Professional Documents
Culture Documents
the Liar
Reflections on the Liar
E D I T E D B Y B R A D L E Y A R M O U R- G A R B
1
1
Oxford University Press is a department of the University of Oxford. It furthers
the University’s objective of excellence in research, scholarship, and education
by publishing worldwide. Oxford is a registered trade mark of Oxford University
Press in the UK and certain other countries.
1 3 5 7 9 8 6 4 2
Printed by Sheridan Books, Inc., United States of America
CONTENTS
Acknowledgments vii
Contributors ix
1. Introduction 1
B r a d l e y A r m o u r -G a r b
v
vi Contents
11. Truth and Transcendence: Turning the Tables on the Liar Paradox 281
Gila Sher
Index 375
ACKNOWLEDGMENTS
About five years ago, I noticed that there were a number of people who do great
work in philosophy yet have said very little, if anything, about the liar paradox
or the projects involving it. The purpose of this volume is to afford those phi-
losophers the opportunity to give what might be described as reflections on the
liar paradox. I want to take this opportunity to thank them for their excellent
contributions.
Many thanks to Peter Ohlin of Oxford University Press both for his patience
and for his judicious oversight of the project. I would like to thank my wife, Allison
Armour-Garb, and our kids, Isabel and Zev, for their love and for their enthusi-
asm. Finally, thanks to my parents, Andrew and Sheila Garb, for their uncondi-
tional support.
vii
CO N T R I B U TO R S
ix
1
Introduction
Bradley Armour-G arb
Consider:
(1) ‘The first full sentence on this page is not true’ is true if, and only if,
the first full sentence on this page is not true.1
But
(2) The first full sentence on this page is ‘The first full sentence on this
page is not true’.
(3) The first full sentence on this page is true if, and only if, the first full
sentence on this page is not true.
(4) The first full sentence on this page is true and the first full sentence
on this page is not true,
1
2 Reflections on the Liar
would contend, since our sentence involves a definite description, it is really about
our domain of discourse and, thus, does not really say of itself that it is not true.
Indeed, it is a contingent fact that our sentence ends up denoting itself at all, as it
is that it yields paradoxicality.
If we want a liar sentence that really refers to itself, and that really says of itself
that it is not true, we might start with a term, λ, such that
λ = ‘~T(λ)’ 2
Now, since λ = ‘~T(λ)’, we can generate paradox, given the following assumptions:
In this argument, we rely on Leibniz’s Law in such a way that we can allow for
the substitution of identicals, together with the laws of classical logic and the
truth rules. We also assume that the sentence for which λ is a name is meaningful,
because, I would contend, it is only meaningful sentences that admit of instances
of our logical principles.
Int roduc tion 3
liar paradox, the next question to ask regards the relevance of that paradox—
more specifically, its philosophical relevance.
The relevance of a paradox—whether it is the liar paradox or some other—
does not concern or involve its apparently resultant inconsistency (or triviality),
though its apparent inconsistency (or triviality) is essential to its paradoxicality.
Rather, the relevance of a paradox involves those underlying features that result,
or in any case appear to result, in inconsistency.7 Paradoxes challenge some of our
most deeply held convictions. Studying the paradoxes is important because they
indicate that something that we thought was correct—something that we were
relying on, or were assuming, which we accepted, either implicitly or explicitly—is
not correct. We study the paradoxes because they reveal certain assumptions that
we have made, which must be questioned, and they force us to rethink some of the
most central commitments that we possess.
The liar paradox strongly suggests that there is something wrong with our
ordinary understanding of certain semantic features of our language or of cer-
tain logical features of our reasoning—that is, with the features that give rise
to the paradox. A central intuition about truth is that, for any meaningful sen-
tence, S, the inference from S to another sentence that expresses that S is true
is warranted, and vice versa. The liar paradox shows that, for a classical language
that contains its own truth predicate, without a revision to our pretheoretic (or
“naïve”) semantic and logical principles, this intuition cannot be preserved.
So far, we have at least the beginnings of an idea as to what problem, or prob-
lems, the liar paradox presents. And we have the beginnings of an answer to the
question of why we do, or, perhaps, should study that paradox. But what does a
study of the liar paradox involve?
because semantic competence with ‘true’ involves a disposition to accept all of its
instances.10
Treatment projects involve saying what we should do in light of the paradox—
how we should treat that paradox. In the case of the liar paradox, they involve
explaining what changes, if any, we would make to our standard, or evident, logi-
cal or semantic assumptions or principles in order to prevent our problematic
conclusion from arising.
There are two different types of treatments to the paradox that have emerged
in the literature, a logical treatment and a semantic treatment. The former sort of
treatment generally requires that the overall logic for a given language will be non-
classical. In many cases, such an approach holds fixed certain semantic principles
while denying that the logic for our language is classical.11 By contrast, a semantic
treatment demands a closer look at the language in which a “liar sentence” can
be formulated, together with an examination of its semantic properties. In many
cases, such an approach holds fixed a given logic while providing a more refined
picture of the formal properties of the truth predicate, or, perhaps more generally,
of providing a more refined picture of semantic values—of what they are, and (or)
of how they behave.12
Although everyone who is engaged in a treatment project for the liar paradox
recognizes these two kinds of treatments, there is another distinction between
types of treatments that seldom gets noted, what we might call hermeneutic ver-
sus revolutionary treatments of that paradox.13
The reasoning in the derivation of the liar paradox shows that not all of our
standard assumptions can be honored, given the presence of a liar sentence. Does
that show that (e.g.) our concept of truth, or the language that contains the truth
predicate, is inconsistent (or, what is perhaps worse, incoherent)? Or does it show
that, while the concept (or the language) is consistent (and coherent), the “liar
reasoning” involves some highly plausible but nevertheless incorrect assump-
tions? Depending on how those questions are answered, anyone involved in a
project of treating the liar paradox will provide either a hermeneutic or a revolu-
tionary treatment.
Treatment projects that are hermeneutic involve explaining how our language
or our logic is such that the paradoxicality that we have considered does not actu-
ally arise—why the paradox is merely apparent, what features of the language or
the logic reveal the flaw in our reasoning. A hermeneutic treatment might involve
providing a semantic theory for our language that shows that certain steps in the
liar reasoning cannot, and, so, should not, arise. Alternatively, it might involve
showing that our logic—the logic, the one that we do, or should, actually employ—
invalidates the reasoning that we (perhaps mistakenly) deployed when we set out
the liar paradox. What is central to a hermeneutic treatment of the liar paradox is
that the appearance of paradoxicality is just that—a mere appearance.
By contrast, a treatment project that is revolutionary involves accepting
the contradictory consequences of the paradox as genuine consequences, and
6 Reflections on the Liar
explaining how we can avoid these results in the future. Thus, for example, some-
one who is involved in a revolutionary treatment of the liar paradox might accept
that what the paradox makes evident is that truth is an inconsistent concept. As
part of her proposed treatment, she might reject one or more of the central prin-
ciples governing the truth predicate and might propose an alternative principle
that enables the truth predicate to do most of what it needs to do without fall-
ing victim to paradox. Alternatively, she might present and propose a logic—not
our actual logic, but one that we should, or might, endorse and employ instead—
which would invalidate the reasoning that we deployed when we set out the liar
paradox. What is central to a revolutionary treatment of the liar paradox is that
the paradox’s resultant inconsistency is a genuine consequence, though it need
not arise (and, presumably, will not remain if or when the proposed changes are
adopted).
Descriptive projects involve identifying our use of the truth predicate in a nat-
ural language like English and providing the meaning of that predicate (together
with the meanings of sentences that contain it).14 More generally, one involved in
a descriptive project will provide the meaning of ‘true’-involving sentences and, if
possible, will provide a means by which we can distinguish the trouble-free ‘true’-
involving sentences from the paradox-inducing ones.15
A descriptive project might reveal that our concept of truth is inconsistent,
which might lead us to a treatment project that is revolutionary, explaining what
changes to our natural language, our conceptual scheme, or our logic are called for
in order to prevent the paradox from arising. Alternatively, a descriptive project
might reveal something new, or at least not obvious at first glance, about our
concept of truth that goes beyond our standard, ordinary (or “folk”) conception
of that notion, something that makes evident that the paradox should not have
arisen in the first place. A treatment project of the liar paradox that is hermeneu-
tic might then explain why and how, given this revelation about our concept of
truth, the paradox is merely apparent.
While those engaged in treatment projects for the liar paradox are not usu-
ally, if ever, clear on whether their proposed treatment of that paradox is (what
I have been calling) revolutionary or hermeneutic, there is an important difference
between engaging in a treatment project that is hermeneutic versus engaging in
one that is revolutionary. The issue turns on ad hockery.
Although, in the case of the Liar, the paradox might require us to abandon
what seemed to be some intuitively plausible, initially attractive principles, one
involved in a treatment project that is hermeneutic cannot reject or revise such
principles for the sole reason that they were implicated in the paradox, for that
would be blatantly ad hoc and, therefore, would be unacceptable.16 For this rea-
son, when a theorist is engaged in a hermeneutic treatment of the liar paradox
and is providing her treatment of that paradox, the theorist will usually go on to
provide independent or principled reasons for rejecting, revising, or replacing the
relevant principles.
Int roduc tion 7
untrue. Or it may be that her account was wrong and liar sentences are not actu-
ally untrue. In either case, we would be compelled to conclude that her account is
incorrect. So, a condition on the correctness of a hermeneutic treatment is that it
be, in the relevant sense, expressively adequate.
The requirement of expressive adequacy is also related to the final sort of
projects, characterization projects. These projects involve semantically charac-
terizing the myriad liar sentences that can be generated, where to “semantically
characterize” a sentence is to assign it a semantic status using vocabulary that is
available.18
Pursuing some version of a characterization project is usually required for one
involved in a treatment project. That is, absent a self-proclaimed commitment to
quietism with respect to the status of liar sentences, which would involve a refusal
to assign a semantic status to any liar sentence, if one has provided a treatment
of the liar paradox for which that paradox does not arise, it is standardly held that
she will still need to be able to semantically characterize liar sentences. But why
must such a theorist semantically characterize liar sentences?
I think that the reason turns on a general acceptance of an aspect of semantic
closure, viz., (1) that all of the sentences of a language possess semantic statuses
and (2) that each of these semantic statuses can be expressed in a language. In the
present context, this means that liar sentences possess (or should possess) some
sort of semantic status and that this status can be expressed in a language. In par-
ticular if we maintain that liar sentences possess (or should possess) a semantic
status, then we should be able to express that they possess (or should possess)
that status.19
For example, if one involved in a treatment project maintained that liar sen-
tences are (or should be taken to be) indeterminate with respect to truth values,
then since indeterminacy is a semantic status it seems that she should be able to
characterize such sentences, viz., as indeterminate with respect to truth values.
If she cannot (consistently) so characterize such sentences, then it seems that
there is a way that the sentences are (or, perhaps, a way that they should be),
by the lights of her own account, that cannot be expressed, and this would be a
problem for the theorist. In general, then, if one engaged in a treatment project
cannot (consistently) express the statuses that she attributes to liar sentences,
then this is a problem for her proposed treatment. Thus, it seems that success
with respect to a characterization project is necessary for the acceptability of a
treatment project.
One of the main challenges for those who take on characterization projects
is the problem of revenge. In a standard case of revenge, a theorist has proposed
a semantic notion as a means for classifying liar sentences, and maintains that
we can use this notion to characterize liar sentences. In so doing, the theorist
commits herself to the coherence of that notion. Revenge then involves show-
ing that the semantic notion is not consistently expressible—that we cannot
use that notion to semantically characterize all cases of liar sentences without
Int roduc tion 9
(We could add more to the description to make it genuinely uniquely identifying,
but I will instead let us assume the rest as read.) Given classical assumptions,
we can show that the sentence marked with an asterisk cannot be true but also
cannot be false (or not true). If we wish to characterize (*) semantically, we seem
forced to postulate a third category exclusive of truth and falsehood, which, for
the sake of argument, we can just entitle ‘meaningless’. Using ‘meaningless’, we
can now characterize (*) (and, indeed, our original liar sentence, ‘λ’) as meaning-
less, and apparently avoid any impending contradiction.
Unfortunately, however, this strategy seems to go nowhere, since we can now
formulate a new liar sentence:
In order to get to get to the problem that (**) presents, we must back up a bit.
It is standard in the literature to endorse what is known as ‘the meaning-to-
truth conditional’, the claim/principle that if a given sentence is meaningful then
the truth schema applies to it, viz.,
Now suppose, as before, that a given theorist who endorses both (MT) and (P) has
declared that (*) is meaningless. So, she maintains that we can characterize liar
sentences as meaningless. She is then presented with (**). Given just (**), together
with (MT), (P), and a modicum of classical logic, we can show that her theory will
be inconsistent. To see this, rather than using (**), for convenience, let σ = ‘σ is
meaningless or is not true’. We can now reason as follows.
Suppose that σ is meaningful. Then, by (MT), σ is true if, and only if, σ is either
meaningless or not true, viz., (T(σ) ↔ (~M(σ) v ~T(σ)). But this can only be the
case if σ is meaningless. Thus, from the assumption that σ is meaningful, it fol-
lows that σ is meaningless. Contradiction. Therefore, σ is meaningless. But if σ
is meaningless, then σ is either meaningless or is not true. So, given that σ is
meaningless, it follows that σ is either meaningless or is not true. But σ is the
sentence ‘σ is either meaningless or is not true’. So, we have what amounts to
a proof the last line of which is σ and, so, by (P), it follows that σ is meaningful.
Contradiction. We have thus shown that our theorist’s theory, containing, as it
does, (MT), (P) and classical logic, is inconsistent.21
Now either (**) (or σ) is in the object language to which our theorist’s theory
applies or it is not. If it is (presumably because ‘meaningless’ is), then what the
above derivation makes evident is that a version of the liar paradox remains in
place for the language in question, and so a new treatment will have to be found.
If, on the other hand, (**) (or σ) is not in the language our theorist’s theory con-
cerns, then while she may have dealt with the language in question to her satis-
faction, her theory will simply be inadequate as a treatment of the liar paradox in
general. And the reason is obvious: (**) is perfectly good English, so a theory that
cannot handle a language that includes (**) cannot explain how it is that English
manages to contain (*) or (**) without being logically trivial.
Moreover, appealing to a more technical-seeming term than ‘meaningless’
(e.g., ‘ungrounded’, or ‘does not settle down to a stable value as you iterate the
jump function into the transfinite’) will not help, since it is a widely accepted view
that newly minted technical terms, in works the nontechnical fragments of whose
languages are English, themselves become part of English. (Indeed, many works
that break new ground in theoretical physics are written in English, for instance.)
This is why so few are convinced by the common dodge of insisting, in the face
of this problem, that a certain semantic theory that is supposed to treat the liar
paradox is not itself expressed in English but is expressible in an extension of
English, English*, which amounts to a different language—convinces, at the very
most, only those who resort to it.
What revenge problems appear to make evident, then, is that we seem forced
to choose between either expressive inadequacy or inconsistency in our own the-
ories or accounts. That is, they show either that we cannot handle expressions
that seem as obviously part of natural language as the expressions that we can
handle—e.g., ‘is false or meaningless’ in relation to ‘is false’—or that by introduc-
ing such expressions we become faced with the very sort of inconsistency that
Int roduc tion 11
we were trying to understand and to avoid. These are the problems that revenge
appears to present.
Having set out some of the important aspects of the projects associated with
the liar paradox and having clarified, at least to some extent, what a paradox
involves, I now shift gears to describe the point of the present volume, after which
I will briefly summarize the contributed chapters.
implication, he recasts his initial attempt at capturing the meaning of ‘true’ and
shows that this is not plagued by the consequences of the liar paradox.
(D) For each Human Language H, some Tarski-style theory of truth for
H is the core of a correct theory of meaning for H,
predicate and argues that to treat the Liar, pluralists should reject infinitary dis-
junction. Wright argues that pluralists should reject the version of pluralism that
Cotnoir assumes and goes on to offer an alternative treatment to the Liar that is
a variant of so-called meaningless strategies—one that shows liar sentences to
be undecidable. The result is a more natural approach to truth predication in real
languages, which, Wright contends, should be a desideratum on any successful
pluralist conception.
Notes
1. In informal terms, the truth schema is
(TS) S is true if, and only if, p,
where p is a sentence and S is a name or is a description of p.
2. It is also worth noting that this is in contrast to some presentations of liar sentences, where,
say, we let ‘L’ be a sentence such that
L ↔~T (‘L’)
This “liar sentence” does not say of itself that it is not true (and the right-hand side in the
biconditional does not refer to itself). What the biconditional provides is a sentence that
is equivalent to a sentence that says of it—of that very sentence—that it is not true. For a
discussion of this, see Heck (2013).
3. Although I shall not be dealing with a propositional version of the liar paradox, it bears
noting that there are such versions and that for those versions one of the assumptions that
generates a paradox is that liar sentences express propositions.
4. If the truth rules are not taken to capture a substitution rule, then we will need a principle of
transparency to the effect that in all extensional contexts, any sentence S may be replaceable,
salva veritate, with ‘S is true’ and vice versa. For more on this rule, see Field (2008).
5. I say “unacceptable,” rather than, say, “inconsistent,” so as to allow dialetheists, who accept
some contradictions, to accept this understanding of what is involved in a paradox. This
enables a dialetheist to grant that if we stick with classical logic, the liar paradox is paradoxi-
cal but that if we were to move to a paraconsistent logic, we would effectively rid the Liar of
its paradoxical element.
6. Triviality, or trivialism, is the view that every sentence is true. This is something that even
dialetheists, who accept certain contradictions, will not tolerate.
7. Although not all paradoxes involve apparent inconsistency (think of the “paradoxes” that
involve counterintuitive aspects of continuous space and time), for our purposes, we can
restrict our attention to those paradoxes that do seem to involve inconsistency.
8. Although my description of the projects involved in studying the liar paradox differ from
those presented by Kevin Scharp (2013), I take the term ‘project’ from his work on that
paradox. For a discussion of diagnostic and treatment projects, see Chihara (1979). For a
discussion of descriptive projects, see Gupta (1982).
9. Charles Chihara (1979, p. 590) explains diagnostic projects as follows:
Alfred Tarski once remarked: “The appearance of an antinomy is for me a symptom of
disease.” But what disease? That is the diagnostic problem. We have an argument that
begins with premises that appear to be clearly true, that proceeds according to infer-
ence rules that appear to be valid, but that ends in a contradiction. Evidently, something
appears to be the case that isn’t. The problem of pinpointing that which is deceiving us
and, if possible, explaining how and why the deception was produced is what I wish to
call ‘the diagnostic problem of the paradox’.
10. For the latter view, see Eklund (2002, p. 260).
20 Reflections on the Liar
11. For a recent logical treatment of the liar paradox, see Field (2008).
12. For a paradigmatic semantic treatment of the liar paradox, see Tarski (1944).
13. Although the terms ‘revolutionary’ and ‘hermeneutic’ are usually directed at various fictional
accounts of some putatively, or apparently, problematic discourse, it is useful to employ
those terms as a way of dividing two distinct aspects of treatment projects for the liar para-
dox. To get a sense for why I co-opt those terms to describe two aspects of treatment proj-
ects, consider the standard understanding of ‘revolutionary fictionalism’ and ‘hermeneutic
fictionalism’. Revolutionary fictionalism is a prescriptive enterprise, telling us how some
putatively problematic discourse should be understood and how it will be understood come
the revolution, when people come to realize that the problematic discourse really is prob-
lematic. Accordingly, while revolutionary fictionalists maintain that some discourse really is
problematic, they also claim that it does not have to be problematic, and maintain further
that it can be made unproblematic, provided its aims are changed. By contrast, hermeneutic
fictionalism is a descriptive enterprise telling us how some discourse actually is to be under-
stood. Hermeneutic fictionalists take apparently problematic discourse to be unproblematic
on grounds that it never had a problematic aim.
Note that John Burgess (1983) distinguishes between hermeneutic and revolutionary
nominalism and that this distinction is later put to use by Burgess and Rosen (1997).
Jason Stanley (2001, p. 36) introduced the terminology of revolutionary versus hermeneutic
fictionalism.
14. As Gupta (1982, p. 2) notes, “It is only when we know the role of truth in our conceptual
scheme that we can evaluate the damage that the paradox does, and the reforms (if any) that
it necessitates.”
15. Treatment projects and descriptive projects are connected, for one might think that it is
only when we know the role of the concept of truth in our conceptual scheme that we can
properly evaluate the damage that the paradox yields and the reform, if any, that it neces-
sitates. For more on this, see Gupta (1982, p. 2).
16. I take it that the reason that ad hockery would be unacceptable for a proposed treatment
that is hermeneutic is that if the proposed treatment is genuinely ad hoc, then we have no
reason for thinking that what is proposed is descriptively correct and, hence, no reason for
endorsing the proposal.
17. I include the parenthetical as a way of accommodating a dialetheic response to the apparent
pathology that the liar paradox appears to present. For the dialetheist, the pathology is logi-
cal as opposed to semantic. For a discussion of this, see Armour-Garb (2004).
18. Nota bene, by saying that liar sentences have (or should have) some semantic status, I am
not saying that they must have some truth value. In general, a sentence may have a semantic
status even if it does not possess a truth value. For example, if someone were to hold that
it is indeterminate as to whether liar sentences possess truth values, we should still see her
as taking liar sentences to have a semantic status, viz., that of being indeterminate with
respect to truth values.
19. The reason for the parenthetical in the last sentence is to accommodate treatment projects
that are revolutionary.
20. Please read the conditional as material and the biconditional as the material biconditional.
21. After writing this, I read in Bacon (2015) a related version of the problem that I am pre-
senting. My presentation has benefited as a result. It bears noting that this is not a knock-
down argument against the meaningless strategy. For a response, see Armour-Garb and
Woodbridge (2012, 2015).
References
Armour-Garb, B. 2004. “Diagnosing Dialetheism.” In Priest, G., Beall, J. C., Armour-Garb, B.,
eds., The Law of Non-contradiction: New Philosophical Essays, 113– 25. Oxford: Oxford
University Press.
Int roduc tion 21
Armour-Garb, B., and J. Woodbridge. 2013. “Semantic Defectiveness and the Liar.” Philosophical
Studies 164: 845–863.
———. 2015. Pretense and Pathology: Philosophical Fictionalism and Its Applications,
Cambridge: Cambridge University Press.
Bacon, A. 2015. “Can the Classical Logician Avoid the Revenge Paradoxes?” Philosophical Review
124 (3): 299–352.
Burgess, J. 1983. “Why I Am Not a Nominalist.” Notre Dame Journal of Formal Logic 24: 93–105.
—— —. 2004. “Quine, Analyticity and Philosophy of Mathematics.” Philosophical Quarterly
54: 38–55.
Burgess, J., and G. Rosen. 1997. A Subject with No Object. Oxford: Clarendon Press.
Chihara, C. 1979. “The Semantic Paradoxes: A Diagnostic Investigation.” Philosophical Review
88: 590–618.
Cotnoir, A. 2013. “Pluralism and Paradox.” In Nikolaj Pedersen and Cory Wright, eds., Truth and
Pluralism: Current Debates. 339–350. New York: Oxford University Press.
Eklund, M. 2002. “Inconsistent Languages.” Philosophy and Phenomenological Research
64: 251–275.
Feferman, Solomon. 1984. “Toward Useful Type-Free Theories, I.” Journal of Symbolic Logic
49: 75–111.
Field, H. 2008. Saving Truth from Paradox. Oxford: Oxford University Press.
Gupta, A. 1982. “Truth and Paradox.” Journal of Philosophical Logic 11 (1): 1–60.
Heck, R. 2013. “More on ‘A Liar Paradox’.” Thought 1 (4): 270–280.
Herzberger, Hans G. 1982. “Naive Semantics and the Liar Paradox.” Journal of Philosophy
79: 479–497.
Lycan, W. 2010. “What, Exactly, Is a Paradox?” Analysis 70 (4): 615–622.
Priest, G. 2006. In Contradiction. Oxford: Oxford University Press.
Quine, W. V. 1966. The Ways of Paradox and Other Essays. Cambridge, MA: Harvard University Press.
Sainsbury, R. M. 1995. Paradoxes. Cambridge: Cambridge University Press.
Scharp, K. 2013. Replacing Truth. Oxford: Oxford University Press.
Schiffer, S. 2003. The Things We Mean. Oxford: Oxford University Press.
Sorensen, R. 2003. Vagueness and Contradiction. Oxford: Oxford University Press.
Stanley, J. 2001. “Hermeneutic Fictionalism.” In French, P. and Wettstein, H., eds., Figurative
Language, 36–71. Oxford: Blackwell.
Tarski, A. 1944. “The Semantic Concept of Truth and the Foundations of Semantics.” Philosophy
and Phenomenological Research 4: 341–376.
Unger, P. 1979. “Why There Are No People.” Midwest Studies in Philosophy IV: 177–222.
———. 1980. ‘Skepticism and Nihilism’ Noûs 14(4), Special Issue on Epistemology (November,
1980): 517–545.
Williamson, T. 1994. Vagueness. London: Routledge.
2
2.1. Introduction
Some years ago, one of us wrote a paper entitled “Why There Are No People”
(Unger 1979a), which dealt with vagueness and the sorites and aimed to establish
a somewhat nihilistic thesis regarding the proper stance to take toward a com-
mon, and seemingly coherent, expression like ‘person’, to the effect that it has an
empty extension. More specifically, he argued that because ‘person’ is an incon-
sistent expression, there are no people. In other work (1979b, 1980), this same
author went on to contend that he does not exist and that there are no swizzle
sticks (which may be true now, in 2016, but certainly seemed to be contrary to
fact when the afore-noted papers were published!), on grounds that a supposi-
tion of the contrary of those claims, when combined with seemingly plausible
assumptions, leads to contradictions. We shall briefly review the form of these
arguments below.
At the end of his paper (Unger 1979a, p. 219), to the effect that there are
no people, this author briefly, and compactly, argued that perhaps certain
other expressions, which do not seem to be soritical, at least in the way that
‘person’ appeared to be, may likewise be inconsistent. Here is the heart of what
he said:
22
From No Peopl e to No Lang uag e s 23
After very quickly drawing out a consequence of supposing that there is such an
expression as ‘expression that does not apply to itself ’, he notes (p. 219) that
“comparative matters merit further examination.”1 We in this chapter set out
to examine further. In particular, after briefly reviewing his general response
to the evident appearance of soritical expressions, we show that the same sort
of reasoning that Unger employed, in order to arrive at the conclusion that
there are no people, can be deployed as a response to Grelling’s paradox and to
the analogue of Grelling’s paradox hinted at above and by Unger (1979a). We
then go on to show that this response to the analogue of Grelling’s paradox
can be used as a response to the liar paradox, where by a “response” we mean a
way of explaining how that paradox does not arise. In particular, we show that
if we grant, for the sake of argument, what led Unger (1979a, 1979b, 1980) to
draw the (nihilistic) conclusions, regarding ‘person’, etc. that he drew, then
we can draw similar (nihilistic) conclusions regarding our semantic expres-
sions, which will thereby provide a response to the family of paradoxes whose
members have been traditionally dubbed the “semantic paradoxes,” just as his
(1979a, 1979b, 1989) conclusions provided a response to the sorites para-
dox. (So, we might conclude from this that the problem of vagueness, while
it is one of the sources of inconsistency, is not the only source of apparent
inconsistency, and that the form of a response to the sorites can be extended
and applied as a response to the semantic paradoxes.) Thus, as we will show,
if Unger’s (1979a, 1979b, 1989) response to the sorites is accepted, then we
should accept our resultant response to the semantic paradoxes. After estab-
lishing this conditional claim, we will conclude by establishing that the same
sort of reasoning that Unger (1979a, 1979b, 1989) employed can be used to
provide a response to a novel sort of liar paradox—one that involves thought,
rather than language.
apply to, or are true of, some things. As a result of this, Unger (1979a, 1979b,
1980) contends that there is nothing to which they apply.
This response might seem extreme, but it occupies a place in logical space.
Russell (1918, p. 101) discusses the barber paradox, which takes off from the
supposition of a barber who shaves all and only those who do not shave them-
selves, in the service of discussing his own paradox, which has come to be known
as “Russell’s paradox.” He there (1918, p. 101) makes the point that although
there are valid forms of Russell’s paradox, the question that drives the barber
paradox—Does the barber shave himself?—is not a valid form. His point, pre-
sumably, is that the problem with the question is that it presupposes precisely
what anyone approaching the paradox will want to deny, viz., the supposition
that there is such a barber.
Grelling’s paradox, which takes off from the supposition of a word, ‘hetero-
logical’, that is true of a word if, and only if, that word is not true of itself,
may seem similar to these paradoxes. And one might conclude that the ques-
tion that drives Grelling’s paradox—Is ‘heterological’ heterological or not?—is
also not a “valid form,” and in the same way that the one that drove the barber
paradox is not a “valid form,” viz., the supposition that there is such a word as
‘heterological’.2
Let us think, for a moment, about these latter two paradoxes. With the barber
paradox, nobody is very disturbed, as pretty much everyone seems happy enough
to conclude that there simply cannot be any such barber. With Grelling’s paradox,
many people are about equally sanguine, also happy to agree that there really isn’t
any such word (or such a property) as ‘heterological’. Why is this so? There are
at least two reasons. The first may lead nowhere, but the second might lead to
somewhere of interest.
First, both with the barber story and with the putative word ‘heterological’
part of the reason for the complacency is that there is nothing plainly presented,
nothing at all, that looks so natural, or that looks so familiar, as to appear quite
certainly real. Perhaps this is most clear in the case of the supposed word ‘het-
erological’, which seems, on the contrary, to be a contrivance and, at that, an
utterly unfamiliar arrangement of letters, at least initially, or at least to those
unfamiliar with much logic or philosophy. And that might well be a large part
of what underlies the general complacency towards these two paradoxes. About
that large part, it seems that only so much can be done beyond what we have
pointed out.3 And, as seems most likely, that is that, without there being any-
thing here that will lead us anywhere else.
Second, there is another large similarity between the two paradoxes. With
both the barber paradox and Grelling’s paradox, the problem is isolated. You throw
away the tumor—or deny the imposter—and that is that; nothing else need be
involved. But, as regards this second part of what makes us complacent, looking
around just a little bit may give us a lead on how we may go from one of these
From No Peopl e to No Lang uag e s 25
it would follow from that supposition that ‘expression that is not true of itself’
is not true of itself. Thus, on the supposition that there is such an expression as
From No Peopl e to No Lang uag e s 29
‘expression that is not true of itself’ and that it is true of itself, it follows, courtesy
of (1), that it is not true of itself.
Suppose, then, that there is such an expression as ‘expression that is not true
of itself’, but that it is not true of itself. In that case, if there really is such an
expression and it is not true of itself, then since
(2) ˹E˺ is not true of itself iff it is not the case that ˹E˺ is E,
it would follow, from our assumption, together with (2), that it is not the case
that ‘expression that is not true of itself’ is an expression that is not true of itself,
which, via some innocuous (albeit classical) reasoning, yields that it is an expres-
sion that is true of itself. Thus, on the supposition that there is such an expression
as ‘expression that is not true of itself’ and that it is not true of itself, it follows
that it is true of itself. So, on the supposition that there is such an expression, we
may conclude that it is true of itself and that it is not true of itself. As a result,
‘expression that is not true of itself’ is app-inconsistent in which case, by (P3),
there is no such expression as ‘expression that is not true of itself’. But if there is
no such expression as ‘expression that is not true of itself’, then the question as to
whether it is true of itself or not does not—because it cannot—arise. This, then,
is our response to the semantic pathology that this variant of Grelling’s paradox
appears to present.8
real noun, we can also get ‘tribe that grows food for itself’. Similarly, from our
noun, ‘tribe’, we can also get ‘tribe that does not grow food for other tribes’ as well
as ‘tribe that does not grow food for itself’.
In parallel, if we have a real noun ‘expression’, we can get ‘expression that is
true of many birds’ and also ‘expression that is true of itself’. And, similarly, we
can get ‘expression that is not true of many birds’, and ‘expression that is not true
of itself’.
This has an important consequence, given our response to the semantic
pathology that ‘expression that is not true of itself’ appeared to present. As we
have seen, in light of our simple and (at least for some) seemingly uncontentious
semantic considerations, if there is an expression ‘expression’, then there is an
expression ‘expression that is not true of itself’, for, as we have indicated, ‘that is
not true of itself’ is a restrictive relative clause, which yields an expression when
added to an expression. This is not to say that if there are expressions, then there
are expressions that are not true of themselves. Rather, it is to say that if ‘expres-
sion’ is an expression—if there is an expression ‘expression’—then appending the
restrictive relative clause ‘that is not true of itself’ to it to form ‘expression that is
not true of itself’ yields, or results in, another expression. So, if ‘expression’ really
is an expression, then the same is true of ‘expression that is not true of itself’. This
unremarkable fact is all that we need to move on to the next step in our argument.
For with this fact, we are getting close to the sort of nihilism familiar from Unger
(1979a, 1979b, 1980), for what we will attempt to establish now is that there are
no expressions and, thus, no languages at all.
Now, the point that we are making is obviously not restricted to ‘cat’, for we are
making a point about expressions more generally. Hence, we propose the fol-
lowing schema, which will be accepted for the same reason that (3) should be
accepted, viz.,
From No Peopl e to No Lang uag e s 31
So far, as we have said, all of this seems unequivocal and, indeed, uncontroversial,
but, as we will show, it leads to something important.
As we have recently noted (section 2.5), if we suppose that ‘expression’ is an
expression—that is, if we suppose that there is an expression ‘expression’—then
we are compelled to contend that ‘expression that is not true of itself’ is likewise
an expression. But if ‘expression that is not true of itself’ is an expression, then,
given our extant commitments, it is either true of itself or is not true of itself. As
we have seen, assuming either disjunct leads to a contradiction. Hence, following
this chain of reasoning, if we suppose that ‘expression’ is true of something, then,
given our extant commitments, together with (5) and the fact that if ‘expression’
is an expression, then ‘expression that is not true of itself’ is likewise an expres-
sion, we eventually land in contradiction. Our supposition has led to a contradic-
tion, which renders ‘expression’ app-inconsistent. Thus, by (P3), our supposition,
that ‘expression’ is true of something, is incorrect, and, as we did in the case of
‘pencil’, we conclude that there are no expressions.10
We now proceed to draw out some consequences of this result in order to make
what will be, for some, a rather obvious, albeit new, point. Now, a language, what-
ever else it may or may not have, possess, or contain, must have (possess or con-
tain) at least some expressions. Hence, if there are no expressions, then there are
no languages, either. So, following the reasoning that we have employed, if Unger’s
response to the sorites is accepted, then there are no languages.
It seems that we have arrived at the place that we sought, having shown that
an Unger-style nihilism can be extended from our response to our variant of
Grelling’s paradox to the conclusion that there are no expressions and, thus, no
languages. But now we get some questions:
• Does the going response to Grelling’s paradox, and to our variant thereof,
which takes off from the purported expression ‘expression that is not true of
itself’, have significance for other semantic paradoxes, such as the liar paradox?
• If not, shouldn’t this response be dismissed, given the standard requirement
on a solution (or: response) to the semantic paradoxes, viz., that the solution
(or: response) be unified—that is, that it apply to all members of a family of
paradoxes?11
32 Reflections on the Liar
• If our response to Grelling’s paradox and its variant does possess some signifi-
cance for other of the paradoxes, then what sort of significance does it possess?
More directly, can our response to these alleged paradoxes shed light on the
liar paradox and its kin (e.g., Curry’s paradox, the truth teller paradox,12 etc.)?
As we will show, it may shed light on this whole family of paradoxes, even given the
app-inconsistency that such sentences reveal.
together with our supposition, it follows that the liar thought is not true. Thus, on
the supposition that there is such a liar thought and that it is true, it follows that
the liar thought is not true. Is that the end of the story here? Not at all, for there
is another part to consider.
In this other part, we again suppose that there is a liar thought, that this very
thought is not true, but this time we suppose that this liar thought is not true. Just
From No Peopl e to No Lang uag e s 33
so, the start of this part of the whole story is, as one might say, the opposite side
of the same coin, whose other side is our whole story’s first part, already consid-
ered, and still displayed just up above. At all events, let us notice what happens
were we to suppose that there is such a liar thought and that this very thought is
not true.
Given the relevant instance of the truth schema for propositions, (6), together
with the fact, under our supposition, that this very liar thought is the thought
that this very thought is not true, it would follow from our supposition that the
liar thought is not true, that the liar thought is true. Thus, on the supposition that
there is such a liar thought and that it is not true, it follows that the liar thought
is true.
Now we have told, and we have heard, both sides of the story here. The moral
of that apparently impeccable story is nothing less than this: If we suppose that
there is such a liar thought, we seem compelled to conclude that it, that very
thought, is both true and not true.
Be that as it may, and as the reader might have predicted, we can provide the
same sort of response to the liar thought that we provided for the variant of
Grelling’s, which emanates from Unger’s (1979a, 1979b, 1980) response to the
sorites. As we have seen, what is needed, in order to arrive at the contradiction,
that the liar thought is both true and not true, is the supposition that there is a
liar thought. That is, on the supposition that there is such a thought, then it will
be both true and not true. Our response, given by now familiar reasoning, cour-
tesy of a variant of (P3), is to deny the supposition that there is such a thought, in
which case the paradox that the supposition of a liar paradox appears to present
is avoided.14
or, perhaps, some variant thereof. The reason for this is that it seems that to
disbelieve, or even just to doubt, an instance of that schema seems to provide a
34 Reflections on the Liar
sufficient condition for lacking full understanding of either the sentence substi-
tuted for ‘p’ or the truth predicate (taking the biconditional as read). And, yet, at
least prima facie, an unrestricted acceptance of the instances of the truth schema
would appear to yield inconsistency, when coupled with seemingly banal, uncon-
tentious facts (e.g., that, in the sentence ‘L is false’, the letter, ‘L’, as it appears in
that sentence, names the very sentence of which it is a part, to wit, ‘L is false’).
In the case of the sorites paradox, the relevant claim seems to be a tolerance
principle, like (P1), noted above. Indeed, although many see tolerance principles
as being the culprit in the sorites paradox, like the instances of the truth schema,
such tolerance principles seem, again, at least prima facie, not merely true, but, in
fact, true in virtue of the meanings of the expressions for which they are tolerance
principles: Someone who doubted (P1) for ‘heap’, for example, would seem posi-
tively to be missing something crucial about that expression. Indeed, as Michael
Dummett (1975) argued, in order to understand the expression ‘looks red’, a
speaker must be willing to accept a tolerance principle that would enable her to
infer from ‘x is visually indiscriminable from y’ and ‘x looks red’ to ‘y looks red’.
And Roy Sorensen (2001) contends that semantic competence with respect even
to sorites-susceptible expressions involves a willingness to accept inferences like
from ‘n seconds after noon is noonish’ to ‘n + 1 seconds after noon is noonish’.
In this chapter, rather than identifying a common feature of the semantic
paradoxes and the sorites paradox, we presented a response to the former para-
doxes, which emerged from one for the latter paradox. In particular, after review-
ing Unger’s (1979a, 1979b, 1980) method for responding to the appearance of
sorties-susceptible expressions, we extended these points and showed that, with-
out much effort, we had the resources sufficient to respond to the putative pathol-
ogy that a number of semantic paradoxes appeared to present. Thus, rather than
identifying a feature common to the two sorts of paradoxes, we have argued that
a particular response to one sort of paradox could be applied to the other sort
of paradox. With that response in hand, so to speak, and with Unger’s (1979a,
1979b, 1980) assumptions clearly in mind, we then went on to offer a response
to the family of semantic paradoxes, including the liar paradox. Thus, rather than
seeing our response to the liar paradox as relevant to these other paradoxes, we
saw those paradoxes—and, in particular, our response to those paradoxes—as
relevant to the liar paradox.
‘autological’, it can be placed in either its own extension or in its own antiexten-
sion, without impending inconsistency. But since there appears to be no good rea-
son for placing it in its extension or in its antiextension, one might conclude that
the expression yields, or results in, indeterminacy, rather than app-inconsistency.
And the former is just a different aspect of semantic pathology.16
So, what about ‘expression that is true of itself’, the analogue to ‘autological’?
Does a supposition that there is such an expression as ‘expression that is true
of itself’ yield app-inconsistency, indeterminacy, or neither? And if it does yield
app-inconsistency (or, for that matter, some form of indeterminacy), what is our
response to the semantic pathology that that (alleged) expression would appear
to present?
