it is not clear how any of the beliefs are justified at all.27 It’s not clear how, in
these cases, the entities under consideration are to come to have the property at issue without the
completion of a certain, relevant process.

But we must ask ourselves, however, why we should assume that conditional property
possession—possession of a property in dependence upon something else—is an inferior mode
of property possession. Perhaps it is the case that being real in dependence upon something else
is as real as anything is, or can be. Perhaps it is the case that being justified in virtue of being
suitably related to further beliefs, that are themselves justified, is as much justification for our
beliefs that we can hope to have. The inclusion of the assumption that the chain must terminate
in a conditional property possessor—the assumption that we must complete a certain process—
forces us to conclude, when the regress does not terminate, that the regress is unacceptable. But
this assumption, at best requires independent justification, and at worst begs the question, for
again, it demands an end point to the chain, which is exactly what we are trying to establish an
argument for.

But perhaps this independent justification is not so far off. One might suggest that, although in an
infinitely regressive chain we have accounted for the conditional possession of the property
under consideration of everything within the chain in terms of that upon which it depends, we
have not yet explained everything that needs an explanation. The function of the categorical
property possessor, then, is to explain how the whole chain has the property at all. The function
of the categorical property possessor is to explain how anything can have said property
whatsoever. In a reality that admitted of no foundations, for example, although everything has its
reality accounted for in terms of that upon which it depends, we have failed to explain how the
whole lot of them—everything—has any reality at all.

Aikin characterizes this feature of the regress argument in terms of a perspective shift: the
regress is only vicious when we shift between the mediate and global perspectives.28 From the
mediate perspective, each phenomenon has its possession of the property under consideration
explained in terms of that upon which it depends. The concern is generated, however, when we
shift to the global perspective and wonder how the whole (possibly infinite) lot taken together
gains the property under consideration.

Legitimate though the thought may be, however, the regress is not designed to answer this
question.29 All the regress can tell us is how each individual member has the property under
consideration, namely, in dependence upon something else. The appearance of an infinite regress
should not lead us to conclude that nothing within the regress has the property under
consideration—nor has its possession of that property unexplained—but rather that not
everything about the possession of the property that needs to be explained has been. To claim
that an infinite regress is vicious because it doesn’t allow us to answer the global question is to
have accused it of having failed to carry out a task it was not designed to complete.

Finally, I would like to address, very briefly, a number of further claims regarding explanations
of what viciousness consists in. Nolan suggests that there is an important connection between
viciousness and reductionism: some theories that aim at being reductive but fail generate vicious
infinite regresses.30 One such theory is the homuncular theory of perception.31 According to this
theory, perception is made possible by an homunculus that sits behind the eye. It receives visual
information, and processes it, before sending it to the brain. How does the homunculus do its
work of perceiving? According to the theory, the homunculus perceives by way of a further
smaller homunculus nested within the original homunculus; and we are off on a regress.

Bennett also notes a connection between vicious infinite regresses and a failure of reductive
explanation. In a discussion of the potential problems associated with grounding being non-
fundamental—with grounding itself being grounded, leading us off on a regress—she states:

Yet surely something is wrong with the idea that it is really grounding in both places. I
suppose the problem is something like this. We are trying to tell a story about how the
grounding relation comes to be, what brings it into the world, what makes it the case that
it exists in the first place. That story cannot invoke grounding anywhere—neither in the
ground itself, nor in the relation by means of which the grounds counts as a ground. The
problem, in short, is a special kind of failure of reductive explanation...32

What is interesting about the connection between viciousness and failed reductive analyses is not
simply that the reductions fail, but the specific way in which they fail. After all, there are many
ways in which a reductive analysis can fail that are not going to be involved with the generation
of vicious infinite regresses. A more illuminating means by which we can approach the
connection between viciousness and failed reductive theories, in my view, is in terms of a
characterization of this failure itself. And this I reserve for Sect. 3.

Nolan also suggests that, certain at least, amongst the regresses commonly taken to be vicious,
can best have their viciousness accounted for in terms of theoretical virtue.33 Nolan suggests that
the turtles’ regress might be considered vicious because of its ontological extravagance. The
turtles’ regress gets going when we try to answer the question of what holds the (physical) world
up in space. The suggestion is that it sits atop the back of a giant world turtle. The world plus
world turtle also requires support, and thus rests on the back of a second world turtle, and so on
ad infinitum. Nolan suggests that it is the ontological extravagance of the regress that is the
reason for which we can consider it to be vicious. Where ten world turtles are obviously silly, an
infinite number of them is just plain ridiculous.

I think this argument is mistaken. First, I think it fails to recognize a feature of the turtles’ regress
—and other regresses mentioned by Nolan—that is shared with regresses whose viciousness we
have a good explanation for. Exactly what these features are will become clearer in the following
section. Second, considerations of theoretical virtue are supposed to help us with theory
preference. An argument from theoretical virtue tells us that, all things considered, we should
employ the theory that is simplest. Whereas what the appearance of a vicious infinite regress has
typically been used to show is that a theory is deeply and inextricably flawed. Not merely
clunky.34 It seems a mistake to think that an infinite regress can be deemed vicious for reasons of
theoretical virtue alone.

3 Viciousness and explanation

A subset of vicious infinite regresses are generated as a consequence of continually attempting to

overcome an explanatory failure that arises at the first level of the analysis.35 The problem in
the case of, at least a subset of, vicious infinite regresses arises at the first level of analysis, and
continues to recur at each level thereafter. Even at infinity, it is not escaped. This point was also
noted by Passmore:

As I have already suggested, it is the first step in the regress that counts, for we at once,
in taking it, draw attention to the fact that the alleged explanation or justification has
failed to advance matters; that if there was any difficulty in the original situation, it
breaks out in exactly the same form in the alleged explanation.36

To see this point, let us return to the homuncular theory of perception. At each stage of the
analysis we are forced to invoke an homunculus further, exactly because we have failed at that
level to explain how the homunculus perceives. The problem is repeated at each level of the
analysis, but it arises at the very first level.

That the problem is repeated at each level of the analysis tells us what leads to the generation of
the regress; it does not tell us what the problem is in the first place. On what the initial
explanatory failure consists in, Passmore states:

Philosophical Regresses, on the contrary, demonstrate only that a supposed way of

explaining something or ‘making it intelligible’, in fact fails to explain, not because the
explanation is self-contradictory, but only because it is, in the crucial respect, of the same
form as what it explains.37

The explanatory failure that occurs at the first level of the analysis consists in the fact that the
explanans is of the same form as the explanandum: the phenomenon for which we are seeking an
explanation reappears as its own explanation. Whilst a comprehensive characterization of this
sameness of form is beyond the scope of the present paper, I offer here a preliminary sketch,
largely by way of example.

According to Ryle’s analysis of an intelligent action:

The consideration of propositions is itself an operation of the execution of which we can

be more or less intelligent, less or more stupid. But if, for any operation to be intelligently
executed, a prior theoretical operation had first to be performed and performed
intelligently, it would be a logical impossibility for anyone ever to break into the circle.38

This regress, as presented by Ryle, is sometimes understood as a temporal regress: where each
intelligent action needs to be preceded by an intelligent thought—and time had no beginning—
no intelligent action could ever take place. The reason for this is that where there is no first
intelligent thought, there can never be any subsequent intelligent action. Passmore points out,
however, that understanding the regress thus is to have missed the deeper import of the problem
to which Ryle is pointing.39 The regress gets its legs when we understand it to be generated
because we are seeking to explain something about the nature of intelligence itself. Passmore
believes that the regress gets going because we are seeking a constitution-explanation of
intelligence.

If we are seeking to understand intelligent action by pointing out that it must be preceded by
intelligent thought, the regress show us that we can never break into the cycle. The reason for
this is that where intelligent thought is constitutive of intelligent action—and we can only
explain intelligent thought in terms of intelligent action—we fail to explain anything about
intelligence at all. The explanatory failure that generates the regress lies in the fact that in order
to explain something about the nature of intelligent action, we invoke intelligent thought: we
invoke the very phenomenon under investigation—intelligence—to explain the phenomenon
under investigation—intelligence. That for which we are seeking an explanation is supplied as
that which does the explaining.40 Or put another way, the analysis of intelligence requires an
understanding of the notion of intelligence—the analysis of intelligence requires an application
of the very concept for which we are seeking an explanation.

Again to the homuncular theory of perception, and the turtle’s regress. In order to explain how
perception occurs we invoke the existence of an homunculus. But in order to explain how that
homunculus facilitates perception, we need a second homunculus whose perceptual activity also
requires explanation, and so on ad infinitum. At each stage of the regress we supply as explanans
the object whose perceptual activity we are seeking an explanation for in the first place. Invoking
homunculus after homunculus only serves to show us that we have explained nothing about how
perception actually works.

As with the turtle’s case. In order to explain what holds the world up in space, we posit the
existence of a world turtle. In order to explain what keeps the world plus world turtle up, we
supply a second world turtle, and so on. What we are seeking to explain is how something stays
up in space, and, yet, at each stage of the analysis we posit the existence of something whose
capacity to stay up in space is in need of explanation. Invoking turtle qua object of support after
turtle qua object of support shows us that we have explained nothing about how anything is
supported at all.

One would not be remiss in wondering if the sameness of form that is troubling in the turtles’
case lies in the infinite number of turtles. The version of the regress that involves the turtle that
sits atop the monkey which sits on the back of the elephant, and so on, would, therefore, be
acceptable.41 But the sameness of form that leads to the charge of viciousness in the turtle’s case
is also present in this second case as well. At least if what the regress is being generated in order
to explain is how the world stays up in space. For no matter how many turtles, monkeys,
elephants or flamingos we conjure up, for each we are left needing to explain what holds it up;
along with everything else it is supporting. What is relevant, then, is the explanatory role such
zoological exotica are playing qua objects of support. And not whether they are the same in other
respects. The version of the regress that involves only turtles just wears its viciousness a bit more
boldly on its sleeve.

Bradley’s regress is also vicious for the same reason. The regress is generated when we are
trying to explain how any unity is a unified object. In an attempt to explain the unity of whatever
is under consideration—a bicycle, a state of affairs—some new phenomenon is introduced—an
instantiation relation or something of the like. But we must then explain how this new relation
creates unity. At each stage of the regress we posit the existence of some fact—or proposition or
relation—to explain the unity of the initial phenomenon under question. But the introduced
phenomenon—the fact, proposition, relation—if it is to have any explanatory force whatever
must, for a brief moment, be assumed itself a unified object, or involved in the creation of unity.
What the regress then shows us is that this new entity must also have its unity explained, and so
on ad infinitum. At each stage of the regress we invoke some object qua unity to explain how the
initial object under consideration is unified, and only end up demonstrating that we have
managed to explain nothing about unity at all.

Each member of the regress is numerically distinct from that which precedes it; qua explanans
and explanandum, however, they are identical. And in so being, we never break out of the
explanatory failure that gives rise to the regress by breaking into the cycle with an adequate
explanation. These notions of sameness of form, constitution explanation and explanatory failure
require much more elaboration. The sense of form in operation here is not that of logical form,
for example. Explanans and explanandum can be of the same logical form without that entailing
explanatory failure. But I leave off this discussion here for these issues are extremely complex,
and turn, instead, to some consequences of the proposed analysis.42

The first of these is that we should take the proposed way of explaining the phenomenon under
consideration to be incorrect: invoking instantiation relations cannot explain the unity of a state
affairs; intelligent acts cannot help explain what the intelligence of intelligent acts consists in.
The appearance of the regress needs not show us that the question we have posed is incorrect but,
rather, that we ought to abandon the proposed way of answering it.

The second consequence of note is that the fact that we have invoked an infinitude of phenomena
is not the problem. A vicious infinite regress, of this kind at least, is problematic because we fail
to achieve the explanation that we are after, not because we require an infinite number of things
to do it. Whatever reasons we might have for disliking infinities, they should have nothing to do
with viciousness. Just as it is a perfectly harmless feature of the world that the set of integers
forms an infinite series, it may well be a perfectly harmless feature of the world that there are an
infinite number of states of affairs, structurally speaking.

4 Viciousness, grounding and the structure of reality

We come finally to the doctrine of metaphysical foundationalism; and arguments in its defense
that employ the notion of a vicious infinite regress. According to the metaphysical
foundationalist, reality is hierarchically structured with chains of phenomena ordered by the
grounding relation terminating in something fundamental. 43 These fundamentalia, thinks the
foundationalist, are ontologically independent or ungrounded.
Philosophers are commonly of the view that grounding is involved with metaphysical
explanation. How, exactly, metaphysical explanations behave, and how they may differ from
other, more familiar, forms of explanation itself stands in need of further explanation. Here, I
assume, however, that if x is grounded in y, then the existence, or being, of y explains the
existence, or being, of x: where the notion of being differs from that of existence insofar as
things with being exist and have natures. We say, then, that y metaphysically explains x.44

Although often not clearly articulated, there are a stock of arguments employed in defense of the
foundationalist claim that there must be something fundamental.45 Amongst these arguments, as
already noted, are arguments from vicious infinite regress. The foundationalist will claim that
there must be something fundamental on pain of vicious infinite regress.

How successful this argument is will turn upon whether or not infinite grounding regresses are
vicious as opposed to benign. As stated in the introductory remarks, that infinite regresses of
grounds are vicious because all vicious infinite regresses are grounding regresses is no reason to
suppose that infinite grounding regresses are necessarily vicious.

A more sophisticated version of a foundationalist argument accommodates the connection, noted

above, between viciousness and underlying dependence relations. The foundationalist might then
think that grounding regresses are unacceptable for all of the reasons mentioned in the discussion
of dependence. But again, we saw that most of those reasons are circular. And where they aren’t
circular, they leave the regress charged with a task that it appears not designed to carry out.

An infinite grounding regress would look as follows: x is grounded in y, y is grounded in z, and

so on ad infinitum. On the understanding of grounding presented above, we should take this to
mean something like: x depends for its existence/being on y, y depends for its existence/being on
z and so on ad infinitum. Or alternatively, x is metaphysically explained by y, y is metaphysically
explained by z, and so on ad infinitum. The question is, then, on the proposed account of
viciousness, are grounding regresses vicious or benign?

If x is grounded in y and y in z, and all that we are seeking for is an explanation of how or why x
exists (as the thing that it is), an explanation of how or why y exists, and so on ad infinitum, the
regress is benign. Why? Because where z explains y and y explains x our explanans and
explanandum are not of the same form. In order to explain facts about my existence, we can
make recourse to the existence of—or facts of the existence of—my parents, my vital organs, etc.
In order to explain these facts, we make recourse to further facts, and so on. At each stage, we
have a satisfactory explanation of that for which we are seeking one.46 The regress—or perhaps,
rather, the series—is benign

The regress is not benign, however, if what we are seeking an explanation for is how anything
exists, or has being, at all. For even at infinity, what the regress shows is that we have not
explained where existence comes from. Even at infinity, we are still invoking things that exist in
order to explain how anything exists at all. Or things with being to explain how anything has
being at all. We encounter, at each level, the explanatory failure characteristic of a vicious
infinite regress: existents whose existence we seek an explanation for are explained in terms of
existents. The existence of y may explain the existence of x but the existence of x, y, z and so on
ad infinitum cannot help us explain how anything exists at all—where being comes from.
Whether or not a regress of grounds is vicious, therefore, will depend upon the question for
which we are seeking an answer.

In this spirit, the looming infinite regress of grounds discussed by Bennett is vicious if it is
generated in an attempt to explain how anything is grounded at all: how grounding comes to be
in the first place. Where we seek to explain grounding itself nothing is achieved by invoking
grounding to this end. For again we are stuck in the same explanatory failure that characterizes a
certain kind of vicious infinite regress. I take it that this is the problem to which Bennett points.

A foundationalist argument to the existence of something fundamental that employs an argument

from vicious infinite regress is only acceptable if the foundationalist assumes that what is at issue
is an explanation of being or existence itself; the foundationalist must assume that the
fundamentalia posited in order to terminate the regress are supposed to help us explain how
anything exists at all. Otherwise, the regress is benign, and the foundationalist loses this
argument to the existence of something fundamental.47

One might wonder if something has gone awry here, for what the foundationalist wants to say is
that there is something fundamental—or a collection of fundamental things—which explain the
non-fundamental things.48 Nowhere does the foundationalist claim to be in the business of
explaining being or existence. But by invoking an argument from vicious infinite regress the
foundationalist is seeking to give us a reason to believe that there is something fundamental—
and with it a division between the fundamental and the non-fundamental—in the first place. She
cannot, therefore, set out by assuming this to be the case. She cannot set out by assuming there is
something fundamental that explains the existence of the non-fundamental. Or that there is
something non-fundamental that stands in need of explanation.

Let us briefly consider some interesting consequences of these claims. Doubtless there are many.
I mention four. The first, a restatement of the conclusion, is that grounding regresses are not
necessarily vicious. Sometimes, regresses of grounds are benign.

Second, let’s grant that some instances of grounding regresses are genuinely vicious. Somewhat
controversially, then, an argument to the existence of something fundamental that invokes the
appearance of a genuinely vicious infinite regress would appear to involve an application of a
Principle of Sufficient Reason. An argument to the effect that being or existence requires an
explanation just is an argument employing the PSR, or so it would seem. This is not the claim
that the foundationalist of this ilk wants, or means, to employ the PSR, but rather the claim that
there would appear to be no good reason to believe that there is an explanation without it.
Moreover, if this is what is driving foundationalism, then the consequences are wide-ranging.
Not only is the foundationalist committed to the PSR; a commitment which, I should imagine,
the foundationalist would shy away from. But suddenly arguments to the existence of something
fundamental start to look more like cosmological arguments to the existence of a necessary
being. This invites all sorts of interesting questions about the modal status of the fundamentalia.49
Not to mention, if, as it is commonly supposed, the fundamentalia are contingent, can they really
explain how anything exists at all? Doesn’t this explanatory task seem more like the work for
necessary existents?

To the third point of note. If this analysis of viciousness is correct, then what the appearance of a
vicious infinite regress is supposed to show us is that the proposed way of explaining some
phenomenon is incorrect. In the case of genuinely vicious grounding regresses, what their
appearance is supposed to show us is that explaining being by way of appeal to things that have
being is unsuccessful. The foundationalist is incorrect to think that halting the regress at some,
likely arbitrary, point will allow them to achieve the explanation that they are after. The solution
to the conundrum of a looming vicious infinite regress is not to terminate the regress but to
abandon the means by which we have set about explaining some phenomenon such that the
regress gets going in the first place.

Fourth and finally, there is an important distinction to be drawn between finite grounding chains
and well-founded grounding chains. A finite grounding chain is one which terminates in
something fundamental, where we can move from any member of the chain to the
fundamentalium in which it terminates in a finite number of steps. A well-founded grounding
chain is one that is grounded in something fundamental, but may, itself, be infinitely long. A
finite grounding chain is well-founded but a well-founded grounding chain need not be finite.

Grounding chains that terminate in something fundamental—whether finite or infinite—we call

maximal grounding chains. Grounding chains that do not terminate in fundamentalia, we call
non-maximal grounding chains. A strong form of foundationalism says that every grounding
chain must be maximal. That is, that every grounding chain must be well-founded. The strongest
form of foundationalism would hold that every grounding chain must be finite. But
foundationalism doesn’t require that every grounding chain is finite, nor even that every
grounding chain is well-founded. It is enough, to secure foundationalism, that there is at least one
maximal grounding chain.

Take Euclidean space. It is composed of points and regions between the points. The regions
divide infinitely into sub-regions. For illustrative purposes, let’s also assume that the regions are
grounded in their sub-regions. The points, on the other hand, are not grounded in anything and
are, therefore, fundamental. On this picture, we have maximal grounding chains—the chains that
terminate in the points—and non-maximal chains—the chains that involve the infinitely divisible
sub-regions.

Arguments from vicious infinite regress would seem to be neutral on whether they establish
finiteness or well-foundedness. What I mean by this is, whether we use an argument from
vicious infinite regress to establish finiteness or well-foundedness will depend upon other of our
assumptions. We should not assume that an argument from vicious infinite regress is necessarily
an argument for finiteness.

It seems that not only would many a foundationalist not want to argue argue for the finiteness of
every grounding chain, they may not even want to establish the well-foundedness of every
grounding chain. But if infinite grounding regresses are necessarily vicious, as the
foundationalist seems to suggest, then the application of arguments from vicious infinite regress
to the conclusion that there has to be something fundamental establishes the well-foundedness of
every grounding chain.

I, however, have argued that infinite grounding regresses are not necessarily vicious. If I am
correct, then the foundationalist need not worry about securing for themselves a foundationalism
stronger than they may want to endorse. But the foundationalist may not be out of the woods yet.

If the foundationalist wants to employ an argument from vicious infinite regress to secure the
well-foundedness of only some grounding chains—the maximal grounding chains—how can she
devise a non-question begging set of criteria for the application of those arguments? After all, the
foundationalist is trying to give us a reason to suppose that there has to be something
fundamental at all, they cannot start out by assuming it. In cases in which we believe ourselves to
have independent reason to suppose there is something fundamental—as with our space-time
points, perhaps—an argument from vicious infinite regress could be used to secure the necessity
of the fundamentality of these points.50 But it is questionable whether the foundationalist actually
has independent reasons to suppose there is something fundamental. And in their absence,
arguments from vicious infinite regress are often invoked. But if the foundationalist wants to
establish that only some grounding chains must terminate in fundamentalia, it is not at all clear
how arguments from vicious infinite regress could be applied non-question beggingly to this end.
The criterion of application cannot be that the chains to which the arguments apply must be
maximal. For whether or not there must be maximal grounding chains is exactly what the
foundationalist is aiming to establish.

5 Conclusion

I have addressed the more general question of what the viciousness of a vicious infinite regress
consists in. Contrary to common assumption, I have denied that accounts which suggest that
vicious infinite regresses are involved with contradictions, dependence relations, reductionism
and parsimony are promising. At least, they are not promising as accounts of why an infinite
regress can be considered vicious as opposed to benign. I addressed this more general question in
order to become clearer on the more specific question of whether or not infinite regresses of
grounds are vicious. This question is particularly important for the metaphysical foundationalist
who moves from assumptions about the nature of grounding regresses to a conclusion about the
fundamental structure of reality. The infinite regress generated failing the existence of something
fundamental is only vicious when it is characterized by the phenomena at each stage of the
regress being of the same form. And the phenomena at each stage of the regress are only of the
same form when the regress is generated in an attempt to answer the question of how anything
exists at all; where being comes from. Otherwise, the regress is benign.

If the foundationalist takes infinite grounding regresses to be genuinely vicious this must be
because they are seeking for an explanation of existence, or being, itself. But of course, exactly
what the appearance of the regress is supposed to show us is that the proposed way of explaining
it doesn’t work. The foundationalist needs to provide us with another means by which to explain
existence, if indeed that’s what they are after; and, ultimately, another reason to be a
foundationalist.
Acknowledgments I would like to thank audiences at the workshop The One and The Many, at
the University of Melbourne, the Melbourne Logic Group and Lingnan University. I would
particularly like to thank Aaron Guthrie, Daniel Nolan, Graham Priest, Greg Restall and
Benjamin Schnieder for comments and discussion.

1 Folks who make recourse to arguments from vicious infinite regress—or something that looks
a lot like them—in the contemporary grounding literature include Schaffer (2010, pp. 37, 63).
Cameron (2008) alsodiscusses the role arguments from vicious infinite regress might play in
arguments to the conclusion that there must be something fundamental. Historical figures of note
who endorse similar arguments include Leibniz (1989), and Bradley (1908).

2 See Brzozowski (2009, chap. 2), Orilia (2009, pp. 6–8), and Vlastos (1954, p. 328, footnote 12)
for a characterization of vicious infinite regresses in something like these terms. I have also
heard this suggestion made by philosophers on several occasions. Whilst this is no explanation of
why grounding regresses are vicious, it secures the necessity of the viciousness of infinite
grounding regresses. Although I know of nowhere that anyone has written this down.

3 Nolan (2001) and Clark (1988).

4 Aikin (2005), Clark (1988), Klein (2003), and Nolan (2001).

5 Nolan (2001).

6 Nolan (2001).

7 There are many different versions of the Bradley regress as it pertains to states of affairs and
propositions. For some discussions see (Orilia 2009), Schnieder (2004) and Vallicella (2000).

8 See Armstrong (1997, pp. 118–119).

9 Armstrong (1997, p. 118).

10 Cameron (2008, pp. 1–3) and Orilia (2009, especially Sect. 6), discuss this interpretation of
the regress.

11 Nolan (2001, pp. 524–530). Clark (1988, p .372) is more reserved about whether or not there
turns out to be a connection, but he mentions the suggestion nonetheless.

12 Vlastos (1954, p. 320).

13 Vlastos (1954, p. 321).

14 Vlastos (1954, p. 324).

15 Vlastos (1954, p. 325)

16 Nolan (2001, pp. 524–525). Italics my own.

17 That the regress is prior to the contradiction—if indeed there is a connection between the two
at all—is hinted at by Clark, who states ‘what is special about valid infinite regress arguments as
instances of species of reduction is the derivation of an infinite regress. But what is not clear is
how, if at all, the regress in turn figures in delivering the required contradiction’ (Clark 1988, p.
372.)

18 Of course, one place to start would be by looking to other theories that harbor both vicious
infinite regresses and contradictions and seeing if we can observe similar patterns.

19 Nolan (2001, pp. 528–530).

20 Clark (1988) and Nolan (2001) discuss viciousness and dependence more generally. In the
debates over foundational epistemology and arguments from vicious infinite regress, the
discussions of inferential justification relations and what is thought to be wrong with the
appearance of an infinite regress seems like a discussion of a specific instance of the more
general phenomenon of the suspected involvement of dependence relations with viciousness. For
discussions of the issue as they take place in foundational epistemology see Aikin (2005, esp pp.
194-202) and Klein (2003).

21 Leibniz (1989, p. 85). Italics my own.

22 Cameron (2008, p. 3).

23 Schaffer (2010, pp. 37, 62).

24 Nolan (2001, pp. 531–532).

25 I borrow this terminology from Clark (1988).

26 Aikin notes this same point. See Aikin (2005, p. 197).

27 The situation with epistemic foundationalists’ arguments from vicious infinite regress is
somewhat more complicated than this. But of the additional concerns expressed by the epistemic
foundationalist, there are also good reasons to be suspicious. I do not discuss these arguments
here because they are not directly pertinent to the matter to hand. For further elaboration see
especially Aikin (2005) and Klein (2003).

28 Aikin (2005, p. 197).

29 Klein (2003, p. 729) also makes this observation.

30 I think there is more to be said about the types of reductions of relevance, amongst other
things. I follow Nolan, however, and only present his very rough sketch of the kinds of reductive
theories involved.
31 Nolan (2001, p. 530) discusses the homuncular theory of perception but it is originally
introduced in Ryle (1949).

32 Bennett (2011, p. 31)

33 Nolan (2001, pp. 533–536).

34 Problems with an argument from theoretical virtue are also discussed by Aikin (2005, p. 196).

35 An example of an infinite regress that seems to be genuinely vicious but does not appear to be
accounted for in terms of explanatory failure is the Weissman regress: the regress generated
when we attempt to establish 2 as rational. See Nolan (2001, pp. 525–526) for a discussion.

36 Passmore (1961, p. 31).

37 Passmore (1961, p. 33).

38 Ryle (1949, p. 30).

39 See Passmore (1961, p. 26).

40 Passmore also points out that the regress assumes that the way in which an action is
intelligent is the same as the way in which a thought is intelligent. Where the way in which
intelligence of thought is expressed differently to intelligence of action, the regress doesn’t get
going. See Passmore (1961, p. 27).

41 Thank you to an anonymous referee for pointing out the need to address this issue.

42 These notions of sameness of form, constitution explanation and explanatory failure are going
to connect up with discussions of circularity—and its cluster of associated notions—along with
requiring a more developed account of both metaphysical explanation and grounding.

43 I do not define a notion of ground. There is disagreement in the literature, which cuts along
several different lines that I do not discuss here. I assume that whether grounding involves facts,
propositions or particulars, one can wonder whether infinite regresses of grounds are vicious or
benign. I also assume that, allowing for slight variations in detail, the reasons for which an
infinite grounding regress is vicious will be similar regardless of what category we take the relata
to belong to.

44 Although philosophers agree that grounding is involved with metaphysical explanation, they
disagree on whether grounding simply is metaphysical explanation, or whether real grounding
relations underwrite, or correspond to, metaphysical explanations. I’m sure many will also
disagree with how I am characterizing the notion of metaphysical explanation here.
45 These arguments include arguments from something like intuition (Cameron 2008; Lowe
1998, p. 158.), arguments from theoretical virtue (Cameron 2008), and arguments that employ
something like a principle of sufficient reason (Cameron 2008; Schaffer 2010).

46 Consider causal explanations. The smashing of the window by the tree is adequately
explained in terms of the storm, plus other relevant details such as the brittleness of the glass, etc.
Although one might wonder why anything exists at all, our explanation of the broken window is
not inadequate because it does not make recourse to the Big Bang or God.

47 This is not to say that the foundationalist doesn’t have other, powerful reasons in defense of
their position.

48 Thank you to an anonymous referee for pointing out the need to address this issue.

49 I am not claiming that the foundationalist is a closet theist but, rather, that there may be a
form of argument in operation here with which we are familiar. The application of which may
have some very surprising and interesting consequences.

50 Ultimately such an argument won’t work because the viciousness of the regress shows us that
the way we have set about explaining something is incorrect in the first place.
EPISTEMIC PRAGMATISM: AN ARGUMENT AGAINST MODERATION
Juan Comesaña
Res Philosophica, Vol. 90, No. 2, April 2013

Abstract: By “epistemic pragmatism” in general I will understand the claim that whether
propositions instantiate certain key epistemic properties (such as being known or being
justifiably believed) depends not just on factors traditionally recognized as epistemic, but also on
pragmatic factors, such as how costly it would be to the subject if the proposition were false. In
what follows I consider two varieties of epistemic pragmatism. According to what I shall call
moderate epistemic pragmatism, how much evidence we need in favor of a proposition in order
to know that the proposition is true depends on our preferences. According to what I shall call
extreme epistemic pragmatism, on the other hand, our preferences influence our epistemic
position at a more basic level, because they help determine how much justification we actually
have in favor of the proposition in question. Simplifying brutally, moderate epistemic
pragmatism has it that the more worried we are about a proposition’s being false, the more
justification we need in order to know it, whereas extreme epistemic pragmatism has it that the
more worried we are about a proposition’s being false, the less justification we have for it.
Recently, Fantl and McGrath have presented an interesting argument for moderate epistemic
pragmatism, an argument which relies on the principle that (roughly) knowledge is sufficient for
action (KA). In this paper I argue that KA, together with a plausible principle about second-order
evidence, entails extreme epistemic pragmatism.

1 Introduction

By “epistemic pragmatism” in general I will understand the claim that whether propositions
instantiate certain key epistemic properties (such as being known or being justifiably believed)
depends not just on factors traditionally recognized as epistemic, but also on pragmatic factors,
such as how costly it would be to the subject if the proposition were false. In what follows I
consider two varieties of epistemic pragmatism. According to a rough formulation of what I shall
call moderate epistemic pragmatism, how much evidence we need in favor of a proposition in
order to know that the proposition is true depends on our preferences.1 According to an equally
rough formulation of what I shall call extreme epistemic pragmatism, on the other hand, our
preferences influence our epistemic position at a more basic level, because they help determine
how much evidence we actually have in favor of the proposition in question. Simplifying
brutally, moderate epistemic pragmatism has it that the more worried we are about a
proposition’s being false, the more evidence we need in order to know it, whereas extreme
epistemic pragmatism has it that the more worried we are about a proposition’s being false, the
less evidence we have for it.

In sections 2 and 3, I introduce the two kinds of pragmatism to be discussed and compare them
with more traditional epistemological theories. Jeremy Fantl and Matt McGrath have recently
argued for moderate epistemic pragmatism.2 As we’ll see, Fantl and McGrath’s argument for
moderate epistemic pragmatism depends on two crucial premises: the falsity of extreme
epistemic pragmatism and a principle to the effect that (roughly) knowledge is sufficient for
action (Fantl and McGrath call this principle “KA”). In sections 4 through 6 I argue, however,
that (given a plausible principle of second-order evidence) KA entails extreme epistemic
pragmatism. Therefore, the argument for moderate epistemic pragmatism fails, for two of its
crucial premises are incompatible with each other. There is no stable intermediate position
between the traditional view that practical matters are irrelevant to epistemology and the extreme
view that practical matters determine the strength of our evidence.

2 Moving Thresholds and Shifty Contexts

One can have more or less evidence for a proposition, but either one knows it or one doesn’t
know it.3 You can have more or less evidence for the proposition that the train that you are about
to board stops in Foxboro, but either you know that it stops in Foxboro or you don’t. Although
knowledge and evidence are thus notions of different kinds (one comes in degrees, the other
doesn’t), they are connected: in order to know that the train stops in Foxboro you have to have
enough evidence for that proposition.4 How much evidence? We can ask the same question using
a slightly different terminology. Let us say that when a subject’s evidence for p is sufficient for
the truth of an attribution of knowledge (that is, if in such a situation the attribution is false that is
not because the subject lacks evidence), then that subject’s evidence for believing that p has
crossed the “knowledge threshold.” The question now is: how high is the knowledge threshold?

We can distinguish the possible answers according to whether they hold that the knowledge
threshold is fixed or mobile—giving us “fixed-threshold” and “moving-threshold” theories.
Traditional epistemological theories are fixed-threshold theories: according to them, there is a
fixed amount of evidence that is needed for the truth of a knowledge attribution in any context.
We can factor whether we have crossed the knowledge threshold for believing that p into two
components: how much evidence we have for believing that p, and how much evidence we need
for believing that p. According to fixed-threshold theories, the second factor is constant across all
cases, and thus whether we have knowledge-level evidence depends exclusively on how much
evidence we have.
Fixed-threshold theories have recently come under attack. The objection is that they simply
cannot do justice to our practices of knowledge attribution. Suppose, for instance, that Jeremy
and Matt are both about to board the train from Boston to Providence. They both wonder
whether the train stops in Foxboro or is instead the “express.” Just a moment ago both Matt and
Jeremy heard the announcement that the train is indeed the express. Matt says to himself: “Oh,
that’s right: they just announced that the train is the express, so I know that it is”, and stops
worrying about the matter—there is not much riding, for Matt, on whether the train stops in
Foxboro or not. By contrast, a lot is riding for Jeremy on whether the train stops in Foxboro or
not. If it does, then he won’t be on time for a meeting that will determine the course of the rest of
his life. Jeremy says to himself: “They just announced that the train is the express, but the
meeting is so important that I better double-check with an agent. Until I do, I do not know that
the train is the express.”5

Fixed-threshold theories entail that either Matt is wrong in attributing knowledge to himself or
Jeremy is wrong in attributing lack of knowledge to himself. Non-skeptic fixed-threshold
theories (that is, fixed-threshold theories according to which the knowledge threshold is
moderately easy to achieve) have it that Matt is right in attributing knowledge to himself and
Jeremy is wrong in attributing lack of knowledge to himself, whereas skeptic fixed-threshold
theories (that is, fixed-threshold theories according to which the knowledge threshold is almost
impossible to achieve) have the opposite consequences. But, the objection goes, what Matt and
Jeremy tell themselves sounds right in both cases: it is what we would tell ourselves if we were
in their (respective) shoes. The cases can be easily multiplied. Therefore, fixed-threshold theories
entail widespread error in our common knowledge-attributing practices. To this extent, it is
argued, they are defective.

Moving-threshold theories can avoid this consequence, for they can claim that the knowledge
threshold is different for Jeremy and Matt. Thus, even though Matt and Jeremy have the same
evidence for the proposition that the train is the express, Jeremy’s knowledge threshold is such
that this evidence is not enough for a corresponding knowledge attribution to be true, whereas
Matt’s knowledge threshold is such that that very same evidence suffices for the truth of the
corresponding knowledge attribution. One prominent moving-threshold view is contextualism.6
According to contextualism, a knowledge attribution of the form S knows that p can express
different propositions in different conversational contexts, and this variance is due to the fact that
the predicate “knows” can express different relations in different conversational contexts. These
relations differ by where they locate the knowledge threshold. Thus, there is a very lenient
knowledge relation that locates the knowledge threshold very low on the scale, and there are
stricter knowledge relations that locate the knowledge threshold higher on the scale. A subject’s
belief may satisfy some but not all of these relations. Thus, a knowledge attribution may express
a true proposition in one conversational context and a false one in another. Contextualism is,
therefore, a “variantist” theory, a theory according to which the predicate “knows” can express
different relations in different conversational contexts. Traditional fixed-threshold theories, on
the other hand, are invariantist: they hold that there is only one relation that is the semantic value
of “knows”.

Let’s consider two different conversational contexts, one where Jeremy’s concerns are somehow
made salient and one where they aren’t (let’s call the first a “high-standards” context and the
second a “low-standards” one).7 Let’s also suppose that, in each of these contexts, someone says
“Matt knows that the train is the express” and someone else says “Jeremy knows that the train is
the express.” Tables 1 and 2 represent what a contextualist theory and a traditional invariantist
(non-skeptic) theory would say about the truth-values of those utterances.

Matt knows Jeremy knows

High standards False False
Low standards True True

TABLE 1. Contextualism

Matt knows Jeremy knows

High standards True True
Low standards True True

TABLE 2. Traditional Invariantism

Contextualism, then, is a moving-threshold theory because it is a variantist theory, and

traditional theories have been both fixed-threshold theories and invariantist. This may suggest
that a moving-threshold theory must be variantist. But this is not the case: there is logical space
for invariantist moving-threshold theories. According to this kind of theories, the semantic value
of “knows” is invariant across contexts (this is the invariantist part), but the amount of evidence
a subject must have in order for a knowledge attribution to be true may vary from case to case.
Table 3 represents that kind of position.

Matt knows Jeremy knows

High standards True False
Low standards True False

TABLE 3. Moving-Threshold Invariantism

A theory represented by Table 3 is invariantist because which proposition is expressed by a

knowledge attribution doesn’t vary with the context of utterance, as shown by the fact that the
rows in the table have the same truth-values. But a theory represented by Table 3 is a moving-
threshold theory because subjects who have the same evidence for some proposition may differ
in whether they know the proposition, as shown by the fact that the columns in the table have
different truth-values. So, it is possible for there to be moving-threshold invariantist theories.
Indeed, one of the versions of epistemic pragmatism that I discuss below is just such a theory.8

3 Moderate and Extreme Epistemic Pragmatism

Like knowledge, practical rationality is all-or-nothing. Either you are rational to act as if the train
is the express or you are not.9 Like knowledge, too, practical rationality is connected to evidence:
in general, in order for you to be rational to act as if the train is the express you have to have
some evidence that it is.10 How much evidence? In this case, we have a well-developed answer:
as much as required by your preferences. Matt’s practical concerns are such that he has enough
evidence to make it rational for him to act as if the train is the express. Jeremy’s practical
concerns, however, are such that his evidence (which is, remember, the same as Matt’s) is not
enough to make it rational for him to act as if the train is the express.

So there are two different binary notions, knowledge and practical rationality, both of which are
connected to the same degree notion, evidence. We have a relatively clear understanding of how
practical rationality is connected to evidence, and no clear understanding of how knowledge is
connected to evidence. Why not explain the unclear by the clear, at least partially? Why not say
that, in order to know that p, the amount of evidence for p that you have has to be at least high
enough to make it rational for you to behave as if p?11 That thought can be captured in the
following principle, which Fantl and McGrath (2002) call “KA”:

KA: For any possible subject S, if S knows that p, then S is rational to act as if p.12

Fantl and McGrath don’t take KA to be obviously true. They present at least two arguments for
it. First, they point out that, when defending courses of action, we routinely appeal to what we
know. Thus, for instance, if I plan to go swimming in the ocean at night with my glasses on and
my wife asks me why, I can sensibly reply “Don’t worry, I know that I won’t lose them.” This
practice of ours of appealing to what we know in defense of what we and others do makes sense
under the assumption that KA is true (and that we are, perhaps implicitly, aware of its truth).
Second, Fantl and McGrath (2007) also argue that a plausible closure inference entails KA with
the aid of the following knowledge-to-action principle (which they call “KB”):

KB: If S knows that A is the best thing for him to do in light of all of his goals, then S is
rational to do A.13

KB really is appealing.14 To deny it one would have to hold, for example, that even though
Jeremy knows that boarding the train straightaway is the best thing for him to do in light of all of
his goals, he nevertheless isn’t rational to board the train straightaway. It is hard to see how to
make sense of that claim.15

Now, most epistemologists nowadays will accept that knowledge is fallible, at least in the sense
described by the following principle:

Fallibilism: It is possible for a subject S to know that p even if there is a subject S’ such
that S’ has more evidence for p than S has.16

KA and fallibilism together entail the following principle, which I will take as the official
definition of moderate epistemic pragmatism:

Moderate epistemic pragmatism: There are possible subjects S and S’ and there is a
proposition p such that (i) S and S’ have the same evidence for p, but (ii) S is in a position
to know that p whereas S’ is not.
What is it to be in a position to know a certain proposition? It is for the amount of evidence for
that proposition to be high enough for knowledge—that is, it is for you to have knowledge-level
evidence.

Under the assumption that knowledge entails justification (and that evidence is what justifies),
KA and fallibilism entail that the knowledge threshold depends on your preference structure. If
there isn’t much riding, for instance, on whether there is a glass of water on the table, then KA
doesn’t impose any very stringent requirement on what it takes for you to know that there is a
glass of water on the table—the knowledge-threshold is set, in accordance with fallibilism, at a
level where a certain non-maximal amount of evidence is needed. But suppose that the issue of
whether there is water or gasoline in the glass is of vital importance to you, because you are
extremely thirsty. In that case, KA entails that, in order to reach the knowledge threshold, you
must have enough evidence for thinking that there is water in the glass. How much evidence? As
much as it is necessary to make it rational for you to act as if there is water in the glass—in the
imagined case, as much as it is necessary to make it rational for you to drink the liquid in the
glass. Crucially, this may mean that you need more evidence to know than you would have
needed if the stakes had been lower. Moderate epistemic pragmatism is thus an invariantist,
variable-threshold theory, a theory of the kind represented by Table 3 above.

We can further illustrate the argument from KA and fallibilism to moderate epistemic
pragmatism with our familiar case: Matt knows that the train is the express, Jeremy has the same
evidence that Matt has for believing that the train is the express, but (because there is a lot riding
on it for him) Jeremy is not rational to straightaway board the train. If we rejected fallibilism we
could hold that Jeremy’s situation is not possible, because he has the same evidence as Matt and
Matt knows that the train is the express. Given the negation of fallibilism, this evidence must be
maximal—it is impossible to have more evidence, in which case it is arguably impossible for it
to be irrational for you to act as if the train is the express. But, given fallibilism, the argument is
now the following:

(1) Matt knows that the train is the express.

(2) Jeremy is not rational to act as if the train is the express.

(3) Jeremy and Matt have the same evidence for believing that the train is the express.

(4) Jeremy is not in a position to know that the train is the express. (From 2 and KA)

(5) Moderate epistemic pragmatism is true. (From 1, 3, and 4).

The argument is valid. The step from 2 and KA to 4 requires some comment. All that
immediately follows from 2 and KA is that Jeremy doesn’t know that the train is the express.
But, as the description of the case makes clear, Jeremy’s lack of knowledge is not due to lack of
belief, falsity of the proposition believed, or failure to satisfy any other of the traditional
conditions on knowledge. Therefore, Fantl and McGrath conclude, it must be due to the fact that
Jeremy is not in a position to know. Although this move gives rise to interesting questions, I
won’t object to the move from 2 and KA to 4. I won’t deny premises 1 and 2 either—that much
can be seen as merely a stipulation about what case we are thinking of. Premise 3 will be the
focus of the rest of this paper. I will make some preliminary comments on premise 3 after
introducing extreme epistemic pragmatism.

Moderate epistemic pragmatism is of course compatible with the claim that how much evidence
you actually have for the proposition, for instance, that there is water in the glass, is independent
of your preferences. That is, someone who accepts moderate epistemic pragmatism is free to
make the following claims about the case: your evidence for thinking that there is water in the
glass is the same whether or not it is vitally important to you that there be water in the glass, but
whether that preference-independent amount of evidence reaches the knowledge threshold does
depend on how important it is to you that there be water in the glass.

Extreme epistemic pragmatism, on the other hand, holds that your preferences help determine
how much evidence you have for a given proposition. For instance, the extreme epistemic
pragmatist will say that if it is vitally important to you whether there is water in the glass, then
you thereby have less evidence for thinking that there is water in the glass than someone who
doesn’t care much one way or the other. More generally:

Extreme epistemic pragmatism: For any possible subjects S and S’ and proposition p, if S
and S’ are exactly alike in all non-epistemic matters except for the fact that S’ is not
rational in acting as if p (whereas S is so rational), then S has more evidence for p than S’
does.

Given the very plausible claim that epistemic properties supervene on non-epistemic properties
(there cannot be a difference in epistemic properties without a difference in non-epistemic
properties), the antecedent of extreme epistemic pragmatism guarantees that if there is an
epistemic difference between S and S’ that difference will be due to the fact that there is a
practical difference between S and S’, and its consequent says that there is indeed such a
difference. We must be careful not to set up matters so that extreme epistemic pragmatism comes
out as incoherent. If we stipulate that how much evidence a subject has is a purely epistemic
matter, in the sense that it depends only on truth-related features of the subject’s situation and
not, for instance, on the subject’s practical situation, then we make extreme epistemic
pragmatism false by stipulation.17 But the difference between moderate and extreme epistemic
pragmatism resides precisely on whether there is such a thing as a purely epistemic domain of
evaluation. No stipulation will make this question go away.

Where does extreme epistemic pragmatism fit in with respect to the pair of distinctions
introduced above—the distinction between fixed-threshold and moving-threshold theories and
the distinction between variantist and invariantist theories? Even under the assumption that
knowing that p entails having evidence for p, extreme epistemic pragmatism, as I have defined it,
is compatible with any of the four possible kinds of theories resulting from the combination of
those two distinctions. A theory that subscribes to extreme epistemic pragmatism can be
variantist or invariantist, and it can posit a moving or a fixed-threshold. Unlike contextualism,
extreme epistemic pragmatism is not primarily a theory about knowledge attributions, and unlike
moderate epistemic pragmatism, it is not a theory about where the knowledge threshold lies.
Extreme epistemic pragmatism is a theory about how practical matters affect evidence. Partly
because of that reason, none of the three tables introduced above can unproblematically describe
extreme epistemic pragmatism. According to extreme epistemic pragmatism, the cases that those
tables aim to represent are not possible. Remember that part of the set-up of the case involving
Matt and Jeremy is that, despite the differences in their practical situation, they have the same
evidence—and that is precisely what the extreme epistemic pragmatist denies. If we concentrate
our attention on invariantist theories that adhere to extreme epistemic pragmatism and we don’t
make the assumption of evidential equality, supposing only that Matt’s and Jeremy’s practical
positions are different, then extreme epistemic pragmatism is compatible with the truthvalues
given in Table 3. But notice that, in this case, if moderate and extreme epistemic pragmatism
give the same answers it may be that they do so for very different reasons: moderate epistemic
pragmatism will be represented by Table 3 because it holds that even though Matt and Jeremy
have the same evidence, Jeremy needs more evidence than Matt in order to know, whereas a
theory that adheres to extreme epistemic pragmatism may be represented by that table for the
same reason, but it also may be represented by that table merely because it holds that Matt has
more evidence than Jeremy.18 The extreme epistemic pragmatist need not say, however, that Matt
has knowledge and Jeremy doesn’t. The extreme epistemic pragmatist is only committed to
saying that the practical situation affects, by itself, how much evidence subjects have, but is free
to remain neutral on the question whether there are cases such that the subject doesn’t know only
because his practical situation made it the case that his evidence didn’t reach knowledge-level.
Thus, extreme epistemic pragmatism doesn’t entail moderate epistemic pragmatism. The relative
extremity of extreme epistemic pragmatism doesn’t reside on the fact that it is logically stronger
than moderate epistemic pragmatism, but rather on the fact that it claims that our practical
situation can by itself affect how much evidence we have, and not merely how much evidence we
need to have in order to know.

Let us re-examine Fantl and McGrath’s argument for moderate epistemic pragmatism in light of
the distinction between it and extreme epistemic pragmatism. Fantl and McGrath call the denial
of moderate epistemic pragmatism (roughly, the claim that if two subjects S and S’ have the
same evidence then S is in a position to know that p if and only if S’ is), “Purism,” because it
allegedly defines a dimension of pure epistemic evaluation, free from any taint of practical
matters. But notice that Purism deserves its name only under the assumption that extreme
epistemic pragmatism is false. For suppose that extreme epistemic pragmatism was true. Then,
Purism might well be true even if practical matters do influence whether you are in a position to
know. In other words, if extreme epistemic pragmatism is true, then practical matters influence
directly how much evidence we have, and so the purity ensured by the truth of Purism is only
apparent. Moreover, premise 3 in Fantl and McGrath’s argument for moderate epistemic
pragmatism presupposes the falsity of extreme epistemic pragmatism.19 But I will argue that KA
entails extreme epistemic pragmatism given a plausible assumption. Therefore, if Fantl and
McGrath are right about KA, then they are wrong about extreme epistemic pragmatism.

4 A Principle of Second-Order Evidence

Evidence of evidence is evidence. If you have evidence that scientists have evidence for the
proposition that there is life outside the solar system, then you thereby have evidence for the
proposition that there is life outside the solar system. If someone is just like you evidentially
except for the fact that she lacks that second-order evidence, then you have more evidence than
she does for the proposition that there is life outside the solar system. A fortiori, if you have
evidence that scientists know that there is life outside the solar system, then you thereby have
evidence for the proposition that there is life outside the solar system. If someone is just like you
evidentially except for the fact that she lacks that second-order evidence, then you have more
evidence than she does for the proposition that there is life outside the solar system. Second-
order evidence need not be about somebody else: the same principles hold if you have evidence
that you have evidence (or that you know) that there is life outside the solar system. Again, if
someone is just like you evidentially except for the fact that she lacks this second-order evidence
about herself, then you have more evidence than she does for the proposition that there is life
outside the solar system.

These considerations motivate the following principle of second-order evidence:

SOE: For any possible subjects S and S’, necessarily,

IF

(i) S has all the evidence for p that S’ has;

(ii) S’ has all the evidence against p that S has; and

(iii) S has more evidence for his belief that p amounts to knowledge than S’ does for her
belief that p amounts to knowledge,

THEN

S has more evidence for p than S’ does.

Let us call any pair of subjects who satisfy clauses (i) and (ii) of the antecedent of SOE,
“evidential cousins.” It is important that SOE contains this restriction to evidential cousins, for it
would otherwise be very implausible. Suppose that Joe has just learned about Goldbach’s
conjecture, and mistakenly but justifiedly thinks that there is a proof of it. Moe, on the other
hand, is a mathematician who has spent his career working on Goldbach’s conjecture. Moe has
evidence that he doesn’t know that Goldbach’s conjecture is true that Joe lacks, but we still want
to say that Moe has more evidence for Goldbach’s conjecture than Joe does. In virtue of having
more evidence than Joe for thinking that he doesn’t know that p, Moe does have evidence against
p that Joe lacks, but that additional evidence is outweighed by his additional independent
evidence for p. In this case, Joe and Moe are not epistemic cousins because Moe has evidence for
p that Joe lacks. A similar case could be constructed that exploits a violation of clause (ii).

But if the pair of subjects in question are epistemic cousins, then SOE is indeed very plausible. I
will not attempt here a full defense of SOE,20 but I will point out the following: even if, as stated,
SOE is false, that is not enough to let the friend of (merely) moderate epistemic pragmatism off
the hook. There may well be counterexamples to SOE, counterexamples which may necessitate a
fair amount of Chisholming to get a correct principle on the table. It may even be the case that it
is impossible to formulate a true SOE-like principle in any neat way (although I don’t believe
that it is). But the use of SOE in the next section is in the context of a very specific kind of case:
a case where the only difference between the two subjects in question is in their practical
situation. Thus, what the friend of KA needs to show in order to block that argument is not just
that SOE is false, and not even that no true SOE-like principle is formulable, but rather that the
basic idea that second order evidence that there is evidence for p is evidence for p fails in the
kind of cases to which the argument applies it. I come back to this issue below, in section 6.

5 The Argument Against Moderation

I will now argue that SOE and KA together entail extreme epistemic pragmatism. Therefore,
Fantl and McGrath’s argument for moderate epistemic pragmatism cannot be sound: if they are
right about KA, then the third premise of that argument (that the two agents have the same
evidence) is false. I will first present the argument and then comment on it:

(1) KA: For any possible subject S, if S knows that p, then S is rational to act as if p.

(2) SOE: For any possible subjects S and S’, necessarily, if (i) S has all the evidence for p
that S’ has; (ii) S’ has all the evidence against p that S has; and (iii) S has more evidence
for his belief that p amounts to knowledge than S’ does for her belief that p amounts to
knowledge, then S has more evidence for p than S’ does.

(3) If S is not rational to act as if p, then S has evidence that he is not rational to act as if
p. [Premise]

(4) If KA is true and S has evidence that he is not rational to act as if p, then S has
evidence that he doesn’t know that p. [Premise]

(5) Matt and Jeremy are exactly alike in all non-epistemic matters except for the fact that
Jeremy is not rational to act as if the train is the express (whereas Matt is so rational).
[Assumption]

(6) Jeremy has evidence that he is not rational to act as if the train is the express. [From 3
and 5]

(7) Jeremy has evidence that he doesn’t know that the train is the express. [From 4 and 6]

(8) Matt has all the evidence for the proposition that the train is the express that Jeremy
has; Jeremy has all the evidence against the proposition that the train is the express that
Matt has; and Matt has more evidence for his belief that the proposition that the train is
the express amounts to knowledge than Jeremy does for his belief that the proposition
that the train is the express amounts to knowledge. [From 5 and 7]

(9) Matt has more evidence for the proposition that the train is the express than Jeremy
does. [From 2 and 8]
 (10) Extreme epistemic pragmatism: For any possible subjects S and S’ and proposition p,
if S and S’ are exactly alike in all non-epistemic matters except for the fact that S’ is not
rational in acting as if p (whereas S is so rational), then S has more evidence for p than S’
does. [From 5 through 9]

The only premises in the argument are KA, SOE, and 3 and 4. Why believe in 3, the claim that if
S is not rational to act as if p, then S has evidence that he is not so rational? Remember that we
are assuming that what is rational for you to do supervenes on your preferences and your beliefs.
It doesn’t take any strong thesis of privileged access to conclude that whenever it is irrational for
you to act as if p you will have at least some evidence that it is.

What about 4, the claim that (under the assumption that KA is true) someone who has evidence
that he is not rational in acting as if p thereby has evidence that he doesn’t know that p? Here it is
important to recall the arguments for KA. One of them was that we (at least implicitly) rely on
the truth of KA when justifying and criticizing actions. The second argument for KA relied on
KB, and Fantl and McGrath’s defense of KB appeals both to its obvious truth as well as to its
role in the criticism and justification of action. Thus, Fantl and McGrath’s own arguments for
KA strongly involve the claim that subjects are (perhaps implicitly) aware of the truth of KA.
The contrapositive of KA says that if S is not rational in acting as if p, then S doesn’t know that
p. Premise 4, therefore, amounts to saying that (given that S is at least implicitly aware of the
truth of that conditional) if S has evidence for the antecedent of that conditional, then S has
evidence for its consequent.

This might seem like a dangerous claim, given the fact that “is evidence for” is not closed under
known logical implication.21 But premise 4 relies on no general principle of evidential closure.
That evidence fails to be closed is not a good argument for the claim that, in a particular case,
evidence for p is not evidence for a q entailed by p. The cases where evidence fails to be closed
all share a common structure, a structure which we have no reason to think is present in the cases
to which the foregoing argument applies. The structure in common to all cases where evidence
fails to be closed is most easily revealed by thinking in probabilistic terms (although one need
not think of evidence probabilistically to appreciate the structure). Suppose that e is evidence for
p for S only if e raises S’s epistemic probability for p. If p entails q, then the probability of q is at
least as high as the probability of p. Therefore, e can raise S’s epistemic probability for p but not
raise S’s epistemic probability for q (that is, e will fail to be closed under entailment) only if,
before receiving e, S’s epistemic probability for q was higher than S’s epistemic probability for
p. But there is absolutely no reason to think that, in the cases that interest us, this will be the case.
That is, there is absolutely no reason to think that, before realizing that he is not rational to act as
if p, S had more evidence for the proposition that he didn’t know that p than for the proposition
that he was not rational to act as if p. Therefore, given that there is no reason to think that the
cases to which 4 applies share the structure of cases where evidence fails to be closed, the initial
plausibility of 4 is not threatened by those considerations.

We could, if we so wished, replace premises 3 and 4 in the argument by a single premise: that if
S fails to know that p in virtue of the fact that S’s evidence for p doesn’t meet the knowledge-
threshold, then S has evidence that he doesn’t know that p. The general principle that if S doesn’t
know that p then S knows that he doesn’t know that p is of course false, but no such general
principle is advocated in the proposed new premise. Rather, the idea is that in the particular case
where the lack of knowledge is due to insufficient evidence, the subject has evidence that he
doesn’t know.22 Given that the subject S’ in question in the argument lacks knowledge precisely
for this reason, we can get from this new premise to step 7 and continue as before.

Step 8 also deserves some comment. We can consider that step in the argument as the
conjunction of two propositions: that Matt and Jeremy are evidential cousins, and that Matt has
more evidence than Jeremy does for the proposition that his belief that the train is the express
amounts to knowledge. Notice, first, that Fantl and McGrath themselves will agree with the
claim that, in the case described, Matt and Jeremy are epistemic cousins—indeed, they think that
Matt and Jeremy have the same evidence for the proposition that the train is the express, from
where it follows that they are epistemic cousins with respect to that proposition. But the claim
also follows from 5: given the supervenience of the epistemic on the non-epistemic, and given
that the only non-epistemic difference between the subjects is that only Matt is rational to act as
if the train is the express, it follows that they are epistemic cousins—for no party to this dispute
(neither the moderate nor the extreme epistemic pragmatist, nor, obviously, the epistemic purist)
thinks that the fact that S is not rational to act as if p means that he has more evidence for p than
another subject who is rational in so acting. For analogous reasons, it follows also from 5 and 7
that the second proposition in 8 is true.

Remember that Fantl and McGrath’s argument from KA to moderate epistemic pragmatism had
as a premise the denial of extreme epistemic pragmatism (see section 3). The argument of this
section shows that KA itself (together with SOE) entails extreme epistemic pragmatism.
Therefore, Fantl and McGrath’s argument for moderate epistemic pragmatism fails.

6 Evidential and Extra-Evidential Conditions on Knowledge

I have just argued that, given SOE, KA entails extreme epistemic pragmatism. I don’t presume
that, as it stands, SOE is free from counterexamples, but I do think that whatever Chisholming
SOE needs won’t affect its use in showing that KA entails extreme epistemic pragmatism. 23 But
there is a worry about SOE which does not consist in claiming that it gets this or that isolated
case wrong, but is rather the objection that it ignores a crucial distinction between evidential and
extra-evidential conditions knowledge. SOE, the objection goes, should apply only to evidential
conditions, and being rational to act as if p is an extra-evidential condition. In this section I
consider three proposals as to how to draw the distinction between evidential and extra-
evidential conditions on knowledge. I find specific problems with appealing to each of them to
argue against my use of SOE. But I close the section by arguing more generally that, whatever
the distinction between evidential and extra-evidential conditions on knowledge comes down to,
SOE should ignore it. SOE applies to cases where subjects receive evidence that they don’t know
that p, not to cases where subjects receive evidence that they fail to satisfy some condition on
knowledge. Even when the former kind of evidence is constituted by the latter, their evidential
impact is still different. Therefore, the distinction between evidential and extra-evidential
conditions on knowledge is irrelevant to SOE.

Let me start by trying to flesh out in intuitive terms what the objection is. According to SOE,
when S has more evidence for the proposition that he knows that p than S’ does for the
corresponding proposition and S and S’ are evidential cousins (which means that S has all the
evidence for p that S’ has and S’ has all the evidence against p that S has), S thereby has more
evidence for p than S’ does. But, the objection goes, there are two ways for a subject to have
more evidence than another subject for the proposition that he has knowledge. First, it may be
that S has evidence that he has more evidence that p than S’ does. But, second, it may be that S
has more evidence than S’ that he satisfies some extra-evidential condition on knowledge. If this
second case obtains, the objection continues, then S and S’ have the same evidence for p. This is
so because one can come to lose some evidence for believing that p in one of two ways: one may
acquire some evidence against p, or one may acquire some evidence that one’s antecedent
reasons for believing that p were not good enough (this is the distinction between “rebutting” and
“undercutting” defeaters—see Pollock [1986]). But learning that one fails to satisfy some extra-
evidential condition for knowing that p is not learning that there is some evidence against p or
that one’s reasons for believing that p are not good. Therefore, the mere fact that S has more
evidence than S’ for the proposition that he knows that p doesn’t entail that S has more evidence
than S’ for p—not even when S and S’ are evidential cousins. Moreover, the argument continues,
that one is rational in acting as if p is clearly an extra-evidential condition on knowing that p.
Therefore, the argument of the preceding section fails, because Matt has more evidence than
Jeremy for the proposition that the train is the express only because Matt’s and Jeremy’s
preferences are such that Matt (but not Jeremy) is rational to act as if the train is the express, and
so this extra evidence doesn’t entail any extra evidence for believing that the train is the express
.
The objection relies crucially on the distinction between evidential and extra-evidential
conditions on knowledge. It is not obvious that the distinction can be made clear without
prejudicing the case against SOE. I will consider three proposals about how to distinguish
between evidential and extra-evidential conditions on knowledge and argue that none of them
can be used against SOE.

First, the most straightforward way of understanding the distinction is as follows: something is
an evidential condition on knowledge if and only if it is a condition on knowing that p having to
do with how much evidence a subject needs to know that p. But on this understanding it is
simply not true that being rational to act as if p is an extra-evidential condition, for a crucial
determining factor of whether one is rational in acting as if p is how much evidence for p one
has.

Second, we may want to try a supervenience thesis: something is an extra-evidential condition on

knowledge just in case it doesn’t supervene on evidence. In other words, if a condition is such
that a subject may satisfy it even though another possible subject who has exactly the same
evidence does not, then the condition is extra-evidential. Under this understanding, being rational
to act as if p is an extra-evidential condition on knowledge. The question now is whether extra-
evidential conditions, so understood, should be excluded from the scope of SOE. In other words,
the question is whether having evidence regarding the satisfaction of an extra-evidential
condition (under this conception of the distinction) can make a difference as to the strength of the
subject’s evidence with respect to his first-order beliefs.

According to this second characterization of the distinction between evidential and extra-
evidential conditions, the anti-Gettier conditions count as extra-evidential. Gettier (1963) showed
that justified true belief is not sufficient for knowledge. One way of putting Gettier’s result is by
saying that what makes the difference between knowledge and true belief can be extra-
evidential: that is, two subjects can have the same evidence for p, while only one of them knows
that p, even if they both truly believe that p.24 Whatever other conditions are necessary for
knowledge we can call “anti-Gettier” conditions. We can then ask whether having evidence
about the satisfaction of an anti-Gettier condition can make a difference to the strength of one’s
evidence for first-order beliefs. That is, if two subjects differ only in that one of them has
evidence that an anti-Gettier condition is not satisfied for him to know that p, do those subjects
have the same evidence for p? Fortunately for our use of SOE in the previous section, the answer
is no.

Let us first recall one traditional Gettier case: Jones is justified in believing that Nogot owns a
Ford (even though Nogot doesn’t own a Ford), and infers from that belief and his knowledge that
Nogot works in his office that someone in his office owns a Ford (which is true, because Havitt
owns a Ford). Let us now suppose that Smith (who also works in the same office as Jones) is in
the same evidential position as Jones: Smith also believes that Nogot owns a Ford, and infers
from that belief and his knowledge that Nogot works in his office that someone in his office
owns a Ford.

Now, what would happen if one (but only one) of our subjects were to become aware that he is in
a Gettier situation? Suppose, for instance, that someone whom Jones trusts tells him “Your belief
that someone in the office owns a Ford constitutes a Gettier case.” Interestingly, even though
what Jones’s friend tells him was true before he told it to him, his telling him makes it false. For
what his friend is telling Jones is that his belief that someone in his office owns a Ford is
justified and true but doesn’t amount to knowledge. But if Jones is to trust his friend that his
belief is true, he cannot trust him that it doesn’t amount to knowledge. Perhaps, then, Jones
acquires information that his original basis for believing that someone in his office owns a Ford
was faulty, but in the same breath he acquires a new basis for believing that someone in his
office owns a Ford (a basis which is good enough for knowledge). That a belief of a subject
constitutes a Gettier case is something that cannot be learned by that subject (although a subject
can learn that a belief of his used to constitute a Gettier case, and somebody else can learn that
my present belief constitutes a Gettier case).25

But Jones can have evidence that his belief doesn’t satisfy some anti-Gettier condition without
thereby having evidence that his belief constitutes a Gettier case. For instance, Jones’s friend
could tell him: “I won’t tell you whether it is true or not that someone in your office owns a
Ford, but your belief that someone does is such that even if it were true you wouldn’t know it”
(or perhaps he could tell Jones “Your belief that someone in your office owns a Ford doesn’t
satisfy some anti-Gettier condition on knowledge”, after explaining to him what an anti-Gettier
condition is). Thus, it is possible for a subject to acquire evidence that his belief doesn’t satisfy
some of the extra-evidential conditions required for knowledge. But this doesn’t constitute a
counterexample to SOE. For, after hearing his trustworthy friend, Jones is clearly less justified
than he was before in believing that someone in his office owns a Ford.

It is, perhaps, surprising that this should be so, but learning that one’s belief fails to satisfy an
anti-Gettier condition on knowledge robs one of justification for that belief. Given that, together
with Fantl and McGrath, we are assuming that how much justification one has supervenes on
one’s evidence, we are led to say that after learning that one’s belief doesn’t satisfy an anti-
Gettier condition one is thereby less justified in that belief because one now has less evidence for
it. This entails embracing the somewhat paradoxical-sounding (but not, of course, really
paradoxical) claim that learning about some extra-evidential condition can by itself affect how
much evidence one has. Therefore, Gettier cases do not represent a counterexample to SOE.

Given that learning that one’s belief fails to satisfy some extra-evidential condition related to the
Gettier problem does affect one’s evidence for the first-order belief, the objection to the
argument of the preceding section based on the idea that SOE fails in precisely this kind of case
is flawed. Of course, it may well be that there are some extra-evidential conditions for which
SOE fails, but the mere fact that the pragmatic condition on knowledge advocated by moderate
epistemic pragmatist is extra-evidential (according to the characterization of the distinction that
we are considering) is not by itself enough to show that SOE fails when applied to it.

Finally, consider a third proposal as to how to draw the distinction between evidential and extra-
evidential conditions on knowledge. Let us say that an alleged condition on knowledge is
evidential with respect to p if and only if it is such that knowledge of it can make a difference to
how strong one’s evidence for p is. The anti-Gettier conditions qualify as evidential under this
definition, for, as we saw, knowledge of them can make a difference as to how strong one’s
evidence is. But, the objection goes, being rational to act as if p is not evidential in this sense, for
knowledge of it doesn’t make a difference to how strong one’s evidence for p is. Given that we
should exclude from the scope of SOE those conditions that fail to be evidential in this sense, we
should exclude from the scope of SOE the condition imposed by KA.

There is at least an air of question-beggingness to this proposal. To be clear, what I suspect is

question-begging is not the characterization of the distinction between evidential and extra-
evidential conditions itself, but rather the claim that being rational to act as if p is extra-
evidential according to this characterization. Isn’t that precisely what is at stake between friends
and enemies of (merely) moderate epistemic pragmatism?26 I don’t want to make much of this
claim of question-beggingness, however, for I think that the proposal is flawed for deeper
reasons. In fact, the flaw in the present proposal applies in general to all the objections to the
effect that the scope of SOE should exclude extra-evidential conditions.

SOE doesn’t say that evidence that one fails to satisfy some condition or other on knowledge that
p makes a difference to the strength of one’s evidence for p. Rather, it says (approximately) that
evidence that one fails to know that p makes a difference to the strength of one’s evidence for p.
All the proposals to restrict SOE’s scope to evidential conditions would have SOE look into the
structure of the subject’s evidence that he fails to know that p. If the subject’s evidence that he
doesn’t know that p is constituted by evidence that he fails to satisfy some extra-evidential
condition then, the objection goes, we should exclude it from the scope of SOE. In other words,
all the proposals considered so far would have SOE say (approximately) that evidence that one
doesn’t know that p makes a difference to the strength of one’s evidence for p only when the
evidence that one doesn’t know that p is constituted by evidence that one fails to satisfy some
evidential condition on knowledge.
The proposed reformulations of SOE are all unwarranted, however. Some epistemologists think
that the project of coming up with illuminating conditions on knowledge is unfounded to begin
with (see Williamson [2000]). But we don’t need to go that far to agree that we just don’t know
which are the conditions on knowledge. A fortiori, we don’t know which of the conditions on
knowledge are evidential and which ones are non-evidential. But we are nevertheless many times
sensitive to whether we know or not. There will be, therefore, cases where we have evidence that
we fail to know some proposition without thereby having evidence about the nature (evidential
or extra-evidential) or even the identity of the condition on knowledge that we fail to satisfy.
Suppose that, nevertheless and unbeknownst to us, the evidence that we fail to know is
constituted by what is in fact evidence that we fail to satisfy some extra-evidential condition on
knowledge. The proposals considered in this section would exclude these cases from the scope of
SOE. But that is the wrong result. Whether evidence that I don’t know that p should make a
difference to the strength of my evidence for p cannot depend on features of the evidence that I
am unaware of. If all that I register is that there is evidence that I don’t know that p, then that
makes a difference to the strength of my evidence for p, even if the evidence that I don’t know
that p is in fact evidence that I fail to satisfy some extra-evidential condition on knowledge.

Of course, it may happen that I do know that the evidence that I don’t know that p is constituted
by evidence that I fail to satisfy extra-evidential conditions on knowledge.27 Shouldn’t SOE fail
to apply in that case? I grant that if I know that the evidence that I don’t know that p is
constituted by evidence that I fail to satisfy some extra-evidential condition on knowledge, then
that shouldn’t make a difference to the strength of my evidence for p. But notice that, in this
case, I would have two pieces of evidence: that I fail to know that p, and that I fail to satisfy
some extra-evidential condition on knowledge.28 We are then free to say the following: the
evidence that I don’t know that p still makes a difference to the strength of my evidence for p,
but this difference is neutralized by the evidence that I fail to satisfy an extra-evidential condition
on knowledge.

Be that as it may, it doesn’t really matter what we should say about SOE in cases where the
subject is aware that the evidence that he doesn’t know that p is constituted by evidence that he
fails to satisfy some extra-evidential condition on knowledge. For it will not be the case that, in
realizing that he has some evidence that he doesn’t know that the train is the express, Jeremy also
realizes that this is so because he fails to satisfy some extra-evidential condition on knowledge. It
may be that Jeremy realizes that the evidence that he doesn’t know is grounded in the fact that he
is not rational to act as if the train is the express, without him realizing that this is an extra-
evidential condition on knowledge.29

To summarize: I have been considering the idea that we should distinguish between evidential
and extra-evidential conditions on knowledge, and that we should restrict SOE to apply only to
evidential conditions. That idea fails to take into account the fact that SOE applies not to
evidence about conditions on knowledge, but to evidence about knowledge itself. Moreover,
even if we did have to restrict SOE so as not to apply it to cases where the subject knows that he
fails to satisfy an extra-evidential condition, this is not what happens to the subjects involved in
my argument from the previous section.

7 Consequences of Impurity
If Fantl and McGrath are right, then KA is true. If I am right, then if KA is true, extreme
epistemic pragmatism is true as well. So, if Fantl, McGrath and I are right, extreme epistemic
pragmatism is true. But extreme epistemic pragmatism has some strange consequences.

Suppose that the chief of the train station announces that he will decide whether the train will
stop in Foxboro or not by flipping a coin: heads, the train stops in Foxboro, tails, it doesn’t. Let
us stipulate that the details of the case are such that Matt and Jeremy both have good evidence to
believe the chief, and therefore also good evidence for the proposition that the train stops in
Foxboro if and only if the coin landed heads. Because Matt doesn’t much care whether the train
stops in Foxboro or not, and given the evidence he has, he has more or less the same evidence for
the proposition that it does stop as for the proposition that it doesn’t (and this reflects the fact that
he has more or less the same evidence for the proposition that the coin landed heads as for the
proposition that it landed tails). Now, because it is vitally important to Jeremy whether the train
stops in Foxboro, extreme epistemic pragmatism has it that he has less evidence for the
proposition that the train is the express than Matt does. So far, this is a direct consequence,
however unpalatable, of extreme epistemic pragmatism. But remember that Jeremy has good
evidence for the proposition that the train stops in Foxboro if and only if the coin landed heads.
Therefore, extreme epistemic pragmatism has the consequence that Jeremy has more evidence
for the proposition that the coin landed tails than for the proposition that it landed heads.

Many people will find this consequence unacceptable: merely being worried about whether a
train stops in Foxboro or not cannot affect your evidence for thinking that a fair coin landed
heads. I sympathize with this purist thought. If the consequence is sufficiently bad, then it might
form the basis for a reductio of whatever entailed it—in this case, the conjunction of KA and
SOE (I assume that the other propositions needed for the entailment are beyond reproach). So,
couldn’t the friend of KA take the untoward consequences of impurity as a reason to reject SOE?

I don’t think they could. True: both KA and SOE are needed to complete the argument for
extreme epistemic pragmatism, and it is extreme epistemic pragmatism that generates the
untoward consequences. But it is clear that the principle primarily responsible for the untoward
consequences is KA, not SOE. It is KA that injects impurity in the epistemic realm, SOE just
spreads it around. SOE is a highly plausible principle about second-order evidence that has
strange consequences only in the presence of KA. On the other hand, KA has strange
consequences (of the same kind as the conjunction of KA and SOE) all by itself.

Notice also that someone who, like Fantl and McGrath, wishes to retain KA but reject extreme
epistemic pragmatism (and, thus, reject SOE) will be forced to say that a subject can have more
evidence for p than she does for the proposition that she knows that p. If the subject has some
evidence that it would be irrational for her to act as if p, then she thereby has some evidence
against the proposition that she knows that p, but she doesn’t thereby have some evidence against
p. If everything else is evidentially equal, then she will have more evidence for p than she does
for the proposition that she knows that p. Maybe it could be argued that this is as it should be, but
is is clearly a prima facie cost of this combination of views.
So, if the consequences of the conjunction of KA and SOE are unbearable, there are strong
reasons to think that it is KA that has to be rejected. But the rejection won’t come easy. This is
not the place to develop a way of coming to terms with the rejection of KA, but I believe that the
best way of doing it would involve taking fallibilism seriously. If we do take fallibilism
seriously, then we could perhaps come to accept that, even though Jeremy knows that the best
thing for him to do is to board the train, Jeremy is not sure that the best thing for him to do is to
board the train, and so it is rational (given the potential payout) for him to make further inquiries.

8 Conclusion

The dialectical situation is the following. Fantl and McGrath argue for moderate epistemic
pragmatism on the basis of KA. But the argument goes through only if extreme epistemic
pragmatism is false. Therefore, the most that can be said for moderate epistemic pragmatism is
that if KA is true and extreme epistemic pragmatism is false, then it is true. But this argument is
unstable, because, given KA, the very plausible SOE entails extreme epistemic pragmatism.
Thus, by entailing extreme epistemic pragmatism, KA and SOE together entail that pragmatic
encroachment goes all the way down. I take the upshot of the discussion to be that we must
either reject KA, or else admit that extreme epistemic pragmatism is true. As in many other
areas, so too in epistemology: purity is not easily lost only partially.

Acknowledgements Previous versions of this paper were presented at the University of Arizona,
the Second Midwest Epistemology Conference, and the 2009 meeting of the Pacific APA. Many
thanks to the audiences at those occasions and to Alan Sidelle and Elliott Sober for very helpful
comments and criticisms. Special thanks to Jeremy Fantl, Matt McGrath and Carolina Sartorio,
each one of whom read several versions of this paper and made it considerably better.

1 Here and in what follows I assume that only evidence can justify, so that if knowledge entails
justification then knowledge entails the having of evidence.

2 See Fantl and McGrath (2002), Fantl and McGrath (2007), and Fantl and McGrath (2009).

3 Some people (Lewis [1996], the entry for “Knowledge” in Quine [1987], Hetherington [2001])
think that knowledge does come in degrees. But even they must admit that there is a minimum
amount of evidence that you must have in order to count as knowing, and the question that sets
up our problem is about the location of this lower bound. I also ignore the vagueness of “knows”
here.

4 Some people (for instance, Lewis [1996] again) think that justification is not (always)
necessary for knowledge. Given my previous stipulation that evidence is what justifies, these
people would disagree with me that knowing that p entails having evidence for p. I would argue
(but not here) that this position stems from an overly intellectualistic conception of justification
(and, correspondingly, of having evidence). But never mind: even those who think that
justification isn’t necessary for knowledge have to admit that a subject cannot know that p unless
his epistemic position with respect to p is strong enough. This notion of the strength of one’s
epistemic position can replace the notion of having evidence throughout.
5 The case is taken from Fantl and McGrath (2002), who in turn take it from early discussion of
knowledge ascriptions by Cohen (1988). It is important to keep in mind that part of the story is
that Jeremy still believes, to the same degree that does Matt, that the train is the express.

6 Three influential contextualists are Cohen (1988), DeRose (1995), and Lewis (1996). Of these,
Cohen is the one whose terminology is closest to the one used in this paper: instead of talking of
the level of justification needed for knowledge, DeRose and Lewis talk about possibilities that
need to be eliminated.

7 Different contextualist theories will differ in the identification of the mechanisms by which
certain features are made contextually salient.

8 Is it also possible for there to be a fixed-threshold, variantist theory? Yes, it is. The kind of
contextualism introduced in the text distinguishes the different knowledge relations that the
knowledge predicate may refer to by where they locate the knowledge threshold, but there may
well be a contextualist theory according to which all the different knowledge relations locate the
threshold in the same place, but differ on some other dimension—for instance, with respect to
what kinds of defeaters are relevant to a knowledge attribution.

9 Following Fantl and McGrath (2002), we will say that you are rational to act as if p if and only
if, for any pair of acts A and B available to you, you are rational to prefer A to B if and only if
you are rational to prefer A and p to B and p. One complication: among the irrational acts that we
might perform, it makes sense to think of some of them as “less irrational” than others—they
have a suboptimal, but not minimal, expected utility. In that sense, practical rationality does
come in degrees.

10 I say “in general” because there may be cases where you do not care at all whether p. In those
cases, for you to act as if p just is for you to act as if not-p, in which case no evidence at all for
believing that p is needed for you to be rational in acting as if p. See also note 11.

11 I say “at least high enough” and not “just as high” because it is not plausible to suppose that if
you have enough evidence for p to make it rational for you to act as if p then you are in a
position to know that p. Sometimes you don’t care one way or the other whether p, in which case
no evidence at all is necessary for you to be rational to act as if p. Sometimes you care only very
little, in which case very little evidence for believing that p is necessary for you to be rational to
act as if p. But some (not very low) amount of evidence is always necessary for you to know that
p.

12 In later work Fantl and McGrath replace “S is rational to act as if p” with “p is warranted
enough to justify you in Ф-ing, for any Ф.” (Fantl and McGrath [2009, 66]) This replacement
doesn’t affect the arguments in this paper.

13 For a defense of related principles, see also Fantl and McGrath (2009, chapter 3).

14 Except, as Fantl and McGrath note, in those cases where whether S does A causally affects
what is the best thing for S to do.
15 But see the last paragraph of section 7 for a suggestion.

16 See Fantl and McGrath (2009, chapter 1) for an illuminating discussion of the problems
surrounding how to define fallibilism. Hawthorne and Stanley (2008) argue that “knowledge
delivers probability 1.” It is not clear whether they mean to deny fallibilism. Williamson (2000)
thinks that if S knows that p then S’s evidential probability for p is 1. However, Williamson also
argues that knowledge is defeasible by misleading evidence, which means that the evidential
probability of p can fall below 1 for a subject.

17 Fantl and McGrath (2007) and Fantl and McGrath (2009) make precisely this stipulation.

18 If, on the other hand, we retained the assumption that Matt and Jeremy have the same
evidence, leaving it open whether their respective practical positions change or not, then a theory
that adheres to extreme epistemic pragmatism may instead agree with the truth-values depicted
in Table 2 (depending on whether the extreme epistemic pragmatist theory is in addition a
moving-threshold or a fixed-threshold theory).

19 Premise 3 is assumption (A2) in Fantl and McGrath (2009, 86).

20 For a defense and application of a similar principle, see Fantl (2003).

21 For instance, that there is a hairless animal in the bush is evidence for the proposition that
there is a hairless dog in the bush, but is not evidence for the proposition that there is a dog in the
bush.

22 Even the restricted principle may need refining. But keep in mind that the cases to which the
premise will apply are ones where the subject in question (Jeremy) is not mistaken about
anything. All we need, actually, is not any general principle, but the claim that, as things stand
with the subject in this kind of case, he has evidence that he doesn’t know that p. Premises 3 and
4, and the proposed revised premise, are reasons to think that this is true, but even if we doubt
those reasons it still seems true that subjects in a situation like Jeremy will have evidence that
they don’t know.

23 Something that we should note is that, in the cases where I apply SOE, the subjects have no
mistaken belief about their epistemic position. Therefore, modifications of SOE that may be
called for if it is to apply to subjects with such mistaken beliefs don’t matter for my argument.

24 This doesn’t mean, of course, that Gettier showed that Purism is false. One may think
otherwise if one forgets that “being in a position to know,” as used in the statement of Purism,
means that the level of justification for the belief in question is high enough for knowledge.

25 That is, “I am in a Gettier-situation with respect to p” is a blindspot—see Sorensen (1988).

Notice also the parallel with Moore-paradoxical propositions such as “It is raining and I don’t
believe that it is raining.”
26 Indeed, one may wonder whether it is not begging the question against SOE to claim that
there are any conditions on knowledge that would qualify as extra-evidential under the proposed
characterization of the distinction. Some may propose as an obvious candidate the belief
condition on knowledge. But I don’t think it’s obvious that knowledge that one fails to believe
that p cannot make a difference to the strength of one’s evidence for p.

27 Of course, in order to know this it would have to be true that there are extra-evidential
conditions on knowledge and I would have to believe that there are extra-evidential conditions on
knowledge. But, as I said before, it is not clear that there are extra-evidential conditions on
knowledge in the requisite sense, and even if there were only a vanishingly small minority of us
would believe that there are (most people, for instance, would never even consider the question).

28 I don’t mean to imply that the two pieces of evidence are independent of each other—we
shouldn’t, for instance, count them both against the proposition that I know that p.

29 In my argument from the previous section I appealed to the fact that, according to Fantl and
McGrath themselves, we “rely on the truth” of KA in defense and criticism of actions. But this
reliance on the truth of KA better not amount to a conscious and explicit belief in KA of the kind
that will be necessary for ordinary folk to realize that it imposes an extra-evidential condition on
knowledge. KA is a controversial philosophical thesis, knowledge of which is at best hard to
come by.
Regress Argument Reconstruction
Jan Willem Wieland
Argumentation (2012) 26:489–503

Abstract If an argument can be reconstructed in at least two different ways, then which
reconstruction is to be preferred? In this paper I address this problem of argument reconstruction
in terms of Ryle’s infinite regress argument against the view that knowledge-how requires
knowledge-that. First, I demonstrate that Ryle’s initial statement of the argument does not fix its
reconstruction as it admits two, structurally different reconstructions. On the basis of this case
and infinite regress arguments generally, I defend a revisionary take on argument reconstruction:
argument reconstruction is mainly to be ruled by charity (viz. by general criteria which
arguments have to fulfil in order to be good arguments) rather than interpretation.

Keywords Infinite regress _ Argument _ Reconstruction _ Charity _ Interpretation

1 Introduction

One text, several reconstructions. The problem of argument reconstruction is just this: If a single
argument admits more than one reconstruction, then which reconstruction is to be preferred? Or
is the relevant text to say several things at once?

In this paper I will address this problem in terms of the well-known infinite regress argument
against the view that knowledge-how involves knowledge-that. This view basically says that any
intelligent action x is accompanied by knowledge that x is to be performed in such-and-such a
way (that reasoning, for example, is accompanied by knowledge of the inference rules). The
infinite regress argument against this view originates in Ryle:

The crucial objection to the intellectualist legend is this. The consideration of

propositions is itself an operation the execution of which can be more or less intelligent,
less or more stupid. But if, for any operation to be intelligently executed, a prior
 theoretical operation had first to be performed and performed intelligently, it would be a
logical impossibility for anyone ever to break into the circle. (1949: 30).

The basic idea is as follows. Suppose I want to perform an action. For example, I want to write a
paper. Yet, I do not want to write a paper in any way whatever. I want to write in an intelligent
way, viz. in such a way that I employ knowledge of how to do such things. Now suppose, as the
so-called intellectualist legend says, that knowledge-how involves knowledge-that. In that case,
if I want to employ knowledge of how to write a paper, I have to apply knowledge that papers
are to be written in such-and-such a way (e.g. that they are to be clear, thoughtful, convincing,
etc.). That is, I have to apply knowledge with propositional content. However, applying
knowledge with propositional content is itself an action that is to be performed intelligently.
Hence, by the intellectualist thesis, I have to apply knowledge that (applying knowledge that
papers are to be written in such-and-such a way) is to be applied in such-and-such a way. This
lands us in an infinite regress. Conclusion: The intellectualist legend entails an absurd regress
and so is to be rejected. Alternatively: If I want to perform an intelligent action, I should not
consequently apply knowledge-that.

Ryle’s argument received recent interest by Stanley and Williamson (2001), Hetherington
(2006), Williams (2008), Fantle (2011), among others. Here is an important disclaimer: I will
leave the problem of argument evaluation unaddressed. So, I will say nothing on whether the
argument is in fact sound and its conclusion true. All I will be concerned with is reconstruction.
Still, what I will say does have implications for argument evaluation: as soon as we know what
kind of premises and inferences are part of the argument, we know what kind of premises and
inferences are to be evaluated.

Before I show how Ryle’s text can be reconstructed in more than one way, I would like to make
some general remarks on argument reconstruction. I think everyone would agree that
reconstruction is to be guided by the following rules:

Interpretation Rule I. One should try to capture the original statement of the argument.

Interpretation Rule II. One should try to capture the context in which the argument
appears (the rest of the text or conversational setting, the author’s intentions, the
background literature, etc.).

If you do not respect these rules, then there is simply nothing that you are reconstructing. You
would just be constructing and setting out your own argument. However, even if these rules are
necessary, it is uncontroversial to say that the Interpretation Rules alone do not always suffice.
By these rules you get paraphrased texts at best, viz. texts in which complex terms and
constructions are replaced with simpler or clearer terms and constructions. Yet, if the original
statement of the argument is rather implicit, i.e. if many premises and inferences are suppressed,
then more is needed. This is exactly the case for infinite regress arguments (henceforth IRAs).
As Gratton observes:
 The typical presentation of infinite regress arguments throughout history is so succinct
and has so many gaps it is often unclear how an infinite regress is derived, or why an
infinite regress is logically problematic. (2010: xi).

To make these gaps explicit, at least two extra rules are needed1:

Charity Rule I. If needed, one should enforce the argument (i.e. modify or supply
premises and inferences) such that it becomes valid.2

Charity Rule II. If possible, one should weaken the premises (i.e. make them less
controversial and true) in such a way that the argument becomes sound.