To see the sort of semantic pathology that the (alleged) expression appears
to reveal, suppose that there is such an expression as ‘expression that is true of
itself’, which is either true of itself or is not true of itself. In particular, suppose
that there is such an expression and that it is true of itself. In that case, on the
supposition that there really is such an expression, then it is an expression that is
true of itself. Hence, on the supposition that there is such an expression and it is
true of itself then, by
(2) ˹E˺ is not true of itself iff it is not the case that ˹E˺ is E,
it is not the case that ‘expression that is true of itself’ is an expression that is
true of itself, which, via some more innocuous reasoning, would yield that it is an
expression that is not true of itself. Thus, on the supposition that there is such an
expression, we seem compelled to conclude that it is true of itself or is not true
of itself. But since there appears to be no good reason for taking it to be true of
itself, as opposed to its not being true of itself, or vice versa, then, as with the case
of ‘autological’, it seems that if there is such an expression, then it would yield,
or result in, indeterminacy. And, although indeterminacy may seem less perni-
cious than app-inconsistency, as we have said, it is another species of semantic
pathology.
So, what is our response to the seeming semantic pathology that this (sup-
posed) expression ‘expression that is true of itself’ appears to present? In order to
get to our response, recall, from the first days of logic class, that within the forma-
tion rules for the language of sentential logic, we find the “negation rule,” viz., that
if φ is a sentence (or a well-formed formula) in the language of a given sentential
36 Reflections on the Liar
logic, then the same is true of ~φ. As is familiar, the language of sentential logic
treats negation as a sentential operator, which is to say that, within the language of
that logic, negation attaches to a whole sentence (or a whole formula).
Within a natural language like English, it is not always held that negation func-
tion only as a sentential operator, as is indicated by the presence of internal, rather
than external, negation. As a result, if we are interested in setting out the syntactic
rules for negation, for a natural language like English, we cannot rest content with
just the negation rule for the language of logic.17
If we do heed this distinction, between internal and external negation, for a
natural language like English (and, for what follows, we will), then, with respect
to the syntactic rules for a natural language like English, we will find two negation
rules.18 The first rule, which will cover external negation, will be just like the one
for the language of sentential logic (although it will use ‘it is not the case that’,
rather than the negation symbol, the tilde) and the second rule, which will be for
internal negation, will have it that appending a negation (e.g., ‘not’) to an expres-
sion within a predicate, in a sentence or expression, will result in a sentence or
expression. So, for example, if ‘country that is impoverished’ is an expression (in,
or for, a natural language like English), then the same will be true of ‘country that
is not impoverished’ (in, or for, that language).
Applying this to the present case, we find that if ‘expression that is true of
itself’ is an expression of, or for, our natural language (in this case, English), then
since ‘expression that is not true of itself’ is, or involves, the internal negation of
‘expression that is true of itself’, the same must be true of ‘expression that is not
true of itself’. That is, if ‘expression that is true of itself’ is an expression, then
‘expression that is not true of itself’ is also an expression. But, as we have recently
seen, the latter is not an expression from which it follows, by elementary logic,
that the former is not, either. But if there is no such expression as ‘expression that
is true of itself’, then the threat of impending indeterminacy dissolves. This is the
case because if there is no expression ‘expression that is true of itself’, then the
question of whether such an expression is true of itself or is not true of itself does
not—because it cannot—arise.
Notes
1. The expression that he (1979a) considers, ‘expression that does not apply to itself’, is a vari-
ant of the one that drives Grelling’s paradox.
2. A related response to the paradox that ‘heterological’ appears to present might deny that
there is a property heterological, which would correspond with ‘heterological’.
3. Well, more can be done, if one is prepared to reject classical logic. For a consistent solution
to Grelling’s paradox that is along those lines, which does not involve jettisoning ‘hetero-
logical’ (or the unrestricted schema, which might serve to characterize this neologism), see
Field (2008).
4. Of course we would say the same about ‘word that is not true of itself’, in addition to the
initially indeterminacy-involving cases, like ‘word that does apply to itself’ and ‘word that
From No Peopl e to No Lang uag e s 37
is true of itself’. Since none of these are words, none of the apparent inconsistency or
indeterminacy—in effect, their apparent semantic pathology—results.
5. We could use ‘applies to’ and ‘does not apply to’, rather than ‘true of’ and ‘not true of’, as we
have been doing. For convenience only, though, we will stick with the latter pair of expres-
sions. Since the pairs are synonymous, as we have thus far seen, everything that we say
using the latter pair can be said using the former pair.
6. Epistemicists will of course refuse to accept (P1), though they will accept (P2).
7. This assumption that if there is such an expression, then it is true of itself or not is justified
by the fact that we are here following Unger’s (1979a, 1979b, 1980) reasoning, which we
take to involve an acceptance of classical logic and classical semantics.
8. Our response to the version of Grelling’s paradox that we presented clearly applies to the
original version of that paradox, which takes off from the introduction of the neologism
‘heterological’, which is such that
On the supposition that there is such an expression as ‘heterological’, classical logic (and
classical semantics), together with the relevant instantiation of (2#), viz.,
(H) ‘heterological’ is heterological if, and only if, heterological is not heterological
yields that ‘heterological’ is, and is not, heterological. Courtesy of (P3), it follows that there
is no such expression. Thus, we offer the same response to the original version of Grelling’s
paradox that we offer to the newer version, recently introduced.
Although it is infrequently mentioned, the dual of ‘heterological’, ‘autological’, also appears
to result in a sort of semantic pathology, as a consequence of the apparently resulting inde-
terminacy that the supposition that there is such an expression makes evident. In order to
see our response to the “paradox” that this supposition appears to present, see the appendix
to this chapter.
9. When we try to combine a restrictive relative clause with a name, it seems that we need a
definite article, which thereby treats the name as a predicate. For example, although ‘Frege,
who introduced the distinction between sense and reference’ is fine, we are not combining
the restrictive relative clause with the name and, when we do try, getting something like
‘Frege who introduced the distinction between sense and reference’, we get something that
is marked because seemingly ungrammatical. That said, there is nothing wrong with ‘The
Frege who introduced the distinction between sense and reference’, though now ‘Frege’ is
functioning as a predicate, rather than as a name. For the claim that names are predicates,
see Graff Fara (2014).
10. One might worry that we cannot articulate our recent conclusion, that there are no expres-
sions, without ourselves using expressions. But recall that what we are really establishing is
that if Unger’s response to the sorites is accepted, then there are no expressions. Since we
are not taking a stance on the status of the antecedent to that conditional, we do not face
the worry that we have to use the very things whose existence we are denying.
11. For arguments to this effect, see Priest (1994).
12. Recall that the truth-teller paradox takes off from and can be shown to result in apparent
indeterminacy, in roughly the way that having the expressions ‘autological’ and ‘expression
that is true of itself’ purport to do. We address the apparent indeterminacy that ‘autological’
and ‘expression that is true of itself’ reveals in the appendix to this chapter.
13. We remind the reader that we are here offering a response to the semantic paradoxes, which
is different from offering a solution to a paradox. A solution to a paradox, to be adequate, will
provide both a diagnosis and a treatment of that paradox. As seems clear, we are offering
neither a diagnosis nor a (complete) treatment of these paradoxes.
38 Reflections on the Liar
15. Armour-Garb would like to thank Doug Patterson for a discussion of the relation between
the sorites and the liar paradox.
16. For the claim that indeterminacy is an element of semantic pathology, see Armour-Garb and
Woodbridge (2015).
17. To see the difference, notice that some might accept that it is not the case that the tooth
fairy is coming without accepting that the tooth fairy is not coming. We could explain this by
noting the difference between, in this case, external and internal negation.
18. We will need to provide “formation rules” for internal and external negation, for a language
like English, even if there is a direct, immediate inference from any occurrence of external
(internal) negation to an occurrence of internal (external) negation, as might well be the
case, given our assumptions of classical logic and classical semantics.
References
Armour-Garb, B., and J. Woodbridge. 2015. Pretense and Pathology: Philosophical Fictionalism and
Its Applications. Cambridge: Cambridge University Press.
Dummett, M. 1975. “Wang’s Paradox.” Synthese 30: 301–324.
Field, H. 2008. Saving Truth from Paradox. Oxford: Clarendon Press.
Graff Fara, D. 2014. “Names Are Predicates.” Philosophical Review 124 (1): 59–117.
McGee, V. 1985. Truth, Vagueness and Paradox: An Essay in the Logic of Truth. Indianapolis: Hackett.
Priest, G. 1994. “The Structure of the Paradoxes of Self-Reference.” Mind 103 (409): 25–34.
Russell, B. 1918. The Philosophy of Logical Atomism, London: Routledge.
Sorensen, R. 2001. Vagueness and Contradiction. Oxford: Clarendon Press.
Tappenden, J. 1993. “The Liar and Sorites Paradoxes: Toward a Unified Treatment.” Journal of
Philosophy 90 (11): 551–577.
Tarski, A. 1944. “The Semantic Conception of Truth and the Foundations of Semantics.”
Philosophy and Phenomenological Research 4: 341–376.
Unger, P. 1979a. “Why There Are No People.” Midwest Studies in Philosophy 4: 177–222.
———. 1979b. “I Do Not Exist.” In Graham F. Macdonald, ed., Perception and Identity, 235–251.
Ithaca, NY: Cornell University Press.
———. 1980. ‘Skepticism and Nihilism.” Noûs 14 (4; Special Issue on Epistemology): 517–545.
3
The story of the liar paradox begins, according to tradition, with Epimenides the
Cretan, who said, “All Cretans are liars.” If all Cretans are liars and Epimenides is
a Cretan, then Epimenides does not tell the truth when he says, “All Cretans are
liars.” In turn, if what Epimenides says is a lie, but at the same time says what
is true, then he is a liar who speaks truly—which, frankly, does not make much
sense at all! It is difficult to not be engaged by the liar paradox once one has com-
mand of what the problem is.
Most work on the liar paradox has treated it as a special kind of logical or
formal problem. Our aim here is to suggest that an understanding of the liar
paradox can be advanced by taking a look not directly at the problem itself,
but rather at how people think about the problem. There is, we maintain, an
underappreciated psychological dimension to the liar paradox that may inform
our work on the logical and formal dimensions of the problem. While the expe-
rience of confronting a simple contradiction is psychologically unremarkable,
the experience of thinking through a version of the liar paradox requires one
to appreciate inferential transitions and aporetic disorientation (cf. Rescher
2001). Armchair speculation alone cannot reveal the particular contours of how
different people think about the liar paradox, but experimental investigation
can, albeit indirectly.
39
40 Reflections on the Liar
This picture of the cognitive inadequacy of the folk strikes us as a plausible arm-
chair psychological hypothesis—yet only a hypothesis. There could be other, rival
plausible hypotheses.
For example, to frame his discussion of how he thinks someone might work
through the liar paradox, Maudlin imagines a hypothetical survey and then artic-
ulates an intuitive model for how one might informally evaluate the liar sentence,
‘ “This sentence is false’.”1 He writes:
Think ing about the Liar, Fast and S l ow 41
To be clear, Maudlin is not at all claiming that everyone, or even most people,
will go through these steps in their mind when they confront LIAR and its ilk.
His model of LIAR-directed thought is not meant to model specific processes of
human cognition; rather, we take him to be presenting a story about what prethe-
oretic, informal resources are available to ordinary cognizers, in confronting
such paradoxes, and their adequacy to arrive at a view about truth-value gaps. As
opposed to Barwise and Etchemendy’s model, Maudlin’s account is an exemplifi-
cation of a folk coherence model of LIAR-directed thought.
On this “Maudlin model,” the conclusion that some sentences are neither true
nor false follows fairly simply from any initial attempt to assign LIAR a truth
value. The hard philosophical work for him comes next in presenting a theory of
42 Reflections on the Liar
semantics and truth that can accommodate that conclusion, but it takes almost
no rigorous philosophical reasoning to get to the conclusion in the first place.
Nonetheless, if the story is a simple one, it is an empirically committed one as
well. The picture of ordinary cognition about LIAR develops in a way that would
have us conforming to some familiar logical principles, e.g., contradiction avoid-
ance, but not to others, e.g., bivalence. We can thus ask both whether his account
of ordinary cognition is accurate in extension—in terms of whether people gen-
erally do have some resources that will bring them to the conclusion that there
are truth-value gaps—and also whether it is accurate in terms of the particular
resources appealed to—in terms of what logical principles people generally can
access.2
So, we have our first motivation for commencing empirical work on the
psychology of the liar paradox. While philosophers have appealed to different
accounts of what our nonspecialist resources are in this area, the accounts them-
selves are speculative at best, and as a group are not obviously consistent with
each other (in a way that even dialetheists should be potentially embarrassed).
Though the philosophers we have cited here are fairly recent, the question of
folk coherence versus folk incoherence—and attempts to answer that question
in an empirically grounded way—have a longer history than that, and we take
ourselves to be following in the footsteps of the Norwegian empirical semanti-
cist Arne Naess. In 1938, after having performed numerous experimental studies
on the ordinary person’s conception of truth (cf. Naess 1938a, 1938b), Naess
reported that people do not typically agree with contradictory or inconsistent
expressions. Just as philosophers abhor inconsistency, so too do ordinary people.
It goes without saying that we likely hold inconsistent beliefs, though we may
not be aware of them. Further, we might not acknowledge this fact even if it were
Think ing about the Liar, Fast and S l ow 43
So, we have reports of Naess’s work seeming to indicate that ordinary people do
not embrace beliefs one might think are inconsistent, contradictory, or confused.
Nonetheless, we only have these reports, and, moreover, one might argue that
Naess’s methods may be aligned more with ethnography than with experimental
philosophy, and, as such, his results would be somewhat limited in how much pur-
chase they could give us on participants’ aspects of cognition. Thus more, newer
work is needed, and using different tools.
We have written thus far of “the folk” in a fairly undifferentiated way. But our
further motivations for empirical work on the liar paradox involve two dimen-
sions of complicating that too-simple picture of a uniform, universal set of cog-
nitive resources and proclivities. One such dimension is that of demographic
difference. Some work in experimental philosophy has indicated interpersonal
differences in judgments in various philosophical domains, including variation by
ethnicity (e.g., Machery et al. 2004) and personality type (e.g., Feltz and Cokely
44 Reflections on the Liar
2009). Although gender effects have proved elusive in the experimental philos-
ophy literature, since there are at least some results indicating possible gender
differences specifically in matters of truth (Barnard and Ulatowski 2013, n.d.),
we took this opportunity to investigate whether gender correlated with different
approaches to LIAR.
Surely other hypotheses of demographic differences are worthy of exploration
in this regard, especially ethnicity (see, e.g., Peng and Nisbett 1999). For this pre-
liminary study, though, we only explored one further demographic variable. We
conjectured that perhaps individuals’ religious commitments could interact inter-
estingly with how they would approach questions of paradox and contradiction,
since many common religious doctrines endorse sets of propositions that are at
least prima facie inconsistent (such as the Trinity). We wished to test the predic-
tion as to whether self-reported theists were more tolerant of contradictions than
self-reported nontheists.
Finally, if our first motivation to do empirical work on LIAR is to document
what resources the folk in general do or do not bring to bear in engaging with it,
our last motivation concerns the difference between the folk on the whole, and
what might be termed more cognitively elite folks, either in terms of reflective
ability or philosophical training. Clearly, trained philosophers have greater for-
mal resources to wrestle with thorny questions about, say, self-reference, or the
theory of types. But while the folk may lack the wherewithal to address formal
issues in and around the liar paradox, when we philosophers too step away from
the areas of our technical training, it is an open empirical question whether our
skills, training, or background as philosophers really give us better tools for unty-
ing that knot, compared with the folk. Even philosophers are prone to err in ways
we like to imagine are only common among the folk. We do not deny that training
and practice in formal methods may produce great differences in how well for-
mal tasks can be performed, but we take it to be an open empirical question just
where and to what extent philosophical training more generally leads to diver-
gences from the cognition of nonphilosophers. Although many philosophers have
suggested sweeping and dramatic differences (cf. Ludwig 2007; Hales 2000, 2004,
2009, 2012; Williamson 2004, 2005, 2007, 2011, 2013), a fair number of stud-
ies have been conducted looking for such differences, and by and large they have
shown philosophers’ minds to work pretty much the same as those of nonphi-
losophers (cf. Schulz et al. 2011; Machery 2009; Schwitzgebel and Cushman 2012;
Schwitzgebel and Rust 2009).
And so a further hypothesis we were interested in considering is whether a
person’s tendency to think reflectively about versions of LIAR could produce
systematic changes in people’s judgments about it. Perhaps LIAR just seems to
strike different people in different ways, and we may diverge in our quick, intui-
tive reactions. Other people generally find LIAR thought-provoking. It is not
easy to discern from the armchair alone the extent to which people’s thinking
about LIAR operates in service of their initial reactions, or to what extent they
Think ing about the Liar, Fast and S l ow 45
are working through considerations that lead them to reflectively modify their
evaluation. There are good theoretical psychological reasons to find both kinds
of consequences of reflection possible here. A central part of the “dual process”
theories of human cognition is that not only do we have distinct intuitive and
reflective systems of judgment, they do not simply operate in parallel on separate
domains. Rather, the reflective activity in “System 2” can serve as a check on the
intuitive products of the cognitive heuristics in “System 1” (Evans 1984; Sloman
1996; Kahneman 2003), and yet the reflective system can also be subject to con-
firmation bias, and can just further “lock in” whatever the agent’s initial take on a
question may be (Mercier and Sperber 2011).5
We should consider as one hypothesis whether we have an initial intuitive reac-
tion that applies a classic truth-value to the liar sentence or liar-like paradox. For
example, perhaps an intuitive matching bias (Evans and Lynch 1973; though, see
also Evans 1998) may lead us to an initial response that LIAR is false, because,
simply put, that’s what it says. We see the word ‘false’, and the mind finds it easi-
est to just use the word right before it as the answer.
Another candidate along similar lines would be the phenomenon of anchoring
(Kahneman and Tversky 1972), in which an initially provided value gets uncon-
sciously recruited as an answer to a question, and it requires some mental effort
to move away from that anchor. This can happen even in situations when one
knows full well that there is no reason to expect the anchor in question to be a reli-
able guide, such as the last three digits of one’s phone number anchoring people’s
estimates of when Attila the Hun departed Europe (Russo and Shoemaker 1989).
It can require cognitive effort to move off of an initial estimate, and when we lack
the will to do so, or when our cognitive resources are depleted, we are particularly
susceptible to the influence of such anchors (Epley and Gilovich 2006). Moreover,
there is also some evidence (e.g., Gilbert, Tafarodi, and Malone 1993) that we tend
to default to accepting presented claims as true, and that it takes at least a bit of
further cognitive activity thereupon to reject them, and so some people may have
an initial intuitive response ascribing truth to LIAR, especially, again, if they lack
the will, interest, or cognitive resources to override such an initial reaction.
The “Maudlin model” of LIAR cognition discussed above fits well with that
picture of reflection engaging with and correcting for initial intuitive responses,
but there are other psychologically plausible models here, and we cannot tell
just by introspecting which is true. Results like Cacioppo and Petty (1982) and
Stanovich, West, and Toplak (2013) indicate that confirmation bias can not only
persist in the face of reflection, it can perhaps be intensified by it. High ability in
reasoning can also be high ability in rationalizing. If the arguments of Mercier and
Sperber (2011) are correct, then this myside bias even in reflective thought may
be a central evolutionary design feature of our minds, and not just some unusual
misfiring cylinder in the engine of cognition: our deliberative capacities have the
function not so much of determining when we ourselves are wrong, but in per-
suading people—both others and, for that matter, ourselves—that we are right.6
46 Reflections on the Liar
Second, and more importantly, while these measures describe the extent
to which some participants engaged, to a greater or lesser degrees, in reflective
thought compared to other respondents, these measures do not tell us about the
actual thinking processes employed by them. In particular, we cannot tell, e.g., the
extent to which they did or did not explicitly consider or try to apply principles
such as the principle of bivalence or the principle of noncontradiction. The results
in aggregate do indicate whether responses tend to conform to such principles,
and this is enough to warrant some cautious speculation.
5.0%
20.7%
17.7%
56.5%
Then, all participants received a CRT and WST before receiving a probe question,
including the liar paradox or liar-like paradox.
As we saw earlier with Maudlin and with Barwise and Etchemendy, attempts to
philosophically analyze problems like the liar paradox tend to make a wide range
of assumptions, implicit or explicit, about the philosophizing subject. One imme-
diate way in which the empirical investigation can be relevant to problems like
LIAR is that we can recognize the meaningful effect that fairly common individual
level differences might have.
The distribution of people’s responses to LIAR reflects in the following pat-
tern: 5.0% TRUE, 56.4% NEITHER, 17.8% BOTH, and 20.8% FALSE (see
figure 3.1.). Initially, two key points emerge from our data: (1) people’s responses,
despite having to choose among just the four we experimenters provided them,
are richly varied and fragmented, and (2) the NEITHER response, which is one of
the four choices offered, is the most common.
Determining whether the distribution of responses reflects explicit differ-
ences of logical or metalogical beliefs or arises from other kinds of individual level
factors is something that interested us in our analysis. So, first, we considered
whether the awareness of philosophical matters in people’s responses to LIAR.
Subjects were individuated according to a measure we deemed “philosophical
awareness.” “Philosophically aware” participants (according to our stipulation)
were those who self-reported as having had more than two college-level courses in
philosophy (N = 93). We compared them with “philosophically unaware” respon-
dents, those who self-reported having had two or fewer college-level courses in
philosophy (N = 143).
While the observed response patterns between the philosophically aware
and unaware appear to differ, the difference is not statistically significant
( χ2 = 5.559, ns) (see figure 3.2).16 This result makes sense, however, if philosophi-
cal training must reach a threshold level in order to have a measurable effect or if
Think ing about the Liar, Fast and S l ow 49
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
Philosophically Sophisticated Philosophically Unsophisticated
the philosophical training must be combined with some other factor. The liar par-
adox is a problem familiar to many philosophers, and one would naturally expect
that a higher level of philosophical training ought to have a meaningful effect on
how one responds to it.
To examine this possibility we considered whether the responses of people
with high levels of philosophical training would differ from the responses of oth-
ers. We compared responses provided by “philosophical experts” (graduate stu-
dents and holders of a Ph.D. in philosophy), N = 61, with those participants who
are non-expert (everyone else), N = 175.
In this case, the responses not only appear to differ but there was a statis-
tically significant difference between the responses of experts and non-experts
( χ2 = 11.701, df = 3, p < .008, Low-to-medium effect size, Cramer’s V = .223) (see
figure 3.3).
Other categorical factors that might influence how one responds to LIAR, of
course, include whether or not one is a reflective thinker. Philosophical exper-
tise and reflective thought are often thought of as complementary, although
exercising one’s expertise and exercising reflective thought are not necessar-
ily the same thing. To test whether reflective thinking makes a difference, we
looked at response patterns to LIAR based upon whether respondents gave a
“passing” response to our modified WST17 and whether they scored a perfect 3
on the CRT.
In the WST case, the responses differ dramatically for the NEITHER and FALSE
responses. Comparing WST passers with WST failers on LIAR, we see that the
50 Reflections on the Liar
80
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
Experts Non-Experts
80
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
WST Passers WST Failers
Figure 3.4. Comparison of WST Passers (N = 52) and WST Failers (N = 184) to LIAR.
difference between them was statistically significant ( χ2 = 16.084, df = 3, p < .001,
low to medium effect, Cramer’s V = .261) (see figure 3.4).
Similar results cropped up when we compared the maximal score of 3 on the
CRT with those who scored lower than that.
Again, we see that the response patterns differ where the two most obvious
differences are NEITHER and FALSE. For CRT passers and CRT failers on LIAR,
Think ing about the Liar, Fast and S l ow 51
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
CRT Passers CRT Failers
80
70
60
50
40
30
20
10
0
0.0 1.0 2.0 3.0
NEITHER responses Other responses
some proposition is either true or false. Because we partly expected that response
to be popular for LIAR, we decided also to ask participants whether they believed
the law of noncontradiction, i.e., “A sentence cannot be both true and false,” is
true or not-true. Interestingly, 68.9% of 90 participants who said that CON is true
believed LIAR was NEITHER, while 48.8% of 146 respondents who agreed CON is
not true responded that LIAR was NEITHER.
Again, as we have found in the other cases, the response patterns differ, espe-
cially for the response NEITHER. Similarly, and as one might expect, respondents
who believed that CON was not true tended to respond BOTH to LIAR, and par-
ticipants who believed that CON is true tended to agree less with BOTH to LIAR.
The difference was statistically significant (χ2 = 12.948, df = 3, p < .005, low to
medium effect, Cramer’s V = .234) (see figure 3.7). However, as will be important
to our arguments later, both CON endorsers and CON rejecters shared the strong
modal response of NEITHER.
We also queried respondents on the principle of bivalence, i.e., the view that
every proposition is either true or false. Among 59 respondents who believed the
principle of bivalence is true, 49.2% responded that LIAR is NEITHER, and 58.8%
of 177 respondents who believed the principle of bivalence is not true responded
that LIAR is NEITHER.
Just as we have discovered with CON responses and others, the response pat-
terns differ for BI. The difference between response patterns was statistically sig-
nificant (χ2 = 10.278, df = 3, p < .016, low to medium effect, Cramer’s V = .209) (see
figure 3.8)—though, again, with many participants choosing NEITHER. In short,
it seems that while some of the overall variation in responses may be explained in
Think ing about the Liar, Fast and S l ow 53
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
CON TRUE CON NOT-TRUE
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
BI TRUE BI NOT-TRUE
seem to indicate that women respondents are on average less likely to believe
truth is correspondence to reality, and women respondents are more inclined on
average than men to believe that some degree of evidence is required for truth.
In turn, we looked at whether there was a corresponding effect of gender on the
response pattern for LIAR. We did find a significant difference between male and
female response patterns to LIAR (χ2 = 16.502, df = 3, p < .001, l ow to medium
effect, Cramer’s V = .264) (see figure 3.9). We did not make an explicit prediction
in advance of what this difference might look like, but we can offer a post hoc
interpretation of our finding. So far as we can tell from the data, women were on
balance more willing than men to respond with FALSE to LIAR.
Given the Barnard and Ulatowski results, the male participants in our cur-
rent study might be comparatively reticent for any claim to count as true or false
unless it, or its negation, corresponds with the way the world is, and LIAR, since
it lacks such a truth maker, would seem a poorer candidate for either one of the
traditional truth values. Moreover, since no evidence plays a role in deciding the
truth value of LIAR, women might be somewhat more willing to offer a rejection
of it as FALSE. We would note, however, that this difference shows up against a
background of NEITHER still being the seemingly preferred response for both
groups. Our claim is not that women and men deploy the notion of truth in radi-
cally different ways, let alone that they have anything like different concepts of
truth, but only that these findings suggest some intriguing differences around the
edges regarding how men and women confront the paradox.
Moreover, also, as mentioned above, we hypothesized that since the theologi-
cal commitments of some religious traditions affirm claims that can be inter-
preted as contradictory, e.g., that in the Holy Trinity 1 = 3, theists might have
an openness to contradiction that in turn might have an effect on responses to
LIAR. Here, too, we found a significant difference, between theist and nontheist
70
60
50
40
30
20
10
0
True Neither True Both True False
nor False and False
Male Female
70
60
50
40
30
20
10
0
True Neither True Both True False
nor False and False
Theist Non-theist
participants (χ2 = 10.033, df = 3, p < .018, low to medium effect, Cramer’s V = .206)
(see figure 3.10).18
These individual-level differences are interesting to contemplate, but their
overall importance in understanding how the folk think about the liar requires a
more rigorous evaluation.
displayed a variety of attitudes toward basic logical principles, and while their
doing so produced interesting effects on their patterns of responses, the prin-
ciples themselves do not seem to be driving the NEITHER response. How could
they, when they are not even endorsed by groups of participants who have that
response? Our data cannot rule out, and perhaps they even suggest, that some
participants are relying consciously on those principles. But it can only be a
minority of them at best.
We thus arrive at the following account of the general difference between those
who gave the NEITHER response and the smaller but not insubstantial number of
our participants who gave such answers as FALSE. We conjecture that the latter
group are primarily participants who simply failed to reflect on the target sen-
tence. When confronted with a sentence that says of itself that it is false, these
participants apparently resisted the urge—or, more likely, failed even to have any
impetus—to reconsider the truth value of the sentence, and simply chose TRUE
or FALSE in effortless accordance with the deliverances of their System 1. As the
data seem to suggest, the pragmatic response to the truth value of LIAR is FALSE
because that is exactly what the sentence says it is—FALSE. Such folks judge the
truth value of the sentence efficiently without necessarily any accompanying
awareness of the source of that judgment.
In contrast, participants who tended to say NEITHER in response to LIAR are
more reflective. Upon reflection, these respondents see that flat-footed responses,
TRUE or FALSE, lead to an unwieldly result: If LIAR is false, then it is TRUE and if
LIAR is true, then it is FALSE. Sometimes explicitly but more often implicitly, such
participants take it that neither TRUE nor FALSE, sans phrase, is a stable response
to LIAR, so they choose NEITHER.
One might still ask: Why do so comparatively few subjects, both reflective and not,
say BOTH? The “Maudlin model” is helpful here. Remember, Maudlin suggested
that once the person comprehends the paradoxicality of TRUE and FALSE, BOTH
is no longer a live option. He writes, “If both [TRUE] and [FALSE] are wrong,
surely [BOTH] is” (Maudlin 2004, p. 1). Our theory is that reflective subjects
intuit that neither TRUE nor FALSE is a legitimate option, and then can draw
the easy inference along those lines Maudlin suggests. For respondents who are
relying only upon the deliverances of their intuitive System 1, BOTH is not a live
option, just as NEITHER is not a live option—their reason seems to be That’s not
what the sentence says of itself. The more complicated answers simply fail to pres-
ent themselves as intuitive candidates, if reflection is not brought to bear.
To discover which interpretation is getting closer to what the data report is to
figure out which of the factors matter most in driving participants’ responses to
LIAR. In order to determine which of the factors that generate distinct responses
to LIAR are most salient, we had to perform a logistic regression analysis. The
regression model included factors that were earlier found to yield significant dif-
ferences in response patterns: respondent gender and theism, CRT score, WST
performance, and acceptance of bivalence and noncontradiction. Significant
Think ing about the Liar, Fast and S l ow 57
results indicate that a factor has an effect on response patterns for LIAR after
controlling for all other factors in the model.19 The β-coefficient reports how
many standard deviations a dependent variable will change per standard devia-
tion increase in the predictor variable.
We performed two multinomial logistic regression analyses. One compared
NEITHER with FALSE (table 3.1), and the other compared NEITHER with BOTH
(table 3.2).
When we originally compared responses from people, e.g., who answered
that CON is TRUE or who self-reported as theist, with those from people in the
complement class, we discovered significant differences. But those particular dif-
ferences do not remain significant when the effects from other factors in the
model are considered at the same time. This is not the case for acceptance of BI,
gender, and WST performance. These three factors have an effect that is suffi-
ciently robust to remain significant when we control for the effect of all the other
factors. Plainly put, these factors seem to drive whether a respondent answers
NEITHER (as opposed to FALSE). A similar regression analysis was performed
looking at what factors would influence respondents to respond NEITHER rather
than BOTH. In this case, the only factor that remained significant was respon-
dent gender.20
Given that the LIAR sentence says of itself that it is false, we treated that
response as the main intuitive default view. Our regression analysis was designed
to see what factors in the model could help explain why participants answered
something other than FALSE. When all of the other factors are controlled for, peo-
ple who reject bivalence are about three times more likely to respond NEITHER
than they are likely to respond FALSE. When all of the other factors are controlled
for, those who can pass the modified WST test (i.e., understand classical logical
implication) are about five times more likely to respond NEITHER than they are
likely to respond FALSE. When all of the other factors are controlled for, men are
roughly 2.25 times more likely to respond NEITHER than women are likely to
respond NEITHER.21
Our work here is at best suggestive, and in particular should not be taken as
in any sense definitive in the absence of further exploration and, especially, rep-
lication. And much more work is required to determine to what extent reflection
is efficacious because we bring logical principles to bear, or instead because it
gets our intuitive mental logic capacities working properly on the problem—and
indeed, whether there are interesting interpersonal differences in the answer to
that question.
5.0%
20.0%
16.7%
58.3%
If we compare the raw descriptive data of responses of LIAR (cf. figure 3.1)
with the responses of SB (figure 3.12), we notice a striking similarity between
them: no univocal response, but with a strong modal endorsement of NEITHER.
Indeed, the overall proportions of the four answers are highly similar across the
two tasks. Once we saw how closely the two sets of data overlapped, we decided to
subject the results of SB to analyses similar to the ones we ran for LIAR.
What we found for SB was largely consistent with our findings for LIAR. For
philosophical expertise, the differences were statistically significant ( χ2 = 8.469,
df = 3, p < .037, medium effect, Cramer’s V = .376) (see figure 3.13).
For WST passers and failers, the difference was statistically significant
( χ2 = 11.854, df = 3, p < .008, medium to large effect, Cramer’s V = .444) (see
figure 3.14).
For CRT passers and failers, the difference was statistically significant
( χ2 = 8.042, df = 3, p < .045, medium effect, Cramer’s V = .366) (see figure 3.15).
When we use respondents’ actual score on the CRT to assess the role of reflec-
tive thinking in a more fine-grained way, the difference was statistically signifi-
cant for SB, just as in LIAR ( χ2 = 24.752, df = 9, p < .003). There is an observable
trend that CRT score and percentage of respondents answering NEITHER to SB
are correlated (Kendall’s τ c = .259, est T = 2.865, p < .004) (figure 3.16). When we
control for philosophical expertise, the correlation remains significant for experts
(experts: Kendall’s τ c = .48, est T = 2.306, p < .02).
60 Reflections on the Liar
90
80
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
Experts Non-Experts
100
90
80
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
WST Passers WST Failers
The collected data on SB resemble the LIAR data, which seems to show that
we have a conceptual replication of the effect concerning reflectivity and liar-
like semantic paradoxes. Primarily, the finding just is that respondents whose
intuitive responses tend to reveal a certain amount of reflection agree with the
response NEITHER for SB, while those participants who responded with FALSE
seem to reveal that their intuitive responses were driven by pragmatic, automatic,
Think ing about the Liar, Fast and S l ow 61
80
70
60
50
40
30
20
10
0
True Neither True nor False Both True and False False
CRT Passers CRT Failers
80
70
60
50
40
30
20
10
0
0.0 1.0 2.0 3.0
NEITHER responses Other responses
and associative responses. The stark difference between low-CRT and high-CRT
scoring subjects is plainly visible in figure 3.15.
These dramatic differences by CRT score lead us to an important cautionary
note about how to interpret what our results seem to be saying about the popu-
lation on the whole. By design, our sample recruited a higher proportion of phi-
losophers and of high-CRT people than is representative of the population on the
62 Reflections on the Liar
whole. (For example, Frederick [2005] reports that about 60% of his college-aged
participants got a score of CRT-0 or CRT-1.) Although NEITHER is still a more
common answer than any other single answer in the low-CRT groups, FALSE is a
very close second behind it in that part of our sample.
Of course, we should not close the case just yet. Even putting aside the fact
that we are only reporting one study here, there are also a great many variations
on the liar paradox, and perhaps there is an iteration of it that yields dissimilar
results, i.e., data that show people tend not to reflect in some response to some
instances of semantic paradox. We believed early on that people’s responses to
LIAR might differ greatly from their responses to other paradoxes, largely because
people might not have an overwhelming reaction to LIAR. After all, think of how
freshmen in an introductory philosophy course often react to the notion that the
principle of moral or epistemic relativism, i.e., “There are no objective moral/epis-
temic principles,” is self-defeating. Sometimes novice thinkers just are not able to
grasp the problem. It is too difficult for the person to step back and get a look at
the principle for what it is saying. This led us to our third and final experiment,
exploring one further way of presenting the liar paradox.
asks Jones whether he may cross. Jones declines. Smith pleads with
Jones to allow him to cross the bridge because he doesn’t want to walk
to the next footbridge many miles from this one.
Jones says, “If the first sentence you utter is true, then I will allow
you to cross the bridge. But if the first sentence you utter is false, then
I will throw you into the water.”
Smith responds, “You will throw me into the water.”
If Jones throws Smith into the water, then the first sentence Smith
uttered was true and Jones should have let him pass. If Jones lets
Smith pass, then the first sentence Smith uttered was false and Jones
should have thrown him into the water.22
is TRUE, BOTH, NEITHER, or FALSE. This item was presented to 114 subjects.
The general distribution looked like figure 3.17.
This overall response pattern for THROW is observably different from the
response pattern for LIAR. We note, for instance, that the modal response to
THROW is BOTH. At a statistical level the response patterns are significantly
different. This was determined using a likelihood ratio analysis comparing LIAR
and THROW (λ = 19.058, df = 9, p < .025, small to medium effect, Cramer’s
V = .213).23
At a minimum, we do now have some reason to think that this presentation
of the liar leads to a different pattern of responses. We are inclined toward the
conjecture that the ability to confront the paradox abstractly is diminished by
the practical factors involved in how the paradox is presented. But the best we
16.7% 14.0%
31.6%
37.7%
can say at present is that our data are consistent with such a conjecture; further
investigation of practical presentations of the LIAR are required.
3.7. Conclusion
With our results now before us, let us conclude by revisiting our empirical hypoth-
eses that motivated our study, as initially discussed in section 3.2.
First, regarding the disagreement among philosophers regarding the coher-
ence or incoherence of folk cognitive resources to address the liar, we do take
ourselves to have some evidence for a particular psychological model of how
people engage the liar paradox. It appears that quick, associative responses to
theoretical forms of the paradox, such as LIAR and SB, tend to be simply “true”
or “false,” mostly “false,” But fast and frugal responses are distinct from more
reflective intuitive responses, and that reflection, especially if aided by some phil-
osophical training, generally leads people to endorse a “neither” response. For
practical presentations of the paradoxes, such as the Buridan’s bridge paradox,
respondents have a greater tendency to stay with that fast, prereflective response.
In all, it seems that initial and prereflective folk responses are perhaps not well
suited to engage with the paradox. The further resources that are put into play by
a process of reflection are able to track the paradoxical nature of sentences like
LIAR, and on the whole, something like the “Maudlin model” seems to be consis-
tent with our findings. Interestingly, it seems that the relevance of reflection may
be less a matter of bringing explicit beliefs about logical principles to bear, and
more a matter of shifting the deliverances of our intuitive systems themselves.
The psychological element of Barwise and Etchemendy’s “breakdown” model is
perhaps disconfirmed, in that under reflection, subjects do not fragment into a
range of responses as if at random; rather, the NEITHER response becomes more
and more attractive, apparently, the more reflective and the more philosophically
trained a person is.
Our second set of hypotheses, regarding some interpersonal variation in
responses to the paradoxes, also received some confirmation: within our sample,
female subjects did give a somewhat different pattern of responses than male
subjects, and likewise for self-identified theists and self-identified nontheists. As
noted above, such results would need to be replicated and extended before taken
as in any way definitive. Those two variables are surely confounded with other
demographic variables as well, such as personality types, political orientation, and
socioeconomic status, and this initial study has only controlled for a few such pos-
sible confounds.
Regarding the question of philosophical expertise, the basic news here seems
to be basically encouraging for philosophers: greater reflection and greater phil-
osophical training were both factors that shifted subjects’ pattern of responses
away from TRUE or FALSE and toward NEITHER. Proponents of a truth-value
Think ing about the Liar, Fast and S l ow 65
gap approach will perhaps find some evidential ammunition in that result, and
we generally incline that way ourselves. Nonetheless, we think it important to
issue two big caveats in that regard. First, there is still at least some variability in
the even the most expert sample, with, e.g., 15% or so answering FALSE (though
our sample apparently contained vanishingly few expert philosophers willing to
offer the dialethic answer BOTH). Second, the fact that philosophical training and
cognitive reflection produce a convergence in our subjects’ answers certainly does
not entail that they have converged on the truth.
In general, it is important to distinguish between a descriptive psychologi-
cal model and any normative implications one might take to follow from that
model. (This is an important part of why, as we emphasized at the start, we
were not at all recommending using x-phi surveys as any sort of replacement
for formal and philosophical work on the paradox itself, and, in this way, we
are assuming Naess’s methodological mantle [cf. Naess 1938a, sec. 97].) In this
case, at least prima facie, our descriptive results are congruent with a particular
normative analysis: since greater reflection or expertise leads to an increased
tendency toward the “neither” response, that is legitimately some reason to
think that that response is more likely to be the correct one. Although we are
basically sympathetic to that normative inference, we would emphasize that
it is indeed an inference, and as such, it needs to be licensed by some further
assumptions about reflection and philosophical training. Those assumptions
themselves need to be subjected to further empirical scrutiny. We do have a bit
of evidence suggesting that there may be some differences in the way that dif-
ferent people do come to grapple with the paradox, including potentially gender
or religious background. And should such differences turn out to be robust, this
could undermine our sense that these data perhaps provide evidence for the
normative psychological claims. Nonetheless, we would emphasize that the nor-
mative import of our findings is conjectural, at best, pending further research.
Despite their preliminary nature, we hope that our findings serve to highlight
the potential value of further inquiry into the psychological dimension of the
liar paradox and similar problems. Research of the kind we discuss here can help
to shed light old problems and lead to an awareness of new areas of potential
research.
Notes
1. For the remainder of the chapter, we will call this sentence “LIAR.” We will call it such even
to its face. That’s just how we roll.
2. We might compare Maudlin’s insistence on formal analysis with Magdziak’s (2000,
2004) work emphasizing a less rigorous alternative response to the liar paradox, but whose
outcome is the same, i.e., “neither true nor false” is the appropriate response to the liar.
While formal definitions of truth are symptomatic of the liar paradox, Magdziak consid-
ers partial and conditional definitions of truth that seem to be operative in colloquial and
66 Reflections on the Liar
scientific discourse and tend to avoid alethic paradoxes. The trouble is that these less rigor-
ous definitions cannot be generalized to cover the more rigorous formal discourse, and a
philosophical resolution of the liar paradox is not forthcoming.
3. Naess defends this view in a number of publications reprinted in Naess (2005).
4. That Naess never published these results is quite disappointing, indeed. We have to take
whatever evidence we have of the Class- E questionnaires reported in Naess’s 1938a
seriously.
5. There has been some recent discussions recognizing the diversity of processes occurring
in both System 1 and System 2, which have suggested further refinements to the two
(Frankish 2009; Stanovich 2009). Perhaps the range of processes involved in System 2 rea-
soning includes Type 3 processes involved in eliciting System 2 activity, e.g., resource alloca-
tion, conflict resolution, and the ultimate control of behavior, and mediating between the
two systems (Evans 2009). We do not expect any such refinements to make trouble for our
analysis here, however.
6. See also Weinberg et al. (2013) for some recent experimental philosophy results suggestive
of ways in which reflection may fail to overcome intuitive biases by introducing its own
biases.
7. There is a debate in the literature as to whether the response of turning over the true-
antecedent card and the false-consequent card really is, in fact, the normatively cor-
rect response to the task. We suspect that readers interested in the liar paradox will not
be tempted by that view, but those interested in it can look at, e.g., Oaksford and Chater
(1994).
8. For example, one question reads: “If it takes 5 machines 5 minutes to make 5 widgets, how
long would it take 100 machines to make 100 widgets?_____minutes.” The answer “100!”
jumps right to mind, but it is not hard to see that the general rate of production here is that
n machines can make n widgets in 5 minutes—and so 5 is, in fact, the correct answer.
9. See also Pinillos et al. (2011) for an application of the CRT in an experimental philos-
ophy context, to distinguish high-reflection from low-reflection subjects in a “Knobe
effect” task.
10. Mutatis mutandis, it is at least possible that some subjects were bored by the somewhat
arbitrary-seeming tasks of WST and CRT, but whose attention—and reflective powers—
were engaged by the deeply interesting question of LIAR.
11. Responses were collected through the University of Mississippi’s Qualtrics Internet por-
tal. Coinvestigators sought and received IRB approval through the University of Mississippi
(Approval #14x-163).
M-Turk is populated by people who have access to the Internet and who are generally com-
puter literate. We can understand why some might resist the use of this respondent pool,
but we foresaw no reason not to employ the M-Turk population. Beyond that, we placed few
restrictions on those who could respond to the survey questionnaire. Only respondents who
self-identified as being a part of sensitive populations, e.g., children under the age of eigh-
teen, prisoners, etc., were forbidden from submitting responses. Published studies of the
M-Turk population suggest that the sample is in many ways superior to the standard social
psychological sample of college students (Buhrmester, Kwang, and Gosling 2011; Paolacci,
Chandler, and Ipierotis 2010).
12. To ensure that we would be able to compare philosophically trained participants to philo-
sophical novices, in addition to using M-Turk, we solicited participants with philosophical
training through announcements on philosophy-oriented social networking sites, profes-
sional blogs, and listservs. That fewer females participated in the study may have been
partly due to the gender imbalance in the sample from the philosophy profession.
13. For an overview of the paradox, see Ulatowski (2003).
14. In order to preserve their anonymity, no other personal information, such as name, address,
email address, or phone number, was collected from respondents.
15. Questions were counterbalanced to detect carryover effects, but none was discovered.
Our suspicion is that not finding an order effect may have been because the number of
Think ing about the Liar, Fast and S l ow 67
participants was too small—though that suspicion cannot be confirmed without running
another study with more respondents.
16. A Pearson χ2 statistic enables researchers to investigate whether the response distributions
of categorical variables, e.g., “What is your gender?,” “Do you have a Ph.D. in philosophy?,”
differ from one another. It evaluates how likely it is that any observed difference between
the sets arose by chance.
17. Of respondents, only people who selected the minimum, ideal answer, i.e., turning over
the card with the true antecedent, and the card with the false consequent, received a “pass-
ing” grade counting as “correct.” If participants selected any other set of answers, they were
counted as getting the WST “wrong.”
18. We collected fine-grained responses from participants, allowing them to self-identify by
broad religious tradition, e.g., Protestant Christian, Jewish, Muslim, etc. These responses
were combined so that people who self-identified in one of these religious traditions were
counted as theists, and those who answered “none” or “other” were coded as nontheist.
(There were only a small handful of those who answered “other.”)
19. The model was not hierarchical or analyzed stepwise, as we had no reason to order the sup-
posed predictors.
20. These regression results should not be overlooked, since the first results show how WST
passers, BI truthers, or being female increases the log odds of choosing NEITHER over
FALSE (cf. table 3.1) and the second results only show being female increases the log odds
of choosing NEITHER over BOTH (cf. table 3.2). The effect of commitment to bivalence and
logical acumen operationalized by WST is what we would, in some sense, hope to find. The
gender result is perplexing, for it suggests that there may be a real, yet unexplained, effect
of gender on how one responds to LIAR. One might try to explain the effect in terms of the
fact that the philosopher sample was heavily male. But if that were the key factor, one would
expect that philosophical expertise would also end up as a significant predictor. According
to our analysis, it is not. (Statistical aside: Because we employed SPSS to produce the logistic
regression data, we ought to remind the reader that instead of t ratios, it calculates using the
“Wald statistic,” which is a test statistic distributed as a χ2 where df = 1 and its value equals
the square of a t ratio.)
21. Regression analysis suggests that although responses to CON and BI play a role in predict-
ing responses to the LIAR, none of these factors has a sufficiently large R2 value to ground
the claim that they drive the response to LIAR.
22. All participants were asked to read the vignette and then answer what Jones should do.
We call it the “practical” version because it asks participants to choose what Jones should
do, rather than tell us whether the sentence “You will throw me into the water” is true
or false. Respondents were able to choose one of the following answers in the “practical”
case: Jones should throw Smith into the water, Jones should allow Smith to pass, or Jones
should not do anything. We have left aside our findings from the “practical” case for this
chapter because we preferred to focus upon the experiment that elicited response to LIAR
and LIAR-like semantic paradoxes.
23. For this comparison, the Pearson’s χ2 test results do not quite reach the threshold of sig-
nificance (χ2 = 15.446, df = 9, p = .079, ns). We suspected that this was a reflection of the
smaller sample size. The likelihood ratio test is more sensitive; it is equivalent to the stan-
dard Pearson’s χ2 test for large data sets and is preferred for smaller sample sizes.
References
Barnard, R., and J. Ulatowski. 2013. “Truth, Correspondence, and Gender.” Review of Philosophy
and Psychology 4: 621–638.
———. n.d. “Just the Facts? Evidence, Gender, and the Ordinary View of Truth.” University of
Waikato.
Barwise, J., and J. Etchemendy. 1987. The Liar. Palo Alto, CA: CSLI Publications.
68 Reflections on the Liar
Buhrmester, M., T. Kwang, and S. Gosling. 2011. “Amazon’s Mechanical Turk: A New Source of
Inexpensive, Yet High-Quality, Data?” Perspectives on Psychological Science 6 (1): 3–5.
Cacioppo, J., and R. Petty. 1982. “The Need for Cognition.” Journal of Personality and Social
Psychology 42 (1): 116–131.
Cook, R. 2013. Paradoxes. New York: Wiley.
Epley, N., and T. Gilovich. 2006. “The Anchoring-and-Adjustment Heuristic: Why the Adjustments
Are Insufficient.” Psychological Science 17 (4): 311–318.
Evans, J. St. B. T. 1984. “Heuristic and Analytic Processes in Reasoning.” British Journal of
Psychology 75: 451–468.
———. 1996. “Deciding before You Think: Relevance and Reasoning in the Selection Task.”
British Journal of Psychology 87: 223–240.
———. 1998. “Matching Bias in Conditional Reasoning: Do We Understand It after 25 Years?”
Thinking and Reasoning 4: 45–82.
———. 2009. “How Many Dual-Process Theories Do We Need? One, Two, or Many?” In Jonathan
St. B. T. Evans and Keith Frankish, eds., In Two Minds: Dual Processes and Beyond, 33–54.
Oxford: Oxford University Press.
Evans, J. St. B. T., and J. S. Lynch. 1973. “Matching Bias in the Selection Task.” British Journal of
Psychology 64: 391–397.
Feltz, A., and E. Cokely. 2009. “Do Judgments about Freedom and Responsibility Depend on Who
You Are? Personality Differences in Intuitions about Compatibilism and Incompatibilism.”
Cognition and Consciousness 18: 342–350.
Frankish, K. 2009. “Systems and Levels: Dual-System Theories and the Personal-Subpersonal
Distinction.” In Jonathan St. B. T. Evans and Keith Frankish, eds., In Two Minds: Dual
Processes and Beyond, 89–107. Oxford: Oxford University Press.
Frederick, S. 2005. “Cognitive Reflection and Decision Making.” Journal of Economic Perspectives
19 (4): 25–42.
Gilbert, D., R. Tafarodi, and P. Malone. 1993. “You Can’t Not Believe Everything You Read.”
Journal of Personality and Social Psychology 65 (2): 221–233.
Hales, S. 2000. “The Problem of Intuition.” American Philosophical Quarterly 37 (2): 135–147.
——— . 2004. “Intuition, Revelation, and Relativism.” International Journal of Philosophical
Studies 12 (3): 271–295.
———. 2009. “Moral Relativism and Evolutionary Psychology.” Synthese 166 (2): 431–447.
———. 2012. “The Faculty of Intuition.” Analytic Philosophy 53 (2): 180–207.
Kahneman, D. 2003. “Maps of Bounded Rationality: Psychology for Behavioral Economics.”
American Economic Review 93 (3): 1449–1475.
Kahneman, D., and A. Tversky. 1972. “Subjective Probability: A Judgment of Representativeness.”
Cognitive Psychology 3: 430–454.
———. 1996. “On the Reality of Cognitive Illusions.” Psychological Review 103: 582–591.
Ludwig, K. 2007. “The Epistemology of Thought Experiments: First Person versus Third Person
Approaches.” Midwest Studies in Philosophy 31 (1): 128–159.
Machery, E. 2009. Doing without Concepts. Oxford: Oxford University Press.
Machery, E., R. Mallon, S. Nichols, and S. Stich. 2004. “Semantics, Cross-Cultural Style.”
Cognition 92 (3): B1–B12.
Magdziak, M. 2000. “Na antynomii semantycznych w języku potocznym.” Kwartalnik Filozoficzny
28 (4): 137–142.
———. 2004. “Prawda w jezyku potocznym i paradoks klamcy.” Kwartalnik Filozoficzny 32
(2): 133–143.
Maudlin, T. 2004. Truth and Paradox: Solving the Riddles. Oxford: Oxford University Press.
Mercier, H., and D. Sperber. 2011. “Why Do Humans Reason? Arguments for an Argumentative
Theory.” Behavioral and Brain Sciences 34: 57–111.
Naess, A. 1937–1938. “Zum Vortrag von Kokoszynska über Einheitswissenschaft.” Erkenntnis
7: 382–384.
———. 1938a. “ ‘Truth’ as Conceived by Those Who Are Not Professional Philosophers.” Jacob
Dybwad Universitetsforlaget, Oslo.
Think ing about the Liar, Fast and S l ow 69
The guiding question for this chapter is what modal system with a sentential
truth-operator—if any—would capture the notion of truth in a semantically
closed natural language. Or, more modestly, what modal system with a senten-
tial truth-operator—if any—would capture the notion of truth in the proposi-
tional fragment of a semantically closed natural language. As a starting point
toward an answer, I consider the basic features of liar sentences that, combined,
create the liar paradox and that threaten the consistency of the notion of truth.
Through distilling the structure from these features, and separating pragmatic
from nonpragmatic elements, I obtain the building blocks for a possible-world
semantics that leads to the normal modal logic KT4M (or S4M or S4.1). This
logic, which so far has gained little attention, makes it possible to represent
truth in the propositional fragment of natural language as a coherent notion.
It also provides some additional insight into the notion of truth, such as that,
pace deflationism, truth does have content and that this content is bivalence;
moreover that it appears that the sets of true sentences and of bivalent sen-
tences are semideterminable on the propositional fragment of natural language.
Accordingly I offer KT4M as the logic of semantic modalities, with truth as
semantic necessity and bivalence as semantic noncontingency. What I suggest
is not an axiomatic theory of truth.
The route from the Liar to KT4M includes the following way-posts. The para-
doxical liar sentences lie at the intersection of three sentential features: first their
ascription of a semantic value, second their self-reference, and third their predi-
cating something that is incompatible with truth. Individually and in pairs any of
these features are unproblematic, and it is an indication of the richness of natu-
ral languages that they include all three. Things go awry when the three features
are exemplified in the same sentence (or in a plurality of sentences related by
anaphora or a successor function).
71
72 Reflections on the Liar
Etchemendy’s ‘simply false’ and ‘simply true’, but instead of a blunt ambiguity,
I suggest a shift in conceptual salience. The chapter adopts McGee’s approach
of treating truth as a sentential operator, but without giving rise to a revenge
problem—or so I hope. Like van Fraassen’s supervaluationist fixed-point models
of truth, it separates bivalence and the law of excluded middle, but unlike his the-
ory, it retains compositionality and does not entail a third semantic value. The
chapter also picks up on—and turns around—recent suggestions that provide a
possible-world semantics for truth and treat necessity and truth in similar ways
(Leitgeb 2003; Halbach 2003; Halbach, Leitgeb, and Welch 2005) and shares with
Billon 2011 the treatment of arguments as contexts. Beyond the amalgamation
of all these factors into one theory, there are the features mentioned in the first
two paragraphs.
Many questions that arise for theories of truth generally are not, or hardly,
touched upon in this chapter: these include related semantic paradoxes such
as Curry’s and Yablo’s, the Gödelized Liar (which gets its own paper!), and the
relations to quantified truth-ascriptions and propositional attitude ascriptions.
This is neither to say that answers to these questions cannot be given nor that
they will not be given. Finally, the chapter is deliberately written in such a way
that it is accessible to nonmathematicians and it keeps formal elements to a
minimum.
(a) Which Gestalt one sees, or sees first, may depend on multiple factors, includ-
ing simultaneous or preceding circumstantial factors.
(b) In ordinary circumstances, whether a person sees the one Gestalt (a duck), or
the other (a rabbit), or experiences a Gestalt shift from one to another, the
person cannot be faulted for what they (say they) see. If someone is presented
with a range of sketched animal representations on cards, with a duck-rabbit
card snuck in, and is asked to say for each card what animal it shows and then
move on to the next card, we have no reason to say that someone who says ‘a
rabbit’ or someone who says ‘a duck’ when faced with the duck-rabbit card is
mistaken. Someone saying ‘a rabbit when looked at one way, a duck looked at
another way’ would not be mistaken either.
(c) Likewise, a person cannot be faulted if, seeing one or the other Gestalt, they
make some inferential observations: when you see the duck, you can make the
inferential observation that part of the figure or drawing you see depicts a
beak.
(d) Some people only see one Gestalt and experience no Gestalt shift. In such
cases, one can sometimes prompt a Gestalt shift in the person’s perception,
e.g., by pointing out some features of the alternative Gestalt. One may be
able to make someone who only sees the rabbit see the duck by saying ‘don’t
you see the beak?’
(e) Such prompts do not by themselves warrant that the perceiver ‘catches on’
and sees the alternative Gestalt. Multistable perceptual illusions can be stub-
born, and are so to different degrees. With some illusions, for some people
it remains impossible, or takes very long, to reverse the Gestalt, even if they
are told what the second stable Gestalt is (or depicts). The Necker cube is a
good example. Even better is the spinning dancer.
76 Reflections on the Liar
(f) As a rule, one or more elements in the multistable figure do double duty: one
and the same part of the figure may depict one thing relative to one stable
percept, another relative to the other—e.g., a beak and a pair of ears.
(g) In multistable perceptual phenomena each of their stable percepts is in
some respect deficient—for example, since something has been left out
(color, shading or missing details, etc.); or due to the two-dimensionality of
the representation of something three-dimensional.
(h) Perceptual multistability can be removed, e.g., by adding visual elements.
In the case of the duck-rabbit illusion some feather details could be added.
Here is an example with the mother-father-daughter figure that illustrates
possibilities of the removal of multistability by adding detail:
The alternative Gestalts of such bistable sentences are less easily detected than
in the case of the ambiguity of (TEL). They are best made apparent contextually;
that is, by means of typical contexts in which each will—almost certainly—occur.
The most effective linguistic contexts are arguments. For the designational under-
standing (i), consider the following derivation:
Here, in order to see that and how 2 is derived in this derivation, one needs to
think of 2—at least temporarily—as that which is denoted by the designator in
1. Or take
Here, in order to see that and how 2 is derived in this derivation, one needs to
think of, and thus make use of, 1 as that which is denoted by the designator in 2.
For the ascriptional option, (ii), consider the following derivation.
This is part of an informal version of the Liar. If you succeed in seeing the—pre-
sumed—contradiction, then at least while moving to line 3, you have understood
(L) in line 1, as ascriptional. You have made use of the fact that (L) self-ascribes a
truth value. The liar paradox also (iii) provides us with an example of a conceptual
Gestalt shift within an argument. Consider this informal version of the simple liar
paradox (SLP):
Someone who grasps (SLP) needs to understand ‘(L) is false’ (i) as the denota
tion of the designator of a semantic-value-ascribing sentence for the steps
from 1 to 2 and from 4 to 5. They need to understand ‘(L) is false’ (ii) as
ascribing a truth value for the step from 7 and 8 to 9. A person who entertains
(SLP) would thus experience at least one Gestalt shift somewhere between
lines 1 and 9.
As in the case of perceptual bistability, the incompatibility of (i) and (ii)
is on the side of the person who entertains the sentence, e.g., as part of an
argument. Sentence (L) lends itself to either understanding, but one can-
not simultaneously entertain (L) in both ways. (Again, this potential for
incompatible understandings seems to be a brute fact regarding semantic-
value-self-ascribing sentences.) Again, the incompatibility at issue seems
psychological. It appears to be psychologically impossible for us to simulta-
neously entertain both understandings. The following deliberation may illus-
trate this. Take
Here (L) can be understood (i) as that which is denoted by ‘(L)’ in 1 or (ii) as 1.
Now, if we entertained (i) and (ii) simultaneously, we would have to include the
fact that what is denoted by ‘(L)’ in 1 is 1. Thus we would have instead of (1)
where (L) is understood simultaneously both (i) as what is denoted by ‘(L)’ and
(ii) as 1′. Then we would have to include the fact that what is denoted by ‘(L)’ in 1′
is 1′. Thus we would have
(a) Whether you understand (L) (i) as designational or (ii) as ascriptional when
you encounter it is likely to depend on various circumstantial factors, such
as steps in arguments (see above).
80 Reflections on the Liar
(b) We would neither fault someone who understands (i), nor someone who
understands (ii), nor someone who states that (L) can be understood both
as (i) and as (ii). This can be seen from our acceptance of the arguments that
illustrate each of these cases.
(c) Someone would not be faulted if they inferred ‘(L) is false’ from ‘(L) is true’,
using (i). Someone would not be faulted either, if they inferred ‘(L) is not
true’ from ‘(L) is false’, using (ii). Nor would someone be faulted for point-
ing out that depending on how they understood the sentence, they could
infer different things.
(d) We may encounter someone who, on their own, managed only to under-
stand (i), or (ii), but can experience the alternative upon prompting.
(e) It can be difficult for someone to shift Gestalt, if they are simply told of the
two options, without context, or if they are simply told that (L) is conceptu-
ally multistable.
(f) The element of the linguistic representation that does double duty is the
whole sentence. Relative to the designational understanding (i), it is what
is denoted by the subject expression of the sentence. Relative to the ascrip-
tional understanding (ii), it is the sentence.
(g) What is the deficiency of (each of) the two understandings?8 Usually, when
we have a sentence, what is denoted by its subject expression and what truth
value is ascribed to it come apart. Not so in semantic-value-self-ascribing
sentences. They are—pragmatically—deficient in that they do not indicate
which of these aspects is at issue.
(h) We can remove the conceptual multistability of content aspect (or saliency)
by adding explicitly whether the focus is on the sentence qua denotation of
the subject expression or qua truth-value ascription.
(i) In order to experience a Gestalt shift from (i) to (ii) or vice versa, someone
would not need any knowledge either of content aspects (or saliency) or of
predication or truth-value ascription; and moreover they may be unaware
that they experienced a Gestalt shift. For instance, anyone who considers
the liar paradox paradoxical experiences a conceptual Gestalt shift. (see
the simple liar paradox (SLP))
(j) As a result of (i), it may be difficult to ascertain in which Gestalt someone
is understanding the sentence, as they may lack the ability to describe, or
even comprehend, what their understanding is.
So, sentences like (L) are conceptually bistable. This kind of salience-based
bistability is present whenever a sentence self-ascribes a truth value. It is a
consequence of the combination of self-reference and truth-value ascription.
Thus we can also have bistability in (T). Since supposition of (T) does not lead
to contradiction, the bistability is harder to detect. Compare the following two
derivations:
G e stalt Shif ts in the Liar 81
This sentence can only be semantically evaluated if either (i) things are as the sen-
tence denoted by the subject expression of (2) (i.e., the sentence ‘s’) says they are,
or (ii) things are not as the sentence ‘s’ says they are. Usually, in the first case (2) is
true, in the second it is false. I call the contexts in which (2) can be semantically
82 Reflections on the Liar
evaluated circumstances. They are (i) the circumstances in which things are as the
sentence designated by the subject expression of (2) say they are, and (ii) the
circumstances in which things are not as the sentence designated by the subject
expression of (2) says they are. The assumption is that a circumstance is of the
second kind precisely if it is not of the first kind, tertium non datur.
Accordingly, it is a necessary condition for the semantic evaluation of the
semantic-value-self-ascribing sentences (T) or (L) that either (i) things are as the
sentence denoted by their subject expression, i.e., the sentence (T) or (L), says
they are, or (ii) things are not as the sentence denoted by their subject expression,
i.e., the sentence (T) or (L), says they are. The contexts in which (T) or (L) can be
evaluated are (i) the circumstances in which things are as the sentence designated
by the subject expression of (T) or (L) says they are, and (ii) the circumstances
in which things are not as the sentence designated by the subject expression of
(T) or (L) says they are. I understand the assignment of a circumstance of one
of the two kinds to the subject expression of sentences like (L), (T), etc., as the
mapping of the subject expression onto a circumstance of one of the two kinds.
This is in line with standard accounts of context sensitivity. Of course, the cir-
cumstances in which things (i) are, or (ii) are not, as the sentence designated by
the subject expression in (L) or (T) says they are, are precisely the circumstances
in which things (i) are, or (ii) are not, as the sentence itself says they are. This is
another manifestation of the self-reference in (L), (T), etc. Therefore, if you prefer
to think of such assignments of circumstances as—possibly random—stipula-
tions regarding the sentences (L), (T), etc., themselves, rather than regarding their
subject expressions, this is fine, too. The difference is immaterial for the formal
representation of truth in sections 4.7 and 4.8. Either way, the assignment satis-
fies a necessary condition for the semantic evaluability of (L) or (T) and thus, if
you like, ‘grounds’ them in some sense. In function (not in kind) you can compare
these assignments of a circumstance to (L) or (T) with the initial assignments of
a semantic value to (L) or (T) in revision theories of truth (Herzberger 1982a and
1982b; Belnap and Gupta 1993).10 Making semantic evaluability possible by fix-
ing the context (or randomly stipulating how things are regarding (L) or (T)) is a
required step that logically precedes the manifestation of assessment sensitivity
in liar sentences to which I turn now.11
is at odds with truth. In this section I describe how the confluence of all three
features endows liar sentences with a specific kind of unsavory assessment
sensitivity.
For this purpose, I introduce four very simple valid argument forms that
produce conclusions (i) from the premise that things are as a sentence ‘p’ says
they are and (ii) from the premise that things aren’t as the sentence ‘p’ says they
are. There is a presupposition that the sentence ‘p’ says something. The name of
the sentence ‘p’ is given in parentheses as (P):
(P) p
I express the relation between sentence and name as ‘(P) says that p’. The four
argument forms are then as follows:
Argument form 1 Argument form 2
1. Things are as (P) says. 1. Things are as (P) says.
2. So (P) is true. 2. So p.
Argument form 3 Argument form 4
1. Things aren’t as (P) says. 1. Things aren’t as (P) says.
2. So (P) is false. 2. So it’s not the case that p.
Argument forms 1 and 3 involve some kind of semantic ascent. Argument forms
2 and 4 involve some kind of disquotation. (I shall use these argument forms,
which I take to be generally accepted, in order to cast a shadow on the bivalence
of the liar sentence.)
Applied to (L), the four argument forms help to bring out the paradoxicality
of (L) in contrast with the nonparadoxicality of (T). I construct four hypothetical
arguments that move from the assumption of a fixed context of the liar sentence
(section 4.4) to a semantic-value ascription to the liar sentence. In the first two
arguments, the context for (L) is the circumstance that things are as (L) says
(context c1), in the last two, the circumstance that things aren’t as (L) says (con-
text c2).12
Argument 1
1. Things are as (L) says. context c1
2. So (L) is true. some kind of semantic ascent
Argument 2
1. Things are as (L) says. context c1
2. So (L) is false. some kind of disquotation
Argument 3
1. Things aren’t as (L) says. context c2
2. So (L) is false. some kind of semantic ascent
84 Reflections on the Liar
Argument 4
1. Things aren’t as (L) says. context c2
2. So it’s not the case that (L) is false. some kind of disquotation
3. So (L) is true. bivalence
The only linguistic expressions at issue at this point are sentences (with the con-
text of evaluation fixed as c1 or c2). And the semantic values at issue are not the
truth and falsehood tout court of these sentences, but are values that function as
constitutive elements of these—stand-alone—values. (This general idea should be
familiar from supervaluationst theories.) Since these values are constitutive of
truth and falsehood tout court, I call them semantic subvalues. We then obtain this
modified account.14
Let me explain the three main expressions of (4) in terms of argument forms 1
to 4: I say that each viewpoint regarding an atomic sentence S houses a legiti-
mate argument with a conclusion that is a semantic-value-ascription to S.15
I call such atomic sentences S input sentences and such arguments viewpoint-
arguments. These arguments are hypothetical deductions in a metalanguage.16
Each viewpoint-argument for S terminates as soon as it reaches a semantic-
value ascription to S. There are four relevant types of viewpoint arguments,
and the argument forms 1 to 4 from above can serve to exemplify these, if we
supplement them with additional uses of the principle of bivalence or seman-
tic ascent where necessary. These arguments can be divided into two kinds, in
line with section 4.3. First, there are semantic-ascent-first viewpoint-arguments.
In these, a statement of the circumstances, (i) or (ii), is followed by the use
of semantic ascent. For (L), arguments 1 and 3 are such viewpoint arguments.
Second, there are disquotation-first viewpoint-arguments. In these, a statement
of the circumstances, (i) or (ii), is followed by the use of disquotation, and then
by the use of further rules or theorems, if necessary. For (L), arguments 2 and 4
would be disquotation-first viewpoint-arguments.
Assessments and semantic subvalues. We can now give a more precise account
of the assessments and semantic subvalues. An assessment is an assessment of an
atomic sentence S at a viewpoint. It is based on the conclusion of a viewpoint-
argument at that viewpoint. Each such conclusion assigns a semantic value to S.
These values are relative to the viewpoint. As such they cannot be the semantic
values of truth or falsehood tout court. Rather, they are the semantic subvalues that
are constitutive for those values. There are exactly two such values: true (i.e., at
a viewpoint) and false (i.e., at a viewpoint). The assessments are thus bivalent.
How viewpoints are related to viewpoints. A viewpoint is defined not just by its
viewpoint arguments, but also by what other viewpoints it has access to. (I only
consider relations in which the context does not change from one viewpoint
to the next.) First, naturally, any viewpoint is such that if someone has taken
it, then they have access to it (reflexivity). Second, some viewpoints are such
that someone who has taken them can, from them, take another viewpoint.
And since once they have taken that other viewpoint they would be able to take
whatever viewpoints that viewpoint can take (if any), such a person would also
have access to those viewpoints (transitivity). There are also viewpoints such
that someone who has taken them cannot take another viewpoint. Such view-
points can be described as satisficer viewpoints. A satisficer is disposed to cease
looking for alternatives once they have found what they were looking for—in
our case a per se irrefutable argument with a conclusion that provides a seman-
tic-value ascription to the input sentence. Nonsatisficers are disposed to continue
looking for an alternative viewpoint. So, with regard to (L), from a non-satisficer
viewpoint someone could have access to arguments 1 and 2, or to arguments
3 and 4, but from a satisficer viewpoint, they would only have access to one
of these arguments each time. Both satisficer and nonsatisficer viewpoints are
86 Reflections on the Liar
Since truth and falsehood (tout court) should not depend on what per se irrefuta-
ble argument we use to infer a sentence’s truth value, I identify truth-regardless
with truth (tout court) and falsehood-regardless with falsehood (tout court) in a
natural language like English. In line with section 4.1, I use the sentential oper-
ator ‘it is true that’ to express truth-regardless and use ‘it is true that (it is) not
(the case that)’ to express falsehood-regardless. (There are clear similarities to
the notions of stable, or nearly stable, truth and falsehood and unstable val-
ues, in revision theories of truth such as Herzberger 1982a; and Belnap and
Gupta 1993.)
So, (L) is assessment sensitive and true-depending, regardless of context. (T) is
assessment insensitive while being constantly true-regardless in one context and
false-regardless in the other. Gestalt shifts occur in disquotation-first arguments
in both cases. (They are required for the recognition that after the use of disquo-
tation, we have a truth-value ascription to the sentence—either when moving
G e stalt Shif ts in the Liar 87
argument 2. The other would house argument 2 and have access to a viewpoint
housing argument 1.20 I call these four viewpoints viewpoints 1, 2, 3, and 4, in
this order. Given c1, someone at one of the two nonsatisficer viewpoints can
make the following further inference from their two viewpoint-arguments:
(L) is true-dependent.
So, with the context of evaluation ‘things are as (L) says’ selected, there are three
conclusions provided by the four possible viewpoints. Each appears to have been
reached by impeccable reasoning. Analogous results are obtained with the con-
text c2, ‘things are not as (L) says’. For contrast, with (T) only different contexts
of evaluation provide different truth values.
The outcome concerning (L) points to a crucial feature of truth. The argu-
ment housed at the satisficer viewpoint 1 concludes that (L) is true. It results
from an irrefutable argument. Nothing suggests the conclusion is come by dis-
honestly. We do not expect someone who produces this kind of argument to
go look for further or alternative arguments. That is, from viewpoint 1, (L) is
true-regardless. Mutatis mutandis the same holds for the satisficer viewpoint
2. From viewpoint 2 (L) is false-regardless. So, with the context fixed, both
the value true-regardless and the value false-regardless of (L) can be rationally
defended on their own. The third possible scenario is that someone takes a
nonsatisficer viewpoint, i.e., a viewpoint that takes an alternative viewpoint,
if there is one. Viewpoints 3 and 4 are examples. In such a scenario, effectively
the overall result is that (L) is true-depending. Since there is nothing irrational
about someone searching for more than one argument that produces a semantic
value for (L), this result, too, can be rationally defended. There are then equally
justifiable viewpoints that result in (L) being true-regardless, false-regardless,
and true-depending. Consequently, it depends on what viewpoint one takes,
whether (L) is assessment sensitive. In other words, it is true-dependent
whether (L) is true-dependent.21
One can show that this result is true-dependent as well. All one needs to con-
cede is that the holder of a rational viewpoint can access and assess (by further
argument) their previously reached results regarding (L) and that truth-regardless
and falsehood-regardless are luminous. Both are points that are generally granted.
Then the following principle holds.
This means that, if we take seriously the fact that the viewpoint-arguments
housed in the various viewpoints are impeccable—and we have no reason not
to, since they are per se irrefutable—the kind of agnosticism it is rational to hold
about the Liar is the following. Not only do we need to suspend judgement w.r.t.
whether liar sentences have truth or falsehood as their semantic value. We also
need to suspend judgement w.r.t whether liar sentences are viewpoint-dependent,
and w.r.t whether it is viewpoint-dependent whether liar sentences are viewpoint-
dependent, and so on. Elsewhere, I call such agnosticism absolute agnosticism
(Bobzien 2010). I propose that the epistemic stance of absolute agnosticism is the
correct one for the Liar.
(‘It is false that S’ and ‘It is true that [it is] not [the case that] S’ are assumed to be
equivalent.) In common understanding, a sentence is bivalent if it is either true or
false. In terms of truth-regardless, a sentence is defined as bivalent if it is either
true-regardless or false-regardless, and that is, if it is not true-depending.
(11) A sentence ‘⌜S⌝’ is bivalent iff either it is true that S or it is true that
[it is] not [the case that] S.
I symbolize the sentential operator ‘it is true that’ with an italicized bold capital
T. Syntactically, T is modeled on the necessity operator □ in normal modal logics
and can be iterated indefinitely. From classical logic I add p, p1, p2, … for atomic
sentences and the connectives ¬, ∧, ∨, → and ↔. I adopt the common method
of describing modal systems with schemata of axioms and theorems and hence-
forth use ‘axiom’ as short for ‘axiom schema’ and ‘theorem’ as short for ‘theo-
rem schema’. A, A1, A2,… are metalinguistic expressions for arbitrary well-formed
formulas.
The results of sections 4.5 and 4.6 suggest that the operator T is governed by
the principles and rules that govern the modal system KT4 (that is, the rules PC,
that all tautologies of propositional calculus are axioms, MP (modus ponens) and N
(Necessitation), and the axioms K, T, 4) plus one additional axiom. The rules MP, N,
and the truth axioms KT ([TA1∧[TA1→TA2]]→TA2), TT (TA→A) and 4T (TA→TTA)
should be uncontroversial. Moves to drop any of them in order to solve the liar
paradox are here disregarded, for obvious reasons. The same holds for dropping
PC. Axiom 4T expresses the luminosity of ‘it is true that’ that was introduced
G e stalt Shif ts in the Liar 91
informally as (7) in section 4.6. Axiom TT expresses (6) from section 4.5. It is the
modal analogue to semantic descent.
For convenience, I define an operator that expresses truth-depending (‘it is not
true that A and it is not false that A’) using the symbol Ŧ.
MŦ ˫ ŦA→ ŦŦA
M ˫ □◊A→◊□A.
Instead of the rule for the necessity operator □ there is an analogous one for the
truth-operator T:
(R4) [VT] For any wff A and for any w ∈ W, V(TA, w) = 1 if for every
w′ ∈ W such that wRw′, V(A, w′) = 1; otherwise V(TA, w) = 0.25
Conditions for the sentential operators ∧, →, ↔, and Ŧ can be derived from the
above in the usual way. In terms of frames, the system KT4M can be character-
ized by the class of frames that are transitive and reflexive and in which every
world can access at least one world that can access only itself. This last condition
is known as finality.26
In line with sections 4.1 to 4.7, one can produce the following model-theo-
retic representation of truth-regardless based on the class of KT4M frames. Its
G e stalt Shif ts in the Liar 93
• Atomic sentences of a natural language like English L (including atomic liar and
truth-teller sentences).
• Standard conditions of value assignments for complex sentences formed with the
operators ¬, ∨, ∧, →, ↔.
• Assessment viewpoints for the sentences. For each atomic sentence, such a
viewpoint houses a per se irrefutable argument with a conclusion that is a
semantic-value ascription to that sentence. Dispositionally, each viewpoint is
either a satisficer viewpoint or a nonsatisficer viewpoint (section 4.5).
• The semantic subvalues true and false.
• Argument-based assessments that assign a semantic subvalue to each sentence
at each viewpoint.
• The subvalues for atomic sentences are obtained as set out in section 4.5.
Subvalues for nonatomic sentences are obtained in accordance with Rules (R2)
to (R4), where (R2) governs natural-language negation, (R3) natural-language
binary inclusive disjunction, and (R4) natural-language sentential-operator
use of ‘true’ and ‘false’ each understood as viewpoint-independent.
• A reflexive, transitive, and final binary accessibility relation of being disposed to take
a viewpoint from a viewpoint, based on section 4.5. (Being disposed to take a
viewpoint is taken to entail being able to take that viewpoint.)
• Truth-regardless, truth-depending, and bivalence as the resultant truth-
modalities of the sentences.
• Circumstances. These ensure that the context of the liar and truth-teller sen-
tences is fixed.
the relation are fleshed out thus. Reflexivity: Each viewpoint can be taken from
itself. Someone who has a viewpoint is understood to be disposed to take it.
Transitivity: Being disposed to take a viewpoint is the same as being disposed
to get to have that viewpoint. Accordingly, one can take every viewpoint that
viewpoint can take. Finality: Every viewpoint is disposed to take a viewpoint
that is only disposed to take itself. That is, by ‘collecting’ viewpoints via the
relation R, from every viewpoint one may start out from, one gets to a view-
point that opens up no further viewpoints. Such a viewpoint is a final viewpoint.
Someone at that viewpoint lacks the disposition to take any other viewpoint.
Given reflexivity, final viewpoints are satisficer viewpoints and vice versa.
Given reflexivity, transitivity, and finality, every nonsatisficer viewpoint is dis-
posed to take a satisficer viewpoint. A frame F encodes pathways of collecting
viewpoints with regard to its atomic sentences p1 to pn. A model M encodes
a possible subvalue assignment over F to the atomic sentences p1 to pn of the
viewpoints of F. The valuated set of points W in a model M encodes a non-
empty set of viewpoints that each comprises an assessment, i.e., subvaluations
for each of its atomic sentences.
Finally, circumstances. In order for liar and truth-teller sentences to be
semantically evaluable, their context needs to be fixed (section 4.4). This
requirement is satisfied in the representation of truth in KT4M via circum-
stances. Note that the conditions for fixing the context of the subject expres-
sions of liar and truth-teller sentences from section 4.5 are identical with the
following conditions: (i) Things are as (L) ((T), etc.) says. (ii) Things aren’t as (L)
((T), etc.) says. I call such conditions circumstance-conjuncts of a sentence. As
it happens, every atomic sentence of LT—whether context sensitive in the man-
ner of (L) and (T) or not—has two such circumstance-conjuncts. The semantic
evaluability of liar and truth-teller sentences is then ensured in KT4MT, if in
every model over every frame each atomic sentence is mapped onto precisely
one of its two circumstance-conjuncts, for example p1 on (i), p2 on (ii), p3 on (ii),
etc. I call the conjunction of those circumstance-conjuncts of a model M the
circumstance of M. In all nontrivial cases there will be a plurality of models with
the same circumstance.28
Validity and consistency of sentences in KT4MT are defined in the standard way.
A sentence A is valid in KT4MT precisely if it is valid on all frames of KT4MT. A
sentence A is consistent in KT4MT precisely if there is a viewpoint w in a model M
of a frame F of KT4MT such that A is true at w.
This concludes the basic representation of truth- regardless and truth-
depending by means of KT4M. Since KT4M is a normal modal logic, the notion
of truth as truth-regardless developed in sections 4.1 to 4.6 and represented
as semantic necessity, symbolized by T, is coherent within a classical, bivalent
logic.