Without these rules in place, it would be possible to say of almost any IRA in the literature that it
is invalid (and thus unsound) as, thanks to their gaps, their conclusions do not follow from their
premises. Such claims are however uninteresting. The point is not whether an argument, as
actually stated by someone in a particular context, is valid and sound, but whether there is a way
in which it can be valid and sound. I have to say much more on this the later on in the paper
(Sect. 4). Still, I would like to be explicit about the assumed aim of argument reconstruction
from the very start. To reconstruct an argument, I will say, is not just to interpret and get
people’s intentions right. It is also, and primarily, to do justice to the subject matter at hand, and
to see what an argument can, rather than does or had to, establish. Or again: argument
reconstruction is to be used to further our inquiries (and not just to learn something about the
arguers, viz. inquirers). The role of the Charity Rules just listed is to be seen from this
perspective.3

Also, it is worth pointing out that the rules are stated as plain commands. They can alternatively
be formulated as decision principles: ‘For any set of reconstructions of a single argument, one
should choose the reconstruction which is the strongest and yet the least controversial’. I have
disregarded this option, because, as I will show later, it need not always be the case that one of
the reconstructions is to be selected (rather than another). Sometimes both reconstructions are
valuable.

In the following I hope to show in a very specific way that the application of the Charity Rules
(however formulated) is a somewhat delicate enterprise. In particular, I will present two
reconstructions of Ryle’s text cited above, and generalize the difference for IRAs in general
(Sect. 2). With this result at our disposal, I address the problem of argument reconstruction: If an
IRA admits more than one reconstruction, then which reconstruction is to be preferred?
Eventually, I will take up a revisionary take on argument reconstruction and defend that Charity
rather than Interpretation is to rule how to proceed in case of such reconstruction problems
(Sects. 3, 4).

2 Two Reconstructions of Ryle’s IRA

In this section I will set out two reconstructions of Ryle’s IRA. I would like to stress that their
difference will not merely lie in the content of their premises. I am well aware that important
qualifications can be made in this respect. However, my point here is that Ryle’s argument
admits two reconstructions which differ not just content-wise, but structurally. That is, the
premises and inferences of Ryle’s argument can take different forms. Particularly, both
reconstructions will be instances of different argument schemas that have been developed in the
meta-debate on IRAs: the Paradox and Failure Schemas (these are my labels and will be
explained below). Both schemas have valid instances and for details and references
I have to refer the reader to the ‘‘Appendix’’. Please note that in order to discuss the problem of
argument reconstruction, I could have used a valid reconstruction as opposed to an invalid one as
well. Nevertheless, the problem is more pressing with two valid reconstructions. Here is the first:

Ryle’s IRA (Paradox Schema Instance)

(1) For any action x, one intelligently performs x only if one employs knowledge that x is
to be performed in such-and-such a way.

(2) For any action x, one employs knowledge that x is to be performed in suchand- such a
way only if one intelligently contemplates the proposition that x is to be performed in
such-and-such a way.

(3) You perform at least one intelligent action.

(4) You perform an infinity of intelligent actions (1–3).

(5) You cannot perform an infinity of intelligent actions.

(C) ~(1): It is not the case that for any action x, one intelligently performs x only if one
employs knowledge that x is to be performed in such-and-such a way (1–5).

(1) and (2) could be integrated into one line in order to simplify the derivation of the regress (cf.
Gratton 2010: 3, 193). Yet, on the basis of one line it is hardly possible to capture the context of
the argument (and we are to account for this by Interpretation Rule II). In this case the context is
the discussion whether knowledge how requires or involves knowledge-that. Moreover, if we
take one rather than two lines, then we may lose interesting parts of the dialectic. That is,
someone who is ready to accept/reject (1), need not accept/reject (2), or vice versa. The number
of lines can also be multiplied. For example, Stanley and Williamson (2001: 413–414) appeal to
three of them.4 Schematically speaking, the number of lines does not matter, as long as the
conjunction of (1) and (2) can be paraphrased as a universally quantified statement of the form
‘For all items x of type i, x is F only if there is another item y of type i that is F’.

Next the second reconstruction:

Ryle’s IRA (Failure Schema Instance)

(1) You have to intelligently perform at least one action.

(2) For any action x, if you have to intelligently perform x, you employ knowledge that x
is to be performed in such-and-such a way.
 (3) For any action x, if you employ knowledge that x is to be performed in such-and-such
a way, then you do not intelligently perform x unless you intelligently contemplate the
proposition that x is to be performed in such-and- such a way first.

(4) For any action x, there is always yet another intelligent action to be performed first,
viz. before performing x (1–3).

(C) If you employ knowledge that x is to be performed in such-and-such a way anytime

you have to intelligently perform an action x, then you never perform any intelligent
action (1–4).

Line (1) is a problem stated as a task in the form of ‘You have to u at least one item of type i’,
line (2) is a proposed solution to this problem stated as a strategy of the form ‘For any item x of
type i, if you have to φx, then you w ψx’. The rationale of such Failure arguments is that the
solution proposed in line (2) fails to solve the problem reported in (1) because it gets stuck in a
regress (hence the label ‘failure’). In Ryle’s case: If you consequently apply knowledge-that,
then you never perform any intelligent action thanks to a regress of intelligent actions.
Furthermore, if you still want to perform an intelligent action, then you should not consequently
apply knowledge-that and have to find another solution to do it.5

So, the two reconstructions prove different things. The rationale of Paradox arguments, by
contrast, is that certain statements cannot hold all at once, i.e. that one of them is to be rejected,
because they jointly lead, via a regress, to contradiction (hence the label ‘paradox’). In Ryle’s
case, the intellectualist thesis is rejected and shown to be false: It is not the case that, for all
actions x, one intelligently performs x only if one employs knowledge-that. It is worth stressing
that, in contrast to Paradox arguments, Failure arguments are not about rejections, not even the
rejection of solutions. More precisely, they do not prove that a solution is false (i.e. they do not
prove the negation of line (2)), but that it is unsuccessful for solving a given problem, and that
another solution be found.6

Interestingly, the ambiguity between the Paradox and Failure reconstruction may already be
found in Ryle’s initial statements of the argument. The following supports the Paradox
conclusion:

The regress is infinite, and this reduces to absurdity the theory that for an operation to be
intelligent it must be steered by a prior intellectual process. (1949: 32).

This supports the Paradox conclusion because it suggests the rejection of a universally quantified
statement. By contrast, the following two texts support the failure conclusion as they suggest that
the given task (viz. to perform an intelligent action) will never be accomplished by the strategy at
hand:

So no rational performance could ever begun. (1945: 10).

 But if, for any operation to be intelligently executed, a prior theoretical operation had first
to be performed and performed intelligently, it would be a logical impossibility for
anyone ever to break into the circle. (1949: 30).7

In the following I will generalize the difference between the Paradox and Failure version of
Ryle’s IRA. I have selected four other cases to illustrate this: one IRA by Sextus Empiricus
(1996) (Outlines of Pyrrhonism, Book 1, §15), one by Juvenal (1992) (Satire 6), one by Carroll
(1895), and one by Sorensen (1995) and Sider (1995). All these cases can be reconstructed along
the lines of both the Paradox and Failure Schema (for two full applications, cf. Wieland 2011).
Here are the corresponding conclusions respectively:

Paradox Conclusions

• It is not the case that for any proposition x, x is justified only if x there is another
proposition that is a reason for x.

• It is not the case that for any person x, x is reliable only if x is guarded by a guardian.

• It is not the case that for any set of premises x, a conclusion follows logically from x
only if x contains the additional premise ‘if the members of x are true, then the
conclusion is true’.

• It is not the case that for any action x, one is obliged to do x only if one can know that
one is obliged to do x.

Failure Conclusions

• If you provide a reason for x anytime you have to justify a proposition x, then you never
justify any proposition.

• If you hire a guardian for x anytime you should have someone x guarded, then you
never have your girlfriend guarded.

• If you appeal to an additional premise ‘if the foregoing premises are true, the conclusion
must be true’ anytime you have to demonstrate that we are forced to accept a conclusion,
then you never demonstrate that we are forced to accept any conclusion.

• If you appeal to an obligation not to make it impossible to know one’s obligation x

anytime you have to secure x, then you never secure any obligation.

Any IRA from the literature can be reconstructed in different ways partly because no text is fully
explicit on its premises and/or inferences. Specifically, IRAs from the literature have many
suppressed premises, if they are not suppressed it is often unclear how they are quantified, and
virtually always the inference steps to the conclusion remain mysterious. If this is right, then
what is to be done? There are at least three options:
 Paradox-Monism. One should always use the Paradox Schema to reconstruct an IRA.

Failure-Monism. One should always use the Failure Schema to reconstruct an IRA.

Pluralism. One should sometimes use the Paradox Schema to reconstruct an IRA, and
sometimes the Failure Schema.

All these positions have their proponents in the meta-literature on IRAs. Prominent Paradox-
Monists are Black (1996) and Gratton (1997, 2010). A prominent Failure-Monist, and one of the
main initiators of the whole debate, is Passmore (1961). Prominent Pluralists are Schlesinger
(1983), Sanford (1984) and Day (1986, 1987). Logically speaking, there is also a fourth position
possible, namely a variety of Pluralism which says: One should always use both the Paradox and
Failure Schema to reconstruct an IRA. I have not found any proponent of this in the literature,
but my own position comes closest, as we will see in the next section.

On first sight, Pluralism appears attractive because it takes into account the reconstructor’s
purposes, viz. what she wants to do with the argument. Is she interested in a Paradox conclusion,
or a Failure conclusion? I think this should indeed be accounted for, but will argue that it need
not give you Pluralism as long as the purposes are always Failure-purposes (for example). This
will be the topic of the next section.

3 Comparison Paradox Versus Failure Arguments

At this point the question is: If Ryle’s IRA can be reconstructed in two different ways, which
reconstruction is better? Or put in general: Are we to use the Paradox Schema to reconstruct
IRAs, or the Failure Schema, or both? In the following I will defend the claim that often Failure
arguments are better arguments, and hence that the Failure Schema is to be employed in most
cases. I have three, related arguments:

(1) The conclusions of Failure arguments are stronger, as they admit fewer options for
resistance.

(2) The conclusion of Failure arguments is automatically relevant, and this need not be
the case for Paradox arguments.

(3) The hypothesized solution of Failure arguments is automatically motivated, and this
need not be the case for Paradox arguments.

Argument (1). In order to explain this first argument, I have to clarify the broader, dialectical
contexts of Paradox and Failure arguments. Suppose that A and B are two arbitrary persons (they
may even stand for one and the same person in case she is reflecting on her own beliefs or
solution), then the dialectical scenarios are the following.

Dialectic Paradox arguments

Line Dialectical context

(1) Hypothesis for reductio A’s position

(2) Premise B defends this
(3) Premise B defends this
(4) Infinite regress B infers this from (1)–(3)
(5) Premise B defends this
(C) Rejection B infers this from (1)–(5)

Dialectic Failure arguments

Line Dialectical context

(1) Problem A wants to solve this

(2) Solution A’s proposal
(3) Premise B defends this
(4) Infinite regress B infers this from (1)–(3)
(C) Failure B infers this from (1)–(4)

In both cases, B, the person who devises the IRA, does all the work. For B to defend something
is not to show that she herself holds it, but to show that A should be prepared to concede it. A, in
turn, can try to resist B’s reasoning at each of these steps. Let me briefly go through the options.
In case of Paradox arguments, A has four options for resistance: she may argue that (1) does not
in fact belong to her beliefs (or at any rate in its fully universally quantified version), defend the
same for (2), reject (3), or deny the extra premise (5). In case of Failure arguments, A has but
three options: she may deny that the given problem is something to be solved in the first place,
she may reject that the solution is something she had proposed (or at least in its fully universally
quantified version), she could deny (3), and this is it.8

The important difference is that in case of Paradox arguments, B needs to defend an extra
premise after the regress (in Ryle’s case, this is the assumption that I cannot perform an infinity
of intelligent actions). There is no such extra step in case of Failure arguments (in Ryle’s case,
you just never perform an intelligent action, whether you can perform an infinity of them or not).
This difference implies that the conclusions of Failure arguments are stronger than their Paradox-
counterparts, as they admit fewer options for resistance. This point should not be underestimated,
for if the extra premise after the regress in case of Paradox arguments fails to hold, then nothing
is to be rejected. In Ryle’s case this possibility may not appear very promising (for in that case
you have to explain how one is able to perform an infinity of intelligent actions), but in other
cases this is not immediately clear. It has been challenged, for instance, that there cannot be an
infinity of propositions which justify one another (to cite a well-known regress, cf. Klein 1999;
Peijnenburg 2010).

Argument (2). To explain the next argument we have to consider the conclusions of the argument
schemas, and see about their relevance. Here is the form of a Paradox conclusion:

It is not the case that, for all items x of type i, x is F only if such-and-such (e.g. there is
another item y of type i and x and y stand in R).

The question is: so what? Well, you now know that the universally quantified statement does not
hold (although its existentially quantified counterpart can still hold). One case where this is
interesting is the IRA from Sorensen (1995) and Sider (1995). They are explicitly concerned
about the question whether the Access principle holds unrestrictedly, i.e. the principle that says
that you are obliged to do something only if you can know this obligation. In that case it is
interesting to find out that the principle does not always hold.

Still, in many cases this sort of conclusion is not relevant, and does not make any difference to a
certain debate. For example, it is uninteresting to find out that it is not the case that for any
person x, x is reliable only if x is guarded by a guardian (also given that this may still hold for
some persons), or that it is not the case that for any set of premises x, a conclusion follows
logically from x only if x contains the additional premise ‘if the members of x are true, then the
conclusion is true’ (also given that this may still hold for some sets of premises).

Next, consider the form of a Failure conclusion:

If you ψ all items of type i that you have to φ, then you never φ any item of type i.

So what? Well, if you have to φ at least one item of type i, then you have to find another solution
to solve this problem. As the problem is a common concern of persons A and B, this is an
interesting result in each and every case. In terms of the examples, this means that you have to
find another solution (another strategy, other means) to perform an action, justify a proposition,
have your girlfriend guarded, demonstrate that a conclusion follows logically, and secure an
obligation. Hence, conclusions of Failure arguments are automatically relevant, and this is my
second argument for why such arguments are better arguments.

Argument (3). The last argument is that the hypothesis, i.e. line (2), of Failure arguments is
immediately motivated.9 This line is motivated because it presents a proposal to solve the
problem as described in line (1). For example, you hire a guardian in order to have someone
guarded, or you appeal to an additional premise in order to demonstrate that a conclusion follows
logically from the premises. But it is not at all clear in all cases why one would introduce and
consider the corresponding lines of Paradox arguments (viz. instances of (1) of the Paradox
Schema). For example, why would anyone believe that, for any person x, x is reliable only if x is
guarded by a guardian, or that, for any set of premises x, a conclusion follows logically from x
only if x contains the additional premise ‘if the members of x are true, then the conclusion is
true’?10

Still, although Paradox arguments are not always good arguments, sometimes they are. Ryle’s
IRA is a case in point. The hypothesis ‘for any action x, one intelligently performs x only if one
employs knowledge that x is to be performed in such-and-such a way’ is worth considering for
anyone who believes that all our intelligent actions are accompanied by knowledge-that.
Likewise, the Paradox conclusion should be interesting as it is the rejection of this claim. To be
sure, this does not mean that the Failure conclusion is not interesting as well. It is still interesting
to find out that you never perform any intelligent action if you employ knowledge-that anytime
you have to perform an intelligent action. For in that case you have to find other means to
perform intelligent actions.11
So, in general it might turn out that you end up with two distinct reconstructions. But, as long as
these are two good arguments, this should not be a problem. Let me conclude this section. I
showed that Paradox arguments do not play the same role as Failure arguments in a broader
dialectical context. Moreover, I presented arguments for why it is often the case that the Failure
Schema produces better arguments: its conclusions are stronger, immediately relevant and its
hypothesis is immediately motivated. Hence my position:

Pluralism*. One should often use the Failure Schema to reconstruct an IRA, and
sometimes the Paradox Schema.

4 Revisionism

So far, I argued that often an IRA is to be reconstructed on the basis of the Failure Schema
because that schema has better instances. What I did not discuss so far is how this pertains to the
general debate on argument reconstruction. So in this section I provide a deeper motivation for
why the Monisms and standard Pluralism from the literature do not suffice.

Recall the problem of argument reconstruction: If there is one text and at least two available
reconstructions, then which reconstruction is to be preferred? My answer can now be formulated
as follows. If there are two reconstructions, then it should be checked which reconstruction is the
best argument. Basically, you should choose the reconstruction with the most plausible premises,
and strongest and most interesting conclusion (i.e. the conclusion which has the least options for
resistance and which makes a difference in the broader debate in which it occurs).

Sounds trivial. It is not trivial, however. It is a rather revisionary take on argument

reconstruction. Revisionism heavily relies on the Charity Rules from Sect. 1, which say that one
should modify arguments in such a way that they become valid and sound.

Of course, Revisionism does not disregard the Interpretation Rules (i.e. that we should capture
the initial statement of the argument plus its context), for in that case no reconstruction would
have any content. They would not be reconstructions of anything. Still, Revisionism is
completely ignorant about what an argument had to be, and focuses solely on what the argument
can be, i.e. what its optimal format is. For example, Charity is not directed at Ryle’s actual
words, nor at Ryle himself and his intentions, but at the most interesting and strongest statement
of the argument. To be sure, in some cases it might also be interesting what Ryle himself thought
of the matter (whether it be ambiguous or obscure). But that would be just another enterprise:
mere interpretation rather than argument reconstruction. Compare Feldman’s motivation of the
Charity Rules:

We should adhere to [them] not because it is nice to do so or because people need or

deserve charity, but because adhering to [them] leads us to consider the best available
arguments and thus gain the most insight into the issue we are studying. (1993: 115).

I think this is right. We should try to obtain the best and strongest reconstructions in order to see
what can be obtained by a given argument, and so gain the most insight into the issue we are
studying (for example about knowledge-how, as in Ryle’s case).
The opposite of Revisionism may be called ‘Conservatism’. The most extreme variant of the
latter would be that the Charity Rules should not be applied at all: one should just do it with what
is in the text and not add nor subtract anything substantially. More moderate variants would be
happy with the Charity Rules except that they should not be applied unrestrictedly (as
Revisionism has it), but only to a certain extent. It is not clear to me what exact restrictions might
be imposed here, but the general thought is that one should not depart too much from the initial
statement.

Consider the debate on IRAs again. I distinguished between three camps, viz. Paradox-Monism,
Failure-Monism and Pluralism, and at this point my position Pluralism* may also be called
‘Revisionary Pluralism’, contrasting it with both Conservative Pluralism and the Monisms.

Conservative Pluralism would be the position which observes that IRAs in the literature take
different forms and draws from this that sometimes IRAs are to be reconstructed Paradox-wise
and sometimes Failure-wise. I think this view is too easy. It has too much respect for the way
IRAs are actually stated and does not apply the Charity Rules. I do think that sometimes IRAs
are to be reconstructed Paradox-wise and that sometimes they are to be reconstructed Failure-
wise, but only because in those cases one of the arguments is better than the other.12

Further, I disagree with the Monisms in that it is not always the case that IRAs are to be
reconstructed Failure-wise (or Paradox-wise for that matter). My view is a Pluralist view exactly
because it takes into account that the purposes of the person who is reconstructing the argument
(i.e. what she wants to do with the argument) may vary. Is she interested in a Paradox or a
Failure conclusion? Still, what is interesting is no subjective or arbitrary matter. Strong
arguments with plausible premises and relevant conclusions are more interesting than weak
arguments with implausible premises and irrelevant conclusions (and I have shown why often
the Failure Schema instances score better at this).

In the literature I have found two important criticisms of the Charity Rules (cf. Walton 1996:
216; Walton and Reed 2005: 341–342; Paglieri 2007: 2–6):

(i) The rules, if taken strongly, may distort the initial statement of the argument (and for
example lead to a straw man).

(ii) The rules, if taken weakly, are not precise enough to select one reconstruction among
the available ones.

A full treatment of them cannot be given here. But let me just briefly point out why they do not
they apply in case of IRAs. Against (i) it can be said that distortion is no problem as long as
IRAs from the literature are gappy (see Sect. 1) and are not meant to be gappy (as perhaps
enthymemes are). Against (ii) it can be said that selection is not to be a problem for IRAs. The
main problem is to get valid and sound reconstructions (this is what the Charity Rules motivate),
and if it turns out that a single argument admits two sound reconstructions (as might well be the
case for Ryle’s IRA), then we just have two sound arguments.
So here is my position on argument reconstruction in a nutshell: Argument reconstruction, at
least in the case of IRAs, is hardly fixed by the initial text, and should rely more heavily on
general criteria which arguments have to fulfil in order to be good arguments.13

5 Epilogue

Ryle claimed at some point (1945: 7) that people can reason well even if they have no
knowledge-that of the inference rules. It this right? I do not know. But maybe the important point
is not whether they can do it, but how we can check whether they can do it (i.e. reason well).
And this, it seems, can only be done on the basis of knowledge-that of inference rules.

Here are two analogies. Analogy no. 1: People can devise good IRAs even if they have no
knowledge-that of the forms IRAs can take (i.e. have knowledge of the argument schemas).
Analogy no. 2: People can reconstruct well even if they have no knowledge-that of the
Interpretation and Charity Rules. It might well be that people can devise good IRAs and
reconstruct well without knowledge-that. Still, the only way to check whether people devise
good IRAs and reconstruct well is on the basis of knowledge-that of the schemas and
reconstruction rules respectively: knowledge which I discussed in this paper.

Acknowledgments Thanks to Anna-Sofia Maurin, Eline Scheerlinck, Maarten Van Dyck, Erik
Weber and the reviewer of the journal for excellent advice. The author is PhD fellow of the
Research Foundation Flanders (FWO).

Appendix

Paradox Schema

(1) For all items x of type i, x is F only if such-and-such (e.g. there is a y and x and y
stand in R) (HYP).

(2) For all items x of type i, such-and-such only if there is a new item y of type I and y is
F (PREM).

(3) There is at least one item of type i that is F (PREM).

(4) There is an infinity of items that are F (1–3).

(5) There is no infinity of items that are F (PREM).

(C) ~(1): It is not the case that, for all items x of type i, x is F only if such-and-such (1–
5).14

To obtain instances of this schema, ‘type i’ is to be replaced with a specific domain, and the
capitals ‘F’, ‘R’ with predicates which express properties and relations (e.g. items of type i =
propositions, F = being justified, R = being a reason for). For details about the inferences (i.e. the
suppressed lines and rules), cf. Wieland (ms). Main supporters of this schema are Black (1996)
and Gratton (1997, 2010). Parts of this schema have also been discussed or suggested, if only
briefly, by Russell (1903), Beth (1952), Yalden-Thomson (1964), Gettier (1965), Schlesinger
(1983), Sanford (1984), Day (1986, 1987), Clark (1988), Jacquette (1996), Nolan (2001), Klein
(2003), Maurin (2007), Rescher (2010: ch. 2).

Failure Schema

(1) You have to u at least one item of type i (PREM).

(2) For all items x of type i, if you have to φx, you ψx (HYP).

(3) For all items x of type i, if you ψx, then there is a new item y of type i, and you φx
only if you φy first (PREM).

(4) For all items x of type i, there is always a new item that you have to u first, viz. before
φ-ing x (1–3).

(C) If you ψ all items of type i that you have to φ, then you never φ any item of type i (1–
4).15

To obtain instances of this schema, ‘type i’ is to be replaced with a specific domain, and the
Greek letters ‘φ’, ‘ψ’ with predicates which express actions involving the items in that domain
(e.g. items of type i = propositions, φ = to justify, ψ = to provide a reason for). For details about
the inferences (i.e. the suppressed lines and rules), cf. Wieland (ms). This schema is my own
contribution to the literature. The pioneer of this schema is Passmore (1961). Parts of this
schema have also been discussed or suggested, if only briefly, by Russell (1903), Rankin (1969),
Armstrong (1974), Johnson (1978), Rosenberg (1978), Schlesinger (1983), Sanford (1984), Day
(1986, 1987), Ruben (1990), Johnstone (1996), Gillett (2003), Maurin (2007).

1 The ‘charity’ label derives from Wilson (1959) and has been made popular by Quine and
Davidson in the debate on radical interpretation. For the application of Charity to argument
reconstruction, cf. e.g. Rescher (1964: 162), Feldman (1993: 115).

2 ‘Validity’ and ‘soundness’ are used in the usual way, so an argument is valid just in case the
conclusion follows from the premises, and it is sound if the premises are true as well.

3 For a clear statement of these two goals, cf. Johnson (2000: 132, cf. 158). For related reasons,
the Charity Rules have been criticized in the literature, and I will turn to this in Sect. 4.

4 Namely: (1) If one Fs, one employs knowledge how to F. (2) Knowledge how to F is
knowledge that φ(F). (3) If one employs knowledge that p, one contemplates the proposition that
p.

5 An alternative reconstruction of Ryle’s IRA in failure format starts from (1*) You have to
explain how S performs at least one intelligent action, and concludes to (C*) If you appeal to the
fact that S employs knowledge-that anytime you have to explain how S intelligently performs an
action, then you never explain how S performs any intelligent action.

6 Two further structural differences concern their take on infinite regresses (series of necessary
conditions vs. series of problem/solution pairs) and the step from the regress to the conclusion
(and what kind of premises and inferences are supposed to license this step). For a full
comparison, cf. Wieland (ms).

7 Ryle uses the word ‘circle’. Yet it is clear that he refers to an infinite regress argument, and not
to a circularity argument. These two kinds of arguments can and should be kept apart, but I will
not go into that here.

8 To be sure, in both cases one may try to resist the suppressed premises and inferences next to
the premises. Yet, in most cases this is not very obvious, if the suppressed premises are general
truths and the inferences licensed by classical rules. See the ‘‘Appendix’’ for references.

9 This last argument relates to the first argument because the strength of the conclusion depends
on the number and plausibility of the premises.

10 A selected number instances might be motivated by Aristotle’s, historically influential,

Principle of the Absolute (cf. Beth 1952: 66–68; Peijnenburg 2011).

11 Or find another way to explain how people perform intelligent actions, if the problem is
theoretical (viz. to explain this) rather than practical (i.e. to perform them yourself), cf. fn. 5
above.

12 I do not think that anyone in the literature exactly fits this description of a Conservative
Pluralist. Still, the studies by e.g. Day (1986, 1987) and Gratton (1997, 2010) are far less
revisionary than what I am proposing here. A clear example of a Revisionist is Black (1996).
Yet, he is no Pluralist, as we have seen in Sect. 2, but Paradox-Monist.

13 What the list of these general criteria exactly is, I did not say. It may even be a slightly
different list for different kinds of arguments. Still, I made three specific suggestions for IRAs in
Sect. 3.

14 This schema has variants where (2) or (3) is rejected rather than (1). See the references below.

15 This schema has a variant with a conclusion of the form ‘You never φ all items of type i’
rather than ‘You never φ any item of type i’. See the references below.
Johanssonian Investigations.
Infinite Regress Arguments*
Anna-Sofia Maurin

Abstract: According to Johansson (2009: 22) an infinite regress is vicious just in case “what
comes first [in the regress-order] is for its definition dependent on what comes afterwards.”
Given a few qualifications (to be spelled out below (section 3)), I agree. Again according to
Johansson (ibid.), one of the consequences of accepting this way of distinguishing vicious from
benign regresses is that the so-called Russellian Resemblance Regress (RRR), if generated in a
one-category trope-theoretical framework, is vicious and that, therefore, the existence of tropes
only makes sense if trope-theory is understood (minimally) as a two-category theory which
accepts, besides the existence of tropes, also the existence of at least one universal: resemblance.1
I disagree. But how can that be? How can Johansson and I agree about what distinguishes a
vicious from a benign regress, yet disagree about which regresses are vicious and which are
benign? In this paper I attempt to answer that question by first setting out and defending the
sense of viciousness which both Johansson and I accept, only to then argue that to be able to
determine if a particular regress is vicious in this sense, more than features intrinsic to the regress
itself must be taken into account. This is why, although the RRR as originally set out by Russell
is vicious, the seemingly identical resemblance regress which ensues in a one-category
(standard) trope-theoretical context is not (provided, that is, that we accept certain views about
how the nature of tropes relates to the resemblance between tropes, and given that we set our
theory in a truthmaker theoretical framework – all of which are standard assumptions for
proponents of (the standard-version of) the trope-theory).2

Infinite regress arguments occupy a unique position in philosophical reasoning: They are wielded
with unusual force and against an impressive number of different views stated in significantly
different philosophical contexts.3 Surprisingly enough, the argument itself is language (for
instance the meaning regress generated from and therefore used against the however
comparatively little discussed.1 Not that surprising, someone might object. Our notion of an
infinite regress is after all one that is well defined and hence well understood.2 Surprising
nonetheless, I insist. For the fact (if indeed it is a fact), that we fully understand the mechanisms
necessary to set into motion generation ad infinitum can hardly guarantee that we fully
understand the proper use and evaluation of an infinite regress in the context of an infinite
regress argument.

1. Infinite Regress Arguments and a Mostly Missing Premise

To see how the infinite regress argument differs from its constituent infinite regress, consider its
ingredients:

1. The premises necessary for the generation of an infinite regress.3

2. Conclusion1: the infinite regress.

3. The premises necessary to show that conclusion1 is unacceptable.4

4. Conclusion2: the rejection of one or more of the premises listed under (1).

The infinite regress argument is a species of the kind reductio ad absurdum that can function as
proof that the position from which a regress is generated should be abandoned, precisely because
it includes premises that show not only that and how a regress is generated but also why what is
generated is a cause for concern (why it is “absurd”). Conclusion2, that is, depends essentially on
whether or not the premises that appear under (3) can demonstrably, relevantly, and with
sufficient strength discredit the position from which the regress is generated. The premises
necessary to show that conclusion1 is unacceptable are obviously nothing we could or should
expect to find included in the infinite regress itself.

In spite of this, in an unexpected number of applications of the infinite regress argument, no

mention is made of premise 3. That the regress is unacceptable (that it is “vicious”) is at most
said to be “plain” or “obvious”.1 This is no innocent omission. For, even supposing that, at least
in some cases, there are regresses which plainly exhibit some potentially vicious-making feature,
this is in itself proof that there is something seriously wrong with (some part of) the position
from which the regress has been generated only if we assume that every time a regress has some
potentially vicious-making feature, it is vicious. But this is the same as to exclude from the outset
the possibility that what potentially vicious-making feature or features actually make a regress
vicious may vary from one theoretical context to another. If the possibility of such context-
dependent variation is accepted, on the other hand, reasons must always be provided for why the
presence of a potentially vicious-making feature, in this particular context, relevantly and
therefore also problematically, interferes with that in the criticized view we wish to reject. This is
precisely the sort of reasons we (ought to) find under 3.

So, what feature or features make a regress vicious? In the next section, I will argue, with
Johansson as we have seen, that a regress is vicious if it instantiates a certain “pattern of
dependence” (in a sense that will be spelled out in more detail below). My argument for this
point is negative. That is, I will argue for this view of viciousness from the claim that alternative
understandings fail, either because the feature they blame is as a matter of fact not problematic,
or because, whether or not it is, it is a feature which characterizes every infinite regress, which
means that it cannot be what distinguishes the vicious from the benign.

2. Against the Standard View: Why Regresses aren’t Substantially Vicious

By arguing that there is but one way of distinguishing vicious from benign regresses, I will be
opposing the “standard view”. According to the standard view, a regress may be (intrinsically)
vicious not only because of the way in which its different steps relate to one another (and to the
position from which the regress is generated), but also more substantially, because of some
feature or features instantiated by the entities generated in the regress (collectively or
individually).

Distinct yet closely related substantial reasons for viciousness have been suggested. They have
this in common: in one way or another, directly or indirectly, they blame as vicious-making the
number of entities to which the regress commits us.1 But the fact that the number of entities
produced in an infinite regress is always the same – infinitely many – prevents anyone intent on
preserving the distinction between a regress that is vicious and one that is benign from resting
content with simply repudiating anything infinitely large. There must be some reason for
distinguishing, and then for preferring, certain infinities over others. One option is to, with
Aristotle, distinguish between so-called actual and potential infinity. According to Aristotle,
potential infinity is acceptable infinity. It is infinity such that:

In general, the infinite is in virtue of one thing’s constantly being taken after another –
each thing taken is finite, but it is always one followed by another; but in magnitudes
what was taken persists, in the case of time and the race of men things taken cease to be,
yet so that [the series] does not give out. (Aristotle, Physics, III.6 206a27-206b2)

Potential infinity is thus infinity in the sense of “capacity” and entails the existence only of
finitely many entities (at a time).1 An actual infinity, on the other hand, is a completed infinity,
all of which members exist. An actual infinity, if such there is, belongs to the furniture of the
universe. Actual infinites, Aristotle maintained, are unacceptable in a very strong sense; they are
impossible. They are impossible, moreover, because their existence entails the existence of
something with proper parts the size of the whole to which they belong.2 That no such thing
could exist, he argued, followed trivially from the fact that if it did it would contradict the
axiomatic Euclidean principle that the whole must be greater than its proper parts (Elements,
book 1, Common Notion 5).

That the distinction between an acceptable and an unacceptable infinity cannot be drawn along
the lines of actual and potential infinity, at least not for the reason urged by Aristotle, is today
almost universally recognized. True, no finite set can be such that its proper parts are the same
size as the whole to which they belong. But, infinite sets are radically different from finite ones.
So different, in fact, that that which made Aristotle deem the actual infinity impossible, is now
singled out as its distinguishing mark. An infinite set is, as noted by Georg Cantor (1932),
nothing other than a set where the whole is equinumerous with its proper parts. To instantiate
this feature, then, does not make a set impossible, it makes it infinite.1 If we want to be able to
draw the distinction between a regress that is vicious and one that is benign along the same lines
as that between an infinity that is actual and one that is only potential, we must find some other
reason for doing so.

Directly after the publication of Cantor’s results discussions on the topic of acceptable and
unacceptable infinities were particularly lively. Although the notion of an actual infinity is
consistent, it was now urged, it does not follow that there can be anything in reality to which it
applies. This was David Hilbert’s view. According to him, actual infinity only had a role to play
as an ideal addition to a finitist mathematics. He concluded that:2

[T]he infinite is nowhere to be found in reality. It neither exists in nature nor provides a
legitimate basis for rational thought – a remarkable harmony between being and thought.
(Hilbert, 1983 [1926]: 201)

But what does it mean to say of an infinity that it is real as opposed to ideal, and why, exactly,
could there not be actual infinity in reality? Now, reality as we experience it while going about
our daily business is finite. However, the fact that what we experience is finite cannot, unless we
want to subscribe to some naïve empiricist principle, be the reason why actual infinity cannot
belong to reality. Instead, what we must look for is something about the real – reality – that
makes it an unsuitable host for actual infinity. An often cited example meant to illustrate why
reality is inapt to harbour actual infinities is that of the so-called paradox of the Grand Hotel:1

An infinitely large hotel (a truly Grand hotel) with infinitely many guests (a “full” hotel, by
finitist standards) can always fit one more guest in, by moving each of the guests already
occupying a room to the room next to it (thereby leaving room one free for the newcomer). In
fact, it can fit infinitely many new guests in (by, this time, moving each guest to a room with a
room-number twice as large as the one they were occupying, thereby leaving all the odd-
numbered rooms free for the infinitely many newcomers). And, if infinitely many guests move
out – it will still be full!

William Lane Craig (1991: 85-96) thinks that the paradox of the Grand Hotel proves that real
actual infinities are impossible.2 It is, however, unclear exactly what in the example proves this.
Craig himself points to two “absurdities”: the task of adding the guests in the manner set out in
the example and the fact that the odd-numbered rooms in the hotel must be as many as its total
amount of rooms. Neither absurdity proves that actual infinity in reality is impossible however.
The fact that, in a truly Grand hotel, the odd-numbered rooms would have to be as many as its
total number of rooms, first of all, only seems absurd because not surprisingly our intuitions are
modelled on that with which we are familiar: the finite. To argue against the possible existence
of real actual infinity on the basis of such intuitions would of course be question-begging. And,
although there is something troubling and perhaps even absurd about the performance of the
requisite room-changing task, only if the absurdity results from the infinity itself can it be proof
that real actual infinity is impossible. But it does not. Moving guests in the required way involves
a “supertask”: a task with infinitely many steps (the adequate changing of rooms) performed in a
finite amount of time. Finite beings (such as the infinitely many guests occupying rooms at the
hotel) could most probably not perform the task of changing rooms in a finite amount of time.
The absurdity comes, not from some particular property had by the infinity in question, but from
the contrast between the properties that the infinity does have and matters of fact concerning
human beings and their capacities. From our perspective this is not interesting because, even if
we agree that supertasking involves us in absurdity, this does not mean that the actual existence
of infinitely many real entities is impossible.

Maybe we should not expect to find an explanation of why reality cannot harbour infinity by
studying the infinite as a whole, but rather by studying the properties of the entities of which the
real (as opposed to ideal) infinities are composed. Suppose that reality consists of the “middle-
sized dry goods” – the objects – with which we interact daily. These are concrete, as opposed to
abstract entities. To say of an entity that it is concrete is, let us suppose, (minimally) to say that it
is such that it occupies only one position in space at each moment in time, that it monopolizes
this position (at least in relation to other concrete objects1), and that it has an identity that can be
retained over time and through at least some changes. To say that there can be no actual infinity
in reality is perhaps just to say that there can be no actual infinity of entities of this kind. To say
this cannot be the same as saying that there can be no actual infinity in reality, however. Both
ontology and natural science are subject-matters which aim at disclosing the “true” nature of
reality. Yet, both ontological and scientific theories posit entities that do not behave like the
entities with which we are accustomed to interact do. They behave, rather, as we would expect
an abstract entity to behave.2 In fact, on many (perhaps even most) theories about the
fundamental nature of reality, to be able to account for the existence of precisely that which we
want to call a concrete object, abstract entities must be posited (think of forces, fields, processes
and the like in natural science, or of relations and properties in ontology).3 Therefore, from the
impossible existence in reality of an actual infinity of concrete entities, it does not follow that
there could be no actual infinity in reality. If we want to distinguish vicious from benign infinity
in terms of concreteness, we must therefore accept that reality can harbour actual infinities of a
kind that is not objectionable.

But why should one think that reality cannot harbor actual infinities of concrete objects? One
reason might be that, since concrete entities are typically such that they monopolize their
position in space-time, there is quite simply no room for an actual infinity of entities of that kind.
As far as I understand, the size of the universe, and especially the question whether, if it is
infinite, this infinity is actual or merely potential, is not yet settled (a fact that in itself would
seems to point to the possible existence of actual infinities). We may therefore reasonably ask: If
the universe is finite in size, could it make room for an (actual) infinity of concrete entities?
Russell, discussing these matters in The Principles of Mathematics, says ‘yes’. As long as we
believe that there are bounded stretches of space (or time) it in fact follows that there are actual
infinities in the world. He points out that those who deny the existence of actual infinity still
admit that what they call finite space may very well be a “given whole”, but:

…such a space is only finite in a psychological sense – it is not finite in the sense that it is
an aggregate of a finite number of terms, nor yet a unity of a finite number of
constituents. Thus to admit that such a space can be a whole is to admit that there are
wholes which are not finite. With respect to time, the same argument holds. (Russell,
1903: 144)
It seems therefore that we must look elsewhere for that which distinguishes entities making up
unacceptable infinities from entities making up acceptable ones. One alternative is perhaps
provided by the principle of ontological parsimony expressed by the so-called “Ockham’s razor”.
Ockham’s razor tells us that we should not postulate entities beyond necessity. Daniel Nolan
explains how this insight may be used to distinguish a regress that is vicious from one that is
benign:

…the boundary might well be this: a regress is taken to be benign when the quantitative
extravagance is a cost worth paying, and vicious when either the quantitative
extravagance is not a cost worth paying, or if it has some more serious fault of which the
regress is evidence. (Nolan, 2001: 536-537)

But what is to decide whether quantitative extravagance is a cost worth paying or not? What
makes whatever the regress commits us to, unnecessary? Suppose, as it is often said, that
unnecessary entities are entities that are idle or inert. You might want to say that an entity is idle
if it exists for no particular purpose, but talk of purposelessness is not of much use here. For one
thing, a kind of purposelessness seems characteristic of every entity generated by a regressive
mechanism (with the possible exception of the entities generated in its first step). It may even
seem as if the regresses that we find most unobjectionable are so more or less because that to
which they commit us is especially purposeless – and so does not substantially interfere with
whatever explanatory task we are for the moment engaged in. To consider what is idle as what is
causally powerless (and equate the razor with a kind of Eleatic principle), is not much of an
improvement. Again, most of the regresses we find unobjectionable turn out to be vicious.
Worse, infinite regresses by many considered as plainly vicious, like the causal regress, become
virtuous almost by definition.1

But if the viciousness of a vicious infinite regress is not situated in some particular feature had
by each of the infinitely many generated entities, individually or collectively, then where does it
reside? What, if nothing “substantial”, makes a vicious regress vicious? In the literature it is
more or less unanimously agreed that, whether or not a regress may be vicious for substantial
reasons, it can be vicious for what we may provisionally call “structural” reasons. That is, and
more precisely, besides (or, as we have just argued, rather than) being vicious (or benign) for
substantial reasons, a regress is vicious or not because of the way its distinct steps relate to one
another, and to the position from which the regress was originally generated.

3. The Structural Understanding of Viciousness

According to Johansson, as we have seen, to accept the “structural” understanding of viciousness

is to accept that in a regress of the problematic kind, “what comes first [in the regress-order] is
for its definition dependent on what comes afterwards.” This way of formulating the distinction
now needs to be somewhat qualified. My first qualification concerns the scope of Johansson’s
formulation. Johansson puts the distinction in terms of dependence for definition, but this is
clearly unfortunate. For, although definition is sometimes what is at stake,1 it is far from always
what is at stake.2 The distinction between vicious and benign regresses is therefore better put in
terms of (direction of) dependence generally: Regresses of the vicious kind are such that the first
step of the regress will depend (for its definition, but also, as the case may be, for its justification,
existence, meaning, etc.) on what appears in the next step of the regress (etc. ad infinitum).
Regresses of the benign variety are characterized rather by the opposite direction of dependence.

That the distinction between vicious and benign regresses ought to be understood in terms of
direction of dependence in the way set out above is a view that I have proposed repeatedly in
past publications (cf. e.g., my 2002; 2007). I still believe that this way of understanding the
distinction allows you to correctly identify as vicious the great majority of the vicious regresses
(and as benign the great majority of the benign regresses). However, I now think that it might
lead you to wrongly identify a regress as vicious (or as benign) in certain (admittedly highly
improbable) circumstances. This is why I now want to propose a second and arguably more
substantial qualification both to Johansson’s account, and to my own generalized version of
Johansson’s account. On this modified view, a vicious regress is vicious if it somehow hinders
the position from which it has been generated from “fulfilling its explanatory (or other) task”.
This way of understanding the distinction between a regress that is vicious and one that is benign
is in fact nicely captured in the following quote from Johnstone:

…an alleged definition (or criterion or explanation, or, for that matter, analysis,
justification of X, or account of the decision to do Y) gives rise to a vicious infinite
regress when instead of defining (or serving as a criterion, explanation, analysis, or
justification of X, or account of the decision to do Y) it merely postpones the definition,
explanation, analysis, or justification, or account of the decision. (Johnstone, 1996: 97)

A good thing about Johnstone’s formulation is that it explains why understanding the distinction
in terms of direction of dependence in the way set out above seems to be on the right track. It
seems to be on the right track, that is, because, in most circumstances, a regress that instantiates a
dependence-pattern of the (potentially) vicious-making kind, is a regress which postpones – and
hence hinders – the original position from constituting the explanation (account, definition,
justified proposition, etc.) it claims to be. Another good thing about Johnstone’s way of putting
things is that while it in this way rationalizes a distinction formulated in terms of direction of
dependence, it also manages to make room for the (admittedly not very likely) possibility that a
regress which instantiates a pattern of dependence of the presumably vicious-making kind is
nevertheless not vicious. This is because what is important for viciousness, on Johnstone’s
account, is whether or not the existence of the regress somehow hinders the position from which
it has been generated from “being” whatever it claims to be: a full explanation, a justified
proposition, a possible existent, a meaningful proposition, or what have you. Clearly, given
certain (probably rather controversial) framework assumptions, neither explanation, justification,
or existence need be incompatible with the existence of an infinite regress, even an infinite
regress which instantiates a pattern of dependence of the seemingly problematic kind.1

But this means that, on the present view, whether a particular regress is vicious (or not) will
depend essentially on what is assumed in the situation at hand. It will depend, that is, on what is
the relevant question to which the regress-generating position purports to provide the (full)
answer and, perhaps even more importantly, it will depend on what in the present context counts
as a full answer to that question. On this view, then, whether or not a particular regress is vicious
cannot be ascertained simply by studying the features (whether substantial or structural)
instantiated by the regress itself independently of the theoretical context in which it appears. It is
this fact, I will next try to demonstrate, which arguably explains the puzzling disagreement that
exists between Johansson and myself.1

4. Assessing Viciousness: The case of the RRR

In “On the Relations of Universals and Particulars” (1956[1911]), Russell examines and
dismisses what he calls “the theory which admits only particulars”. The theory he has in mind
seems to be the same as that advocated by e.g., George Berkeley and David Hume in their
polemic against abstract ideas. The theory is described by Russell as follows:

The general term ‘white’, in this view, is defined for a given person at a given moment by
a particular patch of white which he sees or imagines; another patch is called white if it
has exact likeness in colour to the standard patch. In order to avoid making the colour a
universal, we have to suppose that ‘exact likeness’ is a simple relation, not analyzable
into a community of predicates; moreover, it is not the general relation of likeness that we
require, but a more special relation, that of colour-likeness, since two patches might be
exactly alike in shape or size but different in colour. (Russell, 1956: 111)

A problem arises, says Russell, because the nominalist must, so as to not make the relation of
colour-likeness universal, apply the same analysis as was previously applied to the property
shared by distinct objects to it: “we may take a standard particular case of colour-likeness, and
say that anything else is to be called a colour-likeness if it is exactly like our
standard case.” (ibid) This leads to an infinite regress which, Russell concludes, is “plainly
vicious”.

Now, the context is here ontological, but what are the framework assumptions? The quote offers
us some clues. It is the general term ‘white’ which is defined for a given person at given moment
as a particular patch of white. The same analysis must be applied to (our notion of) exact
resemblance, and then again and again ad infinitum. What Russell is objecting to are the views of
someone who not only believes that there are only particulars, but who also believes that what
there is, is to be decided by a close study – an analysis – of our conceptualisation of reality. If we
appear to be conceptually committed to e.g. universals, proponents of the view under attack must
hold that this appearance can only be rejected as illusory if it can be demonstrably analysed
away. Consequently, if you subscribe to the view criticised by Russell, the trigger – a is exactly
similar to b – can be true, only if the similarity class to which the exact similarity holding
between a and b belongs, exists, and so on for each new level of exact similarity. This is so
because at each step a new general term will appear, and so demand an analysis. This regress is
vicious because at no step are the conditions necessary for a to resemble b ultimately fulfilled,
which means that the existence of the regress hinders the theory from which it has been
generated from providing a full account of (in this case) the fundamental nature of reality.

Notice, however, that what may appear to be the same regress would not be vicious (or so I
would like to claim), if it were based on different framework assumptions. Suppose, again, that
the context is ontological. We are interested in what there is, and, more precisely, we are
interested in arguing that what there is, is particular. On one such view, trope-theory, all there is
are particular properties. What makes it true that two distinct concrete particulars share a
property (e.g. are both red), is that each particular contains a red-trope and that the red-tropes
exactly resemble one another. The exact resemblance of the red-tropes must, however, be given
an ontological account. Given trope monism, the only available ontological characterisation is
one according to which exact resemblance is yet another trope. It is this admission which
generates an infinite regress of the same type as that launched by Russell against classical
Nominalism. On the face of it, the Russellian resemblance regress and its trope theoretical
counterpart will look exactly the same. The trigger in both cases is the state of affairs that a
exactly resembles b.1 One difference is, of course, the nature of a and b. On the view criticised
by Russell, a and b are concrete objects (colour patches, more precisely), whereas the basic
question for trope theory will concern the exact resemblance of tropes, which are a kind of
abstract particulars, or particular properties. The relevant difference is not this difference in
nature, however. To be able to determine if the regress is vicious or not we must consider the
framework assumptions. On most versions of trope theory, objects do not have properties
because they belong to some particular similarity class. Instead, they belong to some particular
similarity class, because they have some particular properties – the tropes, which nature is
primitive (this is the “standard” view among the trope-theorists, a view that is defended by e.g.,
Williams 1953; Campbell, 1990, and; Maurin 2002). The trope theorist can say this, because she
does not assume that in order to account for resemblance she must be able to in one way or
another analyse away each occurrence of a general term. Instead, most trope theorists operate in
a truthmaker theoretical framework. The question is what is required for the truth and not
necessarily also for the meaning of the theory’s central propositions. Therefore, if you are a trope
theorist, the trigger, to obtain, requires no more than the existence of tropes a, b and their trope
of resemblance. The infinite regress does not prevent the trigger from existing. It is rather the
existence of the trigger that sets into motion the infinite generation of exact resemblance tropes.
The trope theoretical resemblance regress is, therefore, benign.

Johansson and I can agree about what makes a regress vicious, yet disagree about which
regresses are vicious, therefore, either because we disagree about what needs to be taken into
account in order to be able to determine if a regress is vicious or not, or, because we disagree
about the framework assumptions given which the relevant (trope-theoretical) regress is
generated.

* This paper is a (sometimes substantially, sometimes not so substantially) rewritten version of

my paper “Infinite Regress: Virtue or Vice” (2007). Thank you Ingvar Johansson for inspiring
me to think more about these matters!

1 That reality contains both tropes and universals is also Johansson’s view. For an introduction to
this his Aristotelian-cum-Husserlian inspired view of reality, cf. esp. his 2004.

2 This text partly repeats, partly continues, a discussion between Johansson and myself which
resulted in some texts (in Swedish), published a few years ago (more precisely: Johansson, 2008;
2010 and Maurin, 2009; 2010).

3 You will find the argument used in epistemology (one prominent example is the justification
regress launched in defense of foundationalism); in metaphysics (one example is the resemblance
regress launched against nominalism; another is the exemplification regress proposed against
universal realism); in the philosophy of language of thought hypothesis); and so on.

1 Among those who have discussed it, you find e.g., Day, T. J. (1987); Gratton, C. (1997; 2010);
Johnstone Jr., H. W. (1996); Nathan, N. (2001); Nolan, D. (2001); Oppy, G. (2006); Passmore, J.
(1961); Priest, G. (2002); Sanford, D. H. (1984); Schlesinger, G. (1983); Waismann, F. (1968),
and; Wieland, J.-W. (2013). A majority of these philosophers are proponents of the standard
view (introduced below).

2 This is apparently A. F. MacKay’s view when he, in trying to understand why Arrow’s
theorem is true, attempts to show that a crucial part of the theorem’s proof can be recast as an
infinite regress argument. This recasting, it is supposed, will further our understanding of the
theorem exactly because our notion of an infinite regress is “the more familiar and perspicuous”
(1980: 367).

3 Here I will not discuss what these premises are, but I take it that Gratton’s (1997; 2010)
exposition comes very close to the truth (cf. also Wieland 2013 for a good overview). According
to Gratton, the relevant premises are: a regress formula (i.e. any statement (or combination of
statements) that entails, or is intended by its author to entail an infinite regress) plus some sort of
triggering statement. To illustrate, an infinity of the relevant kind results from the combination
of “Everything that exists has a cause (which exists)” (regress formula) and “a (i.e. something)
exists” (triggering statement).

4 I assume that an infinite regress is either vicious (“absurd”) or benign. This assumption is
supposed to be uncontroversial. Benign regresses are no cause for celebration – they are
tolerable, just as a benign tumor is (mostly) tolerable (but cf. my 2011, for a discussion of a view
according to which some infinite regresses are not only tolerable, but in fact positively
beneficial). One very good reason for assuming the existence of not only vicious, but also benign
infinite regresses is the existence of what seems to be perfectly good examples of such. The truth
regress, for instance, is considered quite innocent by most people and, if you are not one of those
people, the arithmetic regress presents an even less controversial case. There are some
philosophers who do not want to talk about benign regresses; to them, a regress is always
vicious. Johnstone (1996), for one, talks merely of a benign series. I understand this as a mere
terminological disagreement and will therefore disregard it in what follows.

1 To illustrate, consider the following statement by Russell (1956:112, my italics): “[W]e explain
the likeness of two terms as consisting in the likeness which their likeness bears to the likeness
of two other terms, and such a regress is plainly vicious.” I shall argue later on in this paper that
Russell is demonstrably wrong about this.

1 Precisely because the blame is put on the very feature of a regress which most of us would
agree seems potentially problematic – its infinity – the substantial understanding of viciousness
has been (and still is) one that importantly influences the way we think and talk about infinity
and about the infinite regress. As put by Johnstone (1996: 97-98): “Formulations of both
nonvicious and vicious regresses may make use of the phrase “ad infinitum” ... In both cases, the
hearer is supposed to regard this phrase as a danger signal – a warning of the same magnitude of
seriousness as the phrase “…is a contradiction”.”

1 In Aristotle’s own words (Physics, III.6 206b 33-34): “[i]t turns out that the infinite is the
opposite of what people say it is: it is not that of which no part is outside, but that of which some
part is always outside”.

2 This was not Aristotle’s only reason for repudiating actual infinities. It was, however, the
reason that, for generations to come, was counted as his best reason to do so. For a presentation,
discussion and criticism of some of Aristotle’s other reasons for repudiating actual infinities, see
Priest (2002: 31f.).

1 History was not completely devoid of defenders of actual infinity before Cantor. In a letter to
Foucher, Leibniz wrote already in 1693 that: “I am so in favour of the actual infinite that instead
of admitting that Nature abhors it, as is commonly said, I hold that Nature makes frequent use of
it everywhere, in order to show more effectively the perfections of its Author. Thus I believe that
there is no part of matter which is not, I do not say divisible, but actually divided; and
consequently the least particle ought to be considered as a world full of an infinity of different
creatures.”

2 More precisely, Hilbert distinguished between, on the one hand, a fundamental system of
quantifier-free (i.e. finite) number theory and, on the other hand, a formal addition of transfinite
axioms. The latter were added so as to simplify and complete the theory, but they were added in
a sense analogous to that in which “in geometry, the ideal constructions are adjoined to the
actual” (Hilbert, 1996 [1923]: 1144).

1 This is not really a paradox in the logical sense of the word – it is rather an example exploiting
the extreme unintuitiveness of the idea of an actual infinity as applied to concrete reality.

2 And he then goes on to apply this conclusion to the cosmological argument for the existence of
God (God must exist, for if he does not an actual infinity of causes can be generated into the
past, actual infinities are impossible, so there must be a first cause and this is God).

1 I am here – and for the sake of the argument – disregarding the huge discussion on spatially
coincident concrete particulars (the so-called “statue/clay” debate). But cf. J. J. Thompson 1998,
for a good introduction.

2 To distinguish what is abstract from what is concrete is truly no easy task. For a fuller
treatment of the issue, cf. e.g., my (2002), especially chapter 2.

3 For an interesting discussion of the abstract entities to which science appears to commit us (and
of what these apparent commitments might entail for ontology) see B. Ellis (2005) and S. Psillos
(2005).

1 Nolan might object that his is not a distinction between regresses that commit us to entities that
are idle and those that do not so commit us. His distinction is formulated in terms of costs and
benefits and says that a benign infinite regress is one where the cost of quantitative extravagance
is acceptable and a vicious regress is one where it is not. I think Nolan is right to think that the
standard of parsimony (whether quantitative or qualitative) is not absolute. However, and for the
same reason, quantitative parsimony cannot substantially distinguish a vicious from a benign
regress even if it can be used to distinguish good from bad or acceptable from unacceptable
theories.

1 Cf. Russell (1903: 348) who puts the relevant distinction in terms of definition/meaning: “in
the objectionable kind [of regress] two or more propositions join to constitute the meaning of
some proposition; of these constituents, there is one at least whose meaning is similarly
compounded; and so on ad infinitum.” And, again, “an endless process is not to be objected to
unless it arises in the analysis of the actual meaning of a proposition” (ibid: 51). An infinite
regress of a benign variety, on the other hand, would have the following appearance: “[i]f A be a
proposition whose meaning is perfectly definite, and A implies B, B implies C, and so on, we
have an infinite regress of a quite unobjectionable kind” (ibid: 349).

2 In e.g., an epistemological context, the relevant relations have to do with whether the
proposition that occurs in one step of the regress justifies or is justified by the next. In ontology,
what is at stake is existence and the relevant relations therefore concern existential dependence
between entities; whatever is posited at each step of a regress has its existence determined by
either what precedes it in the regress-order, or by what succeeds it.

1 An example might be if one gives up the requirement that explanation (or, for that matter,
justification) must ground out. Cf. e.g., Orilia 2009 and Gaskin 2008 for two suggestions along
these lines. Cf. Maurin 2011 for a critical discussion of their particular suggestions.

1 The “structural” understanding of viciousness certainly deserves a fuller treatment than it can
be given here. However, hopefully, the admittedly sketchy account provided above is enough to
make sense of the discussion set out in this text’s final section.

1 Or, given that in order to generate the relevant regress, we need at least three resembling tropes
(for an argument to this effect, cf. Johansson 2009), the trigger must rather look something like
this: a, b, and c exactly resemble each other.
Synthese (2014) 191:661–670
On the regress problem of deciding how to decide
Hanti Lin

Abstract Any decision is made in some way or another. Which way? (Have I worked out
enough alternatives to choose from? Which decision rule to apply?) That is a higher-order
decision problem, to be dealt with in some way or other. Which way? That is an even higher-
order decision problem. There seems to be a regress of decision problems toward higher and
higher orders. But in daily life we stop moving to higher-order decision problems—stop the
regress—at some finite point. The regress problem of deciding how to decide is the problem of
explaining what would make it rational to stop the regress. I will give a new solution in the
present paper. The result suggests a new way of looking at standard Bayesian theory and the
more recent theory of adaptive rationality.

Keywords Rational choice ・ Bayesian theory of rationality ・ Adaptive rationality ・

Bounded rationality ・ Deliberation costs ・ Deciding how to decide ・ Externalism ・
Internalism

1 Introduction

“Should I go grocery shopping today?” According to the standard wisdom in decision theory, the
agent should consider some options as alternatives to going grocery shopping today, and evaluate
the options by considering a number of possible states of the world and by considering the
outcomes of those options in each of those possible states. That is, the agent should specify the
decision problem in question. Standard decision theory is silent about which specification is the
best one for the agent to adopt on a particular occasion. It is only after one settles for a
specification of the decision problem that the standard decision theory starts to say something,
recommending which decision rule to apply in order for one to be a rational decision maker.
We have listed some “elements” for rational choice according to the standard decision theory: a
list of alternative courses of action to choose from, a partition of possible states of the world,
some descriptions of outcomes, and a decision rule. To fix those elements is to determine a way
of deciding. When an agent is not satisfied with simply fixing those elements without
deliberation, she may ask: “how should I fix those elements?” And she may deal with the
problem as a second-order decision problem, in which one decides among various fixations of
the elements in order to address the first decision problem. This opens the door to higher-order
decision problems, leading to a regress.

In daily life we stop the regress. The question is what would make it rational for us to stop the
regress. That is what I call the regress problem of deciding how to decide. I will describe what
seems to be the single most important motivation for the regress (in Sect. 2) in terms of the costs
incurred by deliberation. This will allow me to explain in detail what the regress problem is
(Sect. 3). Then I will discuss earlier works that (intentionally or unintentionally) contribute to
solution of the problem, and explain why they do not work (Sects. 4, 5). This discussion will
motivate my solution (Sect. 6).

2 The costs of deliberation

The move to higher-order decision problems can be motivated by a single consideration. Good
decision requires deliberation, but deliberation comes with its own costs—and it is not trivial to
strike a balance.

A good decision requires evaluation of each option on the table in terms of its possible outcomes.
But should a decision maker consider outcomes that might occur in 10 minutes, 10 days, or 10
years? Prediction requires deliberation. Prediction for a farther future requires more deliberation.
And deliberation carries costs. Deliberation may be boring or exhausting, and it usually requires
one to spend the time that could have been spent in a more enjoyable way. If one chooses to
deliberate more, the hope is that one will end up with a better decision than otherwise—a
decision that is sufficiently better to counterbalance the time and effort invested in deliberation.
To deliberate more, or not—for figuring out possible outcomes in a farther future—that is a
question. And it is a second-order decision problem.

With options and their possible outcomes made explicit, the agent can choose a decision rule to
apply, but different decision rules carry different deliberation costs. Let us compare the Bayesian
decision rule “maximize expected utility” with an easier-to apply decision rule, say “maximize
minimal utility” (i.e., the so-called maximin rule). Suppose that the agent has figured out which
outcome is at least as desirable as another. Then she is already in a position to apply the maximin
rule: choose an option whose least desirable outcome is at least as desirable as the least desirable
outcome of any other option. She may either decide to apply the maximin rule or decide to
deliberate more in order to approach Bayesian rationality, namely, to approach coherence among
preferences, beliefs, and desires in a way recommended by Bayesians.1 One way to satisfy
Bayesian coherence is (i) to figure out how probable one takes an event to be, (ii) to figure out
how much more desirable one particular outcome is as compared to another, and (iii) to prefer
acts that have higher expected utilities. So, on a particular occasion of decision-making, should
the agent apply the maximin rule or should she strive to approach Bayesian rationality (or
possibly follow some other decision rules)? Immediate application of the maximin rule may
result in a choice that is inferior according to the Bayesian standard, but the inferiority may or
may not be compensated by the deliberation costs saved for enjoying the rest of the day. To
deliberate more, or not—with the aim of approaching the Bayesian ideal of rationality—that is
the question.

We have noted that we have to make several choices in order to solve a first order problem:
should we think about outcomes in a farther future, and should we strive for the Bayesian ideal
of rationality? There are more choices to make. For example, should I go grocery shopping
tonight? I could read a paper instead, or play video games. And I could do something else, which
is always an option, a “catch-all option” as I shall call it. The unspecificity of a catch-all option
makes it difficult for an agent to evaluate it. What are the possible outcomes of doing something
else? Well, it depends on what one could do specifically in doing something else. So evaluation
of the option “doing something else” requires one to work out possible alternative courses of
actions that one has not thought of. When some specific option is worked out, say “watching a
movie”, it becomes a new option on the table, which gives the term “else” an updated, more
restricted range of application. But, still, the option “doing something else” is unspecific in this
updated sense of “else”. Now, should the agent deliberate further so as to find out what one
could do specifically in doing something else? Or should she stop such deliberation and proceed
to choose from among the options currently on the table (either including or ignoring the option
“doing something else” in the updated sense)? To deliberate more, or not—for evaluating the
catch-all option—that is the question.

In sum, concerns about the deliberation costs for dealing with a decision problem motivate
second-order decision problems. But a second-order decision problem is itself a decision
problem, which requires deliberation for its solution. Then, how much to deliberate? That
motivates a third-order decision problem, and so on. In general, given an n-th order decision
problem D, there are motivations for dealing with an (n+1)-th order decision problem D’—a
problem that concerns how much deliberation to be invested in decision problem D. The
existence of some reasons for dealing with D’ does not mean that one has to deal with D’ in
order to be rational. There are even reasons against it, based on concerns about deliberation
costs, too. Why not ignore the higher-order problem D’ in order to avoid the deliberation it
requires? To weigh the reasons for and against dealing with a higher-order decision problem D’,
ironically, is to deal with an even higher-order decision problem D’’: take D’ seriously, or ignore
it, that is a question of (n + 2)-th order.

3 The regress problem

In daily life we stop moving to higher-order decision problems—stop the regress—at some finite
point. The regress problem of deciding how to decide is the problem of explaining what would
make it rational to stop the regress. I do not know whether we can give a necessary and sufficient
condition for practical rationality that answers the question. But at least we want to be able to say
that it is rational for an agent to decide to stop the regress at a certain point because of certain
features of his decision, which are to be identified in a positive solution to the problem.
It is important to get clear about what the problem is not. The regress problem is not about the
following worry: if one stops the regress at a higher-order decision problem Dn irrationally, then
as a consequence one must make an irrational choice in some of the lower order decision
problems that rely on the choice made in Dn. That worry is unwarranted. Note that an agent
makes a choice given the following inputs into a decision problem: a set of alternative options to
choose from, a list of their possible outcomes, the agent’s beliefs and desires about those
outcomes, etc. Even if the agent’s beliefs are irrational (due to, say, wishful thinking), a choice
can still be made rationally given those inputs (perhaps by forming preferences that cohere with
the agent’s desires and beliefs, as Bayesians would say). In general, a choice can be made
rationally even if there is some irrationality in the inputs into the decision problem in question—
e.g., even if the list of options and their outcomes comes from an irrational choice in a higher-
order decision problem.

There appears to be a quick solution to the regress problem. Resnik invites us to imagine an
agent who is chased by an angry bear (1987, 11). Which tree to climb? It seems to make no sense
for the agent to answer that question by working out a decision table explicitly and applying a
decision rule. Deliberation that lasts more than three seconds would cost the agent’s life. The
agent identifies the closest tree, finds that it is high enough, rushes to it, and climbs it, without
surveying other trees. She is rational.

Throughout the above story, the agent seems to do nothing that can be appropriately called
decision-making, but she acts rationally nonetheless. Perhaps we can think of the agent as
making a decision after all, just in a limiting sense. Then we may say that she makes a decision
on intuitive grounds (Johansen 1977), or that she makes an immediate decision (Resnik 1987), or
an unguided decision (Smith 1991).

These three authors—Leif Johansen, Michael Resnik, and Holly Smith—all appear to draw the
following conclusion: the regress toward higher-order decision problems can be stopped
rationally whenever it is rational for an agent to stop at a decision problem as the consequence of
a decision that is immediate, unguided, and/or based on intuitive grounds. Those three authors
stop at that conclusion. But what they take as a conclusion seems to me rather a starting point
toward a positive solution to the regress problem. What these authors fail to ask is the question as
to what makes an immediate decision a rational one. After all, not every immediate decision is
rational. It is irrational to commit suicide when no reason comes to mind.

To illustrate the point, compare the regress problem of deciding how to decide with the regress
problem of justification in epistemology. If justification of a belief always requires one to adduce
a distinct belief that has not been already adduced, then justification never comes to an end—
hence the infinite regress of justification. The regress problem of justification can be avoided by
applying modus tollens: since we do have justified beliefs without an infinite regress of adducing
more and more justifying beliefs, we can deny the consequent and, hence, deny the antecedent.
Namely, we can conclude that it is sometimes the case that justification of a belief does not
require one to actually adduce a distinct belief that has not been adduced before. But when that is
the case, what justifies the belief in question?2 That is the real problem, which motivates various
positions for solving the regress problem rather than for merely avoiding the problem.
I submit that the same attitude be applied to the regress problem of deciding how to decide. That
is, let us ask: what makes it rational for an agent to stop the regress as the consequence of a
decision that is immediate, unguided, and/or based on intuitive grounds?

4 Bayesian rationality

The Bayesian conception of rationality, which is standard in decision theory, is unable to help us
solve the regress problem. The following argument explains why. Suppose that, after some steps
of regress, an agent stops at an n-th order decision problem, and that she does so rationally. If we
want to employ the Bayesian, coherence conception of rationality to explain such a rational
choice, we have to cite facts about coherence among the agent’s preferences, beliefs, and desires.
Since the agent stops at an n-th order problem rationally, she prefers “stopping at level n” to “not
doing so”. Let us look at the following two cases.

Case (i): the agent comes to have her preference by considering some outcomes of “not stopping
at level n”. So the agent considers some outcomes of “trying to solve a problem at level n + 1”.
In other words, the agent comes to have her preference by evaluating the respective outcomes of
the two options: “stopping at level n” and “trying to solve a problem at level n +1”. But this
amounts to dealing with a decision problem whose level L is greater than or equal to n+1, which
contradicts the fact that she stops at level n. So case (i) is impossible. (In case you are interested:
the present argument relies only on the fact that L is at least n +1, but can L be equal to n +1? It
can if a decision problem Dn+1 can refer to itself so as to decide whether to deal with the very
same problem Dn+1, without resorting to any higher-order problem. But I don’t know whether
this kind of self-reference can be consistent.)

Since case (i) is impossible, its negation must be true, which I shall call case (ii): the agent does
not consider any possible outcomes of “not stopping at level n”. Then, if we are to give a
Bayesian explanation of the agent’s rational stop, we will have to cite the fact that the agent
satisfies Bayesian coherence among the following mental properties:

• the preference of A “stopping at level n” to A’ “not doing so”

• the desirabilities of certain outcomes o1, o2, . . . of A (without the desirability of any
outcome of A’)

• the beliefs about whether those outcomes o1, o2, . . . of A will occur (without the belief
about any outcome of A’).

But it is trivial to satisfy Bayesian coherence among those mental properties. Note that those
mental properties are coherent in the Bayesian sense iff there exist:

• real numbers EU, EU’ (as the expected utilities of alternatives A, A’, respectively)

• real numbers u1, u2, . . . (as the utilities of outcomes o1, o2, . . ., respectively)
 • real numbers p1, p2, . . . in the unit interval (as the probabilities of those outcomes,
respectively)

such that the following two conditions hold:

EU ≥ EU’
EU = u1 p1 + u2 p2 + …

Because the formula for expected utility EU’ is missing, such numbers always exist. So Bayesian
coherence among those mental properties always holds. By sticking to the Bayesian conception
of rationality, we have to conclude that the agent’s stop at level n is rational because of a trivial
truth—which seems implausible.

In sum, if we can successfully employ the Bayesian conception of rationality to explain the
agent’s rational stop at level n, then either case (i) holds, which is impossible, or case (ii) holds,
which is implausible. So I conclude that the Bayesian conception of rationality is unable to help
us solve the regress problem.

5 Adaptive rationality

Concerns with deliberation costs have led to a way of understanding rational choice that is very
different from the standard decision theory. This way is based on the concept called adaptive
rationality and it aims to replace the Bayesian conception of rationality. While Bayesian
rationality is about the internal coherence of one’s mental state, rationality in the adaptive sense
depends only on some external facts that the agent may have no access to. The thesis of adaptive
rationality can be expressed as follows: the rationality of an agent’s choice in a particular
situation depends only on the extent to which the choice method in use is robustly goal-
conducive, in the sense that the choice method that she actually uses would help her achieve her
actual goals in every situation similar to the actual situation.3 Note that it requires goal-
conduciveness that is robust, i.e., holds in the actual situation as well as in other situations
similar to it. That requirement is stronger than goal-conduciveness that is merely actual, i.e.,
holds only in the actual situation. The agent’s goals vary from situation to situation; for example,
goals can be survival, reproductive success, or successful exchanges with others, etc. Recall
Resnik’s angry bear case. It seems correct to say that, when the agent is chased by a bear, the
most important goal is to survive. Theorists of adaptive rationality would say that, if the agent’s
immediate decision is based on a choice method that is robustly goal-conducive, then the
decision is rational. Admittedly, it is not clear whether we can have a definite measure of how
well a choice method helps one achieve her goals. But the general point is that the rationality of a
choice depends only on certain facts—facts about the extent to which the choice method in use is
goal-conducive—which the agent may have no access to. It is a very strong form of externalism
about rationality.

The simplest decision problems are clearly the ones in which there is only one course of action
or strategy to choose from. To deal with such a simple decision problem without moving to a
higher-order decision problem is to adopt the first option that one recognizes without trying to
work out a second act for comparison. When an agent does that, she is said to follow a choice
method called the recognition heuristic. In the angry bear example, suppose that the first strategy
the agent thinks of is: “identify the closest tree and climb it if it is high enough”. What if the
closest tree is not high enough? The agent would not think about that question until she finds that
the closest tree is too short. It seems rational for the agent to act in the angry bear scenario
according to that strategy without further deliberation. Theorists of adaptive rationality would
add: adoption of that strategy is rational because it is based on a choice method, i.e. the
recognition heuristic, that is robustly goal-conducive as a matter of empirical fact.

The recognition heuristic is only one of the many possible choice methods studied by theorists of
adaptive rationality. When a choice method is easy to apply, it is called a heuristic. Theorists of
adaptive rationality aim to identify various heuristics that we seem to follow (explicitly or
implicitly) in different situations, and to determine the situations in which they are robustly goal-
conducive.

The above suggests a kind of externalist solution to the regress problem (which, however, I shall
refute). Imagine an agent who considers a second-order decision problem concerning whether to
employ the maximin rule or to maximize expected utility. Suppose that she stops at that problem
and terminates the regress by the heuristic “Do not address a third-order decision problem”. If
that heuristic turns out to be robustly goal-conducive in that situation, the agent stops the regress
rationally. In general, if the agent’s choice to stop the regress at a certain finite point is based on
a choice method that turns out to be robustly goal-conducive, then the choice is rational,
independently of whether the agent believes or knows the relevant facts about goal-
conduciveness.

We will now see that this solution to the regress problem cannot be correct.

6 The spectrum of rationality

If we want to explain what makes the agent’s response to the angry bear rational, it is not enough
to cite facts about goal-conduciveness. The agent’s immediate reaction to the angry bear would
not be rational if the agent were to believe (truly or falsely) that the heuristic in use does not
satisfy actual goal-conduciveness, let alone robust goal-conduciveness—no matter what heuristic
is in use, and no matter whether the heuristic in use is goal-conducive or not. The lesson is that
rationality always requires something about one’s doxastic state.

The above suggests an easy fix to the approach of adaptive rationality, namely, that an agent’s
decision is rational whenever:

(A) the choice method in use is robustly goal-conducive,

(B) she does not believe that the choice method in use is not actually goal-conducive.

However, the easy fix won’t work. Perhaps (A) plus (B) suffices for rationality if the situation is
very urgent. But what if the situation is not too urgent? In that case, it seems not so rational for
the agent to act by a heuristic and simultaneously have no idea whether the heuristic in use is or
is not goal-conducive. That is because, in such cases, rationality seems to require that the agent
also believes that the heuristic in use is actually goal-conducive. Moreover, given that the agent
does believe that a heuristic is actually goal-conducive, the rationality of acting by the heuristic
seems to be independent of whether the heuristic is robustly goal-conducive as a matter of fact
(although the agent’s belief might be irrational). The standard decision theory takes a similar
stance: when one chooses an act that maximizes expected utility, the rationality of that choice is
independent of whether the act chosen in fact produces a more desirable outcome than the other
acts that are at one’s disposal.4

In sum: A necessary condition for the rationality of following a decision method M is (B): that
one does not believe that M is not actually goal-conducive. As to what additional conditions are
required by rationality, it depends on the situation in which one makes a choice. Depending on
the situation, rationality may or may not require the additional, external condition (A): that M is
robustly goal-conducive. Furthermore, rationality may or may not require the additional, internal
condition that one believes that M is actually goal-conducive. Rationality may or may not require
another internal condition that is more sophisticated and Bayesian in nature: that the agent thinks
about how desirable each relevant outcome is and how probable each relevant state of nature is.
A general (but sketchy) rule is this: the more urgent the situation is, the more the external
conditions for rationality are in force, and the less the internal conditions are in force and, hence,
the less one is required to deliberate. That is, moving to situations with more and more urgency,
rationality will eventually, first, drop all internal conditions except (B) and, second, come to
require the external condition (A). Moving in the opposite direction (toward situations with less
and less urgency), rationality will drop any external condition and require more and more
sophisticated deliberation. For example, in everyday planning, it is rational for the agent to
decide to go grocery shopping today if he just has the binary belief that doing so is actually goal-
conducive;5 but when one moves to cases of less urgency such as deciding which school to go for
the Ph.D. degree, rationality requires taking care of how probable it is that one’s goals will be
achieved and to what extent, as recommended in Bayesian decision theory.

Urgency—or lack of time for deliberation—is only one of the features of situations that tell
against deliberation and tend toward the external condition (A) for rationality. Lack of the
agent’s ability to properly deliberate is another. The agent’s aversion to the very process of
deliberation is yet another.

Now we are in a position to solve the regress problem. When one regresses from a decision
problem to higher and higher-order decision problems, there accumulates the force that pulls one
toward the external condition (A) for rationality. The time for deliberation is less and less and the
materials to deliberate over are more and more, until, first, rationality drops all internal
conditions except (B) and, second, the external condition (A) for rationally becomes in force.
That is, the move toward higher and higher-order decision problems will eventually create a
situation in which rationality requires only conditions (A) and (B). That is one extremal end of
the spectrum of rationality. Complete lack of urgency, unlimited capacity of deliberation, and
unbounded love for deliberation jointly correspond to the other extremal end of the spectrum of
rationality, which requires full Bayesian rationality.

Here is an example. It is 5:00 p.m. when hungry Bob finds that there is little food left in his
house. He wants to have a satisfying dinner. To go grocery shopping today, or to stay at home,
that is the (first order) question. The result of going today is easy to imagine for Bob, who is
familiar with what the store has to offer. He has to compare it with the result of staying at home,
which requires him to work out possible ways to dine with the scarce food available in the house.
Searching in the kitchen for a while, he finds dried penne and mullet roe, a kind of fish roe
considered a delicacy in Taiwan. That reminds him of an Italian inspired Japanese dish, called
mentaiko pasta, which is made of spaghetti and karashi mentaiko (i.e., spicy cod roe). Bob finds
the possibility of cooking (or experimenting with) his own fish roe pasta without grocery
shopping today. Then he asks: “should I continue working out possible ways to dine with the
food available in the house?” That is a second-order question. Now it is 5:20 p.m., 20 minutes
since he started to think about whether to go grocery shopping today. He gets quite a bit more
hungry now than 20 minutes ago (partly because imagination about food causes hunger), so it
becomes very difficult to deliberate. At this point, the situation reaches one end of the spectrum
of rationality, which requires only conditions (A) and (B). Now, the first strategy that Bob thinks
of is this: stop working on the second-order problem, solve the first order problem by choosing
one of the specific options currently on the table (i.e., “going grocery shopping” and “cooking
fish roe pasta”), and choose the option that will lead to an earlier dinner. As it turns out, Bob
adopts that strategy, (explicitly or implicitly) following the recognition heuristic. In this
situation, he does so rationally just in case conditions (A) and (B) hold.

7 Conclusion

Conlisk (1996), when commenting on the regress problem of deciding how to decide,
complained that such an important problem receives so little treatment in the literature. Eighteen
years later the situation has not much improved. I submit that we highlight the importance of the
regress problem by testing any theory of rationality in terms of how well it solves the regress
problem of deciding how to decide. According to the test, the Bayesian theory of rational choice
cannot be the correct theory. Nor can the theory of adaptive rationality that aspires to replace it.
If those two popular theories are both more or less on the right track, then the account of
rationality I sketched above seems to be a first step toward the simplest theory that solves the
regress problem and accommodates those two theories as extremal ends in the spectrum of
rationality.

Acknowledgments I am indebted to participants of the final Workshop on the Frontiers of

Rationality and Decision conference (2012 August, University of Groningen), especially Jeanne
Peijnenburg and Jan-Willem Romeijn for stimulating questions. I am also indebted to Alan
Hájek and Nathan Pensler for discussions, and to an anonymous referee for encouraging me to
clarify some of the points I made.

1 If beliefs and desires are reducible to preferences, Bayesian rationality merely requires
coherence among preferences.

2 There are at least three possible positions: (Infinitism) We can legitimately have an infinite
chain of justifications, in which each belief is justified by a belief that has not been used for
justification, such that the justification relation holds independently of whether the agent actually
adduces the justifying belief. Namely, we can legitimately have an infinite regress of justification
relations, although it is impossible to have an infinite regress in which one actually adduce
beliefs for justification for infinitely many times. (Foundationalism) The regress of justification
relations stops at a certain finite point, reaching beliefs whose justifications do not require further
any beliefs. (Coherentism) The linear, chain-like picture of justification is wrong; instead,
justification of a belief consists in the belief’s being a member of a coherent set of beliefs.

3 See Gigerenzer and Goldstein (1996); Gigerenzer and Todd (1999)

4 Note that the point I make in this paragraph is not specific to any particular heuristic. The point
applies to, for example, the satisficing heuristic proposed by Simon (1955), one of the earliest
pioneers of adaptive rationality.

5 See Lin (2013) for a theory of qualitative decision in which binary beliefs play an important
role. The theory is developed in parallel to Savage (1954) foundation for subjective expected
utility maximization.

IF ALL KNOWLEDGE IS EMPIRICAL, CAN IT BE NECESSARY?

Ísak Andri Ólafsson
April 2012
Department of Social Sciences
University of Bifröst

… A brief digression on the variety of accounts of justification will now be put forth to
demonstrate how notions of justification can differ when confronted with regress arguments.
Foundationalism, coherentism and infinitism are three accounts that will be illustrated along with
a few critiques.

Foundationalism, which both Chisholm and Ayer endorse, is the belief that the endless chain
begins with a justified belief that is not supported by another belief, thus making it a basic
justified belief. Foundationalists claim either that there are beliefs that do not need justification
or that there are beliefs that are justified in themselves. From there it follows that a belief is
justified if it is a basic belief, it is justified by such a belief, or it is justified by a chain of beliefs
that are not basic, but rest upon a basic belief. It is the hope of foundationalists that ultimately,
all beliefs can be traced back to basic beliefs. This position cannot adequately answer the
skeptical challenge. At its best, it can only capitulate to it. However basic a belief might be, there
is no reason to suppose that they are indefinitely true (Fumerton, 2010). Another criticism
directed toward foundationalism is that the foundationalist has two options when questioned
about a supposedly basic belief. The options are that the belief is either true or it is not. If the
foundationalist says that the belief is true, he is still stuck in a regress argument, if he says that it
is not true, his view is arbitrary (Klein, 2010).

Coherentism, another stance against the regress argument, states that the reasoning chain may
eventually form a circle, thus the original justification for a belief justifies it in the end. The
holistic nature of coherentism is appealing; Quine endorses this kind of a position in his paper
Epistemology Naturalized when he talks about rational reconstruction. (Quine, 2004, pp. 294-
295). This however rests upon the ridiculous notion that circular arguments are an accepted way
of reasoning, where “p justifies p”. It seems obvious that this too ends in a regress argument (p
justifies p, which justifies p, which justifies…) and furthermore that the statement “p justifies p”
provokes the question of p itself (Kvanvig, 2007).

A third account is infinitism, which argues that the chain can continue ad infinitum. However,
there is no reason to suppose that there is some justification found within the endless chain
although, because our minds are finite, the endless chain must be endless and as such, infinitism
cannot provide a final conclusive answer to the regress argument. It could thus be argued that
knowledge does not require the issue at hand to be settled completely and a finite subset of the
endless chain of reasons will be adequately close to knowledge. However, within the current
framework of epistemology, knowledge requires a final settlement. Infinitism cannot conclude a
final settlement of that nature, when working within that framework, since it does not have a
strong enough argument to defy the regress argument. It capitulates to skepticism at least as
much as the fallibilist foundationalist (Klein, 2010).

Université de Montréal
Défense et illustration de lʼinfinitisme épistémique
par
Marc-André Lévesque
Département de philosophie
Faculté des arts et des sciences

Abstract
"This dissertation focuses on the problem of epistemic regression which questions the possibility
of justification. For this reason weʼll take interest in the details of epistemic regression and in the
solutions that different critics offer to solve the problem. Generally, two positions oppose each
other : foundationalism and coherentism. The first one proposes to stop the regression at a
foundation, as the second one takes concern about the coherence shared amongst the beliefs.
However the purpose of this dissertation is to present and defend a third position : infinitism.
Introduced in the 1990ʼs by Peter Klein, infinitism is one of the most recent theories of
justification. Although it is quite appealing, Kleinʼs theory is not very popular and few people
defend this position. Infinitism offers to solve the problem of regression by basing the
justification of the beliefs on series of infinite and non repetitive reasons. Consequently,
infinitism reverses the initial problem because infinite regression is often perceived as an issue
for knowledge and a source for scepticism. Our goal is to demonstrate that infinitism is the best
way to solve the problem of epistemic regression. Therefore, weʼll synthesize the arguments in
favor of infinitism and that will mark out three types of infinitism from which weʼll retain one,
an impure form of infinitism, as best suited to answer the problem of regression. Finally weʼll
respond to the main oppositions to infinitism in order to demonstrate that it is in fact a viable
theory of justification.

Keywords: philosophy, epistemology, epistemic regress, epistemic justification, infinitism

INFINITE JUSTIFICATION
San Diego State University
Edward Andrew Greetis
Spring 2012

ABSTRACT OF THE THESIS

In this thesis, I examine the structure of epistemic justification and defend a theory of epistemic
justification termed “infinitism.” The structure of justified belief is commonly analyzed by
beginning with skeptical challenges to justification. Keeping with this method, I begin my
argument for infinitism by examining a skeptical challenge termed “the regress paradox.” I argue
that infinitism is the only theory of epistemic justification that can solve the paradox, which
entails that ceteris paribus infinitism is preferable to other theories of epistemic justification. I
then examine an argument against infinitism called the modus ponens reductio, which is
countenanced by several philosophers as their main reason for rejecting infinitism. I argue that
the scope of the reductio is mistaken: the reductio (slightly modified) applies to the leading
epistemic theories of justification as well as infinitism. Thus, I show that the reductio is a general
skeptical problem for all theories of justification to face. This shows that the leading epistemic
theories of justification are not preferable to infinitism since the reductio (slightly modified)
applies to them as well and because infinitism is the only theory that can solve the problem of
regress. However, if the reductio is correct, it might seem that either infinitism or skepticism
(i.e., only few, if any, of our beliefs are justified) is correct. To support infinitism, I present a
solution to the reductio, which shows that it is not a problem for any theory of epistemic
justification. The reductio should therefore not convince us of skepticism. Since the reductio is
the main reason for rejecting infinitism and since infinitism is the only theory that can solve the
regress paradox, I conclude that epistemic infinitism provides the most tenable account of
justified belief, and we therefore ought to justify our beliefs according to the infinitist structure
of justification.
TABLE OF CONTENTS
PAGE
ABSTRACT..............................................................................................................................v
LIST OF FIGURES ............................................................................................................... viii
ACKNOWLEDGEMENTS..................................................................................................... ix
CHAPTER
1 EPISTEMIC JUSTIFICATION.....................................................................................1
1.1 Foundationalism.......................................................................................................5
1.2 Coherentism.............................................................................................................8
1.3 Infinitism..................................................................................................................9
1.4 A Brief Overview...................................................................................................10
2 A BRIEF HISTORY OF INFINITISM.......................................................................12
3 MODERN INFINITISM..............................................................................................15
3.1 Klein's Infinitism....................................................................................................15
3.2 Fantl's Modest Infinitism.......................................................................................22
3.3 Completion Infinitism............................................................................................25
3.4 Impure Infinitism: Aikin's Theory .........................................................................26
4 THE REGRESS PARADOX.......................................................................................31
4.1 The Regress Paradox and Why Infinitism Is the Only Theory That Can
Solve It ...................................................................................................................31
4.2 Responses to Klein and Fantl.................................................................................32
4.3 Analyzing the Regress Paradox .............................................................................36
4.4 The Sellarsian Dilemma and Bonjour's Foundationalism......................................39
4.4.1 The Sellarsian Dilemma................................................................................40
4.4.2 Bonjour's Foundationalism...........................................................................42
4.5 Why Foundationalism Cannot Stop the Regress ...................................................45
4.5.1 What the Concept/Term Distinction Demonstrates ......................................47
4.5.2 Howard-Snyder and Coffman's Response ....................................................52
4.6 Conclusion .............................................................................................................54
5 THE MODUS PONENS REDUCTIO..........................................................................55
5.1 The Modus Ponens Reductio Ad Absurdum of Pure Infinitism.............................56
5.2 How a Modified MPR Applies to Foundationalism and Impure Infinitism..........58
5.3 How a Modified MPR Applies to Coherentism.....................................................68
6 A SOLUTION TO THE MPR.....................................................................................70
6.1 Infinitism and the Normative MPR Worry ............................................................70
6.2 Response to the MPRc and MPRc* ........................................................................71
6.3 Necessary Conditions of Justification....................................................................78
REFERENCES .......................................................................................................................83

LIST OF FIGURES
PAGE
Figure 1. Bonjour’s account of basic beliefs ...........................................................................44
Figure 2. Why Bonjour’s basic beliefs do not end the regress ................................................51

ACKNOWLEDGEMENTS
I would like to express my indebtedness to the many people who helped make this thesis
possible. I greatly benefited from countless helpful discussions with Professors Steve Barbone
and Angelo Corlett. I would also like to thank Professors Steve Barbone, Angelo Corlett, and
Glen McClish for their many helpful comments on drafts of this thesis. Thanks also to my family
for their unwavering encouragement. Finally, I would especially like to thank my wife, Liana, for
her constant support, helpful comments on and discussions about this thesis, and for helping me
to persevere.