G e stalt Shif ts in the Liar 95
What about the liar property? By this I mean the property commonly expressed,
with predicative semantic-value ascription, in a biconditional, sometimes relative
to a language L, as a semantic relation as follows:
(For the version with ‘untrue’ for ‘false’ see section 4.11.) The modal syntax of
KT4M would allow us to express the liar sentence (L) both as L and as T¬L. (I do
not consider the round brackets in (L) as part of the name of the liar sentence
and omit them in modal formalizations, to prevent readers from inadvertently
parsing the formulas as predicate formulas.) Also, trivially, biconditionals can be
expressed in KT4M. So in principle the modal correlate to the liar biconditional
(12) could be expressed relative to LT (with LT being the semantically closed prop-
ositional fragment of a natural language L like English with atomic liar and truth-
teller sentences) as
(LP) L ↔T¬L
So much for syntax. What about the property itself that is usually meant to be
expressed by the biconditional (12)? The short answer is that the liar property is
already accounted for in KT4M. More specifically, it is accounted for by the combi-
nation of theorem MŦ with the interpretation of the operator Ŧ as true-depending
and a possible-worlds semantics that incorporates both context sensitivity and
assessment sensitivity of semantic-value-self-ascribing sentences. Thus there is
no need for adding to the logic of truth KT4M a biconditional like (LP) to account
for the liar property. In fact, it would be misguided, if one were to add such a bicon-
ditional. I explain why this is so.
First, I do not deny that there is a liar property of a kind such that it might
be tempting to express it in ordinary language versions of (12). Rather, I main-
tain that this property is a pragmatic feature of the liar sentence. Now, prag-
matic features of sentences are not usually features a logic is meant to represent.
Accordingly, KT4M does not represent it: There is neither a logical nor a semantic
relation that sanctions the moves from L to T¬L and from T¬L to L.31 What is the
pragmatic feature that is the liar property? It is the bistability of (L). It is the fact
that an individual who entertains (L) can experience a shift from understanding
it in one way to understanding it in another way (where ‘understanding’ is used
to indicate a pragmatic feature as introduced in section 4.3). This shift is a Gestalt
shift from the designational understanding to the ascriptional understanding, or
vice versa (section 4.3).
How is this feature related to (12) or to (LP)? First, it might seem natural to
express (L), when understood ascriptionally, by ‘T¬L’, and when understood des-
ignationally, as ‘L’. ‘T¬L’ and the atomic ‘L’ would then indicate different ways of
98 Reflections on the Liar
understanding the same sentence, or would express the same sentence from dif-
ferent viewpoints. The sentence expressed by either formula would be the liar
sentence. Just as the two views of the spinning dancer could be taken to repre-
sent the same object, i.e., the same dancer, but from different viewpoints, so ‘T¬L’
and ‘L’ could be used to express the same sentence, but to express it differently.
Assuming a fixed context (circumstance), both the meaning and the extension of
the two expressions ‘T¬L’ and ‘L’ would then be the same.
However, this is not what I suggest. It is true that just as one could say that
the left-spinning dancer is the right-spinning dancer, one could say that T¬L is
L. But just as the expression ‘left-spinning dancer’ is not the same expression as
‘right-spinning dancer’, so ‘T¬L’ is not the same expression as ‘L’ and moreover
they denote different things. So in the logic KT4M, they denote different objects.
Accordingly, the liar property is taken to allow only a prima facie pragmatically
licit transition from one sentence to the other.
Just as there are things we can infer from the left-spinning dancer and from
the right-spinning dancer that are incompatible (you cannot spin left and right at
the same time), so there are things we can infer from the-Liar-qua-understood-
one-way that are incompatible with the-Liar-qua-understood-the-other-way, for
example we can infer T¬L from T¬L but not from L, and can infer L from L but not
from T¬L. Thus pragmatic factors are the reason why incompatible things can be
derived from ‘T¬L’ and from ‘L’.
Now, of course, you are welcome (i) to add pragmatic information to a logical
system, and then (ii) to supplement the rules of that system with whatever prag-
matic restrictions are consequently required to keep it consistent. What matters,
if you do so, is that you are aware of what you are actually doing. For illustration,
I return to the ambiguous sentence (TEL) ‘I saw the woman with the telescope’
from section 4.3. Imagine that, in a system of first-order logic, first you add this
ambiguous sentence to your stack of sentences and then you supplement your
logic with a restrictive rule, to prevent any inconsistencies from being derivable
as a result of your adding this sentence.
Let us add more detail. Take (TEL) and first-order logic (FOL). Start with add-
ing the sentences (TEL-I) for the reading (i) of the sentence (the speaker had the
telescope) (TEL-her) for the reading (ii) (the woman had the telescope). Introduce
the occurrence of a Gestalt shift from one reading to the other during an infer-
ence as (TEL-I)→(TEL-her) and (TEL-her)→(TEL-I) respectively. Remember that
the Gestalt shifts are not in the sentences but in the readers. So, it is at this point
that you introduce a pragmatic element into your logic. Next add a biconditional
combining the two possible shifts between readings of the sentence and call this
biconditional the Telescope property (TELP).
This will allow you to derive inconsistencies in your language. Next, add a
restrictive rule to the logic to prevent any inconsistencies. For instance, you
could add the rule (RTel) that whenever a Gestalt shift occurs, nothing derived
before the Gestalt shift using reading (i) can be employed to derive anything
once you have shifted to reading (ii), and vice versa. Use the following tel-
escope argument as example. (Context: you and the woman are 11 meters
apart.)
1. (TEL-I) assumption 1
2. I use the telescope 1
3. (TEL-I)→(TEL-her) Gestalt shift (discharge
assumption 1)
4. (TEL-her) 1, 3, modus ponens
5. She uses the telescope 4
6. Two people more than 10 meters apart Telescope Theorem
cannot use the same telescope at the same
time
7. ¬[I use the telescope] 4, 6
8. I use the telescope ∧¬[I use the telescope] 2, 7, ∧-introduction
Given rule (RTel), you cannot use 2 in step 7 to 8, since it was derived before
the Gestalt shift. You could also have a fancier rule, say rule (RTel′), that allows
you, after a second Gestalt shift, to access what you had derived before the first
Gestalt shift, and after a third Gestalt shift, to access what you had derived after
the first and before the second Gestalt shift, and so forth. The general method
used should be clear. However, by most logicians, what you are doing would be
regarded as pretty idiotic. Mostly, logicians prefer the method of disambiguating
a sentence before they add it to their logic, and then adding both resulting sen-
tences, and hey pronto.
Back to the liar property (LP). What I suggest is that adding the liar property
as (LP) to KT4M is methodologically similar—and similarly misguided—to add-
ing the ambiguous sentence (TEL) to first-order logic. The main difference from
the telescope example is that where (TEL) is ambiguous and allows two readings,
(L) is bistable for reasons of salience and allows two understandings. I go through
the same steps. The biconditional (LP) would express a pragmatic feature. Each of
its conditionals (LP→) and (LP←) is used to indicate the locus of a Gestalt shift,
one for each direction.
If you add this pragmatic relation to your logic, you will be able to derive incon-
sistencies. Let me illustrate this also. As example, I use the following modal rep-
resentation of a simple liar argument. Applications of the conditionals (LP→) and
(LP←) of the liar property (line 8) are interpreted as representations of Gestalt
shifts.
100 Reflections on the Liar
Next you add a suitable rule of restriction that reins in the pragmatic ele-
ment you added, e.g., rule (RLiar) that whenever a Gestalt shift occurs, nothing
derived before the Gestalt shift using the designational understanding can be
employed to derive anything once you have shifted to the ascriptional under-
standing, and vice versa. Alternatively, you can add a fancier rule (RLiar′) mod-
elled on (RTel′). Both (RLiar) and (RLiar′) are rules that require you to revise
your set of assumptions and/or derived formulas each time a Gestalt shift
occurs, restricting you to those that you obtained with the present Gestalt (or
at your present viewpoint). They are rules of the kind revision theories of truth
might employ. With such a revision-theoretical restriction in place, line 4 can no
longer be used to derive 7.32
Now for the punch line. With the logic KT4M, the introduction of the pragmat-
ically interpreted liar property (LP) followed by pragmatics-based restrictions on
the logic is as misguided—if not more so—as the introduction of the telescope-
property followed by the rule (RTel) into FOL. To extend the analogy, it is like
introducing the telescope-property (TP) together with rule (RTel) into a logic that
already contains the disambiguated sentences.
On the assumption that the two conditionals of the liar-property biconditional
chronicle the occurrence of Gestalt shifts, and thus express a pragmatic relation
that involves a speaker/listener, an appropriate logic for truth would be one in
which (a) the inconsistency-introducing pragmatic elements have been eliminated
and (b) any relevant informational content of the liar sentence is retained—just
as in FOL (a) the ambiguities of sentences are eliminated by disambiguation and
(b) the disambiguated sentences have been added.
Now, if we feed the effects of bistability on the liar sentence as viewpoint sen-
sitivity into the possible-worlds semantics, KT4M does exactly that. In particular,
it is axiom M that, added to KT4, preempts the need for revision-theoretical rules
(or for an N-restriction rule), in the way in which disambiguation preempts the
need for a rule like (RTel).
As regards (a), we have seen above that there is no inconsistency in KT4M,
and in particular truth is not an inconsistent notion. As regards (b): what is the
G e stalt Shif ts in the Liar 101
One rationale behind this is the following. Often, the paradoxicality of the
simple Liar (L) is removed by the introduction of a third semantic status, only
to return with a vengeance, armored in the garb of the strengthened liar. (13)
preempts the revenge in such simple(ton) solutions. So, why not start where
the real problem lies? The reason my approach starts with the simple Liar is
that this made it easier to bring out what I suggest is the structure of the
notion of truth.
102 Reflections on the Liar
Even so, since, in some sense, the proposed theory offers the possibility
of three semantic statuses for sentences, I should say something about the
strengthened liar. The traditional way of introducing revenge would be by
means of a sentence that is a disjunction of the two values that are not the
designated value of truth.34 So let us do that for the representation of truth as
truth-regardless:
(LR') T¬LR′∨ŦLR′
(LR) ¬TLR
(LRP) LR ↔¬TLR
with the two conditionals of (LRP), denoted as (LRP→) and (LRP←), taken to indi-
cate Gestalt shifts. Adding the pragmatic property (LRP) as a theorem to KT4M,
you can then express a revenge paradox, for example
But here is not the place to do this.37 Equally, truth represented by KT4MT
can accommodate multiple-sentence liar paradoxes, propositional versions of
Kripke’s Nixon/Dean example, Curry Sentences, and open sentences such as the
No-no-Paradox. Again, for reasons of space, these cases cannot be discussed. The
procedure is always based on bistability and is similar to those for (T), (L), and
(LR), or combinations of these.
Truth and falsehood. In C-KT4M, neither Tp nor T¬p is valid of any atomic sentence
p. This is as it should be. The truth-regardless (truth tout court) and falsehood-
regardless (falsehood tout court) of atomic sentences are not structural properties
of the notion of truth. Even if we know that a sentence is nonparadoxical, we can-
not infer its truth or its falsehood from that. Accordingly, even in the C-KT4M,
truth and falsehood can only be defined relative to a model: a sentence A is true-
regardless (TA) in a model M precisely if it is bivalent and is valid-in-M (Result 6).
From the definition of true-regardless, together with the notions of negation and
circumstance-conjuncts, it results that a sentence A is false-regardless (T¬A) in a
model M precisely if it is bivalent and its negation is valid-in-M (Result 7). This is in
line with section 4.5 and so is also as it should be.
Assume that it is the mark of a deflationary theory of truth that truth does
not denote a real property of sentences (or propositions); or that there is no such
thing as the nature of truth; or that saying that ‘snow is white’ is equivalent to ‘it
is true that snow is white’ or to ‘ “snow is white” is true’ exhausts what one can
meaningfully say about the truth of ‘snow is white’—and so for all sentences.
Then we would expect deflationists to hold that TRIV is the modal system that
captures the structure of truth and provides the logic of semantic modalities.39
This would be a modal way of expressing that the notion of truth has no content.
The theory I propose is not a deflationary theory. It endows the notion of truth
with content (as presumably would any modal logic stronger than TRIV). This
content is entirely structural, but it is content nonetheless. Let me expound this
content of truth in two steps.
First, the content of truth can be said to be bivalence. As was mentioned in sec-
tion 4.7, the negation of A’s being true-depending, or ŦA, is the bivalence of A. In
modal terms, ¬ŦA expresses that either it is true that A or it is false that A. And
this is precisely that A is bivalent. Making use of the interdefinability of T and Ŧ,
one obtains the following account of truth:
(15) TA↔¬ŦA∧A
Since KT4M can be defined with the basic operator Ŧ instead of T we can also
express this as a definition of truth:
We can then say that every sentence TA has as its content bivalence, but not nec-
essarily every sentence A does.41
Second, the logical structure of bivalence qua content of truth is defined by the
modal logic KT4M. What kind of structure does KT4M equip bivalence (and truth)
with? The simplest way to understand this structure is as representing a kind of
semideterminability, or more precisely proper semideterminability, with respect
to bivalence (and truth) over classical logic as applied to the atomic sentences
of the natural language fragment with truth predicate LT, where these atomic
sentences include liar and truth-teller sentences. By proper semideterminability
I mean the following:
A class of questions is properly semideterminable if and only if there is a pro-
cedure that comes to a halt and says yes if the answer is positive, but there is no
procedure that comes to a halt and says no if the answer is negative.42
Rather than setting out a full mechanism that results in such proper semi-
determinability for bivalence and truth, I offer an informal description that
relies on the partially interpreted possible-worlds semantics for C-KT4M from
section 4.9.
106 Reflections on the Liar
Acknowledgments
I thank Walter Dean, Øystein Linnebo, Stephen Menn, Beau Madison Mount, Ian
Phillips, Ian Rumfitt, Josh Schechter, and Bruno Whittle for helpful comments on
parts of what has become this chapter, Simona Aimar for reading an early draft
and for asking penetrating questions, Bradley Armour-Garb for notes on two
drafts and for his saintly patience and encouragement as editor, and in particular
Kit Fine and Volker Halbach, whose avid discussions of earlier versions of what
has become this chapter have benefited it enormously. I do not thank the anony-
mous referee from Oxford University Press.
Notes
1. Why do I not use (Luntrue) ‘This sentence is untrue’ as the liar sentence? The answer will
become clear later. (Luntrue) is discussed in 4.11. Why do I not use a sentence like (L101)
‘The sentence written on the board in room 101 expresses a false statement’ as an example
of a liar sentence? The answer is brevity. Throughout, you should be able to replace (L) by
(L101), and (T) by a corresponding (T101) without any philosophically significant change
in what I say.
2. I agree with Halbach 2003, p. 79, that “there is hardly any essential syntactical difference
in English between ‘true’ and ‘necessary’, that is, replacing a used occurrence of ‘true’
by ‘necessary’ in an English sentence usually will yield again a sentence and vice versa.”
Where Halbach 2003 and Halbach, Leitgeb, and Welch 2005 suggest treating necessity as a
G e stalt Shif ts in the Liar 109
predicate, I suggest treating truth as an operator that attaches to what is said in a sentence. I
give a rough idea how this works. The sentence ‘the sentence “snow is white” is true’ is short
for ‘the sentence “snow is white” says that snow is white and it is true that snow is white.’
The sentence “Fermat’s last Theorem is true” is short for ‘Fermat’s Last Theorem says that
no three positive integers … and it is true that no three positive integers… .’ The sentence
‘Everything the pope says is true’ is short for ‘for all sayable things p (where p is a sentential
variable), if the pope says (that) p, then it is true that p.’ Or alternatively conjunctively ‘If the
pope says that grass is green, it is true that grass is green and if the pope says that snow is
white, it is true that snow is white, …’ More complex sentences and propositional attitude
ascriptions are dealt with in the same general manner.
3. http://en.wikipedia.org/wiki/Spinning_Dancer.
4. More accurately, this would be the potential for shifts in the perceptual salience of linguistic
stimuli.
5. In this chapter I do not further specify this pragmatic element of attention or focus and
perceptual salience of linguistic stimuli. The details are in part empirical. (For an over-
view of some recent research see Summerfield and Egner 2009.) Let it suffice to say that
what I have in mind is neither a difference in force nor one in presuppositions nor one in
implicatures.
6. There is no formal contradiction here. Rather, it results from the acceptance of the Principle
of Bivalence. If instead of (L) one uses (Luntrue) ‘This sentence is untrue’, one gets a formal
contradiction. A Gestalt shift is still required.
7. See Ryle (1951) for a related infinite iteration.
8. It is not the lack of semantic evaluability due to ungroundedness, for which see section 4.4.
9. Of course, one may propose that we solve the Liar by prohibiting such salience-based
Gestalt shifts. However, first, it seems entirely ad hoc, if one restricts such prohibition
to liar sentences. Second, either one has to grant pragmatics an unusual impact on infer-
ences, or one has to hold that there is a difference in linguistic content between ascrip-
tional and designational understanding. Neither seems a good idea. (Some more on this
in section 4.10.)
10. Needless to say, such circumstances are not what Kripke calls ‘specifiable circumstances’ for
his meaning criterion (Kripke 1975, p. 699).
11. This kind of context sensitivity of (L) and (T) is quite unlike those suggested by Parsons
(1974) and Glanzberg (2004).
12. Since what (L) says is precisely what the sentence denoted by the subject expression of
(L) says, for convenience I use the shorter formulation.
13. Cf. e.g. MacFarlane (2014).
14. This assessment sensitivity is unsavoury, since whether something is true or false should
not depend on which (compelling) argument one uses to obtain a truth-value ascription to
it. I am not at all concerned with assessment sensitivity that is not of this kind and suspend
judgment on theories that claim that assessment sensitivity is a semantic fact for epistemic
modals, expressions of personal taste and the like.
15. In his defense of the assessment sensitivity of liar sentences, Alexandre Billon suggests
that, in the case of sentence types, viewpoints correspond to arguments (Billon 2011). So, in
my view, he gets it almost right. But why arguments? Well, empirical data, perception, and
intuition seem not very promising starting points for liar and truth-teller sentences.
16. The viewpoint arguments are not part of the natural language fragment LT.
17. S is true-depending whenever it is neither true-regardless nor false-regardless. Its nega-
tion not-S is true-depending whenever it is neither true-regardless nor false-regardless.
Whenever there is a viewpoint from which S is false-regardless, S is not true-regardless and
not-S is not false-regardless. Whenever there is a viewpoint from which S is true-regardless,
S is not false-regardless and not-S is not true-regardless. So, whenever there are both a view-
point from which S is true-regardless and a viewpoint from which S is false-regardless, there
are also both a viewpoint from which not-S is true-regardless and a viewpoint from which
not-S is false-regardless.
110 Reflections on the Liar
18. To show (6), suppose: ‘⌜S⌝’ is true-regardless and it is not the case that ⌜S⌝. Then things are
not as ‘⌜S⌝’ says. Then ‘‘⌜S⌝’ is false’ is the conclusion at a viewpoint (see above). Then ‘⌜S⌝’ has
the subvalue false at that viewpoint. Then ‘⌜S⌝’ is not true-regardless.
19. Thus, among rational viewpoints to hold, there always exist satisficer viewpoints with the
viewpoint-arguments 1 and 2.
20. This is only a partial description of rational viewpoints. There are two options for non-
satisficer viewpoints. The viewpoint they access may in turn be a satisficer viewpoint or a
nonsatisficer viewpoint. So, in the first nonsatisficer case, the viewpoint housing argument
2 could be a satisficer or a nonsatisficer viewpoint, and in the second case, the viewpoint
housing argument 1 could be. Since among the (rational) viewpoints there always is a satisfi-
cer viewpoint, each nonsatisficer viewpoint either directly or indirectly accesses a satisficer
viewpoint.
21. Objection: But if one can take a viewpoint which accesses two arguments that come to
conflicting conclusions, then surely each of these arguments alone is no longer defensible.
Reply: This is not so. Given that the context is fixed, the arguments are each flawless. The
facts that, looked at differently, a different conclusion results and that, looked at in two
ways, two conflicting conclusions result do not change the fact that, as it is, the one argu-
ment is legitimate and irrefutable and nobody can be faulted for adhering to its conclusion,
no matter what. Arguing otherwise is basically simply denying the paradoxicality of the Liar.
At the minimum, a flaw in the argument would need to be identified. The fact that, looked
at differently, a different conclusion results is not an indication that the argument is flawed.
Instead, it may be an indication that something is not quite right with (L).
22. Sections 4.12 to 4.14 specify further how truth in natural language is represented.
23. E.g. Bobzien (2015), pp. 85–85.
24. KT4M is also known as S4M and as S4.1. For some historical background see e.g. Hughes
and Cresswell (1996). For its decidability see Segerberg (1968).
25. For a rebuttal of the—misguided—objection that we do not need a possible-world seman-
tics for truth, since there is a truth evaluation ‘built into’ possible worlds see Leitgeb (2003,
p. 129).
26. For soundness and completeness proofs for KT4M see, e.g., Hughes and Cresswell (1996).
Finality has been shunned in theories of truth, though it is discussed in a paper whose
authors, like myself, believe that truth and necessity should be approached in similar
ways—if they do it the other way about than I (Halbach, Leitgeb, and Welch 2005).
27. For the relation between truth and truth simpliciter see section 4.5. My subvalues true
and false would correspond to Leitgeb’s internal truth, and the truth expressed by the
truth-operator to his external truth (Leitgeb 2003, p. 129), though otherwise what I offer is
very dissimilar to his approach.
28. Circumstances can perhaps be elucidated by comparison with possible-world semantics that
have designated actual worlds. Circumstances are somewhat similar to whatever it is regard-
ing a model that makes the actual world the actual world in that model (perhaps truth con-
ditions of the actual world of M in such a semantics). Yet, in the representation of truth by
KT4M, there are no designated actual worlds.
29. Plainly, I am not talking probabilistic logic here.
30. I disregard the relation between the liar property and diagonalization. How the Diagonal
Lemma and Incompleteness Theorem(s) are related to KT4M is the topic of a separate
paper. For present purposes, it suffices to take ‘(L)’ to be the name of (L). See also footnotes
2 and 37.
31. KT4M does cover certain logical relations between L and T, namely those expressible in
propositional modal logic. These include TL→L, T¬L→¬L, and L→¬T¬L, via axiom T and, if
L has been proved, the move from L to TL via N. Because of ŦL (which itself can neither be
proved nor disproved), this move can never happen, though.
32. As an alternative to revision-theoretical rules, if you prefer to incorporate the pragmatically
motivated restriction directly into the axiomatic system, you can place a restriction on a
rule of KT4M⊕(LP). A suitable restriction would be the metarule that you cannot apply the
G e stalt Shif ts in the Liar 111
rule of necessitation N to anything you derive after (LP) has introduced a Gestalt shift. In
KT4MT, N bestows truth-regardless on any formula that has been derived in the system. For
the above reasons, once we have a Gestalt shift, such a move to truth-regardless is no longer
justifiable. Thus Gestalt shifts are accommodated by putting a restriction on a rule of KT4M
for the system KT4M⊕(LP). In the above argument, this restriction on N bans the step from
5 to 6. As a result of supplementing KT4M with (LP) and an (LP)-based restriction on N, you
obtain a logic that is extremely close both to the Kripke-Feferman system (KF) and to the
formal theory Herzberger outlines as emerging from his Naive Semantics (Feferman 1984,
1991; Herzberger 1982a, p. 493). One crucial difference in the suggested theory is axiom M.
33. Thus you can derive, e.g., L∨¬L, T¬L∨¬T¬L. By contrast, from the hypothesis L alone you
can derive nothing, given ŦL (or ŦLR) and axiom M. On T¬L you can apply any axioms or
rules of C-KT4M. Since we cannot prove the semantic value of T¬L, due to ŦL, Mirror, and
axiom M, we can detach no consequent of any conditional in which T¬L is the antecedent (or
from a set of sentences Δ that includes it) and for the same reason we cannot apply N. On
the other hand, from the hypothesis of an assessment-insensitive p you can derive p (Tp, p),
and from the hypothesis of T¬p you can derive ¬p.
34. Revenge arguments tend to be tailor-made to each liar solution. (See, e.g., Beall 2007 for a
whole spectrum of vengeance.) Here I only show that the present theory accommodates the
standard way of introducing revenge.
35. The proof is trivial and is left to the reader. (Hint: use De Morgan, contraposition, ∧-
elimination, def. ↔, def. Ŧ.)
36. Arguments analogous to Arguments 1 to 4 can be constructed, with the context fixed, in
the KT4MT semantics, by means of a circumstance-conjunct. With either circumstance-
conjunct, there are viewpoints at which it can be inferred that (LR) is not true, and others at
which it can be inferred that (LR) is not false.
37. As is obvious from this chapter, I do not believe that the so-called naïve theory of truth is
our theory of truth. So, with regard to diagonalization, it would seem entirely reasonable
to remove Liar sentences from the relevant language (e.g., PA with truth predicate) for the
reasons (i) that it cannot be established that such sentences are semantically evaluable and
(ii) that Tarski biconditionals do not extend to Liar sentences (see section 14). However, it
seems more satisfying to show that liar paradoxes that are based on the diagonal lemma
to prove the inconsistency of some λ, too, make use of the bistability of λ, and that λ is
no less assessment-sensitive than our (L) and (LR). This is possible. The argument leaves
the strengthened and extended diagonal lemma intact. The problem rather lies in the way
the Tarski biconditional for the liar sentence is employed in the paradox. To those who at
this point bring up Montague’s Theorem and the paradox of the knower, I point out that
they are changing the subject. The predicate introduced by diagonalization in the proof of
Montague’s Theorem is different from the truth predicate in the sense that we expect fewer
sentences to be derivable. Perhaps what is knowable or informally provable is true, but the
reverse does not hold. Whereas a modal logic for truth should be normal and complete, the
same is not obvious for a knowability or informal-provability operator. (See also footnote 2.)
38. The proposed theory shares the feature that the principle of bivalence and LEM come apart
with some other theories of truth, such as supervaluationist ones (e.g., van Fraassen 1968
and 1970; also Fine 1975).
39. TRIV is any normal modal system that contains axiom K and in which □A↔A is a theorem.
40. Mutatis mutandis the same holds for falsehood. For a sentence ‘S’ to be false means that ‘S’
is bivalent and that it’s not the case that S.
T¬A =def ¬ŦA ∧ ¬A
42. Some will recognize this as the informal definition of nondecidable semidecidability or
recursive enumerability. Here I am interested solely in this informal nonmathematical defi-
nition, hence the different term.
43. Thus, unlike Liggins (2014), who suggests that the semantic-ascent half of the T-schema
should be renounced, all I say is that judgment needs to be suspended regarding its
validity.
44.
1. ¬TA↔T¬A
2. T¬A→¬A 1, axiom T
3. ¬TA→¬A 1, 2, definition of ↔
4. A→TA 3, contraposition
5. TA→A axiom T
6. A↔TA 4, 5, ∧-introduction, definition of ‘↔’
Having A↔TA as a theorem in the trivial modal fragment, or trivialization, of KT4MT is not
strictly the same as having the T-schema for the relevant sentences A. For this one needs to
add that, for any sentence A we have ‘TA if and only if’ ⌜A⌝ ‘is true’.
References
Barwise, Jon, and Etchemendy, John. 1987. The Liar: An Essay on Truth and Circularity.
New York: Oxford University Press.
Beall, J. C. 2007. The Revenge of the Liar. Oxford: Oxford University Press.
Belnap, Noel, and Anil Gupta. 1993. The Revision Theory of Truth. Cambridge, MA: MIT Press.
Billon, Alexandre. 2011. “My Own Truth: Pathologies of Self-Reference and Relative Truth.” In
Shahid Rahman, Giuseppe Primiero, and Mathieu Marion, eds., The Realism-Antirealism
Debate in the Age of Alternative Logics, 25–45. New York: Springer.
Bobzien, Susanne. 2015. “Columnar Higher-Order Vagueness or Vagueness Is Higher-Order
Vagueness.” Aristotelian Society Supplementary Volume 89: 61–87.
—— — . 2010. “Higher- Order Vagueness, Radical Unclarity, and Absolute Agnosticism.”
Philosophers’ Imprint 10 (10): 1–30.
Feferman, Solomon. 1984. “Toward Useful Type-Free Theories, I.” Journal of Symbolic Logic
49: 75–111.
———. 1991. “Reflecting on Incompleteness.” Journal of Symbolic Logic 56: 1–49.
Fine, Kit. 1975. “Vagueness, Truth and Logic.” Synthese 30: 265–300. Reprinted with corrections
in Rosanna Keefe and Peter Smith, eds., Vagueness: A Reader (Cambridge, MA: MIT Press,
1996), 119–50.
Glanzberg, Michael. 2004. “Truth, Reflection, and Hierarchies.” Synthese 142: 289–315.
Halbach, Volker. 2003. “Modalized Disquotationalism.” In V. Halbach and L. Horsten, eds.,
Principle of Truth, 75–102. London: Ontos Verlag.
Halbach, Volker, and Leon Horsten, eds. 2003. Principle of Truth. London: Ontos Verlag.
Halbach, Volker, Hannes Leitgeb, and Phillip Welch. 2005. “Possible Worlds Semantics for
Predicates.” In R. Kahle, ed., Intensionality: Lecture Notes in Logic, 20–41. Wellesley, MA: A
K Peters.
Herzberger, Hans G. 1970. “Paradoxes of Grounding in Semantics.” Journal of Philosophy
17: 145–167.
Herzberger, Hans G. 1982a. “Naive Semantics and the Liar Paradox.” Journal of Philosophy
79: 479–497.
———. 1982b. “Notes on Naive Semantics.” Journal of Philosophical Logic 11: 61–102.
———. 2011. The Tarskian Turn: Deflationism and Axiomatic Truth. Cambridge, MA: MIT Press.
G e stalt Shif ts in the Liar 113
Hughes, G. E., and Max J. Cresswell. 1996. A New Introduction to Modal Logic. London: Routledge.
Kripke, Saul. 1975. “Outline of a Theory of Truth.” Journal of Philosophy 72: 690–716.
Leitgeb, Hannes. 2003. “Metaworlds: A Possible-Worlds Semantics for Truth.” In V. Halbach and
L. Horsten, eds., Principle of Truth, 129–152. London: Ontos Verlag.
Liggins, David. 2014. “Constructive Methodological Deflationism, Dialetheism, and the Liar.”
Analysis 74 (4): 566–574.
MacFarlane John. 2014. Assessment Sensitivity: Relative Truth and Its Applications. Oxford: Oxford
University Press.
McGee, Vann. 1991. Truth, Vagueness, and Paradox. Indianapolis: Hackett.
Parsons, Charles. 1974. “The Liar Paradox.” Journal of Philosophical Logic 3: 381–412.
Ryle, Gilbert. 1951. “Heterology.” Analysis 11: 61–69.
Segerberg, Krister. 1968. “Decidability of S4.1.” Theoria 34: 7–20.
Summerfield, C., and T. Egner. 2009. “Expectation (and Attention) in Visual Cognition.” Trends
in Cognitive Science 13 (9): 403–409.
Tarski, Alfred. 1935. “Der Wahrheitsbegriff in den formalisierten Sprachen.” Studia Philosophica
1: 261–405.
van Fraassen, B. 1968. “Presupposition, Implication, and Self-Reference.” Journal of Philosophy
65: 136–152.
van Fraassen, B. 1970. “Truth and paradoxical consequence”, in Martin, Robert L. (ed.), The
Paradox of the Liar, Atascadero: Ridgeview. 1970: 13–23.
5
Assumption (A) might seem to capture the meaning of ‘true’ at least for nonin-
dexical sentences of English.
But the liar paradox casts doubt on this. For consider the self-referential liar
sentence (L):
(1) ‘(L) is not true’ is true if and only if (L) is not true.
114
Toward R e s olv ing the Liar Parad ox 115
(DI) P default implies Q iff one is entitled to take P to imply Q (and accept
‘if P then Q’) in the absence of reasons to think the implication does
not hold.
(3) ‘‘(L) is not true’ is true’ default implies and is default implied by ‘(L)
is not true’.
(4) ‘(L) is true’ default implies and is default implied by ‘(L) is not true’.
But (4) is not inconsistent. Since the first quoted sentence is the denial of the
second, there are reasons to think that at least one of the default implications
does not hold.
6
116
Microlang uag e s , Vag uene s s , and Parad ox 117
and others say no. I have written a couple of e-books; do they count as books? But
wait, one isn’t published yet. And the one that is published is only about 50 pages
long. Book? Again the answer I get varies. Was my PhD dissertation a book? By
the way, it was “published,” with minor revisions, by the University of Indiana
Linguistics Club. Book? The same book? What about drafts of books that are sit-
ting on my hard drive? Are they books? It takes a few minutes of asking these
questions before I can answer and tell my conversational partner whether I have
written 2 or 3 or 6 or 10 books.
The story is odd in a way, because ‘book’ is one of the first words we English
speakers learn, and it has been with us for a long time. It comes from the old
English ‘boc’, which seemed to apply to any written document. The shared mean-
ing has evolved over the past thousand years to be somewhat narrower than that
(not every written document is a book) and in some ways broader (think e-book)
but even after a millennium of shared usage the meaning is quite open-ended. And
there are elements of the meaning that can change radically on a conversation-by-
conversation basis. If you think the question of what counts as in the range of
‘book’ is problematic, ‘writing a book’ is even harder. You might think an edited
volume is a book, but did I write it?
Far from being the exception, I think this is typical of how things are with the
words we use. Even for well-entrenched words their meanings are open ended and
can change on the fly as we engage different conversational partners. Consider a
word like ‘sport’. Does it include bowling? Mountain climbing? Darts? Chess? Or
consider words like ‘freedom’, ‘justice’, or (less loftily) ‘snack’ and ‘success’. All of
these words have meanings that are underdetermined, and we adjust or modu-
late their meanings on a conversation-by-conversation basis. Their meanings are
dynamic.
These facts seem to fly in the face of the traditional view of language, which is
more or less the following: Languages like Urdu, German, Polish, and Portuguese
are fairly stable abstract systems of communication that are learned (with vary-
ing degrees of success) by human beings. Those humans in turn use the languages
that they have learned to communicate ideas, perform certain tasks (by giving
orders, instructions, etc.), and in some cases as media for artistic expression. It is
often supposed that the better one learns a language the better equipped one is
to successfully communicate, accomplish complex tasks, etc. Sometimes the stan-
dard view uses the metaphor of language as a widely shared common currency
that agents use to communicate, with individual words being the common coins
of the realm. These common coins are also supposed to be more or less fixed. Of
course everyone believes that language undergoes change, but according to the
standard view the pace of change is glacial. There is a long slow gradual evolu-
tion from Old English to Middle English and on to contemporary English. Word
meanings change slowly, and the change is largely uniform across the population
of language users.
118 Reflections on the Liar
‘hexagonal’). We may sharpen the meaning and we may get clearer on what falls
in the range of these predicates (and we may willingly add or subtract individu-
als from the range), but we never completely sharpen the meaning and we never
completely nail down the extension of a predicate. For example, we might agree
that Jones is safely in the range of ‘bald’, but there are still many cases where the
meaning of ‘bald’ isn’t fixed. We haven’t fixed the meaning of ‘bald’ for people
with more hair than Jones, or for people with about the same amount of hair
as Jones but distributed differently, or for people who shave their heads, or for
nonhumans, etc.
Some theorists think that there is a core meaning for a term that is the abso-
lute sense of the term but that we are pragmatically licensed to use the term
loosely. So, for example, ‘bald’ means absolutely bald—not one single hair,1 ‘flat’
means absolutely flat, etc. There are various ways of executing this idea. For
example Laserson (1999) has talked of “pragmatic halos” surrounding the core,
absolute sense of the terms; Recanati (2004) and Wilson and Carston (2007)
have argued that we begin with the absolute meaning and are “pragmatically
coerced” to modulate to less precise meanings. I don’t believe this view is correct.
In Ludlow (2014) I argue that the “absolute” sense of a term (if it even exists) is
not privileged but is simply one modulation among many—there is no core or
privileged modulation.
This isn’t just the case for predicates like ‘bald’ but, I would argue, all predicates,
ranging from predicates for things like ‘person’ and ‘tree’, predicates for abstract
ideas like ‘art’ and ‘freedom’, and predicates for crimes like ‘rape’ and ‘murder’.
You may think that there is a fact of the matter about what these predicates refer
to, but you would be quite mistaken—even in the legal realm the meanings are
not fixed, not by Black’s Law Dictionary, nor by written laws, nor by the intentions
of the lawmakers and founding fathers.2 Any suggestion that there is a fact of the
matter as to what these terms mean in disputed cases is deceptive—it is appeal-
ing to an alleged fact that simply does not obtain. Indeed, I would argue that this
is also the case with mathematical and logical predicates like ‘straight line’ and
‘entailment’. The meanings of all these predicates remain open to some degree
or other, and are sharpened as needed when we make advances in mathematics
and logic.
You might think that underdetermined meanings are defective or inferior and
perhaps things to be avoided, but in my view they can’t be avoided (even in math-
ematical and logical cases), and in any case there is no point in avoiding them
since we reason perfectly well with words having underdetermined meanings.
I will attempt to show how this works and in particular how we can have a formal
semantics of natural language even though we are admitting massive meaning
underdetermination. The received wisdom seems to be that semantics demands
precision and fully determinate meanings. Whatever the merits of precision and
fully determinate meanings, semantics (paradoxical as it may seem) has no need
for them.
120 Reflections on the Liar
Finally, we will see that once we take on the dynamic lexicon and the idea of
microlanguages, we are in a position to rethink the nature of the Liar—at least
as it appears in natural language. In particular, we will see that microlanguages,
for quite independent reasons, have admissibility conditions that preclude liar-
like sentences. So-called liar sentences are not merely anomalous or semantically
defective or ungrounded—they are not even sentences. Or more precisely, while
they have sentence-like forms, they are not sentences in any microlanguage that
we might construct.
The view of the dynamic lexicon I am adopting here folds in a number of con-
cepts, and a couple of them are worth noting here.
a borderline case. Had Secretariat been a centaur we might have had such an argu-
ment. These were rather arguments about whether ‘athlete’ should be defined in
such a way as to admit horses.
We can, however, say that vagueness is a kind of underdetermination—it is
underdetermination along a scalar dimension.3 So for example, we can argue over
whether I am bald, but in that instance the argument seems to be about whether
I am far enough along on the baldness scale to count as bald (here assuming the
border is underdetermined).
cases in that context. Thus a meaning that is sharp in one context with a few
borderline cases could fail to be sharp in a context with lots of borderline
cases.
While there are plenty of circumstances where we might say that (1) and (2) are
true (or if we are unimpressed by the flatness of Kansas, we might say (1) is
false). Braun and Sider (2007) don’t think utterances of these sentences (or
perhaps any sentences) can ever be strictly speaking true (or false). On their
view the problem is that while meanings are underdetermined, semantic objects
are not:
The idea behind their claim is simple enough. We suppose that the semantic val-
ues introduced by a semantic theory will look something like one of the following
options.
But notice that all of these are precise—even the one that introduces a partial
function!6 (To see this, note that for every object in the domain, it is precisely
determined whether it is in the extension of set of ‘flat’, in the antiextension of
‘flat’, or undetermined.)
124 Reflections on the Liar
If semantic values must look like this, then according to Braun and Sider it fol-
lows that almost all utterances of (1) and (2) are not, strictly speaking, true (or
false). They fail to be true (false) because for them to be true (false) there would
have to have precise semantic objects as their meanings. But there are no such
precise semantic objects. Their conclusion:
They opt for an alternative notion according to which sentences of natural lan-
guage are not true, but are “approximately true.”
and Larson and Segal (1995). I happen to be involved in this project as well. And
as Lepore (1983), Higginbotham (1985), and others have stressed, if you want a
theory of meaning that connects language in the world, then truth in a model is
not sufficient—we will need to deploy truth simpliciter. But is it really the case
that the semantic objects in an absolute semantics must be (or even are) precise?
I don’t think so.
To explain this idea, I need to introduce a toy semantic theory within a
truth-theoretic framework as envisioned by Davidson (1967) and developed by
Higginbotham (1985), Larson and Ludlow (1993), and in considerably detail
Larson and Segal (1995). On such semantic theories, instead of introduc-
ing model-theoretic objects as semantic values we will instead offer a truth-
conditional semantics that makes no reference to such objects but still provides a
recursive semantics that carries all of the information we are interested in for the
conduct of semantic investigation.
In the fragment that follows I define a language L using some simple recursive
rules and then provide a semantics for L that shows how we can compute the
semantic value of the whole sentence from the semantic values of the parts. We
specify the syntax of our toy language as follows:10
Syntax of L
i) S → S1 and S2
ii) S → S1 or S2
iii) S → it is not the case that S1
iv) S → NP VP
v) NP → ‘Michael Jordan’, ‘Kansas’
vi) VP → ‘is flat’, ‘leaps’
To illustrate, successive applications of the rules (iv), (v), and (vi) would yield the
following treelike syntactic structure:
S
NP VP
‘Kansas’ ‘is flat’
To make things easier on typesetters we can represent this tree structure in linear
form as follows: [S[NP ‘Kansas’][VP ‘is flat’]], where the syntactic categories of the
tree nodes are represented by bracket subscripts.