CHAPTER 1

EPISTEMIC JUSTIFICATION

The human mind contains an intricate set of beliefs—some occurent, some dispositional, some
unconscious. We often, and should, critically examine our set of beliefs to determine which, if
any, are worthy of accepting as true, i.e., worthy of belief. During this self-reflection, we attend
to our beliefs and their contraries, as well as newly considered beliefs to determine whether they
are worthy of belief, i.e., justified or not. But what makes a belief justified? That is, how do we
determine which beliefs are worthy of belief?

To answer this question, we need an account of justified belief. This account will inform us
which beliefs, if any, are rational to hold as well as which beliefs are likely to be true.
Accordingly, it is a preponderant view among philosophers, at least since Plato’s Theaetetus, that
knowledge requires justified, true belief rather than simply true belief.1 Thus, an account of
epistemic justification will normatively guide our self-reflection, showing us which beliefs are
rational to hold, which beliefs likely count as knowledge, whether we can attain knowledge, and
how we can attain knowledge (if we can).

As justified belief is one of the most important and discussed topics in philosophy (in part, for
the reasons given above), accounts of justified belief abound. Today, there are two leading
accounts: foundationalism and coherentism. A third account—a theory that I believe warrants
further examination—is infinitism, which until recently has been almost universally rejected. I
will adumbrate each of the three theories below, but before doing so, I want first to explain what
kinds of beliefs and justification I discuss in the essay. This will not only help to set up my
project, it will also help to set up the motivations for the three theories of justification.

First, the beliefs I discuss are defeasible propositional beliefs. I discuss only defeasible beliefs
because if a belief is certain or infallibly known (if there are beliefs known with certainty or
infallibility), then it may not be in need of justification (if “certain” means something like
indubitable), and I assume that knowledge and rational belief do not require certainty (I further
explain why I only discuss defeasible beliefs in §1.1).2 This view—holding that justified beliefs
are defeasible and that knowledge and rationality only require defeasibly justified beliefs—is
often termed “fallibilism.” Next, I discuss only propositional beliefs. A propositional belief is a
belief—that is capable of being expressed propositionally—that such and such is the case. For
example, the belief that “I see a red ball” is a propositional belief. When knowledge is discussed
in philosophy, it is most often propositional, i.e., knowledge that, as opposed to, say, procedural
knowledge or knowledge of how to do some procedure. But why focus only on propositional
knowledge? As Ernest Sosa (2003a) notes, by focusing on propositional knowledge, “we are not
artificially isolating some idiosyncratic sector of the field of human knowledge” since
propositional knowledge permeates most forms of knowledge (p. 102). For example, knowing
how to swim requires much propositional knowledge and may even be expressed propositionally
(e.g., a swimming coach can coach others by using propositional statements even if the coach
lacks procedural knowledge, i.e., the coach can no longer swim [Sosa, 2003a, pp. 100-102]).

For the most part, however, I will attempt only to discuss justified propositional belief rather
than knowledge. I do this for several reasons. The main reason is simplicity; knowledge is a
recalcitrant concept that would take an immense amount of space and time to analyze properly.
Instead, by focusing solely on justification, I can properly analyze the concept. Another reason
for not discussing knowledge is that it is a deeply problematic concept. As Laurence Bonjour
(2003a) asks, it is generally thought that knowledge requires that a belief be justified to an
adequate degree, “But what degree of justification [is required to attain knowledge]?” (p. 21).
Any degree we choose seems completely arbitrary. This does not, however, diminish the worth
of an account of justified belief: the worth of justified belief does not solely rely on knowledge
(as stated above), and, if knowledge is possible, it seems to require justified belief—even if we
cannot provide sufficient conditions for knowledge, this does not entail that we cannot attain
knowledge or that we cannot indicate several necessary criteria for knowledge, e.g., an account
of justified belief.

My next task is to explain the kind of justification I will discuss and the problems that surround
this sort of justification. Considerable disagreement exists on what justifies a belief, what
structure justification takes, which—if any—of our beliefs are justified, and so on.

“What justifies a belief?” seems to be a matter of the reasons we have for the belief. Peter Klein
claims that, although there may be other forms of knowledge, “distinctive adult human
knowledge” requires reasons to justify beliefs.3 There may be some sense of knowledge where
thermometers “know” what temperature it is, and animals “know” that water is necessary for life,
but “carefully examining our beliefs in order to determine which, if any, deserve to be
maintained” (Klein, 2007b, p. 4), that is, holding beliefs produced and sustained by reasoning, is
distinctive of adult humans. Sosa (2001) terms this form of knowledge “reflective knowledge,”
and it is this form of knowledge that Keith Lehrer (2000a) uses “acceptances” (i.e., accepting
something with the purpose of attaining truth and avoiding error) rather than mere beliefs to
describe. Apart from knowledge, epistemic rationality seems also to require that we provide
good reasons for our beliefs to justify them.4 For example, we generally regard the belief that one
will win the lottery as epistemically irrational because there are no good reasons to think that one
will win the lottery. However, a person’s belief that he will win the lotto could be thought
epistemically rational if the person had good reasons for thinking he will win, e.g., he saw the
winning numbers and he was told that the numbers he saw were the winning numbers by a
credible source.

According to Bonjour (2006), this notion—that “epistemic justification requires that the believer
have a good reason for thinking that the belief in question is true”—is a main tenet of
“internalism” (p. 755). Another key tenet of internalism is a requirement on what it is to “have” a
reason for a belief. The internalist claims that to “have” reasons requires that they be cognitively
available to us, i.e., we must have cognitive access to our reasons.

Alternatively, the “externalist” claims that we do not need to be aware of the justification of our
beliefs. And whereas internalism holds that beliefs are justified by the reasons we have for the
belief, “externalism” is the contention that a belief can be justified regardless of the reasons we
have for the belief. In other words, the internalist holds that for a belief to be justified, a person
must be able to reflect on the reasons that justify the belief, i.e., be cognitively aware of the
reasons for a belief and hold the belief because of the reasons supporting the belief. Rather than
the simple producing of truth and avoidance of error, when we inquire, we seek not only truth,
but also conceptual understanding and an understanding of the evidence we have for these
concepts. Furthermore, when we inquire, we endeavor to find reasons to believe that the beliefs
in question are true or false, and we want to be aware of these reasons to satisfy our inquiry.
Rather than simply being reliable producers of truth, we strive to be aware of why we should
hold certain beliefs and what makes the beliefs rational and likely to be true. Internalism captures
this central aspect of justification, which is what I will discuss in the remainder of the essay. I do
not claim that there is no other sense of justification; in fact, I believe that externalist
epistemology explains important aspects of justification. However, I think that internalism aims
to explain a central—perhaps the most central—aspect of justification: justified beliefs require
cognitively accessible reasons. I will therefore analyze only internalist theories of justification to
provide a plausible account of the distinctly internalist aspects of justified belief.

My primary focus will be the structure of internalist justification (since all justification discussed
will be internalist, I will hereafter drop the qualifier). To discuss the structure of justification, it
is helpful to invoke the distinction between propositional and doxastic justification.
Propositional justification is an epistemic property of propositions, whereas doxastic justification
is an epistemic property of concrete doxastic belief states. When S is propositionally justified in
believing p at time t, p is justifiable for S at t: reasons are available to S that would justify p were
S to believe p and provide—and believe—those reasons. (I say that p is “justifiable” rather than
“justified” because S has not yet provided reasons to justify p.)5 Whereas, when S is doxastically
justified in believing p at t, p is justified for S at t: p is propositionally justified for S at t, S
believes p, and S provides—and believes—the relevant available reasons for p. Doxastic
justification, rather than propositional justification, is generally thought necessary for knowledge
and rationality, but propositional justification provides the conceptual structure of reasons that
we should strive to attain in order to justify our beliefs. As argued above, reasons provide
justification for beliefs (and propositions). But how does this work? Do we continue to provide
reasons infinitely? Must we stop the chain of reasons? Can reasons justify themselves, that is,
can chains of reasons be circular? Are chains of reasons linear or are they structured holistically?
Answering questions such as these take up the bulk of the essay and for good reason: they are
some of the most discussed and difficult questions concerning justification. Roughly, these
questions constitute the problem of regress (I further explain the problem of regress in chapter 4),
which, along with related problems, is perhaps the most important for epistemology.6 Thus, the
structure of justification is no small task, which is why it is almost exclusively my focus for the
extent of the essay.
Now that I have set out the scope and some of the aims of the essay, I can briefly adumbrate the
theories of epistemic justification I will analyze.

1.1 FOUNDATIONALISM

First, traditional foundationalism7 holds that “basic beliefs,” which are spontaneous and self-
justified or non-inferentially justified (i.e., justified without the support of other beliefs),
constitute the basis of justification. In other words, basic beliefs are justified and transfer their
justification through logical inference to other beliefs, such that a belief is justified just in case it
is either basic or is inferred from a basic belief.

Traditional foundationalism is either infallible or “modest.” First, infallible foundationalism

holds that basic beliefs are infallible. When person S holds the basic belief “I have a headache,”
S infallibly knows that S has a headache. However, infallible foundationalism is implausible, so I
will not discuss the theory. The main reason for this implausibility is the fact that not many
beliefs can be inferred from infallible beliefs (if there are any infallible beliefs). For example,
empirical beliefs are not infallible: S could be told that a frying pan is very hot and upon
touching the frying pan, believe that it is very hot (a basic belief), when it is actually cold—
perhaps the expectation of a hot pan caused the belief, but the belief is still properly basic (it is a
non-inferentially justified immediate awareness of experience) (Feldman, 2003, p. 55).8 Next,
“modest,” or fallible foundationalism, is the contention that basic beliefs are at least prima facie
justified but can be mistaken (hereafter, all foundationalist theories discussed will be modest, so I
will drop the qualifier). That is, foundationalists hold that even though basic beliefs are fallible,
they are worthy of our trust, or “innocent until proven guilty” in Lehrer’s (2000a, p. 71) terms.

Next, meta-justificatory foundationalists argue that basic beliefs are justified because they have
some truth-conducive property, P. For example, a foundationalist could argue that basic beliefs
reliably produce true beliefs, that is, basic beliefs have a truth-conducive property, P, which
justifies them.

Thus, for all foundationalists, beliefs are doxastically justified if and only if they are basic or
inferred from basic beliefs, and propositions are propositionally justified if and only if they are
either basic or inferred from a basic proposition.

Now that we have a rough outline of foundationalism, we can proceed to describe a few details
of the theory, namely, basic beliefs and the motivation behind basic beliefs.

As stated above, basic beliefs are sometimes described as being “self-justified,” but because of
the apparent implausibility of a belief’s justifying itself,9 most contemporary foundationalists
explain basic beliefs by arguing that some beliefs—basic beliefs—are justified simply by our
immediate awareness of experience. For example, we are immediately aware of experiences such
as “having a headache” and “being appeared to redly,”10 which, according to foundationalists,
justify the basic beliefs that arise from attending to these experiences, namely, the beliefs that “I
have a headache” and “I am appeared to redly.” Because of the immediacy of this awareness, it
is often described as “the given” or “direct apprehension.”
Furthermore, because an awareness of experience is non-inferential and nonpropositional, it
purportedly can stop a regress of propositional and inferential beliefs (a regress of reasons). This
is important because the main reason for accepting foundationalism seems to be the regress of
reasons argument, which is a response to the problem of regress.11 The foundationalist holds that
a regress of reasons cannot be infinite or circular, so the regress must somehow stop, which is the
motivation behind basic beliefs. Sosa provides a helpful illustration of how the foundationalist
structure of justification is supposed to stop a regress of reasons by likening it to a pyramid. The
basic beliefs are the base of the pyramid (the support for all justification), and they transfer their
justification to other beliefs (through inference) upward to the apex of the pyramid, i.e., the
belief being justified (Sosa, 1980, p. 5).

Although we have seen how foundationalists stop the regress, it may be helpful to have a
concrete example. Consider Simpson, the avid drinker. Simpson believes that “it is morning.” If
questioned about the justification of this belief, among the various beliefs that he will provide as
reasons to believe that “it is morning” is true are the following: “since I drink most nights, when
I have a headache, it is usually morning” and “I have a headache.” Since Simpson is a
foundationalist, if questioned why he holds the belief “I have a headache,” he will claim that he
believes this on the basis of his experience of having a headache and that the experience
sufficiently justifies his belief, i.e., since awareness of “having a headache” is non-propositional
and sufficiently justifies his belief, he is done providing reasons.

If Simpson is a meta-justificatory foundationalist, he will claim that his belief that “I have a
headache” is sufficiently justified because it has P (P can be, e.g., the contention that beliefs
formed by an immediate awareness of experience are likely to be true).

1.2 COHERENTISM

Unlike foundationalists, coherentists hold that all beliefs are justified by virtue of their relation to
other beliefs—there are no basic beliefs. Also unlike foundationalists, coherentists hold that
justification is non-linear or holistic: beliefs are not prior to others (recall that the foundationalist
holds that basic beliefs transfer their justification though inference to other beliefs linearly);
instead, beliefs mutually support each other in one’s system of beliefs. But like meta-justificatory
foundationalists, coherentists hold that a belief is justified in virtue of having a property, P—in
this case, the property of cohering with one’s system of beliefs. More precisely, coherentists
purport that a belief that p is doxastically justified for person S at time t just in case p coheres
with S’s system (or system’s)12 of beliefs, and a proposition p is propositionally justified for
person S at t just in case p coheres with a coherent system of propositions available to S at t.13
Thus, for the coherentist, the coherence of a system of beliefs is supposed to be indicative of the
truth of its members: truth-conduciveness arises from a system of mutually supporting beliefs.

The kind of coherence, or mutual support, required is usually stated to be logical, explanatory,
and so on.14 Further, the kind of system of beliefs that is required is usually the set (or subsets) of
the person’s beliefs. Importantly, since we are discussing internalist theories, for a belief to be
justified for S, S must be aware that the belief coheres with S’s system of beliefs, which entails
that S must be aware of her system of beliefs.
Keeping with the example given in §1.1, Simpson (now a coherentist) will justify the belief that
“it is morning” by claiming that it coheres with his system of beliefs. Upon examining his system
of beliefs, Simpson will realize that “it is morning” is explained by several of his other beliefs,
e.g., his beliefs about time, light, hemispheres of the earth, and angles; “it is morning” explains
several of his beliefs, e.g., why his alarm clock is sounding; “it is morning” logically follows
from some of his other beliefs; and so on. Thus, because Simpson is aware that “it is morning”
coheres with his system of beliefs and will perhaps increase the coherence of his system of
beliefs, his belief that “it is morning” is justified according to coherentism.

1.3 INFINITISM

The theory I espouse is infinitism. Infinitism holds that for S to be completely justified in the
belief that p, it is necessary (but insufficient) that p is the first member of an infinite series of
non-repeating reasons for p. Thus, like foundationalists, infinitists hold that justification is the
result of a non-repeating linear series of reasons (not holistic or circular). And like coherentists,
infinitists hold that only beliefs can justify beliefs (there are no basic beliefs).15 Thus, infinitism
holds that it is necessary that for p to be propositionally justified for S at t, p must be the first
member of an infinite chain of non-repeating reasons for p available to S at t, and p is
doxastically justified (as a necessary condition) for S at t to the extent that S provides reasons for
p. In a non-dogmatic fashion, the infinitist holds that although p may never be completely
doxatically justified for S (since no one can provide infinite reasons), providing more reasons for
p further justifies p.

Infinitists therefore hold that Simpson (§1.1) can, and in some epistemic circumstances should,
provide more reasons for the belief that “I have a headache”—the chain of reasons has not
stopped. For example, a skeptic may question Simpson’s belief, he may examine his beliefs in
light of skepticism, or he may simply attempt to justify his belief. Simpson may never
completely justify the belief “I have a headache,” but his belief becomes more justified by
providing more and more reasons for the belief. For instance, Simpson can support “I have a
headache” by claiming that “I drank last night” and “when I drink, I often get a headache,” etc.
Simpson can also further justify his belief that “I have a headache” by providing reasons for the
belief that “I exist,” which would support the belief “I have a headache” by supporting the fact
that there exists an “I” that is capable of having headaches, and so on.

(It should be noted that the theories adumbrated above are not necessarily incompatible. What I
have presented thus far are the pure versions of the theories. I will, for the most part, discuss the
pure versions of the theories because the problems I discuss will generally be solved by one or
two main elements of the theories discussed and not by a combination of the elements of the
theories discussed. However, as we shall see, elements of the theories can be combined, as many
have done.)16

1.4 A BRIEF OVERVIEW

Beginning in chapter 2, I provide a brief exposition of the history of infinitism. I believe that this
helps to illustrate the importance of analyzing infinitism—since it has been almost universally
and often unreflectively dismissed—and helps to set up the problems faced by infinitism and all
theories of justification. In chapter 3, I explicate the leading versions of infinitism and several
arguments for the view, which is necessary for analyzing the theory and each of its variants. In
chapter 4, I examine the regress paradox. I argue that infinitism is the only theory of epistemic
justification that can solve the paradox, which entails that ceteris paribus infinitism is preferable
to other theories of epistemic justification. In chapter 5, I examine an argument against infinitism
called the modus ponens reductio, which is countenanced by several philosophers as their main
reason for rejecting infinitism. I argue that the scope of the reductio is mistaken: the reductio
(slightly modified) applies to the leading epistemic theories of justification as well as infinitism.
Thus, I show that the reductio is a general skeptical problem for all theories of justification to
face. This shows that the leading epistemic theories of justification are not preferable to
infinitism since the reductio (slightly modified) applies to them as well and since infinitism is the
only theory that can solve the problem of regress. However, if the reductio is correct, it might
seem that either infinitism or skepticism (i.e., only few, if any, of our beliefs are justified) is
correct. Thus, in chapter 6, I present a solution to the reductio, which shows that it is not a
problem for any theory of epistemic justification. The reductio should therefore not convince us
of skepticism. Since the reductio is the main reason for rejecting infinitism and since infinitism is
the only theory that can solve the regress paradox, I conclude that epistemic infinitism provides
the most tenable account of justified belief, and we therefore ought to justify our beliefs
according to the infinitist structure of justification.

CHAPTER 2

A BRIEF HISTORY OF INFINITISM

As stated in the first chapter, infinitism has been discussed and almost universally rejected
throughout much of philosophical history. This discussion (and rejection) of infinitism seems to
begin at least with Aristotle.17 In the Posterior Analytics, Aristotle argues for foundationalism by
arguing that if some knowledge results from demonstration, then some knowledge does not.
While arguing for this claim, he dismisses infinitism as an option by claiming that “it is
impossible to go through infinitely many things” (Post. An., 72b10)—presumably because it
takes time to provide reasons and we do not have infinite life spans or infinite intellects.
Aristotle’s objection is commonly termed the “finite intellect” objection and is still commonly
advanced, e.g., Bonjour (1985) states that the holding of an infinite number of beliefs “is
impossible for a creature with only a finite mental capacity and a finite brain” (p. 24).

Aristotle also seems to propound a second argument against infinitism. In the Metaphysics, he
states:

These people [some of those who believe, or merely profess, that man is the measure of
all things] demand that a reason shall be given for everything; for they seek a starting-
point, and they wish to get this by demonstration, while it is obvious from their actions
that they have no conviction. But their mistake is what we have stated it to be; they seek a
reason for that which no reason can be given; for the starting-point of demonstration is
not demonstration. (1011a12-13)
Aristotle’s argument is commonly termed the “no starting point” objection and is most famously
presented by Sextus Empiricus (1990), who writes,

[the] regress ad infinitum is that whereby we assert that the thing adduced as a proof of
the matter proposed needs a further proof, and this again another, and so on ad infinitum,
so that the consequence is suspension [of assent], as we possess no starting-point for our
argument. (pp. 166-167)

In other words, the no starting point objection asserts that to justify a proposition, it must be
inferred from a justified proposition because justification is a property of propositions that is
passed on by valid inference. If a chain of inferences lacks a starting point (as in infinite
justificatory chains of propositions), then there is no justification to transfer between
propositions; hence, infinitism is rejected as an account of justification.

Thus, Aristotle seems to begin the long history of rejecting infinitism. (Although some
Pyrrhonian skeptics, including Sextus, seemed to accept infinitism as required for knowledge,
they used this requirement to argue for the rejection of the entire project of knowledge.)18

Perhaps the first to argue that infinitism is the correct analysis of actual human knowledge or
justification (as opposed to the Pyrrhonians) was Charles Peirce. Peirce argued for what Scott
Aikin terms an “impure infinitism.” Roughly, his view is “impure” because he held that infinite
belief chains are required for knowledge (there must be an inferential feature to all epistemic
support) (Aikin, 2009, p. 79),19 but the infinite chains are also supported by non-cognitive
support (non-doxastic support is required for premises to exist and be justified).20 (I further
explicate “impure infinitism” in §3.4.) Thus, Peirce seemed to allow for structures of justification
other than endless sets of non-repeating reasons but required infinite sets of non-repeating
reasons for justification and knowledge. However, since Peirce’s infinitism was motivated by his
semiotic theory, which he thought to be flawed leading him later to revise the theory, he thereby
lost the motivation for his infinitism. I will therefore not provide a detailed account of his view.21

According to Aikin, subsequent to Peirce, Popper (1935) may plausibly be read as an infinitist
and several other philosophers following Peirce have expressed sympathy for the view, for
example, Reichenbach (1952), Foley (1978), and Hardwig (1988). More recently, philosophers
have explicitly argued for infinitism and have begun developing infinitist epistemologies,
including Peter Klein, Jeremy Fantl, Scott Aikin, John Turri, Jeanne Peijnenburg, and David
Atkinson, among others.22

In the next the chapter, I differentiate and explain the several plausible formulations of, and
arguments for, infinitism.

CHAPTER 3

MODERN INFINITISM

3.1 KLEIN’S INFINITISM

To my knowledge, Klein provides, through a series of papers,23 the most complete extant
formulation and defense of infinitism.24 Because of this, and because many subsequent defenders
of infinitism base their view on Klein’s, I will present his formulation first.

Before explicating Klein’s infinitist formulation of the structure of justification, it is important to

recall that Klein is providing an analysis of what he terms “distinctive adult human
knowledge.”25 Furthermore, Klein notes that this kind of knowledge is “epistemically
responsible” (I explicate the requirements of epistemic responsibility when discussing doxastic
justification below).

Moving on to Klein’s theory of the structure of justification, Klein (2007b) begins by developing
an infinitist formulation of propositional justification: “a proposition, p, is propositionally
justified for S iff there is an endless series of non-repeating propositions available to S such that
beginning with p, each succeeding member is a reason for the immediately preceding one” (p.
11).26 That is, p is inferred from R1, which is inferred from R2, …, Rn is inferred from Rn+1,
such that each previous proposition is a reason for the proposition that follows it. Thus, like
coherentism, only reasons can justify propositions (and beliefs). Furthermore, like coherentism,
Klein holds that justification is an emergent property because it arises for a proposition when it is
a member of a set (i.e., an infinite set of non-repeating reasons beginning with the proposition in
question).

Already, we see that one objection to infinitism—the no starting point objection—is answered:
the no starting point objection begs the question against infinitism and coherentism, which both
hold that justification is a property that arises from a set of propositions and not a property that is
transferred from one proposition to another as the objection assumes.27 To use Sosa’s (1980)
analogy for foundationalism (§1.1), the objection assumes that justification must be like a
pyramid such that the base of the pyramid is the starting point of justification, which is
transferred through valid inference to the apex of the pyramid (the belief being justified).
However, Klein’s infinitism and coherentism hold that justification emerges when certain criteria
are met (e.g., when a belief has infinite reasons supporting it). Thus, the no starting point
objection simply begs the question against infinitism and coherentism by assuming that
justification needs a starting point.

Next, Klein develops an infinitist theory of doxastic justification. According to Klein, S is

doxastically justified in believing that p just in case S is epistemically responsible in believing
that p.28 Since holding an epistemically responsible belief depends on inter alia what S thinks is
required to justify a proposition, “Doxastic justification is parasitic on propositional justification”
(Klein, 2007b, p. 8). In other words, if S thinks infinitism is the correct theory of justification, S
will provide more and more reasons to justify the belief that p (whereas, e.g., a foundationalist
will attempt to find a basic belief). Furthermore, although complete doxastic justification makes
sense within Klein’s theory (i.e., providing every reason along an infinite path of reasons), no
belief is ever completely doxastically justified: it takes time to provide reasons, and no one has
an infinite intellect so no one can hold or provide an infinite number of non-repeating reasons.
However, with more reasons given for their beliefs, people become more justified—not because
their belief chain is approaching infinity but because they are providing more reasons for their
belief (Klein, 2007b, p. 10). In this way, justification is, according to Klein, a relational term. For
example, if S1 provides more reasons for the belief that p than S2, then S1 is more justified in
the belief that p than S2. The quantity of reasons one provides, according to Klein, will depend
on the epistemic context. Thus, S is doxastically justified in the belief that p just in case S
provides “enough” reasons for p along an endless path of reasons, where “enough” is determined
by epistemic context (Klein, 2007b, p. 10). However, it is important to note that context does not
settle justification: recall that no belief is ever completely justified and that justification is
relational, i.e., even if it is not required by the epistemic contextual parameters, S can provide
more reasons for the belief that p and thereby become more justified in the belief that p. 29 It is
also important to note that epistemic responsibility requires more than simply providing reasons;
beliefs must be “held for the ‘right’ reasons” and the epistemic agent must “strive to believe all
and only those propositions worthy of belief” (simply guessing will not produce justification)
(Klein, 2007b, p. 6).30

As we have seen, it is central to Klein’s infinitist theory of justification that an infinite set of
propositions be available to a person. But what are the conditions of availability? Klein

Klein (2007a) further clarifies this account by distinguishing between objective doxastic
justification and subjective doxastic justification. He states:

We can say that a belief is subjectively doxastically justified iff the belief is inferred from
another belief that is a reason for it regardless of whether the reason is propositionally
justified, and we can say that a belief is objectively doxastically justified only if the
reason is propositionally justified. Thus, if a belief is objectively doxastically justified, its
content is propositionally justified and the belief on which it is based is potentially
objectively justified (since there will be reason available for it). (p. 29)

Klein says “potentially” because the belief only becomes (at least partially) doxastically justified
when S actually provides a reason (including sotto voce) for the belief in question. For example,
if S provides reason r for the belief that p, p is objectively doxastically justified only if r is
propositionally justified and r is potentially objectively doxastically justified (i.e., justifiable)
since there will be a reason available for r (since it is propositionally justified), but S has not yet
cited that reason. According to Klein, it is this objective doxastic justification that is necessary
for knowledge.

As we have seen, it is central to Klein’s infinitist theory of justification that an infinite set of
propositions be available to a person. But what are the conditions of availability? Klein holds
that for a reason to be available to S, it must be both objectively and subjectively available.

First, a belief is objectively available only if there is “some normative or non-normative property
or, perhaps, a mixed property that is sufficient to convert a belief into a reason” (Klein, 1999, p.
299). For example, we might say that reason r is objectively available to S for belief p if r “has
some sufficiently high probability and the conditional probability of p given r is sufficiently
high” (Klein, 1999, p. 299). Since infinitism seems to be compatible with any formulation of
objective availability, the infinitist can choose whatever formulation turns out to be right. Klein’s
point is this: only certain propositions can function as reasons for other propositions. Thus, like
foundationalism, infinitism holds that there are features of the world that make a belief a reason,
which means that not all infinite chains of propositions provide justification. (What makes one
proposition a reason for another is a question for coherentists and foundationalists too, so
although it is an important epistemic question, it can be set aside for now.)

Second, reasons must also be subjectively available, that is, they must be properly connected
with a person’s beliefs. Humans have a myriad of beliefs (possibly infinitely many) and not all of
them are occurrent (i.e., not all of them comprise a conscious thought until they are “called
upon”). The infinitist could therefore claim that a belief is subjectively available if it is connected
to S’s beliefs in this non-occurent way or require propositions to be “armchair available,” i.e.,
propositions that are entailed by S’s current beliefs. Alternatively, an infinitist could hold that
“Roughly […] S believes p just in case S would affirm that p, or endorse p in another fashion—
perhaps sotto voce—in some appropriately restricted circumstances” (Klein, 1999, p. 300).
Similarly, the infinitist could require that proposition p is available to S if and only if “there is an
epistemically credible way of S’s coming to believe that p given S’s current epistemic practices”
(Klein, 2007b, p. 13). Klein provides a case of consulting the World Almanac as an example of
the latter. If S’s epistemic practices are such that, if contextually determined strictures required S
to know the capitol of Montana (or to use the capitol of Montana in a chain of reasons), S would
consult the Almanac. Thus, “the Almanac is a reliable source and it lists Helena as the state
capital” would count as “subjectively available” to S (Klein, 2007b, p. 13). Again, Klein claims
that infinitism is not committed to one formulation of subjective availability. His point in arguing
for the fact that propositions are subjectively available to S (whatever the correct view of
subjective availability is) is this: “nothing seems to prevent a finite mind from having an endless
set of propositions available” (Klein, 2007b, p. 13). This, coupled with the fact that Klein
requires only that there is an endless path of propositions and that S provides as many reasons on
this path as required by the epistemic context, sufficiently answers the finite mind objection.31

One final aspect of Klein’s theory is that his infinitism remains uncommitted on whether beliefs
cited as reasons for other beliefs are causes of those beliefs. Even if a kind of naturalistic
foundationalism were true, such that a mental state causes our first belief, which causes the next,
and so on (a finite processes), “it does not follow that the reasons for our beliefs are finite”
(Klein, 2007b, p. 11). Klein’s (2007b) view is therefore compatible with Aristotle’s argument
that if some knowledge is the result of demonstration, then some is not, where “result” is taken
most plausibly to mean a causal result (because the reason Aristotle has for his claim is that
inferring takes time and we cannot “traverse an infinite series” since we are finite) (p. 11).

We now have a sketch of Klein’s infinitism and can see some of the ways it avoids the traditional
reasons for rejecting the theory. However, why should we believe this model of justification
rather than, say, a foundationalist or coherentist model? Klein provides two main reasons: (1)
infinitism provides the only acceptable account of rational belief; and (2) infinitism is the only
theory that can solve the regress of reasons problem. Since the regress problem is, as Bonjour
claims, “perhaps the most crucial in the entire theory of knowledge” (§1), I devote the next
chapter to discussing it. Presently, however, I will provide a brief outline of Klein’s argument for
(1) and explain why it entails (2).

Klein holds that a rational belief is a belief that is held because of reasons. He argues that two
intuitive principles—the “Principle of Avoiding Circularity” (PAC) and the “Principle of
Avoiding Arbitrariness” (PAA)—necessary for forming rational beliefs lead to infinitism (while
excluding coherentism and foundationalism). The principles are as follows:

PAC: For all x, if a person, S, has a justification for x, then for all y, if y is in the
evidential ancestry of x for S, then x is not in the evidential ancestry of y for S. (Klein,
1999, p. 298, 2003, pp. 719 & 725, 2005a, p. 136)

PAA: For all x, if a person, S, has a justification for x, then there is some reason, r1,
available to S for x; and there is some reason, r2, available to S for r1, etc., and there is
no last reason in the series. (Klein, 2003, p. 726, 2005a, p. 136)32

Klein does not argue for PAC because he thinks it is obviously required for good reasoning. He
does say, however, that certain forms of coherentism purporting that a sufficiently large circle of
reasons is acceptable, seems to simply make it harder to detect the error in reasoning, but that
even a difficult to detect error in reasoning is an error nevertheless. PAC therefore rules out
several forms of coherentism (the other “holistic” (§1.2) forms are ruled out by PAA because
they are, as Sosa, Klein, Fantl, and others claim, really foundationalism in disguise).33

PAC captures our intuitions about circular reasoning, whereas PAA captures our intuition that
arbitrary beliefs, i.e., beliefs in which no reason is available, should not be held. Foundationalists
hold that either no reason is available for basic beliefs (hence, they must reject PAA) or that no
reason needs to be given for basic beliefs because they are self-justified (hence, they must reject
PAC because self-justification entails that “p therefore p,” which is circular reasoning). Thus, it
seems that foundationalists and coherentists must reject these intuitive principles. However,
Klein does not simply rely on intuitions with PAA. To support the principle, he defends it against
three objections: “reliabilism,” “meta-justifications,” and “harmless arbitrariness.” I will briefly
explain these objections and Klein’s responses in turn.

First, Klein rules out reliabilism, which holds that a reliable process, rather than reasons, justifies
a belief.34 As noted above (and in §1), we are discussing what Klein terms “adult human
knowledge” and what makes adult human knowledge distinctive is being able to provide reasons
for a belief. Other things may have knowledge that is not rational belief (e.g., animals or
Lehrer’s [2000a] “Mr. Truetemp”),35 but this type of knowledge is not relevant to our purposes
(it is “externalist”), so reliabilism is rejected.

Next, Klein responds to what Bonjour terms “meta-justifications,” which are justifications
purporting that because some beliefs have a certain property, say P, and beliefs with P are likely
to be true, one does not have to provide reasons for beliefs with P.36 (The remaining coherentist
theories belong in this group because they hold that coherent sets of beliefs have P [§1.2].)
However, Klein argues that the Pyrrhonian question is reasonable: “Why is having P truth-
conducive?” Klein (1999) claims that “Either the meta-justification provides reason for thinking
the base proposition is true (and hence, the regress does not end) or it does not (hence, accepting
the base proposition is arbitrary)” (p. 304). Thus, the meta-justification response fails. (I present
Aikin’s response to this argument in §3.4 and discuss several other responses to the argument,
including Aikin’s, in §4.2.)
Finally, Klein (1999, p. 305) addresses Stephen Luper-Foy’s (1990, p. 45) argument that it is
rational to accept arbitrary beliefs because we must if we are to attain our epistemic goal of
finding an accurate and complete picture of the world. Klein responds first by noting that if
infinitism is right, there is no reason to assume that we must accept basic beliefs to attain our
epistemic goal (i.e., Luper-Foy’s argument discounts infinite regresses as an option). Next, Klein
(1999) argues that Luper-Foy is using a prudential formulation of rationality and that it is not
even prudentially rational to accept basic beliefs: if there is no available reason for basic beliefs,
there is no reason to think they will contribute to the epistemic goal cited by Luper-Foy. Thus,
Klein concludes that the leading epistemic theories cannot accept the intuitive reasoning
principles PAA and PAC.

Moreover, PAA together with PAC entail infinitism, that is, that reasons for a belief be infinite
and non-repeating. Thus, Klein concludes that if any beliefs are justified, the structure of
justification must be infinitist. This entails that infinitism is the only theory that can solve the
regress paradox: the problem of regress asks how a belief can be justified (by providing endless
reasons, stopping the regress of reasons, and so on [§1]) and PAA and PAC entail that only
infinite chains of reasons can justify a belief (the chain of reasons cannot stop, cannot be circular,
etc.).

Klein’s arguments have sparked much interest and several other philosophers have begun to
defend the view. In the remaining sections of this chapter, I will discuss several infinitist views
that differ from Klein’s view.

3.2 FANTL’S MODEST INFINITISM

Like Klein, Fantl holds that complete justification for p requires at least one endless set of non-
repeating reasons but that no one is ever completely justified in a belief. Also like Klein, Fantl
holds that the reasons in an infinite justificatory chain must be adequate. For example, “1+1=2”
may be more justified than “she will show up at 8:00pm,” even if the latter is supported by more
reasons (i.e., the latter may need more reasons to be as justified as “1+1=2”). (Like Fantl, Klein
holds that reasons must be adequate, but it is unclear if he holds that different reasons provide
different degrees of justification as Fantl does.) However, ceteris paribus, the addition of another
adequate reason will always increase the justification of a proposition. A further point of
agreement is that Fantl does not present a theory of what makes a reason good or adequate—he
is simply concerned with the structure of justification. Fantl adds, however, that more than
simply providing adequate reasons for a belief is the ability to answer challenges to a belief,
which he claims is an important part of justification (more on this requirement below). A final
point of agreement between Fantl and Klein (a point Klein [2005a, p. 138 and endnote 16]
receives from Fantl)37 is that justification comes in degrees, so as the series of reasons increases,
so does justification. Fantl (2003) states that some beliefs “count as knowledge though only
justified to a degree somewhat less than complete justification,” which he claims makes his
infinitism “modest” (however, since most, if not all infinitists now accept this claim, I will drop
Fantl’s “modest” qualifier) (pp. 558-559). What degree of justification is required for
knowledge? Fantl claims that all fallibilist theories have trouble specifying a non-arbitrary
justification threshold—the infinitist could use epistemic context, or other criteria, but the
threshold is a problem for all theories, not just infinitism (cf. §1).
Now that I have shown some points of agreement, I will explain two crucial differences between
Klein’s and Fantl’s infinitist theories. First, for Fantl (2003), justification is not holistic or
“emergent”: “p’s justification derives from the reasons that support p, whose justification is in
turn derived from the reasons that support them” (p. 558 footnote 24). Second, an increase in the
degree of justification may not require reasons (e.g., some sensations might count), but as stated
earlier, providing adequate reasons, ceteris paribus, always increases justification and complete
justification requires an endless set of reasons.38

Besides developing a theory of infinitism, Fantl (2003) argues for the view by claiming that only
infinitism can account for two important attributes of the structure of justification:

1. The degree requirement: a theory of the structure of justification should explain why or
show how justification is a matter of degree.

2. The completeness requirement: a theory of the structure of justification should explain

why or show how complete justification [where there is no higher degree of justification]
makes sense. (p. 538)

As noted several times (e.g., §1), most epistemologists accept the degree requirement. For
example, William Alston (1986) states that “what justifies one is some evidence one has, one
will be more or less justified depending on the amount and strength of the evidence” (p. 25).
Conversely, the completeness requirement may be more controversial, but to reject the
requirement is to accept infinitism: for any degree of justification, one can always become more
justified.

Infinitism satisfies both requirements. First, Fantl claims that, ceteris paribus, the longer the set
of adequate reasons one has for a proposition, the more justified the proposition is, thus, the
degree requirement is satisfied. Second, complete justification, although impossible for humans,
makes sense: he claims that S is completely justified in the belief that p just in case S has an
infinite array of reasons for p. By “infinite array” he not only includes an infinite number of
adequate reasons but also—because even with an infinite chain of reasons, one may not be able
to answer all objections to a proposition—the ability to answer all objections to the proposition,
the reasons for the proposition, and the inferences used (Fantl, 2003, p. 557). Thus, infinitism
easily fulfills both requirements.

Alternatively, the leading theories of justification (variations of coherentism and

foundationalism) cannot meet these requirements. Fantl separates foundationalism into two
categories: “traditional foundationalism” (i.e., basic beliefs are self-justified, meaning that their
truth is enough to justify them sufficiently) and “meta-justificatory foundationalism” (i.e., basic
beliefs are justified in virtue of having some truth-conducive property, call it P, whether it is
reliability, coherence, etc.).39 (Note that “holistic” versions of coherentism fall into the latter
category and the other coherentist theories—where justification is transferred from one
proposition to another and the justificatory structure is literally circular—are implausible.40
Reliabilism also falls into the latter category: it claims that beliefs with P, i.e., beliefs produced
by reliable processes, are truth-conducive.) First, traditional foundationalism cannot meet the
degree requirement: inferences simply preserve the justification of basic beliefs,41 and self-
justified beliefs cannot admit of degrees. Fantl asks, In virtue of what can self-justified beliefs
admit of degrees? The foundationalist cannot say it is the basic belief’s varying reliability
without becoming a meta-justificatory foundationalist. The foundationalist can only claim, “it is
in virtue of the truth of each reason that it confers the degree of justification it does,” but Fantl
(2003) replies, “truth per se cannot determine which self-justifying reasons are more or less self-
justifying” (p. 544). Next, meta-justificatory foundationalism cannot satisfy the completion
requirement: if you come to have a reason that basic belief, b, completely satisfies truth-
conducive property P, then your belief will become more justified (thus, the regress continues
and foundationalist’s “complete justification” was therefore not complete). Thus, “meta-
justificatory foundationalists cannot plausibly satisfy the completeness requirement without
becoming de facto infinitists” (Fantl, 2003, p. 546). Fantl concludes that infinitism is preferable
to the leading epistemic theories because it can account for two important attributes of the
structure of justification (“completeness” and “degrees”).

3.3 COMPLETION INFINITISM

Another possible version of infinitism, a view we may call “completion infinitism,” requires that
in order for the belief that p to be justified for S, S must complete an infinite number of non-
repeating reasons for p. This view may lead to skepticism, as the arguments discussed thus far
suggest (cf. the Pyrrhonians). In fact, Klein (1998) claims that the requirement of completing
infinite chains of reasons for knowledge “would be tantamount to rejecting infinitism” (p. 920)
and that “no belief could ever be completely justified” (Klein, 2007b, p. 10). However, Atkinson
and Peijnenburg argue that infinite chains of reasons can be completed. Peijnenburg (2007)
states that “against infinitists I will show that beliefs may be justified by an infinite chain of
reasons that can be actually completed” (p. 598). However, is showing that infinite chains of
reasons can be completed different from Klein’s or Fantl’s infinitism?

It seems clear from Atkinson and Peijnenburgs’ arguments that they are not claiming that S’s
belief that p is justified only if S completes the infinite chain of reasons. Instead, they claim
simply that, since justification comes in degrees and is often thought to be a probabilistic
relation, some infinite chains of reasons can be actually completed (using the probabilities of the
propositions in the chain) and that there are several examples of this. This claim is perfectly
compatible with Klein’s and Fantl’s views, which both hold that complete justification (finishing
the chain of reasons) makes sense, but that a belief can be (to a degree or provisionally)
doxastically justified without completing the chain.42

When Klein claims that the requirement of completing the infinite chain “would be tantamount
to rejecting infinitism,” he is only rejecting the completion requirement that the person must
complete the processes of reasoning to the belief for it to be justified. He argues that even when
applied to foundationalism, the completion requirement is too strong: it means that most of our
beliefs are not justified.43 Thus, since we are discussing the actual justification of beliefs or actual
knowledge and most of what we claim to be justified in believing or know does not involve a
completed process of reasoning, I take it that Klein’s rejection of the completion requirement is
correct.44 Thus, it seems that (although we can imagine a strict completion infinitism) no one
holds “completion infinitism.” It is also clear that if the view were held, many or most of our
beliefs would not be justified.

3.4 IMPURE INFINITISM: AIKIN’S THEORY

According to Aikin (2008), pure meta-epistemic theories are committed to exclusively one
formal structure of justification, e.g., pure foundationalism holds that basic beliefs are the source
of all epistemic justification (pp. 175-176),45 whereas impure theories hold that there may be
more than one justificatory structure.46 Thus, impure versions of infinitism hold that justificatory
chains can or must be infinite47 and non-repeating but that other formal structures of justification
are legitimate. I first introduced impure infinitism when discussing Peirce (§1.3), who explicitly
endorses the view in his early works. Today, Aikin is the leading proponent of the view.

Since Aikin uses justification trees (j-trees) to describe his view, I will briefly provide an
example before explicating his infinitism. In the following j-tree, S believes p on the basis of a
and c and believes a on the basis of b and believes c on the basis of d:

p
/\
a c
/ \
b d

Aikin’s view requires the following for fallible external world beliefs to be justified: it is
necessary that at least one branch on every j-tree go on infinitely and that a basic belief be
included in every j-tree. However, Aikin leaves the possibility open whether there are beliefs
completely justified by non-doxastic content only. Thus, Aikin’s infinitism is impure because he
allows other structures—foundationalist structures—of justification.

But why should we be impure rather than pure infinitists? According to Aikin, impure infinitism
has a dialectical advantage over pure infinitism because pure infinitism cannot solve what he
terms the “modus ponens reductio.” Because it is often advanced and considered the most
devastating objection to infinitism, I will discuss the modus ponens reductio in chapters 5 and 6.
Briefly, the argument asserts that infinite chains of reasons are arbitrary when considered as
mere relations between beliefs: there is no way to adjudicate between truth-conducive and non-
truth-conducive chains of beliefs. Aikin (2005) uses basic beliefs to respond to the argument:
basic beliefs are needed to distinguish arbitrary infinite chains of reasons from ones that are
truth-conducive, thus, they are not used as “regress enders” but for “capturing the role that non-
doxastic states play in a subject’s formed and available justificatory support” (p. 201).48 Thus,
responding to the modus ponens reductio is Aikin’s motivation for accepting an impure
infinitism, but as we have seen, his solution to the reductio of infinitism utilizes foundationalist
criteria, so he must argue that foundationalist criteria are tenable. To support his use of
foundationalist criteria, he argues that Klein’s and Fantl’s arguments against foundationalism
fail.
First, recall that Klein’s argument against foundationalism is to ask why basic beliefs are truth-
conducive. The foundationalist must respond—to avoid arbitrariness—that basic beliefs have
some truth-conducive property, P. Klein then counters that either the meta-justification citing P
is a reason for the basic belief’s truth (thus, the regress continues) or it is not (thus, the basic
belief is arbitrary). Next, recall that Fantl’s argument purports that self-justified basic beliefs are
either completely justified or not. If they are completely justified, then the foundationalist cannot
explain how justification admits of degrees; if they are not completely justified, then they must
be justified by something other than themselves (i.e., a meta-justification). Furthermore,
foundationalist theories that allow meta-justifications of basic beliefs cannot satisfy the
completeness requirement: if basic belief, b, is completely justified if and only if it has property
P, then if S provides a reason to think that b has P, S would be more justified and so on, so that
the justification is never complete. On what grounds does Aikin reject these arguments?

Aikin (2008) claims that Klein’s argument and Fantl’s meta-justification argument exemplify the
same confusion: taking the ability to answer questions and challenges about a basic belief for the
continuation of providing supporting reasons for the belief (p. 181). He states that “a subject’s
ability to answer meta-challenges to basic beliefs [e.g., whether the belief has property P] is
necessary for their being justified for S as a reasonable and responsible believer, but they do not
constitute reasons for the belief—they are abilities that make up S’s knowing what she is doing
when she holds the belief” (Aikin, 2008, p. 181). He claims further that Klein’s argument may
show that we must accept infinitism but it does not necessarily eliminate foundationalism. In
other words, a basic belief is immediately and non-inferentially justified (e.g., if S has a
headache, S’s justification is the fact that S is aware of having a headache) for S and S’s
responses to the meta-challenges are “not constitutive of the epistemic relation between S’s
belief and her non-doxastic state” (Aikin, 2008, p. 181). Thus, there are two separate chains of
argument: a chain that terminates in a basic belief and a chain arguing that b is basic (e.g., S
could argue that she responsively regards her non-doxastic psychological states [Aikin, 2008, p.
180]), hence, a new regress emerges; S does not continue the original regress. Furthermore,
Aikin (2008) purports that, contra Fantl, fallibilism is motivated by the thought that complete
justification is unnecessary for knowledge, so an explanation of complete justification would be
an excessive requirement for a theory of justification (p. 182).

Next, against Fantl’s traditional foundationalism argument, Aikin argues that traditional
foundationalists can hold that justification comes in degrees, for example, by partitioning weaker
than deductive inferential support starting with basic beliefs from deductive support starting with
basic beliefs. Thus, Aikin concludes that foundationalist structures of justification remain
legitimate and therefore his response to the reductio is viable.

Finally, among the myriad of arguments defending infinitism that Aikin provides is an argument
that further distinguishes his view; thus, I will briefly explicate it here. Aikin (2005) responds to
what he terms “ought-implies-can” arguments, which claim that we cannot perform the task of
completing an infinite series of reasons for either of two reasons: (1) the quantity is too high to
complete (e.g., we live for a finite amount of time and it takes time to provide reasons) and (2)
the quality of a belief in or about an infinite chain necessary for the chain to provide justification
is too complex to hold (because our minds are finite) (pp. 193-194). Along with a premise
affirming infinitism (assumed for reductio) and a premise affirming ought-implies-can, each
argument respectively forms a reductio of infinitism.

Aikin’s main argument against the ought-implies-can arguments is to deny the ought-implies-
can premise.49 First, he questions whether ought really implies can. To support this doubt, he
maintains that modus tollens should constrain one’s thought, but there ispsychological evidence50
that humans are incapable of constraining their beliefs accordingly; however, it does not follow
that we should reject modus tollens as a rule of reasoning: “When we reason, we do so to do it
right” (Aikin, 2005, p. 206). Finally, he presents a reductio of ought-implies-can:

1. Ought-implies-can, i.e., “it is conceptually necessary that if S ought to do something, S

can do it” (assume for reductio).

2. Determinism is conceptually possible. Thus, for some S, it is possible that S does not
do something because S cannot.

3. Normative inquiry (epistemological, ethical, or aesthetical), only extends to “worlds

where failure to do what you ought is real.”

4. Ought-implies-can is false (1, 2, and 3 form an inconsistent set, where we can only
reject [1]). (Aikin, 2005)51

Thus, Aikin holds that, although providing infinite justificatory chains of reasons is impossible,
we should nevertheless strive for them because ought does not imply can.

CHAPTER 4

THE REGRESS PARADOX

Now that we have an adequate background of infinitism and epistemic justification in general,
we can begin to analyze the arguments surrounding infinitism. In this chapter, I examine Klein’s
and Fantl’s arguments that infinitism is the only theory of epistemic justification that can solve
the regress paradox. In the first section, I provide a common formulation of the regress paradox
and remind the reader of Klein’s and Fantl’s arguments. In the second section, I present several
responses to Klein’s and Fantl’s arguments and conclude that only foundationalist responses are
prima facie plausible. In the third section, I analyze the regress paradox and provide a more
tenable version—a version that many foundationalists purport to solve. In section four, I explain
the Sellarsian dilemma for foundationalism and Bonjour’s theory as an example of a
foundationalist theory, which purports to solve both the Sellarsian dilemma and the regress
paradox. In section five, I argue that any account of foundationalism (including Bonjour’s) that
can avoid rejecting Klein’s PAA and a principle related to his PAC and can solve the Sellarsian
dilemma cannot solve the regress paradox. Finally, in section six, I consider and reject a response
to my argument. I conclude that Klein’s PAA and a variant of his PAC—when it is
acknowledged that a theory must solve the problem of regress and the Sellarsian dilemma—
entail infinitism. That is, infinitism is the only non-skeptical theory of epistemic justification that
can solve the regress paradox.
4.1 THE REGRESS PARADOX AND WHY INFINITISM IS THE ONLY THEORY
THAT CAN SOLVE IT

The regress paradox is commonly presented as four independently plausible, but jointly
inconsistent, statements about epistemic justification:

1. Some beliefs are justified.

2. A belief can only be justified by an additional justified belief.

3. Circular chains of beliefs do not produce justification.

4. Justificatory chains are finite in length.

Although each of these statements seems correct, we can see that they are jointly inconsistent,
that is, all four statements cannot be true. For example, if statements (2)-(4) are true, (1) must be
false; if statement (1) is true, then (2), (3), or (4) must be false. As is easily inferred from the
accounts of foundationalism, coherentism, and infinitism in the first chapter, foundationalism
rejects statement (2), some theories of coherentism reject (3),52 some forms of skepticism reject
(1), and infinitism rejects (4).

Recall that Klein argues that infinitism is entailed by PAC and PAA and Fantl argues that
infinitism is entailed by the fact that justification comes in degrees and the fact that theories of
justification should explain how justification can be complete. Thus, they each argue that if we
are both to avoid skepticism (avoid rejecting statement [1]) and to solve the paradox, we must
reject (4), i.e., accept infinitism.

As I will explain in the next section, many philosophers have argued that Klein’s and Fantl’s
arguments are unsound and therefore infinitism is not necessary for providing a non-skeptical
solution to the paradox.

4.2 RESPONSES TO KLEIN AND FANTL

Since several philosophers have argued against both Klein’s and Fantl’s arguments, and since
several of their arguments are similar to one another, I will attempt to group their arguments into
two general arguments.

First, recall (§3.4) that Aikin argued that Klein’s argument and Fantl’s meta-justificatory
argument may show that infinitism is necessary to solve the regress paradox, but that their
arguments do not show that some beliefs cannot receive non-doxastic justification (i.e., basic
beliefs are still legitimate, but they do not stop the regress of reasons). Thus, this argument by
Aikin does not show (or purport to show) that foundationalism can solve the regress paradox.

However, Aikin’s arguments against Fantl’s main claims (justification comes in degrees and
completeness) endeavor to show that infinitism is not necessitated by these claims. Aikin claims
that traditional foundationalism can explain how justification comes in degrees, namely, by
distinguishing between inductive and deductive inferences (the foundationalist can hold that
deduction provides a higher degree of justification than induction). Aikin also argues that, pace
Fantl, a fallibilist theory of epistemic justification need not provide an account of complete
justification because fallibilism is the view that complete justification is unnecessary for
knowledge. I consider this the first argument that infinitism is not necessary for solving the
regress paradox. However, I believe that Akin’s arguments give us reason to doubt Fantl’s
arguments.53 Thus, for the remainder of the chapter, I will exclusively discuss Klein’s PAA and
PAC.

Moving on to the second argument, Andrew Cling (2004) argues that “foundationalism implies
that some beliefs can be justified independently of propositional support, not that there are
propositions for which propositional support is, in some absolute sense, unavailable” (p. 107).
Klein’s PAA requires that a reason be available for every reason given, but the foundationalist
holds that S’s basic beliefs are sufficiently justified non-propositionally, e.g., by an awareness of
experience that is non-inferentially justified for S, meaning that S is not required to provide more
reasons for basic belief b, even though there may be reasons for b objectively available to S.
Thus, foundationalism can accept PAA.

Although Richard Feldman does not respond to Klein’s arguments, his version of
foundationalism would fall under the heading of argument two. He argues that meta-
justifications for basic beliefs will provide more justification for a basic belief but that it is
unneeded justification because the experience itself (e.g., seeing redly) adequately justifies the
basic belief (Feldman, 2003, p. 77). That is, experience sufficiently justifies the basic belief, so
even though reasons are available to justify the basic belief, they are unneeded.

Similarly, Daniel Howard-Snyder and E. J. Coffman (2006), in one of their arguments54 against
Klein, claim that many philosophers mistakenly attribute the view to foundationalism that basic
beliefs cannot be “overdetermined by the evidence” (this argument also falls under argument
two). In other words, they are claiming that there are available reasons that can support basic
beliefs, i.e., basic beliefs can be “overdetermined by the evidence,” but that this justification is
unneeded. Thus, it seems that pace Klein, Fantl, and Aikin, foundationalism may be able to solve
the regress after all: foundationalism can conform to PAA (there are reasons available to support
basic beliefs) and solve the regress paradox by showing how some beliefs are non-inferentially
justified (basic beliefs are sufficiently justified by experience, so the regress is stopped).

To use the Simpson example of chapter 1, Simpson’s belief that “I drank last night” may be
inferred from his belief “I have a headache,” but the foundationalist claims that the chain of
inference stops there because “I have a headache” is sufficiently justified by his experiencing a
headache. And even though Simpson could provide more reasons to believe that he has a
headache, e.g., he has experienced headaches before and knows how they feel, the regress of
reasons has ended since the chain of inferences used to justify “I drank last night” stops with his
experiencing a headache.

Thus far, we have heard how foundationalists can respond to Klein’s PAA and Fantl’s
arguments. Can coherentists respond as well?
First, it seems that coherentist theories that require circular reasoning to solve the regress
paradox should be rejected. Even if we allow justified beliefs to figure into their own
justificatory ancestry (e.g., p justifies a and a justifies p), we must still agree with Klein’s claim
that circular reasoning cannot solve the regress paradox: for circular reasoning to stop an infinite
regress, a belief must essentially justify itself, which is viciously circular. In other words, if a
belief does not essentially justify itself, other beliefs are needed to justify it, which means that
the regress continues infinitely. Cling (2004) notes that Klein can use a weaker principle simply
to rule out a belief’s essentially justifying itself:

(PAC)* For all x, there is no person, S, such that S has a justification for x and S has a
justification for x only if x is in the evidential ancestry of x for S. (p. 104)

Thus, Klein can allow that by a belief, p, figuring into its own evidential ancestry, it can increase
its own justification, but the belief that p cannot be essential to its own justification and therefore
circular reasoning cannot solve the regress paradox, i.e., theories of justification must accept
PAC*. Because it is plausible that a belief can increase its own justification, rather than claim
that a theory must accept PAC, I will only claim that a theory must accept PAC*, which rules out
using circular reasoning to solve the regress paradox. (Note that PAC* also rules out basic
beliefs that are “self-justified” [§3.1].) Does holistic coherentism suffer the same fate?

Recall (§§1.2 and 3.2) that holistic coherentism, like basic beliefs in meta-justificatory
foundationalist theories, holds that a belief is justified in virtue of having some property, P. For
holistic coherentism, P is a belief’s membership in a coherent system of beliefs. The coherentist
could therefore avoid rejecting PAA by arguing that there are reasons available for the belief that
p but that S need not provide them if p has property P (cf. the foundationalist response). Also
recall that internalist coherentism holds that p is justified for S at t to the extent that p has
property P, i.e., to the extent that p coheres with S’s system of beliefs, and P requires that S be
aware of, or believe that, “p coheres with my system of beliefs.” Thus, p is unjustified for S if S
is not aware that p coheres with S’s system of beliefs. It follows that P requires (or is equivalent
to) the belief that “p coheres with my system of beliefs,” but S cannot justify this belief without
providing reasons—becoming a de facto infinitist—or without begging the question—arguing
that the belief coheres with S’s system of beliefs.55 Although there may be formulations of
coherentism that can avoid rejecting PAA, I simply cannot think of a plausible response
available to the coherentist that will avoid PAA and stop an infinite regress.56 Thus, in the
remainder of the chapter, I will exclusively discuss foundationalism.

Taking stock, we now have two arguments purporting that foundationalism can solve the regress
paradox: (1) Aikin’s arguments against Fantl and (2) the argument maintaining that reasons are
available for a basic belief but are unneeded (i.e., foundationalism can accept PAA).

4.3 ANALYZING THE REGRESS PARADOX

The two foundationalist responses explained in the previous section seem to avoid Klein’s and
Fantl’s arguments. However, the responses must also solve the regress paradox. That is, Klein
and Fantl each conclude that infinitism is the only theory that can solve the paradox; thus, any
response to Klein or Fantl must show how a non-infinitist theory can solve the paradox. Does
foundationalism succeed in solving the regress paradox? In order to answer this question, we
must first carefully examine the regress paradox.

Recall that the first premise of the regress paradox holds that “some beliefs are justified.” This
premise assumes that justification is absolute and non-comparative: some beliefs are either
justified or not justified. But is justification an absolute, non-comparative concept?

As noted earlier (§§1, 3.1, and 3.2), the claim that justification is comparative and admits of
degrees is a preponderant view among epistemologists. I agree with this view. However, I will
not rely on the fact that many epistemologists would agree with this formulation of justification.
Instead, I will provide two examples that—though they will not provide definitive evidence—
will support the intuition that justification is comparative and admits of degrees. Subsequently, I
will further support this intuition by providing a reduction ad absurdum argument against
absolute justification.

In the first example, consider Lisa, an experimental philosopher. Lisa is testing an intuitive belief
of hers: that most Americans, like her, think that justification is an absolute concept. She tests
her intuitive belief by surveying Americans in random regions by phone (e.g., one person
surveyed could be from California and another from Michigan). In her first survey of 2,500
people, she finds that 90% of those surveyed agreed with her belief. Is Lisa’s belief more
justified than before the survey, or absolutely justified, less justified, or not justified after the
survey? Next, suppose that Lisa decides to survey more people, say, 2,500 more people. The
results of this survey are that 96% of those surveyed agree that justification is absolute. With this
further survey, did Lisa’s belief become more justified? Did her belief reach absolute
justification? Was her belief already absolutely justified, so that the latter survey did not affect
the justification of her belief? Finally, envisage that Lisa somehow surveys all Americans and
finds that 98% think that justification is an absolute concept. Considering the whole example at
once, it seems evident that Lisa’s belief becomes more justified with each survey. After the third
survey, it seems that her belief was more justified than it was immediately following the first
survey, and that her belief was more justified after the first survey than it was when it was simply
an intuitional belief of hers.

Next, consider Bart, Selma, and Mel. All three believe that Mel is at the beach. And Mel is at the
beach. Mel sees the ocean and sand, feels the sand, and so on. Selma believes that Mel is at the
beach because she heard that Mel is at the beach from Patty, who is known to be a reliable
source. Bart believes that Mel is at the beach because he heard that Mel was at the beach from
Patty and a moment later, he received a picture from Mel’s phone of Mel at the beach. Ceteris
paribus, it seems natural to say that Mel’s belief is the most justified of the three, Bart’s belief is
justified to a lesser degree, and Selma’s belief is least justified.

The preceding examples seem to indicate that we conceive of justification as a relational concept
that comes in degrees rather than an absolute concept. However, it could still be the case that
justification is an absolute concept, even though the previous examples would seem odd if they
were explained using the notion of absolute justification. Perhaps there is a certain threshold of
evidence that must be provided for a belief before it is absolutely justified.
Let us assume this claim (for the reductio).57 For simplicity, assume that evidence provides a sort
of point system for a belief. For example, some reasons count as 1 point, some as 2 points, and
some as 3 points, while beliefs justified by experience and beliefs justified by their coherence
with other beliefs also count as 3 points. Furthermore, the reliability of a reason or process
increases the points provided by the source in question (or decreases the points in the event that
the source is unreliable). Now assume a threshold of justification, say, 50 points. Next, recall
(§1) that justification is a normative concept that guides our beliefs by informing us which
beliefs, if any, are rational to hold. Would it be rational for a person to leave her house if her
belief that “a fire is going burn my house down” has 49.99 points? If justification is an absolute
concept, according to our example, her belief is not justified and she should act accordingly. This
seems absurd. Finally, consider a person who believes that his boss is stealing from the company
for which they work. He also believes that his boss is not stealing from the company. Both
beliefs are justified: the former scores 50 and the latter scores 888. The employee’s beliefs
obviously have consequences, so he would like to find out which belief is rational to hold.
Setting aside the ethical issue of when one ought to “whistle blow,” what should the employee
believe? If justification is absolute, both beliefs are absolutely justified. This also seems absurd:
the latter belief appears to be more justified than the former.

Thus, the intuition that justification is relational and comes in degrees seems to be well founded.
What does this mean for our understanding of the regress paradox?

It means that we must reformulate the regress paradox: since justification is relational and admits
of degrees, the statement “some beliefs are justified” is dubious. Many foundationalists claim
that the paradox is getting at this: How do we stop a regress of conditional or inferential beliefs?
In other words, the proponents of the paradox were perplexed by justification because it seems
that we must provide reasons to justify our beliefs because our beliefs are inferential in character,
i.e., they admit of truth and falsehood, so they must rely on the truth of other beliefs for their
justification; but the reasons provided will also be inferential and therefore it seems that the
regress cannot be stopped. We arrive at a trilemma: inferences can either be circular, infinite, or
have a foundation. Since foundationalists do not think that infinite regresses of non-repeating
reasons or circular chains of reasons can provide justification, they must show that non-repeating
chains of reasons can have a foundation. Thus, Bonjour, Howard-Snyder and Coffman, among
many other foundationalists, put the problem of regress as follows: How do we stop a regress of
inferential reasons?58

Now that we have a tenable formulation of the regress problem as a trilemma concerning
inferential regress—a formulation foundationalists purport to solve—we can see that in order to
solve the problem of regress, foundationalists must hold that basic beliefs are justified by
something non-inferential, which is what they attempt to do. In the second argument purporting
that foundationalism can solve the problem of regress, the claim was that experience, which is
non-inferential in character because it is non-epistemic (it is not the type of thing that can be true,
false, or justified), justifies basic beliefs and hence stops the regress of inferential justification.
However, much needs to be explained if we are to determine whether foundationalism can solve
the problem of regress, to which I turn in the next section.
4.4 THE SELLARSIAN DILEMMA AND BONJOUR’S FOUNDATIONALISM

Can an immediate awareness of experience stop an inferential regress?59 Perhaps, but the
foundationalist must first solve a dilemma in order to show how this awareness of experience can
stop the regress.

4.4.1 The Sellarsian Dilemma

As we have seen (§1.1), the foundationalist attempts to solve the problem of regress by holding
that a basic belief is justified because it is supported by an aspect of experience via that aspect of
experience’s immediate acquaintance with a person, i.e., the experience is “given.” It seems that
this immediate acquaintance can justify basic beliefs and stop the regress because it needs no
further justification—it is “given,” i.e., it is non-inferential and non-conditional and therefore
non-epistemic (it does not admit of being true or false and so does not rely on the truth of other
beliefs).

However, the “Sellarsian dilemma,” originated by Wilfrid Sellars and further developed by
Bonjour (when Bonjour rejected foundationalism), purports to render foundationalism untenable.
The dilemma is as follows. Either (1) the “immediate acquaintance” with experience or the
“given” is an assertive or judgmental cognitive act (e.g., judging the experience to be one way
rather than another) or (2) it is not a judgmental cognitive act. If (1), then the acquaintance with
experience can justify propositional “basic” beliefs (if it is itself justified). However, if (1), then
this act of judgment—if it is justified—requires a reason to believe it because it is capable of
being true or false. Thus, the acquaintance does not end the regress. If (2), then the content of
this acquaintance with experience is non-judgmental, non-assertive, etc., which means that it is
incapable of being true or false and therefore epistemic justification does not apply to it, i.e.,
acquaintance does not need a reason to justify it, so it can stop the regress. However, if (2), then
it is hard to see how this acquaintance with experience can provide a reason to believe the basic
belief is true. Bonjour (2003a) asks:

How can a state whose content does not in any way say or indicate that things are one
way rather than another nonetheless provide a reason or any sort of basis for thinking that
the propositional content of a belief that they are one specific way is true? (p. 19)60

A related objection to foundationalism is that the content of experience is non-conceptual and

non-propositional (cf. horn [2] of the Sellarsian dilemma) and so cannot justify a basic belief,
which is conceptual and propositional—the relation between the content of experience and
beliefs is merely causal.61

Furthermore, if the connection is merely causal, the foundationalism cannot be internalist. S’s
“basic” belief that b, e.g., “I see a green triangle,” does not include a proposition or an awareness
that b was caused by a green triangle external to S. Thus, b must be supported by the proposition
that “b is caused by an external object,” call it p, which is either an externalist claim (S is not
aware of p or the evidence for p) or it must be justified by S (p does not stop the regress).62
Thus, in order for foundationalists to solve the problem of regress, they must avoid the Sellarsian
dilemma and related objection, i.e., they must explain how a basic belief is inferential and
propositional and so can justify other propositions but is justified by something non-inferential
and non-propositional like acquaintance with experience.

4.4.2 Bonjour’s Foundationalism

Bonjour argues that his theory of foundationalism—and no other theory of foundationalism—

avoids both the Sellarsian dilemma and the related argument and shows how basic beliefs are
justified by non-inferential support, thereby stopping the inferential regress. Because I believe
that Bonjour’s foundationalism is the most clearly formulated and has the best chance of solving
the regress problem, I will use his foundationalism as an example of how foundationalism
attempts to avoid the Sellarsian dilemma and solve the problem of regress.63 However, my
argument that foundationalism cannot solve the problem of regress (in §4.5) is general and
should apply to all empirical foundationalist theories.64

Crucial for Bonjour’s (2003a) foundationalism is the claim that a sensory experience is a
conscious state and therefore “essentially involves a constitutive, or ‘built-in,’ non-apperceptive
awareness of its own distinctive sort of content, namely sensory content” (p. 70). But like the
argument related to the Sellarsian dilemma claims, this experience is non-propositional and non-
conceptual in character, so how does it provide reason to believe a propositional basic belief?
Bonjour claims that the relation between the non-conceptual experiential content and the
propositional “basic belief” is descriptive. The proposition either will describe the experiential
content correctly or incorrectly; thus, the experiential content determines whether the proposition
is correct, that is, determines whether the description correctly describes the experiential content.

Bonjour (2003a) provides the crux of his account in the following passage:

Thus when I have a conscious state of sensory experience, I am, as already argued, aware
of the specific sensory content of that state simply by virtue of having that experience.
And hence if an apperceptive belief that I entertain purports to describe or conceptually
characterize that perceptual content, albeit no doubt incompletely and abstractly, and if I
understand the descriptive content of that belief, i.e., understand what an experience
would have to be like in order to satisfy the conceptual description, then I seem to be in a
good, indeed an ideal, position to judge directly whether the conceptual description is
accurate as far as it goes, and if so, to be thereby justified in accepting the belief. Here
again there is no reason to think that mistake is impossible and thus no reason to think
that such an apperceptive belief is infallible or indubitable. But as long as there is no
special reason for thinking that a mistake is likely to have occurred, the fact that such a
belief seems via direct comparison to accurately characterize the conscious experience
that it purports to describe apparently provides an entirely adequate basis for thinking the
description is correct and hence an adequate basis for justification. (pp. 73-74)

However, Bonjour’s account remains somewhat unclear. What exactly justifies the basic belief?
Elsewhere, Bonjour (2006) claims that a basic belief is “justified by my awareness of the
descriptive fit between the content of the proposition I believe and the relevant aspect of the
content of the experience” (p. 744). This awareness of “fit” is a cognitive act separate from the
propositional content and experiential content, but it is still “grounded” in the relevant
experiential and propositional content. That is, the awareness of “fit” justifies the basic belief,
but the relevant experiential and propositional content are an essential part of this cognitive act
(the awareness of “fit”).

Bonjour’s sufficient condition for empirical justification can be put as follows: S is justified in
the basic belief that p if:

1. S has conscious sensory experience e with experiential content c.

2. S believes p and understands the descriptive content of p.

3. S “apprehend[s] or recognize[s] the agreement or ‘fit’ between the aspect of experience

being attended to and the conceptual description given by the belief” (Bonjour, 2003b, p.
193), that is, S must recognize that the content of p correctly describes c.

4. There is no special reason for thinking that a mistake is likely to have occurred.65

An example may further elucidate Bonjour’s view. In a chapter titled “Reply to

Sosa,”66 Bonjour provides an example where S is looking at a large abstract painting in good
lighting. The painting contains several shapes and colors, but someone suggests (it could be S or
someone else) that in the painting there is an approximately equilateral triangular shape, which is
dark green. After some searching, S spots the shape (and so now has a conscious experience of
the shape). Since S understands the descriptive content of the proposition “I see a dark green,
approximately equilateral triangular shape,” S immediately recognizes the agreement or “fit”
between the content of the proposition and the experiential content of which S is conscious,
which means that S’s belief in the proposition is justified.

Consider Figure (4.1) as a further example. In Figure (4.1), the thin, dark arrows indicate
“essential involvement in the cognitive act of ‘fit’,” and the wide, light arrows indicate “provides
justification for,” where the box at the base of the arrow is a reason (broadly construed) to
believe (or support for) the box at the apex of the arrow. Furthermore, in Figure (4.1) and in the
remainder of the work, I signify “an experience” as follows: x, an experience, is indicated by *x*
(I will continue to use standard quotation marks to signify words).

Basic belief G: “I see a roughly triangular dark green shape”

^ |
| v
Cognitive awareness of “fit”: G correctly describes the experiential content of *G*
^
|
Experiential content of *G*

Figure 4.1. Bonjour’s account of basic beliefs.

This account of foundationalism, if cogent, seems to succeed in stopping inferential regress. A
cognitive awareness of “fit” between the content of a conscious aspect of experience and the
content of a propositional belief justifies the basic belief. Furthermore, this awareness of “fit”—
although it is not entirely non-propositional (it still contains the conceptual content of the
propositional belief)—needs no further justification (hence it seems to avoid the Sellarsian
dilemma and related objection). This is because, Bonjour (2006) claims, I can in one cognitive
act “grasp both the content of a proposition and the non-propositional (though partly conceptual)
situation that makes it true” and see immediately that the truth conditions for the proposition are
satisfied (p. 747). “The upshot is indeed a propositional judgment, but one whose justification is
already present in the cognitive act from which it arises” (ibid.) In other words, by claiming that
some cognitive acts are only in some sense propositional, Bonjour hopes to show that this
cognitive act of “fit” does not require justification, i.e., it is not a propositional judgment.

4.5 WHY FOUNDATIONALISM CANNOT STOP THE REGRESS

First, I should note that although I fail to see how this cognitive act of “fit” is non-judgmental
and therefore not in need of further justification, I will grant that we experience a cognitive
awareness of “fit” between propositional content and sensations and that this cognitive act does
not involve judgment. Nevertheless, so called “basic beliefs” involve more than Bonjour, and
other foundationalists, let on, and for this reason, I think that basic beliefs are inferential beliefs,
which must be inferred from other inferential beliefs to be justified. That is, even if the act of
“fit” does not involve judgment, it does not satisfy the required justification, i.e., the proposition
in question cannot be inferred from an awareness of “fit” alone; thus, the regress continues.

What is left out of the foundationalist explanation of basic beliefs? To answer this question, we
must ask, what is the propositional content of an empirical belief? It contains at least concepts
and terms. But why distinguish concepts67 from terms? The necessity of making this distinction
will become clear by considering the following cases.

First, consider the fact that we can have a concept whose term we have forgotten. For example, I
may use a flash drive every day to save my documents, but on certain occasions—perhaps when
trying to ask someone if she has seen my flash drive—I may draw a blank on what to ask for,
that is, I may have forgotten what term to use even though I still have the concept of a flash
drive. I may be imaging a flash drive, thinking of its purpose, its color, its shape, what I am
going to use it for, and so on and still not remember what to call it.

Similarly, we sometimes call a concept by the “wrong” name. By “wrong,” I simply mean that it
is not the term that is usually used to refer to the concept (i.e., the term that English speakers use
to refer to the concept or—if a private language makes sense—not the term usually used by the
person in question to refer to the concept). For example, I could term my concept black, “red,”
and my concept sour, “sweet” (this could be because I learned the wrong terms or I have simply
confused the terms involved).

Next, think of new words. When a new color, shape, instrument, feeling, etc., was invented or
discovered (or simply never linguistically expressed before), the concept of the color, shape, etc.,
came into existence before the term. For example, a surfboard was likely invented, used, and so
on—meaning that people had the concept of a surfboard—before the term “surfboard” was
attached to the concept.68

Consider also the fact that the terms for concepts often change meaning. For example, the term
“awful” means the opposite of what it once meant: “full of awe,” or wonderful, etc. However,
our concepts of, e.g., wonderful, have not changed.

Finally (although this list of cases is not exhaustive), consider the fact that human children and
some animals may have concepts without linguistic knowledge (or with limited linguistic
knowledge). It seems that a young child has some sort of concept of, e.g., a rattle, before the
child can say, “rattle.” For example, the child may often seek the rattle out of a box of other toys,
know how to get the rattle to make noise, and remember that the rattle is the toy that makes
noise. All of this seems to suggest that the child has the concept of a rattle even though it cannot
linguistically express the concept.69

Thus, it is clear that concepts and terms are not necessarily isomorphic. Moreover, it is clear that
even if concepts require concomitant terms, justified terms (or “correct” terms) are not necessary
to hold a conceptual belief.70

4.5.1 What the Concept/Term Distinction Demonstrates

What does the distinction between terms and concepts mean for the notion of “basic beliefs”?

Consider Bonjour’s painting example. While looking at a painting, S believes that “I see a
roughly triangular green shape.” Bonjour claims that S “must understand the descriptive content
of that belief, i.e., understand what an experience would have to be like in order to satisfy the
conceptual description” (this is part of S’s belief). But why must Bonjour hold the view that an
understanding of the descriptive content of a belief is necessary for holding the belief?

If an understanding of a belief’s descriptive content is not necessarily part of the belief, then the
belief must be inferred from other beliefs, i.e., it cannot be basic; thus, the claim that
“understanding the descriptive content is part of a belief” is crucial for the foundationalist. To
explain why this is so, assume that an understanding of a belief’s descriptive content is not a
necessary part of the belief. Now imagine that you are immediately aware that the descriptive
content of your belief that p “fits” the content of your experience. However, if an understanding
of a belief’s descriptive content is not necessarily required for holding the belief, you must
justify that your belief that p has the descriptive content you think it does—after all,
understanding the belief’s descriptive content is not a necessary part of your belief (you must
judge that the belief has a certain descriptive content). Alternatively, this could show that the
awareness of “fit” cannot be immediate; S must first justify to herself whether the belief that p
has a certain descriptive content (whether she understands the descriptive content of the belief),
which will correctly describe her experience before she can be aware of the “fit” between the
descriptive content and the experiential content. Without a sort of necessary “built-in”
understanding of descriptive content, the descriptive content of a belief (or the claim that the
descriptive content of a certain belief is necessarily understood) admits of truth and falsehood so
it must be inferred from other beliefs. For example, if the descriptive content of S’s belief “I see
a triangular green shape” is not necessarily understood, S must infer the descriptive content from
other beliefs, e.g., “‘green’ describes *green* experiences.” Thus, we see that any account of
basic beliefs must hold that an understanding of the descriptive content of a belief is part of the
belief; otherwise, the belief must be inferred from other beliefs and is therefore not “basic.”

What Bonjour fails to see is that S’s understanding of what an experience must be like for the
belief to describe an experience correctly can either mean that S understands what experience
must be like for the terms of the proposition to describe the experience correctly or S understands
what experience must be like for S’s concepts to describe the experience correctly, or both. That
is, since terms and concepts are not necessarily isomorphic (§4.5), the foundationalist must
explain what describes or “fits” experience (terms, concepts, or both), which yields justification.
I consider each option in turn.

The first option is simply false. S can hold an empirical belief without understanding what
experience would have to be like for the terms of the proposition to describe the experience
correctly. For example, S could understand what experience must be like for the concepts of her
propositional belief to describe experience correctly while not understanding what experience
must be like for the terms of her propositional belief to describe experience correctly. If, for
instance, S mistakenly thinks that “green” refers to her experience *red*, then S does not
understand what experience must be like for “green” to describe her experience correctly.
However, we would still say that S believes that she “sees green” even though she does not
understand what experience must be like for “green” to describe her experience correctly (she
thinks that “green” correctly describes *red*). S clearly believes that she “sees green,” and she is
simply using the wrong term; she has a grasp of the concept she has in mind when she says
“green” (the concept of red) and what experience must be like to “fit” this concept (an
experience of *red*). Thus, if the descriptive content of a basic belief is the terms of the belief,
then an understanding of the descriptive content of the belief is not a necessary part of the belief.
Consequently, the belief is not basic: its descriptive content admits of truth and falsehood (it is a
conditional belief) and therefore must be inferred from other beliefs.

The second option claims that in order for S to hold an empirical belief, S must understand what
an experience must be like to satisfy the concepts of her belief (i.e., satisfy the conceptual
description of her belief). As long as S understands what it would be like for her experience to
satisfy the concepts of her empirical belief, S can have a cognitive awareness of “fit” between
the belief’s conceptual content and S’s experiential content. However, S need not understand
what it would be like for the terms of the empirical belief, e.g., “I see a roughly triangular green
shape,” to describe her experience correctly since concepts and terms do not necessarily match.
For example, S could think that the terms in her basic belief would correctly describe a different
experience than they describe,71 e.g., S could refer to *red* as “green,” but S must (according to
this option) understand what her experience must be like to match her concept of red (it must be
an experience of *red*).

Thus, the analysis of the first and second options so far shows that S can hold an empirical basic
belief without understanding what it would be like for the terms of the belief to describe an
experience correctly. From this analysis, we find that the second option must fail: since terms
and concepts are separate, it does not follow from the fact that the concepts of a proposition “fit”
experience (option two) that the terms of the proposition must “fit” experience. However, it is
necessary that the terms of the propositional basic belief b are justified because it is the terms of
b that justify other propositional beliefs. Consequently, for S’s basic belief to be justified, S must
be justified that the terms “fit” experience (option one). That is, in order to stop the regress of
inferential beliefs, a basic belief’s terms must correctly describe experience. However, recall that
in order to hold an empirical belief, it is not necessary that S understand what experience must be
like to “fit” the terms of the belief, so S must infer from other beliefs that the terms of her basic
belief have a certain descriptive content, which correctly describes the content of her experience.
Thus, the second option fails because the first option (that the terms of a belief constitute its
descriptive content) is necessary to stop the inferential regress and the second option does not
entail the first (i.e., understanding the concepts of a belief does not entail understanding the
terms of the belief, as I have shown above).

Thus, when Bonjour claims that S must understand the descriptive content of an empirical belief,
he must mean that the terms and concepts of the proposition constitute its descriptive content.
Again, since S can hold an empirical belief without understanding what experience must be like
to “fit” the terms of the belief, S must justify the claim that the terms of her belief have a certain
descriptive content that “fits” the content of her experience. And since no one can be sure that
concepts always match terms, we must justify that our concepts and terms match or that our
terms “fit” our experience in order to justify an empirical belief.

Indeed, even if it is claimed that the matching of concepts and terms is an immediate cognitive
act, it is a judgmental cognitive act that admits of truth and falsehood (because concepts and
terms are not necessarily isomorphic), which would violate the Sellarsian dilemma. The third
option therefore fails.72

Considering all of the foundationalist options, my argument can be put succinctly as follows:

1. For S’s basic belief to stop an inferential regress, it is necessary that if S’s basic belief
is empirical (all basic beliefs are empirical), then S understands what an experience must
be like in order for the terms of S’s belief to describe her experience correctly.

2. It is not necessary that if S’s basic belief is empirical, then S understands what an
experience must be like in order for the terms of the belief to describe an experience
correctly.

3. Basic beliefs cannot stop inferential regresses. (1 and 2)

Since the premises of this argument are somewhat complex, I will attempt to clarify them by way
of examples. What I am claiming is this: in order to justify a basic belief, e.g., “I see a roughly
triangular green shape,” S must be justified that the terms in the proposition refer73 to the
experiential objects or concepts involved, are syntactically correct, and so on, which means that
the regress continues.

Even if we grant that a belief’s correctly describing experience provides justification, S’s
propositional basic belief—if the belief correctly describes S’s experience, i.e., if the belief is
justified—implicitly contains several inferentially justified beliefs about the terms involved. For
example, S’s basic belief “I see a roughly triangular green shape”—if it is justified—contains at
least the following justified inferential beliefs about the terms involved: “if appeared to thusly
(experience of *green*), I (and other English speakers) term this aspect of experience ‘green’,”
“if appeared to thusly (experience of a *triangular shape*), I (and other English speakers) term
this experience ‘triangular’,” and “when I have a propositional belief, I generally apply the
correct terms of the belief to the correct experiences, i.e., I often understand the descriptive
content of the terms of my beliefs.”74 Each of these beliefs, if they are justified, must be inferred
from other beliefs and so on ad infinitum. In other words, the terms of colors, cardinality,
ordinality, basic geometrical shapes, relations, etc., are included in basic beliefs and are
inferential empirical beliefs: if they are justified, they owe their justification to other empirical
beliefs. Thus, we do not have the pyramid of reasons promised by the foundationalist (§1.1).
Instead, chains of belief continue infinitely (see Figure [4.2]).

Basic belief G: “I see a roughly <- I reliably understand the descriptive content <- …
triangular dark green shape” of the terms of my empirical beliefs
^ ^ |
| | v
Cognitive awareness of “fit”  Experiential content of *G*
^
|
If I have an experience with content *G*, I (and all English speakers) term the experience
“green”
^
|
…

Figure 4.2.Why Bonjour’s basic beliefs do not end the regress.

In Figure (4.2), by illustrating only two regresses of inferential beliefs, we can see that Bonjour’s
account of basic beliefs cannot stop an infinite regress of inferential beliefs. Nor can any account
of empirical basic beliefs because all basic beliefs must contain terms, which require inferential
justification: any sort of awareness of experience cannot justify a belief unless the terms of the
belief are justified, which requires inference from other beliefs

(I consider a foundationalist response to this claim in the next section). Thus, it seems that we
cannot construe basic beliefs as “basic,” that is, as “non-inferrentailly justified beliefs that are
justified by a non-inferential, non-epistemic, and non-judgmental awareness of the ‘fit’ between
the propositional content of a belief and awareness of experience,”75 and so foundationalism
cannot solve the problem of regress.

4.5.2 Howard-Snyder and Coffman’s Response

The argument I formulated above somewhat resembles a question Sosa poses to the
foundationalist and an argument against foundationalism by Lehrer. Sosa is puzzled how a basic
belief can lack all justificatory support from other beliefs, yet it enjoys the existential support of
other beliefs. That is, in order for a basic belief to exist, it needs a host of other beliefs, so how is
it that these other beliefs do not justify the basic belief (Sosa, 2003b, pp. 208-209)?

Sosa’s question resembles Lehrer’s (2000a) argument against foundationalism, which is as

follows:

• (a) S believes that x is F

• (b) S is justified in believing that x is F

• (c) S lacks the information on this occasion about whether he or she is competent to tell
whether something is an F or not and doubts that he or she is competent to tell the
difference on this occasion. (pp. 80-81)

Lehrer claims that the conjunction of (a) and (c) does not entail (b): S believes that x is F but
doubts that he is competent enough to tell the difference and lacks independent information
whether he is competent to tell the difference, so S is obviously not justified in his belief.76 Thus,
to be justified in a basic belief, S needs to have justified independent information (this
information could either be information rendering S competent to tell when something is F or S’s
belief that he can tell whether something is an F or not).

Perhaps the strongest response to Sosa’s question and Lehrer’s argument, which is propounded
by Howard-Snyder and Coffman (2006),77 is that while other beliefs are existentially necessary
for a basic belief to exist, there is no reason to suppose that a basic belief owes its justification to
these other beliefs. In other words, the mere existence of some other justified beliefs is necessary
to hold a basic belief (e.g., if S believes that “I see a red ball,” S must have justified beliefs
concerning what red is, what a ball is, and so on), but this does not mean that these other beliefs
are reasons for why S believes the basic belief. In fact, Howard-Snyder and Coffman argue that
since some justified beliefs are necessary to hold a basic belief b (and any belief), it seems that
these beliefs are not reasons to believe b.

To support the claim that other justified beliefs are necessary in order to form a belief (cf.
Bonjour’s claim that understanding the descriptive content of a belief is necessarily part of the
belief), Howard-Snyder and Coffman ask us to envisage a person, Evan, whose existential beliefs
that make his basic belief possible to hold are unjustified. They argue that Evan is completely
ignorant of how to use the concepts involved and so has no grasp of them at all; thus, in order to
have a belief, we must have a grasp of the concepts involved, which requires justified beliefs—
one must be disposed to apply the concepts correctly (Howard- Snyder & Coffman, 2006, pp.
561-562).

At first glance, this response to Sosa and Lehrer appears to refute my argument that basic beliefs
must be inferred from other beliefs to be justified; however, my argument remains unscathed.
For consider the fact that I distinguish concepts from terms. Now consider Howard-Snyder and
Coffman’s case of Evan. Evan could be completely ignorant of the terms or concepts of b. If he
is ignorant of concepts, then he could not hold b, he does not “grasp” the concepts involved
(according to Howard-Snyder and Coffman). However, if the concepts contained in Evan’s basic
belief are justified, but he is mistaken about what words refer to what experiences and concepts,
then Evan can form a belief without holding justified beliefs about the terms used (§§4.5 and
4.5.1).78 Thus, Howard-Snyder and Coffman’s argument does not apply to my argument: they
claim that since some justified beliefs are necessary to hold a basic belief b (and any belief), it
suggests that these beliefs are not reasons to believe b, but my argument does not utilize the
beliefs necessary to hold a belief (the conceptual beliefs). Instead, I claim that since justified
beliefs about terms are not existentially necessary to hold a basic belief but are necessary for the
belief to be justified, it suggests that beliefs about terms must justify the basic belief and hence
there are no “basic” beliefs. That is, to stop an inferential regress, the descriptive content of a
basic belief must be the terms of a belief (as shown above), which admits of truth and falsehood
and is not existentially necessary to hold the belief, so it must be justified by other justified
beliefs if it is to be justified at all. Thus, for Evan’s “basic beliefs” to be justified, they must be
inferred from other beliefs, such as “I am applying the correct words to the correct concepts for
the particular language in question,” “I reliably understand the descriptive content of the terms of
my empirical beliefs,” and so on. My arguments therefore withstand Howard-Snyder and
Coffman’s arguments.

4.6 CONCLUSION

Any tenable foundationalist theory must avoid the Sellarsian dilemma, accept PAA and PAC*,
and solve the regress paradox. As I have argued, foundationalism cannot accomplish this task.
The conjunction of PAA and PAC*—when it is acknowledged that the theory must solve the
regress paradox and the Sellarsian dilemma—therefore entails infinitism.

Although I have shown that infinitism is the only theory of epistemic justification that can
provide a non-skeptical solution to the problem of regress, there are several other arguments to
consider before accepting the theory over its rivals to which I turn in the next chapter.

CHAPTER 5

THE MODUS PONENS REDUCTIO

In this chapter, I examine what many philosophers cite as their main reason for rejecting pure
versions of infinitism: the “modus ponens reductio” (hereafter, MPR).79 Although I argued in the
previous chapter that infinitism is the only theory that can solve the regress paradox, if the MPR
is a reductio of infinitism, it may show that other theories of justification are preferable to
infinitism. Thus, in this chapter, I will examine infinitism as well as the leading theories of
epistemic justification (foundationalism and coherentism) to determine whether the MPR shows
that other theories of justification are preferable to infinitism. In section one, I present Aikin’s
version of the MPR and his response to the MPR, which argues that infinitists must adopt
portions of other theories of justification to avoid the reductio (i.e., adopt impure infinitism). In
the second section, I argue that reductio arguments similar to the MPR apply to foundationalism,
including impure infinitism. Thus, pace Aikin, I argue that impure infinitism is not needed to
solve the MPR; what’s more, impure infinitism does not help to solve the reductio, thus, the
motivation for impure infinitism is lost,80 and I examine only pure infinitism in the remainder of
the work. Finally, in section three, I argue that reductio arguments similar to the MPR apply to
coherentism. Thus, in this chapter, I argue that slightly modified versions of the MPR are
reductios of the two leading epistemic theories of the structure of justification (foundationalism
and coherentism), thereby greatly reducing the force of the argument against infinitism by
showing that the MPR is not a special problem for the theory. Furthermore, my argument shows
that the leading theories of justification are not preferable to infinitism because the MPR is a
reductio of the theory, since the MPR (modified) applies to the leading epistemic theories as
well.

5.1 THE MODUS PONENS REDUCTIO AD ABSURDUM OF PURE INFINITISM

The most recent, and I believe most challenging,81 presentation of the MPR comes from Aikin.
He claims that since pure versions of infinitism hold justification to be merely relations between
beliefs, the justification is arbitrary. Aikin illuminates this worry by considering a chain of
beliefs in a support relation, such that B2 supports B1, and B3 supports B2, etc. Consider the
following sets of beliefs:

Set (1) Set (2)

B1: p B1': ~p
B2: q & (q → p) B2': q & (q → ~p)
B3: r & (r → (q & (q → p))) B3': r & (r → (q & (q → ~p)))
B4: …. B4': …. (Aikin, 2008, p. 183)

We can see that each set of beliefs is in a supporting relationship: if modus ponens is truth
preservative, B2 entails B1, and so on. The problem for the pure infinitist is that there is “no
rational way of telling the difference between one justifying set of beliefs that is conducive of
truth and one that is not […] because [infinitist justification] is defined exclusively in terms of
relations between beliefs, [it] loses its moorings with truth” (Aikin, 2008, p. 183). In other
words, the infinitist cannot rationally adjudicate between sets (1) and (2) or between any justified
and arbitrary chains of beliefs. Furthermore, we can justify any belief and its negation using
infinite series of inferential-justification, thus, infinite regresses are not truth-conducive (Aikin,
2005, p. 198).82

There seem to be two worries driving the MPR. First and most important, there is a conceptual
worry (the reductio) that infinite regresses are arbitrary (i.e., non-truth-conducive) because we
can use them to justify all propositions and their negations at all times: assuming infinitist
criteria, every p and ~p is propositionally justified at all times, hence a contradiction is derived
and the MPR is a reductio of infinitism. Importantly, pace the proponents of the MPR, the MPR
is a reductio of infinitist propositional justification, not infinitist doxastic justification.83 Since no
infinitist holds that persons are completely doxastically justified in all (or any) beliefs (provide
an infinite number of beliefs to justify each belief), the “reductio” would not derive a
contradiction if it were applied to doxastic justification. Thus, the reductio concludes that all
propositions (including their negations) are justifiable (propositionally justified) for all persons
at all times, which results in contradiction. Second, there seems to be a doxastic normative worry
that there is no rational way to adjudicate between the affirmation and negation of a belief (or
any two beliefs) using infinitism: a theory should provide a way to attempt to adjudicate between
the truth of two contradictory claims (or any two claims). Because of the reductio, any
proposition can be doxastically justified for S at t if S provides the available reasons to believe p
at t, which means that infinitists cannot adjudicate between any two beliefs since each can be
justified. I will continue to use “MPR” to denote both the conceptual reductio and the normative
worry, but I will also examine the conceptual reductio and normative worry separately, so I call
the conceptual reductio MPRc and the normative worry MPRn.

Aikin (2008) responds to the MPR by adding a necessary condition to infinitist justification. He
argues that since sets (1) and (2) cannot be rationally adjudicated as to the truth of p and ~p using
inferences, another necessary component of justification must be adopted (p. 183). He claims
that foundationalist criteria must be used. For example, we must hold that beliefs based on
analytic entailment are self-justified, such as (A) if S is a bachelor, then S is male. Thus, if we
consider (1) and (2), foundationalist criteria will maintain that, say, (1) is at least prima facie
justified because one of the branches on the j-tree of (1) terminates in a self-justified belief.
Thus, foundationalist criteria are needed (in every j-tree) to solve the problem because something
other than propositions (i.e., beliefs not in need of j-trees) must justify some beliefs in order
rationally to adjudicate between belief chains (to address the MPRn) and separate justification-
affording proposition chains from proposition chains that do not provide justification (to address
the MPRc). Thus, he concludes that impure infinitism is preferable to pure infinitism because
only an impure infinitism can solve the MPR.

5.2 HOW A MODIFIED MPRAPPLIES TO FOUNDATIONALISM AND IMPURE

INFINITISM

Aikin’s solution to the MPR of impure infinitism—or alternatively, the abandonment of

infinitism for another theory of justification—seems to be the only plausible solution until we
consider the fact that modest foundationalism (including impure infinitism) and coherentism are
subject to reductio arguments similar to the MPR. I only claim that the arguments are “similar”
because formally my reductio arguments of coherentism and foundationalism have a different
structure than the MPR since the MPR assumes the infinitist structure of justification (to derive a
contradiction), which differs from foundationalist and coherentist structures. However, since the
reductio arguments are so similar (each argument assumes the theory in question’s structure and
derives a contradiction by showing that all propositions are justifiable at all times, which also
leads to the concomitant normative worry or MPRn), I will simply call them modified MPR
arguments or MPR* (including MPRc* and MPRn*). I propound MPR* arguments of
foundationalism as well as reliabilist and impure infinitist versions of foundationalism in this
section and coherentism in the next section. (Recall that for all foundationalists, beliefs are
doxastically justified if and only if they are basic or inferred from basic beliefs, and propositions
are propositionally justified if and only if they are either basic or inferred from a basic
proposition.)84

First, when traditional foundationalism is carefully considered, we find that the MPR* is a
reductio of the theory. To understand why the MPR* applies to foundationalism, consider two
belief chains, (A) and (B), where both begin with the basic belief (BB) r “I am appeared to
redly” and (A) includes (BB) b “I am appeared to bluely” and (B) includes (BB) g “I am
appeared to greenly.” Both belief chains beginning with justified basic beliefs justify, through
valid modus ponens inferences, the beliefs f and ~f respectively. Filling in the chain of reasons to
provide a more concrete example, we get the following:

Set (A) Set (B)

BB: r BB': r
BB: b BB': g
B1: ((r & b) → f) B1': ((r & g) → ~f)
B2: f B2': ~f

If r is “I am appeared to redly,” b is “I am appeared to bluely,” g is “I am appeared to greenly,”

and f is Friday, then the foundationalist can infer that it is Friday depending on what colors are
seen. Thus, we have:

Set (A)

BB: I am appeared to redly

BB: I am appeared to bluely
B1: ((I am appeared to redly & bluely) → it is Friday)
B2: It is Friday

Set (B)

BB': I am appeared to redly

BB': I am appeared to greenly
B1': ((I am appeared to redly & greenly) → it is not Friday)
B2': It is not Friday

If we envisage that the foundationalist has basic beliefs r, b, and g at the same time, f and ~f are
both justified—assuming modus ponens is truth preservative and B1 and B1' are justified. There
does not seem to be a rational way of adjudicating between (A) and (B) for the foundationalist,
and since belief chains such as this can be infinitely duplicated, there is a normative worry that
foundationalism does not provide a rational way to attempt adjudication between justified
contradictory beliefs. Since basic beliefs are immediately justified, the foundationalist is at
impasse: no logical (since the inference is valid) or foundationalist criteria can adjudicate
between the two belief chains. Thus, MPRn* applies to foundationalism. We can see that this
scenario is possible because modest foundationalists hold that basic beliefs are at least prima
facie justified but can be false (§1.1). Thus, at least one of the basic beliefs is false in (A) or (B).
Furthermore, because basic propositions can be false, it seems that all propositions and their
negations can be inferred from basic propositions at all times (if the foundationalist is right, there
are a myriad—perhaps infinitely many—basic propositions, including false ones). Thus, the
MPRc* is a reductio of foundationalism.85

Moreover, two contradictory proposition chains can terminate in any basic proposition whether
the basic proposition is true—a reductio of infallible foundationalism—or false. Consider two
propositions that are each justified by the same modus ponens inference. If the consequent of one
of the conditional statements is negated, the result is two contradictory conclusions. Thus, we
have:

a ~a
(b & (b → a)) (b & (b → ~a))

Rather than showing that this chain can be infinitely extended (the MPRc of infinitism), if we
stipulate that b is a basic proposition and a represents “any proposition” at all times, we have the
reductio (MPRc*) of foundationalism. However, the foundationalist will protest that one of the
conditional statements must be unjustifiable and therefore all propositions are not justifiable
according to foundationalist criteria. How can the foundationalist argue for one conditional
statement over the other?

The foundationalist holds that propositions are justifiable if they are inferred from a basic
proposition, which means that both conditional propositions in our example (and every
conditional proposition) are justifiable according to foundationalist criteria because we can
deduce conditional propositions from basic propositions. For example, using modus ponens, we
can deduce the conditional statement (b → a) from (c & (c → (b → a))) and (b → ~a) from (c &
(c → (b → ~a))). If c is basic, the foundationalist must accept that both conditionals are
justifiable. Moreover, the foundationalist cannot continue to object to the conditional statements
of my MPRc* argument because every conditional can be deduced from a basic proposition at all
times. For example, (c → (b → a)) can be deduced from (d & (d → (c → (b → a)))) and so on.
Consider a more concrete example:

BP (basic proposition): I have a headache

B1: If I have a headache, then I should take aspirin
B2: I should take aspirin

BP': I have a headache

B1': If I have a headache, then I should not take aspirin
B2': I should not take aspirin

The premises that will be questioned by the foundationalist are obviously the conditionals;
however, we have seen that the conditionals can be deduced from basic propositions. Thus,
consider the justification of the conditional in one of the previous examples:

BP: I am appeared to redly

B1: If appeared to redly, then, if I have a headache, then I should take aspirin
B2: If I have a headache, then I should take aspirin

Thus, using valid modus ponens inferences, I have shown that all propositions and their
negations are justifiable at all times using foundationalist justificatory criteria since the
foundationalist cannot object to any conditional statement used to infer propositions from basic
propositions (because all conditional statements are also justifiable at all times using
foundationalism). We now see that the MPR* applies to foundationalism: all propositions and
their negations are justifiable using foundationalist conditions; hence, foundationalist
justification is arbitrary and the foundationalist cannot adjudicate between any two beliefs. Can
the foundationalist respond?

The foundationalist may respond by claiming that inductive inferences must be used to
adjudicate between proposition chains. It is notoriously difficult to provide an adequate
definition of induction,86 but, roughly, we can say that induction is inferring that past experience
will be the same as future experience and that from several singular past experiences, we can
infer general principles. For foundationalists, past experience is key: they argue that only
inductive inferences from basic propositions (experiential propositions)87 are justifiable.

However, we must first note that “experience” in induction denotes propositions referring to
experience and not the experiences themselves. To see why this is so, consider the interpretation
involved in a propositional “basic” belief of a sensation.88 Take the sensation of water as an
example. A foundationalist may claim that she is appeared to bluely. Already the foundationalist
has focused her attention on one property of the water-sensation to support her belief: the
property of blue supports her belief that she is appeared to bluely and not the properties of cold,
wet, and so on. The focus of attention does not stop there: the foundationalist also hones in on
relevant aspects of this property in order for the sensation to support the propositional belief “I
am appeared to bluely.” Cling (2008) remarks that “So a sensation of [blue] need not support
beliefs about [blue] even if the sensation is evidence for the relevant proposition: [blue] must be
salient in the apprehension” (p. 410). For example, if I am holding my keys, the experience of
*holding my keys* does not support beliefs about my keys unless the experience of *holding my
keys* is salient in the apprehension (it could be the case that, at a certain time, I am looking for
my keys because I believe that I have lost them, even though I am experiencing my keys at that
time because they are in my hand). At the very least, this illustrates the fact that inductions are
inferences from interpreted experiential propositional beliefs to other propositional beliefs, i.e.,
the inductive premises are not experiences: we do not believe everything we experience, so even
if I have experienced, e.g., a *green triangle* several times, I cannot infer any beliefs from those
experiences unless the experiences were salient in my apprehension (i.e., unless I am aware of
my experiencing *green triangles* and have beliefs about *green triangles*). But perhaps the
foundationalist can argue that only experiences salient in one’s apprehension can be used in
inductive inferences, say, because they are conscious experiences (cf. Bonjour). However, any
inductive inference concerning experience must at least involve memory beliefs. For example, to
induce the belief that most triangles are green, I must remember seeing several green triangles.
Further, memories of past experiences are propositional beliefs about past conscious experiences;
they are not conscious experiences themselves (e.g., in remembering a pain, I am not conscious
of experiencing *pain*, I am simply holding a belief that I once had a conscious experience of
pain). Thus, inductions involve propositional beliefs about past experiences, not the experiences
themselves.

Next, note that the infinitist will have the same experiential propositional beliefs as the
foundationalist; the response is therefore also available to infinitism. For example, they will each
believe that they have seen the sun. Thus, if induction involves inference from experiential
propositional beliefs to other beliefs, then ex hypothesi, both foundationalists and pure infinitists
will provide the same inductive inferences: the same experiential propositions will be available
to them (assuming they have the same saliently apprehended experiences). Nevertheless,
inductive inferences can still be contradicted by deductive inferences in both theories (because
any belief can be deductively justified and every proposition is deductively justifiable according
to the MPR and MPR*). Thus, induction is only a response to the MPR/MPR* if it is argued that
inductive inferences based on experiential beliefs (or propositions) are somehow more justified
(or justifiable) than deductive inferences (this response requires the addition of a necessary
condition for justification, which I will consider as a response to the MPR and MPR* in the next
chapter).89

Perhaps the foundationalist could respond by arguing that it is unlikely that the affirmation and
negation of a basic proposition could be justified for the same person at the same time. This
response, if cogent, would hold that while a basic proposition is justifiable (supported by, say, a
sensation), it is unlikely that that proposition’s negation is justifiable at the same time (because it
would have to be supported by the absence of the sensation in question). However, while the
basic proposition, say, “I am appeared to redly,” is available to S at t, also available to S at t is
the proposition “I am not appeared to redly,” which can be inferred from basic propositions using
conditional propositions such as “if I appeared to handly, I am not appeared to redly” (the
conditional proposition is also justifiable since it can be deduced from basic propositions as
well).90

The foundationalist might respond by arguing that propositions about experience can only be
basic, i.e., receive non-propositional support only. This would mean, for example, that
propositions like “there is a red chair” could never be supported by other propositions. It follows
that, since the MPR* still applies to all non-experiential propositions inferred from basic
propositions, the only propositions justifiable for a foundationalist are those that can be
supported by immediate experience, which is implausible. Moreover, the response itself is
implausible. Recall (§4.5.2) that it is generally acknowledged that to hold a basic proposition like
“there is a red chair,” the person must have “supporting” justifiable propositions available to her
that would justify the concepts necessary to form the basic proposition, such as “chair,” “red,”
and so on. The problem with the foundationalist response is this: it must be possible for these
“supporting” propositions to be inferred from basic propositions if they are justifiable (since
concepts, such as “chair,” are not basic), and because these “supporting” propositions, necessary
for the construction and justification of a basic proposition, must be inferred from basic
propositions (if they are justifiable according to foundationalist criteria), all of the “supporting”
propositions are justifiable (since I have shown that all propositions can be inferred from basic
propositions at all times) and therefore all basic propositions are justifiable (since basic
propositions cannot be constructed and are unjustifiable if they are not constituted by justifiable
“supporting” propositions).

Instead, the foundationalist may claim that some conditional propositions are analytic truths, i.e.,
true by definition. For example, if I know the meanings of the words “bachelor,” “male,” etc.,
then I know that “if S is a bachelor, then S is male” simply by the definitions of the terms
involved. (Note that the definitions involved must be known and therefore justified without using
the foundationalist structure since, as shown above, all concepts can be justified using
foundationalist criteria.) This response seems to be prima facie promising because it directly
responds to my argument that all conditional propositions can be deduced from basic
propositions—an important step in applying the MPR* to foundationalism. By adding the
condition that only proposition chains containing analytic truths are justifiable, the
foundationalist may have a response to the MPR*. However, this potential response fails for
several reasons. First, if there are analytic truths, there are not very many, e.g., “if appeared to
triangularly and redly, then I see a triangular red thing” along with most (if not all) empirical
conditional propositions are not analytic truths, so if this response is cogent, only a few
propositions will be justifiable for the foundationalist. Second, the analytic/synthetic distinction
is thought to be severely flawed.91 Finally, this response—using analytic truths—adds a
necessary condition to justification and is equally available to all epistemic theories including
infinitism. For example, an infinitist could add the condition that chains of reasons containing
analytic truths are justifiable whereas chains without analytic truths are unjustifiable. Thus, if
this additional necessary condition successfully defeats the MPR*, it is successful for all
theories, i.e., the MPR and MPR* do not pose problems for any theory, which is what I endeavor
to show. (However, for reasons such as those given above, I think this potential response is
unpromising and will therefore not pursue it in chapter 6 when I discuss the addition of necessary
conditions to solve the MPR and MPR*.)

Finally, the traditional foundationalist may argue that conditional propositions are basic.
However, this response is dubious. To be a response to the MPR*, the argument must at least add
the following: basic propositions are more propositionally justified than propositions deduced
from basic propositions and conditional propositions are basic because they are self-evident. The
first criterion is needed because we can still deduce the negation of a basic proposition (in this
case, the negation of a conditional basic proposition) from another basic proposition. So in order
to avoid the MPR*, the argument must hold that basic propositions are more propositionally
justified than propositions deduced from basic propositions. The second criterion is needed
because conditional propositions cannot be justified by immediate experience, i.e., immediate
experience does not support conditional propositions, so the conditional propositions must be
self-justified in order to be basic.

The first criterion seems doubtful: it is a strong skeptical claim about deduction to hold that it
does not fully preserve the truth of the basic proposition. Although using a long deduction may
make the deduction liable to error (because long deductive chains are difficult for reasoners),
there does not seem to be anything about deduction simpliciter that suggests that it does not
preserve truth. I find the second criterion even more troublesome: it is clearly a case of circular
reasoning (e.g., “p is self-justified” means that “p therefore p”) (§§3.1 and 4.2).

Thus, the MPRc* is a reductio of traditional foundationalism: all propositions and their negations
can be deduced, using modus ponens, from a justified basic proposition at all times, which entails
the MPRn* as well. It seems that either the MPR* applies to foundationalism or foundationalism
is in need of a meta-justification to answer the MPR*.

Do meta-justificatory foundationalist (hereafter, MJF) theories suffer the same fate as traditional
foundationalism? Recall that MJF theorists hold that basic beliefs are truth-conducive because
they have some truth-conducive property, P (e.g., basic beliefs result from a reliable belief-
forming process). I will consider three variations of MJF and Aikin’s impure infinitism in turn.
First, Aikin provides a MJF argument (§3.4) that holds that foundationalist responses to meta-
challenges to basic propositions are separate chains of reasoning supporting the claim that the
reasoner is competent, e.g., S could argue that she responsively regards her non-doxastic
psychological states. Second, recall (§4.2) that Feldman, Cling, and Howard-Snyder and
Coffman argue that a meta-justification will provide more justification for a basic belief (hence,
continuing the regress) but that it is unneeded justification, i.e., the experience itself (e.g., having
a headache) adequately justifies the basic belief (or proposition). Perhaps these varieties of
foundationalism can argue that one of the basic propositions supporting contradictory
propositions is justifiable and the other is not (by continuing the regress), and if so, the
foundationalist may have a means of adjudication. Finally, consider a reliabilist MJF.92 This view
holds that basic propositions are the products of reliable belief-forming processes, i.e., processes
that often produce true beliefs (or propositions) while rarely producing false beliefs (or
propositions).

All MJF theorists face the same choice: either the meta-justifications provided terminate in
another basic proposition (and are therefore justifiable by foundationalist criteria)93 or continue
infinitely (and are therefore justifiable by impure infinitist criteria).94 If the meta-justification is
infinite, we can simply apply the MPR of infinitism to conclude that all propositions and their
negations are justifiable at all times. Thus, the meta-justification cannot extend infinitely. Can
meta-justifications terminate in other basic propositions? To avoid circularity, the basic
proposition must be the result of a different belief-forming process (or, for Feldman’s and
Aikin’s views, the meta-justification must terminate in a different basic proposition). However,
this response is still in danger of circularity: there must be an infinite number of belief-forming
processes and an infinite number of basic propositions to avoid circularity. And even if this
response is not viciously circular, my other responses will apply: all basic propositions are
justifiable at all times because a meta-justificatory proposition chain terminating in another basic
proposition is available to support every basic proposition. For example, (A) and (B) are both
justifiable if the basic propositions in (A) and (B) are propositionally justified by meta-
justificatory proposition chains that terminate in different basic propositions, and so on ad
infinitum.

To clarify this response, consider the use of tea leaves as a belief-forming process. Though most
of us think that reading tea leaves is an unreliable belief-forming process, there is nothing
preventing clever tea leaf readers from providing a meta-justification of this belief-forming
process that they claim produces spontaneous, reliable, and therefore, basic, propositions. The
tea leaf readers will then claim that this meta-justification eventually terminates in another basic
proposition (thereby justifying it), which also has a meta-justification available for it, and so on
ad infinitum. Even if a foundationalist tea leaf skeptic thinks that tea leaf reading is nonsense and
that the meta-justification for it is nonsense and the basic proposition that justifies the meta-
justification is nonsense, the foundationalist has no foundationalist grounds for the claim that it
is nonsense if the inferences used are adequate. Who is correct, the foundationalist tea leaf reader
or the foundationalist tea leaf skeptic? Foundationalist criteria provide both with propositional
justification (the skeptic will provide meta-justificatory defeaters that are justifiable in the same
way as the tea leaf reader) and therefore provide neither with a rational means of adjudication
between doxastic beliefs. Thus, MJF theories cannot respond to the MPR*: MJF arguments
cannot adjudicate between beliefs and cannot distinguish justifiable foundationalist proposition
chains from unjustifiable foundationalist proposition chains.

Finally, we see that Aikin’s response of impure infinitism likewise fails: as we have seen, all
propositions are justifiable at all times using basic propositions, so all propositions are justifiable
at all times using infinite regresses containing basic propositions.

Thus, MJF methods of adjudication and impure infinitist methods simply turn into pure
infinitism by losing the need for basic beliefs: basic beliefs do not aid adjudication normatively.
We saw in the case of (A) and (B) that the foundationalist will be at an impasse: there are no
normative foundationalist means of adjudication available (i.e., the MPRn* applies to
foundationalism). Moreover, we saw that it is conceptually possible that all propositions and
their negations terminate in basic propositions (including meta-justifications) and are therefore
justifiable for the foundationalist at all times (i.e., the MPRc* applies to foundationalism).
Therefore, the MPR* applies to foundationalism (including impure infinitist and reliabilist
versions).

5.3 HOW A MODIFIED MPRAPPLIES TO COHERENTISM

A modified MPR also applies to coherentism. Recall (§1.2) that coherentist views purport that a
proposition p is propositionally justified for person S at time t just in case p coheres with a
coherent system of propositions available to S, and S is doxastically justified in the belief that p
at t just in case p coheres with S’s system of beliefs.

Consider the fact that it is possible that both X and Y are maximally coherent belief systems that
are equally coherent, simple, etc.95 It is also possible that one system consists of mainly false
beliefs, or all false beliefs, while the other consists of mainly true, or only true beliefs.
Furthermore, it is possible that the belief that p coheres with system X and the belief that ~p
coheres with system Y, in which case coherentism provides no rational way to adjudicate
between the two beliefs (or the two systems). Thus, the MPRn* applies to coherentism.

Moreover, it is conceptually possible that all propositions and their negations are justifiable at all
times according to coherentism because every proposition is a member of a maximally coherent
system of propositions—e.g., the proposition “p and ~p” can be explained by (or logically follow
from) other propositions including “the principle of non-contradiction is false.” Thus, the MPRc*
is a reductio of coherentism. Perhaps the coherentist can add a necessary condition to avoid the
MPRc*—I examine responses to the reduction arguments in chapter 6—what is important is that
a reductio similar to the MPR applies to coherentism so, like infinitism, the theory must attempt
to avoid reductio.96 Many philosophers97 have examined problems with coherentism such as this,
so I will not pursue the argument any further.

I have attempted to show that the MPR is not a special problem for infinitism. The proponents of
the theories I have examined may find a way to avoid the modified MPR arguments, but I have
shown that the MPR (modified) is a problem for these theories and therefore not a special
problem for the infinitist.98 I have therefore successfully weakened the MPR by showing that the
leading theories of epistemic justification are not unscathed by some form of the reductio.
Although the problem is weakened for the infinitist, more important to the infinitist is a response
to the MPR. In the next chapter, I consider possible responses to the MPR—which will apply,
mutatis mutandis, to the MPR*—and argue that, with the addition of a justificatory condition,
the reductio arguments do not pose problems for any theory of epistemic justification.

CHAPTER 6

A SOLUTION TO THE MPR

To examine the possible responses to the MPR, I examine each constituent problem separately
(the MPRc and MPRn). In section one, I respond to the normative worry behind the MPR but
conclude that we must first solve the MPRc for the response to apply. Thus, in section two, I
examine possible responses to the MPRc. I first present Sosa and Klein’s response to the MPRc.
Next, I present Cling’s argument that any argument—including Sosa and Klein’s argument—
against the MPRc necessarily fails. Since I argued in the previous chapter that the MPR* is a
reductio of the leading theories of epistemic justification, I have also shown that Cling’s
argument applies to the leading theories. Skepticism threatens.

However, I argue that, with the addition of a necessary condition of justification, any theory of
justification avoids the MPR (or MPR*) and Cling’s argument. Thus, I conclude that the MPR
and Cling’s argument should not convince us, which shows that what many take to be the main
reasons for rejecting pure infinitism (the MPR and Cling’s argument) are not, in fact, reasons to
reject the theory.

6.1 INFINITISM AND THE NORMATIVE MPR WORRY

To avoid the normative worry behind the MPR, the infinitist will provide a normative strategy
for adjudication between doxastic belief chains.

Infinitism holds that a subject should continue to consider reasons for each inference chain (e.g.,
for p and ~p) until a contradiction, repeated belief, etc., is found, which will provide a means of
adjudication. As argued in §§5.2 and 5.3, this method is not available to the coherentist and
foundationalist, who will only seek either coherency or basic beliefs for adjudication, but as I
argued, these methods are unsuccessful: the coherentist and foundationalist will be at an impasse,
whereas the infinitist can always reason further. (If coherentists or foundationalists reason
further, they become de facto infinitists.) However, the normative worry supervenes on the
conceptual reductio (the MPRc). If infinitism cannot distinguish justificatory regresses of
reasons from arbitrary regresses of reasons (the MPRc), then it does not matter if the infinitist
can always reason further: providing more reasons for beliefs will not lead to justified beliefs
since we can provide an infinite number of reasons for every belief (the MPRc). Thus, without
first solving the MPRc, the infinitist response to the MPRn will only show—at most—that
infinitism is a preferable normative theory of rationality (because the infinitist can always reason
further).

6.2 RESPONDING TO THE MPRCAND MPRC*

We have seen that, for the infinitist response to the MPRn to obtain, we must first solve the
MPRc. Thus, the focus of solving the MPR should shift to the MPRc.

Sosa (1980) argues that what I call the MPRc of infinitism fails to distinguish between regresses
of justification that are “actual” and those that are “merely potential” (pp. 12-13). He claims that
actual justificatory regresses contain only justified members. Klein (1999) further expounds this
response by claiming that for every contingent proposition, there is an infinite regress but that
only some regresses contain reasons, i.e., the “actual” regresses that provide justification. In
other words, the propositions in a justificatory chain must be available to a person as reasons
(pp. 311-312). More precisely, some facts (objective and subjective availability)99 make it the
case that a proposition is a reason.100 For example, it may be the case that the proposition “I lost
my keys” is a reason for the proposition “I cannot drive my car to work” (i.e., facts about the
world make it the case that the former proposition is a reason to believe the latter), whereas
“Mars is a planet” is not an available reason for the proposition “I cannot drive my car to work”
(there are no facts that make it the case that the former proposition is a reason for the latter).
Thus, the first two propositions are members of an “actual” regress, and therefore provide
propositional justification, whereas the second group of propositions—even though they may be
members of an infinite regress—do not provide propositional justification. Since the MPRc holds
that all propositions and their negations are justifiable at all times using infinitism, the MPRc no
longer applies to infinitism: Sosa and Klein distinguish between justification-affording regresses
and arbitrary regresses, so every proposition and its negation is not justifiable at all times using
infinitism because some propositions will not be the first members of an infinite chain of
reasons. Since infinitism holds that the infinitist structure of justification is only a necessary
condition of justification, the MPRc is only a reductio if another necessary condition cannot
distinguish between justification-affording and arbitrary regresses: if infinite reasons are not
sufficient for justification, then not every proposition is justifiable by an infinite regress—more
is required for justification. By adding a necessary condition to infinitist justification (that only
regresses that contain only reasons provide justification), the reductio does not apply to
infinitism (this response is also available, mutatis mutandis, to other theories of justification; it
seems that the reductio arguments only apply to justificatory theories if the theory does not
include a condition specifying what counts as a reason for another proposition). Thus, the
proponents of the MPRc are mistaken that it is a reductio of infinitism.

However, when Cling (2008, p. 419, 2004, especially pp. 117-121) presents the MPR, he argues
that any necessary condition (including Sosa and Klein’s condition), call it Φ, that is added to
infinitist justification to distinguish between justification-affording and arbitrary regresses such
that it would explain the connection between infinitist justification and truth would undermine
the regress requirement and therefore infinitism can never avoid the MPR. In other words, since
the MPR shows that infinite regresses cannot explain the connection between justification and
truth because they propositionally justify all propositions at all times, the infinitist must add
another necessary condition, Φ, that will distinguish justified propositions from arbitrary
propositions in a way that explains justification’s connection with truth, but infinitism cannot add
Φ without undermining the regress requirement. Cling (2004) claims that either p’s having Φ
requires justified beliefs (propositions) or it does not (because condition Φ must be truth-
conducive, i.e., “if a belief [or proposition] that p has Φ, then p must be true, likely to be true, or
reasonably believed to be true” [p. 118], Φ must explain why a proposition that has Φ is truth-
conducive). If it is the former, the infinitist account is either (1) viciously circular or (2) Φ is
sufficient for justification, or (3) both. If the following occurs, then (1) applies: if p’s having Φ
requires justified propositions, the infinitist must use the infinitist structure of justification to
justify the propositions, so the account simply assumes the infinitist structure of justification,
which is viciously circular. (2) is problematic because, if p’s having Φ is sufficient for
justification and is not providing infinite reasons, then the motivation for requiring the regress of
reasons is undermined: if Φ is truth-conducive, the infinitist structure is unnecessary for finding
true propositions. For example, if Φ is “propositions produced by tea leaf reading” and is
sufficient for justification, then providing more and more reasons for a proposition is
unnecessary for justification. Finally, it could be the case that (3) obtains: if Φ is “providing
infinite reasons for p” and is sufficient for justification, then the account is both viciously
circular (it simply assumes the infinitist structure) and sufficient for justification. This account is
not only circular, it does not solve the MPR: we saw in the previous chapter that providing
infinite reasons for a proposition does not separate justification-affording regresses from
arbitrary regresses (the MPRc). Cling finds the latter—p’s having Φ does not require justified
propositions—equally unacceptable according to infinitist standards: any response that claims
that p’s having Φ does not require justified propositions undermines the regress condition that
justified propositions require reasons. For instance, if p’s having Φ is a non-doxastic, non-
propositional, or non-mental state, then justified propositions do not require reasons: p’s having
Φ justifies p without reasons, thus, reasons are not necessary for justification as the infinitist
maintains.

Since I have shown that modified MPR arguments apply to the leading theories of justification, I
have also shown that Cling’s argument—that the MPR can never be solved—applies to the
leading epistemic theories because, if Cling is right, the addition of Φ as a necessary condition of
justification to explain justification’s connection with truth would undermine the other necessary
conditions of the theories.

Thus, like infinitists, foundationalists must choose whether p’s having Φ requires justified
propositions or not. If p’s having Φ requires justified propositions, then the account is viciously
circular (foundationalists require that justified propositions are basic or are inferred from a basic
proposition, so if p’s having Φ requires justified propositions, the foundationalist structure must
be assumed), Φ is sufficient for justification (because Φ is truth-conducive, the foundationalist
structure is unnecessary),101 or both. If p’s having Φ does not require justified propositions, then
either p’s having Φ is sufficient for justification, so that the motivation for basic propositions
(and propositions inferred from basic propositions) is undermined (i.e., basic propositions are
unnecessary since p’s having Φ does not require them), or the account is viciously circular (i.e.,
p is basic because it has Φ; and in which case, the MPR* still applies).

Coherentists have the same choice: either p’s having Φ requires justified propositions or not. If it
is the former, then either the account is viciously circular, Φ is sufficient for justification, or both
(since the argument pattern is the same as above, I leave it up to the reader to fill in the details).
If it is the latter, then p’s having Φ undermines the coherence condition and the coherentist claim
that only reasons can justify propositions (if p’s having Φ does not require justified propositions,
then a proposition can be justified without cohering with other propositions and without
providing reasons for the proposition).
Thus, we see that my arguments that modified MPR arguments apply to the leading epistemic
theories of justification, together with Cling’s argument, provide a new general skeptical
problem for all leading theories of epistemic justification: it seems that epistemic theories of
justification can never avoid the MPR or modified MPR arguments. Can theories of epistemic
justification respond to this new skeptical worry?

I believe that theories of epistemic justification have a readily available response to the new
skeptical worry, which shows that, pace Cling’s argument, the MPR (and MPR*) are solvable.
Cling assumes that p’s having Φ must make p truth-conducive if it is to distinguish justification-
affording regresses from arbitrary regresses in a way that explains the connection between
justification and truth. However, the proponents of the theories discussed should hold that truth-
conduciveness arises from sufficient justification.102 That is to say, necessary conditions of
justification—like infinite chains of propositions and Φ103—are not necessarily truth-conducive,
but when a proposition satisfies all necessary and jointly sufficient conditions of justification, the
proposition is likely to be true (i.e., if the proposition satisfies the jointly sufficient conditions of
justification, it is truth-conducive). The infinitist, for example, could hold that providing more
and more reasons for a proposition is necessary for justification, and together with necessary
condition Φ (e.g., all propositions in the infinite chain are available as reasons), is sufficient for
justification and hence truth-conduciveness. Thus, if both conditions of justification are met, then
p is truth-conducive. Φ must distinguish between justification-affording regresses and arbitrary
regresses, but p’s having Φ does not need to be truth-conducive. Thus, pace Cling, it is not
necessary that “if p has Φ, it is true, likely to be true, or reasonably believed to be true.” Instead,
infinitists can claim that p is true, likely to be true, or reasonably believed to be true only if p is
the first member of an infinite set of non-repeating propositions and each member in the set of
propositions has Φ.104 Indeed, most epistemic theories of justification hold that justification arises
when and only when a proposition satisfies the sufficient conditions of justification; I am simply
claiming that if a proposition is sufficiently justified, then it should be thought truth-conducive,
but that it is not necessary that the individual necessary conditions of justification (e.g., Φ) be
truth-conducive, which is plausible because truth-conduciveness should accompany justification.
That is, since justification arises when and only when the sufficient conditions of justification are
met, we should expect truth-conduciveness to arise if the sufficient conditions of justification are
met and not to arise when only one condition of justification is met (e.g., Φ).105

Consider a more concrete example: Klein holds that p has property Φ if and only if p is available
as a reason for another proposition p2. Should Klein hold that if p has Φ, p is truth-conducive?
No. Φ simply means that p is a reason for another proposition, so there is no reason to believe
that p is truth-conducive: p may not be the first member of an infinite chain of reasons that each
have Φ (i.e., p may not satisfy the other necessary condition of justification). It could be the case
that p is a reason for p2, but eventually the chain of reasons repeats, contains contradictory
propositions, etc., which means that p or p2 may not be truth-conducive. Thus, we should think
that a proposition is likely to be true only when the proposition meets the sufficient conditions of
justification. What does this mean for Cling’s argument?

First, recall that Cling argues that since the MPR shows that all propositions are justifiable and
therefore infinite regresses are not truth-conducive, infinitism needs an added necessary
condition to explain why infinitist justification is truth-conducive. However, since I have shown
that only sufficient justification should be thought truth-conducive, Cling’s argument is reduced
to the claim that infinitists must explain why sufficient justification is truth-conducive. This is a
well-known and difficult problem for all theories of justification.

Theories of justification have attempted to explain justification’s connection with truth in many
ways (e.g., arguing that justification entails truth). Despite the many attempts, it seems that no
theory adequately explains justification’s connection with truth. For example, Stewart Cohen
(1984) argues that the extant theories of justification either explain justification’s connection
with truth in a trivial way, which leads to several devastating problems for the theory in question,
or they fail to explain justification’s connection with truth. Thus, if theories of justification must
explain justification’s connection with truth, as Cling, among others, argues, then infinitism must
find a way, like all theories of justification, to explain this connection adequately; but as argued
above, we should not think that a merely necessary condition should be truth-conducive or
explain justification’s connection with truth.

Next, because Cling’s argument is reduced to a well-known problem for all theories since I
argued that we should not expect a necessary condition to be truth-conducive, it follows that
Cling’s argument does not show that the MPR (and MPR*) can never be solved. Recall that the
MPRc is a reductio in the formal sense that it assumes infinitism and derives a contradiction; it is
also a reductio in the non-formal sense that it assumes infinitism and derives the absurd
consequence that all propositions are justifiable at all times. We have seen that necessary
condition Φ, when added to infinite non-repeating regresses, successfully distinguishes between
justifiable and unjustifiable regresses. Moreover, a contradiction cannot be derived from
infinitism since condition Φ shows that not all propositions are justifiable at all times. Thus,
Cling’s argument will show that Φ cannot solve the MPR without undermining the regress
requirement only if (1) justification must be connected with truth, (2) a theory of justification
must explain justification’s connection with truth, and (3) an added necessary condition that
distinguishes between justificatory regresses and arbitrary regresses must be truth-conducive and
explain why justification is truth-conducive (since regresses themselves are not truth-conducive).
Since I have shown that (3) is false, Cling cannot claim that Φ necessarily undermines the
regress requirement.

However, Cling (2008) also argues that infinitism cannot use condition Φ since Φ must be
justified (have reasons available for it) “[i]f it is not arbitrary from the believer’s own point of
view” (p. 407). As we have seen, this leads to the infinitist’s begging the question or violating
the regress condition, including Klein’s intuitive reasoning principle—an important step in
establishing the regress condition—PAA (§3.1).106

Since we are discussing the epistemic justification of propositions, I take “all x” to mean “all
propositions” and the regress condition to apply to justifying propositions (and propositional
beliefs). The property of propositions, Φ, however, is not a proposition, so the infinitist need not
have reasons available to believe Φ: Φ is not arbitrary from the believer’s point of view, nor does
Φ violate the regress requirement. Furthermore, like the justificatory criterion “providing infinite
reasons for a proposition,” Φ is an epistemic concept, a criterion of justification. Thus, if Cling
continues to object that the justificatory criterion Φ is arbitrary from the believer’s point of view
since the infinitist cannot justify Φ without begging the question or undermining the regress
requirement, then he is offering the well-known “problem of the criterion” argument, which is
equally applicable to the criterion “providing reasons for a proposition” and all justificatory
criteria.107

Thus, we see that Cling’s arguments—leading to a “new” skeptical worry—and his possible
responses, collapse into other skeptical arguments, which pose problems for all epistemic
theories and are beyond the scope of this inquiry. Consequently, when necessary condition Φ,
which distinguishes justifiable propositions from unjustifiable propositions by providing criteria
for the concept of “an available reason,” is added to any theory of justification, the theory easily
avoids the MPR and the MPR*.

6.3 NECESSARY CONDITIONS OF JUSTIFICATION

Proponents of the MPR often cite it as their main reason for rejecting pure infinitism in favor of
another theory of justification. However, I have shown that their argument is mistaken at several
points. In the previous chapter, I argued that the MPR* applies to modest foundationalism and
coherentism—two of the most widely accepted theories of epistemic justification—and the MPR
is therefore not a special problem for infinitism. Subsequently, in the present chapter, I showed
that infinitism has a plausible response to the normative worry behind the MPR, but that the
MPRc must be solved before the response applies. I then showed that Cling’s argument—that
infinitism is a hopeless theory because it can never solve the MPR—also applies to the leading
epistemic theories; thus, I provided a new general skeptical worry. Finally, I provided a solution
to the MPRc/MPRc* (and thus the MPRn/MPRn*) and the Cling-type skeptical worry available
to any theory of epistemic justification, thereby dissolving the MPR and the Cling-type skeptical
worry. Thus, I have shown that what many take to be the main reason for rejecting infinitism (the
MPR) is unfairly applied only to infinitism (even though modified versions of the argument
apply to the leading epistemic theories as well), and more importantly, I showed that the
arguments given by the proponents of the MPR (the MPR and Cling’s argument) and the MPR*
should not convince us that we should reject any theory of justification (including infinitism).

Since the MPR need not worry us and since I argued in Chapter 4 that infinitism is the only non-
skeptical solution to the problem of regress, we can conclude that ceteris paribus infinitism is
preferable to the leading epistemic theories of justification (including skepticism).108 But what
kind of infinitism should we accept?

It seems that our inquiry has not led us to one comprehensive infinitist theory. Rather, it has
shown how flexible infinitism is as a theory of justification. In Chapter 3, I discussed several
different approaches to infinitism, but I have not refuted any necessary aspects of these
approaches to render one theory of infinitism. The only conditions of the theories I did refute
were conditions that ruled out other forms of infinitism (e.g., Akin’s condition that infinitists
must include basic beliefs to avoid the MPR). Indeed, there is much that infinitism is not
committed to but can admit if need be. For example, although I argued that basic beliefs cannot
stop an infinite regress and cannot solve the MPR, it does not follow that beliefs cannot receive
non-doxastic justification (I have not refuted impure infinitism). It is possible that some beliefs
can receive a degree of support from non-doxastic sources such as an awareness of experience.
So if there is good reason to believe that beliefs can receive non-doxastic justification, the
infinitist can, and should, account for this. Thus, I have only shown that the infinitist need not be
committed to beliefs receiving non-doxastic support. A further example is the choice between
emergent infinitism and transfer infinitism. We saw (§3.1) that Klein accounts for justification as
an emergent property to avoid the “no starting point objection,” but there may be other ways of
avoiding this objection,109 so infinitism is not committed to the emergent or transfer depiction of
justification. Instead, the infinitist can choose whatever account turns out to be correct. The same
goes for coherency of beliefs, externalist aspects of justification, the ability to answer challenges
to a belief (§3.2), a rejection of ought-implies-can,110 and so on: infinitism can include or exclude
these possible conditions of justification depending on whether they plausibly strengthen the
theory of justification (whether they are correct) or not. Thus, I believe that I have shown
infinitism to be a versatile theory necessary for solving the regress paradox.

We have seen what the infinitist is not committed to, but what are the commitments of
infinitism? Our analysis rendered several necessary conditions of justification. The infinitist must
hold that S is doxastically justified in the belief that p at t to the extent that:

1. S provides reasons for the belief that p along an infinite path of non-repeating
reasons.111

2. All beliefs in the chain of reasons justifying p must have Φ.

3. S is epistemically responsible in believing that p.

And, S is propositionally justified in the belief that p at t just in case:

1. p is the first member of an infinite non-repeating chain of reasons available to S.

2. All propositions in the chain of reasons justifying p have Φ.

3. S is an epistemically responsible in holding beliefs.

Although I have not argued for the third condition, it is clear that some account of epistemic
responsibility is needed to exclude guessing reasons, improperly producing reasons, not striving
for truth and the avoidance of error, and so on.112 Thus, it is clear that some account of epistemic
responsibility is necessary for a theory of justification. Next, as argued above, the second
condition is necessary to separate justification-affording regresses from arbitrary regresses.
Finally, although we have seen several variations of the first condition (in chapter 3), e.g., the
infinitist may hold that the belief that p is more justified than the belief that r even though r has
more reasons supporting it (§§3.2 and 3.4), the infinitist must hold that, ceteris paribus,
providing more reasons increases justification and that the chain of reasons for a justified belief
is endless and non-repeating.

These are the conditions necessary for solving the regress paradox, which concludes that the idea
of having justified beliefs is paradoxical. By showing that obstacles to an account of justification,
such as the regress paradox, can be overcome by accepting infinitism, I have shown that we can
have the system of justified beliefs sought-after since the beginning of the inquiry, albeit not the
indubitable justification, e.g., Descartes desired. Moreover, this account of justification should
normatively guide the structure of our beliefs. When we examine our beliefs or have them
questioned by an interlocutor, we ought to follow the infinitist structure of justified belief. We
should recognize that every defeasible belief is in need of justificatory support from other
beliefs: no defeasible belief is sacrosanct. Although we may never completely justify all (or any)
of our beliefs, this is simply to recognize our finitude. Yes, we are finite, but provisionally (or to
a degree) justified beliefs can still provide the rewards of justified beliefs (§1): we may achieve a
rational system of beliefs which include instances of knowledge. But why strive for knowledge
and rationality?

Epistemic rationality serves many purposes. Some are pragmatic, for instance, some sort of
justificatory community113 of reasoners, where reason arbitrates between conflicting interests,
seems necessary for a just polity—reason allows for this just community because it is impartial,
and all can use reason to judge the claims of others. Other purposes of epistemic rationality are,
inter alia, epistemic. That is, rational belief, or justified belief, seems necessary for knowledge.

Whether we strive to know why there is something rather than nothing, what is the fastest way to
work, or why some finches have larger beaks than others, every person strives for knowledge. As
Aristotle famously claims in the Metaphysics, “All men by nature desire to know” (980b23);
thus, it behooves us to have a tenable account of knowledge to guide our inherent desire. Among
other things, I hope to have contributed both to an account of knowledge and rationality by
showing that justified belief is attainable and by offering plausible conditions for how to attain it.

1 This is according to the standard account of knowledge as justified, true belief.

2 The question of whether there are beliefs known with certain is largely unsettled since there
does not seem to be an unproblematic account of certainty. See, for example, Baron Reed (2008).
I assume that knowledge and rational belief do not require certainty for two reasons: (1) as stated
above, certainty is an obscure concept; and (2) I believe it would be tantamount to accepting
skepticism. That is, if we can attain certainty, it is clear that only few of our beliefs can be known
with certainty, which means that most beliefs that we think are rational and known are irrational
and unknown. Rather than attempt to refute this form of skepticism, I assume that knowledge
and rationality do not require certainty and I attempt to provide an account of defeasible justified
belief, which helps to show that this skeptical view is implausible.

3 He often calls it “real knowledge” in that it is the highest form of knowledge. See Klein (1999),
(1983), and (2007b).

4 I say “epistemic rationality” since it could be pragmatically rational to believe something that
is epistemically irrational. For example, if you are lost without some form of communication
(e.g., a cell-phone) in a remote wilderness with a life threatening injury, it is not epistemically
rational to believe that you will survive: the odds are not in your favor. However, it is certainly
pragmatically rational to believe that you will survive: those who believe they are doomed do not
usually fair well in these sorts of situations—remaining optimistic serves several pragmatic
functions.
5 My use of “justifiable” is similar to Klein’s (2007a). Since the actual justification of beliefs
(doxastic justification) requires that one provide reasons for the belief in need of justification, a
belief’s being propositionally justified means that the belief is justifiable (capable of being
doxastically justified) but is not actually (doxastically) justified until reasons are provided for the
belief. In other words, propositional justification is a purely conceptual way discussing the
structure of justification.

6 For example, Bonjour (1985) claims that the problems surrounding the problem of regress are
“perhaps the most crucial in the entire theory of knowledge” (p. 18).

7 My distinction between “traditional” and “meta-justificatory” foundationalism follows Jeremy

Fantl’s (2003, pp. 539-540).

8 Lehrer (1974) provides a similar example in (p. 96).

9 The possibility of a belief justifying itself is discussed in §3.1 and §4.2.

10 Here I am using Roderick Chisholm’s terminology of using non-comparative terms to

indicate the foundationalist contention that some perceptions are directly apprehended. See
Chisholm (1966, Chapter 2).

11 For example, see Richard Feldman (2003, p. 51).

12 There is no reason to suppose that a system of beliefs must equal the entire set of a person’s
beliefs at a time. Instead, it could be the case that a belief under the purview of science must
cohere with the person’s set of science-related beliefs, and so on. For an exposition of this
account, see, e.g., Bonjour (2003a, pp. 44-46).

13 I will not describe coherentist theories where the reasons are literally circular since these
views are implausible (I discuss this point in §3.1 and §4.2). Also note that the coherentist
condition—like the infinitist condition—is most often taken to be a necessary but insufficient
condition of justified belief.

14 It is generally acknowledged that coherentists have not given a precise account of the nature
of coherence (e.g., see Bonjour, 2003a, pp. 46-48). For our purposes, this sketchy account of the
nature of coherence will suffice.

15 Many coherentist theories are similar to infinitist theories. The main difference is this:
coherentism holds that justificatory chains of reasons are circular or that justification is holistic
(a matter of whole systems of beliefs), whereas infinitism holds that justificatory chains of
reasons are linear and non-repeating.

16 For example, Scott Aikin (2005, 2008, & 2009) argues for an impure infinitism, which is a
mix of infinitism and foundationalism. See §3.4.
17 I will roughly follow Klein’s (2007b) interpretation of Aristotle. He notes other possible
readings in footnote 6 of (2007b).

18 Klein espouses this reading of the Pyrrhonians. See Klein (2004).

19 Cf. Peirce (1931-1958, 5.318).

20 For example, Peirce (1931-1958) states that “all our knowledge rests on perceptual
judgments” (5.142).

21 I have roughly followed Aikin’s interpretation of Peirce. As Aikin notes, Peirce loses the
motivation for his infinitism, and although he no longer explicitly endorsed the view, he may still
have held it in his later works. See Aikin (2009).

22 For example, Dale Jaquette (1996) develops a theory of infinitism, but it is largely based on a
rejection of the semantics of classic logic. Considering Jaquette’s theory, Aikin (2005) remarks,
“I will show that a more modest means of defending infinitism is possible” (p. 193). Following
Aikin, I will examine these more modest means of defending infinitism.

23 The series of papers includes Klein (1998, 1999, 2003, 2005a, 2005b, 2007a, and 2007b).

24 I am not alone in this estimate. Turri (2009a, p. 209, 2009b, 157) makes similar claims.

25 See §1.

26 Klein also claims that reasons must be “undefeated,” but what this entails is beyond the scope
of this work: it is meant to avoid the Gettier problem, which I do not have space to discuss. See
Klein (2007b, p. 6) and Edmund Gettier (2009).

27 Cf. Klein (2007b, p. 16). The distinction between transference and emergence of justification
is madeby, e.g., Bonjour (1985) and Sosa (1980).

28 Klein provides several necessary conditions of epistemic responsibility, but he states that he
does not provide a full account of the concept. For examples of more extensive analyses of the
concept of epistemic responsibility, see Sosa (1974, p. 117) and Angelo Corlett (1996, pp. 98-
104).

29 Recall that justification, for Klein, comes in degrees (is relational) and is an emergent
property. Thus, doxastic justification emerges in degrees as the chain of reasons increases.

30 Turri (2009a, especially pp. 210-215) stresses this point. He argues that simply producing
reasons is not enough; S must “properly produce the relevant reasons” (Turri, 2009a, p. 215).

31 In response to arguments claiming that an infinite number of beliefs are not subjectively
available to persons, Klein claims that although we have a finite vocabulary, we use indexicals
(presumably we can say that an object has color c and have infinitely many objects with color c
available to us) and therefore have the capacity to think about each discrete indexical. See Klein
(1999, pp. 307-308).

32 To my knowledge, Klein first introduces PAA in Klein (1999, p. 299). He then, in Klein
(2003, p. 726), adds “and there is no last reason in the series” to PAA.

33 These “holistic” coherentists hold that propositions are justified in virtue of their membership
in a coherent set of propositions (§1.2), thus, the only reason given for a proposition is that it is a
member of a coherent set, thereby avoiding circularity of reasons, but committing this account to
a kind of foundationalism that must reject PAA. See for example Sosa (2001), Klein (1999, p.
298), and Fantl (2003, p. 540).

34 This can either mean that only beliefs formed from a reliable process are justified (reasons are
not required), or that some beliefs are justified (without reasons) because they are formed by a
reliable process. Klein rejects all forms of reliabilism (i.e., that a belief can be justified without
providing reasons).

35 According to Lehrer, Mr. Truetemp—unbeknownst to him—has an accurate thermometer

implanted in his brain. He thinks unreflectively (no one, including him, knows the reliability of
his temperature gauging), but correctly, that the temperature is such and such degrees. According
to Klein, there is some sense in which Mr. Truetemp “knows” the temperature, but not in the way
that adult humans know. See Klein (1999, p. 302) and Lehrer (2000a, pp. 187-188). It may be
objected that Klein’s argument does not affect all forms of reliabilism. However, I leave this
matter aside because most, if not all, forms of reliabilism are externalist theories, which I do not
discuss. (I consider an internalist reliabilism in §5.2.)

36 See for example, §§1.1 and 1.2, and Bonjour (1985, especially pp. 9-14).

37 As noted in § 1, however, this is a preponderant view among philosophers.

38 The fact that Fantl allows non-reason-giving justification seems to make his an “impure
infinitism.” However, as Aikin notes, “given that Fantl’s requirements are still that infinite
chains of inferential support are the only sources of justification sufficient for complete
justification, the view is exclusive with regards to that element of justification.” Moreover, Fantl
does not “categorically endorse” this view: it is meant as an objection to foundationalists who
allow basic beliefs to be incompletely justified. See Aikin (2008, p. 181 footnote 10).

39 Recall (§1.1) that I closely followed Fantl’s (2003) distinction in my depiction of

foundationalism.

40 Many argue that these coherentist theories are implausible. See for example Fantl (2003, p.
540); Bonjour (1999, p. 123); and recall Klein’s (1999, 2005a) PAC in §3.1. I further discuss
circular reasoning in §4.2.

41 Fantl considers the possibility that different inferences admit of degrees. However, it must be
some property, P, of inferences that make them preserve justification to a certain degree. Thus,
even if the foundationalist claimed that self-justified beliefs are completely justified and
inferences have P completely, the foundationalist would become more justified if he had a reason
to believe that the inference has P maximally and so the regress continues, and the
foundationalist is not completely justified. See Fantl (2003, pp. 551-552).

42 Since Atkinson and Peijnenburg’s view is compatible with Klein’s and Fantl’s, I will not
provide the details of their arguments. See Atkinson & Peijnenburg (2009) and Peijnenburg
(2007).

43 Klein (1998) gives several examples of purported knowledge that we have not completely
reasoned from basic beliefs (if there are basic beliefs): “2+2=4,” “pears don’t typically grow on
orange trees,” etc. (p. 920).

44 I say “actual” as opposed to “potential.” An account of, e.g., potential knowledge, was given
by the Pyrrhonians (§2).

45 For an example of pure foundationalism, see Feldman (2003).

46 For examples of impure theories, see Bonjour’s (1985) coherentism of empirical knowledge
and foundationalism of a priori knowledge; Susan Haack’s (1993) “foundherentism”; Corlett’s
(1996) “social epistemic reliabilism”; and Robert Audi’s (2001) allowance of basic beliefs and
coherence.

47 I say “can be infinite” because Aikin holds that a theory can be weak or strong. Weak theories
hold that some j-trees (see the explanation below) have at least one type of the relevant
justification (e.g., a basic belief or an endless chain) and strong theories hold that all j-trees must
have at least one branch containing the relevant type of justification. See Aikin (2008, pp. 176-
177).

48 Note also that this is a modest foundationalism in that basic beliefs receive defeasible non-
doxastic support and have probabilities less than one.

49 He provides four reasons, but his first and second reasons—that it prevents us from taking
skepticism seriously and that “our cognitive duties do not extend to full-blown epistemic
justification” (Aikin, 2005, p. 206)—are extraneous to my point, as they do not directly deny
ought-implies-can.

50 Aikin gives the example of the selection tasks in Wason and Johnson-Laird (1972). In the
study, the subjects had four cards placed in front of them with e, k, 4, and 7 showing. The
subjects were then told that there are letters on one side of the cards and numbers on the other.
Finally, the subjects were asked to evaluate the validity of the following rule by turning over
only cards necessary to evaluate the rule’s validity: “If a card has a vowel on one side, then it has
an even number on the other side.” 96% of the people in the study gave wrong answers. The
right answer is e and 7: the rule can be invalidated if there is a consonant on the other side of e or
if there is a vowel on the other side of 7 (an application of modus tollens).
51 For Aikin’s (2005) defense of the controversial premise in the set (premise [3]), see (p. 207).

52 To be fair, it is unclear if any philosopher has ever espoused a coherentism that attempts to
solve the regress paradox by rejecting (3). Recall that the holistic variety of coherentism rejects
the linear structure of justification assumed by the paradox (however, the holistic theory may
have to reject premise (2), which is why, e.g., Fantl (2003) claims that holistic coherentism is a
form of meta-justificatory foundationalism [see §3.2]).

53 Recall (§3.2) that Fantl (2003) argues that traditional foundationalists cannot show how
justification admits of degrees and meta-justificatory foundationalism cannot show how
justification can be complete. Thus, if one of these claims is wrong, then a form of
foundationalism (either traditional or meta-justificatory) may be able to solve the regress
paradox. Furthermore, it should be noted that, pace Fantl, some coherentist theories seem to
explain complete justification and explain how justification comes in degrees. See for example
Lehrer (2000a , 2000b) (however, also note that Lehrer’s [2000b] explanation of complete, or
“undefeated,” justification is externalist: certain internal states must match external conditions
for complete justification [p. 651]).

54 Howard-Snyder and Coffman (2006) provide several other arguments against Klein (pp. 535-
556). However, since I argue (in §4.5) that any theory of foundationalism that can avoid rejecting
PAA cannot solve the problem of regress, it is unnecessary for me to provide each way a
foundationalist can avoid rejecting PAA.

55 The coherentist could object that this awareness is part of a system of justified meta-beliefs
specifying the contents of one’s system of beliefs: S is aware that p is a member of S’s system of
beliefs because p is a member of S’s system of meta-beliefs. However, this response leads to
several problems. For example, see Bonjour (2003a, pp. 51-53 & 56).

56 It may be thought that Lehrer’s notion of “self-trust” can solve the regress paradox. Consider
his acceptance argument:

1. I accept p [i.e., I believe p with the purpose of attaining truth and avoiding error].
2. I am worthy of my trust concerning what I accept.
3. I am worthy of my trust concerning my acceptance of p. (from 1&2)
4. I am reasonable to trust my acceptance of p. (from 3)
5. I am reasonable to accept p. (Lehrer, 1997, p. 16)

The key premise in the argument is (2) or “self-trust.” But how is (2) justified? Lehrer states that
(2) is justified by S's other acceptances (it is not "basic") and that these other acceptances must
be supported by self-trust (premise 2). Thus, self-trust is an essential premise in the justification
of self-trust, which violates PAC* (§ 4.2), so self-trust cannot solve the problem of regress.

57 I mean to provide a reductio ad absurdum in the broadest sense of the argument. That is, I
simply want to show how absurd the consequences of this view are (as opposed to deriving a
contradiction from the view).
58 For example, see Bonjour (2003a, pp. 10-12) and Howard-Snyder and Coffman (2006, pp.
542-543).

59 Recall that self-justified beliefs are untenable because they must reject PAC* (§4.2) and that
meta-justificatory foundationalism cannot avoid rejecting PAA without becoming de facto
infinitism (§3.1). Meta-justificatory foundationalists could argue that they accept PAA and argue
(a meta-justification) that the regress ends with basic beliefs, however, this response is plausible
only if there is a plausible candidate for justified basic beliefs that could stop the regress. Thus,
an awareness of experience seems to be the only plausible option for an account of basic beliefs
(meta-justificatory foundationalists could then argue that an awareness of experience justifies
basic beliefs).

60 For a more detailed account of the Sellarsian dilemma, see Sellars (1963, pp. 127-96,
especially pp. 131-132) and Bonjour (2003a, pp. 18-21, 1984, chapter 4).

61 For this argument, see for example, Bonjour (2003a, p. 71), Sellars (1963), and Donald
Davidson (1986, pp. 307-319).

62 Haack claims that her double-aspect foundherentism contains internalist and externalist
elements (e.g., see 1997, p. 31), which is why I do not discuss her theory. However, I should also
note that her explanation of how causation can factor into the justification of a belief cannot be
adopted by an internalist theory. Haack distinguishes between the state and content senses of
belief (S-belief and C-belief respectively). She then argues that a person’s experiential S-
evidence for an S-belief consists of non-belief states of the person (experiences); a causal nexus
of non-belief states causes and sustains the S-belief. However, justification is propositional, so it
must be in terms of C-evidence. Haack (1993) argues that a person’s experiential evidence for p
is the person’s C-evidence that p—a proposition that S is in a certain state or states, states which
constitute the person’s experiential S-evidence for believing that p (p. 80). For example, the C-
belief “A is in the sort of state a normal subject would be in, in normal circumstances, when
looking at a rabbit three feet away and in good light” (Haack, 1993, p. 80), call it p, could be
justified by A’s being in certain states (the perceptual states described in p) that cause and
maintain the S-belief that p (the belief that constitutes A’s C-belief that p). This portion of her
theory may explain how causation can factor into justification; however, it does so at the expense
of rejecting internalism. That a person’s S-evidence is caused by external objects is external to
the person: the S-evidence does not include an awareness of the fact that one is in “normal
circumstances” for an object to cause an S=belief or whether an external object actually caused
an S-belief. As Bonjour (2003a) remarks, “The content of my experience may no doubt incline
me to think that a rabbit is present, but that content obviously cannot by itself reveal that it is in
fact of the sort that is normally (or indeed ever) caused by rabbits” (p. 82). Even if a
foundationalist adopts the double-aspect portion of Haack’s theory (S-evidence and C-evidence)
and characterizes basic beliefs as non-comparative, e.g., “I am appeared to handly,” the
foundationalism will not be internalist. A person with C-belief “I am appeared to handly” is not
aware that the S-belief “I am appeared to handly” is caused by a hand (the awareness in question
is not intrinsic to the C-belief or the S-belief). Nor is the person aware that the S-belief is caused
or sustained by states that are caused by external objects. Thus, I maintain my conclusion that, if
the relation between the content of experience and belief is merely causal, then internalist
foundationalism cannot stop the regress of reasons.

63 A further reason why Bonjour’s foundationalism is a good example is that it closely

resembles a myriad of other foundationalist theories (e.g., Chisholm’s foundationalism and C. I.
Lewis’s foundationalism). See Bonjour (2003a, especially pp. 17 and 75).

64 Recall (§4.2) that non-empirical foundationalist theories are implausible since they must
reject PAC*.

65 This is similar to Sosa’s characterization of Bonjour’s argument. See Sosa (2003b, p. 218).

66 This chapter is in Bonjour (2003b, pp. 173-200).

67 My use of the term “concept” is compatible with the main formulations of “concept”:
concepts as mental representations (mental particulars or symbols, but not necessarily images),
concepts as abilities of cognitive agents, and concepts as Fregean senses (i.e., abstract objects)
(Margolis & Laurence, 2011, section 1). That is, my argument does not rely on one formulation
of “concept” being correct.

68 I simply mean that people had the concept of a surfboard before they regularly began to call
the concept by a name. Also note that many philosophers agree that concepts are independent
and prior to natural language. However, some experiments may indicate that certain tasks (e.g.,
“spatial reorientation that relies on combining landmark information with geometrical
information”) are impaired when the linguistic system is distracted but not when attention is
given to non-linguistic distracters, which leads some philosophers to believe that some of the
concepts involved must be linguistic (Margolis & Laurence, 2011, section 4.2). Even if the latter
philosophers are correct, it does not detract from my point—in the last paragraph of this section
—that terms and concepts are not necessarily isomorphic, i.e., when we think of a concept it is
not necessarily accompanied by a certain term, which is what I hope to bring out by presenting
these cases.

69 Some philosophers argue that the possession of natural language is necessary for having
concepts (e.g., Davidson, 1975). Thus, it is controversial whether animals and children can have
concepts. I believe that the example I give provides some indication that they do have concepts;
however, my point is simply to separate concepts and terms, so my argument does not rely on
children’s or animals’ possessing concepts (this becomes clear in the subsequent paragraph). For
instance, I can hold a conceptual belief, e.g., of what a ball is, while referring to the ball as “box”
(or any unjustified or “incorrect” term)—the mistaken term does not invalidate the conceptual
belief.

70 For brevity, I will call the conceptual aspects of belief “conceptual beliefs” (although
concepts and terms are simply different aspects of propositional beliefs).
71 This could mean that S is using terms she does not usually use to describe the experience in
question (for a private language) or that S is using a term that she and all other English speakers
do not use to describe the experience.

72 John McDowell (1994) proposes avoiding the Sellarsian dilemma and related objection by
holding that experience, e.g., perceptual content, is conceptual. In other words, he argues that the
space of reasons (and justification) is not wider than the conceptual sphere; justification does not,
pace foundationalists, begin with non-conceptual content since experience is conceptual. This
may avoid the Sellarsian dilemma and related objection. However, it does not avoid my
argument: even if perceptual content is conceptual, our terms and concepts are not isomorphic,
and therefore the judgment that our terms match our concepts requires justification.

73 By “refer,” I simply mean that the terms refer to what S generally takes them to refer to (for a
private language) or that English speakers generally use the terms to refer to the experiences,
objects, or concepts in question.

74 Instead of arguing that one’s terms are correctly applied to experiences, one could argue that
his terms correctly apply to or match his empirical concepts. This does not change the argument.

75 To avoid the Sellarsian dilemma, the foundationalist cannot simply claim that an awareness of
experience justifies the basic belief (see §4.4.1). The foundationalist must hold that the belief and
the experience somehow “fit” or “match,” etc., resulting in justification. Since I am applying my
argument to all foundationalist theories, I take the “fit” to be general (i.e., not in the specific way
Bonjour uses it to mean “correctly describes”).

76 If the conjunction of (a) and (c) does not entail (b), then (a) does not entail (b) (if a premise
entails a conclusion that premise conjoined with any other proposition entails the conclusion).
For Lehrer’s argument, see (2000a, pp. 76-81). Sellars (1963) provides a very similar argument
in (pp. 146-148), so I omit it.

77 For the response to Sosa, see Howard-Snyder and Coffman (2006, pp. 556-562). For the
response to Lehrer, see Howard-Snyder (2004, pp. 52-61). The responses are essentially the
same, so I will only discuss the response to Sosa, which is most recent.

78 It is unclear in what sense Howard-Snyder and Coffman use the term “concept.” However, it
seems that any way “concept” is used, there is a distinction between concept and term (see §4.5,
especially footnote 67), and thus, my argument remains the same.

79 See for example Aikin (2008, pp. 82-84); Cling (2004, p. 110, 2008, p. 407 & pp. 418-19);
John Post (1980, pp. 32-35, 1987, pp. 88-91); cf. Paul Moser (1985, p. 67); and cf. James
Cornman (1977, p. 290).

80 Recall that Aikin argues that the motivation for an impure infinitism is that it can solve the
MPR, while pure versions of infinitism cannot. See §3.4.
81 Klein, Aikin, and Cling expose the flaws in the other similar reductio arguments against
infinitism. See Klein (1999, pp. 311-12); Aikin (2005, pp. 198-199); and Cling (2004, pp. 108-
109).

82 Since I use lower case letters for propositions, I changed the letters Aikin uses to lower case
for consistency.

83 I am, as far as I know, the first to notice that the reductio can only apply to propositional
justification. Noting this, I have been careful to keep this distinction straight, especially when
attempting to present charitably the arguments of others (including the MPR), which means inter
alia that when presenting the arguments of others, I explain how they would avoid the
propositional reductio, so I often speak of propositions whereas they speak of beliefs.

84 As stated above, the MPRc applies to propositional justification. Thus, the MPRc* applies to
propositional justification. Many foundationalists and coherentists do not discuss propositional
justification (it is also omitted by many infinitists); however, propositional justification must be
discussed since both the MPRc and the MPRc* are conceptual reductios, they are not reductios
of what people actually believe at a time.

85 The foundationalist may object that the conditional proposition in either (A) or (B) must be
false or that there are finite basic propositions. I address these worries below.

86 For example, see John Vickers (2010, Section 1.2).

87 Although most foundationalists hold that basic propositions are justified by experience, I also
consider non-experiential basic propositions below.

88 I largely follow Cling on this point. However, Cling (2008) uses this notion to argue that
foundationalism finds itself in a fatal dilemma. See Cling (pp. 408-412).

89 However, I believe that there are more promising responses so I do not discuss induction in
chapter 6.

90 The foundationalist could respond by arguing that basic propositions are more justified than
propositions deduced from basic propositions. I consider this response in the next paragraph.

91 See for example, W. V. O. Quine (2001).

92 Since our inquiry involves internalist theories, this will be an internalist reliabilism (if this is
possible). That is, reasons are required to justify the claim that basic propositions are reliable.

93 This is the case if the foundationalist holds that it is necessary that, for a proposition to be
justified, it must be basic or inferred from a basic proposition.

94 That is, an impure infinitist can hold that basic propositions do not end regresses of reasons—
they are only necessary to avoid the MPR (Aikin’s view). An impure infinitist could also hold
that basic propositions end regresses but that (for whatever reason) meta-justifications (different
regresses) are endless.

95 Some philosophers such as W. V. O. Quine (1960) and Wilfred Sellars (1963, pp. 321-358)
add the requirement of simplicity. However, as Lehrer (2000a) argues, any addition to
coherentism muddles the account and makes adjudication more difficult: simplicity must be
balanced against coherency (pp. 118-119). Moreover, even if the coherentist purportedly
rationally adjudicates between two contradictory propositions (e.g., one proposition belongs to a
system of beliefs that is more virtuous or simple than the other), the MPRn* nevertheless applies
to coherentism because the MPRn* supervenes on the MPRc*. That is, since it is conceptually
possible that all propositions are propositionally justified (are members of a coherent and
virtuous or simple system of propositions) for all persons at all times using coherentism,
rationally adjudicating between beliefs is illusory: because the MPRc* applies to coherentism, a
skeptic can, in principle, show how any belief can be justified using coherentism.

96 Lehrer’s notion of “self-trust” might help coherentism avoid the MPR*. However, recall (§4.2
footnote 61) that the notion of self-trust was rejected because it violates PAC*.

97 For example, see Lehrer (2000a, pp. 116-19) and Bonjour (2003a, pp. 53-54).

98 As stated several times, the addition of a necessary condition of justification may help
foundationalism and coherentism avoid the MPR*, but since the infinitist condition is only a
necessary condition of justification, applying the MPR* to foundationalism and coherentism—if
it is held that they are only necessary conditions of justification—is equivalent to applying the
MPR to infinitism. That is, if justification is simply the infinitist, foundationalist, or coherentist
condition, then the condition can be reduced to absurdity. However, I believe that with the
addition of a necessary condition, each theory can avoid the MPR and MPR* (I argue this point
in the subsequent chapter).

99 See §3.1.

100 It should be noted that these facts are not part of the chain of reasoning (Klein, 1999, p. 324
endnote 51).

101 Recall the example where Φ is “propositions produced by tea leaf reading” and is sufficient
for justification. In this case, basic propositions and propositions inferred from basic propositions
are unnecessary for justifying propositions. The account can also be circular and hold that Φ is
sufficient for justification if p’s having Φ is “basic” or is “being inferred from a basic
proposition.” Recall that not only is this option circular, but it does not solve the MPR*: we saw
in §5.2 that the MPR* applies to the foundationalist structure.

102 Theories could hold that truth-conduciveness arises when and only when a proposition meets
the sufficient conditions of justification. However, this is a stronger claim, so I will argue for the
weaker claim that if a proposition meets the sufficient conditions of justification, then it is truth-
conducive.
103 I will not provide conditions for Φ. I am satisfied with Sosa and Klein’s criteria that actual
justificatory regresses contain only objectively and subjectively available reasons. However, it is
my goal to show that the MPR and the Cling-type skeptical worry are not problems, which does
not require me to specify conditions for Φ.

104 Cling assumes that a proposition p must have Φ. Cling is correct that p must meet necessary
condition Φ—the condition that distinguishes between justifiable and unjustifiable regresses—
but he is wrong to suppose that Φ must be a property of a proposition. The infinitist could hold
that Φ is a property of a proposition or that Φ is a property of chains of propositions. However,
since Klein and Sosa hold that propositions have properties that make them available as reasons
for other propositions, I will continue to hold that Φ is a property of propositions. It should be
noted that an interesting consequence of holding that Φ is a property of propositions is that
infinitism is what Sosa (1980) terms “formal foundationalism.” Formal foundationalism is the
view that propositions are justified by, among other things, some property (this property could be
non-normative and nonepistemic, which would mean that normative epistemic properties
supervene on non-normative non-epistemic properties—what Sosa takes to be a desideratum of
an epistemic theory of justification). Since infinitism holds that propositions are justified in part
by their properties (Φ), it is a formal foundationalist theory. Sosa contrasts formal
foundationalism with various theories of “substantive foundationalism,” e.g., modest
foundationalism. However, according to both Klein and Sosa, since coherentism holds that a
proposition is justified if and only if it is a member of a coherent set of propositions with certain
other properties, it is also a form of formal foundationalism. Thus, the leading substantive
accounts (foundationalism, infinitism, and coherentism) are all forms of formal foundationalism,
which Klein claims makes the category of formal foundationalism interesting. However, Klein
(2003) claims that the category seems less interesting when we see that it fails to distinguish
between interesting differences between the substantive views (p. 724 footnote 13). Like Klein, I
leave it to the reader to decide whether the category is significant.

105 This is not to say that if a proposition is not justifiable, then it necessarily cannot be truth-
conducive. I am only claiming that, when a proposition satisfies only one merely necessary
condition of justification, there is no reason to think that it must be truth-conducive, since the
proposition is not justifiable (it does not satisfy the sufficient conditions of justification).

106 Klein (1999, 2003, 2005a) thinks PAA also applies to properties. However, I argue that
infinite regresses of justification only apply to propositions or propositional beliefs, i.e., reasons
—after all, it is the justification of our propositional beliefs that concerns us in epistemology
(propositional knowledge rather than, say, procedural knowledge) (see §1.1). Note that I would
therefore change “all x” in PAA to denote “all propositions”; PAA would still rule-out
foundationalism and holistic coherentism (as Klein intends) since every justifiable proposition
must have a reason available to support it, so if foundationalists and coherentists think that basic
propositions and a belief’s coherence with one’s system of beliefs do not have reasons available
to support them, then foundationalism and coherentism must reject PAA. (And since the truth-
conducive property of foundational beliefs and coherence fails to avoid the MPR*,
foundationalists and coherentists still need Φ.)
107 The problem of the criterion is commonly set up by considering two questions: (1) “What is
the extent of our knowledge?” and (2) “What is the criterion of knowing?” It seems that we need
an answer to one question before we can answer the other, e.g., we might provide a criterion for
knowledge and use this criterion to determine what we know (separate true from false beliefs),
but to know that it is a good criterion for knowledge (leads to many cases of knowledge), we
must know whether it successfully determines what we know (separates true from false beliefs).
Likewise, a criterion of justification can be used to determine what beliefs are justified; however,
to determine whether the criterion is good (separates justified and unjustified beliefs), it seems
that we must determine what beliefs are justified and unjustified—begging the question seems
inevitable for all internalist theories of justification (externalist theories run into different
problems). See for example, chapter 5 of Chisholm (1982) and Stewart Cohen (2002).

108 Skepticism, at certain points of the inquiry, seemed to be the only option (e.g., when the
MPR/MPR* and Cling’s argument were combined). However, skepticism is not preferable to
infinitism if the evidence supporting skepticism is outweighed by the support for infinitism,
which is what I have tried to show in several places (e.g., it was argued that the Cling-type
skeptical worry should not convince us of skepticism). Skepticism may turn out to be right, but
our inquiry has provided reasons against accepting it and reasons in favor of accepting infinitism.

109 For example, Klein provides a possible solution in (2003, pp. 728-29).

110 Aikin rejects ought-implies-can but admits that it is only one way to respond to ought-
implies-can arguments against infinitism (§3.4). (The rejection of ought-implies-can is not a
condition of justification, but it is an argument supporting necessary conditions of justification,
so I added it to the list of conditions.)

111 The infinitist must hold that there is an infinite path or that complete justification is
providing infinite reasons, but it is unclear whether the infinitist needs an account of
propositional justification.

112 See §3.1 for example. Turri (2009a) stresses the point that one must properly produce the
relevant reasons to be justified (especially pp. 210-215). He argues that simply producing
reasons is not enough; S must “properly produce the relevant reasons” (Turri, 2009a, p. 215).

113 Cf. G. A. Cohen (2008, p. 43).

And So On.
Two Theories of Regress Arguments in Philosophy Abstract
Jan Willem Wieland

This PhD dissertation is on infinite regress arguments in philosophy. Its main goals are to explain
what such arguments from many distinct philosophical debates have in common, and to provide
guidelines for using and evaluating them. Two theories are reviewed: the Paradox Theory and
the Failure Theory. According to the Paradox Theory, infinite regress arguments can be used to
refute an existentially or universally quantified statement (e.g. to refute the statement that at least
one discussion is settled, or the statement that discussions are settled only if there is an agreed-
upon criterion to settle them). According to the Failure Theory, infinite regress arguments can be
used to demonstrate that a certain solution fails to solve an existentially or universally quantified
problem (e.g. to demonstrate that a certain solution fails to settle all discussions, or that it fails to
settle even one discussion). In the literature, the Paradox Theory is fairly well developed, and
this dissertation provides the Failure Theory with the same tools.

Contents
Foreword
1. Introduction 1
1.1. Key terms and hypotheses
1.1.1. Argument schema, instance
1.1.2. Reconstruction, evaluation
1.1.3. Regress, regress argument
1.2. Everyday regresses
1.3. The use of schemas
1.3.1. Metaphilosophical arguments
1.3.2. Methodological arguments
1.3.3. Obstacles
1.4. Overview
2. Regress Argument Schemas 27
2.1. Desiderata
2.2. The Paradox Schema
2.2.1. Schema
2.2.2. Validity
2.2.3. Boundaries
2.2.4. Literature
2.3. The Failure Schemas
2.3.1. Schemas
2.3.2. Validity
2.3.3. Boundaries
2.3.4. Literature
3. Instances 67
3.1. Argument reconstruction
3.2. Paradox filling instructions
3.3. Failure filling instructions
3.4. Selected full instances
3.4.1. Sextus Empiricus
3.4.2. Lewis Carroll
3.4.3. Russell
3.4.4. Wittgenstein
3.4.5. Ryle
4. Analysis 91
4.1. Comparison schemas I: Structure
4.1.1. Similarities
4.1.2. Generating the regress
4.1.3. From regress to conclusion
4.1.4. Scepticism
4.1.5. Dialectics
4.2. Regress
4.2.1. Series vs. regresses
4.2.2. Viciousness
4.2.3. Infinity
4.2.4. Circularity
5. Applications 129
5.1. Fallacies
5.2. Regressive Pragmatism
5.3. Carroll’s Tortoise
5.4. Epistemic Infinitism
5.5. Access principle
5.6. Russell’s relations
5.7. Quine/Davidson controversy
6. Meta-Debate 157
6.1. The literature
6.1.1. Three camps
6.1.2. Passmore’s Failure-hypotheses
6.1.3. Gratton’s Failure-hypotheses
6.2. Comparison schemas II: Soundness
6.2.1. Schema choice
6.2.2. In favour of the Failure Schema
6.2.3. Revisionism
7. Epilogue: Two Theories 191
7.1. Taking stock
7.2. The Paradox Theory
7.3. The Failure Theory
7.4. Epilogue
Summary in Dutch 199
References 209

Foreword

The Tortoise was saying, “Have you got that last step written down? Unless I’ve lost count, that
makes a thousand and one. There are several millions more to come.” (Carroll 1895)

[…] and so on ad infinitum, so that, since we have no place from which to begin to establish
anything, suspension of judgement follows. (Sextus Empiricus, Outlines of Pyrrhonism)

If you say ‘and so on’, you yourself do not know more than ‘and so on’. (Wittgenstein 1956)

1, 2, 3, 4,

and so on to infinity.

If there is an infinite regress, people usually write this phrase, or some variant of it, because they
do not want to go to infinity themselves. For myself, I prefer simply ‘and so on’, accompanied by
a bit of handwaving. The phrase suggests that the idea is clear, that from that point onward the
regress is obvious. The reader knows how to go on.

But do we really know how to go on in each and every case? No doubt most of us, except
perhaps Wittgenstein’s pupil (see his 1953: §185), do know how to go on in cases involving
natural numbers. For many regresses, though, I am not so sure. In this dissertation on regress
arguments, I collect and spell out a variety of cases deriving from all corners of philosophy: from
concerns about morality to problems about rationality, from Sextus to Russell to contemporary
philosophers.

Still, even if one does know how to go on, it is not clear what follows. If there is an infinite
regress, then what? What is to be concluded from it? As one philosopher put it after having
described such a regress:
How does “the story go on from here”? (Johnson 1978: 68)

As another philosopher ‘solved’ this issue:

Here the narrator, having pressing business at the Bank, was obliged to leave the happy pair.
(Carroll 1895: 280)

In other words, we want to know not only what regresses are, but also what kinds of conclusions
can be drawn from them. This is the most important query in this dissertation, and I explore two
different answers to it. Specifically, I develop two theories of regress arguments, which will be
called the Paradox Theory and the Failure Theory.

I readily admit that this dissertation is no especially academic production embedded in cutting-
edge, quickly evolving discussions in our field. Also, it is not my own deepest production; not a
resolute attempt to open up our way of looking at the world and ourselves. I have tried to do such
things elsewhere, modestly, in my side project on Pyrrhonian scepticism, i.e. the view that
propagates global suspension of belief.

This dissertation is of a different nature. First, it is intended as a real, scientific project with a
clear and, in some sense, timeless objective: to find out how regress arguments are supposed to
work. This just had to be sorted out. I am of course not the first to address this topic, but my
predecessors are few in number and have left unresolved some rather delicate issues (especially
regarding the Failure Theory, which is my original contribution to the debate).

Second, the results of this study ought to be of interest to anyone who employs regresses in one’s
reasoning; virtually, then, to all philosophers.

Third, it aims to contribute to the progress of philosophy as a discipline. Like any other inquirer
into philosophical methodology (i.e. the branch of philosophy which studies how philosophers
can and should proceed), I hope that my investigation will make a difference to the philosopher’s
practice. Particularly, I hope that from now on disputes about any particular regress argument
(concerning what it establishes, or concerning whether it can be resisted) will be more clearly
motivated, and indeed more clearly framed, in terms of the Paradox or Failure Theories outlined
in what follows.

Several acknowledgements are in order.

I am grateful to: Tim De Mey for sending me to Ghent. To Erik Weber and Maarten Van Dyck
for taking me on board, providing a nice niche for the kind of project that I wanted to pursue, and
advice during the past years. Special thanks to Erik for showing how to do things hypothetically
and for urging me to find more applications, and to Maarten for stimulating my interests in
epistemology through our first-year course.

To Arianna Betti and Anna-Sofia Maurin for advice from a distance and attracting my attention
to the debate on regress arguments in the first place.
To the Research Foundation Flanders (FWO) for funding my research project. To Jon Sozek for
the language edit. To Sanne Rous for the cover painting. To Gitte Callaert for help with the
design. To Mathieu Beirlaen, Frederik Van De Putte and Christian Straßer for checking parts of
the formalisations. To Philipp Blum for reminding me of the Sisyphus case which closes this
dissertation.

To audiences and colleagues at many conferences and seminars for discussion and suggestions.
To the referees of my articles.

To Eline, Annemart, Frijk and my parents for plenty of unrelated things.

And finally to Prof. Dr. Freddy Mortier, the chair of the examination committee, and to Prof. Dr.
Arianna Betti, Dr. Tim De Mey, Prof. Dr. Joke Meheus, Dr. Bert Leuridan, and Prof. Dr.
Benjamin Schnieder for taking part in my reading committee, and to Bert and Benjamin for
helpful feedback in the final stage (Benjamin’s insistence on the formal details of my story led to
substantial improvements of §2).

§1 Introduction

1.1. Key Terms and Hypotheses

In the following I identify six key terms of this dissertation: argument schema, instance,
reconstruction, evaluation, regress, and regress argument. I do not pretend to say anything new or
controversial in this part. My aim is only to explicate what will be meant by the terms. At the end
I set out the main hypotheses concerning the last two terms: regress and regress argument, i.e.
the terms that will bring in controversy.