For the semantics we introduce axioms for the terminal nodes (that is, for the
lexical items) and additional axioms that tell us for the various syntactic struc-
tures, how we can compute the semantic value of a node in a syntactic tree from
the semantic values of the syntactic elements that it immediately dominates.
In effect, we begin by using the axioms for the semantic values of the words,
126 Reflections on the Liar
and then use the non-terminal axioms when we compute the semantic value of
the whole sentence. To keep things clean and consistent with most of formal
semantics, I will suppose that the semantic value of the sentence is a truth value.
We also introduce the predicate Val, where Val(A, B) says that A is the semantic
value of B.
With that, the axioms for our terminal nodes (words) will be as follows.
(3)
a. Val(x, ‘Jordan’) iff x = Jordan
Val(x, ‘Kansas’) iff x = Kansas
b. Val(x, ‘is flat’) iff x is flat
Val(x, ‘leaps’) iff x leaps
So axiom (3a) tells us that x is the semantic value of ‘Kansas’, just in case x is iden-
tical to Kansas. (3b) tells us that x is a semantic value of ‘flat’ just in case x is flat.
The nonterminal nodes now tell us how to compute the semantic value of
higher-level syntactic structures (like the verb phrase and the sentence as a
whole) from the semantic values of the words. Since this is a very simple language
we only need the following rules to cover all the cases.
(4)
a. Val(T, [S NP VP ]) iff
for some x, Val(x, NP) and Val(x, VP)
b. Val(x, [a β ]) iff Val(x, β) (where α ranges over categories, and β
ranges over categories and lexical items)
c. Val(T, [S S1 ‘and’ S2 ]) iff
Val(T, S1) and Val(T, S2)
d. Val(T, [S S1 ‘or’ S2 ]) iff
either Val(T, S1) or Val(T, S2)
e. Val(T, [S ‘it is not the case that’ S1]) iff
it is not the case that Val(T, S1)
Finally, we need some derivation rules that tell us how to mechanically apply the
axioms. In this case we only need two rules. The first one tells us that if we derive
something of the form α iff β, we are entitled to swap in β for α at any point in the
derivation. The second rule does a simple cleanup to get rid of the variables we
have introduced.
Derivation Rule (SoE)
.... α....
α iff β
Therefore .... β....
Microlang uag e s , Vag uene s s , and Parad ox 127
Derivation Rule (SoI)
ϕ iff for some x, x = α and ....x....
therefore ϕ iff .... α....
Using these axioms and the two derivation rules we can, in a straightforward
way, derive theorems like (2).
These types of fragments can be extended to cover the same data covered by a
traditional model-theoretic semantics. That is to say, if you chose to, you could do
semantics in this way without losing anything, modulo some story about repre-
senting entailment relations.11
But now notice that we did all this without introducing the usual machinery
(utilizing properties, sets, functions) that is supposed to be precise (and determi-
nate). What is interesting about our axiom for ‘flat’ is that it takes advantage of
the underdetermined meaning of ‘flat’ in the metalanguage. Notice that there is
no barrier to using axioms like this in a meaning theory in a way that will deliver
theorems like (2), repeated here.
In fact we just proved that we can derive (2) without the technology of sets, prop-
erties, relations, etc.
Let me make it clear that I am not making a claim against contemporary formal
semantics. There is no question but that formal semantics has been extremely
fruitful and that it has provided many profound discoveries and insights into the
nature of language. In the past 40 years we have seen important discoveries in
generalized quantifier theory, in the way that adjectival modification works, in
the way that pronominal anaphora works. We have seen advances in the theories
of tense, aspect, modality, and the list could go on and on. None of these discov-
eries, however, rests upon the assumption of a precise or determinate metalan-
guage, although, as we will see, perhaps some intractable semantic puzzles (like
vagueness) do rest on that assumption.
Are we home free then? Not yet, because Braun (p.c.) has a rejoinder. We can
set it up with our examples (1) and (2).
Now let’s suppose that I’m conflicted about whether Kansas is flat (I’ve been there
and it is definitely hillier than southern Minnesota). If I say that (1) is neither true
nor false, then I have to say that (2) is neither true, nor false. For if (1) is neither
true nor false, then the left-hand side of (2) is false, while its right-hand side is
128 Reflections on the Liar
neither true nor false. In fact, I will have to reject all instances of the T-schema
(T) in which S is neither true nor false.
Braun is raising a good point, because it raises the question of whether I intend
to give up bivalence, and if so, what the consequences are for instances of the
T-schema. I think it would be interesting to pursue that option (giving up biva-
lence), but I have a different story to tell.
To set up my story I first want to make it clear that I take the semantic theory
to be a theory which computes the semantic values of utterances (or, if you prefer,
tokens)—not sentences construed as abstract objects (this is a distinctively anti-
Kaplanian assumption which I won’t defend here).
On my view, in any microlanguage, admissible utterances having the form
of (1) or (2) must be either true or false. How is this possible given that mean-
ings are underdetermined? Let’s return to my claim that when we engage in
conversation we build microlanguages on the fly. Not only are word meanings
dynamically narrowed or broadened in these cases, but there is also a question
of which words (and hence sentences) make it into the microlanguage. My view
is that no sentence is admissible until it is sharp enough to assert a claim that
is clearly either true or false. Let’s state this as a principle just to be clear.
Now that may seem like a bizarre claim, since, after all, we routinely do say things
that are supposed to have indeterminate truth values. We give examples in class
like ‘Ludlow is bald’, after all.
But there is a difference between talking about sentences with indeterminate
truth values and introducing them into a conversation as vehicles for assertions.
I would say that when we talk about sentences with indeterminate truth values,
we are talking about whether utterances of them should be introduced into the
microlanguage. When we are saying the truth value is indeterminate, we are say-
ing that it is not admissible—it cannot be deployed to make an assertion.
Of course, sometimes people say things like ‘Ludlow is bald’ and the listen-
ers may nod even though they may have previously thought it is indeterminate
whether I am bald. To return to an example I gave in the introduction, Chris
Barker (2002) has made the observation that this often is not merely making a
claim about my baldness, but it is actually an instruction about where to push
the meaning of ‘bald’. It is saying “ ‘Ludlow is bald’ is a safe application of ‘bald’—
hereafter, everyone with less hair than Ludlow (modulo head size) is safely bald.
Microlang uag e s , Vag uene s s , and Parad ox 129
All of this raises interesting questions about how we are to think of Semantic
Theory (in capital letters to indicate that I am talking about the scientific enter-
prise of semantics). We normally think of a semantics like the toy version I gave
above as something that works for a stable language. Maybe the theory is some-
thing that a third party deploys for someone else’s language, or maybe it is used
in the interpretation of our own language, but the question is, what is the nature
of a toy semantics like the one I gave above, given our understanding of microlan-
guages and the microlanguage admissibility constraint?12
The answer is that the toy semantics above is a one-off “passing theory” for
computing the semantic values of expressions in a particular microlanguage L (I
believe, based on personal communication, that this was Davidson’s view as well—
one-off T-theories are used to interpret the passing theories like those introduced
in his “Derangement” paper [1986]). On the assumption that Semantic Theory
is interested in the psychological mechanisms (clearly no longer a Davidsonian
assumption) by which we assign meanings (semantic values) to linguistic forms,
then Semantic Theory is interested in understanding the underlying system that
allows us to build such semantic theories on the fly. Put another way, we can think
of ourselves as building little toy grammars for fragments of “languages” all the
time, complete with semantic theories for interpreting them. Semantic Theory is
the study of how we do this.
Presumably, there must be stable elements to the construction of a passing
theory; otherwise it is difficult to see how it could be possible. It is a safe bet that
most of the nonterminal semantic rules are stable across these shifts (they may
well be stable across all human languages). Thus the real dynamic portion would
be in constructing the terminal (lexical) rules on the fly.
F(a)
Therefore F(a)
If the meaning of F(a) shifts, the argument may not be sound even if F(a)
is true—a kind of equivocation might have taken place. Does this mean
that logic goes out the window? Not at all. For expository purposes let’s
sequentially number occurrences of terms in an argument, so that, for
example, in the argument we just gave the form is the following.
130 Reflections on the Liar
F1(a)
Therefore F2(a)
Again, we are saying that the term is F, and that F1 and F2 are occurrences of the
term F within the argument. To keep our understanding of validity stable, let’s say
that soundness calls for a third constraint in addition to the validity of the argu-
ment and the truth of the premises.
It appears that the argument above is sound if the meaning is stable between
F1 and F2 but also if F2 is a broadening of F1 (a narrowing or a lateral shift in
meaning will not preserve truth). Let’s take a concrete example.
Jones is an athlete1
Therefore Jones is an athlete2
If a shift has taken place between premise and conclusion (between the mean-
ing of ‘athlete1’ and ‘athlete2’) it cannot be a shift that rules out individuals that
were recognized semantic values of ‘athlete1’. If ‘athlete1’ admits race car drivers
and ‘athlete2’ does not, then the argument is not sound. If the second occurrence
broadens the meaning of the term ‘athlete’, the argument is sound.
Broadening meaning doesn’t always ensure soundness. Let’s add some to the
argument we just gave.
This time matters are reversed. Assuming the premise is true, the argument is
sound just in case either ‘athlete2’ preserves the meaning of ‘athlete1’ or it is a
narrowing from ‘athlete1’. Negation isn’t the only environment that dislikes the
broadening of meanings. Consider the following.
For the argument to be sound ‘athlete1’ can be broader than ‘athlete2’, but it
cannot be narrower. ‘healthy2’ can be broader than ‘healthy1’ but not narrower.
Notice that this seems to hold if we reverse the order of the premises as well.
Jones is an athlete1
If Jones is an athlete2 then Jones is healthy1
Therefore Jones is healthy2
Microlang uag e s , Vag uene s s , and Parad ox 131
This constraint needs to be generalized and a proof is called for, but we can see
that it works for familiar cases like arguments involving modus ponens (as above)
and the Aristotelian syllogism. Consider the Barbara schema, for example.
Let’s annotate the terms with polarity markers + and − to indicate positive and
negative occurrences as they are traditionally assigned.
Our Dynamic Lexical Constraint on Soundness tells us that the argument is only
sound if ‘collies’ is stable or narrows, ‘barks’ is stable or broadens (by i), and ‘dogs’
is stable or narrows (by ii). This is clearly correct. I leave it as an exercise for the
reader to examine the other forms of the syllogism (and the propositional calculus
for that matter).
6.3. Vagueness
As noted earlier, I consider vagueness to be a special case of meaning underdeter-
mination. Not all cases of meaning underdetermination count as vagueness (con-
sider the question of whether Secretariat is an athlete, as discussed earlier) but
all cases of vagueness are cases of meaning underdetermination. More precisely,
cases of vagueness are those cases of meaning underdetermination that rest on
(at least one) scalar dimension. Vagueness isn’t a threat to bivalence because we
typically don’t admit excessively underdetermined expressions into our micro-
languages, although they may play a role in our metalinguistic discussions about
microlanguage admissibility.
What then are we to say of the sorites argument? Recall the structure of that
argument.
If this is what is going on, then there is a sense in which there is no immediate
puzzle here. Obviously ‘bald’ has been broadened to a ridiculous level, but there
is nothing wrong with that. Let’s now add a premise to the argument to the effect
that having 1,000,000 hairs is not bald. In this case we are alleged to derive a
contradiction, as follows.
are all sharpened enough to be either true or false, then the modulation must be
sharpened to the point that having n hairs is bald and having n + 1 hairs is not
bald, so one of the conditional premises must be false.
Could someone dig in and say the following?: No look, there is a fixed modu-
lation on which all the premises are true! They could say that, but then we can
rightly ask for them to specify the modulation on which all the premises can be
true. Clearly there is no fixed meaning on which this is possible.
If we consider strict conditionals then the same thing holds; in each world
where we evaluate the material conditional, if the modulation is held fixed then
one of the premises must be false in that word.
There are many other accounts of vagueness that claim these premises are
false; what we have now is an explanation of why those premises sounded good.
Statements of the form ‘if x is P then y is P’ are typically metalinguistic devices for
encouraging us to modulate (or not modulate) a term in a particular way (e.g., if
Pluto is in the range of ‘planet’ then so must be many other Kuiper belt objects).
We don’t balk at these statements when we hear them because they are perfectly
reasonable tools for persuading us to modulate word meanings. We should not let
the reasonableness of these statements color our acceptance of stipulated non-
metalinguistic versions, however.
Sentence (4) could give rise to paradox if, for example, Nixon’s utterances about
Watergate were evenly divided between true and false except for an additional
utterance by Nixon of the form ‘Most of the things I said about Watergate are
false’.
The interesting thing about these examples is that in many cases these sen-
tence forms can be perfectly non-paradoxical. So for example, had sentence B
been something like 2 + 2 = 4 there would be nothing at all paradoxical about
A. Accordingly, philosophers like Kripke have supposed that whatever might be
said about examples like these, it makes no sense to say that there is something
Microlang uag e s , Vag uene s s , and Parad ox 135
defective about the syntax or semantics of these utterances. In Kripke’s view, the
sentences are perfectly fine qua sentences of a language; the problem is that they
are not grounded.
Microlanguages and the Microlanguage S-admissibility condition offer us
an alternative way to think about cases like these. In the first place, it is alto-
gether too quick to say that sentence A is perfectly nonparadoxical in alterna-
tive circumstances. The problem is that those cases in which A is referring
to a different sentence could well be a case in which A appears in a different
microlanguage. If so, there is thus a kind of sentence-level equivocation tak-
ing place. The form ‘Sentence B is true’ can be a different sentence in different
microlanguages. A simple instance of this is one where the meaning of ‘sen-
tence’ has shifted, but an alternative case might be one in which B has gone
unanchored.
One way of thinking about the matter is that the S-admissibility condition says
that you are not allowed into the microlanguage unless you are grounded—if you
are not grounded you are not a sentence. Here, of course, the notion of ground-
ing is understood a bit more broadly than Kripke probably intended, since being
grounded not only means that a sentence must be grounded in some nonpara-
doxical claim but that the terms of the sentence must be modulated in such a
way that the sentence is either clearly true or false; there is a recognizable path to
determining its truth or falsity.
This approach has an advantage over Kripke’s, which is susceptible to revenge-
type cases. So for example, we could introduce cases like the following.
(6)
A: B is assertable
B: A is not assertable
(7)
A: B is S-admissible
B: A is not S-admissible
(8)
A*: ‘B is S-admissible’ is (is not) S-admissible
The reason we cannot induce a paradox in these cases is not because of a technical
trick. The reason we cannot induce a paradox is that self-referential and closed-
loop cases are only admitted into the microlanguage if they are appropriately
modulated and grounded.
Now it might be objected that there is something unfair about extending the
S-admissibility criterion to liar cases because it was originally designed to ensure
that words were sufficiently modulated before they were introduced into the
microlanguage. But liar-like cases don’t look like this; Liars aren’t being held back
pending the sharpening of certain terms. That’s the objection, anyway, but I think
the objection is based on a mistaken assumption.
I think a good case can be made that ultimately liar sentences are precisely
of the character of our vagueness cases. I would go further and say that this is
perhaps true for any paradoxical sentence. The proper approach to a paradoxical
sentence is to withhold the sentence—not permitting it to be admitted into our
microlanguage pending some sharpening.
But could it really be that paradoxes like the Liar are simply awaiting the
resolution of meaning underdetermination? That is certainly the view of Jamie
Microlang uag e s , Vag uene s s , and Parad ox 137
Tappenden is thinking of cases in which matters are intentionally left open, but
we can imagine lots of reasons why aspects of word meaning might remain open
as a kind of natural default state—it may simply be too complicated to specify
everything (even for an expert), or it may be that crucial aspects of word meaning
depend upon facts about the world that remain open. Or it may just be that the
language faculty is only accidentally suitable for communication and that for no
reason in particular it just happens not to fix robust lexical meanings.
My point here is not to defend Tappenden’s particular solution to the Liar,
but rather to observe that his approach evinces a general approach to paradox
that I consider to correct. In its pithiest form it is this: Stop, back up, modulate,
try again.
That is, when we encounter paradox we kick our discussion up to a metalevel.
Until the paradox is resolved the paradoxical sentence (or paradoxical string of
sentences) is withdrawn from the microlanguage.
I want to conclude by suggesting that this isn’t just the way we handle the Liar,
but arguably also a broad range of our encounters with paradox. As we saw in sec-
tion 6.3, our solution to the sorites paradox evinced this form—we stopped and
backed up to think about several of the concepts that drove us into the paradox,
including word meaning (choosing a dynamic conception), semantic value (choos-
ing the notion of range instead of extension to characterize semantic values).
138 Reflections on the Liar
How far can we push this understanding of the general strategy for resolving
paradox? Could we think of the “resolution” of the paradoxes of set theory in this
way? The idea would be that encountering paradox, we stopped, stepped back,
and modulated the meaning of the term ‘set’ to include a hierarchy. Similarly we
might think of the resolution to McTaggart’s paradox in the philosophy of time
in the same way. We stop and rethink our understanding of the A-series and
B-series.
Of course we don’t always resolve paradox in this way. Sometimes we find (or
think we have found) a mistake in the reasoning that took us to the paradox. And
sometimes we resolve a paradox by biting a bullet (for example in epistemicism).
But my point isn’t that meaning modulation is the only strategy we have for
resolving paradox, just that it is a quite natural and general strategy that we rou-
tinely deploy when facing paradox. And my more general point is that the theory
of the dynamic lexicon and meaning underdetermination shows why this strategy
is coherent and why it is also often successful.
Notes
1. Of course on this view one presumably needs some absolute sense of ‘hair’, which I think
would be difficult to spell out. Is one cell of hair DNA in a hair follicle a hair?
2. See Endicott (2000) for discussion.
3. I owe this observation to Rebecca Mason.
4. I owe this point to Chris Gauker.
5. They use the term ‘vagueness’ but use it in a broad enough way so that it approximates what
I mean by underdetermination. Their point carries over directly to underdetermination.
6. Paul Pietroski has suggested to me that it might be possible to challenge this assumption.
Given that the characterization of the property/extension/function in each case involves
the use of the underdetermined term ‘flat’, one might wonder if these are precise objects
after all. It is hard to work through this. Would the idea be that properties are underdeter-
mined objects? There is room to maneuver with properties, but what about the set {x: x is
flat}. Does it make sense to talk about an underdetermined set (or function for that matter)?
Note we don’t want to confuse this with the idea of a fuzzy set in the sense of Zadeh (1965),
which is, it its own way, a fully determinate object.
7. It is, however, as they point out, only approximately true that what we say is
approximately true.
8. However Braun and Sider believe Lewis (1993, p. 29) is hedging in the following passage
(their emphasis): “Super-truth, with respect to a language interpreted in an imperfectly
decisive way, replaces truth simpliciter as the goal of a cooperative speaker attempting to
impart information.” I’m unclear on what the force of this point is supposed to be, given
that supertruth can do all the work that truth was supposed to do.
9. Teller (2012, p. 8) provides an example.
There is something I want to emphasize: I do not advocate giving up the familiar frame-
work that includes presuppositions 1) [that semantic values, in particular propositions,
are precise] and 2) [that for a statement to be true is for the statement to express a true
proposition]. The mystique of Kuhnian paradigm shifts to the contrary, science does not
generally discard frameworks that have proven broadly fruitful … [T]he framework of
formal semantics will continue to provide the best way of understanding many impor-
tant features of language.
Microlang uag e s , Vag uene s s , and Parad ox 139
10. Clearly this fragment not only ignores plenty of structure (like the structure of NP and VP
but also takes certain liberties in, for example, classifying ‘is flat’ as a VP when it is prob-
ably better classified as an adjective phrase (AP). This won’t matter for our purposes. More
robust fragments can be found in Larson and Segal (1995).
11. For this you might retain a traditional model- theoretic semantics for that purpose,
or you could run your theory of entailment off of the syntax as, for example, in Ludlow
(1995, 2002).
12. I am indebted to John McFarlane for conversations here.
13. The proviso that it “plays a direct role in the derivation of the conclusion” is designed to
allow us to ignore terms that have multiple occurrences but that are inert in the structure of
the argument. (I am indebted to an anonymous reviewer for pointing out the need for this
proviso.)
So, for example, consider the following argument form:
(P v Q) → R
Q→S
P
Therefore: R
The terms in Q have multiple occurrences, but they do not play a role in the derivation of
R. Matters are a bit more complicated than this example lets on, however, because some-
times there are multiple ways to derive the conclusion in a sound argument. For example:
(P v Q) → R
Q
P
Therefore: R
We can either define soundness relative to a derivation (so that derivations are sound, not
arguments), or we can say that the form of the argument is sound if there is at least one
derivation path to the conclusion such that the argument consisting of just that deriva-
tion path respects the Dynamic Lexicon Constraint on Soundness. Other solutions are,
most likely, available. See van Deemter and Peters (1997) for a survey of articles on the
general problem developing logics that tolerate ambiguity, indeterminacy, and by extension
underdetermination.
14. If you wish, you can (following recent work in dynamic logic) treat each of the conditional
premises as instructions for updating the common ground. In this case the argument
wouldn’t really be an argument as above, but rather a series of instructions on modulat-
ing the meaning of ‘bald’ in the common ground. I remain officially neutral on this way of
executing the idea.
References
Barker, C. 2002. “The Dynamics of Vagueness.” Linguistics and Philosophy 25: 1–36.
Braun, D., and T. Sider. 2007. “Vague, So Untrue.” Noûs 41: 133–156.
Burge, T. 1979. “Individualism and the Mental.” Midwest Studies in Philosophy 4: 73–121.
Davidson, D. 1967. “Truth and Meaning.” Synthese 17: 304–323. Reprinted in Inquiries into Truth
and Interpretation (Oxford: Oxford University Press, 1984).
——— . 1986. “A Nice Derangement of Epitaphs.” In E. Lepore, ed., Truth and
Interpretation: Perspectives on the Philosophy of Donald Davidson, 433–446. Oxford: Blackwell.
Endicott, T. 2000. Vagueness in the Law. New York: Oxford University Press.
Fara, D. G. 2000. “Shifting Sands: An Interest-Relative Theory of Vagueness.” Philosophical Topics
28: 45–81. Published under the name “Delia Graff.”
140 Reflections on the Liar
This chapter is about the relevance of liar paradoxes for truth-theoretic accounts
of linguistic meaning. In “Framing Event Variables” (Pietroski 2015) I develop an
independent objection to such accounts. But in both discussions, the first num-
bered sentence is the one displayed below.
Let ‘Lari’ be a name for the first numbered sentence in that other paper (Pietroski
2015). Since the definite description in (1) can be used to describe Lari, the gram-
matical subject of (1) can be used to talk about that very sentence. So (1) can be
used to say that (1) is false. Likewise, (2)
can be used to say that it isn’t true. In my view, this isn’t paradoxical: it’s true that
(2) isn’t true, and it’s false that (1) is false. But such examples do tell against the
Davidsonian thesis (D);
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H,
141
142 Reflections on the Liar
Though as discussed below, it seems that any truth theory that accommodates
(1) and (2) will be too sophisticated to serve as the core of a plausible meaning the-
ory that applies to (3) and (4); and this concern is amplified if Human Languages
are I-languages in Chomsky’s (1986) sense.
In short, sentences like Lari highlight a deep tension between (D) and an inde-
pendently plausible conception of Human Languages. I think the best response
is to deny that (1) and (2) have truth conditions, and then admit that (3) and
(4) also fail to have truth conditions. While sentences of a Human Language are
often used to express truth-evaluable thoughts, it remains a hypothesis that these
sentences are themselves truth-evaluable. And this hypothesis has implausible
consequences; indeed, it leads to contradiction, given very plausible background
assumptions. In my view, this tells against the idea that sentences of a Human
Language have truth conditions, even if theorists can invent ways of evading or
embracing the reductio.
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
(D′) There are correct Tarski-style theories of truth for Human Languages.
Prima facie, (D′) is quite implausible. But one can try to motivate and defend (D′)
as part of a package that also includes the following conditional claim: if (D′),
then (D). By itself, this conditional is plausible; and (D) offers an attractive
conception of linguistic meaning. Though as I’ll argue in sections 7.2 and 7.3,
the possibility of liar sentences—i.e., the fact that Human Languages can gener-
ate such sentences—tells against conjoining (D′) with a conditional that leads
from (D′) to (D). The net result, I claim, is that both theses remain implausible.
If at least one such theory is correct, at least in the sense that each of its theorems
is true, then one might hope that some such theory can serve as the core of a
theory that correctly specifies the semantic properties of all English expressions.
But one might suspect that any Tarski-style theory of truth for English will gener-
ate incorrect theorems, even restricting attention to declarative sentences. And it
is important to distinguish two kinds of skepticism.
One might concede that many English sentences have truth conditions, but
suspect that certain constructions introduce special complications, with the
result that some sentences either fail to have truth conditions or have truth con-
ditions that preclude a correct Tarski-style theory. Alternatively, one might think
that no English sentences have truth conditions, but that some constructions are
especially useful for purposes of making arguments that apply more generally. My
skepticism is of the latter sort. For example, I think (7) and (8) fail to have truth
conditions.
But my claim is not that (D) is false because of special words like ‘think’ and ‘true’.
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H.
It’s easy to provide implausible theories. For example, given that Hesperus is
Phosphorus, it’s easy to provide theories according to which (7) is true whenever
(8) is true. But prima facie, any such theory is wrong. This can make it tempting to
say that for any context c, (7) is true relative to c if and only if Bert bears a certain
relation R to sentence (9) relative to c,
and that (8) is true relative to c if and only if Bert bears R to sentence (10)
relative to c.
The familiar idea is that depending on the context, a thinker might bear R to
(9) but not (10)—or to (10) but not (9), or to both sentences, or to neither. But
depending on how the alleged relation is described, and how contexts are speci-
fied, the corresponding instances of ‘True(S, c) ≡ F(c)’ may be hard to evaluate—
and in that sense, neither implausible nor plausible.
Let’s focus on (7) and (9), repeated below, and consider a specific proposal:
(7) is true relative to c if and only if Bert endorses the thought expressed with
(9) relative to c. This biconditional, which employs technical notions, is not a tru-
ism.2 Its “left side” concerns the truth of (7); its “right side” concerns Bert and
a thought allegedly expressed with (9). To make an informed judgment about
whether the biconditional is true, we need independent measures of whether
each side is true, given various contexts. But how do we tell whether or not Bert
endorses the thought expressed with (9) relative to a given context?
It’s hard enough to say whether or not (7) is true relative to the contexts associ-
ated with certain episodes of communication, once the relevant notion of context
becomes technical. But we can gather evidence about the semantic properties of
(7) by describing scenarios, and asking competent speakers whether or not Bert
thinks that Hesperus rose at dusk. For we can treat the responses—‘yes’, ‘no’,
‘maybe’, ‘I don’t know’, etc.—as evidence of whether or not, in the conversational
settings we have created, sentence (7) can be used to describe the scenarios cor-
rectly. So given some ancillary assumptions about contexts, and how to control
for various pragmatic effects, we can draw conclusions about whether (7) is true
relative to various contexts. Let’s assume, for now, that such conclusions can be
reliable. Indeed, let’s pretend that for each context c, we know whether or not
I-Lang uag e s and T- Se nte nc e s 145
(7) is true relative to c. That still leaves the harder question: is (7) true relative to c
if and only if Bert endorses the thought expressed with (9) relative to c?
If we can’t tell—because we we’re not sure what it is for sentences to express
thoughts relative to contexts—then providing a Tarski-style theory that assigns
the alleged truth condition to (7) doesn’t show that some such theory accommo-
dates (7) in a plausible way. It’s good for theories to avoid the implication that
(7) is true if and only if Bert thinks that Phosphorus rose at dusk. But one wants
a theory that can be confirmed by noting that its theorems seem to be true. And
for these purposes, it won’t help to exchange ‘endorses the thought expressed
with (9)’ for ‘thinks what (9) says’ or ‘samethinks with (9)’ or ‘thinks* (9)’; com-
pare Davidson (1967a), Segal (1989), Larson and Ludlow (1993). Instances of
‘True (S, c) ≡ F(c)’ invite a picture in which speakers of the relevant language can
evaluate the left side (even if they don’t have the technical knowledge required
to evaluate the right side), and theorists can evaluate the right side (even if
they don’t have the linguistic knowledge required to evaluate the left side); see
Davidson (1967b, 1984). We assume that the speakers in question can under-
stand the object language sentence on the left and make reliable judgments
about whether it can be used to describe various scenarios correctly. But like-
wise, theorists need to understand the metalanguage sentence on the right and
be able to make reliable judgments about whether it is true in those scenarios.
This doesn’t show that any particular proposal is wrong, much less that “atti-
tude reports” are counterexamples to all Tarski-style theories of truth for Human
Languages. But given the range of mental states that (7) can be used to describe,
and the many ways in which the truth of corresponding assertions can depend on
details of the conversation, why think that (7) has a stable truth-theoretic “char-
acter” that maps contexts to truth or falsity? And if there are reasons for doubting
that (7) has a truth condition, then absent countervailing considerations, why
think that any unconfirmed claim of the form ‘(9) is true relative to c if and only
if … c …’ is true?
As discussed below, one can’t just insist that relative to every context, (7)
is true if and only if Bert thinks Hesperus rose at dusk. Disquoting doesn’t ensure
truth. If (7) doesn’t have a truth condition, then (11) isn’t true, not even relative
to particular contexts.
(11) ‘Bert thinks that Hesperus rose at dusk’ is true if and only if
Bert thinks that Hesperus rose at dusk.
If it was clear that “simpler” sentences have truth conditions, then it might be
reasonable to bracket doubts about propositional attitude reports, and conclude
146 Reflections on the Liar
that (7) has a truth condition that is just harder to specify. But upon reflection, it
isn’t clear how to specify a correct instance of ‘True(S, c) ≡ F(c)’ for any sentence
S of a Human Language.
Following Davidson (1967a) and many others, one might suggest that (9)
is true relative to a context c if and only if ∃e[Rise(e, Hesperus) & At-Dusk(e) &
Past(e, c)]; where ‘Rise(e, Hesperus)’ indicates that e is a rising of Hesperus, and
‘Past(e, c)’ indicates that e occurred prior to the time of c. But to evaluate any such
biconditional for truth or falsity, given a context, one needs to evaluate its right
side. So in particular, ‘Rise(e, Hesperus)’ must be clear enough that theorists—
who know that the sun doesn’t rise, but that bread can rise—can say whether or
not the world contains events that satisfy this open sentence. And given any par-
ticular construal of ‘∃e[Rise(e, Hesperus) & At-Dusk(e) & Past(e, c)]’, one needs to
argue that this invented sentence is true relative to c if and only if (9) is.3 Similar
remarks apply to (12).
And while (6) is an even simpler sentence, it exhibits at least two significant
complications.
One is that ‘yells’ suggests a habit of yelling.4 This raises the question of whether
a Tarski-style theory can plausibly accommodate both ‘yells’ in (6) and ‘yell’ in
‘heard Bert yell’ or ‘make Bert want to yell’. Characterizing the relevant notion of
habitual (or “generic”) properties is notoriously difficult, however these alleged
properties are related to meanings; see Leslie (2007) for discussion of examples
like (13).
But even if we characterize a notion of being a chronic yeller, and then stipulate
that the right side of (6T) is true if and only if the entity denoted by ‘Bert’ is a
chronic yeller,
the right side of this biconditional still cannot be evaluated for truth or falsity,
absent a further stipulation about which entity is denoted by the metalanguage
expression ‘Bert’. There are many Berts, none of which is the bearer of the proper
noun ‘Bert’ that appears on left side of (6T); see, e.g., Burge (1973). So whatever
I-Lang uag e s and T- Se nte nc e s 147
‘Bert’ denotes on the right, (6T) mischaracterizes (6), which is not about any par-
ticular Bert. Examples like (14) highlight other questions about names:
and so they are not parade cases that motivate the hypothesis despite examples
like (7) and (12).
Davidson (1967b) initiated a long and fruitful line of work. But so far as I know,
there are still no parade cases of ‘True(S, c) ≡ F(c)’ where S is a sentence of a Human
Language. Every proposed instance of this schema seems like a promissory note,
at least to some extent. For many purposes, that’s fine. But we shouldn’t pretend
that there are good inductive reasons for thinking that for every sentence of a
Human Language has a truth condition. There is no model that yields plausible
instances of ‘True(S, c) ≡ F(c)’ for elementary sentences, and that theorists have
been progressively applying to more complicated constructions.
That said, induction is not the only way to motivate a generalization. If (D′)
is true,
(D′) there are correct Tarski-style theories of truth for Human Languages
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
And if we assume (D), then (D′) becomes a lot more plausible. Put another way,
while (D′) seems grossly implausible if assessed solely in terms of its implications
148 Reflections on the Liar
for particular sentences, one can try to motivate and defend (D′) as part of
a larger conception of linguistic meaning. Given this project, it’s reasonable to
accept various promissory notes as theorists try to formulate truth theories that
yield increasingly better instances of ‘True(S, c) ≡ F(c)’. If an advocate of (D′)
concedes that each sentence of a Human Language presents difficulties for this
hypothesis, but claims that (D′) is still part of an attractive account of linguistic
meaning, then the question is whether the difficulties outweigh the attractions
of using truth theories as meaning theories. When addressing this question, it’s
reasonable to ask skeptics for concessions with regard to specific examples. But
then one can’t assume that (D′) is independently plausible, on empirical grounds,
when faced with objections to the idea that truth theories can serve as meaning
theories.
In sections 7.2 and 7.3, I will be pressing the following question: is any Tarski-
style theory of truth for English sophisticated enough to be correct, given consid-
erations that are easily illustrated with liar sentences, yet naive enough to be the
core of a correct theory of meaning for a language that children can acquire? In
defending an affirmative answer, advocates of (D′) need not provide theories that
yield plausible instances of ‘True(S, c) ≡ F(c)’ for specific examples. But neither
can they assume that sentences of a Human Language have the truth conditions
assigned by some Tarski-style theory.
I’ll return to the idea defending (D) by rejecting (D′) and saying that an incor-
rect theory of truth for H can be the core of a correct theory of meaning for H;
compare Eklund (2002). But for now, let’s bracket this idea, which comes with the
burden of saying which false theorems of the form ‘True(S, c) ≡ F(c)’ are tolerable.
I think that getting clear about why conjoining (D) and (D′) is implausible, even
setting aside objections to (D′) that are based on specific examples, undercuts
much of the motivation for (D). By being concessive about the difficulties that
particular sentences present for (D′), I hope to show that (D) faces a general dif-
ficulty that gets manifested in many ways.
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
Ancillary assumptions are required; and not every truth theory for a Human
Language can be supplemented in the requisite way. Davidson (1967b,
1984) held that only suitably formulated truth theories that meet certain empiri-
cal constraints can play this role. There are at least two clusters of reasons for
saying this. The first, and perhaps most obvious, is that not all sentences are
declarative. The second, and perhaps more important, was stressed by Foster
I-Lang uag e s and T- Se nte nc e s 149
(1976): given one correct truth theory for a Human Language H, there are
boundlessly many such theories none of which is the core of a correct theory of
meaning for H.
If a truth theory Θ is to serve as the core of a meaning theory for English, then
Θ needs to be formulated so that it can be supplemented in ways that accommo-
date expressions like (15)–(20).
One obvious thought, which can be developed in various ways, is that each sen-
tence of a Human Language inflects a sentential “radical” with a grammatical
mood; where the radical has a truth condition, and the mood defeasibly indicates
the kind of speech act being performed.5 The idea is that (15) and (5) share a part
that is true (relative to a context c) if and only if Ernie snores (in c);
But let’s not fuss about this. Given a truth theory that can serve as a decent
theory of meaning for English declaratives, it would be churlish to complain if the
theory doesn’t itself explain everything we would like to explain about the mean-
ings of other sentences.
Still, a correct theory of meaning for a “declarative fragment” of a Human
Language must be extendable to nondeclarative sentences. We can agree to set
(15)–(21) aside for many purposes, just as we can to set (2) and (7) aside.
But for any grammatical mood, to accommodate sentences with that mood in the
way required by thesis (D), a truth theory must be formulated so that it can be
embedded in a larger theory that accommodates sentences with other moods.
Likewise, a correct theory of meaning for any “fragment of a declarative frag-
ment” that excludes words like ‘true’ and ‘think’ must be formulated so that the
theory can be extended to handle sentences that include such words. This turns
out to be a nontrivial constraint, for reasons connected with “Foster’s Problem.”
Suppose that Ernie snores and Bert yells, and that the invented sentence (22)
is thus true.
If we ignore context sensitivity, and pretend that the natural sentences (5) and
(6) are also true,
then the invented sentences (23) and (24) are true instances of ‘True(S) ≡ p’.
But (23) and (24) are uninterpretive in the following sense: the metalanguage sen-
tence, used on the right side, is not a good translation of the object language sen-
tence mentioned on the left. So a theory that generates (23) and (24), rather than
(5T) and (6T),
And if the invented sentence (28) is false, then (29) is also true but uninterpretive.
(28) <(5, 2)
(29) True(‘Every number is even.’) ≡ <(5, 2)
I-Lang uag e s and T- Se nte nc e s 151
Given any interpretive instance of ‘True(S) ≡ p’, there will endlessly many
uninterpretive analogs that replace the instance of ‘p’ with a sentence having the
same truth value. This is unsurprising, since ‘≡’ indicates a truth function, and
sentence meanings are individuated more finely than truth values. But it raises
the question of how any Tarski-style truth theory, even one whose theorems hap-
pen to be interpretive, can serve as a meaning theory for a human language. What
would make any such a theory more than a mere specification of truth/falsity for
sentences of the object language?
As (25)–(29) suggest, it doesn’t help to stipulate that a truth theory is modally
correct only if its theorems are true at every possible world. Moreover, even if
examples involving mathematical/logical necessity are set aside as special cases,
sentence meanings seem to be individuated more finely than sets of possible
worlds; see Kripke (1980). If Hesperus is Phosphorus at every world where Venus
exists, then (30) is true but uninterpretive.6
If there are no worlds where Plato went back in time and killed all his ancestors,
and also none where Edison invented a perpetual motion machine, then (31) is
true but interpretive.
(31) True(‘Plato went back in time and killed all his ancestors.’) ≡ Edison
invented a perpetual motion machine
Advocates of thesis (D) can posit worlds where (30) and (31) are false; see, e.g.,
Lewis (1986).
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
But to defend the idea that a modally correct truth theory will not have uninter-
pretive theorems, one needs to provide independent arguments for positing the
requisite space of possible worlds.
A different idea is that a truth theory with the right axioms will generate only
interpretive theorems; see Davidson (1976). Stressing derivability, as opposed to
modality, can seem more promising. And by aiming for axioms that generate “T-
sentences” like (5T), as opposed to (23),
one might hope to specify theories that speakers naturally but tacitly deploy; see
section 7.2.3 below. But this strategy highlights three further questions for advo-
cates of (D).
First, are there limits to how technical the right side of an interpretive T-
sentence can be? Compare (5T) with (32), which is genuinely disquotational.
The sentence used on the right side of (32) is the sentence mentioned on the left
side; and this sentence of the object language—viz., (5)—is a good translation of
itself.
So (32) is interpretive. But one wants to know how far T-sentences can depart
from the disquotational paradigm, yet still be interpretive, given the need for sig-
nificant formalism in the metalanguage. The importance of this question becomes
obvious given examples that involve context sensitivity and quantification, as
illustrated with (33), which isn’t true if and only if I chased something.
In the spirit of concession, let’s grant that for any context c: True(‘I chased some-
thing.’, c) ≡ ∃e∃x[See(e, Speaker[c], x) & Past(e, c)]; where the right side of this
biconditional is an invented sentence that has a certain (stipulated) interpreta-
tion. And let’s grant that some such T-sentence for (33) can be derived from plau-
sible axioms, given appropriate structural descriptions of English expressions. But
a theorist who says that his “formalized” T-sentences are interpretive, and thus
relevantly like (32) as opposed to (23), owes some account of what distinguishes
interpretive but not disquotational T-sentences from mere specifications of truth/
falsity for sentences of the object language; see Lepore and Ludwig (2007) for
further discussion.
One can reply that right sides of derivable T-sentences can exhibit “logical
forms,” so long as they don’t employ notions that are “foreign” to the object lan-
guage sentences, whose meanings may be obvious to competent speakers but not
manifest in representations like (32). Though as we’ll see, it is far from obvious
that this reply is available to advocates of thesis (D).