1. Argument schema, instance

In this dissertation I will be concerned with argument schemas for regress arguments. Particular
regress arguments are instances of such schemas. Hence the first set of key terms: argument
schema; instance. Broadly taken, argument schemas are general versions of similar arguments.
Consider the following simple schema:

Schema 1

(1) A.
(2) If A, then B.
(3) So, B. [1, 2; Modus Ponens]1

This well-known argument schema (or schema, in brief) is a general version of all Modus Ponens
arguments. The schema contains letters such that you obtain specific arguments (i.e. instances) if
you fill them out. In this case, the letters are to be replaced with sentences. For example, if we
replace ‘A’ with ‘Socrates is a philosopher’ and ‘B’ with ‘Socrates is corrupting the youth’, we
obtain the following instance:
(1) Socrates is a philosopher.
(2) If Socrates is a philosopher, then he is corrupting the youth.
(3) So, Socrates is corrupting the youth. [1, 2; Modus Ponens]

Generally, a collection of lines is an argument schema only if, first, the lines contain letters
which are to be replaced with sentences (and may be connected by logical constants such as ‘if’
or ‘only if’), and, second, each line is either a premise (i.e. taken to be true), a hypothesis (i.e.
taken into consideration), or an inference (i.e. inferred from previous lines). The distinction
between premises and hypotheses will prove to be important in this dissertation. Regress
arguments are typical pieces of hypothetical reasoning. That is, they are arguments where some
claims are considered, yet not taken to be true, for the sake of deriving absurd or other kinds of
consequences from them.2

Instances of such schemas can be defined simply as follows: For any collection of lines x and y,
x is an instance of y iff x can be obtained by systematically replacing all letters of y with
sentences.3

Importantly, in this dissertation I will work with argument schemas stated in predicate logic
rather than propositional logic. Simply put, whereas propositional logic operates with sentences,
predicate logic operates with quantification and predicates. To see the difference, consider the
following schema:

Schema 2

(1) Socrates is F.
(2*) For all x, if x is F, then x is G.
(2) If Socrates is F, then he is G. [2*; Universal Instantiation]
(3) So, Socrates is G. [1, 2; Modus Ponens]

The similarity between schemas 1 and 2 is that both can have the Socrates example as an
instance (leaving premise (2*) aside for a second). This was shown already for schema 1, and the
Socrates example can be obtained also via schema 2 by systematically replacing ‘F’ with ‘a
philosopher’ and ‘G’ with ‘corrupting the youth’.

Nevertheless, the schemas differ in that schema 2 uses an extra line (2*) which quantifies over
all items, rather than some of them. This means that the statement applies to all items of a given
domain. In this case it is said that whenever you find an item in the domain which is F, then that
item is G as well (for example, whenever you find someone who is a philosopher, then you know
she is corrupting the youth). In this dissertation, I will use quantification, as it provides a more
precise tool to analyse arguments. Consider for example the following trilemma:

(A) Our beliefs are unsupported by further beliefs.

(B) Our beliefs are supported by an infinite chain of beliefs.
(C) Our beliefs are supported by a circular chain of beliefs.
This is also known as Agrippa’s Trilemma (e.g. Pritchard 2006: 33), and it purports to show that
none of our beliefs is supported or justified. Still, socalled Foundationalists are happy to resist
the reasoning by pointing to a fourth horn, namely that some of our beliefs are supported by
further beliefs, and some others are not supported by further beliefs. Of course, it is controversial
whether this last horn is as attractive as Foundationalists would have it be, but the important
point here is that this option is expressible only if we are explicit on the relevant quantification.

It is worth noting that in schema 2 predicates are to express properties or relations rather than
actions, though they may express actions as well:

Schema 2*

(1) Socrates φ-s.

(2*) For all x, if x φ-s, then x ψ-s.
(2) If Socrates φ-s, then Socrates ψ-s. [2*; Universal Instantiation]
(3) So, Socrates ψ-s. [1, 2; Modus Ponens]

Here, ‘φ’ may be replaced with ‘philosophize’ and ‘ψ’ with ‘corrupt the youth’. I employ this
distinction between properties/relations and actions later on, so throughout this dissertation
capitals such as ‘F’ and ‘R’ are used when the predicates are to express properties/relations, and
Greek letters such as ‘φ’ and ‘ψ’ are used when the predicates are to express actions.4

Given this, the notions of schema and instance, i.e. that I will employ in this dissertation, need to
be adjusted on two points. First, a collection of lines is an argument schema only if the lines are
suitably quantified (when applicable) and contain letters which are to be replaced with predicates
(which in turn express either properties/relations or actions). Second, a collection of lines x is an
instance of schema y iff x can be obtained by systematically replacing all letters of y with
predicates (rather than sentences).

2. Reconstruction, evaluation

The second set of key terms of this dissertation is the following: reconstruction; evaluation.
Generally, arguments can be reconstructed, and then be evaluated. Next I consider each step in
turn.

2.1. Reconstruction

Everyone will likely agree that argument reconstruction is to be guided by the following rules:

Interpretation Rule I: One should try to capture the original statement of the argument.

Interpretation Rule II: One should try to capture the context in which the argument appears (the
rest of the text, the background literature, the author’s intentions, etc.).

If you do not respect these rules, then you are simply not reconstructing anything, only
constructing your own argument. However, even if they are necessary, it is uncontroversial to
say that these Interpretation Rules alone do not always suffice. On the basis of these rules you
obtain at best paraphrased texts, that is, texts in which certain terms and constructions replace
certain other terms and constructions (where the former are possibly simpler or easier to
understand). However, if the original statement of the argument is rather implicit, i.e. if many
premises and inferences are suppressed, then more is needed. This is exactly the case for regress
arguments. As Gratton observes:

The typical presentation of infinite regress arguments throughout history is so succinct and has
so many gaps it is often unclear how an infinite regress is derived, or why an infinite regress is
logically problematic. (2010: xi)

To make such gaps explicit, at least two extra rules are needed:5

Charity Rule I: If needed, one should enforce the argument (modify or supply premises and
inferences) such that it becomes logically valid.

Charity Rule II: If possible, one should weaken the premises (make them less controversial and
true) such that the argument becomes sound.

The terms ‘logical validity’ and ‘soundness’ are used in the following way. An argument is
logically valid iff the conclusion follows logically from the premises/hypotheses, that is, by rules
of inference that are valid according to a certain logic. Furthermore, an argument is sound iff it is
logically valid and the premises are true as well. In terms of truth: if an argument is logically
valid, then the conclusion is true if the premises are true. If an argument is sound, then its
conclusion is true full-stop.

Without these rules in place, it would be possible to say of almost any regress argument in the
literature that it is logically invalid (and thus unsound) as, thanks to their gaps, their conclusions
do not follow logically from their premises. Suppose someone were to make the following
simple argument:

(1) Socrates is challenging people’s opinions.

(2) So, Socrates is corrupting the youth. [1]

This argument can be regarded as logically invalid (and so as unsound) insofar as there is no
familiar inference rule which takes you from (1) to (2). Still, such an evaluation is rather
uninteresting. The point is not whether the argument is logically valid as actually stated, but
whether there is a way in which it can be logically valid. In this simple case, the argument can
easily be made logically valid by adding an extra premise ‘If Socrates is challenging people’s
opinions, then he is corrupting the youth’ and applying Modus Ponens. This modification is
motivated by Charity Rule I.

A similar case can be constructed for the second rule. Suppose someone were to make the
following argument:

(1) Socrates is challenging people’s opinions and God exists.

(2) If Socrates is challenging people’s opinions, then he is corrupting the youth.
(3) So, Socrates is corrupting the youth. [1-2]

The argument is logically valid by the rules Simplification such that ‘Socrates is challenging
people’s opinions’ can be obtained from (1), and Modus Ponens can be applied as in the previous
case. Still, you may reject the argument as unsound simply because you are an atheist and do not
buy premise (1). Again, this is an uninteresting move, because whether or not God exists is
irrelevant for the conclusion (3). That is, the same conclusion can easily be obtained from a
much less controversial premise, i.e. ‘Socrates is challenging people’s opinions’. Making the
argument sound by weakening (1) is what is motivated by Charity Rule II.

I will have more to say about such charity moves later on. Still, I want to be explicit about the
assumed aim of argument reconstruction from the very start. To reconstruct an argument, I will
say, is not just to interpret accurately and get people’s intentions right. It is also, and primarily, to
do justice to the subject matter at hand, and to see what an argument can establish rather than
merely what it does or had to establish.6 Or again: argument reconstruction is to be used to
further our inquiries, and not just to learn something about the arguers/inquirers. The role of the
Charity Rules just listed is best seen from this perspective.

For related reasons, it is worth pointing out that the Charity Rules are not uncontroversial in the
literature. One of the worries is they may distort the initial statement of the argument. So there
may be limits to Charity. I will address this in §6.2.3. Also, note that all four reconstruction rules
are stated as plain commands. They can alternatively be formulated as decision principles. In the
charity case we would have: ‘For any set of reconstructions of a single argument, choose the
reconstruction which is the strongest and yet the least controversial.’ I have disregarded this
option because, as I will show later, it need not always be the case that one of the reconstructions
is to be selected (rather than another). Sometimes several reconstructions are valuable.

At any rate, in this dissertation I hope to show that the application of the Charity Rules, however
one formulates them, is in case of regress arguments a rather delicate enterprise.

2.2. Evaluation

Once the argument is reconstructed, evaluation is a straightforward matter. The questions to be

asked are, simply:

• Are the premises true?

• Does the conclusion follow from the premises/hypotheses?

Again, if the answer to the second question is affirmative, then the argument is logically valid.
And if the answer to the first is affirmative as well, then the argument is sound and hence the
conclusion of the argument true. So evaluation concerns two things: the premises and the
inferences. Still, evaluation is not as easy as it may seem at this point. There are two
complications.
First, both questions are more complicated than they appear to be. To ask whether the premises
are true is usually not just to ask whether they are true in general, but whether they are true in a
certain dialectical context. Such a context is the broader situation where the premises are part of
an argument devised for or against a certain position of interest. Hence, to ask whether the
premises are true is to ask whether the relevant parties in a given debate would (or should)
subscribe to them. Next to this, to ask whether the conclusion follows logically is usually not just
to ask about the application of a single inference rule, but about a number of different inference
steps. Such, at least, is the case for regress arguments.

Second, evaluation is already involved in reconstruction, for you cannot apply the Charity Rules
(and so know which reconstructions are logically valid and sound) before evaluating the
candidate hypotheses/premises and inferences (i.e. those which could make the reconstructed
argument logically valid and sound). Hence, even if the evaluation of the final reconstruction is
fairly straightforward, the candidate reconstructions complicate the matter.

So much for the second set of key terms.

3. Regress, regress argument

The third and main set of key terms of this dissertation is the following: regress; regress
argument. However one conceives the details of these terms, the two will always be set apart:
regresses are not regress arguments, and regress arguments are not regresses. This is, for
example, assumed here:

An infinite regress in itself neither proves nor disproves anything; an infinite regress argument
does. (Maurin 2007: 1-2)

In other words: regresses are not yet regress arguments and regress arguments are not merely
regresses. Very generally, regress arguments are arguments where a conclusion is drawn from a
regress. Regresses, therefore, are not yet regress arguments, as they are still to be associated with
a conclusion, and regress arguments are not merely regresses, as they also consist of a
conclusion.

In this dissertation I often speak of ‘regresses’ rather than ‘infinite regresses’. The reason is that I
do not want to assume that all regresses need be infinite. For the same reason I speak of
‘regresses’ rather than ‘vicious regresses’ (I address these issues in §4.2). What are good
definitions of regresses and regress arguments? Consider these dictionary definitions of ‘vicious
regress’ for a start:

Since the existence of this regress is inconsistent with an obvious truth, we may conclude that the
regress is vicious and consequently that the principle that generates it is false. (Tolhurst,
Cambridge Dictionary of Philosophy)

A strategy gives rise to a vicious regress if whatever problem it was designed to solve remains as
much in need of the same treatment after its use as before. (Blackburn, Oxford Dictionary of
Philosophy)
These are clearly quite different descriptions, suggesting perhaps that there is no single answer to
the question what regresses and regress arguments are. What regresses and regress arguments are
exactly depends, indeed, on one’s theory of them, and the whole upcoming dissertation will
concern this query. Still, I will provide below two different sets of hypotheses on regresses and
regress arguments. So far there should be nothing new in this dissertation, or for that matter
controversial. From now on this will change. Here is the first set of hypotheses:

Para–A Regresses are series of conditions which result from a number of claims and yield
something absurd.

Para–B Regress arguments are arguments which demonstrate that a number of claims cannot
hold together because they jointly yield an absurd regress.

The second set:

Fail–A Regresses are series of problems which result from certain solutions and which prevent
the success of these solutions.

Fail–B Regress arguments are arguments which demonstrate that a certain solution never solves
a given problem because it gets stuck in a regress.

These hypotheses are based on intuitions concerning what regresses and regress arguments are
about. The Para-hypotheses are based on, what may be called the Paradox Intuition, and the Fail-
hypotheses on the Failure Intuition. The Paradox Intuition in a nutshell: regresses cause
paradoxes. The Failure Intuition: regresses cause failures. I will point out in due course where
versions of both intuitions can be found in the literature. Generally, I take them to be well-
established among those familiar with regress arguments. Not everyone may share both
intuitions, but at least one of them seems present when people reason on the basis of a regress.

To be sure: I do not regard such hypotheses based on intuitions as unimportant. On the contrary,
they are what make regresses and regress arguments interesting objects of investigation. They are
interesting precisely because they are vague guesses and trigger further investigation. The
question is, indeed: Can the Para- and Fail-hypotheses be made precise, and be cashed out in
full-fletched theories?

Hence this dissertation.

1.2. Everyday Regresses

In the following I present five everyday regress arguments, both to give the reader unfamiliar
with regress arguments some idea of how such arguments work, and to illustrate the
philosophical cases which will be my main concern in this dissertation. Each story will be told
twice: the first time in a way reliant on the Paradox Intuition, the second on the Failure Intuition.

1. Thirst
Paradox

Suppose you are thirsty, that you drink a beer whenever you are thirsty, and that you become
thirsty again whenever you drink a beer. This generates a regress which is absurd. Hence, either
it is not the case that you drink a beer whenever you are thirsty. Or it is not the case that you
become thirsty whenever you drink a beer.

Failure

Suppose you want to quench your thirst. As a solution, you drink a beer. Yet, as happens with
beers, this generates a new thirst. So a similar problem occurs: you are thirsty and want to
quench your thirst. As a solution, you drink more beers. Regress. Hence, drinking beer is a bad
way to quench your thirst.

2. Carpet

He is like a man who presses down the bulge in a carpet only to have it reappear elsewhere.
(Armstrong 1978: 21)

Paradox

Suppose there is a bulge in the carpet, that whenever you find such a bulge you press it down,
and that whenever you press it down another bulge appears elsewhere in the carpet. This yields a
regress which is absurd. Hence, either it is not the case that whenever there is a bulge in the
carpet you press it down. Or it is not the case that whenever you press down a bulge in a carpet
another bulge appears elsewhere in the carpet.

Failure

Suppose you want to get rid of the bulge in the carpet. As a solution, you press it down. Yet, as
happens with bulges in carpets, it reappears elsewhere in the carpet. So a similar problem occurs:
there is a bulge in the carpet, and you want to get rid of it. As a solution, you press down more
bulges. Regress. Hence, pressing down bulges in a carpet is a bad way to get rid of them.

3. Guardians

But who will guard the guardians? (Juvenal, Satire 6)7

Paradox

Suppose your girlfriend is unreliable, that all unreliable persons are guarded by a guardian, and
that all guardians are unreliable. This yields a regress which is absurd. Hence, either it is not the
case that all unreliable persons are guarded by a guardian. Or it is not the case that all guardians
are unreliable.
Failure

Suppose you want to have your girlfriend guarded so that she can no longer commit unfaithful
acts. As a solution, you hire a guardian. Yet, as happens with guardians, he cannot be trusted
either. So a similar problem occurs: you want to have the guardian guarded. As a solution, you
hire another guardian. Regress. Hence, hiring guardians is a bad way to have your girlfriend
guarded.

4. Autobiography

Tristram Shandy, as we know, took two years writing the history of the first two days of his life,
and lamented that, at this rate, material would accumulate faster than he could deal with it, so
that he could never come to an end. (Russell 1903: §340)

Paradox

Suppose you write an autobiography of all days of your life, and that it takes you one year to
report one day of your life. This yields a regress which is absurd. Hence, either it is not the case
that you write an autobiography of all days of your life. Or it is not the case that it takes you one
year to report a day of your life.

Failure

Suppose you want a report of your life. As a solution, you write it yourself and start with today.
Yet you are so slow that it takes you one year to write a report of that single day. So a similar
problem occurs: there are 365 new days to be reported. As a solution, you start reporting the new
year, yet again you are so slow that it takes you 365 years to complete this task. Regress. Hence,
reporting all the days of your life yourself, if you are so slow, is bad way to obtain a report of
your life.

5. Revenge

Paradox

Suppose your neighbour killed a cow of yours, that whenever your neighbour kills a cow of
yours, you kill a cow of hers, and that whenever you kill a cow of your neighbour, she kills a
cow of yours. This yields a regress which is absurd. Hence, either it is not the case that whenever
your neighbour kills a cow of yours, you kill a cow of hers. Or it is not the case that whenever
you kill a cow of your neighbour, she kills a cow of yours.

Failure

Suppose you want to get rid of your anger. As a solution, you kill one of your neighbour’s cows.
Yet, as happens with neighbours, she gets angry and kills one of your cows. So a similar problem
occurs: your neighbour killed one of your cows and you want to get rid of your anger. As a
solution, you kill another of her cows. Regress. Hence, taking revenge is a bad way to get rid of
your anger.

1.3. The Use of Schemas

In this dissertation I present several argument schemas for regress arguments. Here, I will take
one step back in order to motivate my inquiry into these schemas, i.e. to offer some arguments as
to why the latter are desirable in the first place. I set forth two kinds of arguments:
metaphilosophical and methodological. It will be useful to start with the metaphilosophical
arguments, even though the methodological arguments are more important. After introducing
both kinds of arguments, I identify and reject some obstacles faced by the search for such
schemas.

1. Metaphilosophical arguments

Regress arguments, just like other philosophical tools such as thought experiments, intuition
pumps, analogies, contradictions, horned dilemmas, transcendental arguments, counterexamples
and paradoxes, show up in all branches of philosophy.8 Consider for instance the following
cases:9

“If the form proves to be like what partakes of it, a fresh form will never cease emerging.”
“That’s very true.”
“So other things do not get a share of the forms by likeness; we must seek some other means by
which they get a share.” (Plato, Parmenides 132a-b)

You can never get rid of the contradiction, for, by the act of removing it from what is to be
explained, you produce it over again in the explanation. And so the explanation is invalid.
(McTaggart 1908: 469)

But this leads at once to an endless regress. Thus the attempt to regard our proposition as
asserting identity of denotation breaks down, and it becomes imperative to find some other
analysis. (Russell 1910-11: 124-5)

And then we should be confronted by a question of the same kind and the game could begin
again. So the attempt to explain truth as correspondence collapses. (Frege 1918-19: 291)

Though all these authors use regress arguments, their concerns are clearly very different: Plato is
trying to explain the connection between us and the Forms, McTaggart is concerned about
paradoxes of time, Russell about the meaning of proper names, and Frege about truth. Many
more examples of regress arguments will be provided in due course. Given that all these cases go
under the same name, ‘regress argument’, it would be surprising if they (or at least a significant
portion of them) had nothing in common. As argument schemas are exactly those things which
arguments can have in common, to say that it is likely that regress arguments have something in
common is to say that it is likely that there is an argument schema, or set of schemas, of which
they are an instance.
It would be likewise surprising if, say, thought experiments (or a significant portion of them)
were to have no common features. Among those who believe that thought experiments are
arguments (or closely related to arguments), these common features have indeed been discussed
in terms of argument schemas. Here is for example Häggqvist’s proposal (2009: 63):

(1) A certain counterfactual scenario C is possible.

(2) Theory T predicts a result W in C.
(3) But W is false in C.
(C) Hence: T is false. [1-3]

A well-known instance of this schema from Putnam (1973) would run as follows (Häggqvist
2009: 68):

(1) It is possible that we have Twin Earth Doppelgängers.

(2) If psychology determines reference, then, if we had Twin Earth Doppelgängers, they would
refer to water with ‘water’.
(3) If we had Twin Earth Doppelgängers, they would not refer to water with ‘water’.
(C) Hence, psychology does not determine reference. [1-3]

Of course, Häggqvist’s proposal is not uncontroversial.10 The main point to be taken from this,
however, is just that the debate on regress arguments as I will present it here (namely as a debate
on argument schemas) may well run parallel to this debate on thought experiments. The first
argument for regress argument schemas is this. Such schemas are desirable because they can
serve as an answer to the question what regress arguments from a variety of philosophical
debates have in common. As Clark puts the query:

Where, we wonder, is the shared, common content to be found in applications as diverse as

these? (1988: 369)

This is a metaphilosophical argument because it unifies debates from epistemology, metaphysics,

philosophy of logic, philosophy of language, ethics, etc., debates that would otherwise be
disconnected from one another. This disconnection is unfortunate, however, if the argumentative
strategies (regress arguments in this case) appear to have to same structure. As we shall see, in
terms of argument schemas this means that what the arguments have in common is the form
taken by their hypotheses/premises and inferences, and what differs in them is the actual content
of the predicates which replace the schematic letters.

It must be admitted that, aside from a few everyday illustrations, I have limited myself to a
consideration of cases only in philosophy. This does not mean, of course, that there are no (or
cannot be) regress arguments in the sciences. Indeed, I see no reason why the schemas have to be
filled out exclusively by philosophical terms (such as truth, knowledge, inference, paradox,
freedom, morality, etc.). Still, I do suspect that regress arguments occur so often in philosophy
because philosophers usually cannot rely on empirical evidence and need other tools to support
and review their positions.11
The question ‘What do the various regress arguments in fact have in common?’ appears to be
purely descriptive, wanting to know just what is the case and not what should be the case. It
should be pointed out, however, that in what follows I will also be interested in its normative
counterpart, i.e. ‘What should the various regress arguments have in common?’ This stress on
the revision of arguments, and the role of the Charity Rules in this revision, will play an
important role later on (in §6.2).

There is a second metaphilosophical argument. The meta-literature on regress arguments, i.e.

literature which tries to say something about regress argument in general (rather than about a
single case only), is small but still divided (and at that interestingly so, as we shall see). Consider
for example the following claim made by one of the main participants in this debate:

The difference between my view and Passmore’s may be put like this. According to me, infinite
regress arguments conclude to the negation of a proposition. […] According to Passmore, they
prove not that a proposition is false, but that an explanation is inadequate. (Black 1996: 111)

According to Black, there is a dispute about the form taken by the conclusions of regress
arguments, specifically concerning whether they conclude to the negation of a proposition or to
the inadequacy of an explanation. Now, to disagree on this is just to disagree on the concluding
line of the regress argument schema. In this dissertation I will show that this disagreement (and
related ones) can be resolved if we relativize the positions of Black and Passmore to argument
schemas. That is, a regress argument can conclude, I will say, either to the negation of a
proposition or to the inadequacy of an explanation depending on which schema is used to
reconstruct the argument. (Note already that the issue of whether an argument should, rather than
can, be reconstructed in one way or in the other is separate from this.)

The second argument for regress argument schemas is this. Such schemas are desirable because
the debate on regress arguments can be clarified and sharpened on their basis. Again, this may be
called a metaphilosophical argument because it brings together different strands in philosophy
(although in a slightly different sense than in the previous argument). Also, the question ‘What
different theories of regress argument are available in the literature?’ appears purely descriptive,
though again I will not deny that at the same time I will be concerned with the normative
question ‘What different theories of regress argument should be available in the literature?’

In sum, from a metaphilosophical perspective regress arguments schemas are desirable because:

• They serve to answer the question of what regress arguments from a wide range of discussions
have in common.

• They can clarify or sharpen several disputes in the literature on regress arguments.

2. Methodological arguments

In this section I want to present three arguments for the use of regress argument schemas which
are somewhat different in nature. They are methodological rather than metaphilosophical
arguments, as I will explain later. The first argument is perhaps the strongest of all. To explain
this argument, I shall use again the four texts cited in the previous section. In each of these we
find:

• a hint at a regress (“a fresh form will never cease emerging”, “by the act of removing it from
what is to be explained, you produce it over again”, “this leads at once to an endless regress”,
“we should be confronted by a question of the same kind and the game could begin again”);

• an indication of a conclusion (“thus” and “so”, marked by the boxes);

• and the conclusion (“other things do not get a share of the forms by likeness”, “the explanation
is invalid”, “the attempt to regard our proposition as asserting identity of denotation breaks
down”, “the attempt to explain truth as correspondence collapses”).

Clearly, regresses are used to establish a variety of conclusions. In the texts just cited we have:
the failure of a version of Plato’s theory of forms, the failure of a certain theory of time, the
failure of a certain theory of meaning, and the failure of the correspondence theory of truth. All
four, it is worth noting, seem to express the Failure Intuition, and indeed I selected them in the
previous section to trigger the search for what those arguments have in common. (I could have
taken a number of cases based on the Paradox Intuition as well.)

In this section, however, I want to focus not on the commonality of the conclusions, but on the
fact that these conclusions are supposed to follow from a regress. All are substantive
philosophical claims, all claims, that is, which make a difference to some certain debate.
Regresses, in other words, are used to make such differences. Consider these general remarks:

Amongst the most powerful weapons in the philosopher’s armoury are reductio ad absurdum and
infinite regress arguments. (Waismann 1956: 26)

It seems natural to assume that Plato’s use of the third man argument […] is an early exemplar,
indeed the first on record, of one of the most effective tools of the philosopher’s trade, i.e. the
vicious regress. (Ranking 1969: 178)

These remarks notwithstanding, it remains far from clear how regresses can be used to establish
anything, be it substantial or not. In each of the four cases listed above (Plato, McTaggart,
Russell, Frege) the regress functions as a sort black box. One may wonder: Does the conclusion
indeed follow from the regress? If so, how?

My claim, now, is that regress argument schemas can clarify and illuminate these inferences.
Specifically, such schemas are useful because they serve to abstract away all information that is
irrelevant to the inference steps (namely, the specific content of a given argument), and to make
explicit all that is relevant. If this is right and if knowledge about the inferences is useful to
evaluate regress arguments, then argument schemas are useful to evaluate regress arguments.
This is the third argument for regress argument schemas.

Philosophy (or at least its logical part) is sometimes characterised as the science of What Follows
From What. It is not directly concerned with content (e.g. whether God exists or not), but with
form (e.g. whether it follows from such and such that God exists or not). This focus on form
underlies this third argument. The questions are: How does anything lead to a regress, and how
does anything follow from a regress?

The next argument is connected to this third one. It holds that regress argument schemas are
useful because some regress arguments in the literature have ambiguous conclusions, i.e. two
different conclusions are meant to follow from what appears to be a single regress, and these
cases can be disambiguated on the basis of an argument schema. Consider the following case
drawn from the ancient sceptics:

In order to decide the dispute that has arisen […], we have need of an agreed-upon criterion by
means of which we shall decide it; and in order to have an agreed-upon criterion it is necessary
first to have decided the dispute about the criterion. […] If we wish to decide about the criterion
by means of a criterion we force them into infinite regress. (Sextus, Outlines of Pyrrhonism,
2.18-20)

Suppose, for example, you want to decide a dispute about whether Juvenal had a wife. To do so,
you introduce another proposition, such as that Juvenal has been banished his whole life and so
could not have had a wife. Of course, this second proposition is disputable too. In order to decide
that new dispute, you introduce a third proposition, say that the sources about Juvenal’s
banishment are highly reliable. Of course, this third proposition is disputable too. Regress. Now
the question is what conclusion can be drawn from it. Does it follow from this regress that one
cannot decide all disputes? Or that one cannot decide even one dispute? The second conclusion
is clearly stronger than the first: if one does not decide all disputes, one might still decide many
of them. Argument schemas should help out here.

Here are three further interesting examples where the conclusion is ambiguous:

Case 1

Hence the attempted analysis of the relation fails, and we are forced to admit what the theory was
designed to avoid, a so-called ‘external’ relation, i.e. one implying no complexity in either of the
related terms. (Russell 1903: §214)

Hence we cannot, without an endless regress, refuse to admit that sooner or later we come to a
relation not reducible to adjectives of the related terms. (Russell 1906-07: 41-2)

Here we have two different conclusions for a regress concerning relations. In the first passage,
Russell suggests that the regress in question (to be spelled out later) demonstrates that it is not
the case that all relations imply corresponding properties of their relata. In the second, he
suggests that that very same regress demonstrates that never all relations can be reduced to the
corresponding properties of their relata. Surely implication is not the same as reduction.
Ambiguity.

Case 2
But now a higher order loophole opens. If I keep ignorant of whether there is an obligation to
ascertain my obligations, I can use the Access principle to evade those epistemic obligations
even if they exist. To close this meta-loophole, the defender of Access must invoke a yet higher
order principle to the effect that we have an obligation to learn whether we have an obligation to
learn our obligations. (Sorensen 1995: 255)

At best, the regress consists of an infinite sequence of cases, none of which refutes Access.
(Sider 1995: 279)

These claims are about the Access principle which states that one is obliged to do something
only if one can know that one is obliged to do that thing. According to Sorensen, Access is
problematic because you can abuse this principle to avoid your obligations, namely by making
sure that your obligations are not knowable to you. Sider responds that this need not be
problematic so long as we have the obligation to refrain from avoiding our obligations. As
Sorensen suggests in turn, however, this new obligation generates a regress (again, to be
explained later on). Finally, Sider agrees that there is a regress here, but denies that it puts any
pressure on Access (see the cited passage). Hence, we have two different assessments of a single
regress.

Case 3

When we ask, “Does ‘rabbit’ really refer to rabbits?” someone can counter with the question:
“Refer to rabbits in what sense of ‘rabbits’?” thus launching a regress; and we need the
background language to regress into. (Quine 1968: 200-1)

The lesson, in my opinion, is that we cannot make sense of truth, reference or ontology
relativized to a background theory or language. The trouble is that we must start a regress we
cannot finish. (Davidson 1979: 234)

Without going into details at this point, Davidson seems to disagree with Quine’s view that the
reference of words (such as ‘Gavagai’) is relative to a background language. Quine seems to
acknowledge that this relativity of reference thesis generates a regress (for the reference of the
background language is relative to yet a further background language, and so on), but not to
regard the regress as vicious. Davidson, by contrast, contends that the regress is vicious and that
it shows that the relativity of reference thesis does not make sense. These are again two very
different assessments.

In these three cases, what is the right conclusion? Or might both make sense? If the latter, then
what premises are to be associated with what conclusions? Such questions can be answered,
indeed, with the aid of argument schemas. Later in this dissertation, I will argue that in each case
above either conclusion makes sense, but only in the context of a different argument schema
(§5). This is the fourth argument in favour of such schemas.

The fifth and final argument for regress argument schemas is that by means of them one can
easily see what to do if you do not want to buy the conclusion of a regress argument. That is, the
schemas display clearly what premises might be attacked (such that the truth of the conclusion is
no longer established by them). Or, to put it in a different way: the schemas define what exactly
can be disagreed about when there is a dispute about regress arguments. For a full version of this
important argument, I again refer to further parts of this dissertation (§4.1.5).

All in all, in this section I have presented three more arguments for regress argument schemas,
this time methodological in nature. Such schemas are useful because:

• Regress arguments have substantive conclusions, and without such schemas we do not know
whether and how they follow.

• Some regress arguments have multiple (potential) conclusions, and these can be disambiguated
on the basis of such schemas.

• By such schemas it can be seen what premises may be attacked if you do not want to buy the
conclusion of a regress argument.

These three points make the search for argument schemas a methodological enterprise. This
means that it takes issue with how certain arguments can and should be used. Or again: regress
argument schemas provide us with a method or tool to evaluate and sharpen regress arguments.
This aspect is broadly normative, rather than descriptive, for it concerns not what is the case (i.e.
unintelligible inferences, ambiguous conclusions, and unclear debates), but rather what can and
should be the case (i.e. intelligible inferences, unambiguous conclusions, and clear debates).

At this point, a word on the literature is in order. How does this search for argument schemas
relate to previous undertakings? As far as I know, three other PhD dissertations have been
devoted to regress arguments in philosophy, i.e. by Black, Day and Gratton (references to their
published work will be provided in due course). Black’s approach comes closest to mine in that
he is explicit that his study on regress arguments is about argument schemas (this is less clear in
Day, and Gratton invokes diagrams rather than schemas). Still, there are three important
differences between these studies and mine.

First, I do not discuss one schema, but two (plus subversions), and consider all claims on how
regresses are to be used and evaluated relative to those two schemas. This pluralist view that I
will advocate will be contrasted with the positions by Black, Day, Gratton, and others.

Second, I include more case studies than do these earlier authors, drawn from all sorts of
philosophical debates, and do so in a way that is both relatively economical and does not
interrupt the main presentation of my results (see §3).

Last, none of the three studies is explicit on the normative aspect of their project in the sense just
explained (though to some degree they clearly are). Compare for example Gratton’s research
questions:

Two very general questions guided this work: (1) How are infinite regresses generated in infinite
regress arguments? (2) How do infinite regresses logically function as premises in an argument?
(2010: xi)
These are descriptive questions and clearly different from their normative counterparts (i.e. from
questions as to how regresses can and should be generated, and how in turn conclusions can and
should be generated from the regresses). Still, both aspects will concern me in this dissertation,
and my position in the end much depends on this normative emphasis.

3. Obstacles

Even if regress argument schemas are to be valued for all these reasons, the search for them is
not without obstacles. Below I will discuss two problems. The first problem is a circularity
problem and is the more pressing of the two.

Take the reconstruction rules again. The Interpretation Rules say: Try to capture the original
statement and context of the argument. The Charity Rules say: Try to revise the argument such
that it becomes logically valid and sound. Clearly, there is a tension between these rules. By the
Interpretation Rules, we should respect how the various regress arguments are actually presented
in the literature. By the Charity Rules, we should revise, rather than respect, these regress
arguments. Or in other words, by Interpretation we should be descriptive (ignoring what can and
should be the case, and capturing just what is the case), and by Charity we should be normative
(ignoring what is the case, and capturing just what can and should be the case). A similar tension
may be found here:

I have given examples from some philosophers whose language suggests that recurring questions
are somehow involved in generating an infinite regress, and have argued that such questions are
not involved. However, they can suggest some parts of a regress formula. (Gratton 2010: 157)

Specifically, Gratton’s contention here is that regresses should be stated not in terms of questions
and answers (even if the original texts suggest so), but directly in terms of regress formulas (to
be explained later on). Gratton’s basic motivation for this is his view that questions do not entail
infinite regresses. I shall not take up this specific issue here. Important here is that he chooses at
this point for Charity rather than Interpretation. I had to make similar choices. Let me note just
two instances where I part ways from Gratton’s choices.

First example. Plato devises a regress argument against a certain conception of ignorance:

[…] Or are you going to start all over again and tell me that there’s another set of pieces of
knowledge concerning pieces of knowledge and ignorance, which a man may possess shut up in
some ridiculous aviaries or waxen devices, which he knows so long as he possesses them though
he may not have them ready to hand in his soul – and in this way end up forced to come running
round to the same place over and over again and never get any further? (Theaetetus 200b-c)

Gratton claims that Plato’s regress argument does not work basically because the new problems
generated by the solution are insufficiently similar to the initial problem (2010: 161). According
to Gratton, the relevant problems and solutions are the following:

Problem: It seems contradictory that sometimes one does not know what one knows.
Solution: The contradiction disappears as soon as we introduce the distinction between active
and passive knowledge: one can fail to have active knowledge that one possesses passively.

New problem: It seems contradictory that sometimes one has active knowledge but is mistaken
about it. As I will explain in §4.2.1, it is correct that regress arguments do not work if the
problems are not similar. As in this case the problems are dissimilar, it is indeed hard to see what
might be a regress argument here. Nevertheless, another reading of the relevant problems is
possible:

Problem: It needs to be explained how one can be ignorant about a piece of knowledge.

Solution: The explanation is that one can be ignorant about what one possesses in one’s soul.

New problem: It needs to be explained how one can be ignorant about what one possesses in
one’s soul.

I do not want to claim that this reading is a better interpretation of Plato’s words (although that
would be nice). I do, however, hope my reading is a more charitable reading of the text, namely a
reading which presents a reconstruction that has a chance of being sound.

Second example. Gratton uses Ryle’s regress argument as the typical case for his regress
argument schema (2010: 2-4). It can be observed that Gratton’s reading ignores the context in
which Ryle presents his argument, i.e. the discussion of whether knowledge-how requires
knowledge-that. This is in line with his general approach described here:

I avoided as much as possible addressing the philosophical content and historical background of
the arguments examined. (2010: xi)

However, as I will show later, if the context is also taken into account, then Ryle’s argument may
well be constructed in a rather different way (which differs not merely qua content, but also qua
form).

At any rate, the contention here is that in many cases argument reconstruction is no
straightforward matter and involves choices between interpretation and charity on behalf of the
reconstructor. But how, then, is the tension to be resolved (i.e. how are choices between
interpretation and charity to be made), and non-arbitrarily at that? Before anticipating my
solution to this, I would like to point to a related circularity problem I encountered.

On the one hand, how is one to set up an argument schema if not by generalizing over actual
cases from the literature? Yet on the other, how is one to generalise over actual cases if not by
using an argument schema? To solve this problem (i.e. of generalizing over actual cases and also
setting up argument schemas), I had to do both at once, switching back and forth between
tentative statements of the argument schemas and the actual cases. This circularity problem, then,
proved manageable.
The circle between Interpretation and Charity, however, cannot be solved by such a reflective
equilibrium. If these principles are in conflict, both cannot have what they want (not even just
half of it). Either the text is to be respected, or the argument is to be respected. There are three
options:

• Defend that Interpretation always wins out.

• Defend that Charity always wins out.

• Defend that sometimes Interpretation wins out, and sometimes Charity.

The first two are radical solutions because they have to show why one aspect of argument
reconstruction is more important than the other in every case. The third is no easier to defend,
because in that case you might need a criterion which non-arbitrarily rules when Interpretation
wins out, and when Charity. I will eventually take up the second horn of this trilemma: whenever
there is a conflict between Charity and Interpretation, then, I will maintain, Charity always wins
out over Interpretation (at least in the case of reconstructing regress arguments).

This was the circularity problem, and at this point it is not yet solved: all I have done is anticipate
my solution. The other, second obstacle goes less deep, but is nevertheless interesting.

Sometimes it was suggested to me (at conferences and other occasions) that the project cannot
but be a failure. In particular, regress arguments would be too diverse, and so it was anticipated
that no general argument pattern would be forthcoming.12 For my research project, of course, I
had to suppose that this scepticism was misplaced. Even if it were to turn out that there are as
many schemas as instances (i.e. that there are no schemas with more than one instance), then the
(negative) result of the project would be that regress arguments are indeed too diverse.

Nonetheless, in this dissertation I will show that there are two main argument schemas for
regress arguments, and that each and every case from the literature can be reconstructed in terms
of both schemas. Accordingly I will present regress arguments as a fairly homogeneous class of
arguments.

Apart from this contingent outcome, there is one argument as to why the scepticism about
argument schemas is misplaced no matter what the outcome of my research. Namely: even if
regress arguments are extremely diverse, there is always the possibility of setting up a schema
which is so general that all fit it. Here is such a general schema (or rather proto-schema, as it
does not contain schematic letters, among other elements):

(1) Some premise/hypothesis.

(2) Regress. [1]
(C) Some conclusion. [2]

Of course, this schema is rather uninformative, as at least the following questions remain
unanswered: What kind(s) of premises/hypotheses lead to a regress? What kind of line is a
regress? What kind(s) of conclusions can be drawn from a regress? How do those conclusions
follow? The question, basically, is whether more specific schemas are available, that is, whether
there are schemas specific enough to be informative and non-trivial (i.e. such that not almost
anything can be an instance of it), but still general enough to capture many instances.

A final kind of scepticism which I had to judge misplaced is that even if I were to identify some
general feature of many regress arguments, it might still be the case that regress arguments
happen to be based on a logical mistake. Merely to note that philosophers from Plato to Hume to
Russell have used these arguments would be no help here, because logical mistakes can occur
anywhere. So let us see.

1.4. Overview

This dissertation is divided into seven main parts. In the following I summarise briefly the
contents of each.

Part Summary

1. Introduction In this part, I introduce regress arguments by several everyday examples, define
the key terms of this dissertation, and motivate its main goal: the search for regress argument
schemas. I also formulate two sets of hypotheses on regresses and regress arguments, which will
be crystallised into full theories in what follows.

2. Schemas In this part, I set out the two main argument schemas (plus subversions) which are
labelled as the Paradox Schema and the Failure Schema. The former schema derives from the
literature, and the latter is my original contribution. This constitutes the key part of the
dissertation as all further aspects of the two theories of regress arguments rely on these schemas.

3. Instances In this part, I present instances of the schemas from the previous part. First I make
my approach explicit, identify the filling instructions for both schemas, and then spell out a
selected number of classic instances in full.

4. Analysis In this part, I compare the two schemas on the basis of the case studies, and pin down
both what they have in common and what makes them different. Specifically, I compare their
premises/hypotheses, their conclusions and their dialectics. After that, I explain how each
schema pertains to the questions: What are regresses, when are they vicious, and do they have to
be infinite?

5. Applications In this part, I employ the schemas to clarify some important debates, among them
the debates on Carroll’s Tortoise and Epistemic Infinitism.

6. Meta-debate In this part, I first defend the Failure Schema against a number of concerns
anticipated in the literature. Next I answer the query: If regress arguments can be reconstructed
along the lines of both the Paradox and the Failure Schema, then which way is to be preferred?
My position called Revisionist Pluralism basically says this: If you want to refute a universally
or existentially quantified statement, then use the Paradox Schema. And if you want to show that
your opponent fails to solve a universally or existentially quantified problem, use the Failure
Schema.

7. Epilogue I conclude the dissertation with what the two theories (each based on one of the
schemas) have to say about how regress arguments are to be used and evaluated.

§2 Regress Argument Schemas

2.1. Desiderata

In the following I specify what kind of argument schemas I am looking for. Ten desiderata will
be listed. Many of these follow directly from §1.3, that is, the schemas must fulfil such and such
desiderata in order to be useful.

1. First set

The argument schemas have to fulfil a number of desiderata. To begin with, the schemas must be
such that:

• a regress occupies one of their lines;

• the regress is derived from premises/hypotheses;

• a conclusion is drawn from the regress.

These desiderata derive from Gratton (2010: xi), and can be motivated in two ways. First, they
can be used to demarcate regress arguments from nonregress arguments. This means that nothing
is a regress argument unless it fulfils these three desiderata (for example, if no regress occupies
one of the lines), and, vice versa, that anything which does fulfil them is a regress argument. In
terms of argument schemas: given that regress arguments are a specific kind of argument, the
schemas must fulfil these desiderata in order to have regress arguments as their instances.13

Second, the schemas must fulfil these desiderata in order to be useful from a methodological
point of view. That is, on the basis of such schemas we can check whether and, if so, how
substantive conclusions follow from regresses, disambiguate these conclusions, and see what
premises may be attacked if someone committed to (some of) them does not want to buy their
apparent conclusion (see the methodological arguments from §1.3).

Please note that at this stage there are no further restrictions on the schemas. In principle, then, it
is possible to set up an argument schema which is logically invalid, which has no example from
the literature as an instance, and which cannot explain any dispute about regress arguments in the
literature (e.g. identify the exact point of controversy between Quine and Davidson regarding a
regress noted in §1.3.2). However, a schema which is logically invalid is uninteresting from the
point of Charity Rule I, and a schema which has no actual instances and cannot explain (or help
to resolve) disputes about regress arguments (such as Quine/Davidson dispute) is useless from a
methodological point of view. Here, then, are three further desiderata. The schemas should:
• be logically valid;

• have several cases from the literature as an instance;

• be able to explain disputes about regress arguments.

Now we have six desiderata. They will be the desiderata that will form my concern in a larger
part of this dissertation: §§2-5. Specifically, I will discuss two main arguments schemas (plus
subversions) which fulfil all six. These schemas will be labelled the Paradox Schema and the
Failure Schema. As we shall see, each of the schemas can be coupled with one of the intuitions
about regress arguments identified in §1.1.3, that is, respectively, with the Paradox Intuition and
the Failure Intuition. To my knowledge, no other argument schemas which fulfil these six
desiderata are on the market.

2. Second set

The follow-up question is this: If several regress argument schemas fulfil the foregoing
desiderata, then which is the most fruitful? This is a new question as yet unaddressed in the
literature on regress arguments. But in fact it is just a version of the problem of argument
reconstruction: If there is more than one way in which an argument can be reconstructed, then
how is it to be done? (If there are different argument schemas available according to which an
argument can be made explicit, then which is the most fruitful?) The problem is especially
pressing if both schemas happen to be logically valid.

I will approach the issue by introducing some further desiderata for the argument schemas.
Particularly, the schema should:

• be as simple as possible;

• produce plausible premises/hypotheses;

• produce conclusions that are hard to resist;

• produce conclusions that can play an interesting role in a debate.

These desiderata are plausible, because a schema which fulfils them produces better arguments
than a schema which does not.14 And by Charity Rule II, we should look not only for logically
valid arguments (as Rule I motivates), but for sound ones, that is, for the best possible
arguments. Here, I assume without further argumentation that simple arguments are better than
complex ones, arguments with plausible premises/hypotheses better than those with implausible
ones, arguments which are hard to resist better than arguments which are not, and arguments
with interesting conclusions better than those with uninteresting ones. They basically work thus:

Schema choice. For any two argument schemas, if it is possible to reconstruct an argument on the
basis of either, then one should use the schema that is such and such (e.g. the simplest).
These last four desiderata will form my concern in §6.2 of this dissertation. To be sure: it need
not be the case that the desiderata converge. For example, the schema which produces the
arguments with the most plausible premises and most interesting conclusions may not be the
schema which produces the simplest arguments. As a consequence, it may turn out that the
Paradox Schema and the Failure Schema score well on different points, so that they are fruitful
in their own right.

I will nevertheless turn the tables in favour of one of the schemas. That is, although I will argue
that both schemas can have good, sound instances, I will also explain why regress arguments that
take the form of the Failure Schema are often better arguments, and hence that often the latter
schema is a more fruitful way of reconstructing regress arguments. (Such is my qualified
pluralist view.)

2.2. The Paradox Schema

In the following I present one of the regress argument schemas: the Paradox Schema. Next, I
address queries about its logical validity and its boundaries. Finally, I refer to the literature where
parts or versions of the Paradox Schema have been investigated.

1. Schema

Recall the following hypothesis from §1.1:

Para–B Regress arguments are arguments which demonstrate that a number of claims cannot
hold together because they jointly yield an absurd regress.

Gratton’s diagram (1997: 205, 2010: 4) can be taken as a specification of this hypothesis. Here is
my reading of his diagram:

Proto Paradox Schema

(1) Regress formula 1.

(2) Regress formula 2.
(3) Trigger.
(4) Regress. [1-2]
(5) Result. [3, 4]
(6) (5) is false.
(C) (1) is false. [1-6]

From now on the premise/hypothesis distinction will be important (see §1.1). In this proto-
schema, there is one hypothesis (for Reductio Ad Absurdum, as I will explain later): line (1).
There are three premises: lines (2), (3) and (6). And finally three main inferences: lines (4), (5)
and (C). We shall see about the ins-and-outs of the inferences soon, but the main idea is: given
that (1), (2) and (3) together lead to a contradiction with (6), we have to reject the hypothesis,
which is line (1). In this sense, this schema is a specification of Para–B above.
Still, the above is only a proto-schema, for something is a full argument schema only if its lines
are quantified statements with schematic letters. In the following, therefore, all lines are replaced
with such statements. From now on I shall refer to this as the Paradox Schema. Whenever I talk
about Paradox arguments in this dissertation, I talk about instances of this semi-first-order
schema. A full first-order rendering will be presented below.15

Paradox Schema

(1) For all items x of type K, x is F only if there is a new item y of type K and x and y stand in R.
(2) For all items x and y of type K, x and y stand in R only if y is F.
(3) There is at least one item of type K that is F.
(4) Regress:
(a) a is F.
(b) a and b stand in R, where a≠b. [a, 1]
(c) b is F. [b, 2]
(d) b and c stand in R, where a≠b≠c. [c, 1]
(e) c is F. [d, 2]
And so on.
(5) An infinity of items of type K are F. [3, 4]
(6) (5) is false: No infinity of items of type K are F.
(C) (1) is false: It is not the case that for all items x of type K, x is F only if x stands in R to a
new item y of type K. [1-6]

My interpretation of Gratton’s proto-schema, here, largely overlaps with the proposal by Black
(1996: 100-1).16 To obtain instances of this schema ‘items of type K’ is to be replaced with a
certain domain, and the capitals ‘F’ and ‘R’ are to be replaced with predicates which express
properties of and relations between the items in that domain. Before explaining the inferences, let
me provide one simple instance of the schema:

Guardians (Paradox instance)

(1) For all people x, x is reliable only if there is a guardian y and x is guarded by y.
(2) For all people x and y, x is guarded by y only if y is reliable.
(3) At least one person is reliable.
(4) Regress:
(a) My girlfriend is reliable.
(b) She is guarded by guardian no. 1. [a, 1]
(c) No. 1 is reliable. [b, 2]
(d) No. 1 is guarded by guardian no. 2. [c, 1]
(e) No. 2 is reliable. [d, 2] And so on.
(5) There is an infinity of reliable persons. [3, 4]
(6) There are not so many reliable persons.
(C) (1) is false: It is not the case that anyone who is reliable is guarded by a guardian. [1-6]
Hence, if we take Juvenal’s case as an instance of the Paradox Schema we obtain an argument
for the claim that there is at least someone who is reliable but not guarded by a guardian. Further
instances, but of a philosophical nature, will be provided in §3.

2. Validity

Is the Paradox Schema logically valid? That is, is its conclusion (C) true if the premises (2), (3)
and (6) are true? It can be shown that it is valid in a classical way. In this section I will explain
the inferences of the Paradox Schema, i.e. lines (4), (5) and (C), both in an informal and formal
way. I shall identify the rules on which the inferences rely and draw the attention to a suppressed
premise. This premise is suppressed in the Paradox Schema above, because it is virtually never a
point of discussion: its truth is accepted by both proponents and opponents of a given instance.

I should note that a sketch of the inferences can be found in Black (1987, 1988: 421-2, 1996: 99-
103), and some steps are identified in Gratton (2010: ch. 2). It seems important, however, to
check whether all steps of the Paradox Schema are logically valid. For if this were not the case,
then all regress arguments which are instances of this schema may be based on a logical mistake
and, moreover, the relevant conclusions may not be established by the premises at hand.

2.1. Informal explanation Paradox’s (4) The regress in (4) follows from the first two lines (1) and
(2):

(1) For all Ks x, x is F only if x stands in R to a new K y.

(2) For all Ks x and y, x stands in R to y only if y is F.

Both (1) and (2) state necessary conditions. Line (1) has it that the fact that x and y stand in R is
a necessary condition for the fact that x is F, and line (2) that the fact that y is F is a necessary
condition for the fact that x and y stand in R. I shall explain the meaning of the term ‘necessary
condition’ later on (§4.2.1). Here it suffices to say that ‘A only if B’ means at least ‘if A is true,
then B is true as well’, such that it allows us to derive ‘B’ from ‘A’ and ‘A only if B’ by Modus
Ponens. The conditions may be sufficient as well as being necessary. Their sufficiency, however,
can play no role in generating a regress. Consequently, if you want to generate a regress, the
conditions may not be merely sufficient.

The lines (1) and (2) may have variants, and may be multiplied as well. For example, the
following set of lines would also generate a regress: For all Ks x, x is F only if x is G; x is G only
if a new K is H; and x is H only if x is F.

It is important however that the regress formulas taken together, no matter their number, should
at least entail that: For any K, K is F only if there is a new K that is F. For example, it is possible
to generate a regress by the formula that anyone is reliable only if there is someone else who is
reliable. Indeed, (1) and (2) can always be integrated into one in order to simplify the derivation.
However, in many instances this is not very useful given that relevant parts of the dialectic might
be lost in the simplification. In particular, someone who is ready to accept/reject (1) (e.g. that
someone is reliable only if guarded by a guardian), need not accept/reject (2) (e.g. that someone
is guarded by someone else only if the latter is reliable), and vice versa.
Importantly, in my construction regresses are generated hypothetically, i.e. on the basis of the
hypothesis (a) of (4) (such as ‘my girlfriend is reliable’). The latter hypothesis is meant to be
completely arbitrary. For example, in the guardian case one could run the same argument on the
basis of the hypothesis that my brother is reliable. To my knowledge, this hypothetical
construction has not been used before in the literature. On Gratton’s account, for example,
regresses are not generated on the basis of a hypothesis, but on the basis of a premise, namely
premise (3) (i.e. the trigger). In my construction, in contrast, (3) plays no role in generating a
regress, but in the step from regress to conclusion, as we shall see next.17

Paradox’s (5)

This line is called the result, i.e. what follows from the regress. It follows from (3) and (4) in two
main steps. First, and as just noted, it follows from the regress that: For any K, K is F only if
another K is F. Logically speaking, we do not need the whole regress for this: as a and b are
arbitrary items (e.g. my girlfriend and guardian no. 1), it follows at once from ‘a is F only if b is
F’ that ‘For all Ks, K is F only if a new K is F’. Hence, it is slightly misleading to depict
regresses as an open list of lines (ending with ‘and so on’). Nevertheless, I have decided to stick
to this presentation as it is illustrative for reconstructing particular cases.

Second step: Paradox arguments rely on the following suppressed premise:

(7) If at least one K is F and any K is F only if a new K is F, then an infinity of Ks are F.

Given that we just obtained the second conjunct of the antecedent, and that the first conjunct is
supplied by premise (3), (5) follows at once: An infinity of Ks are F. (7) says for example: If
there is in fact a reliable person and if for any reliable person there is a new reliable person, then
there is an infinity of reliable persons.

It is important to point out that ‘y’ in line (1) is to range over new items, i.e. items not yet used in
the derivation (cf. Gratton 2010: 33). So in the guardian case, (1) should read: For all people x, x
is reliable only if there is a new, not yet used guardian y and x is guarded by y. For if the item
introduced were one already used in the derivation, the regress would run into a bigger or smaller
loop (e.g. guardians that would guard one another). Furthermore, if this were the case, it would
not follow that an infinity of items are F.18

As noted at the beginning of this section, this premise (7) is suppressed in the schema, because
its truth is virtually never a matter of dispute: it is often a common background assumption in the
debate. Moreover, if one finds (7) controversial, then it is worth noting that the infinity is not
really an issue in most cases. Important for the next step to (C) is that it follows that too many
items of type K are F, i.e. more items than there are in fact (e.g. that it follows that there are too
many reliable people, i.e. more reliable people than are in fact available). More about this below.

One last qualification regarding line (5). Namely: more results entailed by the regress might be
taken into consideration. Specifically, in many cases it follows from the regress not only that an
infinity of Ks are F, but also that an infinity of pairs of Ks stand in R. The regress of guardians,
for example, not only entails that there is an infinity of reliable persons, but also that an infinity
of persons are guarded by a guardian, and indeed that there is an infinity of guardians. This is
relevant because the argument might also run that the regress entails a bad result because there is
no infinity of guardians. This brings us to the last step.

Paradox’s (C)

The last step to (C) consists of three parts. First it is shown that the result entailed by the regress
is unacceptable, as it conflicts with independent considerations, i.e. something else that one does
not want to give up. For example, in the guardian case I assumed that there is no infinity of
reliable persons. This may be the case for various reasons. Perhaps an infinity of reliable persons
is psychologically, physically or even logically impossible. Perhaps you have empirical evidence
as to whether or not this is the case. What route you eventually take to line (6) is unimportant for
the main line of reasoning.

Second step: lines (5) and (6) form a contradiction. This is the paradox that the name of the
schema speaks of: the claims (1), (2), (3) (i.e. which together entail (5)) and (6) are jointly
inconsistent. By the classic definition, paradoxes consist of claims which appear individually
plausible, yet jointly inconsistent (Sainsbury 1987: 1).19 In this dissertation I will mainly retain
the ‘jointly inconsistent’ part of this definition, and not assume that it is always the case that
regress arguments concern claims which appear individually plausible (even though they may, of
course). For example, line (1) of the guardian case does not, I think, appear very plausible: For
all people x, x is reliable only if there is a guardian y and x is guarded by y.

Last step: given the contradiction, at least one of the assumptions has to be rejected by Reductio
Ad Absurdum. In the schema it is given that lines (2), (3) and (6) are premises, and that line (1)
is the hypothesis to be rejected. There are, to be sure, variants of the Paradox Schema which do
not conclude with the rejection of (1), but rather the rejection of (2), (3) or (6). This debate
would lead us into the question of what is to be taken as a premise, and what as a hypothesis (i.e.
for Reductio Ad Absurdum). For example, if one wants to reject (3), i.e. the trigger, then (3) is
the hypothesis and (1), (2) and (6) premises.

Importantly, (3) has to be an existentially quantified statement. It does not say: some particular K
is F (e.g. one particular person that one may have in mind is reliable). It rather says: an arbitrary
item of type K is F (e.g. at least one person is reliable, whoever it is). In the guardian case, for
example, if you want to keep (1), (2) and (6) in place, it is not enough to reject the statement ‘my
girlfriend is reliable’ because similar arguments can be constructed on the basis of ‘I am
reliable’, ‘you are reliable’, etc. Instead, what is needed in that case is a rejection of ‘at least one
person is reliable’.

At this point, I have explained the inferences of the Paradox Schema in a semi-formal way. As
we shall see next, all relevant inference rules are valid according to classical predicate logic such
that the Paradox Schema is valid by this logic. The Paradox Schema hosts no logical mistakes,
meaning that (C) can indeed be obtained via (1)-(6).

2.2. Formal explanation

Here, I present a formal rendering of the Paradox Schema. Some explanations are in order. To
begin with, I will employ standard natural deduction abbreviations of the inference rules (see the
table below). I will use the propositional calculus by Nolt et al. (1988: ch. 4), and the first-order
extension by Gamut (1982: 142-7). This means that I will employ a strict distinction between
premises and hypotheses (cf. Gamut 1982: 132-3, Nolt et al. 1988: 87, Batens 1992: 67, Woods
et al. 2000: 129). All portions of hypothetical reasoning will clearly be marked by vertical lines.20

PREM Premise
HYP ¬I Hypothesis for ¬I
HYP →I Hypothesis for →I
¬I Reductio Ad Absurdum: AHYP, …, B∧¬B ⊢ ¬A21
→I Conditional Proof: AHYP, …, B ⊢ A→B
→E Modus Ponens: A, A→B ⊢ B
∧I Conjunction: A, B ⊢ A∧B
∧E Simplification: A∧B ⊢ A
∃I Existential Generalisation: [a/x]A ⊢ ∃xA
∃E Existential Instantiation: ∃xA, [a/x]A→B ⊢ B
∀I Universal Generalisation: [a/x]A ⊢ ∀xA
∀E Universal Instantiation: ∀xA ⊢ [a/x]A

The rules of Existential Instantiation (∃E) and Universal Generalisation (∀I) should be applied
carefully as they do not apply across the board. For example, from ‘there is a philosopher in the
room’ one cannot just conclude ‘Socrates is in the room’ and from ‘Socrates is in the room’ one
cannot just conclude ‘all philosophers are in the room’. Nevertheless, sometimes it is allowed to
use ∃E and ∀I. In this case, we would be allowed to apply ∃E and ∀I if we would use the label
‘Socrates’ to refer to an arbitrary philosopher. Here are correct, formal applications of ∃E and
∀I:22

(1) ∀x(Fx→∃yGy) PREM

(2) ∃xFx PREM
(3) Fa HYP →I
(4) Fa→∃yGy 1; ∀E
(5) ∃yGy 3, 4; →E
(6) Fa→∃yGy 3-5; →I
(7) ∃yGy 2, 6; ∃E
(8) ∀x(Fx→∃yGy) 6; ∀I

In this case, we may conclude from (2) and (6) to (7) by ∃E, given that ‘a’ in (6) forms an
arbitrary item, that is, given that ‘a’ does not occur in the premises (such as (2) itself), nor in any
undischarged hypothesis, nor in the conclusion, i.e. (7). Also, we may conclude from (6) to (8)
by ∀I, again given that ‘a’ in (6) forms an arbitrary item, that is, given that ‘a’ does not occur in
the premises, nor in any undischarged hypothesis (for these specific restrictions, cf. Gamut 1982:
142-7, Batens 1992: 170, Woods et al. 2000: 209-11).
Next, the numberings of the lines (1)-(C) correspond to the numberings of my semi-first order
schemas presented above (the latter line up closely with the dialectic of regress arguments, see
§4.1.5). Given that I wanted to keep my semi-formal schema as readable and applicable as
possible, at two points it will slightly differ from the upcoming formalisation. First, in the latter
all suppressed premises and inferences are made explicit (i.e. no lines will be suppressed or
superfluous). Second, in the formalisation the regress is no longer taken as an open list of lines.
The latter is illustrative for reconstructing particular cases, yet, as we shall see, the ‘and so on’
serves no logical purpose. Finally, some of the predicates and premises will need some additional
explanation (particularly line (7)). These explanations are provided right after the formalisation.

Key
Kx: x is of type K
Fx: x has property F
Rxy: x stands in relation R to y
IR: an infinity of Ks are F

Example:
Kx: x is a proposition
Fx: the dispute about x is settled
Rxy: the dispute about x is settled by y
IR: the dispute is settled about an infinity of propositions

Paradox Schema

(2) ∀x∀y((Ky∧Rxy)→Fy) PREM

(3) ∃x(Kx∧Fx) PREM
(7) (∃x(Kx∧Fx)∧(∀x((Kx∧Fx)→∃y(Ky∧Fy∧Rxy))))→IR PREM
(6) ¬IR PREM
(1) ∀x((Kx∧Fx)→∃y(Ky∧Rxy)) HYP ¬I
(i) Ka∧Fa HYP →I
(ii) (Ka∧Fa)→∃y(Ky∧Ray) 1; ∀E
(iii) ∃y(Ky∧Ray) i, ii; →E
(iv) Kb∧Rab HYP →I
(v) ∀x((Kb∧Rxb)→Fb) 2; ∀E
(vi) (Kb∧Rab)→Fb v; ∀E
(vii) Fb iv, vi; →E
(viii) Kb∧Fb∧Rab iv, vii; ∧I
(ix) ∃y(Ky∧Fy∧Ray) viii; ∃I
(x) Kb∧Rab→∃y(Ky∧Fy∧Ray) iv-ix; →I
(xi) ∃y(Ky∧Fy∧Ray) iii, x; ∃E
(xii) (Ka∧Fa)→∃y(Ky∧Fy∧Ray) i-xi; →I
(xiii) ∀x((Kx∧Fx)→∃y(Ky∧Fy∧Rxy)) xii; ∀I
(xiv) ∃x(Kx∧Fx)∧(∀x((Kx∧Fx)→∃y(Ky∧Fy∧Rxy))) 3, xiii; ∧I
(5) IR xiv, 7; →E
(8) IR∧¬IR 5, 6; ∧I
(C) ¬(∀x((Kx∧Fx)→∃y(Ky∧Rxy))) 1-8; ¬I
As noted, lines (1)-(3) may have variants in terms of one- or many-place predicates and their
number (this does not hold for the Failure Schemas that will be presented in §2.3). Also, it is
easy to see how variants of the Paradox Schema can be constructed where line (2), (3) or (6),
rather than (1), is the hypothesis for reductio. Note that the distinction between types (here: ‘K’)
and properties (here: ‘F’) is not meant to be a fundamental distinction (i.e. that types and
properties are different kinds of things). The type ‘K’ just refers to the domain at issue (such as
the domain of propositions). This domain has to be explicit in the schema in order to capture
Failure I arguments (to be discussed below). For the sake of uniformity, I decided to include ‘K’
in the formalisation of all schemas.

Line (7) requires some explanation. First, ‘IR’ is a placeholder for the schematic sentence ‘there
is an infinity of Ks that are F’. I should flag that this is no full first-order expression (as ‘infinity’
is no familiar first-order term). As we shall see below, it can nevertheless be stated in first-order
terms when the whole line (7) would be true.

In my construction line (7) is taken as a premise for two reasons: first, to assure logical validity
and, second, in actual debates it is virtually always taken for granted, i.e. those who wish to resist
the charge of a regress do not usually attack (7), but other premises (see §4.1.5).

As it stands, though, (7) has many false instances. For it merely says (assuming all items are
from the same domain K): if there is one item that is F and if for any item x that is F, there is a y
that is F and x stands in R to y, then an infinity of items are F. Here is for example one dubious
instance: if there is at least one bald person, and if for any bald person x there is a bald person y
and x has exactly the same DNA as y (such that x=y), then there is an infinity of bald persons.

To be sure, the fact that instances of (7) may be false forms no problem for the logical validity of
the argument. Rather, the problem is that (7) does not say what it should say, namely that ‘y’ has
to be a new item in the domain. Unfortunately, the phrase ‘there is a new item y’ cannot be
captured by a familiar logical constant. For often it does not merely mean ‘there is an item y and
y is distinct from x’, but rather ‘there is an item y that is distinct from all other items mentioned
earlier in the regress’. To capture this formally, we have to assume certain properties of the
relation ‘R’ which block all loops (i.e. that may prevent the regress from being infinite). Now,
the properties that will do the job are that R is asymmetric and transitive:23

(a) ∀x∀y(Rxy→¬Ryx)
(b) ∀x∀y∀z((Rxy∧Ryz)→Rxz)

Yet, the problem with this is that not all relations in regress arguments are transitive (consider
e.g. the relation ‘the dispute about x is settled by y’). In the following, I shall explain two distinct
solutions to this problem.

First solution. We could require that the transitive closure of R, i.e. R*, rather than R itself, is
irreflexive. Roughly, items x and y stand in the transitive closure R* of a relation R just in case
there is a chain of R-pairs of arbitrary length between x and y. Furthermore, if R* is irreflexive,
then no item can be reached via R to itself, whether this be directly or indirectly via a whole
series of pairs related by R (e.g. no guardian can guard himself, directly or indirectly via other
guardians; or the dispute about a proposition cannot be settled by that same proposition, directly
or indirectly via other propositions). Black (1996: 100), who employs this notion in this context,
provides a useful example that shows that R having a transitive closure R* does not entail R
itself being transitive: fatherhood is not transitive, even though the transitive closure, let us call
this relation ‘ancestor’, is irreflexive (nobody is an ancestor of him- or herself). Here is how one
can describe the transitive closure of R, i.e. R*, being irreflexive in first-order logic:

(c) ∀x∀y(Rxy→R*xy)
(d) ∀x∀y(∃z(R*zy∧Rxz)→R*xy)
(e) ∀x¬R*xx

Second solution. We could introduce an additional relation ‘<’ distinct from R, whose only job is
to order the Ks, and make sure that all items introduced in the regress are new items (such that
they form an infinite, non-circular series). To do this, ‘x<y’ can be read as ‘x occurs earlier in the
regress than y’ and has to satisfy the following conditions:24

(f) ∀x∀y((x≠y∧Kx∧Ky)→(x<y∨y<x))
(g) ∀x∀y(x<y→(Kx∧Ky))
(h) ∀x∀y∀z((x<y∧y<z)→x<z)
(i) ∀x¬x<x

Moreover, the advantage of this solution is that it allows us to formulate both ‘IR’ and ‘¬IR’ in
first-order terms:

IR ∃x(Kx∧Fx)∧∀x((Kx∧Fx)→∃y(x<y∧Ky∧Fy))
¬IR ∃x(Kx∧Fx∧∀y((Ky∧x<y)→¬Fy))

For example: The dispute about at least one proposition is settled and the dispute about any
proposition is settled only if there is a new proposition which is such that the dispute about it is
settled vs. For at least one proposition x, the dispute about x is settled and, for all new
propositions y, the dispute about y is not settled.

Now back to our problem: though (7) makes the schema valid, it has many false instances. At
this point we can say, in more precise terms, that (7) is true whenever (a)-(b), (c)-(e), or (f)-(i)
are true (or all).25

Another final option, as suggested by Cling (2009: 343), would be to drop the idea of ‘infinity’,
and to replace ‘there is an infinity of Ks that are F’ with ‘there is an endless regress of Ks that are
F’ (where the latter, but not the former, includes finite, circular regresses). If we change this in
both (6) and (7), then we would not need to block loops and place restrictions on the relation, and
yet we still obtain a contradiction in (8) such that we can apply ¬I.26 This solution will work in
all cases where infinity is not really an issue (for this, see §4.2.3 below).

3. Boundaries
The question of boundaries is this: Why does the Paradox Schema begin at line (1) and stop at
(C)? Why not more lines, or fewer for that matter? There are two arguments here.

The first argument is that the Paradox Schema had to consist of at least (1)-(5), because
otherwise it would not fulfil the first three desiderata from §2.1. To recall those: the schema had
to be such that the regress occupies one of the lines, i.e. (4); that it is derived from
premises/hypotheses, i.e. (1)-(2); and that from it some conclusion is drawn, i.e. (5).27

The second argument, due to Gratton (2010: 9), is that the Paradox Schema has to consist of (1)-
(C) if it is to have Reductio Ad Absurdum arguments as instances. Reductio Ad Absurdum
arguments are arguments where some hypothesis is assumed for the purpose of deriving
unacceptable consequences from it (which in most cases is a straight contradiction, in others
merely something disadvantageous). The hypothesis is rejected on that basis. Take the
Omnipotence Paradox as a simple example:

- Omnipotent beings should be able to do anything whatever.

- It is logically impossible to accomplish the following two tasks at once: create a stone that one
cannot lift, and lift it.

By these, we obtain a paradox: omnipotent beings should be able to but cannot do anything
whatever. To avoid or resolve the paradox, something needs to be rejected. So either we reject
that omnipotent beings should be able to anything whatever (i.e. even logically impossible
things), or we reject that no one can create a stone that one cannot lift, and then lift it.28

Now, if regress arguments are to be special cases of such arguments, then the boundaries of the
former are the boundaries of the latter. That is, in such a case, regress arguments should contain
both the hypothesis to be rejected and the conclusion where it is actually rejected (plus what lies
in between). These are (1) and (C).

It is worth pointing out, however, that Gratton (2010: 9-11) acknowledges the existence of
regress arguments that consist of lines (1)-(5) only. These arguments are the so-called benign
regress arguments. The latter have in common with full Paradox instances that a regress is
entailed by premises, and a result in turn by the regress, but they differ in that benign regress
arguments do not show that this result is unacceptable (and so neither show that something is to
be rejected). Take for instance the following simple argument:

(1) There is Socrates.

(2) Anything has proper parts (i.e. parts not identical to the whole of which they are a part).
(3) Regress: Socrates has proper parts, the proper parts of Socrates have proper parts, the proper
parts of the proper parts of Socrates have proper parts, etc. [1-2]
(4) Socrates has an infinity of proper parts. [3]

This is a regress argument in that it fulfils the three basic desiderata for something to be a regress
argument. Still, this can be expanded to a full instance of the Paradox Schema only if it is also be
shown that Socrates’ proper parts are not infinite.
It may be useful here to invoke the following distinction between positive and negative regress
arguments:

• A regress argument is negative if it consists of instances of lines (1)-(C) of the Paradox Schema
(where (C) is a rejection).

• A regress argument is positive if it consists of instances of lines (1)-(5) of the Paradox Schema
(where (5) is not a rejection).

Such negative/positive terminology is admittedly somewhat arbitrary (cf. Sanford 1984: 100,
Nolan 2001: 523); at least, that is, if negative arguments are meant to be negations and positive
arguments are not. For regress formulas are universally quantified statements, and, provided that
the domain is non-empty, the negation of such a statement (e.g. ‘it is not the case that for all
persons x, x is reliable only if x is guarded by a guardian’) is equivalent to an existentially
quantified statement which is not a negation (‘there is at least one person x such that x is reliable
and not guarded by a guardian’).

There is however one area where this equivalence may fail, namely in cases with vague
properties (also known as Sorites cases). Consider the following argument adapted from Sanford
(1975: 521-4). I have reframed the case in Paradox-format:29

Sorites (Paradox instance)

(1) For any possible person x, if x is short, then there is another possible person y and y is one
millimeter taller than x.
(2) For any possible persons x and y, if x is short and y is one millimeter taller than x, then y is
short.
(3) Regress:
(a) Person no. 1 is short.
(b) Person no. 2 is one millimeter taller than no. 1. [a, 1]
(c) Person no. 2 is short. [a, b, 2]
(d) Person no. 3 is one millimeter taller than no. 2. [c, 1]
(e) Person no. 3 is short. [c, d, 2] And so on.
(4) Person no. 9999 is short. [3]
(5) Person no. 9999 is tall.
(C) (2) is false: It is not the case that for any possible persons x and y, if x is short and y is one
millimeter taller than x, then y is short. [1- 5]

From this one may want to conclude:

(C*) There are at least two possible persons x and y such that x is short and y is one millimeter
taller than x, and y is tall.

But this assumes the existence of a clear borderline case between shortness and tallness.
Sanford’s point here is that universally quantified statements such as (2) can be false even if
there are no clear counterexamples. Yet, apart from such Sorites cases, equivalence may be
presumed between universally and corresponding existentially quantified statements.

Likewise: the negation of a trigger statement is the negation of an existentially quantified

statement (e.g. ‘it is not the case that there is an x such that x is reliable’), and, provided that
anyone is either reliable or not (i.e. provided the Excluded Middle is applicable), this is
equivalent to a universally quantified statement (‘for all persons x, x is unreliable’). In both
cases, there is a negative and a positive way of expressing the same thing.

Finally, the distinction between positive and negative regress arguments is also relative to the
schema of which they are an instance. So in §2.3.3, we shall see that the positive/negative
distinction can be cashed out in another way as well.

4. Literature

Finally a note on the literature. The Paradox Schema has been discussed most extensively by
Black (1996) and Gratton (2010) (both draw from earlier work, see the references list). Of
course, they did not obtain the schema from scratch but took their inspiration from cases in the
literature. Black (1996: 96) cites for example the following of Aristotle:

[…] we do not choose everything for the sake of something else (for at that rate the process
would go on to infinity, so that our desire would be empty and vain). (Aristotle, Nicomachean
Ethics 1094a)

This would be an instance of the Paradox Schema because Aristotle concludes to the negation of
the claim that everything is chosen for the sake of something else (in particular, the Highest
Good is not chosen for the sake of anything else). And Gratton (2010: 2) refers to the following
text by Ryle:

The regress is infinite, and this reduces to absurdity the theory that for an operation to be
intelligent it must be steered by a prior intellectual process. (1949: 32)

This would be an instance of the Paradox Schema because Ryle concludes to the negation of the
claim that all intelligent actions are steered by prior intelligent processes.

As I shall explain later (in §6), there is a clear difference between acknowledging the Paradox
Schema and holding that all regress arguments can or should be reconstructed in terms of it.
Black and Gratton uphold both claims. Yet, there are others who acknowledge or suggest
something close to the Paradox Schema, yet deny that all regress arguments can or should be
reconstructed in terms of it. They hold that certain regress arguments take a different form.
Among this group are, most prominently, Sanford (1984), Schlesinger (1983) and Day (1986).
Further references will be provided in due course.

2.3. The Failure Schemas

In the following I present a second regress argument schema: the Failure Schema. I shall
distinguish between two variants of this schema, i.e. I and II. Next, I address queries about the
logical validity and boundaries of each variant. Finally, I identity the origins of the schemas;
though the schemas themselves are my own contribution to the literature.

1. Schemas

Recall the following hypothesis from §1.1: Fail–B Regress arguments are arguments which
demonstrate that a certain solution never solves a given problem because it gets stuck in a
regress.

Here is a first step towards the specification of this hypothesis:

Proto Failure Schema

(1) Problem.
(2) Solution.
(3) Extra premise.
(4) Regress. [1-3]
(5) Failure. [4]
(C) If (2), then (5). [2-5]

The reasoning here starts from a certain problem that is to be solved, and a certain solution that is
considered for this problem. It can then be shown that this solution entails a regress if an
additional premise is in place. Finally, it follows from the regress that the problem is never
solved by the solution under consideration (and hence that the problem calls for some other
solution). Again, the premise/hypothesis distinction proves to be important here. This time, lines
(1) and (3) are the premises, and line (2) is the hypothesis (now for Conditional Proof, as I shall
explain). The conclusion (C) basically says: If you assume the solution proposed in (2), then you
will fail to solve the problem at issue. In this sense, this schema is a specification of Fail–B
above.

Crucially, this schema has two varieties, which I shall give the extra label ‘I’ and ‘II’. Informally,
they can be rendered thus:

• In Failure Schema I, a general problem is never solved because the solution under
consideration generates a regress of more instances of the general problem.

• In Failure Schema II, a particular problem is never solved because the solution under
consideration generates a regress of more and more problems which are to be solved in order to
solve the initial one.

So the main difference is this: in one case it is concluded that a general problem is never solved,
in the other that a particular problem is never solved. Before explaining what is meant by
‘general’ and ‘particular’ here, and how the difference between the two conclusions can be
explained in terms of full argument schemas, let me first set out an instance of both:
Beer (Failure I instance)

(1) You have to quench all your thirsty feelings.

(2) For any thirsty feeling x, if you have to quench x, you drink a beer to quench x.
(3) For any thirsty feeling x, if you drink a beer to quench x, then the beer generates a new thirsty
feeling y.
(4) Regress:
(a) You have to quench thirsty feeling no. 1.
(b) You drink beer no. 1. [a, 2]
(c) You have to quench thirsty feeling no. 2. [b, 3, 1]
(d) You drink beer no. 2. [c, 2]
(e) You have to quench thirsty feeling no. 3. [d, 3, 1]
And so on.
(5) You always have to quench a new thirsty feeling in addition to quenching any thirsty feeling.
[4]
(C) You will never quench all your thirsty feelings if you drink a beer every time you have to
quench a thirsty feeling. [1-5]

This is an instance of Failure Schema I because there is a general problem in line (1) and the
solution under consideration in line (2) entails more instances of the general problem (i.e. more
thirsty feelings to be quenched), such that it is eventually concluded in (C) that, given the
solution under consideration, the general problem is never solved.

Guardians (Failure II instance)

(1) For at least one person x, you should have x guarded.

(2) For any person x, if you should have x guarded, you hire a guardian for x.
(3) For any persons x and y, if you hire a guardian y for x, then you first should have y guarded
in order to have x guarded.
(4) Regress:
(a) You should have your girlfriend guarded.
(b) You hire a guardian no. 1 for your girlfriend. [a, 2]
(c) You should have guardian no. 1 guarded first. [b, 3]
(d) You hire a guardian no. 2 for no. 1. [c, 2]
(e) You should have guardian no. 2 guarded first. [d, 3] And so on.
(5) For any person x, you first should have a regress of persons guarded in order to have any
person guarded. [4]
(C) You will never have any person guarded if you hire a guardian every time you should have
someone guarded. [1-5]

This is an instance of Failure Schema II as there is a particular problem in line (1) and the
solution under consideration in line (2) entails new, similar problems which are to be solved
before the initial one is solved (i.e. more persons to be guarded before you will have your
girlfriend guarded). Thus it is eventually concluded in (C) that, given the solution under
consideration, the initial, particular problem is never solved.
It could be noted that line (2) of either argument will be proven false as soon as the beers or
guardians run out. However, this is unimportant for the following reasons. First, I am interested
in the general line of reasoning and use these cases only as illustrations. Second, (2) is assumed
only as a hypothesis, i.e. it is not regarded as true, but only considered in order to derive a failure
from it. So the reasoning is that if you would consequently apply a certain solution, then you
would never solve the problem, no matter whether or not in fact you apply that solution. Also,
this reasoning holds whether or not it is possible for you to apply the solution (e.g. to invoke so
many beers or guardians). I will return to this point several times in this dissertation.

To obtain the argument schemas, all variable parts of the arguments must be replaced with
schematic letters. Specifically, in the schemas ‘items of type K’ stands for a specific domain, the
Greek letters ‘φ’ and ‘ψ’ stand for predicates which express actions involving the items in that
domain, and ‘you’ stands for an arbitrary person or agency that can solve problems. The two
schemas are presented below. Whenever I talk about Failure arguments in this dissertation, I talk
about instances of these semifirst-order schemas. A full first-order rendering will be presented
soon.30

Failure Schema I

(1) For all items x of type K, you have to φ x.

(2) For all items x of type K, if you have to φ x, then you ψ x.
(3) For all items x of type K, if you ψ x, then there is a new item y of type K.
(4) Regress:
(a) You have to φ a.
(b) You ψ a. [a, 2]
(c) You have to φ b, where a≠b. [b, 3, 1]
(d) You ψ b. [c, 2]
(e) You have to φ c, where a≠b≠c. [d, 3, 1]
And so on.
(5) For all items x of type K, you always have to φ a new item of type K in addition to φ-ing x.
[4]
(C) If you ψ all items of type K that you have to φ, then you will never φ all items of type K. [1-
5]

Failure Schema II

(1) For at least one item x of type K, you have to φ x.

(2) For all items x of type K, if you have to φ x, then you ψ x.
(3) For all items x of type K, if you ψ x, then there is a new item y of type K, and you first have
to φ y in order to φ x.
(4) Regress:
(a) You have to φ a.
(b) You ψ a. [a, 2]
(c) You first have to φ b, where a≠b. [b, 3]
(d) You ψ b. [c, 2]
(e) You first have to φ c, where a≠b≠c. [d, 3]
etc.
(5) For all items x of type K, you first have to φ a regress of new items of type K in order to φ x.
[4]
(C) If you ψ all items of type K that you have to φ, then you will never φ any item of type K. [1-
5]

Three main differences are apparent between these two schemas. First, and as we already knew,
the problems in (1) differ. The problem in Failure I is a universally quantified task (or, as I called
it, ‘general problem’), whereas the problem in Failure II is an existentially quantified one (or
‘particular problem’). In the set-up of the schemas, therefore, I decided to take problems as tasks
that are not yet, but have to be, accomplished, and solutions as potential strategies/actions to
accomplish those tasks.31

Second, even though the regresses in both schemas look the same (i.e. if we ignore the term
‘first’, then the lines (a), (b), (c), etc. are simply identical), they are generated differently. The
main line responsible for this difference is (3), which is substantially longer in Failure II. I shall
explain this in the next section where I consider the inference steps of the two schemas.

Third, the conclusions in (C) differ. Moreover, Failure II’s conclusion is considerably stronger
than Failure I’s conclusion in that the former entails the latter, but not vice versa. That is, if you
will never φ any K, then (provided that there is at least one such item) it cannot be the case that
you φ all Ks. For example, if you will never have anyone guarded, then it cannot be the case that
you have your girlfriend guarded (for example). Yet, if you will never φ all Ks, it may still be the
case that you φ some of them. If you will never quench all of your thirsty feelings, it may still be
the case that you quench some (or even many) of them.

2. Validity

Are the two Failure Schemas logically valid? Is their conclusion (C) true if their premises (1) and
(3) are true? It can be shown that they are valid in a classical way. Again, I shall explain the
inferences, i.e. lines (4), (5) and (C), both in an informal and formal way.

2.1. Informal explanation

Failure I’s (4)

The regress in (4) follows from (1)-(3). Lines (2) and (3) could be integrated into one in order to
simplify the derivation, namely as: For all Ks x, if you have to φ x, then there is a new K-item y
(e.g. ‘for any thirsty feeling x, if you have to quench x, then there is a new thirsty feeling y’). In
that case, however, we would lose the dialectic and fail to distinguish what is the solution for the
problem from what is additional. In contrast to the Paradox Schema, which has variants
regarding one- and many-place predicates as well as variants regarding the number of lines that
generate the regress, the Failure Schemas have no such variants: all predicates and lines are fixed
(as they have a strict problem/solution structure).
Again, the regress in (4) is generated hypothetically, i.e. on the basis of the hypotheses (2) and
(a) of (4) (such as ‘you have to quench thirsty feeling no. 1’). The latter hypothesis is meant to be
completely arbitrary, as one could run the same argument on the basis of the hypothesis that you
have to quench thirsty feeling, say, no. 7. The manner in which regresses are generated Failure I-
wise differs significantly from how they are generated Failure II-wise (to be explained soon). In
this Failure I case, any problem is generated because of an interplay between lines (1) and (3):
(3) generates a new K such that it is to be φ-ed by (1) (as (1) says that all items of that type have
to be φ-ed). This is depicted on the left side of Figure 1 (where (a) refers to an initial problem
‘you have to φ a’, (b) to the solution ‘you ψ a’, and (c) to the newly generated problem ‘you have
to φ b’).

(a) (a)
| |
v v
[2] [2]
| |
v |
(b) (b)
| |
v v
[1], [3] [3]
| |
v v
(c) (c)

Failure I Failure II

Figure 1: Entailment of problems and solutions

Failure I’s (5)

(5) follows from (4) at once. First it follows from the first steps of the regress that if you have to
φ a, then you have to φ another K, namely b. Given that a and b are two arbitrary items here, we
may generalise this for any K:

(5) For all Ks x, you always have to φ a new K in addition to φ-ing x.

Hence, just as in the Paradox case, we do not need the whole regress: a few steps suffice to
obtain (5). Again, I have decided to stick to the presentation of regresses as open lists of lines as
this is illustrative for reconstructing particular cases.

Failure I’s (C)

(C) follows in two steps. First assume:

(6) If for at least one K you have to φ it and if for all Ks x, you always have to φ a new K in
addition to φ-ing x, then you will never φ all Ks.

This premise is suppressed in the schema, for its truth seems completely general. That is,
whether or not (6) holds does not seem to depend on specific instances. Consider for example the
following instances: If you have to write down at least one number and if there is always yet
another number to be written down, then you will never write down all numbers. Or: If you have
to report at least one day and if there is always yet another day to be reported, then you will
never report all days. Such instances of (6) are virtually never a point of dispute. Still, I will
point to some controversial instances later on in §4.2.2.

Now if we also assume that for at least one K you have to φ it (which is fairly uncontroversial
given premise (1) that you have to φ all Ks), then given that we obtained the second conjunct of
(6)’s antecedent in the previous step (i.e. (5)), it follows that:

(8) You will never φ all Ks.

That is: you fail to solve the problem from line (1). Please note that this result is not inconsistent
with (1): it is not inconsistent to say that a problem has to be solved and that a given solution
never solves it (i.e. that you have to φ all Ks, but never succeed in this). To be sure, (1) would
conflict with (8) if it would imply that you φ all Ks. However, (1) does not imply this.

The final step of Failure I arguments is that (C) follows from (2) to (8) by Conditional Proof. It
says: If you consider the solution of line (2) as a hypothesis, then you obtain line (8).

Failure II’s (4)

In this case, the regress in (4) follows from lines (2) and (3) only: all of the regress-generating
work is done by premise (3), which is substantially longer than the parallel line in Failure I (i.e. it
comprises the extra clause ‘you first have to φ y in order to φ x’). That is, unlike Failure Schema
I, it generates the new problems in the regress without the interference of line (1) (cf. the right
side of Figure 1). No Failure II regress is entailed without the complete premise (3). That is, the
regress would not follow if you have to φ x of type K, you ψ x, yet one of the following would
fail:

• there is a new item y of type K;

• you first have to φ y in order to φ x.

Take the guardian case where you should have your girlfriend guarded. But now suppose that it
fails to be the case that there is a new person (e.g. because the guardian is not a person), or that it
fails that you first should have the guardian guarded in order to have your girlfriend guarded (e.g.
because the guardian is reliable and not in need of any further guardian). In neither case would a
regress ensue. The phrase ‘you first have to φ b in order to φ a’ derives from Sextus:
In order to decide the dispute that has arisen […], we have need of an agreed-upon criterion by
means of which we shall decide it; and in order to have an agreed-upon criterion it is necessary
first to have decided the dispute about the criterion. (Outlines, 2.20)

Note that ‘you first have to φ b in order to φ a’ admits of some alternative phrasing such as ‘you
φ a only if you φ b first’, ‘φ-ing b is a necessary means to φ-ing a’ and ‘your φ-ing a depends
upon your φ-ing b, but not vice versa’. These phrases indicate a instrumental take of ‘first’, i.e.
the latter term induces an instrumental (i.e. means/end) order, rather than a temporal order. For
example, it need not be the case that the problem of φ-ing b needs be solved earlier in time. What
matters is the asymmetry between the problems: φ-ing b is to be a necessary means to φ-ing a,
and not the other way around. For example, making sure that the guardian of your girlfriend is
guarded is meant to be a necessary means to making sure that your girlfriend is guarded, and not
the other way around.32

I have to clarify something about premise (1) (‘for at least one K, you have to φ it’) here. It does
not mean: You have to φ some particular K (e.g. to have one particular person guarded that one
may have in mind). It rather says: You have to φ an arbitrary item of type K (e.g. to have at least
one person guarded, whoever it is). In my construction, (1) plays no regressgenerating role and
will be used only later on in the argument. Hence, line (a) of (4) (e.g. ‘you should have your
girlfriend guarded’) is not to be seen as following from (1) (but rather as a hypothesis on the
basis on which a regress can be generated).

Failure II’s (5)

First it follows from the first steps of the regress in (4) that if you have to φ a, then you first have
to φ another K, namely b, in order to φ a. Given that a and b are two arbitrary items here, we may
generalise this for any K: For all Ks x, you have to φ a new K first in order to φ x. Now assume:

(7) If for all Ks x, you first have to φ a new K in order to φ x, then you first have to φ a regress of
new Ks in order to φ x.

By this we immediately obtain:

(5) For all Ks x, you first have to φ a regress of new Ks in order to φ x.

In this dissertation I take (7) as a premise. Still, its motivation might be sketched as follows.
Suppose you first have to φ b in order to φ a, and that you first have to φ c in order to φ b, then
by transitivity you first have to φ c in order to φ a. For example: If you should have guardian no.
1 guarded in order to have your girlfriend guarded, and if you should have guardian no. 2
guarded in order to have no. 1 guarded, then you should have no. 2 guarded in order to have your
girlfriend guarded. Now, if you repeat this procedure, then it follows that you should have a
whole regress of guardians guarded in order to have your girlfriend guarded. And this is a
specific instance of (7).

This premise (7) might also be formulated in terms of the notion of a ‘supertask’: If for all Ks x,
you first have to φ a new K in order to φ x, then you first have to carry out a supertask in order to
φ x. Later in §4.2, I shall argue that Failure regresses can indeed be seen as a special kind of
supertask.33

Failure II’s (C)

(C) follows in two steps. First assume:

(6) If for at least one K you have to φ it and if for all Ks x, you first have to φ a regress of new
Ks in order to φ x, then you will never φ x.

This states that if you have to carry out a single task and if it so happens that there is a regress of
means required to accomplish it, then you will never accomplish that task. Just like the
corresponding Failure I line, this premise is suppressed as its truth is virtually never a matter of
dispute. Instances are, for example: If there is a regress of persons to be guarded in order to have
anyone guarded (which you should), then you will never have anyone guarded. Or: If there is a
regress of decisions to be made in order to make any decision (which you should), then you will
never make any decision. But again, I shall point to a few controversial instances of (6) in §4.2.2.

Now given that the first conjunct of (6)’s antecedent is supplied by premise (1), and that we
obtained the second conjunct in the previous step (i.e. (5)), it follows that:

(8) You will never φ any K.

That is: you fail to solve the problem from line (1). As in the other case, this result is not
inconsistent with (1): it is not inconsistent to say that a problem has to be solved and that a given
solution never solves it (or, in this case, to say that you have to φ any K, but will never succeed
in doing so). The final step of Failure II arguments is that (C) follows from (2) to (8) by
Conditional Proof. It says: If you consider the solution of line (2) as a hypothesis, then you
obtain line (8).

As I will argue later in §4.2.3, the term ‘never’ is to be construed non-temporally (just as ‘first’ is
no time indicator, see above). Saying that you will never φ any K is saying that at no point in a
regress (i.e. as governed by lines (1)-(3) of the Failure Schemas) will it be the case that you φ at
least one/all K(s). Or again, for any point in the regress will it be the case that a solution entails a
new problem that must be solved in order to solve the initial one. And this holds even if the
solutions take no time at all.

Alternatives for ‘never’ might be ‘do not’ and ‘cannot’. Replacing ‘never’ with ‘do not’ seems
plausible: if you will never φ, then you do not φ either. Yet, replacing ‘never’ with ‘cannot’ is
not plausible: if you will never φ, then it might still be the case that you can or cannot φ. Still,
sometimes ‘cannot’ is used. Compare the following two passages:

This question cannot be answered until we move up to the next level, and so on ad infinitum.
What this means is that our original question cannot be answered. (Dodd 1999: 150)
If we continue in this way, of course, we are led to an infinite regress and we will never have an
answer to our original question. (Chisholm 1982: 64)

I have opted for the ‘never’ version for several reasons. First, the ‘never’ version does not
depend on the issue whether you can or cannot φ so many Ks (e.g. answer to so many questions).
You just never φ (e.g. never answer to the original question), no matter what your capacities.
This point is easily overlooked, so in the following I shall spell out a ‘cannot’ reconstruction of
the guardian case.34

Guardians (Cannot)

(1*) You can have someone guarded only if you can hire a guardian for x.
(2*) You can hire a guardian for x only if you can have him guarded.
(3*) So, you can have someone guarded only if you can hire a whole regress of guardians. [1*-
2*]
(4*) You cannot hire a regress of guardians.
(5*) So, you cannot have anyone guarded. [3*-4*]

In contrast to this, Failure arguments do not rely on premise (4*), i.e. your capacity to hire a
regress of guardians. Rather, they rely on the idea that if you should have a regress of persons
guarded in order to have anyone guarded, then you will never have anyone guarded in the sense
that at no point in the Failure regress will it be the case that a solution entails no new problem
that must be solved in order to have any initial one solved.

To explain this further, let us distinguish between objective and subjective failures. S
subjectively fails when she follows a strategy to solve a certain problem yet she does not succeed
because the strategy is not something she manages to handle (but there need not be something
wrong with the strategy apart from that). In contrast, S objectively fails when she follows a
strategy solve a certain problem yet she does not succeed because the strategy is simply a bad
one (in this case, she might well be able to handle the strategy). Now, Failure arguments appeal
to the latter, objective kind of failures.

A second main reason to work with Failure arguments, rather than such cannot-arguments, is that
it is not clear how to obtain all-conclusions (i.e. corresponding to Failure I conclusions) such as
‘you cannot have all persons (rather than any person) guarded’ on the basis of cannot-arguments.

So far, I have explained the inferences of the two Failure Schemas in a semi-formal way. Next
we shall see, in a formal way, that all relevant inference rules are again valid according to
classical predicate logic such that the two schemas are valid by this logic. In both cases, (C) can
be obtained via (1)-(5).

2.2. Formal explanation

All the same preliminaries apply as in the Paradox case (i.e. the abbreviations of the rules, the
calculi used, etc.). Again, some of the predicates and premises will need some additional
explanation (particularly Failure I’s (6) and Failure II’s (7)). These explanations are provided
right after the formalisations.

Key
Kx: x is of type K
Tx: S has to carry out task T regarding x
Ax: S performs action A regarding x
FAIL-I: S fails to carry out T regarding all Ks

Example:
Kx: x is a dispute x:
S has to settle x
Ax: S invokes a proposition to settle x
FAIL-I: S fails to settle all disputes

Failure Schema I

(1) ∀x(Kx→Tx) PREM

(3) ∀x(Ax→∃y(Ky∧x≠y)) PREM
(7) ∃x(Kx∧Tx)35 PREM
(6) (∃x(Kx∧Tx)∧(∀x((Kx∧Tx)→∃y(Ky∧Ty∧x≠y))))→FAIL-I PREM
(2) ∀x((Kx∧Tx)→Ax) HYP →I
(i) Ka∧Ta HYP →I
(ii) (Ka∧Ta)→Aa 2; ∀E
(iii) Aa i, ii; →E
(iv) Aa→(∃yKy∧a≠y) 3; ∀E
(v) ∃yKy∧a≠y iii, iv; →E
(vi) Kb∧a≠b HYP →I
(vii) Kb vi; ∧E
(viii) Kb→Tb 1; ∀E
(ix) Tb vii, viii; →E
(x) Kb∧Tb∧a≠b vi, ix; ∧I
(xi) ∃y(Ky∧Ty∧a≠y) x; ∃I
(xii) (Kb∧a≠b)→∃y(Ky∧Ty∧a≠y) vi-xi; →I
(xiii) ∃y(Ky∧Ty∧a≠y) v, xii; ∃E
(xiv) (Ka∧Ta)→∃y(Ky∧Ty∧a≠y) i-xiii; →I
(5) ∀x((Kx∧Tx)→∃y(Ky∧Ty∧x≠y)) xiv; ∀I
(xv) ∃x(Kx∧Tx)∧(∀x((Kx∧Tx)→∃y(Ky∧Ty∧x≠y))) 7, 5; ∧I
(8) FAIL-I xv, 6; →E
(C) ∀x((Kx∧Tx)→Ax)→FAIL-I 2-8; →I

Additional key
Rxy: S first has to carry out T regarding y in order to carry out T regarding x
STx: S first has to carry out supertask ST in order to carry out T regarding x
FAIL-II: S fails to carry out T regarding any K
Example:
Rxy: S first has to settle y in order to settle x
STx: S first has to settle an infinity of disputes in order to settle x
FAIL-II: S fails to settle any dispute

Failure Schema II

(1) ∃x(Kx∧Tx) PREM

(3) ∀x(Ax→∃y(Ky∧Ty∧Rxy)) PREM
(7) ∀x((Kx∧Tx)→∃y(Ky∧Ty∧Rxy))→∀x((Kx∧Tx)→STx) PREM
(6) (∃x(Kx∧Tx)∧∀x((Kx∧Tx)→STx))→FAIL-II PREM
(2) ∀x((Kx∧Tx)→Ax) HYP →I (i) Ka∧Ta HYP →I
(ii) (Ka∧Ta)→Aa 2; ∀E
(iii) Aa i, ii; →E
(iv) Aa→∃y(Ky∧Ty∧Ray) 3; ∀E
(v) ∃y(Ky∧Ty∧Ray) iii, iv; →E
(vi) (Ka∧Ta)→∃y(Ky∧Ty∧Ray) i-v; →I
(vii) ∀x((Kx∧Tx)→∃y(Ky∧Ty∧Rxy)) vi; ∀I
(5) ∀x((Kx∧Tx)→STx) vii, 7; →E
(viii) ∃x(Kx∧Tx)∧∀x((Kx∧Tx)→STx) 1, 5; ∧I
(8) FAIL-II viii, 6; →E
(C) ∀x((Kx∧Tx)→Ax)→FAIL-II 2-8; →I

In all three schemas, I am operating with conjunctions of predicates (rather than a single one):
‘K’ and ‘F’ in the Paradox Schema, and ‘K’ and ‘T’ in the Failure Schemas. The reason for
doing so should be clear from Failure Schema I: conjunctions allow me to generate regresses
Failure I-wise (that is, on the basis of lines (3) and (1) of Failure I, while Failure II arguments
only employ (3) of Failure II, cf. Figure 1). This technique has not been used before in the
literature. Yet, without conjunctions it does not seem possible to capture Failure I arguments
(such as Tarski’s answer to the Liar Paradox, or McTaggart’s attack on the A-theory of time, as
we shall see later on).

Just like ‘IR’, ‘FAIL-I’ and ‘FAIL-II’ are placeholders for schematic sentences: ‘S fails to carry
out T regarding all Ks’ and ‘S fails to carry out T regarding any K’ respectively. Again, these are
no full first-order expressions. Yet, it is easy to remedy this by introducing a new predicate ‘C’
where ‘Cx’ stand for ‘S carries out T regarding x’. In terms of this, ‘FAIL-I’ and ‘FAIL-II’ can
be expressed simply as follows:

FAIL-I ¬∀x(Kx∧Cx)
FAIL-II ¬∃x(Kx∧Cx)

For example: It is not the case that, for all disputes x, S settles x vs. It is not the case that, for at
least one dispute x, S settles x. Importantly, this additional predicate ‘C’ cannot be fully
expressed in terms of the predicates ‘T’ and ‘A’. For example, ‘¬∃x(Kx∧Tx)’ (i.e. ¬(1) of
Failure Schema II, and ¬(7) of Schema I) means ‘It is not the case that for at least one K, S has to
carry T regarding that K’, and ‘¬∀x((Kx∧Tx)→Ax)’ (i.e. ¬(2) of Failure I and II) means ‘It is
not the case that for all Ks x, if S has to carry T regarding x, then S performs A regarding x’. As
we can see, these do not say what FAIL-I and FAIL-II say.36

Another important thing to note here is that ‘T’ (i.e. the task that S has to carry out) does not
carry modal or deontic connotations. At least, in my construction none of the inferences relies on
such considerations.37 For example, the inferences do not make use of the consideration that
oughtimplies-can (i.e. that if S has to carry out T, then S should be able to carry out T).
According to the Failure Theory, to press this point again, ‘S fails to carry out T regarding
any/all K(s)’ does not mean ‘S lacks a certain ability’, but rather ‘S never carries out T in the
sense that there is always a further task of the same sort to be accomplished in order to
accomplish T regarding any/all K(s)’.

One main difference between the two Failure Schemas lies in the predicates R and ST. R and ST
cannot be expressed purely in terms of the predicate T, given that R and ST induce an ordering
on tasks (i.e. something that the tasks themselves do not have). ST refers to one kind of
supertask, namely the kind where an infinite series of tasks is generated in terms of the
predicates T and R, i.e. on the basis of the following line in Schema II:

(vii) ∀x((Kx∧Tx)→∃y(Ky∧Ty∧Rxy))

which is the antecedent of:

(7) ∀x((Kx∧Tx)→∃y(Ky∧Ty∧Rxy))→∀x((Kx∧Tx)→STx) Again, I have taken this line as a

premise in order to ensure logical validity, and because in actual debates it is virtually always
taken for granted, i.e. those who wish to resist the charge of a regress do not usually attack (7),
but other premises (see §4.1.5). Nevertheless, a worry can be constructed similar to the one about
Paradox’s (7), namely that it has many false instances, given that its antecedent, i.e. (vii), is too
readily satisfied (consider e.g. ‘if for all persons that you have to count, there is a person that you
have to count and you have to count them at once, then for all persons that you have to count,
you have to carry out a supertask’). So the question is whether we can state in first-order terms
when (7) is true.

This time, we may safely assume that R is both asymmetric and transitive in all cases: if S first
has to settle d2 in order to settle d1, then it is not the case that S first has to settle d1 in order to
settle d2; and if S first has to settle d2 in order to settle d1 and S first has to settle d3 in order to
settle d2, then S first has to settle d3 in order to settle d1. These two properties suffice to block
potential loops of tasks and to obtain: S first has to settle an infinity of disputes in order to settle
d1, that is, S first has to carry out a supertask in order to settle d1. For the first-order formulation
of these two properties, see (a)-(b) in §2.2.2. By this, we are able to say when (7) is true: (7) is
true whenever (a)-(b) are true.

As we can see, Failure I’s premises (3) and (6) explicitly assume that x and y are distinct items.
First, (6)’s antecedent would automatically be satisfied without this assumption (which is
undesirable, because in that case FAIL-I would follow at once). Second, in this schema we have
no asymmetric or irreflexive relation R that can assure that x≠y.
However, there still remains a problem about (6), as its antecedent does not say what it should
say. It should say that there is always a new task of the same kind to be carried out, while in fact
it merely says that for each task, there is a distinct task of the same kind to be carried out (and in
the latter case two tasks of the same kind would be sufficient). To solve this, we could employ
the ‘<’-solution from §2.2.2. If this additional relation ‘<’ satisfies the properties (f)-(i), then it
makes sure that all Ks introduced in the regress are new items. Importantly, this problem does
again not affect the logical validity of the schema: even if (6) would not really say what we want
to express with it, it does make the schema logically valid.38

3. Boundaries

Why do the Failure Schemas not have more or fewer lines? In this case, the question is more
specific: Why start with problems, and end with failures? As in case of the Paradox Schema,
there are two considerations.

First, the Failure Schemas had to consist of at least (1)-(5), because otherwise it would not fulfil
the first three desiderata from §2.1. To recall again these three: the schema had to be such that
the regress occupies one of the lines, i.e. (4); that it is derived from premises/hypotheses, i.e. (1)-
(3); and that from it some conclusion is drawn, i.e. (5).

Still, the question is why not stop at line (5), where it is concluded that there is always another
problem to be solved, i.e. another K that you have to φ? The basic answer here is that line (5)
does not itself state that the solution under consideration has failed. This is brought out only by
the further step to (C): You will never φ any/all K(s) if you carry out the considered solution.

Clearly this answer is not as strong as in case of the Paradox Schema (where the latter borrows
the boundaries of familiar Reductio Ad Absurdum arguments), yet still it corresponds to the
intuition behind this schema: the Failure intuition.

If this is right, then the Failure Schemas should not have fewer lines. But the question remains:
should it have more lines? Specifically, the issue is why no extra steps need to be added, such as
the following two:

(i) If the solution never solves the given problem, and if it is also shown that there is an
alternative regress-free solution that does solve the problem, then this favours the alternative
solution.

(ii) If the solution never solves the given problem, and if it is also shown that no alternative
solution is possible, then it can be concluded that the problem cannot be solved.

Let me first explain (i). If the argument demonstrates that you will never solve the given problem
if you carry out solution such and such, then from this you may draw the conclusion that you
have to find and favour an alternative, better solution to the given problem. In terms of the
schemas: ψ- ing all Ks that you have to φ is a bad solution to φ-ing all/any K(s), and you have to
find another solution to φ-ing all/any K(s). For example: hiring a guardian for all persons that
you should have guarded is a bad solution to have anyone guarded, and you should find another
solution to have your girlfriend (or anyone else) guarded.

To further explain (i), let us again invoke the distinction between negative and positive regress
arguments. According to Sanford, the distinction is this:

• A regress argument is negative if its conclusion is that “a certain philosophical account,

definition, theory, or explanation will not do because it leads to an infinite regress.” (1984: 100)

• A regress argument is positive if its conclusion is that “something of a special sort must exist. If
something of this special sort did not exist, there would be an infinite regress.” (ibid)

I would like to generalise these claims as follows:39

• A regress argument is negative if its conclusion is that a certain solution will not do because it
leads to a regress.

• A regress argument is positive if its conclusion is that an alternative solution will do because it
does not lead to a regress.

In terms of the Failure Schema, negative regress arguments are arguments that stop at (C), and
positive regress arguments are arguments that expand the argument along the lines of (i). In the
former case it is merely concluded that a considered solution will never solve the given problem,
and in the latter it is added that this favours an alternative solution that does solve the problem. A
selected group of passages hint at such expanded arguments. Consider for example: 39 To be
sure, the second citation can also be read in terms of the Paradox Schema, i.e. that there is at
least one item to which the regress formula does not apply. See §2.2.3 above.

[…] and such a regress is plainly vicious. Likeness at least, therefore, must be admitted as a
universal, and, having admitted one universal, we have no longer any reason to reject others.
(Russell 1911-12: 9)

The links are united by a link, and this bond of union is a link which also has two ends; and these
require each a fresh link to connect them with the old. The problem is to find how the relation
can stand to its qualities; and this problem is insoluble. (Bradley 1893: 28)

Russell’s text is an example of the (i) expansion: he concludes against Resemblance Nominalism
and in favour of Realism about Universals in order to solve the problem of how it is possible that
distinct items can have the same property/relation (i.e. the Problem of Universals). Bradley’s text
is an example of (ii): there is, according to Bradley, no non-regressive solution that can explain
how relations can form a unity with their relata (the Unity Problem, as it is sometimes called),
and so the problem must be left unsolved. Another famous example of (ii) is the following:

The one based on infinite regress is that in which we say that what is offered as support for
believing a given proposition is itself in need of such support, and that support is in need of other
support, and so on ad infinitum, so that, since we have no place from which to begin to establish
anything, suspension of judgment follows. (Sextus, Outlines, 1.166-7)

Sextus argues that no solution will ever solve the problem of how our beliefs are justified or
supported, and so the problem must be left unsolved. Moreover, if no belief is ever justified to
anyone, or so Sextus and the ancient sceptics concluded, we must suspend them (more about
scepticism in §4.1.4).

More possible expansions are conceivable:

(iii) If the solution never solves the given problem, and it is also shown that the solution serves
no other purposes next to this, then that solution is never to be used.

(iv) If the solution never solves the given problem, and if theory T is committed to this solution,
then T entails that the problem is never solved, and this then forms a reason to disfavour T.

I have found no philosophical examples of (iii), but it terms of the two practical illustrations
cited above, the idea is this. If drinking beer and hiring guardians were useless for whatever
purpose, then the idea would be that those actions are just never to be performed.

Expansion (iv) is more common in the literature. For example, in Russell’s case just cited the
theory which is committed to the regress is Resemblance Nominalism. So, by the regress
argument, Resemblance Nominalism entails that the Problem of Universals is never solved (be it
a general or particular problem), and so this constitutes a reason to disfavour Resemblance
Nominalism. I will return to (iv) when I address the issue of viciousness (§4.2.2), i.e. the issue of
when regresses are bad for theories.

The moral for the question of boundaries is as follows. As we just saw, the course of regress
arguments (i.e. those which are instances of one of the Failure Schemas) may differ after (C),
and so it is better to stop at what they have in common, namely (C). To be sure, if it can be
shown that all instances of the Failure Schemas can be expanded in the four ways sketched
above, then in that respect the Failure Schemas have four varieties and the expansions may be
added as optional lines to the schema (or rather, as a disjunction of options).

4. Literature

As noted, the Failure Schemas are developed only in this dissertation. In my view, this is
substantial progress for two reasons. First, several philosophers have thought that the task could
not be done, and have expressed various worries about Failure-like ideas. Yet, as I shall argue
later in §6.1, in my view these worries can be countered. Moreover, they can be countered on the
basis of the two Failure Schemas just set out. Second, as I will show below, the Failure intuition
underlying these schemas, namely the Failure Intuition, is far from uncommon and has been
expressed several times in the literature.40 Now, the Failure Schemas demonstrate that this
intuition can be made precise.
To begin with, if we ignore some general remarks by Russell (1903: §329) to be discussed later
in §6.1.1, then it can be said that Passmore (1961: ch. 2) initiated the metaphilosophical debate
on regress arguments. Passmore’s proposal of what to expect from regress arguments is very
close to the Failure take. According to him, regresses

demonstrate only that a supposed way of explaining something or ‘making it intelligible’ in fact
fails to explain. (1961: 33)

As ‘you have to explain x’ and ‘you have to make x intelligible’ are examples of tasks,
explaining and making things intelligible might be taken as cases of problem solving. As a
consequence, Passmore’s claim may be read more generally (which is exactly the Failure
rationale): regresses demonstrate that a considered solution in fact fails to solve a given problem.
I will have much more to say on Passmore later on (in §6.1.2).

As a second source, the distinction between the Paradox and the Failure Schema may be traced
back to Day’s distinction between product vs. process regress arguments (1986: 51-2).
According to Day:

• Product regress generating arguments demonstrate that a given set of premises entails an
infinity of items.

• Process regress generating arguments demonstrate that a given procedure (analysis,

explanation, definition, etc.) can be iterated endlessly.

So, Day’s product arguments correspond to the first two steps of the Paradox Schema (i.e. the
entailment of the regress and the specific result of an infinity of items), and his process
arguments correspond to the first step of the Failure Schema (i.e. the entailment of a regress of
similar solutions). He does not make the distinction explicit, however, in terms of argument
schemas.

As a third source, I would mention Gratton’s problem and response regresses (1994b, 1997: 216-
7, 2010: ch. 6). Note that he uses the term ‘responses’ rather than ‘solutions’. His reason is as
follows:

It is odd to speak of solutions because it suggests success in solving a problem, but if the same
type of problem keeps recurring, then there does not seem to be a genuine solution. (2010: 159,
cf. 1994b: 314)

So Gratton makes the following assumption:

• If a solution entails a problem which is of the same type as the problem it is meant to solve,
then it is no genuine solution.

This may look plausible, yet all regresses generated in Failure Schema I are counterexamples to
it. Take for example the particular problem that I have to get rid of a bulge in the carpet. Pressing
it down will solve this problem even if it generates a similar kind of problem, namely that
another particular bulge appears elsewhere in the carpet. In this case, of course, we have to
assume that we can clearly discriminate among bulges. Yet, as I may well have chosen a
philosophical example to show where this holds, we can safely assume it for purposes of
illustration. Philosophical counterexamples will be provided later on (in §4.1.3). A

nother way of pressing this point is to ask: No genuine solution to what? What exactly is the
problem that is to be solved? Do you want to get rid of all the bulges in the carpet? Or only one
particular bulge (in an especially unfortunate place, say)?

Also, even if the assumption above were to hold, it seems fine to speak of solutions, for bad
solutions are solutions after all.

At any rate, Gratton’s position is that most regress arguments are not of this variety. Compare:

I have not examined these infinite regress arguments in this paper because I have found only a
few, and my goal is to identify and clarify the characteristics common to most infinite regress
arguments. (1997: 222, n. 35, cf. 2010: 159)

At this point, we do not yet know whether this is right or wrong. Basically the question is which
schema has the most regress arguments from the literature as an instance. We shall see about this
soon in §3. It might be useful to know that even though Gratton’s problem and response regress
arguments resemble instances of the Failure Schema, there are also some crucial differences,
which I shall identify in §6.1.3.

Lastly, let me point out that the distinction between Failure I and II arguments, which is quite
crucial given that one kind is much stronger than the other, has gone wholly unnoticed in the
literature. Yet, attempts to capture the reasoning of either can be found in various places. I have
selected two texts each for Failure I and II. For Failure I:

A philosophical explanation of predication must, if it is to be successful, explain all instances of

predication. The theory of forms fails to do this. […] The argument purports to show that no
matter how often you iterate the explanation in order to include the predication just introduced,
you will always introduce a new, unexplained predication. (Day 1987: 156-7)

Thus, if the regress is vicious, it is vicious because it prevents Resemblance Nominalism from
accomplishing its explanatory project of accounting for all properties in terms of resembling
particulars: such a project remains forever incomplete. (Rodriguez-Pereyra 2002: 108)

These two texts appeal to the reasoning of Failure I because they mention a universally
quantified problem (which is marked). It is also worth pointing out that Day and Rodriguez-
Pereyra use variants of ‘never’ (marked as well). The next two texts appeal to the reasoning of
Failure II because they mention an existentially quantified problem (marked).

If, to establish the truth of any proposition, A, one must establish, as a necessary condition, the
truth of an infinite number of propositions B…n, the regress is vicious, because one is, as a
consequence, logically incapable of ever establishing the truth of A. (Johnson 1978: 80)
A regress is said to be vicious if, for example, in order to have something, there is always an
additional something one is first required to have. In general, in a vicious regress, one could
never be in a position to have anything at all, or the requirements for having the first or any
additional thing could never be met. (Ruben 1990: 127)

All in all, the Failure Schemas, too, have their traces in the literature.

Looking ahead

This closes the key part of this dissertation. In the remainder of this dissertation, I shall proceed
as follows. Next, in §3, I will present classic instances of the argument schemas presented in this
part. Then, §4 will provide further theoretical details of the schemas. In §5, the schemas will be
put to work: I shall show how they can be used to clarify existing controversies about particular
regress arguments. In §6, I will defend my Failure Theory from objections levelled in the
literature. Finally, in §7, I shall briefly summarise the two theories of regress arguments
developed in this dissertation.

§3 Instances

In the following I present philosophical instances of the schemas from the previous part. First, I
explicate my approach, i.e. say what rules my reconstructions shall follow. Second, I provide
filling instructions for the argument schemas and a number of full instances (from Sextus,
Carroll, Russell, Wittgenstein, and Ryle).

3.1. Argument Reconstruction

At this point I regard as finished the work set by Charity Rule I. That is, in the foregoing I have
set out two ways in which the missing steps in actual regress arguments can be supplied such that
logically valid arguments are obtained. Importantly, this does not mean that the work set by
Charity Rule II is finished as well, and I will leave the latter unaddressed until §5. This also
means that all I say below about argument reconstruction does not yet touch upon the issue of
argument evaluation (except at one point, namely the tentative choice between Failure I and
Failure II, on which I will say more below).

In this part I will use the argument schemas to reconstruct a number of case studies. I will not
each time apply Charity Rule I, but rather immediately check whether the argument can be
reconstructed as an instance of the schemas. In other words, I will start from the schemas and try
to fill them out. To reconstruct, then, is to fill out an argument schema, and the latter is to fill out
all its schematic letters. Let me specify this for the Paradox and Failure Schemas respectively.

Paradox Reconstruction

Steps Instructions
1 Identify the regress formulas and trigger.
2 Draw a regress (the first five steps, say).
3 Identify what follows from the regress.
4 Determine whether these results are unacceptable.
5 Identify what is eventually to be rejected.

At each step choices are demanded of the reconstructor. For example, in the last step it must be
decided what to reject, i.e. which of the regress formulas (or perhaps the trigger) has to go. In my
reconstructions in §3.4, below, I simply follow the position of the relevant philosopher (if
available). A full reconstruction, however, would have to incorporate further argumentation to
justify this choice. That choices are involved means that the reconstruction procedure as just
specified does not determine a unique reconstruction of an argument. Still, in most cases this
does not affect the conclusion, as the conclusion is to be a rejection of the hypothesis identified
in the first step.

On first sight, the tasks for filling out a Failure Schema appear even simpler:

Failure Reconstruction

Steps Instructions
1 Identify the problem, considered solution, and extra premise.
2 Draw a regress (the first five steps, say).
3 Draw the conclusion that the solution fails.

However, this is a bit more complicated than it seems, because it also needs to be checked
whether a given case should be taken as an instance of Failure I or II. It is important to point out
already that, logically speaking, any case can be stated in terms of both Failure Schema I and II.
However, whether a given case should be stated in specifically one of or the other must be
decided on the basis of the plausibility of one specific line. Namely:

• Does premise (3) of Failure Schema II apply? If Yes, then do it Failure II-wise. If No, do it
Failure I-wise.

The idea is as follows. If the instance of ‘For any K x, if you ψ x, then there is a new K y, and
you first have to φ y in order to φ x’ looks plausible (or at least worth considering), then the
given case is to be spelled out in terms of Failure Schema II. If the instance is implausible, then
the given case is to be spelled out in terms of Schema I.

Basically, the motivation of this criterion is that Schema II yields stronger, and so more
interesting, conclusions (again: its conclusion entails its I-counterpart, but not vice versa), and so
by Charity Rule II any case should be stated in II-format wherever plausible. I will say more on
this point later on (see §6.2), and will here provide just two examples: Sextus’ regress argument
for the claim that a natural way of justifying propositions fails, and Tarski’s regressive solution
to the Liar Paradox.

First, Sextus’ case is taken as an instance of Failure Schema II, because the following instance of
line (3) of Failure II looks plausible:
- For any propositions x and y, if you provide a reason y for x, then you first have to justify y in
order to justify x.

Suppose you have to justify the proposition that Socrates corrupts the youth. In Sextus’ case,
justification is explicitly dialectical: to have a justification for a proposition is to justify it and to
have resolved the disagreements about it.41 So in this case you have to resolve the disagreements
about the proposition that Socrates corrupts the youth in order to justify it. Now suppose that you
provide the following reason: the youth no longer respect authority. Surely you cannot resolve
any disagreement on the basis of this proposition if your opponents do not accept it. And so you
justify the initial proposition only if you first justify its reason (which is what (3) of Failure II
says).

If this is right, Sextus’ case is to be taken as a II-case which has the conclusion that you will
never justify any proposition. One could state it in I-format as well, though this would be
uninteresting given that the IIconclusion entails the I-conclusion: If you will never justify any
proposition, then it is trivial that you will never justify all propositions.

Second example: Tarski’s case is taken as an instance of Failure I, because the following does
not look plausible:

- For any language L, if you deny that L is semantically closed, then you first have to resolve the
Liar Paradox in L’s metalanguage M in order to resolve the Liar Paradox in L.

Put simply, if a language is semantically closed, then it is possible to state whether or not
sentences are true within that very same language. English, for example, is such a language, as I
can say in English that the English sentence ‘Socrates is mortal’ is true (or not). Furthermore,
this feature is exactly one of the key elements generating the Liar Paradox. Namely, it is possible
to state the following sentence in English:

This sentence is false.

The Liar Paradox is as follows: If the above sentence is true, then what it says is the case: it is
false (i.e. as well as true). If the sentence is false, then it says what is the case and so it is true
(i.e. as well as false). To block the Liar Paradox, consequently, one could deny that the given
language is semantically closed (i.e. deny that it can speak of its own sentences as to whether or
not they are true), and introduce a metalanguage to be able to express this. However, if there are
worries about the initial language arising from the Liar Paradox, then the same worries should
arise for the metalanguage, which invites a regress.

Now the issue is whether this is a case of Failure I or II. If the above instance of (3) of Failure II
is plausible, it is a II case. Yet this does not seem plausible: Why should you resolve the Liar
Paradox in M in order to solve it in L? As soon as you deny that L is semantically closed the Liar
Paradox is blocked within L, no matter whether M is paradoxical or not. If this is right, then
Tarski’s case must be taken as an instance of Failure I which has the conclusion that you will
never resolve all, rather than any, Liar Paradoxes (even if you resolve many of them).
As we shall see in §3.3, thanks to the criterion that we should look for Failure II-arguments
whenever they seem plausible, Failure Schema II will have many more instances than Failure
Schema I. 42 Still, it is important to keep Failure Schema I apart, as certain cases are only
plausible in this format. Tarski’s case is a good example. Other familiar cases which happen to
be Failure I cases include Plato’s Third Man, Russell’s regress argument for the irreducibility of
asymmetric relations, and McTaggart’s regress argument against the so-called A-theory of time
(to be explained later on).

To be sure: my upcoming collection of instances is incomplete. I have found other regress

arguments, and am sure there are (and surely can be) many more. Still, my collection, which
builds on Gratton (2010), is more extensive than anything in the literature. I think it is important
to consider many cases at once as they form the data (so to speak) of the theories about regress
arguments. Any general theory about regress arguments would be irrelevant if there were not a
great variety of specific cases.

Sometimes I have included a case where the initial text does not speak of ‘regress’ (e.g. the cases
by Plato and Wittgenstein). In that case, I have followed the secondary literature in assuming the
presence a regress argument (and perhaps such moves can be justified by Charity as well). I have
also included several odd cases, such as Plato’s aviary case. It does not matter that no one
believes that knowledge can be compared with birds in aviaries. My focus is on the general line
of reasoning.

Importantly, and related to the last point, I do not endorse any of the regress arguments in this
dissertation (nor for that matter reject them). Hopefully this will be clear later on when I identify
possible strategies to resist these arguments (§4.1.5), and especially when I turn to the topic of
argument evaluation (§6.2). For reasons of space, I will not present all instances in full, but
restrict myself to five classic cases, and merely identify, in the other cases, relevant instances of
the schematic letters.

3.2. Paradox Filling Instructions

Consider the first line of the Paradox Schema:

• For all Ks x, x is F only if such and such (e.g. there is a new Kitem y and y is G or x and y
stand in R).

To obtain instances of such a line (and of the others), ‘items of type K’ is to be replaced with a
specific domain, and the capitals ‘F’, ‘G’, ‘R’ with predicates which express properties/relations
of the items in that domain. In the following I provide a table with examples (references to the
main sources are included). The cases are ordered historically, rather than thematically.

items of type K x is F x is G/ y is G/ x and y stand in R main sources

distances to Achilles Achilles runs x/ Zeno of Elea,
the Tortoise traverses x the half of x Aristotle, Physics, 239b,
cf. Huggett 2008
sets of the members of the members of x Plato, Parmenides,
large things x are large participate in the form
Largeness Plato, Parmenides,
132a-b, cf. Vlastos
1954, Rickless 2007
pieces of one is ignorant one does not know Plato, Theaetetus,
knowledge about x that one has x in 200b-c
one’s soul
actions x is good x is performed for Aristotle,
the sake of y Nicomachean Ethics,
1094a, Metaphysics,
994a
principles x is demonstrated x is proved by y Aristotle,
Metaphysics, 1006a,
cf. Johnstone 1994
beliefs x is justified one has a reason y for x Sextus Empiricus,
to someone Outlines, 1.166-7, cf.
Klein 1999, 2007
disputes x is decided there is an agreed upon Sextus Empiricus,
criterion to decide x Outlines, 2.18-20, cf.
Chisholm 1982,
Amico 1993
things x moves/is caused x is set in motion/is Aquinas, Summa, I,
caused by y v.2, §3, cf.
Reichenbach 2004
ideas x is indubitable x is clearly and Descartes,
distinctively perceived Meditations, cf. Van
Cleve 1979, Newman
& Nelson 1999
sovereign powers x is subjected to there is a judge who Hobbes, Leviathan,
legal limits monitors whether x Part 2, ch. 29, cf.
transgresses the laws Hampton 1986
anything x exists (analogy: x designed by God Hume, Dialogues, §4,
x is supported) (analogy: x is supported cf. Locke, Essay,
by a Tortoise) Book 2, ch. 23, §2,
Russell 1927
ideas x is inductively x is derived from past Hume, Enquiry, §4,
justified facts and a Uniformity Treatise, Book 1, ch.
principle 3, cf. Norton 2013
mental states one is conscious of x x taken as an object by a Brentano 1874: Book
higher-order mental state y 2, §2, cf. Textor 2006,
Zahavi 2006
relations x and y stand in R R is related to x and y by R* Bradley 1893: chs. 2-
3, cf. Vallicella 2002,
Maurin 2012
sets of premises a conclusion follows x contains the additional Lewis Carroll 1895,
 from x premise ‘if the members of cf. Thomson 1960, x
are true, then the conclusion Smiley 1995
must be true’
asymmetric relations relata a and b a and b have properties Russell 1903: §214,
stand in x corresponding to x 1906-07, 1959
propositions x is of the form x states an identity of Russell 1910-11
‘[proper name] is denotation
[description]’
A-series x is noncontradictory the members of x are past, McTaggart 1908,
present and future at 1927: ch. 33, cf.
different times Dummett 1960,
Rankin 1981
propositions x is true x corresponds with reality Frege 1918-19, cf.
Dummett 1973,
Künne 2003: §3.3.2
languages x is free from the no sentence in x speaks of Tarski 1944, cf. Beall
liar paradox its own truth 2007
actions x is performed the agent employs Ryle 1945, 1949: ch.
intelligently knowledge that x is to be 2, cf. Stanley &
performed in such and Williamson 2001
such a way
actions x is performed the agent acts on a volition Ryle 1949: ch. 3, cf.
voluntarily to do x Locke, Essay, Book 2,
ch. 21, §25
linguistic or the meaning of x is there is a method of Wittgenstein 1953:
mental items fixed projection between a mental image of x §141, cf. Stokhof
and that to which x applies 2000: §2.6
rules one follows x rather one specifies x mentally or Wittgenstein 1953:
than another rule y linguistically §§185-6, 201, cf.
Kripke 1982: ch. 2,
Varzi 2008
heterological x is resolved sentences which generate x Mackie & Smart
paradoxes are meaningless 1953, cf. Grelling &
Nelson 1908
words the reference of x the reference of x is fixed by Quine 1968, cf.
is fixed a background language Davidson 1979
expectations person A has x about B has an expectation Lewis 1969, cf.
person B’s behaviour about A’s x Binmore 1987
persons x is wholly mistaken x is being deceived by Lehrer 1971, cf.
an evil demon Johnson 1978
sets of items members of x have the members of x instantiate Armstrong 1974,
the same property/ the same universal/ belong 1978
relation to the same class/ etc.
languages one is able to learn x one possesses a language Fodor 1975, cf.
of thought to learn x Laurence & Margolis
1997
actions one performs x one performs all actions Danto 1979, Sneddon
by which x is performed 2001
rankings of x is set there is majoritarian MacKay 1980, cf.
candidates aggregation Hardin 1980
device which sets x
visual perceptions one has x of oneself one sees oneself looking at Ratford 1983,
in the mirror x in the mirror Haldane 1983, Garrett
1983
experimental results x is the correct x is confirmed by an Collins 1985
outcome of an experiment
experiment
regularities x is explained there is a law of nature Van Fraassen 1989
which necessitates x
decisions x is optimal a costly algorithm is used Smith 1991, Mongin
to make x 2000: §3
contexts x is recognised the computer is programmed Dreyfus 1992
by a computer in such a way that it selects
the relevant features of x
actions x is morally x is a function of one’s Strawson 1994
responsible mental make-up
obligations one is obliged to one can know that one Sorensen 1995,
do x is obliged to do x Sider 1995
sets of attitudes one has to perform x contains the extra belief Blackburn 1995,
an action A given that one has to perform A Schueler 1995
one’s x given x
sets of attitudes one must intend to one has a pro-attitude Railton 1997, Dreier
perform A given towards the instrumental 2001, Brunero 2005
one’s x rule that one must intend to
perform A given x (i.e. one
believes that A is a means to
one’s intended ends)
obligations one has x to perform A is what one regards as the Lazar 1999, cf.
an action A best option, all things Davidson 1970
considered
omnipotence x is resolved omnipotent beings can gain Zamir 1999
paradoxes one ability by
losing another
actions one is frequently one is frequently free to Vander Laan 2001
free to do x generate a desire to refrain
from performing x
sets of circumstances x entails a x contains the extra Maitzen & Wilson
Newcomb’s problem circumstance that the 2003
Predictor has made a
prediction of how many
boxes you will take
physical causes a mental cause there is another physical Kim 2003
 excludes x as a cause cause of E
of physical event E
strategies persons have they send each other Floridi 2004
common knowledge messages about x
about x
epistemic rules one is entitled to x x is inferentially justified Boghossian 2001,
2006: chs. 5-7, cf.
Wright 2001, Philie
2007
seemingly absolute x obtains x obtains according to a Boghossian 2006: ch.
facts certain theory 4
dispositional x’s identity is x’s identity is fixed by its Bird 2007
properties determinate relation to other properties
thoughts one thinks x one wants that x is in Boghossian 2008,
accordance with rules Glüer & Wikforss
2009
actions one is culpable for x one does x from ignorance Rosen 2004; Peels
2011
obligations one has x to perform one’s evidence that one Kiesewetter 2011
an action ought to seek supports x
external or one might be one is sceptical about x Wilson 2012
mental states deceived about x

3.3. Failure Filling Instructions

Consider the second line of the Failure Schemas:

• For all Ks x, if you have to φ x, you ψ x.

To obtain instances of such a line (and of the others), ‘items of type K’ is to be replaced with a
specific domain, and the Greek letters ‘φ’ and ‘ψ’ with predicates which express actions
involving the items in that domain. Again, in the following I provide a table with examples. The
references remain the same as above. This time, I have used the fourth column to indicate
whether the argument (presumably) takes the form of Failure Schema I or II (i.e. ‘F-I’ or ‘F-II’).

items of type K φx ψx schema

distances to the traverse x run x F-II
Tortoise
sets of large things explain why members appeal to the fact that all F-I
of x are all large members of x partake in the form
Largeness
pieces of knowledge explain how one is appeal to the fact that one F-II
ignorant about x does not know that one has
x in one’s soul
actions explain the purpose of x identify the sake for F-II
which x is done
principles demonstrate x provide a proof of x F-I
beliefs justify x provide a reason for x F-II
disputes decide x employ an agreed-upon F-II
criterion to decide x
things explain why x moves appeal to something else F-II
that sets x in motion
ideas demonstrate that x is appeal to the fact that x is F-II
indubitable clearly and distinctively
perceived
sovereign powers subject x to legal limits install a judge who monitors F-I
whether x transgresses the
laws
anything explain why x exists appeal to the fact that x is F-II
(why x is supported) designed by God (that x is
supported by a Tortoise)
ideas justify x inductively appeal to past facts and a F-II
Uniformity Principle to
derive x
mental states explain how one is appeal to the fact that x is F-II
conscious of x taken as an object by a
higher-order mental state y
relations explain how x is unified appeal to the fact that x is F-II
with its relata unified with its relata by
another relation
sets of premises demonstrate that a introduce an additional F-II
conclusion follows logically premise ‘if the members of x
from x are true, then the conclusion
must be true’ to the argument
asymmetric relations reduce x reduce x to properties of x’s F-I
relata
propositions of the analyse x appeal to the fact that x states F-I
form ‘[proper name] an identity of denotation
is [description]’
A-series eliminate the hold that the members of x F-I
contradictions in x are past, present and future at
different times
propositions decide whether x is true decide whether x F-II
corresponds with reality
languages resolve the Liar Paradox in x hold that no sentence in x can F-I
speak of its own truth
actions perform x intelligently employ knowledge that x is F-II
to be performed in such and
such a way
actions perform x voluntarily act on a volition to do x F-II
linguistic or explain what fixes the appeal to the fact that x F-II
mental items meaning of x is fixed by a method of
 projection
rules explain why x is followed invoke a specification of x F-II
rather than another rule
heterological resolve x decide that sentences which F-I
paradoxes generate x are meaningless
words fix the reference of x appeal to a background F-II
language of x
expectations have x about your use information about your F-II
opponent’s behaviour opponent’s expectations
about your having x
persons explain how it could be appeal to an evil demon F-I
that x is wholly mistaken which could deceive x
sets of items explain how it is possible appeal to the fact that the F-I
that members of x have members of x instantiate the
the same property/relation same universal/ belong to
the same class/ etc.
languages explain how we can learn x appeal to a language of F-II
thought by which we can learn x
actions perform x perform all actions by F-II
which x is done
rankings of settle x use a majoritarian F-II
candidates aggregation device to set x
visual perceptions explain how you have x appeal to the fact that you F-I
of yourself in the mirror see yourself having x
experimental results check whether x is the carry out an experiment to F-II
correct outcome of an confirm x
experiment
regularities explain x appeal to the fact that there F-II
is a law of nature which
necessitates x
decisions make x optimally use a costly F-II
algorithm to make x
contexts recognise x programme x in such a way F-II
that it selects the relevant
features of x
actions be morally act on the basis of your F-II
responsible for x mental make-up that
favours x
obligations secure x against possible appeal to the obligation not F-II
abuse of the Access principle to make it impossible to
know one’s obligations
sets of attitudes explain why one should appeal to the extra belief that F-II
perform action A given one has to perform A given x
one’s x
sets of attitudes demonstrate that one must appeal to one’s pro-attitude F-II
intend to perform A given towards the instrumental rule
 one’s x that one must intend to
perform A given x (i.e. one
believes that A is a means to
one’s intended ends)
obligations demonstrate that one has x appeal to the rule that one F-II
to perform an action A should perform A is that is
what one regards as the best
option, all things considered
omnipotence resolve x resolve x by appealing to the F-I
paradoxes fact that omnipotent beings
can have one ability by losing
another
actions explain how one can appeal to the fact that one is F-II
frequently be free to do x frequently free to generate a
desire to refrain from
performing x
sets of circumstances demonstrate that x entails a add to x the circumstance F-II
Newcomb’s problem that the Predictor has made a
prediction of how many
boxes you will take
physical causes show that a physical event appeal to the fact that the F-I
has a mental cause rather mental cause excludes x
than x
strategies obtain common knowledge send one another message F-II
about x about x
epistemic rules be entitled to x justify x by using further F-II
epistemic rules
seemingly absolute explain away x appeal to the fact that x is F-I
facts relative to a certain theory
dispositional fix the identity of x appeal other dispositional F-II
properties properties to fix the
identity of x
thoughts think x want that x is in accordance F-II
with rules
actions explain that one is culpable show that x was performed F-II
for x from ignorance
obligations show that someone has x appeal to the fact that one’s F-II
evidence that one ought to
seek supports x
external or be sceptical about x invoke the fact that one F-I
mental states might be deceived about x

3.4. Selected Full Instances

In the following I present some full instances of the schemas. Specifically, I will reconstruct
cases by Sextus, Carroll, Russell, Wittgenstein and Ryle. I have selected these five cases because
they are classics and have inspired many other regress arguments. Sextus’ case is still a central
concern in epistemology; Carroll’s case proved inspiring not only in philosophy of logic but also
in ethics and debates of practical reasoning; Russell’s case was a major event in metaphysics
which established the category of relations (i.e. as distinct from properties); and Wittgenstein’s
and Ryle’s cases still today inspire new regress arguments in philosophy of language and mind.

For each case I offer a brief introduction, then fill out all the lines of the Paradox Schema and
one of the Failure Schemas such that we get complete arguments with a conclusion. I am well
aware that in each case longer discussions of the arguments are desirable. At this point, however,
the plain instances of the schemas should suffice. I shall consider debates which centre on the
cases by Sextus, Carroll and Russell in some more detail later on in §5.

1. Sextus Empiricus

Consider any of your beliefs. Are you justified in holding it? Arguably, you are justified only if
you have a reason for it. But are you justified in holding that reason? By parity of reasoning, you
are justified in holding it only if you have a reason for it as well (i.e. a reason for the reason of
your initial belief). Regress. According to the ancient sceptics, the regress demonstrates that we
will never be able to show that our beliefs are justified and that we would do better to suspend
them:

The one based on infinite regress is that in which we say that what is offered as support
for believing a given proposition is itself in need of such support, and that support is in
need of other support, and so on ad infinitum, so that, since we have no place from which
to begin to establish anything, suspension of judgement follows. (Outlines 1.166-7)

Paradox instance

(1) For any proposition x, x is justified to S only if S has a reason y for x.

(2) For any propositions x and y, S has a reason y for x only if y is justified to S.
(3) At least one proposition is justified to S.
(4) Regress:
(a) p1 is justified to S.
(b) S has a reason p2 for p1. [a, 1]
(c) p2 is justified to S. [b, 2]
(d) S has a reason p3 for p2. [c, 1]
(e) p3 is justified to S. [d, 2]
And so on.
(5) S has an infinity of reasons. [3, 4]
(6) S does not have an infinity of reasons.
(C) (3) is false: No proposition is justified to S. [1-6]

Failure II instance

(1) For at least one proposition x, you have to justify x.

(2) For any proposition x, if you have to justify x, then you provide a reason for x.
(3) For any propositions x and y, if you provide a reason y for x, then you first have to justify x
in order to justify y.
(4) Regress:
(a) You have to justify p1.
(b) You provide a reason p2 for p1. [a, 2]
(c) You have to justify p2 first. [b, 3]
(d) You provide a reason p3 for p2. [c, 2]
(e) You have to justify p3 first. [d, 3]
And so on.
(5) For any proposition x, you first have to justify a regress of propositions in order to justify x.
[4]
(C) You will never justify any proposition if you provide a reason every time you have to justify
a proposition. [1-5]

2. Lewis Carroll

Consider this simple argument:

(A) You are either here or somewhere else.

(B) You are not here.
(Z) Hence: You are somewhere else.

As Carroll’s dialogue between Achilles and the Tortoise goes, the Tortoise is willing to accept
(A) and (B), but not (Z) just because she denies that (Z) must be accepted if (A) and (B) are. So,
to demonstrate that (Z) follows from (A) and (B), Achilles adds an extra premise to the
argument:

(C) If (A) and (B) are true, (Z) must be true.

Still, the Tortoise is unsatisfied. This time she is willing to accept (A), (B) and (C), but not (Z)
just because she denies that (Z) must be accepted if (A), (B) and (C) are. So to demonstrate that
(Z) follows from (A), (B) and (C), Achilles adds yet another premise (D) to the argument:

“You should call it (D), not (Z),” said Achilles. “It comes next to the other three. If you
accept (A) and (B) and (C), you must accept (Z).”
“And why must I?”
“Because it follows logically from them. If (A) and (B) and (C) are true, (Z) must be true.
You don’t dispute that, I imagine?”
“If (A) and (B) and (C) are true, (Z) must be true,” the Tortoise thoughtfully repeated.
“That’s another Hypothetical, isn’t it? And, if I failed to see its truth, I might accept (A)
and (B) and (C), and still not accept (Z), mightn’t I?” (Carroll 1895: 279)

Paradox instance

(1) For any set of premises x, a conclusion (Z) follows logically from x only if there is a bigger
set of premises also containing the premise ‘if the member of x are true, then (Z) must be true’.
(2) For any set of premises x, x contains ‘if the member of x are true, then (Z) must be true’ only
if (Z) follows logically from x.
(3) A conclusion follows logically from at least one set of premises.
(4) Regress:
(a) (Z) follows logically from (A), (B).
(b) There is a bigger set also containing the premise ‘If (A), (B) are true, (Z) must be
true’, i.e. (C). [a, 1]
(c) (Z) follows logically from (A), (B), (C). [b, 2]
(d) There is a bigger set also containing the premise ‘If (A), (B), (C) are true, (Z) must be
true’, i.e. (D). [c, 1]
(e) (Z) follows logically from (A), (B), (C), (D). [d, 2]
And so on.
(5) There is an infinity of premises. [3, 4]
(6) There is no infinity of premises.
(C) (3) is false: There is no set of premises such that a conclusion follows logically from it. [1-6]

Failure II instance

(1) You have to demonstrate for at least one set of premises that a conclusion follows logically
from it.
(2) For any set of premises x, if you have to demonstrate that (Z) follows logically from x, then
you add a premise ‘if the members of x are true, then (Z) must be true’ to the argument.
(3) For any set of premises x, if you add such a premise to the argument, then you first have to
demonstrate that (Z) follows logically from a new set of premises (comprising x plus the
additional premise) in order to demonstrate that (Z) follows logically x.
(4) Regress:
(a) You have to demonstrate that (Z) follows logically from (A) and (B).
(b) You add the premise ‘if (A), (B) are true, (Z) must be true’ (C) to the argument. [a, 2]
(c) You have to demonstrate first that (Z) follows logically from (A), (B) and (C). [b, 3]
(d) You add the premise ‘if (A), (B), (C) are true, (Z) must be true’ (D) to the argument.
[c, 2]
(e) You have to demonstrate first that (Z) follows logically from (A), (B), (C) and (D). [d,
3]
And so on.
(5) For any set of premises x, you first have to demonstrate that (Z) follows logically from a
regress of premise-sets in order to demonstrate that (Z) follows logically from x. [4]
(C) You will never demonstrate that a conclusion follows logically from any set of premises if
you add a premise every time you have to demonstrate that a conclusion follows logically. [1-5]

3. Russell

Before Russell, philosophers held that relations are reducible to properties of their relata. If you
love God, to borrow Russell’s example, then presumably this is nothing but a property of you,
and not a relation between you and something else. This does however not work for all relations.
Consider the relation ‘earlier than’:
 If A is earlier than B, then B is not earlier than A. If you try to express the relation of A to
B by means of adjectives of A and B, you will have to make the attempt by means of
dates. You may say that the date of A is a property of A and the date of B is a property of
B, but that will not help you because you will have to go on to say that the date of A is
earlier than the date of B, so that you will have found no escape from the relation.
(Russell 1959: 54-5, cf. 1903: §214, 1906-07) \

Paradox instance

(1) For any relation R and items x, y, if x and y stand in R, then there are properties F and G such
that x is F and y is G.
(2) For any asymmetric R and items x, y, if x and y stand in R and there are properties F and G
and x is F and y is G, then F and G stand in R.
(3) At least two items stand in an asymmetric relation.
(4) Regress:
(a) a is earlier than b.
(b) a happens at t1, and b at t2. [a, 1]
(c) t1 is earlier than t2. [b, 2]
(d) t1 happens at second-order t*1, and t2 at second-order t*2. [c, 1]
(e) t*1 is earlier than t*2. [d, 2]
And so on.
(5) There is an infinity of time-orders. [3, 4]
(6) This is absurd.
(C) (1) is false: It is not the case that any relation implies corresponding properties of its relata.
[1-6]

Failure I instance

(1) You have to reduce all asymmetric relations.

(2) For any asymmetric relation R between items x, y, if you have to reduce R, then you reduce
R to a property F of x and a property G of y.
(3) For any asymmetric R between items x, y, if you reduce R to property F of x and property G
of y, then F and G stand in R.
(4) Regress:
(a) You have to reduce the earlier-than relation from a to b.
(b) You reduce it to the property ‘happening at t1’ of a and the property ‘happening at t2’
of b. [a, 2]
(c) You have to reduce the earlier-than relation from t1 to t2. [b, 3, 1]
(d) You reduce it to the property ‘happening at t*1’ of t1 and the property ‘happening at
t*2’ of t2. [c, 2]
(e) You have to reduce the earlier-than relation from t*1 to t*2. [d, 3, 1]
And so on.
(5) For any asymmetric relation R, you always have to reduce yet another asymmetric relation in
addition to reducing R. [4]
(C) You will never reduce all asymmetric relations if you reduce any of them that you have to
reduce to properties of their relata. [1-5]
4. Wittgenstein

What fixes the meaning of a word? Take the word ‘cube’. Wittgenstein argues that its meaning is
not fixed by a mental image of a cube, and also considers a somewhat expanded proposal:

Suppose, however, that not merely the picture of the cube, but also the method of
projection comes before your mind? – How am I to imagine this? – Perhaps I see before
me a schema showing the method of projection: say a picture of two cubes connected by
lines of projection. – But does this really get me any further? Can’t I now imagine
different applications of this schema too? (1953: §141)

Figure 2: Methods of projection

Paradox instance

(1) For any linguistic or mental item x, the meaning of x is fixed only if there is a method of
projection between a mental image of x and that to which x applies.
(2) Methods of projection are themselves mental items with fixed meanings.
(3) The meaning of at least one word is fixed.
(4) Regress:
(a) The meaning of the word ‘cube’ is fixed.
(b) There is a method of projection1: lines of projection between a mental image of a
cube and a cube. [a, 1]
(c) Method of projection1 is a mental item with a fixed meaning. [b, 2]
(d) The meaning of method of projection1 is fixed by a method of projection2: lines of
projection between a mental image of method of projection1 and method of projection1.
[c, 1]
(e) Method of projection2 is a mental item with a fixed meaning. [d, 2]
And so on.
(5) There is an infinity of methods of projection. [3, 4]
(6) It is not possible for us to handle so many, rather complex mental items.
(C) (1) is false: It is not the case that for any linguistic or mental item x, the meaning of x is fixed
only if there is a method of projection. [1-6]

Failure II instance

(1) You have to explain for at least one word what fixes its meaning.
(2) For any linguistic or mental item x, if you have to explain what fixes the meaning of x, you
appeal to the fact that it is fixed by a method of projection.
(3) For any linguistic or mental items x and y, if you appeal to a method of projection y, then you
first have to explain what fixes the meaning of y in order to explain what fixes the meaning of x.
(4) Regress:
(a) You have to explain what fixes the meaning of the word ‘cube’.
(b) You appeal to the fact that it is fixed by a method of projection1: lines of projection
between a mental image of a cube and a cube. [a, 2]
 (c) You have to explain first what fixes the meaning of method of projection1. [b, 3]
(d) You appeal to the fact that it is fixed by a method of projection2: lines of projection
between a mental image of method of projection1 and method of projection1. [c, 2]
(e) You have to explain first what fixes the meaning of method of projection2. [d, 3]
And so on.
(5) For any linguistic or mental item x, you first have to explain what fixes the meaning of a
regress of mental items in order to explain what fixes the meaning of x. [4]
(C) You will never explain what fixes the meaning of any word if you appeal to a method of
projection every time you have to explain what fixes the meaning of a word. [1-5]

5. Ryle

Suppose I want to write a dissertation. Of course, I do not want to write a dissertation in just any
way whatever. I want to write in an intelligent way, i.e. in such a way that I employ knowledge
of how to do such things. Now suppose, as the so-called intellectualist legend maintains, that
knowledge- And So On 88 how necessarily involves knowledge-that. If this is so, and if I want
to employ knowledge of how to write a dissertation, then I have to apply knowledge that
dissertations are to be written in such and such a way (e.g. that they are to be clear, thoughtful,
well-organised, etc.). That is, I have to apply knowledge with propositional content. Applying
knowledge with propositional content is itself, however, an action to be performed intelligently.
This lands us in a regress:

The crucial objection to the intellectualist legend is this. The consideration of

propositions is itself an operation the execution of which can be more or less intelligent,
less or more stupid. But if, for any operation to be intelligently executed, a prior
theoretical operation had first to be performed and performed intelligently, it would be a
logical impossibility for anyone ever to break into the circle. (1949: 30)

Paradox instance

(1) For any action x, one intelligently performs x only if one employs knowledge that x is to be
performed in such and such a way.
(2) For any action x, one employs knowledge that x is to be performed in such and such a way
only if one intelligently contemplates the proposition that x is to be performed in such and such a
way.
(3) You perform at least one intelligent action.
(4) Regress:
(a) I intelligently write a dissertation.
(b) I employ knowledge that [dissertations are to be written in such and such a way]. [a,
1]
(c) I intelligently contemplate the proposition that [dissertations are to be written in such
and such a way]. [b, 2]
(d) I employ knowledge that [the proposition that [dissertations are to be written in such
and such a way] is to be contemplated in such and such a way]. [c, 1]
 (e) I intelligently contemplate the proposition that [the proposition that [dissertations are
to be written in such and such a way] is to be contemplated in such and such a way]. [d,
2]
And so on.
(5) I perform an infinity of intelligent actions. [3, 4]
(6) I cannot perform an infinity of intelligent actions.
(C) (1) is false: It is not the case that for any action x, one intelligently performs x only if one
employs knowledge that x is to be performed in such and such a way. [1-6]

Failure II instance

(1) For at least one action x, you have to intelligently perform x.

(2) For any action x, if you have to intelligently perform x, then you employ knowledge that x is
to be performed in such and such a way.
(3) For any action x, if you employ knowledge that x is to be performed in such and such a way,
then you first have to intelligently contemplate the proposition that x is to be performed in such
and such a way in order to intelligently perform x.
(4) Regress:
(a) You have to intelligently write a dissertation.
(b) You employ knowledge that [dissertations are to be written in such and such a way].
[a, 2]
(c) You have to intelligently contemplate the proposition that [dissertations are to be
written in such and such a way] first. [b, 3]
(d) You employ knowledge that [the proposition that [dissertations are to be written in
such and such a way] is to be contemplated in such and such a way]. [c, 2]
(e) You have to intelligently contemplate the proposition that [the proposition that
[dissertations are to be written in such and such a way] is to be contemplated in such and
such a way] first. [d, 3]
And so on.
(5) For any action x, you first have to intelligently perform a regress of actions in order to
intelligently perform x. [4]
(C) If you employ knowledge that x is to be performed in such and such a way every time you
have to intelligently perform an action x, then you will never perform any intelligent action. [1-
5]

§4 Analysis

4.1. Comparison Schemas I: Structure

In the following I compare the Paradox and Failure Schemas on the basis of the instances from
the previous part, and show how they differ. Specifically, I compare their premises/hypotheses,
conclusions and dialectics. Throughout this part, I will focus on such structural features, and
postpone issues related to soundness until §6.2.

1. Similarities
As suggested in the previous part, any case from the literature can be reconstructed both as an
instance of the Paradox as well as an instance of the Failure Schema. There are several reasons
for this. First, there is much left implicit in actual regress arguments found in the literature.
Specifically, these arguments either have many suppressed premises, or, if not suppressed, it is
often unclear how their premises are quantified, and virtually always the inference steps to the
conclusion are left mysterious. Hence, it is not surprising that regress arguments can be made
explicit in different ways (i.e. as instances of different schemas).

Second, so long as there are predicates available to fill the schematic letters, it is always possible
to obtain instances of the schemas. For the same reason I doubt that it is possible to find a case
that cannot be reconstructed Paradox or Failure-wise. Of course, whether both reconstructions
are also plausible (i.e. in addition to being possible) is a separate question, and in the present part
I will continue to postpone it.

In this part I compare the two argument schemas in detail, and identify their structural
similarities and differences. Specifically, we shall look at their premises/hypotheses, conclusions,
and dialectics (i.e. the ways their instances function in a dialogue between opponents).

The main similarities are at least three-fold. First, and just noted, the schemas are similar in that
they can have all cases from the literature as instances. Second, they are similar in that they both
fulfil a number of basic desiderata listed in §2.1: namely, a regress occupies one of their lines,
the regress is derived from premises/hypotheses, and a conclusion is drawn from the regress.
Third, in both schemas all inferences from premises/hypotheses to regress and from regress to
conclusion are valid according to classical rules of inference. At some points this is not
immediately clear, yet, as I argued in §2, this worry can be met as soon as some suppressed
premises are made explicit.

Later we shall see some further similarities (e.g. that both schemas can account for certain
disputes concerning regress arguments). Given all these similarities, the question arises: What
actually are the differences?

The striking difference between the schemas is their rationale. The rationale of the Paradox
Schema is that a set of statements cannot hold at once because they jointly lead, via a regress, to
consequences which conflict with something else that we believe (that is, with independent
considerations). The rationale of the Failure Schemas is in contrast that a certain solution never
accomplishes a given problem because it gets stuck in a regress (that is, of similar problems that
have to be solved first).

Yet one might still think on the basis of these informal characterizations that the two schemas do
the same thing, but just in different terms (especially since both can have all cases from the
literature as an instance). Now, one may wonder why this issue is so important. Why is it
important that the schemas are really different? First, for my purposes their distinction is
important because the schemas are the basis for the two theories of regress arguments (i.e. the
Paradox vs. Failure Theory), and if the schemas are not really different, then there are
presumably no such different theories either. Second, in §6 I will address the query: If regress
arguments can be reconstructed in two different ways (which differ in format, not merely in
content), then how should they be reconstructed? Clearly, this query makes no sense if the
schemas are not really different.

In the following two sections I shall compare Paradox and Failure arguments in two stages:
considering first those parts of them which generate the regress, and then those parts connecting
regress to conclusion. I will argue that although Paradox and Failure arguments generate
regresses in comparable ways, the parts of these arguments which associate a conclusion to a
regress are structurally different. In both cases I first consider (and reject) a suggestion by
Gratton on this issue, i.e. on whether arguments based on problem/solution regresses can be
taken as a variety of Paradox arguments. After that, I present my own analysis.

2. Generating the regress Next I compare the first parts of Paradox and Failure arguments, i.e. up
to the first result entailed by the regress. Before presenting my own analysis, I shall consider a
suggestion by Gratton concerning a parallel between Paradox’s regress formulas and Failure’s
extra premises (i.e. line (3), which is partly responsible for the new problems in the regress):

For any problem U of the kind V, there is a response W that entails a contradiction that
functions as a premise in a new problem X of the kind V. (2010: 166)

The suggestion may be spelled out along the following lines:

Trigger An arbitrary problem of type K is to be solved.

Regress formula 1 Any problem of type K is solved by a solution of type Q.

Regress formula 2 Any solution of type Q entails a problem of type K.

Result An infinity of problems of type K and solutions of type Q.

Gratton himself rejects this mapping because, according to him, the goal of Failure arguments is
not to reject such regress formulas, and hence there is no need to include them in the argument. I
shall discuss this point at some length in §6.1.3.

I agree that these parallels are problematic, yet for reasons unrelated to Gratton’s analysis. My
first problem: Where is the difference between Failure Schemas I and II? A second problem: It is
unclear why all the problems of type K must be solved. It is clear that solutions of type Q entail
new problems of type K, yet so long as it is not stated that all such problems have to be solved,
no worrisome regress is generated. Third problem: The result is rather different from the
conclusion of a Failure argument. In order to address these problems, we should rather look for
two different two sets of parallels, i.e. parallels which depend on whether Paradox arguments are
compared with Failure I arguments or with Failure II arguments. Here they are.

Parallels Paradox vs. Failure I

Regress formula 1 All problems of type K are to be solved.

Regress formula 2 Any problem of type K is solved by a solution of type Q.

Regress formula 3 Any solution of type Q entails a problem of type K.

Result There is always yet another problem of type K to be solved.

Parallels Paradox vs. Failure II

Trigger At least one problem of type K is to be solved.

Regress formula 1 Any problem of type K is solved by a solution of type Q.

Regress formula 2 Any solution of type Q entails a problem of type K, and the latter is to be
solved in order to solve any initial problem.

Result There is always first yet another problem of type K to be solved in order
for at least one problem of type K to be solved.

Let me illustrate these parallels by reference to Sextus’ case. The first lines of the Failure I and
Paradox reconstruction can both be rendered informally as follows: You have to justify all your
beliefs; You justify all of your beliefs by providing a reason for them; Reasons are also among
your beliefs; Result: You always have to justify yet further beliefs.

By contrast, the first lines of the Failure II and Paradox reconstruction can both informally be
rendered like this: You have to justify at least one of your beliefs; You justify all of your beliefs
by providing a reason for them; You have in turn to justify those reasons in order to justify any
initial belief; Result: You always have to justify yet further beliefs in order to justify any initial
belief.

On this account, the three problems with Gratton’s first suggestion concerning the link between
Paradox and Failure arguments are accounted for. To begin with, the difference between Failure
Schemas I and II is made explicit.

Second, it is made clear why all problems must be solved (and hence why a regress is generated
at all). In the Failure I case, the first regress formula accounts for this, as it simply states that
each and every problem of a certain type K is to be solved. In the Failure II case, the second
regress formula accounts for this, as it states that the newly generated problems are to be solved
in order to solve any initial problem.

Third, the results drawn match with the results of Failure regresses. Compare: ‘For any K that is
F, there is another K that is F’ (Paradox) vs. ‘For any K that is to be φ-ed, there is another K to
be φ-ed’ (Failure). For example: ‘For any proposition that is justified to you, there is another
proposition that is justified to you’ vs. ‘For any proposition that you have to justify, there is
another proposition you have to justify’. The reason for this last similarity is that the inferences
from premises/hypotheses are very much the same (i.e. the rules used are mainly Modus Ponens
and Universal Instantiation), and so it not surprising that similar kinds of results follow from
Paradox and Failure regresses. Hence, according to this analysis, the first lines of Paradox and
Failure arguments run parallel.

Nevertheless, the similarities are not always immediately apparent. Consider for instance my
reconstructions of Wittgenstein’s case from §3.4.4: ‘The meaning of the word ‘cube’ is fixed’
(Paradox) vs. ‘You have to explain what fixes the meaning of the word ‘cube’’ (Failure II). Here,
the Paradoxline does not describe the state to be obtained according to the Failure II-line. If it
did, it would say ‘It is explained what fixes the meaning of ‘cube’’. There are more of such
irregularities, and all may have their own explanation. In this case, the irregularity is due to the
fact that the problem is theoretical (of the form ‘explain, find out or demonstrate how such and
such can be the case’) rather than practical (of the form ‘do such and such’). I have chosen for
the theoretical reading, here, for it seems already settled what the meaning of ‘cube’ is. You just
have to explain something that is already settled. Contrast this with Quine’s case: ‘You have to
fix the reference of the word ‘Gavagai’’ (which would, then, be a practical problem) vs. ‘You
have to explain what fixes the reference of the word ‘Gavagai’’ (a theoretical problem). Here, I
would opt for the practical version instead, because it seems not already settled what the
reference of ‘Gavagai’ is.43

Yet, these are irregularities and do not invalidate the suggestion that Paradox and Failure
regresses are generated in a similar way. Despite this fact, Paradox and Failure arguments are
different kinds of arguments, as their endings do not run parallel. As I will explain next, the
important structural difference between the schemas is the connection between regress and
conclusion.

3. From regress to conclusion

Again I will first consider Gratton’s suggestion concerning the link between Paradox and Failure
arguments. This time, it does not concern the manner in which regresses are generated, but the
portion following the regress:

If we believe that a solution is […] acceptable only if it does not generate the same kind
of problem it is supposed to resolve, then the regress entails a statement that conflicts
with a belief. If we do not abandon the belief, then the regress entails an unacceptable
statement, and so the regress is vicious. (1997: 217)

This reasoning can be unpacked as follows:

(1) Regress.
(2) A solution S generates the same kind of problem it is meant to solve. [1]
(3) If S generates the same kind of problem it is meant to solve, then S is unacceptable.
(4) S is unacceptable. [2-3]

The main similarity between this reasoning and the Paradox Schema is that a result is drawn
from the regress and shown to be unacceptable. Yet the suggestion is problematic, because (3) is
problematic. Let us consider (3) in more detail. At first, it seems a tempting and not uncommon
assumption. Compare:
 No candidate solution to a philosophical problem should raise another problem which
appears just as intractable and which requires the resolution of an issue similar to that
which made the original problem so intractable. (Noordhof 1998: 223)

Now, why would anyone accept such a premise? Consider this instance:

• If a solution to eliminating the Liar Paradox entails a similar problem, then that solution is
unacceptable.

Recall the Liar sentence:

L L is false.

If L is true, then what it says is the case: it is false (i.e. as well as true). If L is false, then it says
what is the case and so it is true (i.e. as well as false). Paradox. One proposal to resolve the
paradox, from Tarski, is the following.44 For a given language X, assume that of no sentence in X
it can be stated within X that it is true or not, yet that a sentence in X is true or not can be stated
in a metalanguage of X. By this, L is not possible given that it says of some English sentence,
namely itself, that it is false within the same language. Yet the same paradox can still be put in
the metalanguage, and so a similar problem can be generated (or, as it is also put in this debate,
the paradox takes its revenge).

Also consider a slightly different proposal to resolve the paradox (cf. Cook 2007). The idea is to
introduce a third semantic value (whatever its specific nature), call it ‘pathological’, and to
regard L as pathological (and hence not as true or false). By this, the Liar Paradox is blocked as
L is to be neither true nor false. But now consider the following sentence:

L* L* is either false or pathological.

If L* is true, then what it says is the case: it is either false or pathological (i.e. as well as true). If
L* is false, then it says what is the case and so it is true (i.e. as well as false). If L* is
pathological, then again it says what is the case and so it is true (i.e. as well as pathological).
Again, the paradox takes its revenge.

The question is: If a solution to the Liar Paradox generates a similar paradox, is it then a bad
solution? This very much depends on what the problem is. If the problem is to resolve the Liar
Paradox in any language whatever, and concerns alike sentences L, L*, etc., then the solution is
not effective. Yet, if the problem is merely to resolve the Liar Paradox stated in English (and not
stated in any other language, or not concerning any other sentence such as L*), then the solution
may well be effective.

In terms of the Failure Schemas: If the problem is a general one, solutions should not generate
more instances of that problem. But if the problem is merely particular, such vengeance should
cause no worry. Furthermore, counterexamples similar to Gratton’s (3) can be set up for all
Failure I instances. So, if (3) is problematic because it structurally admits counterexamples, then
Gratton’s suggestion is problematic.45 Moreover, next I will argue that any such link between
Paradox and Failure arguments will fail, as the portions of these arguments following the regress
differ structurally.

Specifically, in case of Paradox arguments there is a mediate connection between regress and
conclusion. Namely: in order to obtain a rejection of one of the initial hypotheses, you need an
extra premise which conflicts with anything entailed by the regress (e.g. a premise that
contradicts that an infinity of Ks are F). In Sextus’ case, you need an extra premise like ‘It is not
the case that an infinity of propositions are justified to S’ or ‘It is not the case that there is an
infinity of reasons’ or ‘S is psychologically unable to have an infinity of reasons available’. If no
such extra premise holds, then nothing can be rejected. Indeed, the view known as ‘Epistemic
Infinitism’ maintains that the regress entails no unacceptable results and that nothing should be
rejected (more on this view later: §5.4).

In the case of Failure arguments, by contrast, there is an immediate connection between regress
and conclusion. No extra premise is needed to obtain the conclusion that the problem is never
solved by the considered solution (or at least no extra substantial premise, as I will say later on
in §6.2.2). In Sextus’ case, you just never justify any proposition (so long as you provide a
reason every time you have to justify a proposition), no matter whether or not an infinity of
reasons can be available to you. In Ryle’s case, you just never perform any intelligent action (so
long as you employ knowledge-that every time you have to perform an intelligent action), no
matter whether or not you can contemplate an endless series of propositions.

A similar point has been noted by Day:

This version of Passmore’s argument works even if there is not an infinity of forms. All
that matters is that when we analyse any case of predication it will be possible for us to
find another to analyse. (1986: 50)

Day does not strictly speaking talk about Failure arguments. However, as I regard the Failure
Schemas as an explication of what Passmore had to say, Day’s claim may well be taken to apply
to Failure arguments. The specific point about (a version of) Plato’s Third Man can be put in
Failure terms like this: You will never analyse all cases of predication if you invoke a form every
time you have to analyse a case of predication, whether or not there is an endless number of
forms. Generally: You will never φ all/any K(s) if you ψ all Ks that you have to φ, whether or
not there is an endless number of Ks for you to ψ.

Later on I will argue, furthermore, that Failure arguments show that it is never the case that
all/any problem(s) are solved in the sense that there is always yet another problem to be solved,
whether or not there is in fact a solution for all of them (see §6.1.2.4).

I should say that Failure arguments do require a suppressed premise. However, this kind of
premise differs from the one required by the Paradox Schema. Specifically, Failure I and II
require respectively:46
• If for at least one K you have to φ it and if for all Ks x, you always have to φ a new K in
addition to φ-ing x, then you will never φ all Ks.

• If for at least one K you have to φ it and if for all Ks x, you first have to φ a regress of new Ks
in order to φ x, then you will never φ x.

For example: ‘If you have to justify a proposition and there is always a new proposition that you
have to justify, then you will never justify all propositions’; ‘If you have to justify a proposition
and there is always a new proposition that you have to justify in order to justify any proposition,
then you will never justify any proposition’. Indeed, these are rather different from the premise
required by the Paradox Schema (as ‘S does not have an infinity of reasons’). This means that
you have to assume different things to get the arguments going. It demonstrates that the last parts
of Paradox and Failure arguments do not run parallel, and that they are distinct kinds of
arguments.

To support this reasoning we may check whether the relevant conclusions entail each other. For
if they do still entail one another, then perhaps it might not matter much how you obtain the
conclusion. If we consider the concluding lines of the two schemas, they appear very different:

PARA It is not the case that for all Ks x, x is F only if such and such.

FAIL You will never φ all/any K(s), if you ψ all Ks that you have to φ.

For example: ‘It is not the case that for any proposition, x is justified to S only if S has a reason y
for x’ vs. ‘You will never justify any/all proposition(s), if you provide a reason for all
propositions that you have to justify’. Neither of these entails the other, which means that either
can hold without the other holding. Or again: there is no parallel between Paradox’s rejection of
regress formulas and Failure conclusions.

It is worth stressing this point. In contrast to Paradox arguments, Failure arguments are not about
rejections, not even the rejection of solutions. More precisely, they do not prove that a solution is
false, but that it is no good for accomplishing a given problem (and furthermore that another
solution has to be found). As a consequence, it is implausible to take the concluding line of the
Failure Schema as the negation of its hypothesis (2), i.e. as: ‘It is not the case that for all Ks x, if
you have to φ x, you ψ x’ (e.g. ‘It is not the case that for all propositions x, if you have justify x,
then you provide a reason for x’). Rather than negating hypothesis (2) on the basis of Reductio
Ad Absurdum, Failure arguments merely discharge the hypothesis by Conditional Proof.47

To be sure, the concluding line of the Failure Schema is incompatible with ‘If you ψ any K that
you have to φ, then you φ all/at least one K(s)’ (e.g. ‘if you provide a reason for any proposition
that you have to justify, then you justify all/at least one proposition(s)’). Yet this is not part of the
Failure Schema. If it were assumed, then no solution could be bad or unsuccessful.

4. Scepticism
There is nonetheless one point where Paradox and Failure arguments converge: namely,
scepticism. Sceptics often employ regresses in their arguments (cf. Sextus and the ancient
Pyrrhonists). Here is their position in terms of both schemas:

Paradox Scepticism
It is not the case that at least one K is F.

Failure Scepticism
You will never φ any/all K(s), no matter your solution.

For example: ‘It is not the case that at least one proposition is justified to someone’ vs. ‘You will
never justify any proposition, whatever you do’.

Paradox sceptics maintain that the regress is unacceptable such that something needs to be
rejected, yet, because it holds onto the regress formulas, it concludes by rejecting the trigger.
Further examples:

• No dispute is decided.

• No inductive inference is justified.

• The reference of no word is fixed.

• No one is entitled to any epistemic rule.

Failure sceptics cannot stop at the last line (C) of the Failure Schema which says that a certain
solution never solves the given problem, but have to add the extra step that no alternative
solution can do the trick either (i.e. by showing that all other solutions have regressive
consequences as well, or fail for a different reason). Further examples:

• Whatever you do, you will never decide any dispute.

• Whatever you do, you will never justify any inductive inference.

• Whatever you do, you will never fix the reference of any word.

• Whatever you do, you will never be entitled to any epistemic rule.

Furthermore, Paradox Scepticism entails Failure Scepticism (or more precisely: Failure II
Scepticism), and vice versa. For example, ‘you will never justify any proposition, whatever you
do’ entails and is entailed by ‘it is not the case that at least one proposition is justified to you’
(modulo a qualification on the notion of justification, see §5.4).

Even though Paradox and Failure arguments converge at this point, the ways in which each has
arrived there have followed very different routes. As just explained, Paradox Scepticism needs to
show that the regress entails an unacceptable result and that the trigger rather than the regress
formulas is to be rejected. Failure Scepticism, in contrast, needs to show that all solutions to a
given problem are unsuccessful.

5. Dialectics

In the following I will show that there is yet another difference between Paradox and Failure
arguments. Specifically: they differ not only with respect to their premises and conclusions, but
also play different roles in a broader dialectical context. Any argument is devised for or against a
certain position. Dialectical contexts, then, are such things where all parties concerned about that
position are identified, and where for all steps of the argument it is made clear which of the
parties subscribe to them.

If ‘NN1’ and ‘NN2’ stand for two arbitrary persons, then the dialectical contexts of Paradox and
Failure arguments are the following:

Paradox Dialectic

Step Context

Regress formula 1 NN1’s belief

Regress formula 2 NN2 argues that NN1 has to concede this

Trigger Common sense claim

Regress NN2 infers this from the foregoing

Extra premise NN2 argues that NN1 has to concede this

Conclusion NN2 concludes that NN1’s belief has to go

Failure Dialectic

Step Context

Problem Common concern of NN1 and NN2

Solution NN1’s proposal

Extra premise NN2 argues that NN1 has to concede this

Regress NN2 infers this from the foregoing

Conclusion NN2 concludes that NN1’s proposal fails

To press the point again from §4.1.3 above, the important structural difference between the two
schemas is the following. In the Paradox Schema, NN2 needs to make an extra step after the
regress. In the Failure Schema, by contrast, there is no such extra step.

In both cases, NN2, i.e. the one who devises the regress argument, does all the work, and NN1
may try to resist the reasoning at any of the reasoning steps proposed by NN2. As always, there
are two ways to challenge an argument: to attack one of its premises, or attack one of its
inferences.

In the case of Paradox arguments, the main options for NN1 to resist the argument are to defend
(i) that something else is to be rejected (for example, another regress formula), or (ii) that
nothing needs to be rejected in the first place (because the regress entails no unacceptable result).
Consequently, debates on Paradox arguments may centre on the following questions:

(Q-1) Does the regress entail something unacceptable?

(Q-2) If it does, then what should be rejected?

Consider Sextus’ case taken in terms of the Paradox Schema:

Justification (Paradox instance)

(1) For any proposition x, x is justified to S only if S has a reason y for x (where y is a new
proposition).
(2) For any propositions x and y, S has a reason y for x only if y is justified to S.
(3) At least one proposition is justified to S.
(4) Regress: […] [1-2]
(5) S has an infinity of reasons. [3, 4]
(6) S does not have so many reasons.
(C) (3) is false: No proposition is justified to S. [1-6]

The (Q-1) debate turns on premises like (6). Is it plausible to assume that S does not have so
many reasons? If this is plausible, then the regress indicated by (5) is acceptable and nothing is to
be rejected on this basis. To be sure, in order to give (Q-1) a full treatment all other results from
the regress need be taken into consideration as well.

By contrast, the (Q-2) debate turns on whether (3) is to be rejected rather than lines (1) or (2).
The sceptic would hold that (3) should be rejected. Yet Foundationalism and Coherentism both
differ from scepticism on this, and propose to reject (1) instead, though for different reasons.
Foundationalism rejects (1) because of its view that certain propositions (i.e. the basic ones) are
justified to someone independently of their relation to further propositions. Coherentism (or at
least a simple version of it) rejects (1), or accepts only a modified version of (1), because of its
view that reasons need not always be new, i.e. they may previously be used in the regress.
Consequently, Coherentism allows for circles of reasons, in which case line (5) does not follow,
i.e. it does not follow that there is an infinite series of propositions.
In the case of Failure arguments, the main options for NN1 to resist the argument are to defend
(i) that the solution was never meant to be fully general, or (ii) that the extra premise does not
hold (in Failure II cases), or the problem is too general (in Failure I cases). Consequently,
debates on Failure arguments may centre on the following questions:

(Q-3) Is the solution to apply to each and every problem of a certain kind?

(Q-4) Are the problems generated in the regress relevant for the initial problem?

Consider the following principle from Davidson (1970):

(R) If action x is what you regard as the best option, all things considered, then you ought to
perform x.

Lazar (1999) shows that it is possible to set up a Carroll-style regress argument against the use of
(R). Take the following example. If I take it that it is the best option to respect other people, all
things considered, then by (R) I ought to respect other people. Now suppose I accept that I
indeed take it that respecting other people is the best option, yet deny that I am obliged to respect
them just because I deny (R). One possibility to force me to accept the obligation is to appeal to
the following meta-principle:

(R*) If respecting (R) is what you regard as the best option, all things considered, then you ought
to respect (R).

But of course I can resist this meta-principle in the same way, and a regress has begun. Yet, the
message here is not that (R) is useless. On the contrary, it is that (R) is simply more basic than
other rules, and should not be treated on the same footing. (R) can be used to generate
obligations from all-thingsconsidered judgments, and so the obligation to comply with (R)
should not itself depend on further all-things-considered judgments and meta-principles. Or
again: (R) is to be obeyed not because we have an obligation to do so, but because it is meant to
be constitutive for all our obligations, i.e. without it no obligations would be possible in the first
place.

This reasoning may be controversial.48 Still, the important point for my purposes is that this
debate is a (Q-3) debate: it centres on the issue of whether the solution (here: appealing to (R))
applies to all problems of a certain kind (here: explaining how I have an obligation). And the
suggestion is in the negative: not all my obligations are to be explained by rules, only obligations
which do not concern the obligation to respect (R) are thus to be explained.

Next consider the debate concerning the Problem of Universals as construed by Armstrong
(1974, 1978). Basically, this is the problem to explain how distinct items can nevertheless have
the same property or stand in the same relation. Armstrong argued that virtually all solutions to
this problem (i.e. Realism about Universals, Resemblance Nominalism, Class Nominalism, etc.)
are regressive:
 If a’s being F is analysed as a’s having R to a ø, then Raø is one of the situations of the
sort that the theory undertakes to analyse. So it must be a matter of the ordered pair <a,
ø> ‘ having R* to a new ø- like entity: øR. If R and R* are different, the same problem
arises with R* and so on ad infinitum. (1978: 70-1)

Specifically, all solutions fall prey to a Failure I regress and never solve all (rather than any)
instances of the Problem of Universals. Yet, if all solutions are regressive in this sense, then no
solution is successful. Lewis draws from this the conclusion that the problem must be too strong:

But the clincher, the one argument that recurs throughout the many refutations, is the
relation regress. […] Doing away with all unanalysed predication is an unattainable aim,
and so an unreasonable aim. No theory is to be faulted for failing to achieve it. (1983:
353-4, cf. Oliver 1996: 33)

Indeed, if no solution is successful, perhaps there is something wrong with the problem. This is a
(Q-4) debate: it concerns the issue of whether all newly generated problems in the regress need
to be solved in the first place.

Hence, the schemas are accompanied by different possible debates. This difference between the
dialectics of Paradox and Failure arguments will prove important when I turn to matters of
soundness (§6.2). Important here is that (Q-1) to (Q-4) are the four main points that can be
disagreed about when there is a dispute about a regress argument. That is, if you do not buy the
conclusion of a regress argument, then you should start one of the debates (Q-1), (Q-2), etc.

Alongside these four, there are a few further possible debates. In principle, any premise
(suppressed or not) or inference (suppressed or not) can be put into question. Yet these latter
debates occur less often, as most suppressed premises concern common background assumptions
of the debate, and the inferences are licensed by classical rules. Nevertheless, some extra options
will be identified in §4.2.