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
For even if a truth theory does not generate implausible T-sentences, it may
employ notions that are foreign to the languages that children naturally acquire.
I-Lang uag e s and T- Se nte nc e s 153
This point is related to a second question: does any truth theory for a Human
Language have enough “deductive power” to generate interpretive T-sentences,
but not enough to generate uninterpretive analogs? Tarski envisioned a truth
theory as an addition to a nonsemantic theory—e.g., of arithmetic—already for-
mulated in a language that is governed by a logic, which licenses deduction of
theorems from axioms. From this perspective, it isn’t a problem if (29)
can be derived from the enlarged theory. On the contrary, it would be more wor-
risome if some analog of (29) wasn’t derivable from a theory according to which
the right side is false. The left side is true if and only if every number is even; and
it’s not the case that every number is even. More generally, uninterpretive theo-
rems seem unavoidable once a truth theory is embedded in a theory governed
by a logic that licenses some inferences that preserve truth but not meaning.
So one might wonder how any plausible theory can generate only interpretive
T-sentences.
Perhaps the right response is that a truth theory for a Human Language has its
own limited subject matter, and that at least for purposes of distinguishing inter-
pretive T-sentences from uninterpretive analogs, a “pure” truth theory isn’t part
of a larger theory-of-the-world. I think it’s odd to view theories of truth as segre-
gated, in this way, from theories of the things that words and phrases are alleg-
edly true of.7 But let’s not worry here about whether truth theories for Human
Languages can be isolated from theories of numbers, space-time, matter/energy,
life, animal psychology, Human Language syntax, etc. In particular, let’s grant
that a truth theory for English need not be interwoven with premises that would
support a derivation of (34) from (5T).
Even this much concession still leaves the question of whether any theory Θ that
generates interpretive T-sentences like (5T) can fail to generate uninterpretive
T-sentences of the forms shown in (35) or (36); where Γ is itself a theorem of Θ.
The scope of this point depends on the relevant background logic. But consider a
Fregean (second-order) logic that licenses instances of (37); where A is a theorem
of arithmetic,
(37) HP ⊃ A
154 Reflections on the Liar
and ‘HP’ stands for (the conjunction of suitable Fregean definitions and) the
Humean Principle that some things correspond one-to-one with some things if
and only if the former have the same cardinality as the latter; see Heck (2011) for
discussion. Replacing Γ in (35) and (36) with instances of (37) yields T-sentences
that are wildly uninterpretive. One can try to formulate theories of truth against
the background of a weak logic that does not generate interesting theorems of its
own. But a truth theory that has (5T) and (6T) as theorems will also have (38) as
a theorem
if the theory permits replacement of a sentence on the right side of ‘≡’ with the
conjunction of that sentence and a theorem. So the background logic needs to be
very weak. I’ll return to this point in section 7.2. But note that the requisite isola-
tion of a truth theory is radical. Given some elementary principles of logic, a truth
theory will generate uninterpretive T-sentences.
We are, recall, reviewing some questions raised by the following idea: if a truth
theory for English has no uninterpretive theorems, then it can serve as the core of
a theory of meaning for English. The first two questions concerned the plausibil-
ity of this conditional’s antecedent. But as Foster (1976) discussed, one can ask a
third question about the conditional itself.
Given a theory that generates (5T), one can easily construct an extensionally
equivalent theory whose theorems include (39), where ‘TRUTH’ stands for any
truth you like.
For illustration, suppose that (5T) follows from a theory that has (40) and (41)
as axioms,
Let’s not worry about the formal gap between (5T) and the conjunction of (40)–
(42). The point is that however this gap is filled—so that (5T) is actually derivable
from (40)–(42) via some suitably general principle—one can replace (40) with
(43), and then derive (44).
Alternatively, one can replace (42) with (45), and then derive (44).
since another truth theory has (44) as a theorem. At least one of these truth theo-
ries cannot be the core of a correct theory of meaning, since (5) and (46) differ in
meaning.
Both theories may generate (47). But the point remains that (44) is uninterpretive.
One can say that (43) and (45) are not interpretive axioms. Indeed, that’s one
way of putting Foster’s point. But in my view, it’s no defense of (D)
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
to say that a theory of truth can serve as a theory of meaning so long its axioms
are interpretive. If a truth theory is to explain why Human Language expressions
mean what they do, then we need some account of what makes (40)–(42) bet-
ter than alternatives like (43) and (45), as theoretical descriptions of the semantic
properties exhibited by the English expressions.
156 Reflections on the Liar
will be as derivable as (5T), regardless of what ‘Lari’ and ‘Linus’ name. But if the
psychological hypothesis is correct, one can try to explicitly formulate the theo-
ries that speakers somehow encode and tacitly deploy. And as theorists progress
toward this goal, they can say that for any Human Language H, a proposed truth
theory Θ is more interpretive than alternative proposals if Θ better reflects how
speakers of H represent the semantic properties of H-expressions.
The appeal to a language of thought abandons certain reductive ambitions
for theories of meaning; compare Dummett (1976). And by focusing on how
speakers actually represent the alleged truth conditions of English sentences,
one goes beyond the task of specifying a theory such that knowledge of it would
suffice for knowing that English sentences have those truth conditions; compare
I-Lang uag e s and T- Se nte nc e s 157
Davidson (1976), Foster (1976). But these are not yet objections to a cognitivist
gloss of (D).8
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
So absent a better response to the questions about how a truth theory could ever
serve as the core of meaning theory, let’s grant that a cognitivist gloss of (D)
provides a way of distinguishing interpretive from uninterpretive T-sentences.
At this point, we have conceded a lot. We’re supposing that some theory Θ,
yet to be explicitly formulated, is such that Θ assigns truth conditions to English
declaratives (modulo liar sentences) in a descriptively plausible way that accom-
modates the respects in which the alleged truth conditions are context sensitive;
Θ can be supplemented with ancillary hypotheses in a way that accommodates
nondeclarative sentences; and because Θ meets certain formal and empirical
constraints, its theorems are interpretive. But (D) goes farther, and not merely
because of the generalization to other languages. While an interpretive truth
theory represents truth conditions in a special way, it still assigns truth conditions
to sentences. And sentences of a Human Language, along with their constituent
expressions, may be significant in other ways.
Davidson was inspired by invented languages whose expressions have no
interpretive properties apart from the semantic properties characterized by
truth theories. By stipulation, the expressions of such a language have no mean-
ings that go beyond the truth-theoretic properties specified by some Tarski-
style theory. But the expressions of a Human Language may—perhaps as
devices for accessing and combining concepts in certain ways—have meanings
that are prior to and often presupposed by their varied uses, including uses as
tools for making truth-evaluable claims in contexts; see Pietroski (2005, 2008,
2010, 2015 forthcoming). I think liar sentences point in this direction. They
reveal a fundamental problem for the project of characterizing linguistic mean-
ings in terms of truth.
sensitivity and quantification. As we’ll see in section 7.2.3, sentences like (33)
heighten the basic concern.
and that ‘Lari’ is a name for this sentence. Given plausible assumptions, Lari is
false if Lari is true, and Lari is true if Lari is false. So prima facie, Lari is neither
true nor false.9
This conclusion is not paradoxical. Many things are neither true nor false: frogs,
numbers, rocks, stars, rock stars, etc. We can also invent languages whose sen-
tences exhibit three truth values—T for true, ⊥ for false, N for neither—where
a sentence is false if and only if its negation is true; see Kleene (1950). But the
truth-evaluable things, which may include many acts of judgment/assertion, may
not include sentences of Human Languages.
Another possibility is that such sentences are typically true or false, relative to
contexts, but that (1) and (50) are among some special cases that are neither true
nor false.
If (50) is neither true nor false, then both sides of (48) are false, and so (48) is true.
One might think that (51) is also true, along with its right side, and hence that
(52) is true.
So one might conclude that a truth theory for English can and should have theo-
rems like (48) and (51). I disagree. In my view, (51) is false because (52) is neither
true nor false; the right side of (51) is true, but its left side is false. More impor-
tantly, I don’t think a theory of meaning for English should generate theorems
like (48) or (51). Such a theory will go wrong in many ways.
I-Lang uag e s and T- Se nte nc e s 159
If (2) is true, then (2) is not true. So (2) is not true. But if sentences of the form
‘… is not true’ are among the truth-evaluable things, then presumably, (2) is true;
and if a correct theory of meaning has theorems like (51), then presumably, sen-
tences of the form ‘… is not true’ are among the truth-evaluable things. I con-
clude that a correct theory of meaning for English should not have theorems like
(51). But it is worth going slowly and being explicit about the relevance of these
points for the idea that truth theories can serve as meaning theories.
Suppose that Kermit, the largest frog in Walden Pond, is neither red nor blue.
Then modulo context sensitivity, (4) is true if any sentences containing the word
‘red’ are true.
And if (53) is true, along with its right side, then (4) is true.
So one might conclude that (4) is true, along with both sides of (53). Likewise, if
Linus is a frog, then Linus is neither true nor false. In which case, (54) is true if
(49) is true.
And if Linus is a frog—or even if Linus is a sentence that isn’t true—one might
be tempted to say that both sides of (49) are true. But suppose that Linus is
sentence (2).10
Again, if Linus is true, then Linus is not true. So like Lari, Linus is not true. But
since Linus is the second numbered sentence in this very chapter, (54) is true if
and only if (2) is true; and so (54) is true if and only if Linus is true. Hence, (54) is
not true. No surprise there. But given who Linus is, (49) implies that the sentence
‘Linus is not true.’—a.k.a. (54)—is true if and only if (2) is not true. And (2) is not
true. So while (49) looks plausible, it implies that (54) is true. Hence, (49) is not
true, and so no true theory has (49) as a theorem.
160 Reflections on the Liar
That’s not a paradox. That’s an argument against any theory that has (49)
as a theorem. The right side of (49) is true, since Linus is not true. But the left
side of (49) is false, since (54) is not true. That’s OK: (49) is an invented sentence
of a metalanguage that lets us express certain theoretical claims about Human
Languages; and theoretical claims are often false.
Perhaps (55) is true, and (54) is not TruthEvaluable in the relevant sense.
That’s also OK.
One can say that (54) fails to be true in the way that Kermit fails to be true—viz.,
by not being the sort of the thing that has a truth condition. If (54) doesn’t have
a truth condition any more then Kermit does, then we need not (and should not)
say that (54) is true if and only if Linus isn’t true. You can say that (56) is true,
since both sides of the biconditional are false.
But don’t replace the right side with a truth, as in (49), and try to maintain the
biconditional.
then it’s hard to see how the theory could fail to have (49) as a theorem. So evi-
dently, a theory of meaning had better not have (51) as a theorem. Likewise, even
though (48) is true,
it seems that a theory of meaning had better not have (48) as a theorem, on pain
of having (51) and (49) as theorems. It doesn’t follow that (51) is false. Theorists
can still appeal to three-valued logics and say that Lari—a.k.a. (1)—has the third
truth value N.
But absent reason for thinking that (51) follows from a good theory of meaning,
why think (51) is true? Indeed, one might suspect the left side of (51) is just as
I-Lang uag e s and T- Se nte nc e s 161
false as the left side of (49). For if (54) doesn’t have a truth condition any more
than Kermit does, why think (52) is different?
These points ramify. If (52) doesn’t have a truth condition, why think (50) is
different?
Upon reflection, the truth of (48) provides no reason for thinking that (50) has a
truth condition.
then it’s hard to see how the theory could fail to have (48) as theorem. Likewise,
if the theory has (53) as a theorem, then it’s hard to see how the theory could fail
to have (49) as theorem.
This is an indirect argument that no correct theory of meaning for English has
theorems like (53) and (57). It would be tendentious to assume that the left sides
of these biconditionals are false, like the left sides of (48) and (49). But even if one
can consistently maintain (53) and (57), it doesn’t follow that one should do so.
Perhaps the left sides of (53) and (57) are false—though not as obviously false as
the left side of (49)—because (4) and (58) don’t have truth conditions.
For we can say that (59) doesn’t have a truth condition any more than (58) does.
162 Reflections on the Liar
where ‘S’ is replaced by an English declarative and ‘S’ is replaced with that declara-
tive quoted. For even if ‘modulo context sensitivity’ can be cashed out plausibly
for examples like (61),
(61) ‘I am hereby quoting him.’ is true if and only if I am hereby quoting
him.
saying that instances of (60) are true, modulo context sensitivity, is tantamount
to adopting (D).
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
Davidson and others may be right that if there are endlessly many true
instances of some context-sensitive version of (60), then the obvious and perhaps
only explanation for all those truths is that they follow from a finite number of
more basic principles that constitute a truth theory for English. But the anteced-
ent of this conditional is as much in question as (D) itself. Note that (60) has end-
lessly many instances like (62),
(62) ‘Bert thinks that Ernie snores.’ is true if and only if Bert thinks Ernie
snores
whose quoted sentence can be used to report (some of) what Bert thinks. If all
these English biconditionals are true, they presumably follow from a finite num-
ber of basic principles that constitute a truth theory for English. But that’s a
terrible argument for the hypothesis that some theory of truth for English accom-
modates belief ascriptions. The argument takes immediate consequences of the
hypothesis as premises, when those consequences are also at issue.
So while it might seem that (D) explains the “fact” that instances of (60)
are true,
(63) ‘Linus is not true.’ is true if and only if Linus is not true.
This instance of (60) might seem to be true if Linus is a frog, or a sentence that
has the third truth value N. But if Linus is sentence (2), then (63) looks like a
counterexample to (60).
where both conditionals are true, but the antecedent of (49L) is false, and the
antecedent of (53L) is true. Let’s not worry here about how the technical notion
‘Legit’ gets spelled out—e.g., via fixed points (Kripke 1975), revisions (Gupta and
Belnap 1993), or whatever. For even if all the troublemakers can be filtered out,
two related concerns remain if the goal is to defend thesis (D).
First, it seems that any such filtering will yield consistency at the cost of expla-
nation. If endlessly many expressions of the object language H are such that a
proffered theory of meaning for H fails to assign interpretations to those expres-
sions, then we seem to be left with the same kind of explananda that we started
with. With regard to Human Languages, the theoretical task is to explain how
humans understand linguistic expressions, not merely to explain how a finite
mind could understand boundlessly many expressions. We want theories of how
we understand the (boundlessly many) expressions that we do understand. If a
theory of meaning accommodates ‘chased every cow’ and ‘cow that saw brown
dogs’ but not ‘every dog that chased brown cows’, then it does not accommodate
the shorter phrases in the right way. And if a truth-theoretic account of meaning
classifies (4) but not (54) as truth-evaluable,
The second concern is that theorems like (49L) and (53L) don’t seem to be
interpretive. One can say that competent speakers recognize the antecedent of
(53L) as true, that they discharge it to obtain (53), and that (54) is somehow
understood by analogical extension.
But if speakers can discharge the antecedent of (53L), then we need to revisit the
concession that speakers tacitly deploy some truth theory Θ that is governed by a
very weak logic that does not generate uninterpretive theorems like (36);
Once the last bit of paper is deposited, someone utters (66). Unlike the arithmetic
sentences (64) and (65), (66) is context sensitive if only because it is tensed. But
it’s hard to see how any context-sensitive element of (66) could track which sen-
tence the ninth person chose to inscribe. Yet the utterance of (66) is false, true, or
neither depending on this choice.
In my view, (66) does not itself have a truth condition—not even relative
to a context that determines the relevant bits of paper, and hence the relevant
inscribed sentence. If the ninth person inscribed (64) or (65), it might seem that
(66) is straightforwardly false or true. But if the ninth person inscribed (66), the
problem becomes manifest. One can say that (66) is Legit in many but not all
I-Lang uag e s and T- Se nte nc e s 165
contexts. But if one offers a theory of meaning that generates theorems like (5L),
with some context relativization,
one needs to say how and under what conditions speakers discharge the ante-
cedents of such conditionals without deploying knowledge that would also yield
uninterpretive instances of (36).
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
This may not preserve the idea that each language in the hierarchy has a truth
predicate. But advocates of (D) may not care, if only a few truth predicates are ever
used in ordinary discourse.
A deeper concern is that on Lycan’s view, sentences of λ can be sentences of λ′,
but sentences of λ′ containing a word like ‘true’ cannot be true-in-λ. As discussed
in section 7.3, I think this conflicts with the idea that Human Languages are
generative procedures. Chomsky (1986) spoke of “I-languages” in part to distin-
guish these procedures, characterized intensionally, from sets of expressions or
external manifestations of internalized generative rules. Though as Lycan’s dis-
cussion reminds us, Davidson held that a single utterance might be transcribed
with/empedɔkliyz liypt/, classified by one interpreter as an utterance of the
English sentence (67), and classified by another interpreter as an utterance of
the German sentence (68).
But it isn’t clear that an expression of one of my I-languages can also be an expres-
sion of another one of my I-languages. And it isn’t clear that Lycan’s λ and λ′ are
distinct Human Languages, rather than arbitrarily distinguished sets of expres-
sions. Though let me return to this point after briefly discussing context sensi-
tivity and the idea that truth theories for Human Languages will have theorems
concerning utterances.
But as we’ll see, the two main strategies for accommodating sentences like (33)—
relativizing the truth of sentences to contexts, or specifying truth conditions
for sentential utterances—heighten the worry that no truth theory for a Human
Language H is the core of a correct meaning theory for H. Focusing on truth
favors the second strategy, while focusing on meaning favors the first. Though
either way, the net result is not good.
We sometimes speak as if ‘true’ might itself be the central predicate in a theory
of meaning for a Human Language, or at least for an invented language that pro-
vides a decent model of a Human Language. But this is a simplification, as Tarski’s
own semantics for the first-order predicate calculus reminds us. One can invent a
language in which (69)–(71) are sentences,
(69) ∀x∃x′Pxx′
(70) ∃x∀x′Pxx′
(71) Pxx′
and then provide a semantics with the following consequences: (69) is true if
everything precedes something, and otherwise (69) is false; (70) is true if some-
thing precedes everything, and otherwise (70) is false; (71) is neither true nor
false, but not because it has a neutral truth value. In a propositional calculus, the
truth or falsity of atomic sentences determines the truth or falsity of complex
sentences. But Tarski showed how to characterize truth, for a predicate calculus, in
terms of a more basic notion of satisfaction; where closed sentences like (69) and
(70) have satisfaction conditions that are specified in terms of the satisfaction
conditions of open sentences like (71).
Truth is satisfaction by all sequences. So one T-sentence for (69) is simply (69a).
If you like, add the syntactic requirement that only closed sentences are true. But
the work of the “truth” theory is done by the axioms that license derivations of
“S-theorems” like (69b);
where for any sequence F and variable i, F(i) is the ith element of σ, and for each
sequence F*, F* ≈i F if and only if F* differs from F at most with respect to the ith
element. Let variables be characterized recursively: ‘x’ is the first; and for each
variable α, α′ is the next.
168 Reflections on the Liar
One might wonder if any analogous S-theorems for sentences like (72) are
interpretive.
Even if (72a) is true, its right side may employ notions that are “foreign” to (72).
And given sentences like (73)–(75), whose meanings are not “firstorderizable,”
(76) It is here.
(33) I chased something.
As desired, sentences like ‘Lπpt’ are not true or false simpliciter, since they are
satisfied by some but not all assignments. But one can introduce a notion of truth
relative to contexts that select assignments. One can say that a context c selects
assignment A if and only if: A(s/p/t) is the speaker/place/time of c; and for each
pointer πi, A(πi) is the ith thing demonstrated in c. That makes room for (77),
according to which ‘Lπpt’ is true relative to some but not all contexts.
(76D) True(u, Lπpt) ≡ the thing demonstrated with the (first) deictic act
in u is located in the place of u at the time of u
(33D) True(u, ∃x(Bxt & ∃x′(Cxsx′)) ≡ some A* such that A* ≈x A is such
that A*(x) occurs before the time of u, and some A** such that A**
≈x′ A* is such that A**(x) is a chase of A**(x′) by the speaker of u
See Burge (1974), Larson and Segal (1995), Lepore and Ludwig (2005). For many
purposes, the two approaches are interchangeable. But here, I want to stress a
respect in which the K(aplan)-strategy and the D(avidson)-strategy reflect funda-
mentally different ways of thinking about how Human Languages and their users
are related to truth and contexts.
On the K-strategy, a theory of truth for a language is a theory of a language
in a sense familiar to logicians and grammarians: a set of well-formed formulae,
specified by some (typically recursive) procedure; or a physically realizable pro-
cedure that generates expressions of some kind from finitely many lexical items
and combinatorial principles. As discussed below, I follow Chomsky (1986) in
taking Human Languages to be procedures rather than sets. But the important
point is here that on the K-strategy, theories of truth are theories of linguistic
objects. Since the use of a sentence on a particular occasion determines which
170 Reflections on the Liar
assignments are licensed on that occasion, truth depends on use. But the context-
sensitive truth conditions of sentences are specified without reference to use. So
a K-style theory ascribes truth conditions to sentences.
By contrast, on the D-strategy, a truth theory for a language turns out to
be a theory of certain actions of using the language. Such theories ascribe truth
conditions to spatiotemporally located utterances of generable sentences. In my
view, this strategy is misguided, and not only because it is hard to formulate axi-
oms from which biconditionals concerning utterances follow, given a weak logic
that avoids uninterpretive theorems. (While Larson and Segal [1995] endorse
a D-strategy in their text, their derivations of theorems usually employ the K-
strategy.) Given the independent need for sequence relativization, one might
hope for a theory that delivers theorems like (76K) and (33K) above; these are
hard enough to derive. But more importantly, specifying meanings in terms of
utterances seems to confuse linguistic competence with language use.
That said, the D-strategy can seem reasonable in other respects. With regard to
Human Languages, it may well be that truth is more plausibly ascribed to certain
actions of using sentences than to sentence-context pairs. Indeed, “contingent liar
sentences” like (66)
uses of sentences. But the sentences used need not be truth-evaluable. For some
purposes, it does no harm to simplify and say that Human Language sentences
are true or false relative to contexts. We can also speak of “propositions” that
are unrelativizedly true or false; see, e.g., Cartwright (1962). Given episodes of
thinking, we can speak of thoughts thought; given episodes of asserting/endors-
ing/judging, we can speak of propositions asserted/endorsed/judged. But it
doesn’t follow that truth-evaluable things are truth-evaluable by virtue of hav-
ing propositional contents, much less that theories of meaning should specify
such contents.
If truth is downstream from meaning—in that certain uses of meaningful
expressions are candidates, subject to review, for being true or false—then good
theories of meaning for Human Languages may not deploy the notion of truth
even if humans enjoy a mental predicate true(_) that applies to certain events
or mental sentences. That leaves the hard task of describing the context sensi-
tive conditions on being true or false, and explaining why these are conditions
on truth/falsity; though see Glanzberg (2004, 2015) for a good start. In any case,
one can share Davidson’s suspicion that any plausible theory of truth for a Human
Language will have theorems that specify properties of spatiotemporally located
actions of using expressions, while also suspecting that a plausible theory of
meaning for a Human Language will have theorems that specify use-independent
properties of expressions; where these properties constrain how the expressions
can be understood, and hence how the expressions can be naturally used.
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
7.3.1. I Before E
Let’s be very generous, and say that a language associates signals of some kind
with interpretations of some kind. An expression of a language pairs a signal
(type) with an interpretation (type). An expression might pair a certain sound
or gesture or inscription with a certain object or property or concept. Sounds, of
172 Reflections on the Liar
But (79) is homophonous even if we hold the meaning of ‘duck’ fixed. It can mean
that duck was ready to dine, or that the duck was prêt-a-manger. By contrast, (80)
has the meaning indicated with (80a) but not the equally coherent meaning indi-
cated with (80b).
And while (82) is ambiguous, it fails to have the third meaning indicated
with (82c).
reasons for identifying Human Languages with any alleged sets of expressions
that might be generated in various ways. In thinking about whether such reasons
are likely to emerge, it is useful to follow Chomsky (1986) in applying Church’s
(1941) intensional/extensional distinction—regarding mappings from inputs to
outputs (compare Frege [1892b])—to the study of Human Languages and the
aspects of cognition that support acquisition of these languages.
We can think of functions as procedures (intensions) that determine outputs
given inputs, or as sets (extensions) of input-output pairs. Consider the set of
ordered pairs <x, y> such that x is a whole number, and y is the absolute value of
x–1: {… (–2, 3), (–1, 2), (0, 1), (1, 0), (2, 1) …}. This set can be characterized in
many ways. Using the notion of absolute value, one can say that F ( x ) = x − 1 .
One can also use the notion of a positive square root: F(x) = +√(x2 –2x + 1). These
descriptions of the set correspond to distinct procedures for computing values
given arguments. And a mind might be able to execute one algorithm but not the
other. In Church’s idiom: we can use lambda-expressions to indicate functions-
in-intension, with λx.|x –1| and λx.+√(x2 –2x + 1) as distinct but extensionally
equivalent procedures; or we can use lambda-expressions to indicate functions-
in-extension, saying that λx.|x –1| is the same set as λx.+√(x2 –2x + 1). Though
as Church stressed, specifying the space of computable functions requires talk of
procedures.14
Echoing this point, Chomsky contrasted I-languages with E-languages: an I-
language is a procedure that pairs signals with interpretations; an E-language is a
language in any other sense—e.g., a set of signal-interpretation pairs, or a cluster
of dispositions to make certain utterances. Note that to even specify a set with
endlessly many elements, one must somehow specify a procedure that determines
the set. (In the previous paragraph, I was able to talk about the infinite set only
because the context included reference to a determining procedure.) And while
many procedures determine sets—arithmetic procedures defined over numbers
being paradigm cases, along with invented procedures that generate sets of well-
formed formulae—a biological system might pair pronunciations with meanings,
yet not determine any set of expressions, if only for lack of a fixed domain of
inputs.
Consider (83).
Speakers of English know that this string of words is defective, but meaning-
ful. In particular, speakers hear (83) as having the interpretation of (83a) and
not (83b);
see Chomsky (1965), Higginbotham (1985). The defect does not preclude under-
standing (83), which is neither word salad like (84)
And in terms of the discussion in section 7.2.3, it seems implausible (to me) that
a theory of truth for an I-language will have theorems concerning utterances. This
seems to confuse the distinction that Chomsky (1986) was emphasizing when he
contrasted questions concerning the linguistic knowledge/competence that chil-
dren acquire with questions concerning how that knowledge/competence is put to
use. Given (C), it also strikes me as optimistic to think that Human (I-)Language
sentences—expressions that children can generate and comprehend—are them-
selves truth-evaluable. Like many philosophers in the Rationalist tradition, I find
it amazing that humans can make any claims that rise to the level of falsity: once
a thinker achieves this level of clarity and contact with her environment, truth is
176 Reflections on the Liar
a mere negation sign away. Obviously, humans can often use their I-languages to
make truth-evaluable assertions and judgments. That’s a wonderful thing, but it
may require many ancillary cognitive capacities (and in many domains, hard cog-
nitive work); see Pietroski (2010, forthcoming).
One can hypothesize that humans enjoy an ancillary system that specifies truth
conditions of uses of I-language expressions that already pair pronunciations with
meanings, which are to be specified truth-theoretically via some version of the
Kaplan-strategy for accommodating context sensitivity. This interesting hypoth-
esis embraces (C). But on this view, one can’t deal with liar sentences by saying
that truth and falsity are really evaluative properties of spatiotemporally located
assertions/judgments, and that theories of truth for Human Languages assign
semantic properties to utterances rather than sentences relativized to contexts.
is not an expression of λ, and hence is not true-in-λ.15 Though given that Linus
is not true, or least not true in λ or λ′, (54a) is true-in-λ′, true-in-λ″, etc. If λ and
λ′ are sets of expressions, with λ as a proper subset of λ′, this proposal is coher-
ent: whatever (4) is, it can be an element of both sets while (54a) is an element
of λ′ but not λ. But if λ and λ′ are Human I-Languages, then it isn’t clear how the
suggestion is supposed to work.
Suppose that (4) and (54a) are generable expressions—certain pronunciation/
meaning pairs—and that these expressions are generated by some procedures but
not others. Let λ be a procedure that generates (4) but not (54a). Let λ′ be a proce-
dure that generates both. So far, this is just an intensional version of the idea that
λ is a proper subset of λ′. But if children who “acquire English” acquire both λ and
λ′, then it is hard to see what λ can be apart from an arbitrarily restricted variant
of λ′: the same generative procedure, apart from some lexical items included in
the specification of λ′. And it is very hard to see how λ′ can be an I-language dis-
tinct from λ″, which allegedly includes a truth predicate that is pronounced ‘true’
I-Lang uag e s and T- Se nte nc e s 177
but not a lexical item of λ′. (It is very very hard to see how λ″ can be an I-language
distinct from λ′′′, and so on.)
Lycan is surely right that ‘English’ is not a name for any particular Human
Language. Even a single speaker of Standard American English may have acquired
various I-languages that are similar but not identical for communicative or socio-
political purposes. A person can be monolingual in the ordinary sense, yet have
several overlapping I-languages. Likewise, as a typical American child acquires a
mature version of English, she acquires various I-languages that differ from those
of local adults. For a while, her I-language(s) will differ grammatically from “adult
English” in ways that may be striking. But soon enough, the child will acquire
at least one I-language that corresponds to adult competence, apart from the
absence of certain (open-class) lexical items.
It is unclear when children acquire semantic vocabulary, in part because one
can have lexical items without articulating them. It is also unclear what counts
as semantic vocabulary. Do words like ‘correct’ count? What about “factive” verbs
like ‘know’? But suppose we had a characterization of the troublemaking words
that must excluded from the lexicon of a Lycan-style “core” language that does not
itself generate liar sentences. It still doesn’t follow—and it isn’t plausible—that
speakers who understand (4) and (54),
have one I-language that generates only (4) and another I-language that gener-
ates (4) and a suitably subscripted analog of (54). We can imagine a child who
hasn’t acquired the word ‘true’, but who does have an I-language λ that generates
(4). Later, the child may have an I-language λ′ that also generates (54). We might
describe this change by saying that the child added some mental analog of (86) to
an expandable list of semantic axioms.
Or we might say that the child replaced λ with a extended language λ′. But why
think that when the child has λ′, she also has λ? Once the child has λ′, it seems that
(for her) λ is an arbitrarily defined object. Indeed, λ seems gruesome: λ′ except for
lexical items acquired after some time t.
Theorists are free to describe a lexicon L and a combinatorial system C as the
result of adding one or more lexical items to an I-language specified in terms of
C and a slightly smaller lexicon L–. One can then introduce, by stipulation, exten-
sionally distinct notions like true-in-CL and true-in-CL–. But whatever seman-
ticists choose to do, and however they choose to talk, the empirical questions
remain: is there a Tarski-style theory of truth for the generative procedure CL;
and if so, is some such theory the core of a correct theory of meaning for CL?
178 Reflections on the Liar
Certain arguments for a negative answer will not apply to CL– if the “smaller” lan-
guage does not generate the apparent counterexamples. But that does not make
the restriction principled. (Clauses like ‘thinks he snores’ present concerns even if
some languages have no such clauses.)
Adding ‘true’ to an I-language seems no different than adding ‘snores’ or
‘Linus’. And if λ′ is a generative procedure, then prima facie, a truth theory
for λ′ is also a truth theory for any generative procedure λ that is just like λ′
except for having a slightly smaller lexicon. So to repeat, why think that when
our imagined child acquires the word ‘true’, she both acquires a slightly more
productive procedure and retains the slightly less productive procedure? One
needs reasons, independent of (D), for thinking that Human I-languages exhibit
Lycan’s hierarchy.
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
(63) ‘Linus is not true.’ is true if and only if Linus is not true
(63a) True(‘Linus is not true.’) ≡ ~True(Linus)
can be a correct theory of meaning for English. The idea is that (63) and the axi-
oms it follows from can be interpretive, even if the axioms—and endlessly many
theorems like (63)—are false. In which case, a truth theory Θ doesn’t have to be
true to be a correct theory of meaning for a Human Language H. It’s good enough
if Θ generates an interpretive T-theorem for each sentence of H, given a back-
ground logic that does not also yield uninterpretive theorems; compare Eklund
(2002). I am suspicious of the idea that false theories of truth can be correct theo-
ries of meaning. In this respect, I appreciate the motivations for Lycan’s strategy,
even if my conclusion is that (D) is the troublemaking thesis that we should reject.
But given my objections to Lycan’s strategy, let me briefly address Lepore and
Ludwig’s defense of (D).
While they don’t explicitly reject thesis (C),
my sense is that Lepore and Ludwig have some other conception of what Human
Languages are. Or perhaps they hope to remain neutral on this score. But in my
view, (C) is more plausible than any extant defense of (D). Advocates of (D) may
I-Lang uag e s and T- Se nte nc e s 179
not want to assume (C). But I see no reason not to assume (C) when evaluating
(D). So let’s consider the very bold thesis (CDF),
provide the overall best translations of the object language sentences, given cer-
tain restrictions on the metalanguage. One can argue about how to character-
ize the relevant notion of translation (or “samesaying”). But my concern is more
basic. Absent an independently motivated condition on being interpretive, I don’t
see why a set of inconsistent axioms should be regarded as a correct theory of
meaning for a Human Language just because the axioms are better than other
truth-theoretic axioms with regard to yielding theorems that can be used to
translate sentences that a child could understand into sentences of an invented
metalanguage.
Given a consistent theory that recursively pairs each English expression E with
a plausible S-theorem of the form shown in (87),
I can see why one might say that the theory is meaning-specifying. Even if the
theorems involve considerable formalism, one can argue that the apparent “mate-
rial adequacy” of the axioms is best explained by the hypothesis that the axioms
do indeed capture the core semantic properties of English expressions—at least
in the absence of any alternative explanation for why the theorems seem to be
true. Correlatively, one might think that if at least one (suitably formulated) the-
ory of truth for English is true, that itself is an argument that the core semantic
explananda concern Tarski-style relations that linguistic expressions bear to the
environments that speakers share. But one needs independent reasons for think-
ing that false theorems capture explananda. So even if the false theorems (of a
false theory) are “interpretive” in some technical sense, why take that to be evi-
dence that the theorems (and the axioms) are meaning-specifying? Perhaps the
180 Reflections on the Liar
Lepore and Ludwig defense of (D) is the best one available. But if so, that may tell
against (D).
with the idea that speakers of H represent the axioms of a truth theory for H.
Perhaps speakers of English tacitly endorse a false theory whose theorems include
liar T-sentences like (63a).
then Θ is plausibly viewed as a theory of meaning, even if it has some false axi-
oms/theorems. But if the best theory of truth for H differs from the best theory of
understanding for H, the latter may be the better candidate theory of meaning for
H. Indeed, one might think that (D)
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
is attractive only insofar as it remains plausible that a truth theory for a Human
Language can serve as a theory of how expressions of that language are under-
stood by competent speakers; see Dummett (1976). And given liar sentences, per-
haps the best way to defend (D) is by defending a cognitivist version of (CDF). But
this strategy comes with a justificatory burden. Why retain the idea that correct
theories of meaning/understanding for Human Languages will take the form of
truth theories if the best candidate theories have false axioms/theorems?
I-Lang uag e s and T- Se nte nc e s 181
One has to ask what explanatory role the appeal to truth—or a mental truth
predicate—plays in a theory according to which axioms like (40) and (86)
are interpretive because speakers have mental analogs of these axioms and
thereby misrepresent linguistic expressions as having certain truth-theoretic
properties. If a truth theory Θ implies that certain expressions of H are true
of certain things, but (on pain of contradiction) the expressions are not true of
those things, why think that speakers of H tacitly endorse Θ?
One can posit “error” theories. But note that (D) is usually embedded within
a broader view of the subject matter and methodology of semantics: complex
expressions of a human language have the properties ascribed by a certain truth
theory; competent speakers of the language often recognize that the expressions
have these properties; and clever theorists—who suitably control for various
complications when creating settings in which speakers are asked to assent or
dissent, given a sentence and a situation—can cite judgments as evidence for or
against proposed truth theories, because in suitably controlled settings, compe-
tent speakers are reliable judges of whether or not the truth conditions that sen-
tences have are met in various situations.
Advocates of (D) who reject this broader view, and grant that sentences of
a Human Language are not conditionally true or false as specified by a correct
Tarski-style theory, thereby concede much to critics of (D). So what justifies the
residual appeal to truth in (CDF)? Perhaps in the end, (CDF) will be part of the
best account of how humans understand expressions, and how our judgments can
be evidence for or against proposed theories of understanding. But recall that in
order to focus on sentences like (54), we set aside many other concerns about (D).
So if the response to (54) is that false truth theories can still serve as correct
meaning theories, one wants to hear more about how proposals regarding specific
constructions like (7) and (12)–(14)
If the reply is that (59) isn’t true, but that (CDF) is, the next question is obvi-
ous: why retain (D)? At this point, what motivates the alleged connection between
understanding and truth?
I don’t doubt that some theories of meaning are better than others. But the
question is how advocates of (CDF) can adjudicate between alternative theories,
yet plausibly maintain that the cited evidence is evidence that favors one truth
theory over another. This is not the place to explore other conceptions of mean-
ing. But axioms like (40)
1893). This requires an interesting logic, and presumably one that licenses many
inferences not licensed by the first-order predicate calculus. And for purposes of
figuring out what follows from a theory, mathematical implications usually count.
But not even arithmetic is reducible to “pure logic”; analysis, geometry, and topol-
ogy seem to go well beyond logic. Moreover, we often extend the notion of impli-
cation to include inferences that are not justifiable a priori, so long as the tacit
premises are beyond reasonable doubt for the purposes at hand.
In any case, as theorists, we can infer from (5T) to (34).
then this capacity cannot support extraction of (34) or other instances of (35);
where Γ is itself an extractable theorem. But as we’ve seen, this requires a simple
(and naive) notion of extraction that seems to be odds with the sophistication
required to keep semantic paradoxes at bay. In short, there is a tension between
(i) characterizing knowledge of meaning in terms of truth and (ii) supposing that
this knowledge is, unlike truth, closed only under very weak deductive principles.
In retrospect, it seems that Davidson (1967a) was remarkably brief about
paradoxes.
The semantic paradoxes arise when the range of the quantifiers in the
object language is too generous in certain ways. But it is not really clear
how unfair to Urdu or to Wendish it would be to view the range of
their quantifiers as insufficient to yield an explicit definition of ‘true-
in-Urdo’ or ‘true-in-Wendish’…. In any case, most of the problems of
general philosophical interest arise within a fragment of the natural
language that may be conceived as containing very little set theory.
(pp. 28–29)
then (49) is false. So prima facie, a theory of meaning for English should not
imply (49).
This leads to problems for thesis (D) that are not addressed by restricting the
range of quantifiers.
(D) for each Human Language H, some Tarski-style theory of truth for H
is the core of a correct theory of meaning for H
I’m not sure how Davidson was counting “problems of general philosophi-
cal interest.” But one question of interest to many philosophers, among others,
is whether (D) is true. So we should ask if (D) is plausible in light of sentences
like (54).
If not, then (D) may not be plausible for any Human Language that generates
(4) or (5).
Notes
1. For these languages, truth can also be characterized in terms of satisfaction, without treat-
ing sentences and predicates as devices that denote truth values (T or ⊥) and functions; see
section 7.2.3. But for other languages, a Tarski-style theory may characterize truth in terms
of truth values and context-sensitive denotation conditions: True(S, c) ≡ Denotes(S, T, c);
cp. Frege (1892a), Church (1941), Montague (1974). And for these purposes, let’s be gener-
ous about what contexts can be: sequences of entities, possible worlds, centered worlds,
situations, etc.
2. Compare: Sadie is a mare if and only if Sadie is a mature female horse.
3. One can grant that verbs are associated with “eventish” variables without granting that
verbs have Tarski-style satisfaction conditions; see Pietroski (2015, forthcoming).
4. For simplicity, let’s ignore “reporting uses” of the present tense, which introduces other
complications.
5. See Dummett (1976) on the need for a theory of “content” that meshes with a theory of
“force,” regardless of how the relevant notion of content is related to classical notions of
truth; see also Segal (1991), Lepore and Ludwig (2007), and Lohndal and Pietroski (2011)
and references there.
6. We can individuate expressions semantically, so that (30) is not obviously false at worlds
where the pronunciation of ‘Hesperus is Phosphorus’ has the meaning of (26). And we can
say that each side of (30) has the same content; cp. Stalnaker (1984). But then it seems that
meanings are not contents; see Pietroski (2006).