Gratton defends the Paradox Schema by stating that it can explain why there is often
disagreement as to whether a certain regress is problematic:

The relativized portion of my account helps to explain why there can be disagreement in
establishing the viciousness of some infinite regresses: some people hold certain beliefs
that conflict with the result entailed by an infinite regress, while others do not hold those
beliefs. (1997: 216)

This is what I labelled a (Q-1) debate. Regresses entail something unacceptable as soon as they
entail a result which conflicts with something else we believe. But the latter is usually
controversial, and opens up room for disagreement.

The question is: does this favour the Paradox Schema over the Failure Schema? No. First, that
there is disagreement in the literature may of course just be due to the fact that people are
confused about what to draw from a regress. But more importantly, as we have just seen with (Q-
3) and (Q-4) debates, there is room for disagreement about Failure arguments too.
Final point. In some places, it has been stressed that regress arguments are not as strong as they
appear to be because they admit resistance (cf. Waismann 1956: 28-30, Passmore 1961: 35-7).
This is, in some sense, a strange point, as it could be expected that anything with premises can be
resisted. I take it that regress arguments can be strong arguments in the sense that they can
establish substantive conclusions, i.e. conclusions that make a difference to a relevant debate
(again: given that the premises hold).

4.2. Regress

In the following I turn to two related queries: what are regresses, and when are they bad? Each
question will be answered twice: once within the context of the Paradox theory, and once within
the context of the Failure theory. I also discuss how regresses differ from circularities.

1. Series vs. regresses

When is a series a regress? Recall the following hypotheses from §1.1:

Para–A Regresses are series of conditions which result from a number of claims and yield
something absurd.

Fail–A Regresses are series of problems which result from certain solutions and which prevent
the success of these solutions.

At this point these can be made precise.

1.1. Paradox regress

By the Paradox Schema, a series is a regress iff

(i) the series consists of necessary conditions; and

(ii) each necessary condition is entailed by the previous condition in combination with a regress
formula.

Consequently, anything which does not fulfil both conditions is not to be a regress (or at least no
Paradox regress). So, by (i), not all series are to be regresses. Only those series are regresses
which consist of necessary conditions (this clause has been mentioned by Black 1996: 115). By
(ii), not all series of necessary conditions are to be regresses. Only those series are regresses
where each necessary condition is entailed by the previous condition plus a regress formula (this
clause has been mentioned by Gratton 2010: 18).

Surely not all infinite series fulfil both conditions. For example, any regress such as the
following can also be taken in a non-entailed way, though in that case it would not be a regress
(simply because no step is entailed by the previous one):
- p1 is justified to S;
- S has a reason p2 for p1;
- p2 is justified to S;
- S has a reason p3 for p2;
etc.

What, then, are necessary conditions?49 In this dissertation, I assume that ‘A only if B’ or ‘B is a
necessary condition for A’ involves two relations. First, the relation between A and B comprises
the material conditional: A materially implies B, which means that B is true whenever A is true
(i.e. this relation holds whenever B is true or A is false). Second, the relation between A and B
comprises dependency: A somehow depends on B. The second relation is more intimate than the
first. That is, there are more facts which materially imply something than facts which depend on
something. For example, my girlfriend’s being reliable materially implies that Socrates is a
philosopher (the latter is true whenever the former is), but it seems implausible to think that her
reliability somehow depends upon anything about Socrates (modulo far-fetched scenarios).

Consider the following case: My girlfriend is reliable only if she is guarded by a guardian. This,
then, comprises two relations. First, my girlfriend’s being reliable materially implies that she is
guarded, which means that whenever it is true that she is reliable, it is true that she is guarded
(and whenever it is false that she is guarded, it is false that she is reliable). Second, my
girlfriend’s being reliable depends upon her being guarded by a guardian. For example, if
reliability means that she will not run away or commit unfaithful acts, then this dependency can
be taken causally: her being guarded is a cause of her being reliable.

But dependency relations need not always be causal. Consider Sextus’ case: a proposition p is
justified to S only if S has a reason for p. Here, p’s being justified to S depends upon S’s having
a reason for p. It is not plausible to regard this as a causal dependency: S’s having a reason for p
does not cause p’s being justified to S. Still, it may be regarded as a constitutive dependency: p’s
being justified to S is partly constituted by the fact that S has a reason for p (however one further
unpacks this).

At this point we have specified the first part of hypothesis Para–A. First, regresses are series of
consequences in the sense that each necessary condition is a consequence of the former. Second,
they result from a number of claims in that they are entailed by regress formulas.

So far I have not discussed how regresses in this sense yield something absurd. I will turn to this
below in the section on viciousness. But it is worth stressing already here that not all regresses
entail something absurd. Or in other words: not all infinite series which fulfil both conditions (i)
and (ii) are also vicious regresses. Vicious regresses are just a subclass of regresses, and the
criterion of viciousness is meant to distinguish them strictly from non-vicious ones.

1.2. Failure regress

By the Failure Schema, a series is a regress iff

(i) the series consists of similar problems and similar solutions; and
(ii) each problem and solution is entailed by a previous problem and/or solution in combination
with one or more of lines (1)-(3).

Anything which fails to fulfil both conditions is not to be a regress (or at least no Failure
regress). So, by (i), not all series are to be regresses. Only those series are regresses which
consist of similar problems and similar solutions (this clause has been mentioned by Schlesinger
1983: 221). By (ii), not all series of similar problems and similar solutions are to be regresses.
Only those series are regresses where each problem and each solution is entailed by previous
steps in combination with one or more of lines (1)-(3) (depending on whether the regress is
generated in terms of Failure I or Failure II, cf. Figure 1).

Surely not all infinite series fulfil both conditions (i) and (ii). Suppose, for example, that you
have to write down all natural numbers, and as a solution do write them all down. In that case we
have the following series of similar problems and similar solutions:

You have to write down number 1;

You write down number 1;
You have to write down number 2;
You write down number 2;
etc.

Yet, as these problems and solutions are given at once, and not entailed along the way, this series
is not to be a regress. In the same way, the following argument is not to be a regress argument:

(1) You have to write down all numbers.

(2) There is an infinity of numbers.
(3) You will never write down all numbers. [1-2]

You can get from (1) and (2) to (3) by assuming that if you have to perform an infinity of
actions, then there is will always be a new action to be performed such that you will never
perform all of them. This resembles a Failure I argument, but is not one for the simple reason
that no line is occupied by a regress.

It is instructive to compare Failure regresses with supertasks at this point. Both consist of an
infinity of tasks (or actions that have to be performed), yet they should not immediately be
identified. To see this, consider a well-known supertask: Thomson’s Lamp (Thomson 1954). The
scenario is such that you perform the following infinite series of tasks:

You turn the lamp on;

After a minute, you turn it off;
After a half a minute, you turn it on;
After a quarter of a minute, you turn it off;
After an eighth of a minute, you turn it on;
etc.
Again and again, you turn the lamp off if it is on, or on if it is off at increments declining in
length by one-half.

Now, there are two main differences between this kind of supertask and Failure regresses. First:
Failure regresses do not merely consist of tasks that have to be accomplished, they also consist of
actions that are supposed to accomplish those tasks. Second, and as just explained, the tasks in
Failure regresses are not given at once, but entailed along the way. Specifically, they are entailed
by the actions that are supposed to accomplish those tasks. Both aspects are missing out in
Thomson’s Lamp: it does not have a task/action structure, and as a consequence the new tasks
are not generated along the way but given at once.

Nevertheless, this does not show that certain other supertasks cannot fit into the Failure
framework. That is, Failure regresses might still be seen as a kind of supertask. More
specifically, they are supertasks to which condition (i) and (ii) above apply. I shall give examples
of such supertasks soon (see §4.2.2 below).

Problem Decide whether p is true.

Solution Use a criterion c.
Similar problem Decide whether c is true.

Figure 3: Problem shifting

Here is a nice, appropriate description of the relation between the problems in a regress:

What the two infinite regresses bring out is that the predicate nominalist does not in fact
solve his problem, he simply shifts it. (Armstrong 1978: 21)

That is, a solution shifts the problem iff it entails a problem which is similar to the one it is meant
to solve. This relation of problem shifting is depicted by the arrow in Figure 3.

The question is why all problems must be similar, i.e. of the same type. For you might expect it
still to follow that the problem is never solved given that each time you have to solve yet another
problem (no matter what kind of problem). The explanation is that without the similarity there is
no link between lines (1)-(3) of the Failure Schema, and no regress would be entailed. For
example, no regress results from the following set of lines:

(1) For all items x of type K, you have to φ x.

(2) For all items x of type K, if you have to φ x, you ψ x.
(3) For all items x of type K, if you ψ x, then there is a new item y of type Q.

In this case, if there is an item a of type K, then by (1) you have to φ a, by (2) you ψ a, and by (3)
there is a new item y of type Q, call it ‘b’. Yet then, by (1), it does not follow that you have to φ
b. Here is a real example of this scenario:

(1) For all natural languages x, you have to explain how we are able to learn x.
(2) For all natural languages x, if you have to explain how we are able to learn x, you appeal to
the fact that we master a language of thought by which we can learn x.
(3) For all natural languages x, if you appeal to the fact that we master a language of thought by
which we can learn x, then there is another language that we master.

No regress is entailed because, as it stands, (1) applies only to natural languages such as English,
not to mental languages (cf. Laurence & Margolis 1997). This shows why the problems in a
regress should be similar to each other. Moreover, a similar story holds for why the solutions in a
regress should also be similar to each other.

At this point we have specified the first part of hypothesis Fail–A, i.e. the sense in which
regresses are series of problems which result from solutions. As in the previous case, I did not
specify how regresses in this sense prevent the solutions from being successful. Or in other
words: so far I have not discussed how Failure regresses can be vicious.

2. Viciousness

Here is a familiar argumentative move:

It could be held that the paradox itself shows that there is this infinite hierarchy of types
of meaninglessness. […] This is the solution to which a certain sort of logician would be
forced. But it is none the less absurd. (Mackie & Smart 1953: 65)

What virtually all regress-theorists share is their denunciation of such moves. No regress is just
absurd, i.e. absurd full-stop without any further explanation. Sometimes the debate on regresses
is presented in such a way that its chief concern is to divide vicious regresses from
virtuous/benign ones. In that case, the basic task is to provide a criterion which distinguishes
regresses which are harmful for a theory (and which have to be stopped by a proponent of that
theory) and those which are harmless (and so cause no worry for a proponent of that theory).
Compare:

An infinite regress is either vicious or virtuous. Virtuous regresses are not virtuous in the
sense that their existence is cause for celebration. Instead, they are virtuous because you
can tolerate them. (Maurin 2007: 3, cf. Nolan 2001: 523)

According to this approach, the various reasons to regard regresses as vicious are to be classified
and evaluated. In contrast to this focus on viciousness, I have opted for a slightly different
approach. As stated explicitly in §1.3.2, my focus falls not directly on theories and whether
regresses might be bad for them, but rather on argument schemas in which regresses are
associated with a certain conclusion. These two projects are of course connected, but need not
coincide if those conclusions happen to be useless for theory choice.

Nonetheless, even if the question of viciousness need not be answered to obtain regress
arguments schemas, it is still possible to answer it. In the following, I will associate the Paradox
and Failure Schemas each with their own criterion of viciousness. That is, I will identify a
criterion which divides vicious from virtuous Paradox regresses, and likewise one which so
divides Failure regresses. In both cases the idea is rather simple:

(Vic) For any regress R and theory T, R is vicious for T iff (i) T is committed to the
premises/hypotheses by which R is generated, and (ii) R entails a result which is bad for
T.

Still, as we shall see, what exactly this bad result amounts to depends on the schema in terms of
which the regress is generated.

2.1. Paradox viciousness

The criterion of viciousness for the Paradox Schema is the following (cf. Gratton 1997: 216,
2010: 101):

(Vic-P) For any regress R and theory T, R is vicious for T iff (i) T is committed to the
regress formulas by which R is generated, and (ii) R entails at least one unacceptable
result.

Some clarifications are in order. First, regresses may entail more than one unacceptable result.
But one is enough for viciousness.

Second, results are unacceptable if they are false or (what is more likely in philosophy) if they
are inconsistent with something else that theory T is unwilling to abandon. Or again: results are
unacceptable if the benefits of rejecting one of the propositions that generate the regress
outweigh the costs of the regress and its results (cf. Nolan 2001, Cameron 2008). Whenever this
is the case, T is in trouble, for its commitments are inconsistent: it is committed to claims that
entail R and to claims inconsistent with what R entails. Consider for example the regressive
claim that for every x, there is a singleton {x} (i.e. the set of itself). This generates regresses such
as: Socrates, {Socrates}, {{Socrates}}, and so on. The question is whether the benefits of
rejecting the claim that anything has a singleton outweigh the costs of such regresses (in this case
infinitely many sets). The regress of singletons is vicious, then, only if the costs are too high.

This criterion is very general and can capture many specific criteria such as the following:
Regresses are vicious whenever they entail something absurd or counter-intuitive, paradoxical or
logically impossible, beyond human capacities, uneconomical, etc.

To illustrate these points, let us consider Vander Laan’s regress argument in Paradox-format:

Restrictivism (Paradox instance)

(1) For any action x, one is frequently free to perform x only if one is frequently free to generate
a desire to refrain from performing x.
(2) One is frequently free to perform at least one action.
(3) Regress:
(a) I am frequently free to pick up the phone.
 (b) I am frequently free to generate a desire to refrain from answering the phone. [a, 1]
(c) I am frequently free to generate a desire to refrain from generating a desire to refrain
from answering the phone. [b, 1]
(d) I am frequently free to generate a desire to refrain from generating a desire to refrain
from generating a desire to refrain from answering the phone. [c, 1]
And so on.
(4) I am frequently free to generate an infinity of desires. [2, 3]
…

Vander Laan argues that the result (4) is unacceptable (and hence that the regress is vicious by
(Vic-P)) for three reasons (2001: 208). First, one cannot be frequently free to generate an infinity
of desires as generating so many desires are beyond anyone’s capacity. Second, one cannot be
frequently free to generate an infinity of desires as they are increasingly complex. It is hard
ANALYSIS 113 to imagine even just the third-order desire described in (d), let alone to possess
it. Third, if generating desires takes time, and if this happens successively, then in order to
generate an infinity of desires one needs an infinite amount of time. As this is not available, one
cannot generate an infinity of desires. Hence, according to Vander Laan, the regress is thrice
vicious.

Furthermore, as (1) is to follow from more basic principles which are granted in the overall
discussion, Vander Laan infers that (2) must be rejected: It is not the case that we are frequently
free. And so did he establish his view: Restrictivism.50

Also consider another case: the regress argument against Cartesian scepticism by Wilson (2012):

Cartesian scepticism (Paradox instance)

(1) For any external or mental state x, if one might be deceived about x, then one is sceptical
about x.
(2) For any external or mental state x, one is sceptical about x only if one might be deceived
about one being sceptical about x.
(3) Regress:
(a) I might be deceived about there being an external world.
(b) I am sceptical about there being an external world. [a, 1]
(c) I might be deceived about my being sceptical about there being an external world. [b,
2]
(d) I am sceptical about my being sceptical about there being an external world. [c, 1] (e)
I might be deceived about my being sceptical about my being sceptical about there being
an external world. [d, 2]
And so on.
(4) I am both sceptical about something and, at the same time, sceptical about being sceptical
about that thing. [3]
…

Wilson argues that the result (4) is unacceptable (and so that the regress is vicious by (Vic-P))
because it is psychologically unstable. One cannot both at once be sceptical about something and
be sceptical about being sceptical about that thing. Suppose I am sceptical about there being an
external world. In that case, I claim that one can neither claim that there is an external world, nor
claim that there is not. Suppose I am sceptical about my being sceptical about there being an
external world. In that case, I claim that one can neither claim that I am sceptical about there
being an external world, nor claim that I am not sceptical about there being an external world.
Yet, for me to claim both at once that I am sceptical about there being an external world and to
claim that one cannot claim that I am sceptical about there being an external world is
psychologically unstable. Generally, each pair subsequent to (b)/(d) above suffers from the same
problem.

Furthermore, as the regress is vicious, Wilson rejects (1) which is to express Cartesian
scepticism: the possibility of deception should not be a reason for scepticism.

Yet, regresses are not always vicious. They are non-vicious exactly when they do not entail any
unacceptable result. That is, they are non-vicious when it is shown that each and every necessary
condition for the trigger situation is unproblematic.

Consider, for example, McTaggart’s attack on the existence of time. His reasoning rests on the
distinction between A and B-series of time: A-series are orderings on the basis of the temporal
properties being past, present and future. B-series are orderings on the basis of the temporal
relations earlier-than (or later-than) and simultaneously-with. Part of McTaggart’s argument here
is based on a regress, namely the part against the A-theory of time, i.e. the view that posits only
A-series. According to McTaggart, any such theory generates a regress:

If we avoid the incompatibility of the three characteristics by asserting that M is present,

has been future, and will be past, we are constructing a second A-series, within which the
first falls, in the same way in which events fall within the first. […] The second A-series
will suffer from the same difficulty as the first, which can only be removed by placing it
inside a third A-series. The same principle will place the third inside a fourth, and so on
without end. (1908: 469)

Now, against McTaggart’s assessment, this regress may be classified as nonvicious for the A-
theorist so long as an infinity of A-series is not regarded as absurd, impossible, uneconomical, or
unacceptable in any other way (for an argument along these lines, cf. Smith 1986). In §5.5, I will
present one extended example that can plausibly be regarded as non-vicious in the sense of (Vic-
P).

2.2. Failure viciousness

The criterion for the Failure Schemas:

(Vic-F) For any regress R and theory T, R is vicious for T iff (i) T is committed to the
solution S by which R is generated, and (ii) R entails that S fails.

Again, some clarifications are in order. First, this criterion looks more specific than (Vic-P). But
this is only apparent. It is true that (Vic-P) can apply to all case studies reconstructed Paradox-
wise, but it is equally true that (Vic-F) can apply to all case studies reconstructed Failure-wise.51
Second, suppose there is a theory which satisfies clauses (i) and (ii). In that case, T is committed
to a regressive solution that fails, and this is why the regress is bad for T. Consider the following
case:

Thus, if the regress is vicious, it is vicious because it prevents Resemblance Nominalism

from accomplishing its explanatory project of accounting for all properties in terms of
resembling particulars. (Rodriguez-Pereyra 2002: 108)

This example can be spelled out as follows:

(1) You have to account for all properties/relations.

(2) For any property/relation x, if you have to account for x, then you appeal to the resembling
particulars which have x.
…
(C) If you appeal to resembling particulars every time you have to account for a
property/relation, then you will never account for all properties/relations. [1-5]

Resemblance Nominalism is committed to the solution specified in (2), and so this theory entails
that the problem specified in (1) is never solved. The regress is bad exactly because it plays a
crucial role in this entailment. Rodriguez-Pereyra has put this nicely: the regress is bad because it
prevents the theory from solving the given problem. This is a well-suited term for Failure
regresses: at each step yet another problem needs be solved (in order for the initial one to be
solved), and this prevents the initial problem from ever being solved.

Importantly, solutions might fail in different ways: they might fail to solve all problems of a
certain kind (‘fail’ in the sense of Failure I), or fail to solve any problem of a certain kind (‘fail’
in the sense of Failure II). When it comes to theory choice, this difference matters: Failure II
failures are worse. Suppose there are two theories, T1 and T2, that both have to accomplish the
same universally quantified task, but that T1 fails in the Failure I-way while T2 fails in the
Failure II-way. That is to say: T1 solves lots of problems of a given type, yet never all of them,
whereas T2 never solves even one of them. Clearly T1 is to be preferred over T2 in this case.
Both fail in a certain way (which is why both kinds of regresses are vicious by (Vic-F)), but T2’s
failure is worse.52

Again, by (Vic-F) regresses are not always vicious. They are nonvicious exactly when they do
not entail that the given solution fails. Yet nonvicious regresses are here somewhat more
complicated than those that turn out non-vicious by (Vic-P). The reason is that, within the
Failure Schema, once a regress is entailed, you immediately obtain a failure: no further,
substantial premises are needed to get you there.

Still, in some selected cases the failure does not follow because one of the suppressed premises
(6) or (7) fails. As this is not easy to see, I will spell out some examples below. Specifically,
there will be three kinds of scenarios where the solution’s failure is not established by the regress
(such that the regress is non-vicious by (Vic-F)). Namely, this will be the case whenever all the
problems described in the regress are in fact identical, whenever they are different but still can
be solved at once, or whenever they tend to become less and less important. Let me explain these
in turn.

2.2.1. Exception (1)

Consider, for illustration, Frege’s regress argument against the definability of truth. Here is a
Failure II reconstruction:

Truth (Failure II instance)

(1) You have to decide for at least one proposition whether it is true.

(2) For all propositions x, if you have to decide whether x is true, then you decide whether x
corresponds with reality.
(3) For all propositions x, if you decide whether x corresponds with reality, then you decide
whether x is true only if you decide first whether the proposition that x corresponds with reality
is true.
(4) Regress:
(a) You have to decide whether the proposition that I am mortal (p1) is true.
(b) You decide whether p1 corresponds with reality. [a, 2]
(c) You have to decide first whether the proposition that p1 corresponds with reality (p2)
is true. [b, 3]
(d) You decide whether p2 corresponds with reality. [c, 2]
(e) You have to decide first whether the proposition that p2 corresponds with reality (p3)
is true. [d, 3]
And so on.
(5) For all propositions x, you first have to decide whether a regress of propositions is true in
order to decide whether x is true. [4]
(C) You will never decide whether any proposition is true if you decide whether it corresponds
with reality every time you have to decide whether a proposition is true. [1-5]

Now suppose that there is a certain conception of propositions according to which p1, p2, p3, etc.
are identical. The proposition that I am mortal (i.e. p1) is identical to the proposition that p1
corresponds with reality. Likewise, both are identical to the proposition that [the proposition that
p1 corresponds with reality] corresponds with reality. And so on. In that case, the problems (a),
(c), (e), etc. in the regress above are identical as well. Whether the problems are indeed identical
depends on one’s ideas about propositions. However, the main point is that if they are identical,
then the regress reduces to one task only and, in that case, it does not follow that you will never
decide whether the proposition that I am mortal is true (for example).

Counterexamples of this sort can be blocked by explicitly assuming (i.e. in the

premises/hypotheses that generate the regress) that the problems are distinct. For example, in the
Failure I case the instance of the premise (3) should exactly be an instance of: ‘For all Ks x, if
you ψ x, then there is a new item y of type K’, where ‘new’ means ‘distinct from any other item
already encountered in the regress’.
A similar analysis can be applied to Danto’s case and Zeno’s Dichotomy Paradox which concern
the following series of tasks (I provide one more extended example in §5.2):53

(a) Perform action a (e.g. write a dissertation).

(c) Perform action b by which a is performed (e.g. write a chapter).
(e) Perform action c by which b is performed (e.g. write a section).
etc.

(a) Traverse a distance (e.g. 1 meter).

(c) Traverse half the initial distance (0.5 meters).
(e) Traverse half of half the initial distance (0.25 meters).
etc.

The tasks in each of these cases are not wholly identical, but in each the new tasks are already
part of the initial one. In Zeno’s case, each new distance to be traversed is part of the initial
distance to be traversed. In Danto’s case, each new action to be performed is part of the initial
action. There may be instances of the Danto series where this is not the case. However, the main
point is that if the tasks generated in the regress are partly identical to the initial task, then again
no failure follows from the regress. For if you carry out the initial task, then you accomplish
along with it, for free as it were, the whole regress of partly identical tasks.54

2.2.2. Exception (2)

The next exception: The solution’s failure follows only if all problems described in the regress
cannot be solved at once. Take Frege’s case again. Let us assume this time that all problems are
different (i.e. that the relevant propositions are distinct), but can still be solved at once with the
aid of the following equivalence schema:

For any instance of p: ‘p’ is true iff ‘‘p’ corresponds with reality’ is true.

For example, if you decide whether ‘I am mortal’ is true, then by this schema you can at once
decide whether ‘‘I am mortal’ corresponds with reality’, ‘‘‘I am mortal’ corresponds with reality’
corresponds with reality’, etc. are true. If that is right, then it again does not follow from the
regress that you will never decide whether ‘I am mortal’ is true. Compare Dummett on a slightly
different case:

[…] there is no harm in this, as long as we recognize that the truth of every statement in
this series is determined simultaneously: the regress would be vicious only if it were
supposed that, in order to determine the truth of any member of the series, I had first to
determine that of the next term in the series. (1973: 443, cf. Künne 2003: 131)

It remains to be seen, of course, how one can motivate such an equivalence schema, but the
general point is that for Failure arguments to work the problems described in the regress should
not be solvable at once. To account for counterexamples of this sort one needs to be explicit in
the extra premise that each problem generated in the regress depends on the next, and not vice
versa. Consider line (3) in Frege’s case:
- For all propositions x, if you decide whether x corresponds with reality, then you first have to
decide whether the proposition that x corresponds with reality is true in order to decide whether x
is true.

The role of ‘first’ here is to indicate that the problem of deciding whether x is true depends upon
the problem of deciding whether the proposition that x corresponds with reality is true, and not
vice versa. Without ‘first’, there would not be such an order and all the problems might well be
solvable at once. It is worth noting again that the term ‘first’ indicates an instrumental order,
rather than a temporal order. For example, it need not be the case that the problem of deciding
whether the proposition that x corresponds with reality is true needs be solved earlier in time.
What matters is the asymmetry between the problems: Deciding whether the proposition that x
corresponds with reality is true is to be a necessary means to deciding whether x is true, and not
the other way around.

Unfortunately, the same solution does not work for Failure I arguments, as the latter do not use
the term ‘first’. This means that regresses of problems may be generated Failure I-wise that are
solvable at once (such that no failure follows even though there is a regress). Take Frege’s case
once again. We have just seen that the Failure II reconstruction might be blocked. However, we
could still set up a Failure I argument which concludes: You will never decide whether all
propositions are true (that is, if you decide whether a proposition corresponds with reality every
time you have to decide whether it is true). Yet this conclusion fails as well if all problems in the
regress are solvable at once. As we shall see below, in such exceptional cases the conclusion is
false, not because Failure I arguments are invalid, but because the suppressed premise (6) is
false.

2.2.3. Exception (3)

The last exception: The solution’s failure follows only if all problems described in the regress do
not become less and less important. Specifically, there are distinct two places where Failure
arguments might go wrong if the problems described in the regress become less and less
important, namely at the inference of (5) from the regress (i.e. that there are always further
problems to be solved), and also at the inference of the failure from (5) (i.e. that a failure follows
if there are always further problems to be solved). Let me explain these in turn. From (4) to (5)
Take Sextus’ case, which concludes that no proposition is ever justified to you if you provide
reasons after reasons. The argument rests on the assumption that no matter how many reasons are
added to a chain of reasons, the justification of the initial proposition will still depend on any of
the last added reasons (i.e. will still depend on the last generated problems). If this were not the
case, and if after a certain point the justification of the initial proposition were to remain
unaffected (and the further generated problems irrelevant), then the Failure conclusion might fail
to follow.

I take it that Peijnenburg (2007, 2010) and Peijnenburg & Atkinson (2008) demonstrate exactly
this. They start from the assumption that justification is no all-or-nothing affair (such that a
proposition is either flat-out justified to someone, or not), but rather can come in degrees such
that propositions can be more or less justified or probable (i.e. relative to certain And So On 120
reasons). On the basis of this assumption they show that justification can be gained on the basis
of infinite regresses of reasons. Namely, it can be gained in such a way because, after a certain
point in an infinite chain, the reasons no longer make a significant difference to the initial
proposition. As Peijnenburg puts it:

In general it is the case that, as the chain becomes longer, the justification provided by the
conditional probabilities increases, while the justification given by the source of the chain
decreases. And as the n of En grows larger and larger […] the influence of P(En) on
P(E0) diminishes with each link, until it finally vanishes entirely. (2010: 1131-2)

In terms of the Failure Schema, the extra premise (3) seems granted (in order to generate the
regress), namely that if you provide a reason y for a proposition x, then you first have to justify y
in order to justify x. However, doubt is casted on the repetition of this step in the inference of
line (5), which reads that there is always (no matter how far you go) a new proposition that you
have to justify first, i.e. in order to justify any initial proposition. Put more generally: doubt is
casted on the idea that no matter how many problems in the regress are solved, the initial
problem will still depend on any single one of them. Moreover, if (5) fails to follow, then the
remainder of the argument (i.e. that you will never justify any proposition) fails as well.

Yet, if the regress follows from the premises/hypothesis and (5) fails to follow from the regress,
then does not this pose a problem for the logical validity of Failure II arguments? It does not. For
recall that the step from the regress to (5) relies on suppressed premise (7) (see §2.3.2):

• If for all Ks x, you first have to φ a new K in order to φ x, then you first have to φ a regress of
new Ks in order to φ x.

In this case: If for all propositions x, you first have to justify a new proposition in order to justify
x, then you first have to justify a regress of new propositions in order to φ x (and this implies, as
we shall see in §4.2.3, that you have to justify an infinity of new propositions in order to justify a
single one). Thus, if in a certain case (5) fails to follow from the regress, then this does not mean
that the step is logically invalid, but that the instance of (7) is false.

From (5) to (C)

Suppose (5) does follow: you always have to solve yet further problems in order to solve any/all
problems of a given kind. Does it always follow that you fail in these circumstances? Certain
supertasks suggest that this is not the case. It is instructive, here, to consider Zeno’s Dichotomy
Paradox and his Achilles and the Tortoise Paradox (i.e. two cases that are usually regarded as
supertasks) in Failure format:

Dichotomy (Failure II instance)

(1) Achilles has to traverse at least one distance to the Tortoise.

(2) For all distances to the Tortoise x, if Achilles has to traverse x, then he runs x.
(3) For all distances to the Tortoise x, if Achilles runs x, then he first has to traverse the half of x,
namely distance y, in order to traverse x.
(4) Regress: […] [2-3]
(5) For all distances to the Tortoise x, Achilles first has to traverse a regress of distances in order
to traverse x. [4]
(C) If Achilles runs all distances to the Tortoise that he has to traverse, then he will never
traverse any distance the Tortoise. [1-5]

Achilles and the Tortoise (Failure I instance)

(1) Achilles has to traverse all distances to the Tortoise.

(2) For all distances to the Tortoise x, if Achilles has to traverse x, then he runs x.
(3) For all distances to the Tortoise x, if Achilles runs x, then there is a new distance y to the
Tortoise.
(4) Regress: […] [1-3]
(5) For all distances to the Tortoise x, Achilles always has to traverse yet another distance in
addition to traversing x. [4]
(C) If Achilles runs all distances to the Tortoise that he has to traverse, then he will never
traverse all distances the Tortoise. [1-5]

These conclusions are no doubt counterintuitive. Why would Achilles fail to catch the Tortoise
by running to the latter? Achilles is much faster and will certainly catch the Tortoise eventually.
However, in both cases (5) does seem to follow from the regress: Achilles always has to traverse
yet another distance (modulo a worry in the Dichotomy case that I identified above: see
exception 1). But if (5) is true, and (C) false, then the step from (5) to (C) is invalid.

Do these kinds of supertasks pose any problem for the logical validity of Failure arguments
generally? Fortunately not. Recall that in both Failure Schemas, the step from (5) to (C) relies on
a suppressed premise, namely line (6) (see §2.3.2). In Zeno’s cases, the relevant instances are the
following:

• If Achilles has to traverse at least one distance to the Tortoise, and if he first has to traverse a
regress of new distances in order to And So On 122 traverse any distance, then he will never
traverse any distance to the Tortoise.

• If Achilles has to traverse at least one distance to the Tortoise, and if he always has to traverse a
new distance in addition to traversing any distance, then he will never traverse all distances to
the Tortoise.

I would like to suggest, now, that these lines fail to hold in Zeno’s cases. Generally, if (certain)
supertasks are possible, then this does not seem to form a problem for Failure inferences, but it
does cast doubt on the general truth of these suppressed premises.

There might furthermore be an explanation of why instances of the suppressed (6) may fail in the
case of supertasks. As noted, Failure arguments do not operate with a notion of time. For
example, in the cases just set out Achilles will never reach his goal (in the Failure I or II
meaning) in the sense that, given the regress, there are always further tasks to be accomplished to
reach that goal. This holds whether or not those tasks take up time, and whether or not they take
up the same amount of time (if they take up any time at all). For most regress arguments
discussed in this dissertation this nontemporal kind of reasoning is well-suited, but this is not the
case for supertasks: time plays a crucial role in the latter. Compare the following encyclopedia
definition:

A supertask may be defined as an infinite sequence of actions or operations carried out in

a finite interval of time. (Laraudogoitia 1999: §1.1)

The same holds for Zeno’s cases. Achilles has to traverse the infinite number of distances within
a finite interval of time. The Failure reconstructions just given are completely silent about that.
Also, in these reconstructions nothing is said about Achilles’ speed and that he will traverse the
distances to the Tortoise in less and less amount of time (such that, in this sense, the problems
generated in the regress become less and less important). Now, if this point is correct, then this
might explain why (6) does not hold in the case of supertasks. Thus, to account for
counterexamples to the step from (5) to (C) is to defend (6) (whenever possible). In §6.1.2, I will
say some more about the connection between supertasks and Failure regresses.

To sum up, in order for a regress to be vicious by (Vic-F) and to entail a failure, the problems in
it should not be identical, not be solvable at once, and not become less and less important.

3. Infinity

Must regresses be infinite to be vicious? Again, this depends on the schema in which the
regresses are generated and on what, in this context, is meant by ‘viciousness’.

3.1. Paradox infinity

All Paradox regresses are infinite, as the regress formulas of the Paradox Schema have infinite
consequences. That is, in the case of Paradox regresses it holds that for any necessary condition
generated in the regress there is another necessary condition. For any K that is F there is another
K, not previously mentioned in the regress, that is F. Or again: it is not the case that for one K
that is F there is not a further, new K that is F.

However: must Paradox regresses be infinite to entail an unacceptable result? Consider the
following example:

(a) Dispute d1 is settled.

(b) d1 is settled by a criterion c1.
(c) Dispute d2 about c1 is settled.
(d) d2 is settled by a criterion c2.
(e) Dispute d3 about c2 is settled.
etc.

This regress entails an unacceptable result as soon as it violates the number of criteria that are in
fact available. This easily shows that a Paradox regress need not be infinite to entail an
unacceptable result, and so to be vicious (by (Vic-P)).
However, the fact that Paradox regresses need not be infinite to entail unacceptable results does
not mean that they cannot entail unacceptable results thanks to their infinity. For example,
paradoxes of infinity can only be worrisome for an infinite regress, not for a finite one. Here is
one example of such a paradox (in terms of the beer regress from §1.2): How is it possible that
when two persons both drink an infinity of beers, it could still be the case that one of them drinks
more than the other? Here is a second one (a variant on Hilbert’s famous Hotel): How is it
possible that when you have an infinite number of glasses and fill all of them with beer, it could
still be the case that there are empty glasses?55

3.2. Failure infinity

All Failure regresses, too, are infinite, as the considered problem and solution (in combination
with an extra premise) have infinite consequences. Namely, in the case of Failure regresses it
holds that for any problem generated in the regress there is a solution, and for any generated
solution there is another problem to be solved. Or again: it is not the case that for some problem
generated in the regress there is no solution, and it is not the case that for some solution there is
no new problem to be solved.

Now: must Failure regresses be infinite to entail that a given solution fails? Consider the case
mentioned earlier in Failure-version:

(a) You have to settle dispute d1.

(b) You invoke a criterion c1 to settle d1.
(c) You (first) have to settle dispute d2 about c1.
(d) You invoke a criterion c2 to settle d2.
(e) You (first) have to settle dispute d3 about c2.
etc.

Suppose this is a Failure I regress. Must this regress be infinite in order to entail that the
considered solution to the problem of settling all disputes fails? On the one hand, it seems the
regress need not be infinite, for it becomes clear after just a few steps that the problem will never
be solved. At (c), another instance of the general problem (i.e. of settling all disputes) is entailed
by the solution at (b). At (e), it turns out that a similar solution (d) will not change this situation
and generates yet another instance of the general problem.

However, there is a counterexample to this thought. Suppose you would solve the problem as
stated in (e) in a different way. For example, suppose you settle dispute d3 by rhetorical tricks,
rather than on the basis of a new criterion. In that scenario, no new dispute is entailed, and your
problem of settling all disputes is solved. This suggests that the Failure I regress has to be infinite
in order to entail a failure, for counterexamples like this would be blocked if the regress were
infinite (and a similar analysis applies to Failure II regresses).

Still, I think there is a clear sense in which Failure regresses need not be infinite (i.e. in order to
entail a failure) as well as a clear sense in which Failure regresses need be infinite. On the one
hand, Failure regresses need be infinite in the sense that for any problem there has to be a new
problem that must be dealt with (e.g. that for any dispute that is to be settled there is a new
dispute that must be dealt with). For this is exactly why failures follows from such regresses:
solutions that fall prey to Failure regresses never succeed in solving the given problem because
at no point in the regress will it be the case that there is no further problem that must be solved in
order to succeed.

On the other hand, (a), (b), (c), etc. are arbitrary instances, that is, the reasoning works no matter
the dispute one starts with. So if it follows from (a)-(c) that you have settle a new dispute d2 if
you have to settle d1, then by Universal Generalisation we may immediately conclude that you
will have to settle a further dispute for any dispute that you have to settle. This suffices to obtain
a failure,56 and indeed shows that Failure regresses need not be infinite (i.e. in order to obtain a
failure, and be vicious by (Vic-F)) in this second sense.

Still, I have three qualifications. First, one might think that the solution fails as soon as the tasks
exceed one’s abilities. This might have all sorts of reasons: you might be unable to handle more
than a certain number of beers given your physiological condition, or drinking more than a
certain amount of beers might be too expensive, or might take too much time, etc. Any obstacle
would do (practical or theoretical) that impedes your ability to complete the newly generated
tasks. The point of Failure arguments, however, is precisely that you will never solve the initial
problem (e.g. never quench all of your thirsty feelings) even if there are no such obstacles (see
§2.3.2).

Second, something must be clarified here about ‘never’. We may distinguish between a temporal
and structural reading of this term:

• At no point in time do you solve the problem.

• At no point in the regress do you solve the problem.

The first reading assumes that the solutions take time, and must precede one another temporally.
The second reading does not, which makes the latter weaker than the former. But is the stronger,
first reading needed? Yalden-Thomson, for example, seems to suggest in the following that it is:

If, however, a proponent of Schools of Education were to argue that in order to qualify to
teach, someone must teach you how to teach and he, in turn, must be taught by someone
how to teach to teach to teach, etc., then a logically vicious regress is created: owing to
the temporal factor, no one could ever teach. (1964: 509, cf. Rescher 2010: 23-7)

Here is a version of the example (just one version). Suppose you want to learn how to teach. As a
solution, you consult an unqualified person who promises to teach you how to teach. Of course,
this person, as she is unqualified, teaches you how to teach only if she herself learns how to teach
how to teach. As a solution, you consult an unqualified person who is going to teach her how to
teach how to teach. Regress. The conclusion: You will never learn how to teach if you look
among unqualified teachers for anyone who has to learn how to teach.
The question is, then: Does it matter that consulting people takes time? Suppose for the sake of
the issue that it does not take time (or that, if it does, an infinite amount of time is available).
Suppose too that there is an infinity of unqualified persons around. The conclusion still follows:
You will never learn how to teach if you look among unqualified teachers for anyone who has to
learn how to teach, where ‘never’ means ‘at no point in the regress as governed by lines (1)-(3)
of the Failure Schema’. Even if the solutions take no time at all, at no point in the regress will it
be the case that a solution (of consulting an unqualified person) entails no new, particular
problem (of teaching an unqualified person how to teach) which must be solved in order for the
initial one to be solved (i.e. the problem of my learning how to teach). If this is right, then
regress arguments are not committed to the temporal reading of ‘never’.57

Last qualification: this infinity issue got some attention in the literature, and I shall examine
some discussion points in §6.1.58

4. Circularity

Regress arguments are sometimes accompanied by circularity arguments. Well-known examples

include Sextus’ Problem of the Criterion, Boghossian’s Problem of Relativism, and Hume’s
Problem of Induction:

Can a warranting, contingent fact be justified inductively? In the case of a formal theory
of induction, this horn of the dilemma yielded a circularity or an intolerable infinite
regress. (Norton 2013)

However: regresses are not circularities. To explain the difference between regresses and
circularities, let us consider the Problem of the Criterion (as construed by Amico 1993: 35-6).
The scenario is that you have to decide whether a certain proposition p is true. You can do this
critically, i.e. by a proof, or uncritically. If you do it uncritically, then your decision is arbitrary
and will be discredited. But if you do it critically and use a criterion c1 to decide whether p is
true, you need first to decide whether c1 correctly rules what is true and what is not. Again, there
are two options: you can do this critically, or not. If the latter, your decision will be discredited.
So you do it critically and have two options.

Option 1: You prove that c1 correctly determines what is true and what is not by showing that it
gives the right results. In this case, you already know what is true and what is not (and hence
whether p is true or not). But this is impossible, because we started from the situation where you
still have to decide whether proposition p is true. This is a circularity.

Option 2: You prove that c1 correctly determines what is true and what is not by appealing to a
meta-criterion c2 able correctly to determine what criteria correctly determine what is true and
what is not. But now you need first to decide whether c2 correctly determines the correct criteria.
Again, there are two options: you can do this critically, or not. If the latter, your decision will be
discredited after all. So you do it critically and have two options. Either you prove that c2
correctly determines what is true and what is not by showing that it gives the right results, which
again is a circularity. Or you prove that c2 correctly determines what is true and what is not by
appealing to yet another meta-criterion, c3. This is the regress.
dispute p -> use c1 -> dispute c1 -> use p (circularity)
|
v
use c2 -> dispute c2 -> use c1 (circularity)
|
v
use c3 -> … (regress)

Figure 4: Regress vs. circularity

What, then, is the difference between a circularity and a regress? There are at least two important
differences. First, the two have a slightly different structure. Circularities follow the following
pattern. There is a problem (a): Define term x. Explain fact x. Argue for conclusion x. (a) is to be
solved by a solution (b): Use x in the definition, explanation or argument. (b) does not work
unless (a) is already solved: One cannot use x unless one has already defined, explained or
argued for it. As we can see, this basic pattern works alike for circular definitions, circular
explanations and circular arguments.

The Problem of the Criterion is somewhat more complicated as it uses two extra steps: There is a
problem (a): Determine whether p is true. (a) is to be solved by a solution (b): Use c1. (b) does
not work unless a problem (c) similar to (a) is solved: One cannot use c1 unless one determines
whether c1 is correct. (c) is to be solved by a solution (d): Use p. (d) does not work unless (a) is
already solved: One cannot use p unless one determines whether p is true. But the pattern is the
same. In general, a solution which falls prey to a circularity fails exactly because sooner or later
it implies that the problem it is meant to solve is already solved. But as this solution was meant
to be a solution for that very problem, the latter cannot already be solved, and so the solution
fails.

Regresses, on the other hand, follow a slightly different pattern. There is a problem (a):
Determine whether p is true. (a) is to be solved by a solution (b): Use c1. (b) does not work
unless a problem (c) similar to (a) is solved: One cannot use c1 unless one determines whether c1
is correct. (c) is to be solved by a solution (d): Use c2. (d) does not work unless a problem (e)
similar to (c) is solved: One cannot use c2 unless one determines whether c2 is correct. And so
on.

The main structural difference between circularities and regresses, therefore, is the following. If
a solution is circular, then it does not work unless the very same problem it attempts to solve is
already solved. If a solution is regressive, in contrast, then it does not work unless problems
similar to the one it attempts to solve are already solved.

The second significant difference between circularities and regresses is that the former are
problems, whereas the latter are problem/solution pairs. More precisely, a circularity is a
problem, and a regress (at least as understood within the Failure framework) is a never ending
attempt to get rid of similar circularities (cf. Figure 4; cf. also Barnes 1990: 214, 216).59
§5 Applications

In the following I show in more detail how the regress argument schemas can be used to clarify
six important debates: the ethics of belief debate, the debate on rationality and the role played in
it by Carroll’s Tortoise, the debate on Epistemic Infinitism, the Sorensen/Sider controversy over
the Access principle, Russell’s defence of relations, and finally the controversy between Quine
and Davidson concerning reference. I start by making explicit three sorts of general fallacies that
may occur in any debate which centres on a regress argument.

5.1. Fallacies

In the following I will show how the schemas distinguished in this dissertation can be used to
clarify existing debates centred on regress arguments. Specifically, I will show that there are
three main points that may cause confusion. First, two parties may disagree on whether there is a
vicious regress in the first place. Or, if they agree on this, they may disagree on the form the
argument takes, i.e. whether it has a Paradox or Failure structure. Last, if they agree that it is a
Failure argument, they may disagree on its strength, i.e. whether it has a Failure I or Failure II
structure. These three points correspond to the following fallacies:

No Vicious Regress Fallacy

When you criticise a view as if it were to generate a vicious regress (whether Paradox or Failure-
wise), whereas your opponent need not accept this.

Paradox/Failure Fallacy
When you criticise a regress argument taken Paradox-wise, whereas your opponent takes it
Failure-wise. Or vice versa.

Failure I/II Fallacy

When you criticise a regress argument taken Failure II-wise, whereas your opponent takes it
Failure I-wise. Or vice versa.

One significant qualification is in order. According to the Charity rule, it is no mistake to take
your opponent’s argument in its strongest form, whatever the argument may be. To do this,
indeed, is exactly what Charity motivates. In such a case, however, you might at least make it
explicit that you are changing the argument. We shall consider the prospects of Charity soon
(§6.2).

The first fallacy has been made explicit earlier in the literature on regress arguments (cf. Nolan
2001, Maurin 2007), and the latter two are my own contribution. I have implicitly stressed and
provided illustrations of these three sorts of fallacies throughout this dissertation. In the current
part we shall consider six examples explicitly and in more detail. For example, we shall see that
applications of Carroll’s Tortoise commit the No Vicious Regress Fallacy, that Russell’s defense
of relations incorporates and ambiguity between Paradox/Failure, and that the controversy
between Quine and Davidson on reference derives from a Failure I/II confusion (the latter were
announced in §1.3.2). In each case the controversies or ambiguities can be resolved by showing
that the incompatible claims rest on different arguments, and that sometimes they are not
incompatible after all. Indeed, if there is to be a debate about anything, all interested parties have
to agree on the argument that is at issue.

5.2. Regressive Pragmatism

Should the practical value of a belief enter into the evaluation of its rationality? Pragmatists say
Yes, Evidentialists say No.60 Consider the belief that this dissertation will be accepted by the
reading committee, and suppose that this belief makes me really happy. Moreover, suppose for
the sake of the thought experiment that the acceptance of this dissertation will make the whole
world a better place (plenty of nice drinks and music, etc.). Should I adopt this belief, in order to
be rational, even if its epistemic value is really quite low (i.e. even if I have no reason to think it
will be accepted)?

Suppose that the practical value (PV) of a belief should indeed enter into the evaluation of its
rationality, i.e. on top of its epistemic value (EV). This means that I should believe something iff
the sum of its evidential credentials plus its expected utility outweighs the sum of the evidential
credentials plus the expected utility of not having the belief in question:

(R) For any belief x, I should believe x iff EV(Bx) + PV(Bx) > EV(¬Bx) + PV(¬Bx).

According to this Pragmatist principle, I should believe that this dissertation will be accepted,
given the scenario I sketched above. Yet, as Zemach (1997) showed, proponents of (R) have to
deal with a regress argument. Here is my Failure II reconstruction:

Pragmatism (Failure II instance)

(1) For at least one belief x, I have to decide whether to hold x.

(2) For any belief x, if I have to decide whether to hold x, then I apply (R) to x.
(3) For any belief x, if I apply (R) to x, then I first have to decide whether to believe that PV(Bx)
> PV(¬Bx) in order to decide whether to hold x.
(4) Regress:
(a) I have to decide whether to believe p1, i.e. that this dissertation will be accepted.
(b) I apply (R) to p1. [a, 2]
(c) I have to decide first whether to believe p2, i.e. that PV(Bp1) > PV(¬Bp1). [b, 3]
(d) I apply (R) to p2. [c, 2]
(e) I have to decide first whether to believe p3, i.e. that PV(Bp2) > PV(¬Bp2). [d, 3]
And so on.
(5) For any belief x, I first have to decide whether to hold a regress of beliefs in order to decide
whether to hold x. [4]
(C) I will never decide whether to believe that this dissertation will be accepted (or anything
else) if I apply (R) to a belief x every time I have to decide whether to hold x. [1-5]

This is a significant result, as it constitutes a serious problem for Pragmatism. Weintraub states
the moral of this argument as follows:
 Practical considerations […] cannot be invoked. Or rather, if they are, no prescription can
ever be engendered. (2001: 63)

Despite its significance, this result is not often discussed (exceptions are Weintraub and Percival,
considered below). Moreover, there seems to be a serious problem with it: for one might readily
design a similar argument against Zemach’s own Evidentialism, i.e. against the view that does
not accept practical reasons for belief (a view according to which, therefore, I should not believe
that this dissertation will be accepted, given the circumstances). Zemach himself seems aware of
this possibility (1997: 526- 7), yet quickly sets it aside. Weintraub also considers this option, and
concludes, by contrast, that even if one can block the argument against Pragmatism one cannot
block the analogous argument against Evidentialism (2001: 65). So let us try to make things
more precise, and identify the exact dispute.

There are two main structural differences between the initial argument against Pragmatism and
the analogous argument against Evidentialism. First, the following Evidentialist principle
substitutes for (R):

(R*) For any belief x, I should believe x iff EV(Bx) > EV(¬Bx).

Second, the following line substitutes for (3):

(3*) For any belief x, if I apply (R*) to x, then I first have to decide whether to believe that
EV(Bx) > EV(¬Bx) in order to decide whether to believe x.

Again a regress is generated, and I have to decide whether to believe any member of the
following series:

p1 This dissertation will be accepted.

p2 EV(Bp1) > EV(¬Bp1).
P3 EV(Bp2) > EV(¬Bp2).
etc.

Again, the conclusion of the argument is that I fail to decide whether to believe anything. Now
the question is: Why should this argument against Evidentialism fail whereas the initial one
against Pragmatism does not? What is the difference?

Evidentialism’s problem is basically this: that in order to apply (R*), one needs already to hold a
belief that certain facts about evidential credentials obtain. Yet, whether one should adopt any
such belief itself depends, according to Evidentialism, on the belief’s epistemic value. What is
the epistemic value of the belief that EV(Bp1) > EV(¬Bp1)?

Well, its epistemic value is the same as the epistemic value of p1, whatever that is. For to say
that I think it is unlikely that [this dissertation will be accepted] (i.e. p1) is just to say that I think
it is unlikely that [the epistemic value of p1 is greater than the epistemic value of ¬p1] (i.e. p2)
(and indeed that it is unlikely that [the epistemic value of p2 is greater than the epistemic value
of ¬p2], and so on). All problems generated in the Evidentialist’s regress, therefore, seem to
reduce to one problem (i.e. to one decision that is to be made), and so cause no worry for
Evidentialism.

Matters are different for Pragmatism. In order to apply (R), one needs already to hold a belief
that certain facts about expected utility obtain. Yet, whether one should adopt any such belief
itself depends, according to Pragmatism, on the belief’s practical value (at least in part). What is
the practical value of the belief that PV(Bp1) > PV(¬Bp1)?

Unfortunately, this cannot be determined on the basis of the practical value of p1. It might be
likely for me that the practical value of believing that this dissertation will be accepted is greater
than the practical value of not believing this, if I am fully aware of the circumstances. Yet,
likelihood concerns epistemic value, not practical value. Who knows the expected utility of the
belief that PV(Bp1) > PV(¬Bp1)? Will this belief make me happy? Will it make the world a
better place? These matters should be determined separately, and so on into the regress.

Hence it could be shown that only Zemach’s initial argument against Pragmatism holds water.
The parallel argument against Evidentialism (endorsed by Weintraub) fails because of
considerations identified in §4.2.2: For any Failure argument to go through, the regress at hand
should consist of problems that are distinct.

To be sure, Pragmatism still has some options. First, if Pragmatism applies only to ordinary
beliefs, and not beliefs about expected utility, then the regress could be blocked. Still, though, in
that case Pragmatism would need to find some motivation for why line (2) of the regress
argument does not hold in its unrestricted version, which seems no trivial task (cf. Zemach 1997:
527, Weintraub 2001: 64, Percival 2011: 139-40).

Second, one could object that Pragmatism is not correctly captured by line (2). The latter states
that in deciding whether to hold a belief, one needs to apply a rule to it, i.e. the rule (R), and do
so consciously at that. Yet as Percival (2002: 139) argues, it is also possible for Pragmatism to
say that in deciding whether to hold a belief one need merely proceed in accordance with (R),
and need not consciously apply the rule. By this proposal, then, it would be rational for me to
adopt the belief that this dissertation will be accepted, as this accords with (R) even if I cannot
determine that I am rational in this case. In my opinion this is a drawback, but it is a second way
in which Pragmatism can resist the regress.

5.3. Carroll’s Tortoise

Suppose I intend to finish this dissertation today, that I believe this requires me to stay home
tonight, and yet that I refuse to intend to stay home (while sticking to my initial intention and
belief). I am being irrational. To be rational is to have at least a consistent set of propositional
attitudes (beliefs, intentions, etc.). Yet it is not wholly clear that my three attitudes here are
inconsistent. Some extra story must be supplied to clarify what sorts of attitudes cannot be
combined on pain of irrationality. The question is what licences the step from (A) and (B) to (Z):

(A) I intend to finish this dissertation today.

(B) I believe that this requires me to stay home.
(Z) I ought to intend to stay home.

Here is another example adapted from Blackburn (1995: 708):

(A) The Tortoise intends to get the lettuce.

(B) The Tortoise believes that getting the lettuce requires her to move to the other side of the
street.
(Z) The Tortoise ought to intend to move to the other side of the street.

Why should I and the Tortoise accept (Z) given (A) and (B)? There are two classes of possible
solutions. Either we have to accept (Z) because of an extra premise (which specifies that we
should take the supposed necessary means to our intended ends), or because of a rule which
takes us directly from (A) and (B) to (Z) (cf. Schwartz 2010: 89-90). There are, related to this,
various kinds of regress worries, and in the following I will focus in particular on one regress
argument from the literature meant to demonstrate that a variant of the rule solution fails. This
solution, specifically, which I will call the ‘Internal Rule’ solution, states that I and the Tortoise
must accept (Z) on the basis of (A) and (B) along with our pro-attitude towards the rule of
instrumental rationality, stated as follows:

(R) S intends to φ; S believes that φ-ing requires S to ψ ⊢ S ought to intend to ψ.

The status of this rule is meant the same as that of Modus Ponens and the other familiar rules of
inference. By (R), (Z) follows logically from (A) and (B). Furthermore, the Internal Rule
solution proposes that (R) is not in force unless the person who has the intention in (A) and belief
in (B) also has a pro-attitude towards (R). For me (or for the Tortoise) to have a pro-attitude
towards this rule means not only that I reason in accordance with (R) (which might be
coincidental or just a regularity), but that I let my reasoning be governed by (R), i.e. that I have a
desire to comply with (R) and to apply the rule to (A) and (B).

According to Dreier (2001: 38-42) and Brunero (2005: 561-4), the Internal Rule solution falls
prey to a regress. Basically, the suspicion behind the regress argument suggested by Carroll’s
Tortoise is that whatever is to govern our attitudes must not be something additional to our
attitudes. For if our actual attitudes do not suffice for our obligations, then why suppose that
additional attitudes (pro-attitudes in this case) will be of any help? Compare Carroll’s initial
case: the suspicion is that whatever is to take us from premises to a conclusion should not itself
be an extra premise. For, again, if any given set of premises fails to entail a conclusion, why
suppose additional premises will help? (cf. Thomson 1960, Clark 2002: 87-8)

Still, the question is how this line could be made precise, and next I will show that the regress
that Dreier and Brunero have in mind is in fact not generated. The Failure II reconstruction
would run along the following lines:

Instrumental Tortoise (Failure II instance)

(1) You have to accept an attitude (e.g. the intention to move to the other side of the street; ‘I’ in
short) given at least one set of other attitudes that you have.
(2) For any set of attitudes x, if you have to accept I given x, then you have a pro-attitude
towards a rule that commits you to I given x.
(3) For any set of attitudes x, if you have a pro-attitude towards a rule that commits you to I
given x, then you first have to accept I given another set of attitudes y, i.e. the set of x plus that
pro-attitude, in order to accept I given x.
…
(C) If you have a pro-attitude towards a rule whenever you have to accept I given any set of
attitudes, then you will never accept I given any set of other attitudes that you have. [1-5]

This is the conclusion that Dreier and Brunero draw from the regress: Proattitudes are useless to
make one accept certain attitudes given other attitudes that one has.

This conclusion cannot be reached from (1)-(3), however, as (3) fails to hold. Or at least: the
Internal Rule solution subscribes to (2), but need not be committed to (3) as well. The Internal
Rule says, simply, that if you need to accept the intention to move to the other side of the street
given your intention to get the lettuce and your belief that this requires you to move to the other
side of the street, then you have a pro-attitude towards (R). That is, in that case you have a desire
to comply with the rule which takes you from your initial two attitudes to the new one.

However, it does not also say (or need to concede) that your obligation to accept a new attitude
given a set of attitudes depends on a And So On 138 further obligation to accept that attitude
given a bigger set of attitudes, i.e. comprising that pro-attitude. Surely, if you would have the
latter obligation as well, then by (2) you would appeal to further pro-attitudes, i.e. proattitudes to
follow the following series of rules:

(R*) S intends to φ; S believes that φ-ing requires S to ψ; S has a proattitude towards (R) ⊢ S
ought to intend to ψ.

(R**) S intends to φ; S believes that φ-ing requires S to ψ; S has a proattitude towards (R); S has
a pro-attitude towards (R*) ⊢ S ought to intend to ψ.

etc.

This whole series is irrelevant, however, so long as you do not have to accept the intention to
move to the other side of the street because of all those proattitudes. Moreover, if (3) fails then
no regress is generated and then the argument cannot be used against Internal Rule.61

Generally, the fallacy I have stressed here and in the previous section is the No Vicious Regress
Fallacy: One should not criticise a view by saying that it has problematic, regressive
consequences, if in fact it has no such consequences (or, at least, does not have them so long as
further things are not debated first).

5.4. Epistemic Infinitism

No regress has received more attention than the regress of reasons. Consequently, all possible
aspects of the Paradox and Failure I and II variants of the regress argument have in fact been
discussed or at least mentioned in the literature. In the following I will explain why it is useful to
clearly separate Paradox, Failure I and Failure II issues. Unless noted otherwise, in my
discussion I will be speaking about propositional rather than doxastic justification. That is, I will
be concerned with the conditions under which someone is justified in believing a proposition,
whether or not that person actually believes that proposition (cf. Klein 2007: 6).

The Paradox debate. A proposition is justified to someone when certain conditions are fulfilled.
According to Infinitism, a proposition is justified to a subject S only if there is a reason for that
proposition, a reason for that reason, and indeed an infinite series of reasons, and only if all of
them are available to S. Consequently, if such a series is not available, the proposition is not
justified to S. Opponents of Infinitism have argued that these requirements are incorrect, given
that justification hardly ever obtains in such circumstances. How can there be an infinity of
relevant reasons? How can they all be available to S?

Infinitists have two options at this point. Either, first, they can argue that all necessary conditions
obtain in certain cases after all (not only trivially, of course, for in that case any arbitrary
proposition could be justified to S). Or else, second, they can bite the bullet and defend the claim
that the Infinitist’s conditions on justification are the right ones, and that, if it so happens that
they are not fulfilled, then it so happens that justification nowhere obtains. In sum, this first
debate centres on the question of whether or not the Infinitist’s conditions on justification are
problematic.

The Failure debate. A proposition is justified to someone depending on one’s strategy in

accomplishing the task. According to Infinitism, the strategy is to appeal endlessly to reasons
which themselves must be justified by further reasons. As there is always some further reason to
be justified in order for any single proposition to be justified, one will never justify any
proposition. Opponents of Infinitism have complained that this must be the wrong strategy, given
that, on its basis, justification is never reached. Indeed, the Foundationalist’s strategy is to stop
the regress of reasons at basic beliefs which need no further reasons to be justified. And the
Coherentist’s strategy is to admit circular reasoning or else to drop local justification altogether
in favour of a more holistic viewpoint (according to which a proposition is justified when it can
be shown that one’s total system of beliefs is more coherent with that proposition than without).

Infinitists, again, have two options at this point. Either, first, they can argue that the Infinitists’
strategy is successful in certain cases after all (and not just trivially successful, again, for in that
case one could justify any arbitrary proposition). Or else, second, they can bite the bullet and
defend the claim that the Infinitist’s strategy is the only right one, and that if it so happens that it
does not do what it seemingly has to do (i.e. justify propositions), then it so happens that this task
of gaining justification is nowhere ever accomplished. In sum, this second debate centres on the
question of whether or not the Infinitist’s strategy to reach justification is problematic.

The difference between these two debates is clear: The first centres on the Paradox
reconstruction, and the second on the Failure II reconstruction (both reconstructions are provided
above in §3.4.1). It is important, accordingly, to keep these two apart, as one cannot criticise or
defend Infinitism only by discussing the Paradox debate. Let me explain.
A common worry about Infinitism is that we, i.e. human beings with limited mental capacities,
can never have an infinity of reasons available to us. This worry has been countered by Klein
(1999: 306-10). According to him, reasons can be available to a person in two ways, i.e.
objectively and subjectively, and both of these are required for justification. A reason is
subjectively available to S if it is, in some sense, among S’s beliefs. A reason is objectively
available to S, by contrast, if it does indeed support the proposition for which it is a reason,
whether or not it is linked in any way to S’s beliefs. Clearly, the latter kind of availability is
unproblematic to the extent that it has nothing to do with S’s limited mental capacities. And
neither is the former, subjective kind of availability problematic, so long as it does not require
that the infinite set of reasons be entertained consciously in one’s mind.

Still, a defense of Infinitism requires more than this. For Klein’s argument here serves to show
only that, if it is successful, Infinitism can overcome one Paradox argument. Yet, first, there may
be other Paradox arguments which show that other aspects of Infinitism’s requirements on
justification are problematic. And furthermore, second, there is the Failure argument which relies
on no such step at all. The Failure argument against Infinitism, which Klein calls the ‘Specter of
Scepticism’ (1999: 312), basically holds that even if all necessary conditions for justification are
unproblematic (i.e. even if Infinitism can overcome all objections in the Paradox debate), and so
even if all these conditions may well be in place, still justification may never be reached via the
Infinitist’s strategy.

Next I will clarify the Failure argument and defend it from some objections. Before invoking the
Failure Schemas, I will take up the discussion between Gillett (2003) and Klein (2003). Their
overall positions are basically the following: Gillett challenges that justification can be reached
via a regress of reasons, while Klein defends Infinitism against this objection. Consider the
following principle from Gillett (2003: 712):62

IV Anything of a certain kind has a property H depending on another item of that kind having H.

In the Infinitism debate, ‘H’ is the property of being justified, and the things at issue
propositions, such that the instance of IV becomes: Any proposition is justified depending on
another, distinct proposition being justified. Clearly, this seems to generate a regress as soon as
one considers the justification of any arbitrary proposition:

p1 is justified depending on p2 being justified;

p2 is justified depending on p3 being justified;
p3 is justified depending on p4 being justified;
etc.

Following Gillett, let us call any regress governed by an instance of IV a ‘IV regress’. Here is
their controversy about such regresses:

The question consequently arises how it could ever come to pass that any member of the
chain has the property H? (Gillett 2003: 713)
 But the IV regress is not designed to answer that question. It is designed to answer the
quite different question ‘How does it come to pass that each member of the chain – taken
individually – has property H?’ (Klein 2003: 729)

In terms of the case at issue, the controversy is the following. Gillett asks: How could it ever be
that any proposition (in a IV regress) is justified? Klein rebuts: This is the wrong question. The
right question is: How could it be that each proposition (in a IV regress) is justified? And the
latter question, according to Klein, has an easy answer: Each proposition is justified simply
thanks to the next member of the chain of propositions being justified. Furthermore, he argues
that questions like ‘How could it be that all propositions in an IV regress, taken collectively, are
justified?’ are wrongheaded, as they are asking for nothing other than a Prime or Unmoved
Mover. Compare: How could it be that all events in an IV regress, which are one another’s
causes, taken collectively, exist? Because of God? To ask for such starting points is bad not
because there are no such starting points, but because the question already assumes
Foundationalism and begs the question against Infinitism.

Klein’s response may look sensible, and it has been taken to be sound by Aikin (2011: 52-7) and
Bliss (2012) (cf. also Orilia 2006: 232, Peijnenburg & Atkinson 2008: 336). In my view,
however, Gillett’s worry (which he calls the Structural Objection) is a real challenge, and does
not fail for the explanation provided by Klein: I do not think that Gillett implicitly And So On
142 assumes an Unmoved Mover. To see this, three questions may be distinguished:

• How is at least one member of an IV regress H?

• How are all members of an IV regress, taken individually, H?

• How are all members of an IV regress, taken collectively, H?

These questions are clearly distinct. Klein assumes that Gillett is concerned about the third
question, yet on what seems to me the most interesting reading of Gillett’s challenge the first
question is far more important. Gillett does not ask, that is, ‘How are all propositions in an IV
regress, taken collectively, justified?’ but rather ‘How is at least one proposition in an IV regress
justified?’. Furthermore, this last question does not fall prey to the Unmoved Mover objection,
i.e. it does not assume that there is a first proposition which is the source of all justification.

Compare Juvenal’s case. The corresponding IV instance would read as follows: Any guardian is
reliable depending on the next guardian being reliable. If we ask, ‘How are all guardians, taken
collectively, reliable?’, it is tempting to appeal to someone who is the source of all reliability,
and not itself not dependent for its reliability on anyone else being reliable. Yet if instead we ask,
‘How is at least one guardian of the series reliable?’, then it is not clear what to say. If all
guardians are reliable only so long as further guardians are reliable, then how does the reliability
of any one of them come into the picture at all? Again it may be tempting to appeal to a Prime
Guardian.63

This assessment is mistaken, however, as it turns the dialectic on its head. The problem in itself
(i.e. how to make sure that at least someone is reliable) does not assume that only a Prime
Guardian can be the answer to it. The point, rather, is that Infinitist solutions are bad ways of
responding to such problems. This can easily be seen if we reconsider the reasoning in Failure II
format:

Guardians (Failure II instance)

(1) You have to be sure for at least one person that she is reliable.
(2) For any person x, if you have to be sure that x is reliable, you hire a guardian for x.
(3) For any persons x and y, if you hire a guardian y for x, then you first have to be sure that y is
reliable in order to be sure that x is reliable.
…
(C) You will never be sure that anyone is reliable if you hire a guardian every time you have to
be sure that someone is reliable. [1-5]

Therefore, if you want to have your girlfriend guarded because you suspect her of being
unreliable, this argument demonstrates that your intended solution will not work. And one cannot
save the solution, i.e. line (2), by saying that the problem, i.e. line (1), already assumes a Prime
Guardian. For it assumes no such thing, and admits of all sorts of other solutions, as there are
various other ways (other than hiring an endless series of unreliable guardians) to make sure that
people are reliable: you can lock them up, keep them under camera surveillance, start a normal
relation with them, etc. Now if we replace ‘having x guarded’ with ‘having x justified’, and
replace persons with propositions, we can set up a similar argument against Infinitism:

Justification (Failure II instance)

(1) You should have at least one proposition justified.

(2) For any proposition x, if you should have x justified, you have another proposition available
which is a reason for x.
(3) For any propositions x and y, if you have a reason y for x available, then you first should
have y justified in order to have x justified.
…
(C) You will never have any proposition justified if you have a reason available every time you
should have a proposition justified. [1-5]

Infinitism is the position which proposes line (2). Klein (2007: 5-6) defends this line by saying
that all our justified beliefs need a reason because that is what makes us epistemically
responsible (if they have no reason, then it is irresponsible to hold them). Aikin (2011: ch. 5)
defends (2) by saying that all our justified beliefs need a reason because that is what makes us
nondogmatic (if they have no reason, it is dogmatic to hold them).

The conclusion of this Failure II argument is that Infinitism, i.e. the view that maintains (2),
cannot explain how even one proposition is justified to someone. This, I think, is how Gillett’s
worry should be spelled out. Furthermore, my point here is that one cannot save Infinitism by
saying that the problem in (1) (i.e. to have at least one proposition justified) assumes that only a
Prime Proposition can be the answer to it. For it assumes no such thing, and admits other
solutions (such as Foundationalist or Coherentist solutions), i.e. that do not suffer from a Failure
II attack.

If we set the suppressed premises aside,64 then there are just two ways for Infinitists to counter
the argument: by denying premise (1) or premise (3). The first strategy might be promising for
those who believe that justification is not something that needs to be shown (cf. Alston 1985,
Rescorla 2009, 2013). For clearly, the Failure attack relies on the assumption that justification is
an activity, i.e. something that has to be accomplished. If this assumption is denied, and
justification is conceived as something that, so to speak, might just happen to you, then the attack
fails.

Two qualifications. First, the state/activity distinction was designed to support Foundationalism
about perceptual belief, not Infinitism. The suggestion was that I could be justified in believing
that I have a hand if I see that I have a hand, even without doing anything to show or convince
others (or myself) that I have a hand. The suggestion was not that I could have an endless series
of reasons for my belief that I have a hand without doing anything. Second, even if the
state/activity distinction is in place, it is still possible to launch another Failure attack on
Infinitism that starts from the problem ‘You have to show for at least one proposition that it is
justified to you’ and concludes that Infinitism will never show that anyone is justified in
believing anything under its conditions.

Next, in order to evaluate (3), it is worth comparing this Failure II attack on Infinitism with its
Failure I counterpart (i.e. which does not make use of premise (3)):65

Justification (Failure I instance)

(1*) You should have all of your beliefs justified.

(2*) For any belief x, if you should have x justified, you have a reason for x available.
(3*) For any belief x, if you have a reason y for x available, then y is among your beliefs.
…
(C*) You will never have all of your beliefs justified if you have a reason available every time
you should have a belief justified. [1*- 5*]

Clearly (C*) is weaker than (C) (though still problematic for Infinitism). If it has already been
shown that Infinitism fails to explain how you have anything justified, then it need not also be
shown that Infinitism fails to explain how all of your beliefs are justified.

The main difference is to be found in line (3) of the arguments. Suppose you want to be justified
in believing a proposition p1. In that case, you trace a reason p2 for p1. As line (3) of the Failure
II reconstruction has it, you should not only have p2 available, but should have a justification for
p2 as well (and eventually justifications for p3, p4, etc.). This is not required by (3*) of the
Failure I reconstruction. Of course, by the latter in combination with (1*) you will have to justify
an endless series of beliefs as well, but these problems are independent of the problem of
justifying p1.
Why should the stronger (3) apply? One motivation might be the following: that one should have
a justification for p2 in order to have a justification for p1 if the role of this justification (i.e. for
p2) is to settle disputes about p1. Suppose some interlocutor questions p1 (and possibly upholds
¬p1 for such and such reasons), and suppose you justify p1 by appealing to p2. Now suppose
further that you have no justification for p2. In that case, your interlocutor may disregard p2 as
arbitrary, and hence likewise disregard your justification for p1. According to this dialectical
take on justification, (3) of Failure II holds.66

Now, if both (1) and (3) hold, the argument against Infinitism seems quite a good one (at least
for one who wishes to remain non-sceptical about justification). Moreover, the Failure attack
applies regardless of any worries we may have related to the Paradox reconstruction (e.g. about
our mental capacities). Here we have, then, my first extended illustration of the point that it is
useful to keep Paradox, Failure I and Failure II issues apart.

5.5. Access Principle

We have many obligations. I, for one, have the obligation to work on my dissertation, to refrain
from making too much noise in the office, to have true beliefs about where my office is, to drink
fair-trade coffee, to call my mother, etc. Surely there are restrictions on what we ought to do.
One such candidate restriction is epistemic of sort:

Access For any obligation x, one has x only if one can know one has x.

Let us consider the example from Sorensen (1995: 254): I am obliged to donate some of my
inheritance to charity only if I can know I am obliged to donate. This appears plausible. For if I
am in no position to find out that I am obliged to donate, then why should one expect me to
donate? Importantly, Access does not say that we actually have to know our obligations, but it is
only required that it is possible to know them. Surely I might be very ignorant (e.g. if I lack any
kind of curiosity), yet this does not excuse me from my obligations insofar as Access is
concerned.

Still, Sorensen argues against Access because it ‘dumbs down ethics’ (among other
considerations). To see this, consider the following consequence of Access: If one cannot know
one has a certain obligation, then one does not have that obligation. Moreover: If one eliminates
one’s possibility to know whether one has an obligation, then one eliminates one’s (potential)
obligation. So, if I burn the will before reading whether I am obliged to donate, then (assuming
the will was my only access) I eliminate my obligation to donate. In general, Access might be
abused by shirkers who avoid their obligations by eliminating their possibility to know them.
This is a bad consequence of Access.

In response, Sider (1995: 278) suggests that this problem does not follow if we assume that one
cannot just eliminate one’s possibility to know things. That is, one has the obligation to refrain
from making it impossible to know one’s obligations.67

According to Sorensen, in turn, this solution will not do as it invites a regress. Consider the fact
that, by Access, one has the obligation to refrain from making it impossible to know one’s
obligations only if that further obligation itself is knowable. Now consider the problem with the
shirker again. Surely she will work herself in such a situation that this new obligation is not
knowable to her. The question is: How can we say that she is doing something wrong here, given
that we are proponents of Access? Should we again say that she ought to refrain from making it
impossible to know her new obligation? This, indeed, lands us in a regress.

Now, Sorensen and Sider agree that there is a regress, but only the former takes it as a problem
for Access. Here is the situation according to Sorensen. Scenario 1: I ought to donate. As a
shirker, I make it impossible to know this obligation. This is possible by Access and there is no
explanation of where I go wrong. Scenario 2: I ought to refrain from making it impossible to
know whether I ought to donate. As a shirker, I make it impossible to know this new obligation.
This is possible by Access and there is no explanation of where I go wrong. And so on.

Sider disagrees with this picture. According to him, the situation is rather the following. Scenario
1: I ought to donate. As a shirker, I make it impossible to know this obligation (let us label it
‘o1’). Where do I go wrong? I violate the following obligation o2: I ought to refrain from making
it impossible to know whether I have o1. Scenario 2: As a Shirker, I also make it impossible to
know whether I have o2. Where do I go wrong? I violate the following obligation o3: I ought to
refrain from making it impossible to know whether I have o2. And so on. In contrast to what
Sorensen suggests, each time there is an explanation of where I go wrong. Sider concludes that
none of these scenarios puts any pressure on Access:

At best, the regress consists of an infinite sequence of cases, none of which refutes
Access. (1995: 279)

Now the question is: Who is right? Is Access a good restriction on our obligations, or not?

In the following I shall invoke the regress argument schemas, and on the basis of a Paradox
reconstruction argue that Sider is right that in a certain sense the regress is harmless. After that,
on the basis of a Failure reconstruction I also show that there remains a serious worry about
Access. First, the Paradox reconstruction:

Access (Paradox instance)

(1) For any action x, S ought to perform x only if S can know she ought to perform x.
(2) For any action x, S can know she ought to perform x only if S ought to refrain from making it
impossible to know whether she ought to perform x.
(3) S ought to perform at least one action.
(4) Regress:
(a) I ought to donate.
(b) I can know I ought to donate. [a, 1]
(c) I ought to refrain from making it impossible to know whether I ought to donate. [b, 2]
(d) I can know I ought to refrain from making it impossible to know whether I ought to
donate. [c, 1]
(e) I ought to refrain from making it impossible to know whether I ought to refrain from
making it impossible to know whether I ought to donate. [d, 2]
 And so on.
(5) I have an infinity of obligations, and can know all of them. [3, 4]
(6) This is absurd.
(C) (1) is false: It is not the case that for all actions x, S ought to perform x only if S can know
she ought to perform x. [1-6]

This forms a direct argument against Access, at least in its universally quantified version. If (2),
(3) and (6) are true, then (1), i.e. Access, is false. Is it sound?

In my view this is not the case, because (6) is highly controversial. The reason, here, is that
obligations and possibilities of knowing often come for free. Why should there be any limit to
the obligations that I have? This is a serious worry because I can easily respect all obligations
generated in the regress in one go, namely by donating some of my inheritance to charity. Also,
why should there be any limit to the things I can possibly know? Again, this is a serious worry
because if I am in the position to know the general truth (if it is a truth) that I ought to refrain
from making it impossible to know whether p (for an arbitrary instance of ‘p’), then I am in the
position to know the truth of all its instances (including those generated in the regress).

If this reasoning is right, then (6) is false, the regress at hand nonvicious,68 and then Access
cannot be refuted on the basis of this argument. Moreover, this would provide additional support
for Sider’s position in this debate.

Still, there remains a worry about Access, namely that Sider’s attempt to catch the shirker, i.e.
his attempt to show that one cannot abuse the Access principle, fails. This worry can be phrased
nicely in Failure terms:

Access (Failure II instance)

(1) For at least one obligation of the shirker x, you have to secure x.
(2) For any obligation x of the shirker, if you have to secure x, then you appeal to an obligation
to refrain from making it impossible to know x.
(3) For any obligation x of the shirker, if you appeal to an obligation y to refrain from making it
impossible to know x, then you first have to secure y in order to secure x.
(4) Regress:
(a) You have to secure her obligation to donate.
(b) You appeal to her obligation to refrain from making it impossible to know her
obligation to donate. [a, 2]
(c) You have to secure first her obligation to refrain from making it impossible to know
her obligation to donate. [b, 3]
(d) You appeal to her obligation to refrain from making it impossible to know her
obligation to refrain from making it impossible to know her obligation to donate. [c, 2]
(e) You have to secure first her obligation to refrain from making it impossible to know
her obligation to refrain from making it impossible to know her obligation to donate. [d,
3] And so on.
(5) For any obligation x of the shirker, you first have to secure a regress of obligations of the
shirker in order to secure x. [4]
(C) If you appeal to an obligation to refrain from making it impossible to know her obligation
every time you have to secure one, then you will never secure any obligation of the shirker. [1-5]

This regress argument appears sound.69 It proves that consequent shirkers form a problem for
Access.

At one point, Sorensen (1995: 255) and Sider (1995: 279) suggest that we need not take such
shirkers seriously. For what might it mean, for instance, to violate one’s obligation to refrain
from making it impossible to know one’s obligation to refrain from making it impossible to
know one’s obligation to donate (see line (d) in the reconstruction above)? But I am inclined to
disagree here. For if we link our obligations with issues of responsibility, and say that we are
responsible (and possibly blameworthy) for what we do only if we are in the position to know
our obligations, then based on the loophole just identified, I could make it the case that I am
never responsible and to be blamed for what I do. If this is right, then Access is no good
restriction on our obligations after all.

In any case, this Sorensen/Sider debate is another extended example of where it proves useful to
separate Paradox and Failure concerns.

5.6. Russell’s Relations

Before Russell, the view prevailed that all relations are in fact properties of their relata. The
relation of greater-than between me and the universe, for example, was conceived to be nothing
but our size properties. Russell defended the new, incompatible view that an important group of
relations do not behave in this way. This was a major achievement in the history of philosophy.70
Aside from certain selected philosophers (e.g. Campbell 1990: 101-4), virtually everyone today
accepts Russell’s conclusion (cf. Mulligan 1998: 325-7).

Russell’s main argument was a regress argument. In the following, I basically show that this
argument is ambiguous as it has two different conclusions, but demonstrate how both can be
drawn if the regress argument is set up in two different ways (i.e. Paradox vs. Failure-wise).

In order to explain Russell’s argument, we should first be clear on some terminology.

Specifically, two pairs of distinctions are relevant: relations can be internal or external, and
reducible or irreducible. The view that Russell is attacking, which he called Monadism, is the
view that all relations are internal and reducible. In Russell’s view, some relations are external
and irreducible. Examples of the latter class include the relations of part and whole, before and
after, greater and less, and cause and effect (1899: 139).

Importantly, externality and irreducibility do not coincide, and neither do internality and
reducibility. Here are two definitions. For all items x, y which stand in a relation R,

• R is internal iff there are corresponding properties F and G such that x is F and y is G (and
otherwise external).
• R is reducible iff there are corresponding properties F and G such that R is nothing but F of x
and G of y (and otherwise irreducible).

The first distinction derives from Russell (1899: 143, 1903: §214, cf. Bradley 1893: 514), and
the second from Russell (1906-07: 41, 1959: 54-5). (Note that in both cases the definitions are
restricted to Monadism. Russell also considers another theory of relations, i.e. Monism, which
reduces R not to distinct properties of x and y but rather to a single property of the sum of x and
y. In order to discuss the regress argument against Monadism, I will ignore the Monist view in
what follows.)

The main difference between the two, reducibility and internality, is that the reducibility claim is
stronger than the internality claim. That is, even if all relations are internal and imply that their
relata have corresponding properties, then still it may be the case that they are not reducible to
those properties. As I shall argue in the following, Russell’s argument against Monadism can be
spelled out both as an argument against the claim that all relations are internal (and none
external) and as an argument against the claim that all relations are reducible.

Here is one of Russell’s examples: a is earlier than b. If the earlierthan relation is internal, there
should be corresponding properties F and G such that a is F and b is G. If the earlier-than relation
is also reducible, it is nothing but F of a and G of b. What might F and G be in this case? Russell
considers two options.

First option: ‘F’ stands for the property ‘being earlier than b’, and ‘G’ for the property ‘being
later than a’, such that a has the property of ‘being earlier than b’, and b has the property of
‘being later than a’. This option may have some initial plausibility, but Russell quickly dismisses
it as follows: properties like ‘being earlier than b’ involve reference to another item (b, in this
case), and if properties involve such references they are properties relative to those items, and so
involve relations to those items (1903: §214).

Second option: ‘F’ stands for ‘existing at t1’, and ‘G’ for ‘existing at t2’, such that a has the
property of existing at t1, and b has the property of existing at t2. Here we meet Russell’s regress
argument, which he repeats in several places. Perhaps its clearest expression is here:

You may say that the date of a is a property of a and the date of b is a property of b, but
that will not help you because you will have to go on to say that the date of a is earlier
than the date of b, so that you will have found no escape from the relation. (1959: 54-5)

Why is it the case that the dates of a and b stand in an earlier-than relation if a and b themselves
do not stand in it? Russell’s motivation is the following (1903: §214). The earlier-than relation
between a and b is asymmetric. If you reduce this relation to the properties of a and b, then the
only remaining relation between a and b will be ‘difference in property’, which is symmetric.
But the feature of asymmetry should be retained, for without it ordered series such as numbers,
space, time, etc. are not possible. Hence, if there is no asymmetric relation between a and b, there
should at least be one between their two properties.
Importantly, the above text is directed against the view that all relations are reducible (cf. also
1906-07: 41-2). Yet, Russell’s original statement is directed against the view that all relations are
internal:

[…] Hence the attempted analysis of the relation fails, and we are forced to admit what
the theory was designed to avoid, a so-called ‘external’ relation, i.e. one implying no
complexity in either of the related terms. (1903: §214)

In other words, Russell’s regress argument is ambiguous. It proves both that not all relations
imply corresponding properties of their relata, and that not all relations are reducible to such
properties. As noted above, these claims are not identical. So what is the proper conclusion of
Russell’s argument?

On the basis of the reconstructions from §3.4.3, it can easily be seen that both conclusions make
sense. Russell’s anti-internality claim refers to the conclusion of the Paradox reconstruction: It is
not the case that for any relation R and items x, y, if x and y stand in R, then there are properties
F and G such that x is F and y is G. By contrast, Russell’s anti-reducibility claim refers to the
conclusion of the Failure I reconstruction: You will never reduce all relations if you reduce them
to properties of their relata.

Strictly speaking, of course, this is not a Paradox/Failure Fallacy, but the debate will be confused
unless we distinguish the two arguments. Most importantly, the two arguments have different
premises and so cannot be evaluated in the same way. For example, in the Paradox case Russell
needs a premise which states that an endless series of time-orders is absurd. No such premise is
required in the Failure I case.

It is worth noting that this result is by no means accidental. There is a reason for using the
Paradox Schema in one case and the Failure in the other. The Paradox Schema is used against the
internality claim because the issue there concerns whether or not this universally quantified
claim holds unrestrictedly, and the Paradox reconstruction demonstrates that it does not. By
contrast, the Failure Schema is used against the reducibility claim because there we encounter a
problem: namely, to reduce all relations (this may be a problem because relations appear to be
strange entities, or because, as per Occam’s Razor, it is always better to have no more kinds of
entities than are necessary). The Failure reconstruction demonstrates that the Monadistic solution
does not solve this problem.71

5.7. Quine/Davidson Controversy

Let us consider Quine’s radical translation scenario. A field linguist meets a native who utters the
expression ‘Gavagai’ in an unknown language. The only further information available to the
linguist is that the native points to a rabbit. In this scenario, ‘Gavagai’ might mean several things
such as ‘This is a rabbit’, ‘Here is potential food for tonight’, and so on. Yet, in the following we
shall not be concerned with the meaning of the native’s expression, but rather with its reference.

To what does ‘Gavagai’ refer? What object does it pick out? There are a number of possibilities.
It might for example refer to (i) a complete rabbit which persists in time, (ii) an undetached part
of a rabbit, or (iii) a temporal stage of a rabbit. These are all different things: complete rabbits are
different from their undetached parts (so (i) and (ii) differ), rabbits which persist through time are
different from their temporal stages (so (i) and (iii) differ), and undetached parts of rabbits,
unlike their temporal stages, may or may not persist in time (so (ii) and (iii) differ as well). It
does not really matter how you cut the piece. What matters is that it is possible to take different
items, differing in their spatial or temporal parts, as the referent of ‘Gavagai’.

The point of this story is that, according to Quine, ‘Gavagai’ does not refer to anything full-stop.
‘Gavagai’ refers only to this or that item relative to a background language which, in this case,
specifies the individuation criteria for objects like rabbits. So, ‘Gavagai’ refers to the complete
rabbit which persists in time (say) if it is specified in the background language that singular
terms refer to such items rather than to their undetached parts or temporal stages. This claim can
be called the Relativity of Reference thesis.72

Quine is well aware that this thesis generates a regress. If the reference of the words in the object
language is to be fixed by a background language, then the reference of the words in the
background language (e.g. which specify that ‘Gavagai’ refers to a complete persisting rabbit) is
to be fixed by a further background language. The regress that Quine has in mind can be stated in
Failure I format:

Reference (Failure I instance) (1) You have to fix the reference of all words.
(2) For any word x, if you have to fix the reference of x, you use a background language to
specify it.
(3) For any word x, if you use a background language M to fix the reference of x, then M
contains further referring words.
(4) Regress:
(a) You have to fix the reference of ‘rabbit’.
(b) You use a background language1 in which you specify that ‘rabbit’ refers1 to a
complete persisting rabbit. [a, 2]
(c) You have to fix the reference of ‘refers1’. [b, 3, 1]
(d) You use a background language2 in which you specify that ‘refers1’ refers2 to the
relation between ‘rabbit’ of the initial language and ‘a complete persisting rabbit’ of
background language1. [c, 2]
(e) You have to fix the reference of ‘refers2’. [d, 3, 1]
And so on.
…

Yet, Quine does not take the regress to be problematic. He suggests that the relativity of
reference can be compared with the relativity of position. In the latter case, it similarly does not
make sense to ask where you are, full-stop (or absolutely) but only where you are relative to a
system of coordinates. For example, I am in the middle of my empty room relative to the walls
and everything not in my room. Of course, one may further ask about the position of my room,
and this question can be answered in a similar way relative to a system of coordinates. But it
makes no sense to ask where my room is, full-stop.
Likewise: it makes no sense to ask what ‘rabbit’ of the object language or ‘refers1’ of the
background language refers to (i.e. what their extension is), full-stop. In other words, even if we
never stop asking about the position of all coordinate systems, it may be no problem that we
never complete this task. For we may well be asking for things that are not there (i.e. absolute
positions). Likewise: even if we never stop asking about the reference of all words, it may be no
problem if we never complete this task. For there may well be no absolute referents (e.g. no fact
of the matter as to what ‘rabbit’ refers to).