7. It’s even odder to then say, as Davidson (1986, p. 446) does, “[T]here is no boundary
between knowing a language and knowing our way around in the world.” If uninterpre-
tive T-sentences can be distinguished from theorems of meaning theories, then pace Quine
(1951), there seems to be a theoretically interesting analytic/synthetic distinction.
8. Some ambitions are unattainable, and some tasks are unduly modest. It may be hard to find
out how speakers represent truth conditions. But there may be evidence of various kinds;
see Evans (1982), Peacocke (1986), Davies (1987), Lidz et al. (2011). Heck (2007) offers a
different cognitivist gloss on (D), drawing on Higginbotham (1991) to suggest that one
overt truth theory for H is more interpretive than another if the former better reflects
how speakers of H use their semantic competence in communication. I have doubts about
emphasizing communication if Human Languages are I-languages in Chomsky’s (1986)
I-Lang uag e s and T- Se nte nc e s 187
sense. But this debate is intramural. Like Higginbotham and Heck, I think the best defense
of (D) is via some cognitivist gloss. Though in the end, cognitivism about meaning may be
more plausible than (D).
9. Let ‘FEV’ be a name for that other paper, and let ‘iff’ abbreviate ‘if and only if’. If Lari is
either true or false, then Lari is true iff the first numbered sentence in FEV is false. So if
Lari is true: the first numbered sentence in FEV is false; and since Lari is that sentence, Lari
is false. But if Lari is false: the first numbered sentence in FEV is false; and since Lari is true
iff the first numbered sentence in FEV is false, Lari is true. So if Lari is true or false, then
Lari is true and false. One can say that some claims are true yet contradictory; see Priest
(1979, 2006). But even if that is so, it isn’t yet a reason for concluding that Lari is both true
and false. The hypothesis that Lari is true or false is much less plausible than the implied
contradiction.
10. There are many ways in which ‘Linus’ could end up being a name for (2). But let’s explic-
itly introduce ‘Linus’ as a name for the second numbered sentence in this very chapter,
“I-Languages and T-Sentences.” Compare the introduction of ‘Julius’ as a name for whoever
invented the zipper; see Evans (1982). For many purposes, we could use ‘(2)’ instead of
introducing of a name whose pronunciation does not connote the thing named. But the
issue here concerns Human Languages. And ‘(2)’, an invented numeric description, is not a
name in any natural sense.
11. Cp. Davidson’s (1986) description of communication in terms of shared “passing” theories
that speakers deploy, in contexts, by adjusting “prior” theories that may never be used with-
out adjustments; see also Ludlow (2011).
12. For these purposes, instructions include strings of ‘1’s and ‘0’s used in a von Neumann
machine to access other such strings and perform certain operations on them: ‘0100110101’
might be executed by performing operation two (010) on the number (0) fifty-three (110101),
while ‘1101011010’ calls for operation six (110) on the number stored in (1) register twenty-
six (011010). And one can imagine operations like conjunction on accessible/generable con-
cepts. For further discussion, see Pietroski (2011), from which some of this subsection is
adapted.
13. See Chomsky (1965), Higginbotham (1985). For extended discussion in the context of
broader issues concerning language acquisition and the nature of Human Languages, see
Crain and Pietroski (2001), Berwick et al. (2011).
14. For Frege (1892b), functions as procedures are logically prior to the more set-like courses of
values of functions. Marr (1982) likewise distinguished “Level One” descriptions, of func-
tions computed, from “Level Two” descriptions of the algorithms employed to compute
those functions. And while Level One descriptions can have a certain primacy in the order
of discovery, making it fruitful to ask what a system does before worrying about how the
system does it, a function cannot be computed without being computed in some way. So
one must not confuse the methodological value of Level One descriptions, in characterizing
certain cognitive systems, with any suggestion that the corresponding extensions are them-
selves targets of inquiry.
15. Lycan speaks in terms of grammar. But even if ‘Linus is not true(-in-λ′)’ is ungrammatical
in λ′, whose only truth predicate is ‘true-in-λ’, it does not follow that λ′ does not assign the
truth condition of (54a) to the sound of ‘Linus is not true(-in-λ′)’; cp. ‘seems sleeping’.
16. Lewis followed Tarski in this respect; see Burgess (2008), who notes that Tarski’s use of
‘semantics’ was also stipulative and nonstandard.
References
Austin, J. 1961. Philosophical Papers. Oxford: Oxford University Press.
———. 1962. How to Do Things with Words. Oxford: Oxford University Press.
Berwick, R., et al. 2011. “Poverty of the Stimulus Revisited.” Cognitive Science 35: 1207–1242.
188 Reflections on the Liar
———. 2007. “Meaning and Truth-Conditions.” In D. Greimann and G. Siegwart, eds., Truth and
Speech Acts: Studies in the Philosophy of Language, 349–376. New York: Routledge.
———. 2011. Frege’s Theorem. Oxford: Oxford University Press.
Heim, I., and Kratzer, A. 1998. Semantics in Generative Grammar. Oxford: Blackwell.
Higginbotham, J. 1985. “On Semantics.” Linguistic Inquiry 16: 547–593.
———. 1991. “Truth and Understanding.” Iyyun 40:271–288.
Higginbotham, J., and R. May. 1981. “Questions, Quantifiers, and Crossing.” Linguistic Review
1: 47–80.
Kaplan, D. 1978a. “Dthat.” In P. Cole, ed., Syntax and Semantics, 221–243. New York: Academic Press.
———. 1978b. “On the Logic of Demonstratives.” Journal of Philosophical Logic 8: 81–98.
———. 1989. “Demonstratives.” In J. Almog, J. Perry, and H. Wettstein, eds., Themes from Kaplan,
481–563. New York: Oxford University Press.
Karttunen, L. 1977. “Syntax and Semantics of Questions.” Linguistics and Philosophy 1: 3–44.
Katz, J. 1981. Language and Other Abstract Objects. Totowa, NJ: Rowman and Littlefield.
Katz, J., and J. Fodor. 1963. “The Structure of a Semantic Theory.” Language 39: 170–210.
Katz, J. and Postal, P. (1964). An Intergrated Theory of Linguistic Description. Cambridge MA: MIT Press.
Kleene, S. 1950. Introduction to Metamathematics. Princeton, NJ: Van Nostrand.
Kripke, S. 1975. “Outline of a Theory of Truth.” Journal of Philosophy 72: 690–716.
———. 1980. Naming and Necessity. Cambridge, MA: Harvard University Press.
Larson, R., and G. Segal. 1995. Knowledge of Meaning. Cambridge, MA: MIT Press.
Larson, R., and P. Ludlow. 1993. “Interpreted Logical Forms.” Synthese 95:305–355.
Lepore, E., and K. Ludwig. 2005. Donald Davidson: Meaning, Truth, Language, and Reality.
Oxford: Clarendon Press.
———. 2007. Donald Davidson’s Truth-Theoretic Semantics. Oxford: Oxford University Press.
Leslie, S. 2007. “Generics and the Structure of the Mind.” Philosophical Perspectives 21: 375–403.
Lewis, D. 1970 “General Semantics.” Synthese 22: 18–67.
———. 1975. “Languages and Language.” In K. Gunderson, ed., Minnesota Studies in the Philosophy
of Science, 7:3–35. Minneapolis: University of Minnesota Press.
———. 1986. On the Plurality of Worlds. Oxford: Blackwell.
Lidz, J., et al. 2011. “Interface Transparency and the Psychosemantics of ‘Most’.” Natural
Language Semantics 19: 227–256.
Lohndal, T., and P. Pietroski. 2011. “Interrogatives, Instructions, and I-languages: An I-Semantics
for Questions.” Linguistic Analysis 37: 459–510.
Ludlow, P. 2011. The Philosophy of Generative Linguistics. Oxford: Oxford University Press.
Lycan, W. 2012. “A Truth Predicate in the Object Language.” In G. Preyer, ed., Donald Davidson on
Truth, Meaning and the Mental, 127–147. Oxford: Oxford University Press.
Marr, D. 1982. Vision. San Francisco: Freeman.
Montague, R. 1974. Formal Philosophy. New Haven: Yale University Press.
Parsons, C. 1974. “The Liar Paradox.” Journal of Philosophical Logic 3: 381–412.
Peacocke, C. 1986. “Explanation in Computational Psychology: Language, Perception and Level
1.5.” Mind and Language 1: 101–123.
Pietroski, P. 2005. “Meaning before Truth.” In G. Preyer and G. Peters, eds., Contextualism in
Philosophy, 253–300. Oxford: Oxford University Press.
———. 2006. “Character before Content.” In J. Thomson and A. Byrne, eds., Content and
Modality: Themes from the Philosophy of Robert Stalnaker, 34– 60. New York: Oxford
University Press.
———. 2008. “Minimalist Meaning, Internalist Interpretation.” Biolinguistics 4: 317–341.
———. 2010. “Concepts, Meanings, and Truth: First Nature, Second Nature and Hard Work.”
Mind and Language 25: 247–278.
———. 2011. “Minimal Semantic Instructions.” In C. Boeckx, ed., Oxford Handbook of Linguistic
Minimalism, 472–498. New York: Oxford University Press.
———. 2015. “Framing Event Variables.” Erkenntnis 80:31–60.
———. Forthcoming. Conjoining Meanings: Semantics without Truth Values.
Priest, G. 1979. “Logic of Paradox.” Journal of Philosophical Logic 8: 219–241.
190 Reflections on the Liar
As this passage shows, Tarski was aiming for an outright definition of truth
that could be applied to all the sentences of a given language, but he would have
imposed the same requirement on any theory that aspires to articulate the clas-
sical conception of truth as it applies to a language. A theory of truth as it applies
to the sentences of English, then, must imply the biconditional just displayed—
along with many others, of course. Tarski calls any biconditional of the form
191
192 Reflections on the Liar
to put into precise form the conditions under which we will consider
the usage and the definition of the term ‘true’ adequate from the mate-
rial point of view: we wish to use the term ‘true’ in such a way that all
equivalences of the form (T) can be asserted, and we shall call a defini-
tion of truth ‘adequate’ if all these equivalences follow from it. (Tarski
1944, sec. 4)
The sentence printed on lines 18 and 19 of page 192 of this book is not
true
and let S be a name of the token sentence just displayed. If we were entitled to
assert the T-sentence for S, then we would be entitled to assert
which is “an obvious contradiction” (Tarski 1944, sec. 7). For Tarski, then, the
significance of the Liar is that it reveals a latent contradiction in what he takes to
be the natural way of spelling out the intuitive, Aristotelian conception of truth.
On Tarski’s diagnosis, the source of the contradiction is the implicit assump-
tion “that all the sentences which determine the adequate usage of the term
The Liar w ithout Tr uth 193
false as potentially uttered by you at noon. With that point noted, we might gloss
Aristotle’s dictum as follows:
Similarly,
The theory of truth and falsity that comprises (T), (F) and their logical conse-
quences is formally consistent. As Timothy Williamson points out, we can show
this “by constructing an unintended model … in which formulas are treated as
referring to truth-values, the propositional quantifiers range over truth-values,
and all formulas of the forms ‘Say (A, c, P)’, ‘True (A, c)’, and ‘False (A, c)’ are
treated as false” (Williamson 1998, p. 14).
How, though, does the present elaboration of Aristotle’s dictum avoid the con-
tradiction that crippled Tarski’s? Let us revert to our paradoxical sentence S, as
potentially used in its actual context of inscription c*. We can be sure that if S says
anything in c*, it says that the sentence printed on lines 18 and 19 of page 192 of
this book is not true in c*. That is the only thing it could say, given what its com-
ponent words mean in the actual context of inscription. On the assumption that
S says something in c*, then, (T) yields
(1′) True (S, c*) ↔ the sentence printed on lines 18 and 19 of page 192 of
this book is not true in c*.
(2′) S = the sentence printed on lines 18 and 19 of page 192 of this book
Again as before, (3′) is a contradiction, but there is a crucial difference with the
earlier deduction. In the earlier case, we were able to assert (1) because (1) is a
T-sentence and Tarski takes it that an adherent of Aristotle’s conception of truth
will be willing to assert any T-sentence. By itself, however, our new articulation
of that conception, namely (T), does not yield (1′). It yields it only given the addi-
tional assumption that S, when uttered in c*, says something. So it is open to the
defender of (T) to take (3′), not to reveal any problem with (T), but to refute this
additional assumption. In other words, the deduction shows that S does not say
anything, when uttered in c*. That is, the deduction proves
No doubt this is initially surprising but, given the truth of (T), it appears to follow
ineluctably. On this view, “[t]he semantic paradoxes are transformed into sound
arguments for constraints on what can say what in what contexts” (Williamson
1998, p. 15).
196 Reflections on the Liar
truth—a large philosophical question if ever there were one. In fact, though, the
style of quantification to be found in (T) itself suffices to generate Liar-like para-
doxes. Paradoxical as it may sound, the Liar Paradox has nothing essentially to do
with truth.
To show this, let me introduce my colleague Professor Brainstorm. Brainstorm
has days when he forgets that he is Brainstorm and, on some of these days, he
denounces Brainstorm as a peddler of untruths. In making these denunciations,
he sometimes says ‘Nothing Brainstorm is now saying is true’. This sentence—the
so-called Strengthened Liar, because it attributes lack of truth, rather than false-
hood, to itself—involves a truth predicate, and is no better for our purposes than
S. However, some of Brainstorm’s unwittingly self-directed denunciations have
a different form. In his youth, Brainstorm had the good fortune to study logic
with Prior, who taught his pupils to use ‘However things may be said to be’ and
‘There is a way things may be said to be’ as devices of sentential quantification.
Brainstorm remembers this, and at noon on day d we duly hear him utter the fol-
lowing English sentence:
(L) There is no way things may be said to be such that Brainstorm is now
saying that they are thus, and they really are thus.
The letter ‘L’ is my label for the type sentence that Brainstorm then utters; I shall
also use the symbol ‘λ’ to stand for the particular utterance of L that Brainstorm
produces at noon on day d.5
The question is whether λ expresses a proposition—that is, whether, in pro-
ducing λ, Brainstorm says that anything is the case.
L is a well-formed type sentence of English. Indeed, L is a type sentence that an
English speaker could use, on a favorable occasion, to say something true. Thus,
if I utter L at a time when the only thing Brainstorm is saying is ‘2 + 2 = 5’, I shall
thereby say something true. These facts about L, though, do not entail an affir-
mative answer to our question; they do not show that the particular utterance λ
says that such-and-such is the case. An utterance of a grammatical type sentence
can fail to say anything, even though other utterances of the same type express
truths. Philosophers of language have advanced various examples to illustrate
this general point; perhaps the most persuasive involve demonstrative expres-
sions. You dye an elephant pink, thereby enraging it. In these circumstances your
utterance of
says something true. Consider, though, another utterance, µ, of the very same
type sentence M. Utterance µ is produced by Dr Lush, the departmental dip-
somaniac, as he points into thin air. There are strong arguments for the thesis
that, if an utterance of M says anything at all, it expresses a ‘de re thought’—a
198 Reflections on the Liar
thought that exists (i.e., is available to be thought) only if an elephant (or some
other large animal) is identifiable in the context of utterance (see Evans 1982,
esp. chap. 6, and McDowell 1982). In the context of µ, though, there is no
such animal: Lush is hallucinating. Accordingly, µ says nothing; it expresses no
proposition.
Our question about λ, then, is not settled affirmatively by the fact that other
utterances of L express truths. It seems, indeed, as though logical deduction suf-
fices to settle it negatively. For suppose that λ does express a proposition. Given
what L means, the only proposition that λ could express is the proposition that
there is no way things may be said to be such that Brainstorm is now saying that
they are thus, and they really are thus. Let us call this the proposition that P, for
short. Our supposition that λ says something, then, amounts to the supposition
that λ says that P. Suppose further that P. Then there is a way things may be said
to be—viz., the way they are said to be when someone says that P—such that λ
says that they are thus, and they are in fact thus. This is contrary to P, so we may
infer that it is not the case that P. That is, there is a way things may be said to be
such that Brainstorm is now saying that they are thus, and they really are thus.
Since the only thing that Brainstorm could be saying is that P, it follows that it
is the case that P, after all. This contradiction appears to reduce to absurdity the
initial supposition that λ says something, so it seems we may conclude that λ
expresses no proposition.6
Although apparently reached by rigorous logical deduction, this conclusion is
not the end of the matter. The conclusion is that λ expresses no proposition, i.e.,
that there is no way things may be said to be such that λ says that they are thus.
Given that λ is Brainstorm’s utterance at noon, it would appear to follow that
there is no way things may be said to be such that Brainstorm says at noon that
they are thus, and hence a fortiori that there is no way things may be said to be
such that Brainstorm says at noon that they are thus, and they really are thus.
But in reaching that conclusion we appear to have expressed a proposition—
indeed, a true proposition—using the very words that, in Brainstorm’s mouth,
say nothing. The claim that λ says nothing, then, generates a classic revenge para-
dox. We have at the least to explain how it can be that when we—the commenta-
tors on the paradoxical situation—utter the type sentence L we say something
(indeed, say something true), whereas when Brainstorm utters L he fails to say
anything at all.
This way of taking revenge confirms the view that this version of the Liar is
centrally about the limits of what can be said, and is only derivatively about truth.
Accepting that view, though, does not appear to bring us any closer to a solu-
tion. Let us grant for a moment that in uttering λ, Brainstorm does not succeed
in saying anything; even so, the revenge reasoning still appears to ensnare us in
paradox. In these matters, however, appearances can deceive. Next, we need to
formalize the reasoning of the previous two paragraphs, and check whether each
step is valid.
The Liar w ithout Tr uth 199
In these terms, the assumption that λ says something may be formalized as ∃RδR .
We may, then, formalize our argument that ∃RδR yields a contradiction as follows:
This derivation spells out our informal reductio of the assumption that
Brainstorm, at noon, says that things are thus-and-so, given as premise the
200 Reflections on the Liar
specification of what he would then have said if he had succeeded in saying any-
thing. It is noteworthy that all the rules of inference appealed to in the deduction
are acceptable to an intuitionist, as well as to a classical logician.
So far, then, we appear to have a proof that Brainstorm’s paradoxical utterance
λ says nothing, even though it is the utterance of a type sentence that could in
other circumstances be used to express a truth.7 The problem, however, is that the
logical rules used in reaching this conclusion may be further applied to show that
things are as Brainstorm says them to be, thereby ensnaring us in the revenge
paradox. For, as it seems, our derivation may be validly extended as follows:
Line 25 is just our formalization of Brainstorm’s sentence L. So, it seems, the argu-
ment that shows that Brainstorm says nothing may be extend to establish the truth
of what he said. (Again, the extension uses logical rules that are acceptable to an intu-
itionist.) In other words, our treatment faces an acute form of the Liar’s Revenge.
Line 25 of the extended derivation does not produce a formal contradiction.
Rather, the problem it presents is that we seem to have derived a formula that
says precisely what Brainstorm’s utterance attempted but failed to say. This is
puzzling and calls for explanation. We expect the conclusion of a logically correct
derivation from a true premise to be true, but if line 25 expresses a truth, it must
say something. And then the puzzle is: how come we succeed in saying something
when we write down line 25, whereas Brainstorm fails to say anything when he
utters precisely the same sentence (or formula) with precisely the same meaning?
There would seem to be only two ways of dealing with this problem. The first would
lay the puzzle to rest by identifying a semantically relevant d ifference between
Brainstorm’s utterance and the inscription of line 25—a difference sufficient to
account for the fact that his utterance expresses nothing while the inscription
expresses a truth. The second would be to identify some flaw in the derivation of
line 25, so that we are not, after all, committed to asserting the sentence which,
in his mouth, says nothing. I explore these two strategies in turn.
one might at first think to ensure that such a shift always provides a way out of
paradox.
What we need is for line 25 to express a proposition that Brainstorm is unable
to express by producing λ—presumably because he is not in a position to express
that proposition at all. We should readily grant that the range of ‘ways things
may be said to be’ over which Brainstorm quantifies when he utters L at noon
may well fail to include the way line 25 says things are: the range of a quantifier
is often restricted, and the restriction operative in λ might leave what line 25
says out of range. We shall not have a resolution of the revenge paradox, though,
unless the way line 25 says things are is always out of the range of Brainstorm’s
sentential quantifier, and it is mysterious what prevents Brainstorm from speci-
fying the range of his quantifier in such a way that it does cover whatever line
25 might express. We normally take a quantifier’s range to be determined by the
relevant speaker’s intentions. Let us imagine, then, that Brainstorm glosses L as
follows: “In saying ‘There is no way things may be said to be …’, I really do mean
‘no way’. That is, no matter how people’s capacities for saying that things are thus-
and-so may be expanded, I maintain that there will never be a way things may be
said to be such that Brainstorm says they are that way when they actually are that
way.” Yet, if Brainstorm were to gloss his utterance of L by saying all that, it would
seem that anything that line 25 might express will fall within the scope of L, and
we shall be back in paradox.
It will not do simply to insist that some factor or other must block
Brainstorm’s quantifier from reaching so far, on the grounds that if it did we
should be unable, on pain of contradiction, to recognize that line 25 expresses
a truth. To adopt a useful phrase of Dummett’s (which he coined in a related
connection), simply to insist on this would be “to wield the big stick” when
what is wanted is an explanation. What we need to find are principles regulat-
ing the range of a sentential quantifier which are well supported by the linguis-
tic evidence in more straightforward cases, and which explain why Brainstorm
cannot quantify as widely as his gloss shows that he intends to.
An intriguing suggestion in this direction was put forward by Charles
Parsons (1974) in the paper which first developed the present approach to the
Strengthened Liar with modern logical rigor. As Parsons notes, we need to per-
suade ourselves that an inscription such as line 25
means that the sentential quantifiers in lines 21 to 25 range more widely than do
the corresponding quantifiers in lines 1 to 20.
There are, however, two problems with this claim. First, it is quite unclear why
the relation of expressing a proposition, or of saying that such-and-such, only
becomes salient at line 20. That line, Glanzberg notes, is “the first point in the
proof where there are no undischarged [assumptions]. Hence, I suggest, it is the
point where the relation [of expression] is accepted as salient in the discourse”
(2004, p. 39). As far as I can discern, though, Glanzberg gives no argument to jus-
tify the “hence,” and without an argument, his claim is pretty implausible. After
all, the notion of expression—implicit in our symbol ‘δ’—is central to the deduc-
tion from its first appearance in premise (1)—and so, one would have thought, it
is salient right from the outset. Second, if the range of the sentential quantifiers
really does expand immediately after line 20, then the argument from 21 to 25 is
fallacious. If the range of quantification has expanded between lines 20 and 21,
then the occurrence of ∃RδR at line 24 is entirely consistent with the occurrence
of ¬∃RδR at line 20, just as ‘There is a red book’ (uttered in a context where the
quantifier’s range extends over the whole house) is entirely consistent with ‘There
is no red book’ (uttered in a context where the quantifier’s range is restricted to
things in the music room). Accordingly, we cannot apply reductio to reach line 25,
and we then have, as far as I can see, no reason to judge that line 25 is true. Yet
Glanzberg’s project is precisely directed to explaining how lines 20 and 25 can
both be true.
There are, indeed, more general worries about Glanzberg’s attempted solution.
He shows how the range of a restricted sentential quantifier can expand as new
relations become salient. He does not show how, or why, a maximally inclusive
domain of sentential quantification—perhaps corresponding to a maximally
inclusive range of salient relations—is illegitimate. But if such a maximally inclu-
sive domain is legitimate, then we have no solution to the revenge paradox that
we are confronting.
Second, even if a wholly unrestricted domain of sentential quantification is
illegitimate, some restricted domains still cause trouble. Let us adjust the inter-
pretation of ‘δP’ so that it now means ‘At noon on day d, Brainstorm utters a
sentence or formula which Rumfitt will interpret as expressing the thought that
P’. Let us again suppose that Brainstorm utters ¬∃P(δP ∧ P ) at noon on day d. As
before, the first twenty lines of the derivation yield ¬∃RδR . Thus ¬∃P(δP ∧ P ) ,
as I interpret it, says nothing. At line 25 of the derivation, though, I affirm the
very same formula, ¬∃P(δP ∧ P ) . That is, I am led in the course of the whole deri-
vation to affirm a formula that, as I interpret it, says nothing. Even if something
stops us from interpreting Brainstorm’s quantifiers as absolutely unrestricted, on
the present interpretation of ‘δP’, it is hard to see why ‘P’ should not range over all
the ways I might say that things are.
Third, even if we find some reason to say that the sentential quantifier
Brainstorm utters cannot be interpreted as ranging over the proposition expressed
204 Reflections on the Liar
by writing line 25, there remains a problem in the realm of thought. Brainstorm,
it will be recalled, glossed L by insisting that ‘There is no way things may be said to
be …’ really does mean ‘no way’. That is, he clearly intends to quantify over any
proposition that line 25 might express. So, even acknowledging that he cannot
quantify as widely as he wants in language, he can still do so in thought. But then
we can construct variant paradoxes using propositional attitudes instead of indi-
rect speech. Thus suppose Brainstorm believes that there is no way things could be
believed to be such that he believes things are thus, and they really are thus. Then
we can prove, as before, that this belief—or, better, this pseudobelief—is without
content while also proving the content it would have, if it had any. Important as
they may be for some purposes, principles regulating the interpretation of quan-
tifiers in our languages do not get to the heart of the present paradox.10
8.5. A Way Out?
In the previous section, I examined and found wanting two of the attempts that
logicians and philosophers have made to reconcile the truth of ‘comments’ on
paradoxical utterances—comments such as our line 25—with the presumed
vacuity of the paradoxical utterance itself. While other attempts would need to be
scrutinized, it is clear that the contextual approach faces serious problems. In this
concluding section, I turn to consider, all too briefly, the second of the two ways
of addressing the revenge problem that I distinguished at the end of section 8.3.
In a chapter that distills decades of reflection on logical and semantical para-
doxes, J. L. Mackie considers a case very similar to that of Brainstorm (Mackie
1973, chap. 6). He invites us to consider a remark by the Cretan Epimenides, who
has forgotten his nationality: ‘No sentence uttered by a Cretan, standardly con-
strued, makes a true statement’. Mackie continues:
Let us assume, what might be the case, that no other sentence uttered
by a Cretan, standardly construed, makes a true statement. We cannot
without contradicting ourselves allow that Epimenides’ remark makes
a true statement. And yet if it fails for whatever reason to make a true
statement, we must ourselves say exactly what Epimenides has said;
how then can we deny that this is a sentence uttered by a Cretan which,
standardly construed, makes a true statement? How can we avoid con-
tradicting ourselves? This example … seems to be by far the toughest
version of the liar. (Mackie 1973, p. 294)
Mackie, we may assume, would say the same about Brainstorm’s utterance, λ.
Like Epimenides’s remark, λ does not succeed in saying that things are thus-and-
so: there just is no how λ says things to be. So the conclusion reached at line 20 of
our derivation is correct. However, “[w]e cannot now drive a wedge between what
we say about that sentence [in our case: what we say about λ] and what we allow it
to say about itself: standardly construed, it would say just what our own use of the
same type sentence would” (Mackie 1973, p. 295). Since λ says nothing, the same
must go for the inscription of ¬∃P(δP ∧ P ) that concludes our deduction: that
formula also fails to say that things are thus-and-so.
In the absence of any plausible explanation of how the final line of our deduc-
tion can say something while λ does not, I think we must grant Mackie’s con-
clusion: the inscription of ¬∃P(δP ∧ P ) at line 25 says nothing. There remains,
though, a residual problem that seems more than merely ‘awkward’. We reasoned
our way to line 25 by applying widely accepted logical rules to a true premise,
namely (1). Just as we do not want our logic to take us from true premises to a
false conclusion, so we do not wish it to take us from a true premise to a conclu-
sion that fails to say anything at all. On Mackie’s analysis, though, the deduction
from line 1 to line 25 accomplishes precisely that. So it would appear that at least
one of the logical rules applied in the course of the deduction must be revised.
The need for logical revision has been contested by some of those who accept
Mackie’s analysis of the Liar. Thus Timothy Smiley speaks of grammatically accept-
able sentences malfunctioning: “[i]n a particular context, perhaps in any context,
they fail to convey any coherent message” (Smiley 1993, p. 23). Like Mackie, he
holds that the ordinary Liar sentence malfunctions. For the Strengthened Liar, he
invites us to
consider A: ‘A is not true’, where this is spelt out as ‘A is false or mal-
functions’. Concluding that A malfunctions, how can we avoid the
206 Reflections on the Liar
further conclusion that A is not true, with the consequent re-entry into
paradox? Answer: this conclusion depends on the inference ‘A malfunc-
tions. Therefore A is false or malfunctions’. But it’s not that A is false
or malfunctions; it’s that ‘A is false or malfunctions’ malfunctions. The
fact that the conclusion of the inference fails to express the proposition
which its form would suggest, undercuts the appeal to form on which
the inference relies for its validity. Sod’s law trumps the law of or-
introduction, just as it does when sentences are ambiguous or context-
dependent…. As Priest says of his own solution [to the Liar], “If this is
all disconcertingly non-algorithmic, this is just an unfortunate fact of
life.” (Smiley 1993, p. 26)
On Smiley’s view, the laws of logic are the familiar laws of the classical system. The
Liar calls for no revision to that system. But we have to recognize that the correct
application of those laws to a true premise will sometimes result, not in a true
conclusion, but in a sentence or well-formed formula that does not say anything
at all. There are no rules by following which we can be sure to avoid falling through
the trap door and uttering or inscribing words that say nothing. “The mistake is
to think of malfunctioning as being like failure to be a wff, something perceptible
and inherent and all-or-nothing, whereas it may be inferred and fortuitous and
perhaps a matter of degree” (Smiley 1993, p. 26).
Priest has objected that Smiley here “waves goodbye to the project of formal
logic, that is, of determining a (non-empty) class of inferences that are guaran-
teed to be truth-preserving in virtue of their form” (Priest 1993, p. 43). Priest’s
characterization of the logical project is tendentious. Whether an utterance
is paradoxical or not may depend on its context. Epimenides’s remark above is
paradoxical if no other utterance by a Cretan expresses a truth; otherwise, it is
straightforwardly false.11 Even in a formalized language, then, we cannot expect
to find a syntactic filter that excludes paradoxical utterances and inscriptions, and
thus ensures that when we utter or inscribe a well-formed formula we will be say-
ing something. All the same, Smiley’s talk of Sod’s Law trumping the laws of logic
is disconcertingly blithe. If at all possible, we would do better to reformulate the
rules of logic so that they at least warn us of what pitfalls to look out for before
we take an inferential step.
In order to see what form that reformulation might take, it helps to stand back
and reflect on the approach to the Liar that is being recommended. Central to that
approach is the distinction between a sentence or well-formed formula’s express-
ing a falsehood and its failing to express a truth. In a logical system of the familiar
kind (whether classical or intuitionistic), we can signify that a sentence is false
by asserting its negation. But how can we signify that it fails to express a truth?
One common approach is to postulate two negation operators. We have the
familiar ‘internal’ negation operator ‘¬’, understood so that ‘¬A’ is false when A is
true, true when A is false, and which fails to express any proposition when A so
The Liar w ithout Tr uth 207
fails. Alongside this, some logicians have postulated an external operator ‘~’. As
with ‘¬’, ‘~A’ is taken to be false when A is true, and true when A is false. However,
when A fails to express any proposition, ‘~A’ is taken to be true. It is, however,
far from evident that this attempt to endow ‘~’ with a sense is coherent. If ‘~A’ is
true, it must surely say that such-and-such, i.e., must express a proposition, but
it is quite unclear how it can do so when its component part, A, says nothing.12
For this reason, I recommend another approach to the problem of formal-
izing the difference between a sentence’s being false and its not being true. Let
us use the notation +A to signify that A is accepted as true. Then we can express
our belief that A is false by accepting its (internal) negation: we shall write
down +(¬A). Cognate to the act of accepting a sentence as true is that of reject-
ing it as untrue. We signify that A is untrue by writing down −A. The difference
between +(¬A) and −A symbolizes that between judging A to be false and judging
it to be untrue. Unlike those who postulate an external negation operator, I do
not suppose that −A expresses any propositional content when A does not. By
writing down −A, one simply performs the speech act of rejecting A is untrue—
an act one may well wish to perform when A itself fails to say anything. As indi-
cators of forces with which speech acts may be performed, + and − may not be
iterated. While +(¬¬A) is well formed, −−A is not.
What we shall then need is a system of logical rules which regulate inferential
transitions between these signed formulae, as I call them.13 We want the rules to
preserve correctness: a positively signed formula +A is correct if and only if A is
true, while a negatively signed formula −A is correct if and only if A is untrue. We
may take over the rules of the positive logic in the obvious way; thus modus ponens
takes the form: From +(A → B) and +A infer +B. But we also need rules that mix
positively and negatively signed formulae. Unlike the dialetheists, I suppose that
deeming A to be false commits one to rejecting it as untrue. So I accept the rule:
from +(¬A), infer −A. In general, however, the converse rule fails. From the prem-
ise that A is untrue, we cannot always infer that A is false. One may correctly reject
A as untrue because it fails to say anything, rather than because it says something
which is not the case, as correct acceptance of ¬A requires.
What form does the rule of reductio take in such a system? When the supposi-
tion that A is true yields (in tandem with background assumptions) the conclu-
sions that B is true and that it is untrue, we may infer that A is untrue (given the
same background assumptions). Thus we should accept the following rule of proof
(in which X is an arbitrary set of signed formulae, perhaps empty):
From the fact that +A yields a contradiction, we cannot always infer that A is false,
but in special circumstances we can deduce instances of +(¬A). For present pur-
poses, we may suppose that a sentence is untrue but not false only when it fails to
say anything. Let us say that a set X of signed formulae determines a bare formula
208 Reflections on the Liar
A when the correctness of all the members of X logically guarantees that A says
something. Then we have the following rule for affirming a negated sentence:
What light do these rules cast on our problematical derivation? What (Red)
directly yields at line 10 is −R. However, we may apply +¬-intro to deduce +¬R:
among the undischarged assumptions at this stage in the derivation is δR,
which certainly guarantees that R says something. As for the inference to line
11, (Red) and modus ponens combine to yield − ¬∃P(δP ∧ P ) . This combines
with + ¬∃P(δP ∧ P ) at line 18 to yield − ∃P(δP ∧ P ) at line 19. However, we can-
not assert + ¬∃P(δP ∧ P ) , for the assumptions left undischarged at that stage of
the derivation do not guarantee that ∃P(δP ∧ P ) says anything. Correspondingly,
the conclusion we reach at line 20 is simply − ∃RδR , not + ¬∃RδR . That weaker
conclusion, however, is quite sufficient to vindicate the present approach to the
Strengthened Liar. The supposition that Brainstorm says something at noon
is demonstrably untrue; that is enough for us to be warranted in rejecting the
hypothesis that he did then succeed in saying something.
Under the revised logical rules, then, the argument goes through as
far as the weaker version of line 20. What, though, of the next stage
of the derivation? The positive logic takes us to + ∃RδR at line 24, and
(Red) duly yields − ∃P(δP ∧ P ) at line 25. We cannot, however, move for-
ward to + ¬∃P(δP ∧ P ). The remaining undischarged assumption of the
derivation, viz. premise (1), certainly does not determine ∃P(δP ∧ P ).
To the contrary, line 20 establishes that − ∃RδR, so ¬∃P(δP ∧ P ) says nothing.
Accordingly, the derivation’s final conclusion − ∃P(δP ∧ P ) is entirely consistent
with our rejecting the hypothesis that Brainstorm says something when he utters
¬∃P(δP ∧ P ), or its English translation.
This analysis vindicates a slightly emended version of Mackie’s solution to the
problem of the Liar’s Revenge. He wrote that the Strengthened Liar’s “indetermi-
nacy with respect to truth is of a kind which prevents our saying even that it is
not true, and therefore from arguing, by a further step, that it is true.” I say: its
indeterminacy prevents us from asserting its negation and therefore from argu-
ing that it is true after all; but we may, and should, reject the Strengthened Liar
as untrue.
The idea that we should deal with the Strengthened Liar by rejecting is as
untrue, rather than by affirming its negation, is not new. It was propounded
more than 30 years ago by Terence Parsons (Parsons 1984). What I hope to have
contributed is something Parsons omitted to provide: a precise statement of the
logical rules by which we can justify rejecting λ as untrue without falling into the
trap of asserting its negation. That statement enables us to improve on Smiley’s
insouciant acknowledgment that the laws of logic may be trumped by Sod’s Law.
The rule of + ¬-introduction states the condition which must be satisfied if we are
to affirm a conclusion in the form +(¬A). For reasons already explained, there can
The Liar w ithout Tr uth 209
be no syntactic test for satisfying that condition. All the same, our analysis shows
us what we have to check for: we need to verify that the correctness of the opera-
tive premises and assumptions guarantee that A says something. To that degree,
we keep faith with the central project of formal logic: we do as much as can be
done by formal methods to ensure that our deductive arguments preserve truth.
To check that A says something, we shall have to go beyond the merely formal. But
we also need to do that when we check that the proponent of an argument has not
equivocated between two senses of a term, or that the passage of time does not
make a material difference to an argument formulated in the present tense.
We would be ensnared in paradox again if it were possible to introduce an exter-
nal negation operator ‘~’ for which + ~A were equivalent to −A. For in that case,
Brainstorm could cause renewed trouble by uttering the formula ∃P(δP ∧ P ) .
As we saw, though, there is no good theoretical reason why it should be possible to
introduce such a negation operator (see further Tappenden 1999). We would also
get into trouble if we applied our theory to ‘Whatever Brainstorm is saying should
be rejected as untrue’. I think we just have to concede that problems would arise
if the entire semantic machinery of the present chapter were to be projected into
the object language: the limitations on such projection mean (as in Kripke’s the-
ory) that “the ghost of the Tarski hierarchy is still with us” (Kripke 1975, p. 714).
But the ghost is far less inhibiting than the hierarchy proper: the rules proposed
enable us to do a great deal of semantic theorizing within our system. At any rate,
we have here a promising way of dealing with the conundrum presented by the
Strengthened Liar.15
Notes
1. In his essay “The Semantic Conception of Truth and the Foundations of Semantics” (Tarski
1944). That essay draws heavily on the earlier work on the definition of truth for formalized
languages reported in Tarski’s Wahrheitsbegriff paper (Tarski 1935). But it is in the later
paper that Tarski addresses the philosophical questions with which I shall be concerned.
2. For an illuminating discussion of the relationship between Convention (T) and Tarski’s gen-
eral metaphysical views about truth, see Patterson 2012, chap. 4.
3. I shall follow Tarski in using ‘true’ as a predicate of sentences, even though ‘expresses a
truth’ would be better. Tarski’s argument is not materially affected by replacing ‘true’ with
‘expresses a truth’.
4. The distinction is sharp, but it is relative. If “we become interested in the notion of truth
applying to sentences, not of our original object-language, but of its metalanguage, the
latter becomes automatically the object-language of our discussion; and in order to define
truth for this language, we have to go to a new metalanguage—so to speak, to a metalan-
guage of a higher level. In this way we arrive at a whole hierarchy of languages” (Tarski 1944,
sec. 9).
5. The case is inspired by the example of Buridan’s in which Plato says, ‘Whatever Socrates is
now saying is true’ at a time when Socrates says ‘Whatever Plato is now saying is not true’.
But it will be more convenient to treat of a single paradoxical utterance than a paradoxical
pair of utterances.
6. More precisely, what this argument shows is that there is no one proposition (up to
material equivalence) which λ expresses or says. For suppose, for a contradiction, that
210 Reflections on the Liar
References
Dummett, M. A. E. 1981. Frege: Philosophy of Language. 2nd ed. London: Duckworth.
Evans, M. G. J. 1982. The Varieties of Reference. Oxford: Clarendon Press.
Feferman, S. 2010. “On the Strength of Some Semi-constructive Theories.” In S. Feferman and W.
Sieg, eds., Proofs, Categories, and Computation: Essays in Honour of Grigori Mints, 109–129.
London: College Publications.
The Liar w ithout Tr uth 211
Semantics for Semantics
James R . Shaw
What drives our investigations of semantic circularity are logical puzzles, stem-
ming from the Liar. No wonder, then, that such investigations center predomi-
nantly on the question of how semantic circularity impacts our logical theorizing.
But this focus tends to defer an important question, that I take as my topic
here: what relevance does the Liar, and semantic circularity more generally, have
for the shape of our compositional semantic theories?