This analogy may look plausible, yet Davidson cannot agree. Indeed, he uses this very same
regress to argue that the whole Relativity of Reference thesis does not make sense. So the
controversy between Davidson and Quine may be put as follows. On the one side, the relativity
of reference thesis generates a regress, but this is accepted as unproblematic by Quine. On the
other, Davidson argues that the relativity of reference thesis does not make sense as it, due to the
regress, does not explain how we fix reference in the first place. These positions seem to conflict.
So who is right?

In the following I will argue that Davidson actually switches the argument to its Failure II
counterpart, which is far more problematic. Thus reconstructed, the regress argument reads as
follows:

Reference (Failure II instance)

(1*) You have to fix for at least one word its reference.
(2*) For any word x, if you have to fix the reference of x, you use a background language to
specify it.
(3*) For any word x, if you use a background language M, then you first have to fix the reference
of the words in M in order to fix the reference of x.
…
(C*) You will never fix the reference of any word if you use a background language every time
you have to fix the reference of a word. [1*-5*]

Indeed, the conclusion of the argument in this latter form is that the reference of no word is ever
fixed using background languages. Hence, the Quine/Davidson controversy reduces to a
controversy about (3*) of the argument just given. For this is the main premise which
distinguishes a Failure II argument from a Failure I one. Basically this premise says that the
introduction of a background language is of no use in fixing the reference of a word like
‘Gavagai’ unless we are already clear on the referents of the background language. Why does
Davidson think that this holds?

“Wilt”
_____
/ \
| ^
V |

Figure 5: Davidson’s rabbit

To make his point, Davidson slightly changes Quine’s scenario. In this new scenario, there are
three persons A, B and C, and A utters ‘Wilt’ in the presence of a rabbit (and his shadow). B
comes along and wants to specify the reference of A’s expression. B sets up two theories, and
says that A’s expression refers to an object relative to these two theories. According to the first
theory, ‘Wilt’ refers to the rabbit. According to the second, ‘Wilt’ refers to the shadow of the
rabbit. So, B successfully fixes the reference of A’s ‘Wilt’ (i.e. relative to the two theories) only
if the third party C is able to pick out the two referents of A’s expression given B’s two theories.
However, if A’s words do not refer unless relativized to a background language (i.e. B’s
theories), then neither do B’s words refer unless relativized to a further background language. So
C tries to fix the referents of B’s words, and sets up two different theories about B’s word
‘refers’, i.e. identifies different relations between A’s expression and B’s theories. According to
the first theory, ‘refers’ refers to the relation between A’s ‘Wilt’ and B’s ‘rabbit’. According to
the second theory, ‘refers’ refers to the relation between A’s ‘Wilt’ and B’s ‘shadow’. But C, of
course, can pick out unique objects only when her words are themselves relativized to yet further
background languages. And so on. The moral:

A can talk distinctively and meaningfully about Wilt and shadows. B can talk
distinctively and meaningfully about two different relations between A’s words and
objects. But at no point has anyone been able uniquely to specify the objects of which a
predicate is true, no matter how arbitrarily or relatively. (Davidson 1979: 234-5)

If Davidson is right that the regress argument takes the Failure II form, then Quine’s analogy
between the relativity of reference and the relativity of position breaks down. In the latter case, a
position can be fixed relative to a coordinate system whether or not the position of the coordinate
system itself is fixed. In the former case, the reference of a word cannot be fixed relative to
background language unless it is already clear what the words of that background language refer
to (cf. Figure 5; the arrow stands for the reference predicate that is relativized to background
theories).

The role of the metaphilosopher is not to decide who is right in this debate. The important point
for my purposes is rather this: Davidson and Quine do not draw different conclusions from the
same regress, but rather different conclusions from two regresses generated in different ways (i.e.
from different premises): Failure I or Failure II-wise. So here we have another, final example of
how the distinction between these two argument schemas is useful in clarifying existing debates.

§6 Meta-Debate

6.1. The Literature

In the following I first divide the literature on regress arguments into three camps: Paradox-
Monists, Failure-Monists, and Pluralists. Next, I defend the Failure Schemas from some
objections raised and anticipated in the literature.

1. Three camps
As we have seen, any regress argument from the literature can be reconstructed in two different
ways: as an instance of the Paradox Schema, or an instance of the Failure Schema. This is
possible partly because no text is fully explicit concerning its premises and/or inferences. Yet if
this is right, if both ways are possible, then what is to be done? Which schema should be used in
a given case? There are at least three options:

Paradox-Monism
Do it always Paradox-wise, never Failure-wise.

Failure-Monism
Do it always Failure-wise, never Paradox-wise.

Pluralism
Do it sometimes Paradox-wise, sometimes Failure-wise.

All three of these camps have their proponents. That is, the literature on regress arguments hosts
both Paradox-Monists, Failure-Monists and Pluralists. It is worth stressing that my classification
is somewhat anachronistic, as the Failure Schemas have been presented for the first time only in
this dissertation, and only very basic versions are presently to be found elsewhere in the literature
(mainly one finds only its rationale). I have used the following criteria for my categorisations:

• Does the author think that her analysis applies to all regress arguments, or only to some?

• Does the author think that there is a direct connection between a regress and a failure (as
Passmore suggested), or does she see a need for some intermediate step to show why the regress
is bad such that something has to be rejected?

The first criterion should distinguish Monists and Pluralists, and the second Paradox and Failure-
Monists. So here we go, starting with the ParadoxMonists:

Paradox-Monists: Gettier (1965); Clark (1988); Black (1996); Post (1993); Gratton (1997,
2010); Rescher (2010); Roy (2010)

The classification of Gettier, Black and Gratton will become clear in the next section, where I
discuss their objections to Passmore’s Failure-hypotheses. Still, a note on Gettier and Gratton is
in order. Gettier (1965) is strictly speaking only a review of Passmore (1961), but as he spends
some time criticizing the latter’s ideas on regress arguments I have listed him among the
Paradox-Monists. Here is a revealing passage:

After all, the argument purports to prove the impossibility of the occurrence of the first
member, not the impossibility of an explanation of its occurrence. (1965: 268)

As we shall see, the fact that Gettier does not accept that the regress argument at issue is about
explanation might be explained by the fact he does not accept the Failure Schema as a legitimate
form for a regress argument.
One might think that Gratton is a Pluralist rather than an ParadoxMonist, because he holds that
one group of regresses (i.e. the so-called problem/response regresses) do not fit into his schema
(I will discuss this at some length below). Still, I have classified him as a Paradox-Monist,
because the Paradox Schema is his point of reference for virtually all topics he discusses
(viciousness, circularity, regresses of questions/answers, etc.). Pluralists, by contrast, would not
take one of the two schemas as their sole point of reference.

Another Paradox-Monist is Clark:

Vicious infinite regress arguments are, all of them, instances of a special kind of Reductio
Ad Absurdum. (1988: 372)

First, this is an expression of Monism, as the analysis is meant to apply to regress arguments
across the board. Second, this is an expression of ParadoxMonism because regress arguments
contain an additional step stating how the regress conflicts with something else we believe.
According to Clark, this ‘something else’ always has the same form: infinite regresses entail that
all items of a certain type are F only conditionally (they are F depending on whether something
else is F), whereas it is believed that at least some of them are F categorically (they are F
independent of whether something else is F). For example, the regress of reasons entails that all
propositions are justified to someone depending on another proposition that is justified to that
person, yet this conflicts with the belief that at least some propositions are categorically justified
to someone.73

Next the Failure-Monists:

Failure-Monists: Passmore (1961); Rankin (1969); Johnstone (1996); Gillett (2003); Maurin
(2007)

The classification of Passmore was explained above in §2.3.4 where I identified the origins of
the Failure Schema. I have classified Rankin and Johnstone as Failure-Monists because they rely
on Passmore’s analysis. It is worth noting that Johnstone does talk about inconsistencies, but
these are not conflicts between the regress and something else we believe, but rather conflicts
between what had to be accomplished and what actually has been accomplished (or rather
postponed). As Johnstone puts it:

An alleged definition [etc.] gives rise to an infinite regress when instead of defining [etc.]
it merely postpones the definition, analysis, criterion, or justification, or account of the
decision. (1996: 97)

The same holds for the Failure Schema: its conclusion that the problem is never solved (or its
solution forever postponed) is inconsistent with the claim that the problem has been solved. This
further step plays however no role in Failure arguments (as their goal is not to reject anything on
the basis of a contradiction, see §4.1.3).

Gillett and Maurin are arguably Failure-Monists for their defense of the claim that what is
important about regresses is the procedure which generates them, not the results they entail
(results which may or may not conflict with something else we believe). For example, the regress
of reasons would be bad not because it entails there being an infinity of reasons (which may not
exist), but because at no point can justification be obtained if it is to be achieved only via the
procedure which governs the regress. This is what Gillett calls the Structural Objection (see
§5.4).

Maurin, in turn, provides an overview of accounts which identify viciousness at locations other
than the regress and its generating procedure, but she takes them to be less interesting. Her
Monistic position is nicely expressed here:

This pattern, furthermore, is completely general. No matter which context, no matter

what trigger, what distinguishes a vicious from a virtuous infinite regress is its pattern of
dependence. (2007: 21)

This can be spelled out in terms of the Failure Schema so long as those patterns of dependence
are associated with problem/solution patterns (as Maurin does). In the vicious case, the initial
problem depends on a whole regress of problems (such that the latter must be solved in order to
solve the initial problem). In the virtuous case, the dependence is the other way around: the
whole regress of problems depends just on the initial problem (i.e. if you META-DEBATE 161
solve the initial problem, you solve the whole regress of problems in one go). This difference is
depicted in Figure 6.74

(a) -> (c) -> (e) ---->

Vicious pattern

(a) <- (c) <- (e) <----

Non-vicious pattern

Figure 6: Dependence of problems

Further Failure ideas have been expressed, as we have seen in §2.3.4 and §4.2.1, by Johnson
(1978), Armstrong (1974, 1978), Ruben (1990), and Rodriguez-Pereyra (2002). Yet, their
remarks are too brief to count as full Failure views. The same holds for the following passage by
Chisholm, who seems to describe a Failure regress (yet does not generate it from
premises/hypotheses, nor link it to a Failure conclusion):

What does it mean to be confronted by such a regress? One is confronted with a vicious
infinite regress when one attempts a task of the following sort: Every step needed to
begin the task requires a preliminary step. (1996: 53)

Finally the Pluralists:

Pluralists: Schlesinger (1983); Sanford (1984); Day (1986, 1987); Jacquette (1996); Nolan
(2001)
Most of these are Pluralists simply because they classify regress arguments in different
categories. For example, Schlesinger has four categories, Sanford has two, and Jacquette three.
Of course, these categories vary depending on what they regard as important distinctions and the
level of generality of those distinctions (i.e. whether they themselves can be grouped or not). I
shall not go into this, as for my purposes it is important just that all classifications include at least
both Paradox and Failure arguments (this being the distinction I regard as important). Here are
two typical expressions of Pluralism:

I have not said whether regresses are vicious or not. There is a good reason for this. I
hope it is apparent by now. There is no single answer. (Day 1986: 273)

In providing a taxonomy of different sorts of vicious infinite regresses, I take issue with
those who try to provide the same aeitiology for all vicious infinite regresses. (Nolan
2001: 524)

Some participants in the debate on regress arguments resist categorisation. I would highlight two
of them: Russell and Rosenberg. Here are Russell’s pioneering ideas:

Thus wherever the meaning of a proposition is in question, an infinite regress is

objectionable, since we never reach a proposition which has a definite meaning. But
many infinite regresses are not of this form. If A be a proposition whose meaning is
perfectly definite, and A implies B, B implies C, and so on, we have an infinite regress of
a quite unobjectionable kind. […] Thus there is no logical necessity, as there was in the
previous case, to complete the infinite regress before A acquires a meaning. (1903: §329)

Here is how I would cash out and generalise Russell’s distinction:

• A regress is unobjectionable if it is generated Paradox-wise and entails no unacceptable result.

• A regress is objectionable if it is generated Failure-wise such that each time yet another
problem needs be solved before the initial problem is solved.

For example, regresses like the following which are generated on the basis of ‘for any instance of
p, p iff it is true that p’ (plus a trigger) are unobjectionable because they entail nothing
unacceptable:

Socrates is mortal.
It is true that Socrates is mortal.
It is true that it is true that Socrates is mortal.
etc.

On the other hand, if I have to define the meaning of ‘Socrates is mortal’ and my solution falls
prey to a regress where each time the meaning of yet another proposition needs be defined first,
then the regress is objectionable. For in that case I will never define the meaning of ‘Socrates is
mortal’.75
At first sight, Rosenberg hints at a Failure regress as well:

That is the essence of the criticism. The question does not go away. (1978: 72)

This is about Failure regresses if we assume that a question is just one form a problem can take,
and that problems do not go away in the sense that, throughout the regress, they are never solved
by the solution under consideration. Yet, Rosenberg proceeds:

This is what makes the challenge dialectical rather than logical. It disqualifies the
proposed answer as an answer for something qualifies as an answer to a question only if
one can understand it without already knowing the answer to the question. The
philosopher who offers this answer, therefore, violates a canon of rational practice.
(1978: 72)

The relevant principle would be this: that if someone asks you a question Q, then you should not
give an answer which can be understood only if that person already knows the answer to Q. True
enough: when I ask you to explain what a philosopher is, then you should not answer ‘someone
who is a philosopher’, for in that case I can understand your answer only if I already know the
answer to my question. Furthermore, the idea is that violating this principle does not make one
illogical (because it is not a logical principle), but rather irrational (because it would be a
principle of rationality). You are not rational if you refuse to obey such things as this principle.
Here is Clark’s criticism:

It is not at all obvious that the reason, on Rosenberg’s analysis, that ‘the question does
not go away’ is not based in formal properties of the argument. (1988: 379)

This seems right to me, yet Clark develops the point no further. I take it that the Failure Schema
shows that Rosenberg’s phrase can be given a logical treatment. Namely: the question/problem
does not go away, in case of Failure regresses, as each time yet another problem needs to be
solved before one can say the initial problem is solved.

In sum, both Russell and Rosenberg have come close to Failure ideas, yet as they do not say that
all regress arguments have this format they are not classified as Failure-Monists.

As announced from the outset of this dissertation, I am a Pluralist of a certain sort (and indeed
maintain that everyone should adopt this position, which I take to be the correct one). Before
explaining this position in §6.2, I will discuss Passmore’s and Gratton’s Failure-hypotheses in
the remainder of this part in order to show that the set-up of the Failure Schemas as presented in
this dissertation was no trivial enterprise. Others in this metadebate have made different choices.

2. Passmore’s Failure-hypotheses

Passmore’s Failure-hypotheses have been criticised on two fronts, namely by those who
recognise only the Paradox Schema (the Paradox-Monists), and by those who limit the Failure
Schema to a certain group of regress arguments (the Pluralists). In the following I take issue with
five of the most pertinent worries expressed by these critics:
(1) That Passmore’s criterion for distinguishing between regresses which demonstrate that a
certain explanation fails, and those which do not, itself fails.

(2) That the reason Passmore gives to explain why regresses entail inadequate explanations, fails.

(3) That regresses are not merely rhetorical tools, as Passmore assumes them to be.

(4) That the Failure Schema is involved in an ambiguity, and that it is possible to draw another,
more optimistic conclusion as well.

(5) That regress arguments are not about inadequate explanations, but rather about rejections.

Having the Failure Schema at my disposal, unlike these critics, I will argue that all these
objections are based on misunderstandings.

2.1. Objection (1)

The first objection I will discuss is from Gettier (1965). He points out that Passmore’s criterion
for distinguishing between regresses which demonstrate that a certain explanation fails, and
those which do not, does not work. Consider Passmore’s examples:

(P) Every event has a cause.

(P*) To know that an event has happened, one must know how it came about.

According to Passmore, two different regresses can be generated on the basis of these statements
and only the second would demonstrate that the relevant explanation is inadequate. Gettier’s
criticism, I take it, is that it is unclear how Passmore’s criterion yields this result.

Reply

Here is the text that Gettier is referring to:

It is easy to construct similar pairs of assertions, of which one commits us to the view that
some procedure can be carried on ad infinitum, the other commits us to the view that an
infinite regress would have to be completed before the procedure could be carried out at
all. (Passmore 1961: 29)

It is already worth noting that Passmore speaks of procedures rather than explanations, but let us
set this aside for the moment. Next I will show that this criterion can be made precise in the
context of the Failure Schema, and that it can indeed be applied to (P) and (P*). Here are the
cases spelled out in terms of that schema:

(P) Argument (Failure II instance)

(1) You have to explain for at least one event why it exists.
(2) For any event x, if you have to explain why x exists, you appeal to another event which is the
cause of x.
(3) For any events x and y, if you appeal to y which is the cause of x, then you first have to
explain why y exists in order to explain why x exists.
…
(C) If you appeal to a cause for x every time you have to explain why event x exists, then you
will never explain why any event exists. [1-5]

(P*) Argument (Failure II instance)

(1*) You have to know for at least one event whether it happened.
(2*) For any event x, if you have to know whether x happened, you appeal to another event
which is the cause of x.
(3*) For any event x, if you appeal to another event which is the cause of x, then you first have to
know whether y happened in order to know whether x happened.
…
(C*) If you appeal to a cause every time you have to know whether an event happened, then you
will never know whether any event happened. [1*-5*]

In both cases the regress is entailed and it follows that the given explanation or procedure fails.
Yet, Passmore claimed that the failure holds only in the case of (P*), not (P). Thus where is the
difference? The difference can be explained by the fact that Passmore rejects premise (3) of (P),
but not premise (3*) of (P*). Rejecting premise (3) of (P) means that if you appeal to event y
which is the cause of event x, then you need not explain why y exists in order to explain why x
exists. Compare:

We can, if we want to, go on to look for the cause of the cause and so on ad infinitum, but
we need not do so. (Passmore 1961: 29)

We may therefore explain event A by appealing to its cause B, explain event B by appealing to
its cause C, etc., and indeed break off our investigation whenever we like. Furthermore, if (3) is
rejected, then the truth of (P)’s conclusion, i.e. that the explanation is ineffective, is not
established. By contrast, in the case of (P*) Passmore concedes (3*) and endorses the conclusion
that the procedure is ineffective. So the idea is that we may come to know whether event A
happened by appealing to its cause B, but in that case we should first know whether or not
indeed B happened.

Hence, the distinction between regresses which demonstrate that a certain procedure is
ineffective and those which do not comes down to the question of whether line (3) of the Failure
Schema II is applicable. If one cashes out Passmore’s suggestions in the way I have done here,
the first objection against his Failure-hypotheses is met.

2.2. Objection (2)

Black (1996) introduces three important worries about Passmore’s account. His first objection is
that the reason Passmore’s gives to explain why regresses entail inadequate explanations does
not work (1996: 113). This reason is the following. Regresses demonstrate that a certain putative
explanation fails to explain

not because the explanation is self-contradictory, but only because it is, in the crucial
respect, of the same form as what it explains. (1961: 33)

For similar a suggestion, compare:

 Each member of the regress is numerically distinct from that which precedes it; qua
explanans and explanandum, however, they are identical. And in so being, we never
break out of the explanatory failure that gives rise to the regress. (Bliss 2012)

As we can see, Passmore contends that explanations fail if they are, in some crucial aspect,
similar to what they explain (and that regresses would demonstrate this). Yet according to Black
this contention is false, for it allows of counterexamples. His own counterexample is the
following: one has blue eyes because one’s parents have blue eyes.76 In this case, the explanans
and explanandum are similar, but the explanation is not inadequate (or at any rate need not be
so). So fails the reason Passmore gives to explain why regresses entail inadequate explanations.

Reply

I reject Black’s objection, not because I think that his counterexample is a poor explanation but
because I regard it as an unsuccessful counterexample against Passmore. To see this, let us
consider the example in terms of the Failure Schema:

Eyes (Failure II instance)

(1) You have to explain for at least one person why she has blue eyes.
(2) For any person x, if you have to explain why x has blue eyes, you appeal to the fact that x’s
parents have blue eyes.
(3) For any person x, if you appeal to the fact that x’s parents y and z have blue eyes, then you
first have to explain why y and z have blue eyes in order to explain why x has blue eyes.
…
(C) If you appeal to the fact that x’s parents have blue eyes every time you have to explain why x
has blue eyes, then you will never explain why at least someone has blue eyes. [1-5]

Hence the explanatory failure. Of course, it is not difficult to resist the conclusion by rejecting
(3), as the latter is a rather strict requirement on explanation (Do we really want to require that
we have an explanation for the colour of the parents’ eyes in order to explain the colour of the
child’s?). The point, however, is that the conclusion follows logically and therefore that regresses
can demonstrate the inadequacy of an explanation (i.e. if the premises are true).

According to this reconstruction, Passmore is right to claim that the explanation does not fail due
to a contradiction. The explanation that appeals to x’s parents conflicts with none of our
assumptions. Also, the explanation that appeals to x’s parents fails not because it is similar to the
explanatory problem described in line (1) (i.e. to explain why x has blue eyes), but rather
because it gives rise to an explanatory problem similar to the initial one (i.e. to explain why x’s
parents have blue eyes) and which, given premise (3), must itself be solved first in order to solve
the initial problem.

Hence, if one cashes out in this way the reason Passmore gives to explain why an explanation
may be inadequate due to a regress, this reason does not fail as Black’s first objection has it.
Gratton voices two worries in the same direction (2010: 145-6, cf. 1994a, 1994c). Specifically,
he takes issue with the claim that regresses can be vicious because a given explanation is
circular, and with the claim that regresses can be vicious because they specify some endless task.
Yet, even if these claims do not make sense in the context of Gratton’s preferred Paradox
Schema, they do make sense in the Failure-context. Let me briefly show this for either claim.

In the Paradox Schema, nothing can function both in the generation of a regress and in the
argument of why that regress is vicious (in the sense of entailing something unacceptable). These
are two different steps. Yet this is exactly what seems at issue: the circularity of the explanation
is used both to generate the regress and to argue that the regress is vicious. Take the following
example:

(1) For any event x, x is explained only if x is explained by another event y and y itself is
explained.
(2) At least one event is explained.
(3) Regress. [1-2]
(4) The explanation is circular. [3]
(5) Circular explanations are unacceptable.
…

In this context, the explanation of some event x is not circular in the sense that it explains itself,
but rather in the sense that it is explained by y, which is another event (events explain events). I
take it that Gratton holds that this results not from the regress, but rather from line (1), i.e. one of
the regress formulas which is used to generate the regress. The regress would contribute nothing
to an argument that the explanation is circular.

Gratton’s point is correct, yet his objection is a straw man. His opponents are Passmore and
Rankin:

a supposed way of explaining something […] in fact fails to explain, […] because it is, in
the crucial respect, of the same form as what it explains. (Passmore 1961: 33)

Regresses are vicious when, and only when, symptomatic of circular explanation.
(Rankin 1969: 178)

We have seen Passmore’s text above, where I explained it in terms of the Failure Schema, not
the Paradox Schema. To review this: that an explanation is circular means, in this context, that
the explanandum gives rise to a similar explanatory task which needs to be met before the initial
task can be met. This in itself is no problem. Were one to hold onto the same kind of explananda,
however, the initial task would never be met, and it is exactly this that makes Failure regresses
vicious. In this way, therefore, regresses can be vicious because a given explanation is circular.77

Next consider Gratton’s second objection, here expressed at some length:

Can a regress entailed by a circular explanation be vicious because of some obligation to

go through each step of the regress, thereby having to end an endless task? This approach
misfires […] for if it were possible to complete the process, the explanations would
 remain circular, and nothing would be gained. Similarly, nothing is lost by not
completing the process. (2010: 145)

Here Gratton considers an alternative to line (4) in the reconstruction above: the result taken
from the regress is now, namely, that ‘we have to end an endless task’. This may turn out to be
impossible, such that the regress is vicious, yet Gratton’s point is that this would be irrelevant
given that the circularity of the explanation still follows from line (1).

All of this, to say again, make perfect sense in the context of the Failure Schema II. Consider for
instance the argument (P) discussed above, where the problem is explanatory. In that case, you
explain an event by other events and are obliged to go through each step of the regress even
though you will never succeed in your explanation, for the regress is endless.

All in all, then, the second set of worries about Passmore’s Failurehypotheses rest on
misunderstandings as well.

2.3. Objection (3)

Black’s next objection (1996: 113-4) is that regresses are not mere rhetorical tools, as Passmore
assumes them to be. According to Black, regress arguments do not work, logically, unless the
relevant regresses are infinite. Passmore writes, in contrast:

It is the first step of the regress that counts, for we at once, in taking it, draw attention to
the fact that the alleged explanation or justification has failed to advance matters; that if
there was any difficulty in the original situation, it breaks out in exactly the same form in
the alleged explanation. (1961: 31)

Reply

Though Passmore never explicitly claims that regresses are mere rhetorical tools, he does say
that the first step of a regress is the important one and that its further steps (indeed its infinity of
further steps) serve merely to bring out the same worry expressed in the first step. This same idea
appears elsewhere in the literature:

The real trouble arises already at the first step: if it is rightly diagnosed there, we can
forget about the regress. (Geach 1979: 100)78

The real trouble arising already at the first step is that of making no progress. We should
see this straight away. (Sanford 1984: 96)

Hence, it is not unfair to ask: Is everyone here really mistaken, as Black claims? Again, I shall
show that Passmore’s claim makes sense once we have both schemas at our disposal.

In the Paradox case, if the first step of the regress is the crucial one, then already the regress is in
conflict with something else that we believe. For example, Plato’s Third Man Regress is already
in conflict with the (possible) belief or commitment that there is only one form of Largeness as
soon as the second form of Largeness appears on the scene (cf. Vlastos 1954: 328, Geach 1979:
100-1). Or consider Ryle’s regress in Paradox format, from which it follows that I perform an
infinity of intelligent actions (see §3.4.5). Arguably it is already beyond my capacities to
contemplate the proposition that [the proposition that [dissertations are to be written in such and
such a way] is to be contemplated in such and such a way], let alone to contemplate it
intelligently. This would mean that line (4e) is itself already sufficient to obtain an unacceptable
result from the regress. (If the steps are the same as the lines in the argument, then this is strictly
speaking not the first step, though at any rate it is the beginning of the regress.)

In the Failure case, if the first step of the regress is the crucial one, then it is already clear that
strategies similar to the initial one will never make it the case that the initial task is
accomplished. Take Ryle’s regress in Failure format:

(a) You have to write a dissertation intelligently.

(b) You employ knowledge that [dissertations are to be written in such and such a way]. [a, 2]
(c) You have first to contemplate intelligently the proposition that [dissertations are to be written
in such and such a way]. [b, 3]
(d) You employ knowledge that [the proposition that [dissertations are to be written in such and
such a way] is to be contemplated in such and such a way]. [c, 2]
(e) You have first to contemplate intelligently the proposition that [the proposition that
[dissertations are to be written in such and such a way] is to be contemplated in such and such a
way]. [d, 3]
etc.

From line (c) onwards it is clear that similar solutions (in this case: employing knowledge-that)
shall always entail similar problems (in this case: tasks to perform intelligent actions) which
must be accomplished in order for the initial task to be accomplished.

In both cases, therefore, it makes sense to say that the beginning of the regress is what is
important (and not the infinity). Moreover, as I argued in §2 and §4.2.3, the beginning of the
regress is all that is needed from a logical point of view: only a few steps of the regress are
needed to derive a Paradox or Failure conclusion from it. So much, then, for objection 3.

2.4. Objection (4)

Before turning to Black’s final objection, let us consider one objection posed by Schlesinger
(1983: 221-7). He claims that the Failure Schema is involved in an ambiguity, and that it is
possible to draw another, more optimistic conclusion as well. According to him, what is to be
drawn from regresses consisting of problems and solutions is a matter of debate. If NN1 and
NN2 are again two arbitrary persons, then Schlesinger takes the dialectical situation to be the
following (also cf. Oppy 2006: 289):

Step Context

Initial problem P1 Common concern of NN1 and NN2

Initial solution S2 NN1’s proposal

Similar problem P2 NN2 shows that NN1 has to solve this

Similar solution S2 NN1 solves P2 in a similar way

Similar problem P3 NN2 shows that NN1 has to solve this

etc. etc.

The question is what follows:

Are we to say that, since essentially the same problem keeps arising no matter how far we
progress along the regress, we are faced with an ineradicable problem, or that, since
every time we raise a problem we can at once come up with a solution, we are left we no
difficulty? (Schlesinger 1983: 221)

Hence, Schlesinger identifies two conclusions:

Pessimism. Every solution entails the same (kind of) problem, so in some sense the problem is
ineradicable.

Optimism. For every problem there is a solution, so in some sense we are left we no difficulty.

This is a sensible query. Take McTaggart’s example (also discussed by Schlesinger):

P1 You have to eliminate the contradictions in the A-series.

S1 You appeal to a second-order A-series.
P2 You have to eliminate the contradictions in the latter.
S2 You appeal to third-order A-series.
P3 You have to eliminate the contradictions in the latter. etc.

Furthermore, as some have suggested (Prior 1967: 5-6, Schlesinger 1983: 127), Pessimism
prevails if we stop at one of the problems (for in that case a problem is left unsolved), and
Optimism if we stop at one of the solutions (for in that case no problem is left unsolved).

Reply

Still, granting all this, in my view Failure instances do not fall prey to the Optimism/Pessimism
ambiguity. This seems clear for four reasons.

First, so long as all problems and solutions are entailed, it is not possible just to stop at a problem
(and so land in Pessimism) or at a solution (and so end with Optimism). And here it seems that
all problems and solutions are entailed, at least within the context of the Failure Schema (i.e. by
lines (1)-(3)).
Second, the dialectic invoked by Schlesinger is not exactly the one which I myself presented in
the discussion of Failure arguments above (in §4.1.5). Schlesinger’s dialectic differs, that is, in
its being between someone who poses problems and someone who proposes solutions for them.
The Failure dialectic I presented, by contrast, is between someone who purports to solve a
problem and someone who shows that the former never succeeds in doing so (as each time yet
another, similar problem has to be solved in order for the initial problem to be solved). In the
latter case, the Optimism/Pessimism controversy does not apply. Either the problem is ever
solved, or it is not: it is not solved half of the time.

I consider this second point decisive. However, one might still suspect that Optimism could
apply to Failure I instances, which are considerably weaker than Failure II instances. Let us
consider again McTaggart’s case, which is arguably a Failure I instance. So long as there is an
endless number of A-series, the contradictions in all A-series can be eliminated by appealing to
further A-series. In general: so long as the series of solutions is endless, all problems can be
paired off with a solution.

Also compare Russell’s Tristram Shandy case (introduced above in §1.2). To recall, Shandy
thinks his life is so interesting as to merit being reported. This problem can be solved by writing
an autobiography. As it happens with Shandy, he writes so slowly that it takes him one year to
report one day of his life. At the end of that year, of course 365 interesting, unreported days
await him. The same story holds for the next 365 days, and so on. Hence: as soon as Shandy
starts working on his autobiography, the unreported days of his life increase explosively.

Now Russell’s claim is that if Shandy were to have the eternal life, then no day of his life would
remain unreported. This may appear counterintuitive, yet it is correct. If Shandy’s life were never
to end, then all the unreported days of his life and autobiographical reports can be paired off
against one another:

Reported days 12345…

Working years 12345…

Now, if all problems can thus be paired off with a solution, does this constitute a worry for
Failure I arguments? It would do so only if the fact that all problems have a solution were to
conflict with something stated in the Failure I argument. Yet here this may not be the case.
Consider the following two claims:

(i) There is a solution for all problems.

(ii) There is always yet another problem to be solved.

The first claim is Optimism, and might apply to a selected number of Failure I arguments. The
second claim is what is demonstrated by a Failure I regress, in all cases. Here there is no conflict,
for neither claim excludes the other. Still, the implications of these claims do seem to conflict:

(iii) If (i), then all problems are solved.

(iv) If (ii), then it is never the case that all problems are solved.

Indeed, it cannot be both that all problems are solved and that it is never the case that all
problems are solved. Nonetheless, I do not think that Failure I arguments are afflicted by this
conflict. What Failure I arguments show is that it is never the case that all problems are solved in
the sense that there is always a further problem to be solved, whether or not there is a solution
for all of them. In McTaggart’s case, therefore, this would mean that even if all contradictory A-
series can be eliminated by appealing to further contradictory A-series, still it would follow from
the regress that there are always further contradictory A-series to deal with and that you will
never deal with all such series in this sense.

My fourth and final point is that similar queries have been raised in the discussion of supertasks.
For various supertasks seem possible, even though they consist of an infinity of tasks.

As explained, though, the problems with supertasks do not automatically carry over to Failure
arguments (see §4.2). It is right that both supertasks and Failure regresses consist of an infinity of
tasks, yet the relevant issues about them are distinct. In the case of a supertask, the main question
is whether it can be performed in a finite period of time. More specifically, the question is
whether such tasks are possible both conceptually and physically. To illustrate, consider again
Thomson’s Lamp (see §4.2.1 above). In this case, one of the issues is whether the lamp is on or
off after two minutes. Initially, Thomson (1954) argued that any answer to this question entails a
contradiction such that supertasks like this had to be conceptually possible. Benacerraf (1962)
responded that Thomson’s argument was invalid. No matter the details of this debate, for my
purposes it suffices to point out that these are not the main issues that pertain to Failure
arguments. For in the case of a Failure regress, the main question is not whether it is
conceptually and physically possible to accomplish all generated tasks, but rather whether the
first, initial task will ever be accomplished (given a certain solution). In addition, the time factor
plays no role in Failure arguments (as constructed in this dissertation). This means that worries
about supertasks do not automatically carry over to Failure arguments.79

Given these four points, there seems to be here no room for Optimism, and so the fourth
objection fails as well.

2.5. Objection (5)

The final objection that I will discuss is Black’s main point (1996: 111, 114- 5). His concern is
that regress arguments are not usually connected with explanatory concerns, and that regress
arguments are not about inadequate explanations, but rather about rejections, i.e. about
demonstrating the falsity of a proposition.

Reply

On first sight, this objection would seem to fail simply because it begs the question against
Failure-Monists and Pluralists. For to say that all regress arguments are about rejections and not
about inadequate explanations (or bad solutions generally) is to assume, rather than to
demonstrate, that all regress arguments should take the form of the Paradox Schema. Yet this is
exactly what is under discussion. Still, I have three more precise points.

First, I agree that not all regress arguments have as rationale that a given explanation fails. Take
Ryle’s case in its Paradox format. Here the lines which might possibly express an explanatory
claim are the regress formulas. The following, for example, might express such a claim (however
schematically): For any action x, one intelligently performs x only if one employs knowledge
that x is to be performed in such and such a way. In this case, one’s employment of knowledge-
that is a necessary condition for performing an action intelligently. But is it also an explanatory
condition? According to Black it need not be, for according to him it is controversial to assume
that all the necessary conditions of something are involved in the explanation of that thing.

Furthermore, if the regress formulas do not express explanatory claims, then it is hard to see how
demonstrating that they are false (on the basis of a Paradox argument) amounts to demonstrating
that an explanation has failed.

A similar point is made by Schnieder (2010) in response to Gaskin (2008). The latter claims that,
in the following series, each step is explained by the next (and previous) step because each is
necessary for the other:80

(a) a is F;
(b) a stands in the instantiation1 relation to F-ness;
(c) a and F-ness stand in the instantiation2 relation to instantiation1; etc.

In response, Schnieder says:

Why should we believe that there is any explanation going on? (2010: 296)

Schnieder’s main argument is that explanations usually involve epistemic goals. For example,
explanations may improve our understanding. In that case, if A explains B, then A somehow
contributes to our understanding of B. Yet, in the present case it is not immediately clear how (b)
contributes to the understanding of (a), and (c) to the understanding of (b), etc. The point is,
basically, that if A is a necessary condition for B and if A is to explain B as well, then one needs
a separate story as to why this latter, explanatory relation obtains (cf. also Klein 2003: 722).81

Even if regress formulas do not always express explanatory claims, sometimes they might.
Consider the following, simple example: For any event x, x exists only if x is brought into
existence by another event y. In this case, it is plausible to think that the fact that an event is
brought into existence by another event is not only a necessary condition for the fact that the
former event exists, but also, on top of that, reports an explanation. On the basis of a Paradox
argument we could obtain the negation of this formula: It is not the case that any event x exists
only if, and because, x is brought into existence by another event y. As we can see, however, this
still shows not that an explanation is inadequate, but merely that a universally quantified claim
stating an explanatory condition is false.
Nevertheless, regress arguments can sometimes be about inadequate explanations, namely (i) if
they are taken Failure-wise, (ii) if the problem is an explanatory problem, and (iii) if the
considered solution fails due to a regress. This does not, for example, apply to the Failure-
version of Ryle’s case (where the problem concerns our performing an action in a certain way,
and not our explaining anything). It may, however, hold for Bradley’s case, if the latter is taken
to be concerned to explain (and to contribute to our understanding of) how any two things are
related by a relation (or, as in Gaskin’s case, how any proposition forms a unity). It also holds for
the Passmore case (P) discussed above. This is my first point.

Second, and explained already above in §4.1.3, instances of the Failure Schema are not about
rejections, not even the rejection of solutions. They do not, that is, prove that a solution is false,
but just that it is no good for solving a given problem (and therefore that another solution must
be found).

Third, even Passmore himself is not committed to the claim that all regress arguments are about
inadequate explanations. For he speaks about procedures (1961: 29), and ‘procedures’ is just
another term for the ‘solutions’ of the Failure Schema. Examples of procedures/solutions
mentioned by Passmore include: to provide a criterion, a justification, an explanation, a
definition.

In summary, I have shown that all five of the main worries about Passmore and the Failure
Schema misfire. This does not serve to establish Failure-Monism, though it does establish the
possibility of Failure-Monism, and indeed too the possibility of Pluralism (my own position).

3. Gratton’s Failure-hypotheses

In the following I consider Gratton’s so-called ‘problem and response regress arguments’ (2010:
ch. 6) and compare them to instances of the Failure Schema. This comparison is useful because
regresses generated in the Failure Schema also consist of problems and responses/solutions, yet
differ in certain important respects. This will allow me, then, to clarify from another angle my
choices in the set-up of the Failure Schema. Between Gratton’s arguments and the Failure
arguments there are four significant differences (henceforth I abbreviate ‘problem and response
regresses’ to G regresses and ‘problem and response regress arguments’ to G arguments):

(1) G regresses take problems as arguments, Failure regresses take them as tasks.
(2) Failure regresses are always entailed, G regresses are not (at least usually not).
(3) Failure regresses can be infinite, and yet Failure arguments work; G arguments do not work if
G regresses are infinite.
(4) G arguments conclude with rejections, Failure arguments do not.

So, even though Gratton’s view that arguments based on regresses consist of problems and
responses/solutions happens to differ from Failure instances, it may still be called a ‘Failure-
hypothesis’ insofar as it is meant as a proposal to make such arguments explicit.

Before explaining these differences, let us consider an example of a full G argument, i.e.
McTaggart’s case (see his 1908: 468-9, 1927: §§325- 33). I will present this case in two parts:
first the regress, then the argument. The regress resembles Gratton’s Diagram 3 (2010: 164)
except that in Gratton’s Diagram the problems P1 and P2 appear to be identical rather than
merely similar. Yet importantly, recurring problems are to be merely similar to one another (i.e.
of the same type), not identical. Were they identical, it would follow immediately that response
R1 fails to solve problem P1. For if R1 entails P2 and P1=P2, then R1 entails the problem it is
meant to solve.

McTaggart G-wise

P1 Any event has the incompatible characteristics of being past, present and future
simultaneously. [Further premises.] So, time does not exist.
R1 Any present event is present at the present moment, future at a past moment, and past at a
future moment (any past event is present at a past moment, etc.). So, events have these
characteristics successively, and the first premise of P1 is false.
P2 So, any moment has the incompatible characteristics of being past, present and future
simultaneously. [Further premises.] So, time does not exist.
R2 Any present moment is present at the second-order present moment, future at a second-order
past moment, and past at a second-order future moment (any past moment, etc.). So, firstorder
moments have these characteristics successively, and the first premise of P2 is false.
P3 So, any second-order moment has the incompatible characteristics of being past, present and
future simultaneously. [Further premises.] So, time does not exist.
etc.

Here is the argument: (i) The regress cannot go on forever. So, (ii) for some natural number n,
problemn remains unresolved. For example, assume that P3 above remains without solution. (iii)
If, for some n, problemn is unresolved, then the responsen-1 which entails problemn entails an
unresolved problem. So, (iv) R2 entails an unresolved problem. (v) If something entails an
unresolved problem, it is unacceptable and must be rejected. Therefore, (vi) R2 is to be rejected.
In other words: It is not the case that first-order moments are past, present and future
successively. (And this in turn would mean that first-order moments are contradictory and that
time does not exist.)

3.1. Difference (1) The first striking difference between Failure and G arguments is that
problems in the latter are arguments rather than tasks to be accomplished:

McTaggart’s infinite regress argument of recurring problems and responses consists of a

succession of exchanges between opposing arguments. (2010: 163, cf. Mellor 1981: 94,
Rankin 1981: 337)

More specifically, problems are arguments with counterintuitive conclusions, and solutions are
counterarguments against such arguments. Given this alternative conception, one may wonder
why I did choose for the tasksformat in my Failure Schemas. Basically, the argument is that this
allows problems to figure as premises in arguments. In Failure arguments, problems had to
occupy the first line, and this does not work with arguments with counterintuitive conclusions.
At least it is not obvious to me how it would work. Also, as I shall explain below, G arguments
are not able to do what Failure arguments can do.
One may think that this, first difference is not very pertinent. Yet it is worth noting that Gratton
himself distinguishes among three kinds of regress: regresses of definitions and explanations
(2010: ch. 4), regresses of questions and answers (2010: ch. 5), and indeed regresses of problems
and responses (2010: ch. 6). Furthermore, no such distinctions are needed if providing
definitions, providing explanations, and answering questions are all regarded as a variety of
tasks. Why should we want to treat all these cases in the same way? My answer here is simple: If
there are no structural differences between certain groups of regress arguments, then we should
not treat them differently, even if they are presented differently in the literature. Better to have
one single and widely applicable story as to what regress arguments are, rather than a bunch of
them. Of course, this dissertation is about two theories of regress arguments, not one, but that is
because I think there are significant differences between Paradox and Failure arguments.

Indeed, in my view there are no structural differences between regresses of definitions and
explanations, of questions and answers, and of problems and responses. Take one of the cases
discussed by Gratton himself (2010: 149-50): at issue is what free action consists in. The regress
here runs in terms of questions and answers:

Q1 What makes act1 voluntary?

A1 A voluntary act2 of the will.
Q2 What makes act2 voluntary?
A2 A voluntary act3 of the will.
etc.

Nothing seems lost, however, by restating it as follows:

P1 You have to explain what makes act1 voluntary.

S1 You appeal to a voluntary act2 of the will.
P2 You have to explain what makes act2 voluntary.
S2 You appeal to a voluntary act3 of the will.
etc.

The same applies to other cases, and so I take it that all must be treated in the same way, namely
as a variety of tasks and actions.

3.2. Difference (2) An important difference between Failure and G regresses is that the latter are
usually not entailed, whereas the former always are.

According to Gratton, problems can ‘recur’ in several ways depending on the entailment relation
between the problems and responses, i.e. (i) whether or not the responses entail new problems,
and (ii) whether or not the problems entail new responses (making for four logical options). For
example, McTaggart’s case is one in which all responses entail new problems, but no problem
entails a new response. If the responses are not entailed, then neither is the whole regress
entailed, and it stops once you recognise that the same pattern would continue, or once you stop
out of “exhaustion, boredom, or insanity” (Gratton 2010: 166).
To begin with, I have doubts about Gratton’s contention that any given case can belong to one
category only. For if the ways in which problems can recur are structural patterns, then it can be
expected that any case can follow any pattern depending on how it is reconstructed. More
importantly, Failure regresses differ from G regresses in that they are always entailed. They are
entailed, that is, by the previous solution or problem in the regress in combination with the first
lines (1)-(3) of the Failure Schemas. Among the Failure instances these patterns are completely
general. The only variation is that between Failure I and II. In the case of Failure I, any problemn
is entailed by solutionn-1 plus the extra premise (3) and the general problem (1). In the case of
Failure II, any problemn is entailed by solutionn-1 plus the extra premise (3). In both cases, any
solutionn is entailed by problemn plus the general solution (2).

3.3. Difference (3)

While Failure arguments can be good arguments if the regresses at hand are infinite, G
arguments do not work if G regresses are infinite. The reason for this is simple: If G regresses
were infinite, then there would be a solution for every problem and therefore no problem left
unresolved (cf. Schlesinger’s Optimism discussed above). As we have seen in the reconstruction
of McTaggart’s case, Gratton assumes that G arguments take the following form:

(1) Assume that for some n, problemn is unresolved.

(2) If problemn is unresolved, then the solutionn-1 which entails problemn entails an unresolved
problem.
(3) If something entails an unresolved problem, it is unacceptable and to be rejected.
(4) Hence, solutionn-1 is to be rejected. [1-3]

If it is necessary for this argument that a problem in the regress remains unresolved, then it is
necessary for this argument that the regress be finite, namely in the sense that it stops at the
unresolved problem.

Failure arguments differ completely from this. Here is a reconstruction of McTaggart’s case in
Failure I format:

Time (Failure I instance)

(1) You have to eliminate the contradictions in all A-series.

(2) For any A-series x, if you have to eliminate the contradictions in x, you appeal to a higher-
order A-series y such that the members of x are past, present and future at different members of
y.
(3) For any A-series x and y, if the members of x are past, present and future at different
members of y, then y is contradictory as its members are both past, present and future.
…
(C) You will never eliminate the contradictions in all A-series if you appeal to a higher-order A-
series every time you have to eliminate the contradiction in an A-series. [1-5]

None of these lines correspond. In this Failure reconstruction, we neither assume that some
problem remains unresolved, nor specify a sufficient condition of unacceptability, nor conclude
with a rejection of one of the solutions in the regress (more on this last point below). Hence, the
issue as to whether or not Failure regresses must be infinite for the argument to work cannot be
decided on the same basis. This latter issue I addressed above in §4.2.3.

3.4. Difference (4)

According to Gratton, the goal of G arguments is not to reject regress formulas (as in Paradox
arguments), but rather to reject responses located within the regress. In McTaggart’s case, for
instance, we rejected R2: It is not the case that first-order moments are past, present and future
successively. This differs completely from Failure arguments, which conclude that the initial
problem is never solved by the given solution (see (C) just above).

6.2. Comparison Schemas II: Soundness

In the following I address the problem: If regress arguments can be reconstructed along the lines
of both the Paradox and Failure Schemas, then which schema is to be preferred? To address this
question I will evaluate both schemas, and eventually conclude that the Failure Schema often has
better (i.e. more useful) instances.

1. Schema choice

At this point the Paradox Schema (which captures what I called the Paradox Intuition) and the
Failure Schema (which captures the Failure Intuition) are on par with one another. Both, we have
seen, have a regress on one of their lines, which is derived from other lines and has a conclusion
associated with it (§2). And both, moreover, have logically valid instances (§2), both can have all
instances from the literature as an instance (§3), and both can account for disagreements about
regress arguments (§5).82 Yet they differ structurally (§4), leaving us with the question: Which
schema is more fruitful? Which should be used to reconstruct regress arguments? Or should both
be used?

By the Interpretation Rules, the choice of schema depends on the text and context of the initial
statement of the argument. By the Charity Rules, the choice of schema depends on which schema
has better arguments as instances. Here the rules need not converge and this is a problem, as
noted in §1.3.3. Three possible solutions were identified: Either defend that Interpretation always
wins out, or that Charity always wins out, or that sometimes Interpretation wins out and
sometimes Charity. I also anticipated my preference for the second option: Charity is more
important.

My argument is simple: So long as our goal is to reconstruct an argument (and not merely to
interpret it), it is preferable to reconstruct good arguments rather than bad ones. Why? Here is
what Feldman says to motivate adherence to the Charity Rules:

We should adhere to [them] not because it is nice to do so or because people need or

deserve charity, but because adhering to [them] leads us to consider the best available
arguments and thus to gain the most insight into the issue we are studying. (1993: 115)
This is exactly right. We should try to obtain the best argument from a given case in order to see
what can be demonstrated by it, without being restricted by what the source text actually says or
what the author actually meant. For in that case we stand to gain the most insight into the issue
we are concerned about (e.g. about justification as in Sextus’ case, about induction as in Hume’s
case, or about meaning as in Wittgenstein’s case).

But what are good arguments? Here the remaining desiderata for the schemas (from §2.1.2)
become relevant. The schema should:

• be as simple as possible;

• produce plausible premises/interesting hypotheses;

• produce conclusions that are hard to resist;

• produce conclusions that can play an interesting role in a debate.

Perhaps there are other such desiderata, but these seem to me the most important. The debate
between Paradox-Monism (reconstruct always Paradox-wise), Failure-Monism (do it always
Failure-wise), and Pluralism (do it sometimes Paradox-wise, sometimes Failure-wise) will turn
on these desiderata.

The basic idea is this (and here I am assuming an unrestricted use of the Charity Rules, as just
discussed). If it turns out that Paradox arguments are always simpler, and always have more
plausible premises, more interesting hypotheses, and stronger and more interesting conclusions,
then Paradox-Monism holds. If, by contrast, it turns out that Failure arguments are always
simpler, and always have more plausible premises and all the rest, then Failure-Monism holds. If
Paradox and Failure arguments both score well, though on different points, then some sort of
Pluralism holds.

I will argue next that the Failure Schema scores better on most points (that is, its instances are
often better arguments than their Paradoxcounterparts), which in any case rules out Paradox-
Monism. We shall see about the other positions later.

Before arguing that the Failure Schema scores better on most of the desiderata, let me present a
quick argument for the Paradox Schema. Namely this: it is simpler. Perhaps it does not have the
least number of inferences, but it has the simplest ones in the sense that they are easier to
understand. By contrast, some steps of the Failure Schemas are not immediately clear. The
reason here is that its suppressed premises are more complex, particularly the following (which
are lines (6) of Failure I and II respectively):

• If for at least one K you have to φ it and if for all Ks x, you always have to φ a new K in
addition to φ-ing x, then you will never φ all Ks.

• If for at least one K you have to φ it and if for all Ks x, you first have to φ a regress of new Ks
in order to φ x, then you will never φ x.
Admittedly, if this is the only thing, then the two schemas do not differ very much on this point,
but as I regard simplicity as an important virtue of arguments, I do regard this as an argument in
favour of the Paradox Schema.

Still, simplicity works only if all the rest is in order as well. Simple but implausible or
uninteresting arguments are surely not to be preferred over arguments that are slightly more
complex, but plausible and interesting. So let us see about that.

2. In favour of the Failure Schema

In the following I shall argue that the Failure Schema is the better of the two schemas because it
produces better arguments (even if they are, as just noted, slightly more complex). I will present
three arguments, each connected to one of the remaining desiderata:

(1) The conclusion of Failure arguments is immediately relevant, and this need not be the case in
Paradox arguments.
(2) The hypothesis of Failure arguments is immediately motivated, and this need not be the case
in Paradox instances.
(3) The conclusion of Failure arguments is stronger, as it allows of fewer options for resistance.
The first argument will be the most important.

The three arguments are related in that the strength and relevance of the conclusion partly
depends on the number and plausibility of the premises. All three will be explained with the aid
of the dialectical scenarios outlined for both schemas in §4.1.5, i.e. the scenarios where all steps
of the schemas are linked to two opponents, NN1 and NN2. At the end I shall explain what these
arguments entail with respect to the debate between the Monisms and Pluralisms.

2.1. Argument (1)

The Paradox Schema concludes that one of NN1’s beliefs is false, and the Failure Schema
concludes that NN1’s solution fails. Basically the point here is that the falsity of NN1’s belief
need not be relevant if it plays no role in a certain debate, whereas the failure of NN1’s solution
is automatically relevant in at least one debate, i.e. the debate on how to deal with the problem it
tries to solve.

To explain this point in some more detail we have to consider the conclusions of the schemas,
and the relevance of each. Let us first consider the Paradox conclusion:

PARA It is not the case that any K is F only if such and such (e.g. it stands in R with another K).

The question is: So what? Well, in short, now you know that the universally quantified statement
does not hold (although its existentially quantified counterpart can still hold). One case where
this seems interesting is the regress argument from Sorensen (1995) and Sider (1995). Their
discussion concerns the Access principle which says that you are obliged to do something only if
you can have knowledge of that obligation. Now, if you are interested in whether this principle
holds unrestrictedly, then on the basis of a Paradox argument it is interesting to find out that the
principle does not always hold (see §5.5 above for discussion of this argument).

Also, Beth (1952: §4, cf. Peijnenburg 2011) shows that in the history of philosophy the Paradox
conclusion has sometimes been used to establish absolute entities, i.e. entities which stop the
regress at some point and form an exception to the universally quantified statement at hand, such
as the Unmoved Mover, which sets the series of moved things in motion but is not itself moved
by anything else.

Still, in many cases this form of conclusion is irrelevant, i.e. it makes no difference to any
particular debate. It is uninteresting, for example, to find out that it is not the case that for any
person x, x is reliable only if x is guarded by a guardian (especially given that this may still hold
for some, if not many, persons), or that it is not the case that for any set of premises x, a
conclusion follows logically from x only if x contains the additional premise ‘if the members of
x are true, then the conclusion is true’ (again, given that this may still hold for some, if not many,
sets of premises).

Next consider the Failure-conclusion:

FAIL You will never φ any/all K(s) if you ψ any K that you have to φ.

Again the question arises: So what? Well, again in short, now you know that if for at least one K
you have to φ it, then you have to find another solution to solve this problem. If the problem is a
common concern of both NN1 and NN2, then this is an interesting result in each and every case.
This means that you have to find another solution (another procedure, other means) to perform
an action; to justify a proposition; to have your girlfriend guarded; to demonstrate that a
conclusion follows logically; to secure an obligation; and so on for the other cases.

I regard this point as decisive. I should note that not all Paradox arguments are supposed to refute
universally quantified statements. If the regress formulas happen to hold, then the trigger
statement is to be rejected instead. Some examples: no proposition is justified; no dispute is
decided; the reference of no word is fixed. Such conclusions are in many cases surprising as they
usually concern a commonsense scenario, and so may be interesting in scepticism debates. Yet
the first argument in favour of the Failure Schema does not hinge on this. All that matters for that
argument is that Paradox conclusions are not always relevant in a broader dialectical context,
whereas Failure conclusions often are. There are two further arguments. META-DEBATE 185

2.2. Argument (2)

The second argument for why the Failure Schema produces better arguments is that the
hypothesis of Failure-instances (i.e. their line (2)) is immediately motivated. This line is
motivated because it presents a solution to deal with the problem described in line (1). For
example, you hire a guardian in order to have someone guarded, or you appeal to an additional
premise in order to demonstrate that a conclusion follows logically from the premises.
But it is not always clear why one would introduce the corresponding lines of the Paradox
Schema. For example, why would anyone believe that for any person x, x is reliable only if x is
guarded by a guardian, or that for any set of premises x, a conclusion follows logically from x
only if x contains the additional premise ‘if the members of x are true, then the conclusion is
true’? It is hard to see where these necessary conditions come from and how they can be
motivated (apart from the fact that without them no regress is generated).

2.3. Argument (3)

This last argument follows up on the observation that the Failure Schema has no extra
(substantial) step after the regress. For this difference also implies that Failure arguments are
stronger: there are fewer places where these arguments might be resisted (i.e. they cannot be
resisted by denying that extra step). In the following I will argue why this is a real difference.

Recall the main resistance options for Paradox arguments: (i) reject something else, or (ii) deny
that anything is to be rejected in the first place (by defending that the regress entails no
unacceptable result). It can be shown that one of these options is often possible.

Consider for example Bradley’s case. The conclusion of the Paradox argument is the rejection of
the trigger (i.e. at least two items are related by a relation). Yet, it is quite possible to resist this
argument by rejecting one of the regress formulas instead of the trigger (e.g. by rejecting the
formula that for any relation R, R is unified with its relata only if there is another relation R*
which relates R to its relata). Something needs to be rejected, but logic itself does not decide
what.

Additionally, one might resist the argument by denying that something must be rejected in the
first place, and accept that there is nothing unacceptable about an infinity of relations.

For Failure arguments the situation is different. The main resistance options for the latter are the
following: (i) make an exception to the solution, or (ii) deny the step which leads to more
problems (this differs for Failure I and II). Unlike the options for resisting Paradox arguments,
these options for resisting Failure arguments are in many cases unmotivated.

Take Bradley’s case again, now in Failure II format. To resist it you may try two things. First,
you may reject the extra premise (i.e. ‘for any relation R, if you appeal to a relation R*, then you
first have to explain how R* is unified with its relata in order to explain how R is unified with its
relata), though one cannot reject this without a good motivation. For if it is unclear how R is
unified with its relata, then we seem to have no reason to suppose that an additional relation R*
can explain this (unless of course R* can do something R cannot do, but then the difference must
be motivated). Here is a useful analogy:

Suppose I am given the task of making a chain out of some loose metal rings, and when I
come to join any two of them, I respond by asserting that we need a third ring to do the
job, so that the most I can achieve is just the addition of more rings to the collection. It is
quite clear that no matter how many rings I add, I shall never get a chain. (Candlish 2007:
170)
That is: if no two rings link together unless something additional is supposed to do the job, then
it is plausible to suppose that no three rings link together unless something additional is supposed
to do the job (or again, in all further cases too an explanation must be supplied).

Second, you might make an exception to the solution (i.e. ‘for any relation R, if you have to
explain how R is unified with its relata, then you appeal to a relation R* which unifies R with its
relata’), though again, this lacks proper motivation. If you explain how R1 is unified with a and b
by appealing to R2, then why would you explain, for some n, why Rn is unified with its relata
otherwise than by appealing to Rn+1, but by a different solution?

Let us consider a second case: the Failure I reconstruction of Tarski. To resist it you may attempt
two things. First, you may reject the generality of the problem. For example, you may propose
that the task is to eliminate the Liar Paradox in a certain given language, but not in any language
whatever. Yet this move needs a good motivation. If you are interested in solving the Liar
Paradox itself (rather than in obtaining a certain contradiction-free language), then the paradox
seems to be a problem no matter what language it is formulated in.

Second, you might make an exception to the solution (i.e. ‘for any language x, if you have to
resolve the Liar Paradox in x, then you hold that for no sentence in x it can be said that it is true
or not’). This move lacks motivation as well. If you resolve the Liar Paradox in one language by
a certain solution, then why would you resolve the Liar Paradox in some other language by a
different solution? The motivation cannot simply be that with such an exception the regress is
not blocked, because that would be ad hoc. There needs to be an independent motivation or
explanation for exceptions (cf. Haack 1978: 139). Compare Russell in a different context:

If everything must have a cause, then God must have a cause. If there can be anything
without a cause, it may just as well be the world as God, so that there cannot be any
validity in the argument. (1927: 183)

All in all, Failure arguments are harder to resist. There are at least three possible explanations for
this. Most importantly, as noted in §4.1.3, Failure arguments require no extra substantial premise
after the regress to get to the conclusion. Paradox arguments, by contrast, do require extra
substantial premises for the step that the result is unacceptable such that something can be
rejected. And indeed, one extra substantial premise implies one extra resistance option.

Second, given that Paradox hypotheses may be unmotivated (as pointed out in the previous
argument), it does matter which one is eventually rejected. If so, a Paradox argument can be
resisted simply by rejecting something else.

Third, the conclusions of Failure arguments are in a way more limited. That is, they demonstrate
not that a solution is bad full-stop, but only that it does nothing to help solve a given problem.
This might also explain why Failure arguments are harder to resist.

2.4. My position
Although the Paradox Schema does not always produce good arguments, sometimes it does.
Ryle’s case is a clear case of this. The hypothesis ‘for any action x, one intelligently performs x
only if one employs knowledge that x is to be performed in such and such a way’ is worth
considering for anyone who believes that all our intelligent actions are accompanied by
knowledge-that. Likewise, the Paradox conclusion should be interesting in that it is the rejection
of this claim.

This does not mean, of course, that the Failure conclusion could not be interesting as well. It is
still interesting to find out that you will never perform any intelligent action if you employ
knowledge-that every time you have to perform an intelligent action, for in that case you have to
find some other means to perform an intelligent action (i.e. some other means to solve the
problem stated in line (1) of the argument). In general, then, it might turn out that you end up
with two distinct reconstructions. Yet so long as these are two good and interesting arguments,
this should not be a problem. Let me conclude this section. I just showed that in a broader
dialectical context Paradox arguments play a different role than Failure arguments. Moreover, I
presented arguments for why the Failure Schema often produces better arguments: its
conclusions are stronger and immediately relevant, and the hypothesis is immediately motivated.

How does all this relate to the Monism/Pluralism debate? First, I am not a Monist in the sense
that I take both Paradox and Failure arguments to be logically valid arguments. Also, I am not a
Monist because I think it is possible that both a Paradox argument as well as its Failure
counterpart can be sound arguments in addition to being logically valid. Still, my position is no
standard Pluralism. As I just argued, Failure reconstructions are more likely to be sound for
structural reasons, and so I take it that in most cases we should consider these Failure
reconstructions. In the next section I will single out another apparent distinction between my
view and existing Pluralisms. For the moment, let me label my view thus:

Pluralism*: Do it often Failure-wise, sometimes Paradox-wise.

3. Revisionism

In the foregoing I have done basically two things. First, I have set out the debate on regress
arguments as a debate concerned with two different argument schemas, i.e. two different forms
that a regress argument can take. Next, I have argued that we may often prefer to reconstruct an
argument on the basis of the Failure Schema because that schema produces better arguments. I
have not yet discussed how this pertains to the general debate on argument reconstruction, and I
have not yet provided a deeper motivation for why the various Monisms and standard versions of
Pluralism found in the literature do not suffice (as I defend only a very specific version of
Pluralism). This I will do in the following.

Recall the problem of argument reconstruction: if there is one text and two available
reconstructions (differing structurally qua premises, inferences, and dialectic), then which
reconstruction is to be preferred? My answer can now be formulated as follows. If there are two
reconstructions, then it should be checked which reconstruction will produce the best argument.
Basically, you have to choose the reconstruction with the most plausible premises, and the
strongest and most interesting conclusion (i.e. the conclusion which admits of the fewest options
for resistance and which makes a difference in the broader debate in which it occurs).

This sounds trivial, but it is not. For indeed, it is a rather revisionary take on argument
reconstruction. Revisionism relies heavily on the Charity Rules from §1.1.2, which hold that one
should modify arguments in such a way as to make them logically valid and sound.

Revisionism does not, of course, disregard the Interpretation Rules (i.e. that we should capture
the initial statement of the argument plus context), for in that case no reconstruction would have
any content; they would not be reconstructions of anything. Still, Revisionism is completely
ignorant about what an argument was supposed to be, and focuses solely on what the argument
can be, i.e. on finding its optimal format. For example, Charity is directed not at Ryle’s actual
words, nor at himself or his intentions, but rather solely at the most interesting and strongest
statement of his argument. In some cases, to be sure, it may also be interesting to know what
Ryle himself thought of the matter, whether it be ambiguous, contradictory or just obscure. But
that would be another enterprise altogether: interpretation rather than argument reconstruction.

The opposite of Revisionism may be called ‘Conservatism’. The most extreme variant of the
latter would be that the Charity Rules should not be applied at all: one ought to do with the text
as one finds it, and not add or subtract anything substantial. More moderate variants would
accept the Charity Rules yet hold that they should not be applied unrestrictedly (as Revisionism
has it), but only to a certain extent. It is not clear to me precisely what restrictions might be
imposed here, but the general thought in any case is that one should not depart too much from
the initial statement.

Consider again the debate on regress arguments. Above I distinguished three camps, i.e.
Paradox-Monism, Failure-Monism and Pluralism, and my own position, formerly labelled
‘Pluralism*’, may also be called ‘Revisionary Pluralism’, in order to contrast it both with
Conservative Pluralism and with the Monisms.

Conservative Pluralism would observe that regress arguments in the literature take different
forms, and draw from this the conclusion that sometimes regress arguments are to be
reconstructed Paradox-wise and sometimes Failure-wise. But this view is too easy. First, it
neglects the fact that regress arguments can always be reconstructed in different ways (as I have
shown in §3). One needs only to identify relevant instances of the schematic letters. Yet to
recognise that plurality is possible is not to prove it is desirable. Relatedly, too, Conservative
Pluralism has too much respect for the way in which regress arguments are actually stated and
does not apply the Charity Rules. I do think that regress arguments are sometimes to be
reconstructed Paradox-wise and sometimes Failure-wise, but only because sometimes one of the
reconstructions is better than the other.
No one in the literature, I think, really fits the label of a pure Conservative Pluralist. Yet studies
by Day (1986, 1987) and Gratton (1997, 2010), for example, are far less revisionary than the
view I am proposing here. Compare for instance Day’s comment here:

Bradley’s argument then generates a process regress. I do not mean to suggest that there
is no way in which this argument could be interpreted as generating a product regress. I
am only claiming that it is most faithful to what Bradley says that we interpret him in this
way. (1986: 52-3)

A clear example of a Revisionist is Black (1996). Yet he is no Pluralist, as we have seen, but
rather a Paradox-Monist.

Furthermore, I hold against the various Monisms that it is not always the case that regress
arguments are to be reconstructed Failure-wise (or Paradox-wise for that matter). My view is a
Pluralism exactly because it takes into account that the purposes of the one who is reconstructing
the argument (i.e. what she wants to do with the argument) may vary. Is the reconstructor
interested in producing a Paradox conclusion or a Failure one? That is:

• Does she want to refute an existentially or universally quantified statement?

• Or does she want to show that a solution fails to solve an existentially or universally quantified
problem?

If the former, she should use the Paradox Schema for her reconstruction. If the latter, she should
use the Failure Schema.

Still, this matter of ‘what is interesting’ is not merely subjective or arbitrary. Strong arguments
with plausible premises and relevant conclusions are more interesting than weak arguments with
implausible premises and irrelevant conclusions (and I have shown in the previous section why
Failure arguments often score better at this).

In the literature I have found two important criticisms of the Charity Rules:83

(i) The rules, if applied too strongly, may distort the initial statement of the argument (and e.g.
yield a straw man).

(ii) The rules, if applied too weakly, may not be precise enough to select one reconstruction
among the available ones.

A full treatment of these cannot be given here. Let me briefly point out, however, why they do
not apply in the case of regress arguments. Against (i) it can be said that distortion is no problem
so long as regress arguments from the literature are full of gaps (see §1) and are not meant to be
full of gaps (as perhaps enthymemes, i.e. syllogisms with an unstated assumption, are). Against
(ii) it can be said that selection should pose no problem for regress arguments. The main problem
is to produce sound, or at least logically valid, reconstructions (which is what the rules motivate),
and if it turns out that a single argument allows of two sound reconstructions (as may well be the
case for Ryle’s argument, for example), then you just end up with two sound arguments.

So here is my position on argument reconstruction in a nutshell: Argument reconstruction, at

least in the case of regress arguments, is hardly fixed by the initial text, and should rely more
heavily on the general criteria which arguments have to fulfil in order to be good arguments.84

§7 Epilogue: Two Theories

In the following I conclude this dissertation. First, I check to make sure that everything promised
at the outset has been done. Second, I briefly review what the two theories of regress arguments
set out above can tell us about how regress arguments are to be used and evaluated.

7.1. Taking stock

Here my investigation ends. Let us see whether everything promised at the outset of this
dissertation has been done. Basically, I have tried to tell a coherent story about two theories of
regress arguments, and discuss all major topics relevant to regress arguments (the nature of
regresses and regress arguments, the criterion of viciousness, disputes and meta-disputes on
regress arguments, regress argument reconstruction and evaluation) in terms of these two
theories.

Both theories confirm the idea that regress arguments are pieces of hypothetical reasoning.
According to the Paradox Theory, a certain claim X is considered, yet not taken to be true, for
the sake of deriving a consequence from it that conflicts with independent considerations, such
that X has to be rejected by the hypothetical rule Reductio Ad Absurdum. According to the
Failure Theory, i.e. my original contribution to the debate, a certain solution X to a given
problem is considered, yet not taken to be true, for the sake of deriving a failure from it, such that
‘if X, then failure’ follows by the hypothetical rule Conditional Proof.

The key parts of the two theories are their argument schemas. In §1.3, I argued that such schemas
are worth having for both metaphilosophical and methodological reasons. Now I will indicate
briefly whether the schemas presented above satisfy these needs.

The methodological reasons were three-fold. First, the schemas are useful because regress
arguments have substantive conclusions, and without such schemas we do not know whether and
how these follow from the premises. I tackled this in §2. Second, the schemas are useful for
regress arguments have ambiguous conclusions, and these can be disambiguated on the basis of
such schemas. I discussed some important cases in §5. Third and finally, the schemas are useful
because with them it can easily be seen what premises may be attacked if you do not want to buy
the conclusion of a regress argument. I listed the main resistance options in §4.1.5, and will
return to these in a minute.

The metaphilosophical reasons were two-fold. First, the schemas are useful because they are
what regress arguments from a wide range of discussions (can) have in common. This I showed
on the basis of many instances in §3. Second, the schemas are useful because by them several
disputes in the literature on regress arguments can be clarified and sharpened. I showed this in
§6.1.

Before turning to the two theories, let us return briefly to a worry identified in §1.3.3. The worry
was that regress arguments would be too diverse to exhibit a general argument pattern. As we
have seen, however, there happen to be two such patterns (plus subversions).

First follow-up worry: Why two? To be sure, I did not expect there to be two different patterns
when I started the project. But I think that all I have said in this dissertation suggests that the
Paradox/Failure distinction is useful and unavoidable (and not to be bridged by a unified theory).
Also, nothing in my investigation suggested a need for further argument patterns unrelated to the
Paradox and Failure Schemas. There is no single theory of regress arguments; nor are there, say,
seven. There are exactly two of them.

Second follow-up worry: If the schemas are so general that one can find instances of them from
ethics to epistemology, then one might wonder whether they are really non-trivial. I think these
schemas are quite non-trivial, of course, exactly because they present precise, but still different
guidelines for using and evaluating regress arguments. These guidelines are listed below. As we
shall see, now that everything is in place (particularly the Charity Rules, the regress argument
schemas, their first-order details, and their corresponding dialectics), things need not be
complicated.

I take it, moreover, that any debate about a specific regress argument should follow one of these
two sets of guidelines. For this is how regress arguments work according to the two theories of
regress arguments by now on the market.

7.2. The Paradox Theory

2.1. Use

According to this theory, regress arguments can be used to demonstrate that a certain universally
quantified statement is false. So if you, NN2, want to cast doubt on a universally quantified
statement X of your opponent NN1, you may devise a regress argument. Here is a four-step
recipe.

Step Instruction

1 Make NN1’s X explicit as an instance of ‘For all Ks x, x is F only if such and such’.

2 Show that NN1 has to concede the corresponding instance of ‘For all Ks x, such and such
only if there is a new K-item y which is F’ as well.

3 Introduce a simple trigger situation (‘a is F’) and generate a regress from these three
ingredients.
4 Show that the regress entails something unacceptable, such that NN1’s X must be
rejected.

2.2. Evaluation

NN1 has in turn the following main options to resist your reasoning:

Option Instruction

1 Defend that X (as fully universally quantified) is not something she believed in
the first place.

2 Defend that the regress entails no unacceptable result such that nothing needs be
rejected.

3 Defend that the other universally quantified statement used in the derivation of the
regress, rather than X, is to be rejected.

4 Defend the sceptic option that the trigger rather than X is to be rejected (such that
‘no K is F’).

7.3. The Failure Theory

3.1. Use

According to this theory, regress arguments can be used to demonstrate that a certain solution
fails to solve a given problem (which may be existentially or universally quantified). So if you,
NN2, want to cast doubt on a solution X of your opponent NN1 (and perhaps turn this into an
argument in favour of your own solution to the given problem), you may devise a regress
argument. Here is a four-step recipe.

Step Instruction

1 Make the given problem explicit as an instance of ‘For all/at least one K(s) x, you have to
φ x’.

2 Make NN1’s X explicit as an instance of ‘For all Ks x, if you have to φ x, you ψ x’.

3 Show that NN1 has to concede the corresponding instance of ‘For all Ks x, if you ψ x,
then there is a new K y/ and you have to φ y first in order to φ x’ as well.

4 Draw the regress, and conclude that NN1’s X fails (whether it be Failure I or II-wise) and
that another solution needs to be found.

3.2. Evaluation
NN1 has in turn the following main options to resist your reasoning:

Option Instruction

1 Defend that the problem is not something to be solved in the first place.

2 Defend that an exception to X can be made (such the regress can be stopped at
some point).

3 Defend that the extra premise does not hold (such that no further problems are
generated).

4 Defend that it does not follow from the regress that X fails (e.g. because the
problems described in the regress are in fact identical).

7.4. Epilogue

Outside philosophy, I like it, in a certain sense, when people have nothing to say. Inside
philosophy, I like it when people, if they have something to say, make an effort to explain their
words. In this respect, I hope I have done better than some famous philosophers have in
addressing the topic of regresses:

It is obvious beyond all possibility of doubt that if the conditioned item is given, then a
regress in the series of all its conditions is set as a task. (Kant, Critique of Pure Reason,
Book 2, ch. 2, §7)

‘Knowledge’ is a referring back: in its essence a regressus in infinitum. That which

comes to a standstill (at a supposed causa prima, at something unconditioned, etc.) is
laziness, weariness. (Nietzsche, The Will to Power, §575)

The reasoning that leads to an infinite regress is to be given up not ‘because in this way
we can never reach the goal’, but because there is no goal; so it makes no sense to say
‘we can never reach it’. We readily think that we must run through a few steps of the
regress and then so to speak give it up in despair. Whereas its aimlessness (the lack of a
goal in the calculus) can be derived from the starting position. (Wittgenstein, Zettel,
§693)

I would like to finish this dissertation with one of my favourite nonphilosophical regresses: the
Sisyphus case.85 As Greek mythology tells us, Sisyphus had provoked the gods and so was
punished by them accordingly. He had to roll a huge boulder up a steep hill in the underworld.
Yet, the boulder was so heavy that it always rolled back as soon as he succeeded in making some
progress. Sisyphus fell prey to an infinite regress without ever being able to complete his task.

According to some philosophers, modern life is just like the situation in which Sisyphus finds
himself: day in, day out we work on the very same tasks (at the office, in the factories), yet in
fact are stuck in regresses and will never accomplish anything meaningful. Life is absurd, as
Camus concludes:

The workman of today works every day in his life at the same tasks, and this fate is no
less absurd. (1942: 117)

Yet this analysis seems incorrect to me. If life is absurd, then there are no tasks to be
accomplished (the gods, in that case, would not provide us with instances of line (1) of the
Failure Schemas). And if that were so, then neither could we get stuck in a regress problem as
Sisyphus did (for regress problems are generated partly on the basis of line (1)). Therefore: either
we are stuck in regresses, in which case we do it to ourselves (i.e. live by regressive instances of
(1) and (2)), or else we do it to ourselves, but are not stuck in regresses (i.e. live by non-
regressive instances of (1) and (2)). If nothing else, the new Failure Theory of regress arguments
teaches us this.

1 Throughout the dissertation, the square brackets ‘[‘, ‘]’ indicate how the line is obtained from
previous lines. Sometimes I also identify a rule of inference; in some other cases, I identify the
rule separately in the main text.

2 This is possible, as we shall see later, on the basis of the hypothetical inference rules Reductio
Ad Absurdum and Conditional Proof.

3 This all is commonplace. See e.g. Kitcher (1981: 516), who has it that the filling instructions
(i.e. directions for filling out the schematic letters) and classification (i.e. the description of
inferential characteristics) are independent of the schemas (which then are mere sets of
schematic lines). But all components are the same.

4 The fact that I use this grammatical distinction does not mean that I think properties/relations
are ontologically different from actions, or, for that matter, that I think they are identical.

5 The ‘charity’ label has been invented by Wilson (1959) and popularised by Quine and
Davidson in the debate on radical interpretation (cf. Davidson 1973). For the application of
Charity to argument reconstruction, cf. Rescher (1964: 162) and Feldman (1993: 115).

6 For a clear statement of these two goals, cf. Johnson (2000: 132, cf. 158).

7 For context and contemporary discussion, cf. Hurwicz (2008).

8 For an introduction into this area, cf. Baggini & Fosl (2003).

9 The boxes indicate the inference step from regress to conclusion, see §1.3.2 below.

10 Cf. e.g. Gendler (2000: 18-27) and Williamson (2007: ch. 6) for alternative proposals.
11 In addition, philosophers, more than scientists, seem concerned with universally quantified
claims of a certain sort (e.g. ‘can every proposition be justified by a further proposition?’, ‘what
if every event is explained by a further event?’, etc.).

12 For similar criticism to argument schemas for thought experiments, cf. Cappelen (2012: 197-
9).

13 As we shall see, though, regresses will play only a minimal logical role in regress arguments:
a few steps will suffice to draw conclusions from them.

14 Speaking literally, of course, schemas do not produce anything. What is meant is that
instances of the relevant lines have such and such features.

15 Throughout the dissertation, I try to be a clear as I can. Still, if certain instances of the schema
are ambiguous, then their meaning should be determined on the basis of the formalisation.

16 There are three main differences. First, in Black’s schema, lines (1) and (2) are integrated into
one line. Second, Black employs an extra assumption that specifies certain properties of the
relation ‘R’. Third, he seems to employ only hypotheses (rather than premises) given that his
concluding line runs ‘either (1), (2), (3) or (6) is false’. I shall motivate my choices in due course.

17 I do think that (3) can play a role in generating regresses (e.g. when the latter are generated
non-hypothetically; or when one does not make use of (7) below), yet in the construction I
propose, (3) does not play this role.

18 Another, more formal way to block loops would be to impose some properties on the relation.
See below.

19 The term ‘paradox’ has explicitly been used in connection with regress arguments by Black
(1996: 101) and Cling (2009).

20 It should be noted that this distinction, and the corresponding hypothetical rules ¬I and →I,
are not part of all formalisations of the propositional calculus (such as the one by Copi 1953).

21 In the table, ‘A’ and ‘B’ are placeholders for formulas, the subscript ‘HYP’ indicates that the
formula is a hypothesis (rather than a premise), and ‘[a/x]A’ refers to the formula which results
when all free occurrences of the variable ‘x’ in A are replaced with occurrences of the constant
‘a’.

22 In these first-order proofs, all numbers on the right side of the page indicate how a line is
obtained from previous lines (I shall drop the square brackets here).

23 Or irreflexive and transitive. Cf. Sanford (1975: 534-5, 1984: 109-13), MacKay (1980: 374-
5), Black (1987, 1988), Post (1993), Gratton (2010: ch. 2), and Aikin (2011: 23-32).
24 (f) ensures that all Ks stand in <; (g) that only Ks stand in <; (g) that only Ks stand in <; and
(h) and (i) that < is transitive and irreflexive.

25 All credits for (c)-(e) go to Bert Leuridan, and all credits for the second solution (f)-(i) to
Christian Straßer.

26 Cling (2009) proposes a schema that is very much in the spirit of the Paradox Schema. Rather
than using (6) (i.e. ‘¬IR’), he uses a premise of the form ‘IR→¬∃x(Kx∧Fx)’. This also leads to a
contradiction, namely with (3).

27 For the motivation of these desiderata, see again §2.1.

28 For many more examples of paradoxes, cf. Rescher (2001) and Clark (2002).

29 Given that we do not use a step to ‘there is an infinity of short possible persons’, we do not
need a trigger in this case.

30 The same disclaimer applies: if certain instances of the schemas are ambiguous, then their
meaning should be determined on the basis of the formalisations. Also, for the latest statement of
these schemas, cf. Wieland (2013d).

31 It is also possible to frame problems in a different way, namely as arguments with

counterintuitive conclusions, and solutions, then, as counterarguments against such arguments.
Cf. §6.1.3 below.

32 For this point in a different context, cf. Van Cleve (2003: 50, n. 12). The fact that time plays
no essential role, i.e. in regress arguments generally, does not mean of course that it cannot play
any role in some of them (cf. some of the everyday cases in §1.2).

33 I also use this notion in my formalisation below.

34 For this format, cf. Wieland (2011d).

35 Strictly speaking, the second conjunct is superfluous given (1). Yet, it simplifies the
derivation (and puts no extra demands on the premise set).

36 This partly explains, as we shall see in §4.1, why Failure arguments are not Paradox
arguments.

37 Still, it seems worth exploring whether a different Failure Theory can be constructed in
deontic logic.

38 Next to this, perhaps it is not entirely implausible to suppose that one never solves all
problems of a given kind if there is a regress that loops between two (or any other finite number
of) tasks that have to be accomplished.
39 To be sure, the second citation can also be read in terms of the Paradox Schema, i.e. that there
is at least one item to which the regress formula does not apply. See §2.2.3 above.

40 Note that in §1.3.1 I already provided four actual cases that express the Failure Intuition.

41 Cf. Lammenranta (2008). This dialectical conception is not uncontroversial, see §5.4.

42 Of course, if my choices in cases like these are incorrect, then the cases must be restated in
either version I or II.

43 Of course: in each case, if my choices are incorrect, then one may restate the arguments in
their practical or theoretical format.

44 I introduced this case in §3.1 above.

45 For a brief version of this argument, see §2.3.4 above.

46 These were identified in §2.3.2 above.

47 Cf. my formalisations in §2.3, where the Failure conclusion cannot be fully captured in terms
of the predicates ‘T’ and ‘K’.

48 For one thing, it is controversial to say that (R) is constitutive for our obligations. For this
would imply that all our obligations are, so to speak, wholly up to us.

49 Cf. Brennan (2003) for an overview of the literature on this question.

50 These principles are: ‘S is free to perform an action x only if (i) S has a desire to refrain from
performing x, or (ii) S is free to generate a desire to refrain from performing x’ and ‘S rarely has
a desire to refrain from performing x’ (i.e. S rarely fulfils the first disjunct) (cf. Van Inwagen
1994, Fischer & Ravizza 1996).

51 Note that, as they stand, (Vic-P) and (Vic-F) are incompatible. For (Vic-F) implies that a
regress is virtuous so long as it does not entail that the given solution fails (even if it does entail
other unacceptable results, whatever they may be). There are two solutions to this query. First,
the criteria could be turned into sufficient conditions. Second, the criteria could be relativized to
Paradox and Failure regresses respectively (i.e. ‘for any Paradox regress R, R is vicious iff etc.’).
The latter strategy seems preferable, for by the former’s sufficient conditions it is no longer
possible to determine whether regresses are virtuous.

52 Even if T1 wins out relative to T2, this does not automatically form an argument in favour T1.
For there are alternative conclusions that one might draw: that yet another solution next to T1
and T2 must be found, or that the problem must be left unsolved for the time being.
53 Zeno’s Dichotomy Paradox should not be confused with his Achilles and the Tortoise
Paradox. They are closely related, but in the latter case, all distances to be traversed are distinct
from each other. See below.

54 Rescher (2010: 24-5, 53-4) makes a comment into this direction by distinguishing between
actions and accomplishments: between something that an agent does and something that results
from what she does (i.e. for free).

55 Moreover, if infinities are assumed to be absurd from the start (a minority position), then all
Paradox regresses entail unacceptable results so long as they are infinite. For the notion of
infinity generally, see e.g. Russell (1914), Priest (1995), and Oppy (2006).

56 See §2.3 for the formal details.

57 This argument can be strengthened by the fact that my formalisations in §2.3 operate without
temporal notions.

58 One pertinent issue will be: If Failure regresses are infinite, then there is a solution for all
problems, and in that case one may wonder why the solution fails.

59 Jacquette & Johnstone (1989) identify a similar duality, as they call it, between self-
application paradoxes and infinite regresses.

60 For an overview of the debate, cf. Marušić (2011).

61 No analogous problem arises for the Failure II construction of Carroll’s initial case provided
in §3.4.2, where its line (3) does seem plausible if (2) is in place. For further discussion and
context of this new Tortoise problem, cf. Wieland (2013c).

62 I have simplified it somewhat: ‘IV’ derives from ‘in virtue of’ which is to mean the same as
‘depending on’.

63 Cf. the analogy of royal people in Brzozowski (2008: 200).

64 But cf. §4.2.2 above.

65 For the moment I am switching to doxastic justification in order to render (1*) credible. It is
plausible to suppose that all our beliefs need a justification, though not to suppose that all
propositions whatever need a justification. But, of course, this is not the difference I am stressing
here, which is the difference between (3) and (3*).

66 For further considerations, cf. Cling (2004, 2008, 2009).

67 At one point, Sorensen (1995: 255) considers the stronger, alternative obligation that we
ought to learn our obligations. Here, I will stick to Sider’s weaker suggestion, as that will do to
discuss the regress worry.
68 On (Vic-P), see §4.2.2.

69 Nevertheless, premise (3) deserves further discussion.

70 For a historical reconstruction, cf. Griffin (1991: ch. 8).

71 Of course, Russell’s view is that no solution whatsoever can solve the problem, and hence
that at least some relations are irreducible entities. To establish this, however, requires
argumentation that extends beyond the current regress argument, cf. his (1903: §§212-6).

72 Not to be confused with the Indeterminacy of Translation thesis, and other similar claims.

73 Further, Rescher (2010: ch. 2) is a Paradox Monist as he follows Clark, and Post (1993) and
Roy (2010) because they follow Black.

74 Cf. the cases that are non-vicious by (Vic-F) in §4.2.2.

75 One might think I am reading too much into Russell’s suggestions here, but at least this is one
way of making things explicit. For other comments, cf. Taylor (1934: 47-9), Maurin (2002: 100-
2), Rescher (2010: 153-7).

76 This is a non-causal explanation, as the colour of one’s and one’s parents’ eyes are both
effects of a common cause: the parents’ genotypes.

77 I take it that Rankin can be defended in a similar way. Also, it is worth stressing that I myself
would not speak of ‘circularity’ in these contexts, see §4.2.4.

78 Geach seems to refer to a cryptic passage by Wittgenstein that I will cite in §7.4.

79 And if they do carry over, then my view is that Failure inferences are not logically invalid, but
that the suppressed premises (6) fail to hold: see §4.2.2.

80 I have simplified Gaskin’s case somewhat. First, Gaskin speaks of ‘a is said to be F’ rather
than ‘a is F’. Second, on Gaskin’s account, each step is not only necessary for the previous one,
but also sufficient, as well as vice versa.

81 Nevertheless, if ‘A is necessary for B’ involves ‘B depends on A’ (as suggested in §4.2.1

above), and if ‘B depends on A’ involves ‘A explains B’ (as e.g. Correia (2005: ch. 3) and
Schnieder (2006: 402-12) argue), then necessary conditions automatically involve explanatory
conditions.

82 This means that both schemas fulfil the first six desiderata from §2.1.1.

83 Cf. Walton (1996: 216), Walton & Reed (2005: 341-2), and Paglieri (2007: 2- 6).
84 To be sure, I have not indicated what exactly this list of general criteria will consist of. It may
even be a slightly different list for different kinds of arguments. Still, I have offered a few some
suggestions in the previous section: good arguments should have plausible premises and
hypotheses, and interesting conclusions that are hard to resist.

85 A suggestion that I shall not explore here is that, in fact, all punishments of the Greek gods
concern regresses. Cf. Tityus, Tantalus, and indeed Sisyphus from Homer, Odyssey, 11.576-600.