Questions about what a compositional semantics for a self-applicable truth-
predicate should look like, as compared to questions about the logic of paradox,
have received relatively little attention. The current literature on truth and cir-
cularity, for example, reveals nothing like the state of research in epistemic log-
ics, which are accompanied by highly active parallel research programs centering
directly on the compositional semantics of epistemic modal language, or knowl-
edge ascription.1 There is, I think it is fair to say, no comparably active, linguis-
tically driven research program centering on the compositional semantics of
truth-talk.2
There are very good reasons, I believe, why investigations into truth and
semantic circularity have been lopsided in this way. Compositional semantic
investigations tend, methodologically, to be descriptive in character. And they
aim, in part, to account for data that includes relatively stable acceptability
judgments, like truth-value judgments. But because circularity can lead lay-
man, like theorist, into confused or contradictory pronouncements, it’s not
clear that there are stable acceptability judgments to explain or, if so, what
they are. And the problems are serious enough to raise worries that the usage
of ordinary speakers simply evinces incoherence, so that there is no descrip-
tive theory of a coherent use of ‘true’ to begin with.3 Moreover, even those
investigations that treat semantically circular languages with tools conducive
to compositional investigation, like the model-theoretic tools of Tarskian or
Kripkean theories, tend to apply those tools in ways that treat ‘true’ compo-
sitionally on a par with other predicates by associating it with an extension,
or extension/antiextension pair. If these theories are on the right track, there
212
Semantic s for Semanti c s 213
compositional semantics for belief ascriptions ends up being tied to the success of
distinct accounts in foundational semantics.
Terms like ‘true’, ‘means’, or ‘represents’, however, generate an even more
intricate two-way entanglement between foundational and compositional seman-
tics. This arises as follows:
(A), in a way, merely restates one key explanatory goal of foundational seman-
tics: to characterize the kinds of semantic properties that would be attributed in
a formal theory, like a compositional theory.9 Formal theories, whether merely
logical or compositional, have no significance independently of an interpretation
of the kinds of semantic properties (like truth, or validity) that the formalism
ascribes.10 And foundational semantics is charged, in part, with the task of telling
us what those properties are like. (B), on the other hand, follows from the claim
that the theories within which we give our foundational accounts themselves
have compositional structure. If so, there is some story to be told as to how the
statements of our foundational accounts gain significance, and that story can be
given in compositional terms.
(A) and (B) represent, as I say, a kind of entanglement of compositional and
foundational semantics, with each playing a role in illuminating or explaining the
other. This intertwining requires the sentences of our completed foundational
theories to discuss foundationally significant semantic properties, which in turn
are attributed in the compositional formalisms that show how those very com-
posite sentences acquire the semantic properties they do. That is where the need
for semantic reflexivity, and a compositional theory modeling its behavior, comes
from.11
Abandoning (A) or (B) is in danger of crippling linguistic theorizing. If (A) fails,
properties attributed in some of our compositional semantic theories outstrip
those detailed in foundational semantics, and our compositional theory becomes
mere uninterpreted formalism, and will cease to do explanatory work.12 If (B) fails,
properties mentioned in our foundational semantics outstrip those which can be
given a compositional semantics, so that there is no systematic account of the
significance of the foundational account itself, as if its sentence-meanings were
had by magic. This raises strong worries about the coherence of the foundational
account.
Of course, if one works on the paradoxes, one may reasonably be concerned
that either (A) or (B) does fail, in spite of all this. I think the paradoxes are so
216 Reflections on the Liar
recalcitrant that we should bear this in mind as a drastic option to contend with.
But I want to stress that the damage accruing to this option can go unappreci-
ated by those positing expressive and logical restrictions in their formal theories
of truth. The foundational programs just discussed won’t recognizably survive
weakenings of the assumptions I’ve been making. If the semantic paradoxes show
the reflexivity required for (A) and (B) to be incoherent, then we should prob-
ably abandon hope of resolving the problem of intentionality, especially in any
naturalistically satisfying way. Some intentional, representational, and linguistic
phenomena are essentially inexplicable, unlike any other known object of study.
Moreover, whatever formalisms we give in supplying our theories of truth will be
essentially uninterpretable, or open ended. We’ll lack any justifiable theory draw-
ing explicit ties between the formalism and the mental and linguistic phenomena
that the formalism is clearly supposed to somehow illuminate. This threatens the
whole point of giving a formalism in the first place.
So: better to proceed, provisionally, on the assumption that such a composi-
tional semantics admitting circularity can be found. Let’s turn now to the ques-
tion of what it should look like.
Sighted English-speaking consultants report (1) is true, even if they have never
encountered this sentence before. Others are not as reliable. What explains these
patterns?
A natural explanation appeals to two theories. First, a theory of nonlinguistic
cognition: the sighted consultants have the perceptual and cognitive capacities to
tell the shapes and colors of objects in their near environment. Second, a theory
of linguistic competence for English—a competence appropriately modeled by a
theory associating semantic values with the words ‘gray’ and ‘square’, and com-
prising rules for composing those meanings in line with English syntax.
218 Reflections on the Liar
We find the same regularity for (2) as for (1). Why? Again, our compositional
theory should determine the obvious truth-conditions for (2): that the sentence
labeled ‘(1)’ on the questionnaire express a truth. But here, unlike with our sen-
tence about shapes or colors, the explanatory work of the compositional theory
in explaining the judgments about (2) is not quite done. We need to explain the
speakers’ prior acquaintance with (1), and how speakers compositionally settled
its truth, to fully explain how they have the information which, paired with the
information modeled by the cognitive theory, entails (2)’s truth. The composi-
tional theory, not merely the separate cognitive theory, is partly responsible for
explaining how our speakers are aware of (1)’s truth, so as to settle (2)’s truth. If,
to take a strange case, our consultants knew English up the meanings of ‘gray’
and ‘white’, they would obviously not be able to settle (2)’s truth in the way they
typically do. The fact that the compositional theory re-engages with (1) in the
explanation of the judgments about (2) is what helps explain why and how this bit
of linguistic competence with ‘gray’ and ‘white’ figures in the consultants’ judg-
ments about (2), which contains neither of those words.19
The simple case involving (2) reveals a basic pattern in using compositional
theories to explain regularities in truth-value judgments about sentences con-
taining semantic terms. To explain the regularities, the compositional theory may
be set up to require speakers to have information about truths. But to explain how
speakers have that information, we may need to appeal to the compositional the-
ory again. For any set of attributions, these roundabout explanatory moves may
iterate many times before bottoming out in our nonlinguistic cognitive theories,
which in turn ground the explanation in a satisfactory way.
But note that the structure of these explanations reveals a possibility, and hid-
den danger, unique to our compositional theories of semantic terms. The danger
is that we structure our theory in such a way that explanations fail to bottom
out. Our theory may be structured so that using it to explain various truth-value
assessments unendingly redirects us to the compositional theory, thus affording
no explanation of the assessments at all. This outcome is obviously unaccept-
able provided it arises for an empirically confirmed regularity in compositionally
mediated, correct truth-value assessments. I’ll call such a situation one in which a
(vicious) compositional circularity arises in our theory.
Though it may be hard to appreciate now, in the abstract, it is worth stressing
that the question of whether a theory exhibits compositional circularities is not
settled by the theory’s being compositional. In particular, a theory may ensure
that semantic values of parts and wholes satisfy a compositional principle of
220 Reflections on the Liar
any desired strength, while that theory nonetheless fails to avoid compositional
circularities in the sense I’ve just described. If so, avoidance of compositional
circularities would place an additional desideratum on our compositional theo-
ries of semantic vocabulary. As I say, this may not be easy to see merely from
the definitions given here. But rather than continue in abstraction, the best
way to appreciate the separability of the constraints is to examine some actual
theories which generate compositional circularities, while seemingly satisfying
compositional constraints on relations between parts and wholes. Let’s turn to
this task now.
9.3. Defaulting
In section 9.1, I argued that we have a very strong need for a compositional
semantics for languages that admit semantic circularities. What should such the-
ories look like?
One of the most successful semantic programs that fits into the paradigm
given by section 9.2 is the broadly model-theoretic tradition informed by genera-
tive grammar. Working within this program, a tempting way to model the com-
positional effects of ‘true’ would be to assign it a familiar semantic value borne by
any common predicate: one which at a context-index pair (say) fixes an extension
consisting of the entities which bear the property it is used to record.20 I’ll call this
view ‘Truth-Extensionalism’.21
(3) I made a promise which I’ll have kept just in case everything I say
today is true.
all, if there were such relevant, true, reflexive content, what possible grounds
would there be not to treat (4) as expressing it? If this is the motivation for
an appeal to pragmatics, though, such appeals are essentially self-defeating,
because they are themselves helpless to avoid positing such content. Sid could
have equally well promised that what he said or implied (or conveyed, etc.) would
be true. If he reports his success on Monday, it continues to seem true, quantifi-
ers seemingly interpreted broadly just as with (4). What is the explanation of this
judgment? We face a dilemma. We could grant that the new utterance literally
and truthfully expresses a bit of reflexive content. But then there is no reason
not to attribute similar content to (4). Alternatively, we could try to say that the
new utterance merely implicates a truth about all things said and implicated. But
then we will have a bit of content implicated that is both reflexive (an implicature
in part about the implicature) and true. There is no ‘hyperpragmatic’ content
to avoid the reflexivity in the implicated content. And so the motivations for
appealing to pragmatics in the first place are undermined—it needlessly shifted
the bump in the rug.
Speakers don’t seem to be confusing the truth value of (4) with that of impli-
cated content. But this doesn’t yet tell us speakers are making correct judgments.
There is still room for a broad error theory. Perhaps speakers are systematically
mistaken about what is said or implicated in defaulting cases. On this view, for
example, speakers are simply in error if they take Sid to communicate something
that, on its own, entails that his promise will be fulfilled.
This maneuver is quite radical. After all, we would be claiming that speakers
systematically make mistakes about what they all stably think that they are com-
municating to each other (whether by saying, or implicating, and so on). How in
this case could a speaker think they are conveying something with certain truth
conditions (remember: not necessarily literally conveying them, merely convey-
ing them somehow), a hearer think that those truth conditions are conveyed, and
both be in error? One might be concerned that it is close to constitutive of com-
municating certain truth conditions, in a broad sense, that it is common belief
between communicator and interpreter that precisely those truth conditions are
being conveyed.
The radical nature of this position means that it calls for strong justifica-
tion. There is perhaps one set of considerations that could play this justifica-
tory role: reflections on reflexivity in relation to paradox. The thought goes
roughly as follows: Speakers may think they are conveying certain truth con-
ditions. But we know from discussions of reflexivity and paradox that some-
times reflexive statements can’t obviously have truth conditions at a given
world (in the case of the Liar) owing to logical problems, or they have truth
conditions that are awkward to assign (in the case of the truth teller). Perhaps
similarities between the problem cases and (4) should lead us to assimilate the
latter to the former, in spite of how speakers seem to use sentences like (4) in
communication.
Semantic s for Semanti c s 223
The problem is that there is a clear and striking asymmetry in how speakers react
to (4) and (T).23 This asymmetry needs to be accounted for somehow. Granted, we
have existing theories that treat (T) as untrue (often ‘defective’). And these theo-
ries will likely treat (4) as untrue as well. But overturning the judgments about
(4) with an error theory for these reasons is putting the cart before the horse. If
we are creating a compositional theory, it should be responsive to the data extrap-
olated from speaker judgments, not overturning it. We should of course allow
irregularities to be smoothed over by appeals to other plausible explanations
of the data—most notably pragmatic ones. But when such explanations aren’t
available, we can’t simply pretend the data isn’t there in order to hold on to our
current favored theories, even if they are simpler. The data must be accounted
for somehow, so the error theory, even bolstered by reflections on paradox, must
be accompanied by some further explanation of the regularities in speaker judg-
ments. I’m not sure what this further explanation looks like, given that pragmatic
avenues seem to be ineffectual.
There is another, I think slightly better, way of pressing this kind of worry.24 The
position grants that (4) is typically treated as a truth by ordinary speakers. But it
goes on to contend that the speaker judgments are inherently unstable. Perhaps,
for example, speakers haven’t properly reflected on the similarities between cases
like (4) and those like (T), and if they did, they would be willing to retract their
judgments, or at least become confused.
I think this is a more plausible route to pursue with the theoretical reflections
on paradox—not to simply overturn speaker judgments, but question their stabil-
ity. A problem is that the methodology here is a little murky. The paradoxes are
extremely complicated. Even careful, highly trained theorists can make mistakes
about how they operate. And we do not yet have anywhere near an uncontrover-
sial theory of how paradox is to be treated. Consultants may be unduly influenced
by the complexities of the paradoxes, or prejudices of theorists, in forming ‘sen-
sitized’ judgments (whatever the proper methodology for producing them would
be). For example, I think it would not be entirely surprising if discussions of
paradox could (without proper coaching) problematize speakers’ intuitions about
clear truths like those used in the famous Dean/Nixon cases of Kripke (1975),
discussed below. If they did, I think it would be clear we should ignore those judg-
ments. More generally, if we get confused judgments after complex theoretical
tutoring, there is an open question about which of the two conflicting sets of
judgments should be empirically fundamental. Perhaps more importantly, even if
the ‘sensitized’ judgments are the more fundamental ones, the original judgments
224 Reflections on the Liar
(in particular the original asymmetries between (4) and (T)) have still been
left largely unexplained. We only have the claim that speakers get confused by
circularity—but why so much so with (4), and not at all with (T), even when in the
former case all nonsemantic facts are broached, and speakers are consulted about
the range of quantification? So even this more nuanced form of error theory still
needs to be supplemented with some additional explanatory considerations.
Still, I see that the contention that we should be wary of sensitized judgments
may be controversial. And though I do not know what kind of theory can be paired
with the error theory to systematically predict the regularities in speaker judg-
ments, I do not want to presuppose that no such theory can be devised. We’ve
bottomed out in a mix of empirical and theoretical considerations I can’t hope to
resolve here. Thankfully, there’s no strong need to. As I flagged earlier, the cases
I’ll consider in the next section are much more theoretically costly to ignore in
the way proposed here, though they effectively require a ‘benign’ reflexivity just
like (4) seems to exhibit. This will actually break the alleged symmetry between
(4) and (T) on which the error theory relies. Still, we shouldn’t turn to those cases
immediately. Our discussion so far has presented at least a noteworthy prima
facie case for (4)’s truth. It will be helpful to first discuss why this simple feature
of (4) would matter to our compositional theories, before turning to more com-
plex examples.
Now, working provisionally on the assumption that speakers are correctly
assessing (4) as true, I take it to be uncontroversial that those assessments are
compositionally mediated. Speakers who judge (4) true haven’t, for example, col-
luded with some third party who has told them its truth value. Awareness of the
truth of (4), if speakers have it, depends at least in part on appreciating (4)’s struc-
ture, and the meanings of its parts.
Granting these claims, let’s first note a minor consequence. If (4) is true, an
understanding of ‘true’ in English is not exhausted by a grasp of the Tarski bicondi-
tionals. This is because one can never use a Tarski biconditional for (4) to establish
(4)’s truth. This biconditional, along with basic empirical premises, can establish
that (4) is true if several utterances, including that of (4), are true. But this mild
circularity ensures (4)’s truth cannot be established using only uncontroversial
empirical premises and truth biconditionals. This may have implications for those
deflationist positions which take something like a grasp of the Tarski bicondition-
als to exhaust an understanding of ‘true’. Such theories would require added rules
to capture defaulting conditions.
I mention this logical fact because it is connected with the separate claim that
I want to make my focus. If (4) is true, truth-extensionalist views cannot explain
the speaker judgments about (4) because those theories generate compositional
circularities.
Recall the discussion of section 9.2: compositionally mediated speaker judg-
ments are to be explained as the product of linguistic competence, modeled by
our semantic theory, in conjunction with the nonlinguistic information speakers
Semantic s for Semanti c s 225
Kripke noted that while utterances of (J) and (N) by Jones and Nixon respectively
can be paradoxical, under favorable circumstances they are not. If Nixon utters
only clear falsehoods besides (N), and Jones only utters (J), then Jones’s utter-
ance is false and Nixon’s is true.
One might think that such utterances resemble defaulters. The interdependent
semantic ascriptions make (N) in some sense, in part, ‘about itself’. But none of
this casts doubt on Truth-Extensionalism. Every normal case where (J) and (N) are
truth-evaluable is one where we can settle, with the information contained in a
suitable compositional theory, enough truth values to settle the semantic status
of one utterance before the other.
For example, our compositional theory will explain that speakers can settle
(J)’s truth if they can settle that the majority of Nixon’s utterances are untrue. If
Nixon has only uttered falsehoods besides (N), our cognitive and linguistic theo-
ries are in a position to explain how speakers get that information in a noncircular
way, before using it to settle (J)’s truth, and then using that information to settle
(N)’s truth. There is no threat of a circular explanation here. There are no grounds
to think that the information modeled with an extension assignment to ‘true’,
paired with the information speakers have by other means, will not entail the cor-
rect truth-value allotments to (J) and (N) (in the relevant circumstances).
So (J) and (N), though semantically interdependent, do not generate com-
positional circularities for truth-extensionalist theories the way defaulters do.
Utterances that rely more directly on their own truth come closer to generating the
problem. Simple truth-tellers like (T), discussed above, at least have the right kind
of circular structure. But (T) can’t be used to argue against Truth-Extensionalism,
for two reasons. First, it’s unclear that there is speaker concurrence on (T)’s sta-
tus. In fact, it seems there is no concurrence for (T) on a ‘standard’ semantic sta-
tus. Second, and relatedly, (T) is commonly taken to be semantically defective. But
generating an unavoidable explanatory circularity seems like a perfectly legiti-
mate way to explain the presence of semantic defect. It’s not clear why the status
of defective utterances would have to be determined by nondefective composi-
tional means.26
So the problem generated by (4) requires a special confluence of features,
borne neither by Kripke-style interdependences, nor by self-dependent truth-
tellers. Note that the problem is not that truth-extensionalist theories fail to be
compositional. A model-theoretic semantics, like that associated with a Kripkean
fixed-point construction, may show how the semantic value for (4) harmoniously
depends on (is a function of) the semantic values belonging to (4)’s expressions,
in essentially the standard way (up to any special accommodations needed for
antiextensions). The problem is that such a theory does not afford an explana-
tion of how speakers exploit these compositional dependencies to establish the
Semantic s for Semanti c s 227
semantic facts that they can. The information encapsulated by the theory isn’t
sufficient, given the other information speakers have, to determine the relevant
semantic facts.
Recall that these problems for the truth-extensionalist can be framed as a
dilemma in explaining judgments concerning (4), once the truth of other utter-
ances (4) speaks about is established. Either she can try to derive (4)’s truth
within the compositional semantics, representing speakers’ linguistic compe-
tence, which is hopelessly circular, or she can appeal to some bit of nonlinguis-
tic cognitive awareness to settle (4)’s truth, which I’ve claimed to be implausible.
Before concluding our discussion of defaulters, I want to address one last attempt
to avoid trouble by gripping the second, cognitive horn of this dilemma.
I’ve claimed nonlinguistic cognition can’t get us far enough to settle (4)’s
truth. Perhaps it could (coupled with information possessed in virtue of lin-
guistic competence) explain how we arrive at the truth of the other utterances
(4) speaks about. But we don’t know (4) is true in the same way that we know, for
example, that the sky is blue—we don’t simply ‘perceive’ (4)’s truth without any
compositional work.
But the analogy with perception here, though I think forceful, only points out
that speakers’ a posteriori nonlinguistic knowledge is unhelpful. But what if (4)’s
truth is known with the help of nonlinguistic a priori, conceptual knowledge?
Consider: Most English speakers may correctly judge ‘green resembles blue more
than red’ a true sentence of English. Plausibly, this regularity isn’t explained by
their linguistic competence, combined only with knowledge gained by a posteriori
cognition. Instead, it is the product of their linguistic competence and a priori
cognition of conceptual truths about colors and their relative similarities. The
objection I want to consider asks whether there could likewise be a priori, concep-
tual truths about the concept of truth that help speakers settle the value of (4).
For example, perhaps it is just an a priori, conceptual fact about truth that, when a
sentence like (4) is uttered and all other utterances (4) speaks about are true, then
(4) itself expresses a truth. The idea is that linguistic, compositional resources
and a posteriori forms of cognition are generally needed only to get us informa-
tion about the truth of the other utterances (4) speaks of. But the final step of
settling (4)’s truth owes neither to a posteriori cognition, nor to the composi-
tionally mediated competence with the language in which (4) is couched. Rather,
it involves an appeal to conceptual facts about truth that speak to the semantic
properties of defaulters under special circumstances.
This view at least improves on our other suggested appeals to cognitive facul-
ties in that it leaves room for compositional work to play some role in settling
(4)’s truth, before hypothesized conceptual truths come into play. However, there
are two worries for the proposal. The first is simply that introducing irregular
conceptual truths about the linguistic properties of special uses of ‘true’ seems ad
hoc. But I want to press a second, more instructive problem with the strategy: it
is clearly a conventional matter whether, and when, defaulting occurs. There is
228 Reflections on the Liar
there is a sense that the term ‘true’ is defectively applied to them. Giving a seman-
tic value for ‘true’ that fails to specify how it is to be applied in such cases may be
the perfect way to model the relevant speaker behavior. So the real problems for
Truth-Extensionalism only arise when we have a use of ‘true’ whose application
is nondefective (simply true or false) while the application of ‘true’ is underdeter-
mined by the stipulation that ‘true’ apply to all and only the truths. When these
two conditions are met, a truth-extensionalist theory will not be in a position to
explain how speakers correctly allot the truth values in question. There will be a
nondefective, conventional use of ‘true’ which the theory cannot help predict.
So examining defaulters helps reveal the conditions under which a composi-
tional circularity arises for truth-extensionalist views. We need a sentence S for
which two claims hold.
(i) Empirical/Semantic Claim: Speakers correctly judge S true (or false) through
normal compositional means.
(ii) Logical Claim: S would be nontrue (nonfalse) in a language which differed
from ours only in adopting a different conventional use of ‘true’, compatible
with its correct application to all and only truths.
Let me spell out why these two conditions are sufficient to generate composi-
tional circularities. If speakers can become aware that some sentence S contain-
ing ‘true’ is true (or false), by conventional compositional means (condition (i))
we need an explanation of how they can become so aware. Let I denote the sum
of the information encapsulated by the theory of speaker competence excluding
that mediated by the semantic value of ‘true’, along with all nonlinguistic a priori
and a posteriori information a given speaker possesses. To this we add the aspect
of speaker competence modeled by the semantic value of ‘true’—call it i—to I.
Together, I and i should conjointly determine that S is true (false), if we’ve cho-
sen our semantic value for ‘true’ correctly. After all, I and i together exhaust the
sources of information speakers have, both tacit, linguistic information had in vir-
tue of competence with their language, and all nonlinguistic information gained
by a priori and a posteriori means. Now, suppose i is the information supplied
in a truth-extensionalist theory: only that ‘true’ applies to all and only truths,
or that it has the truths as its extension at a world. Then, if condition (ii) holds,
there is a possible language which differs from that of our target speakers only in
conventions for the use of ‘true’, but consistently with both i and I. That is, in the
second language ‘true’ also applies to all and only truths. If S has a different truth
value, then obviously I and i cannot settle that S is true (false). What the possibil-
ity of this alternate language shows is that I, i, and S’s nontruth (nonfalsehood)
are consistent. So this would reveal that our hypothesis—that the semantic value
of ‘true’ determines merely an extension consisting of all and only truths, repre-
sented with i—was incorrect. We need something besides i to model the aspect of
speaker competence relevant to understanding the meaning of ‘true’.
232 Reflections on the Liar
Note that these two conditions (i) and (ii) also specify the sense in which sen-
tences generating compositional circularities for Truth-Extensionalism are ‘spe-
cial kinds’ of semantic circularities. We saw that truth tellers like (T) arguably do
not generate such compositional circularities. If so, it is because (i) fails to hold
of them: speakers don’t judge truth tellers true or false (whether by normal com-
positional means, or otherwise). We also saw that Kripke’s interdependences, like
(J) and (N), do not generate compositional circularities. This is because (ii) does
not hold of them. In the relevant circumstances, there is no way of adopting a use
for ‘true’, applying to all and only truths, which assigns (J) and (N) truth values
other than those they actually have.
In this section, I’ll use conditions (i) and (ii) to pinpoint a second class of
claims that generate compositional circularities for truth-extensionalist theories.
These claims are drawn from semantic generalities: sweeping statements about
patterns of truth-value allotments, which those very statements may instan-
tiate. Semantic generalities should sound familiar. These are the very kinds of
statements we discussed in section 9.1, which make providing a compositional
semantics for words like ‘true’ imperative. In section 9.1, I effectively argued that
conventionally valued semantic generalities cannot be turned aside with an error
theory, the way defaulters like (4) might be. Jettisoning defaulters stops Sid from
talking about how good he is at keeping his promises. But jettisoning semantic
generalities undercuts the expressive resources we need to do foundational work
in the philosophy of language. When we tamper with such statements, we tamper
with the theoretical goals in giving formal theories of truth quite generally. So if
we find semantic generalities that create compositional circularities, the truth-
extensionalist cannot relegate the judgments to an error theory. To do so would
not merely add to their theoretical and expressive costs, but could well be self-
defeating, by threatening to undermine the coherence and purpose of the very
compositional theory they put forward.
As before, we need to be careful when looking for the relevant semantic
generalities—this time, for two interrelated reasons. First, semantic generalities
pose a well-known stumbling block for formal theories of truth. Second, the rea-
sons for this likely have to do with getting the semantics of expressions other than
‘true’ correct.
Consider, for example, that some of the simplest and most obvious generalities,
like (5) and (6), elude expression by quite sophisticated formal theories of truth.
Surely (5) and (6) (setting aside dialetheist views) should come out as truths
of English. What precludes their truthful expression? A candidate culprit here is
the presence of some projective truth-value other than mere truth or falsity. (By
‘projective’ I mean the value is ‘infectious’, or tends to be inherited by logical com-
pounds.) Such a value is present in both the strong and weak Kleene schemes. But
there are other possible culprits worth considering.
For example, the languages used in standard fixed-point constructions render
(5) using a unary quantifier from first-order logic, when a conservative binary gen-
eralized quantifier more appropriately helps capture its semantics. Introducing
conservative generalized quantifiers could allow (5) to be true for intuitive rea-
sons, even within a theory that otherwise admits projective truth values other
than truth and falsity. After all, (5) would become a logical truth, even given the
presence of a third value, on such a construal.
Also, problems with (6), though doubtless connected with projective truth-
values, may also concern the meanings of logical connectives. Bracketing para-
dox, (6) might come out true in a language admitting truth-value gaps as long as
it also boasts a form of ‘exclusion negation’—a use of negation that is true if its
compliment is anything but true. That connective may be needed to capture the
intuitive, true reading of (6).
I review these obstacles to the truthful expression of (5) and (6) because one
may be tempted to think they could generate compositional circularities for truth-
extensionalist theories. Ordinary speakers judge (5) to be true, ostensibly by nor-
mal compositional means, seemingly satisfying my semantic claim (i). And some
formal theories of truth in which ‘true’ accurately applies to all and only truths,
like fixed-point constructions, render (5) untrue, seemingly satisfying my logical
claim (ii).
But the appearance here is a little deceptive. Though the semantic claim (i) is
probably satisfied for (5) and (6), whether the logical claim (ii) is also satisfied
should be controversial. For the logical claim to hold, the theories in which (5) is
untrue (e.g., various fixed-point theories) must give that sentence a semantics dif-
fering from the English use of (5) only in the conventions governing the behavior
of ‘true’. It was a stipulation of the logical claim that only the behavior of ‘true’
be changed, and with good reason. After all, if we get a truth-value shift in a sen-
tence S by shifting around not the use of ‘true’, but instead the logical form of the
sentence, or the meaning of separate words, then a plausible explanation of how
speakers settle the truth value of S is that they are aware of its logical form and
the meanings of the relevant words. If this were to happen, then in my argument
above the alternate language constructed would be in conflict with I (which sub-
sumes all the information in the model of speakers’ linguistic capacities vis-à-vis
words other than ‘true’). And if so, the argument that S generates a compositional
circularity wouldn’t go through.
Just such worries cast doubt on whether (5) satisfies my logical claim (ii). If
there are projective forms of semantic defect belonging to applications of ‘true’,
234 Reflections on the Liar
it is very plausible that speakers who judge (5) true are recognizing the use of
a conservative binary generalized quantifier. No truth-extensionalist language
in which ‘every’ is given the appropriate meaning will allow (5) to be untrue—
it will become a logical truth even in the presence of a third value.29 So truth-
extensionalists may have a possible legitimate explanation of how speakers
recognize (5)’s truth. (5), though ‘about itself’, doesn’t generate a compositional
circularity. Similar remarks will apply to (6).
What this shows is that to uncover semantic generalities that create composi-
tional circularities, we may have to look for somewhat more specialized cases. And
arguing those cases generate circularities will require a great deal of care: we not
only have to show how the truth value of the sentence can shift owing to different
conventions governing the truth predicate, but we also have to argue that other
logical words aren’t in any way contributing to the shift.
The work here is tricky, so I will focus on just one case. And in discussing it,
I will be appealing to the following assumption.
(C) There is a projective form of truth value other than simple truth or simple fal-
sity (like the third value in the strong or weak Kleene schemes) which predica-
tions of ‘true’ may bear owing to reflexive applications.
(C) can be motivated in several ways. It could perhaps be motivated just by con-
sideration of odd semantic reflexivities like (T).30 But the real reason for taking
(C) on board is because in the model-theoretic tradition on which I’m focusing,
devising a language with a truth predicate whose extension consists in the true
sentences of the language (essentially, a fixed-point in the sense of Kripke [1975])
is not obviously, or at least not easily, achievable, owing to the presence of liar-like
circularity, without the assumption of something like (C). Accordingly, I take it as
satisfactory for my purposes here if I can argue against candidate compositional
theories in the model-theoretic tradition that take (C), or something like it, on
board as an assumption.31
Now, on any reasonable construal of the semantics for disjunction, (7) should
be simply false.
(7) Some disjunction which is not true is such that all of its disjuncts are
true.
But, on the assumption of (C), there are formal theories of truth in which the
extension of ‘true’ consists of all and only the truths in which (7) is not false. For
example, it is neither true nor false in the minimal Kripkean fixed-point for a
suitably expressive language containing (7) on both the weak and strong Kleene
schemes. As we’ve just seen, we have to be careful about what this shows. Perhaps
(7)’s nonfalsity derives primarily from the treatment of quantification, or the logi-
cal connectives, in these fixed points, as was argued for (5) and (6). But, in this
Semantic s for Semanti c s 235
case, there seem to be good arguments which show that we cannot trace the pos-
sibility of the truth-value shift to these other expressive resources. To see this, it
is easiest to focus on a different problematic sentence: an equally false disjunction
of (7) with itself.
(8) Some disjunction which is not true is such that all of its disjuncts
are true or some disjunction which is not true is such that all of its
disjuncts are true.
Supposing (8) can be false consistently with ‘true’ applying to the truths,
let us now ask whether it could also bear our hypothesized third projective
value, call it ‘u’, consistently with ‘true’ applying to the truths. This does seem
possible, regardless of how we construe the semantics of other terms in the
sentence.
First, as effectively noted before, if the semantics for (8) makes use of unary
quantifiers and is interpreted relative to either the weak or strong Kleene schemes,
we have constructions which show how (8) and both its disjuncts can bear u con-
sistently with the stipulation that ‘true’ hold of all and only the truths. But unlike
with (5) and (6), the absence of generalized quantifiers or exclusion negation can’t
be the source of the problem.
In particular, even if we accommodate both resources, we can show (in some-
what painful detail) that (8) can consistently bear the value u owing to semantic
circularity. To do this, we start with three assumptions:
(a) ‘true’ used in (8), as applied to (8), or either of its disjuncts, bears the projec-
tive value u (owing to circularity, as per (C)),
(b) ‘Some’ and ‘all’ in (8) are binary generalized quantifiers, and
(c) ‘not’ in (8) expresses exclusion negation.ge
(9) Every disjunction which is not true is such that all of its disjuncts are
true.
(10) Every disjunction which is not true is such that all of its disjuncts
are true or every disjunction which is not true is such that all of its
disjuncts are true.
(9), though similar to (7), will not generate a compositional circularity. Intuitively
(9), like (7), should come out false. But this will in fact occur in any language
which has the logical resources to capture what (9) would typically be used in
English to say. In such a language, all we need to falsify (9) is to find one counter-
instance, like (11).
We can compositionally settle that (11) is false, and therewith that it satisfies the
restrictor of (9) and falsifies its matrix. That’s sufficient to show (9) is false (and
(10) as well), independently of what statuses (9) or (10) have. In this respect (9)’s
truth value will never hinge on its own truth value as does (7) or (8).
(9) doesn’t ‘significantly’ concern its own status because its own status isn’t
necessarily relevant to its evaluation. But there are other ways in which a gener-
ality might not be significantly about itself. Sentence (5)—‘Every true sentence
is true’—is an example. If it uses a conservative generalized quantifier, (5) will
be true as long as ‘true’ has any extension at all. (5)’s truth is secured on logical
grounds by the meaning of its quantifier. So there is no danger that its own status
could interfere with assessing its truth value.
So, like we saw with defaulters, not just any semantic generality which ‘con-
cerns itself’ generates compositional circularities. But this doesn’t detract from
the point that (presumably infinitely) many sentences do meet the conditions
required to generate such circularities provided (C) holds.
Also, as already noted, an error theory concerning these judgments—for exam-
ple that (7) and (8) are not actually false—is very hard to swallow, and comes with
some quite striking theoretical consequences that I highlighted in section 9.1. So
we can’t take the recourse we saw might be available for defaulters.
Indeed, it is worth flagging that examining semantic generalities also strength-
ens the case for the truth of defaulters like (4). Cases like (8) show we will need
‘benign’ reflexivity (that is, reflexivity that results in ordinary truth-evaluability)
for certain semantic generalities anyway. This effectively undercuts the force of
the objection to defaulting that makes use of an error theory, which relied essen-
tially on an analogy between (4) and (T). If we think (7) or (8) is false, we con-
cede that some reflexivity is benign, even if some, like that in (T), is not. There is
then no reason to think (4) must be assimilated to the problem cases like (T) as
opposed to the truth-evaluable cases like (7) or (8)—indeed, all evidence points
to the opposite assimilation.
238 Reflections on the Liar
9.5. Truth-Proceduralism
I’ve argued that compositional theories that merely assign ‘true’ an extension at a
context are inadequate. This is not because such theories fail to be compositional,
nor because they assign incorrect truth-values. Nor is it necessarily because there
is no extension for ‘true’ in the sense of there being no set of truths—I’ve nei-
ther assumed nor argued that there isn’t. It is because those theories generate
compositional circularities inconsistent with their explanatory purpose: they fail
to explain productive uses of compositional tools by natural-language speakers.
Truth-extensionalist theories encounter these problems because they lack the
resources to represent special compositional conventions that govern the uses of
terms like ‘true’. We need a liberalized conception of the semantic value of ‘true’
that is capable of encoding such conventions.
There is no hope of supplying such a semantic value here, even in a ‘toy’ model.
Even a simple version of this model would have to be at least as complex, if not
more so, than something like a Tarskian or Kripkean theory. So what follows is, of
necessity, said merely by way of advertisement.
Even if we can’t give the details of such a theory here, we can use the lessons
of the preceding sections to draw out some general and instructive information
about what a semantic value for ‘true’ overcoming the problems with Truth-
Extensionalism will have to look like. We’ve learned that uses of ‘true’ are gov-
erned by special kinds of conventions. If we reflect just a little on what exactly
those conventions are, they should give us insight into what work a semantic
value that represents them has to do. In particular, when we reflect on the con-
ventions, we see that they can instructively be divided into two separable compo-
nents. Let me start with the first.
Consider again our defaulter (4), and our semantic generality (8).
240 Reflections on the Liar
What we want to know is how speakers are positioned to arrive at the conclusion
that Sid’s utterance of (4) is true or that (8), and its disjuncts, are false. We need
a better model of their semantic competence with ‘true’ which gives us a picture
of how the information they are able to acquire by other linguistic and cognitive
means is sufficient, in conjunction with that competence, to settle the truth of
(4) and (8).
Let’s focus on (4). Clearly, to settle the utterance’s truth a speaker must
have the information that all other utterances Sid made the day he speaks
(4) are true. We know how the speaker can get that information, so suppose
she has it. Now, does she need any further information? No. Or, at least, she
needs no further information about the truth values of various utterances.
The thing that needs to be registered by our model of her semantic compe-
tence with ‘true’ is that that information is ‘enough’, for compositional pur-
poses, to establish the truth of Sid’s (4). In other words, our model of her
competence should encapsulate the information that if Sid’s other utterances
are true, his utterance of (4) is as well. Something roughly analogous is true
of (8): our model of a speaker’s competence needs to subsume the informa-
tion that the possibility of (8)’s defectiveness is excluded from consideration,
in settling whether the generality, of which (8) might be an instance, holds.
We can, as it were, exclude (8) and its disjuncts (and perhaps other claims as
well) from consideration when looking for counterexamples to the disjunctive
existential claim (8) makes.
As argued in section 9.3, the aforementioned conditional pertaining to (4)
(‘if the rest of Sid’s utterances are true, Sid’s (4) is true’) is no conceptual truth.
It is a conventional truth—a linguistic convention. This is a piece of information
that needs to be represented in modeling speakers’ competence with English.
As such, we need to be able to extract this conditional from the semantic value
of ‘true’ (in conjunction with other linguistic and nonlinguistic information).35
This is the information that was nowhere encoded in a mere function from an
index-context pair to an extension, or to the property of truth.
We can begin to frame this conventional information in a slightly more tech-
nical way, appealing to the notion of semantic dependence. Consider sentences
(12a)–(12c).
To settle that what (12a) says is true, you need only know facts about roses. (12a)’s
truth depends, if on anything, only on ‘the facts’—that is, the nonsemantic facts.
To settle that what (12b) says is true, by contrast, you need to know some seman-
tic facts: facts about what (12a) says and whether what it says is true. Likewise,
the semantic status of (12c) depends on the semantic status of (12b).
Iterated ascriptions of semantic terms, like (12a)–(12c), create a natural kind
of ‘hierarchy’ in which the semantic properties of ascriptions ‘higher up’ in the
hierarchy intuitively depend on the semantic properties of those below. At the
‘bottom’ of the hierarchy (if all goes well) things depend on ‘the facts’. Saying that
one sentence ‘semantically depends’ on another is effectively to place it above the
other in this hierarchy. Something like this notion of dependence is implicit in
the notion of groundedness in Kripke (1975). Other versions of the concept are
also treated more directly in Yablo (1982), Gaifman (1992), Maudlin (2004), and
Leitgeb (2005).
The foregoing characterization of semantic dependence has obviously been
informal and schematic. It can be fleshed out in many competing ways, as some of
the above authors have done, depending on how one interprets the nature of the
relation, and its formal properties. For example, is the relation metaphysical—
a kind of grounding? Is it logico-semantic—merely a means of recording which
items a sentence predicates ‘true’ of? And as regards its structure: Is it transitive—
does (12c) depend on (12a) because (12b) does? Does (12a) semantically depend
on nothing? ‘The facts’? All, or merely some?
Once we ask these questions, we should recognize that there is probably not
just one notion of semantic dependence here, but a family, with different struc-
tures and theoretical purposes. What we should be interested in is a semantic and
perhaps quasi-epistemic version of such a dependence relation—what I will call
a compositional semantic dependence relation. Very, very roughly, the relation could
be specified as follows.
means (e.g., by being told), you could still compositionally settle the value of
(12c) without knowing anything about (12a). So the compositional semantic
dependence relation as I’ve characterized it is nontransitive. The truth teller
(T) arguably compositionally semantically depends on itself. That, arguably, is
why it is defective in some way: it compositionally requires information about its
own truth value, where there is no ‘antecedent’ way of obtaining it. Finally, Sid’s
utterance of (4) compositionally semantically depends on all his other utterances
spoken that day. That is, even though Sid’s utterance attributes truth to itself, as
I argued in section 9.3, his utterance of (4) does not compositionally semantically
depend on itself.
Why? If (4) is knowably true, this last claim is just a more technical way of
saying that to settle (4)’s truth, one doesn’t need antecedent noncompositional
awareness of (4)’s truth. Otherwise one could never know it was true, and
(4) would probably be defective (as (T) seems to be).
We can also use this new terminology to restate our more recent lesson about
the meaning of ‘true’. I pointed out that the semantic value of ‘true’ needs some-
how to register the information that the truth of all Sid’s other utterances is, by
itself, sufficient to compositionally settle the truth value of Sid’s utterance. In our
new technical terms: ‘true’ needs to somehow encapsulate the information that
(4) does not compositionally semantically depend on itself. Again, an extension
assignment in standard model-theoretic frameworks simply won’t model that
information, even implicitly.
So: the first component of the conventions that contribute to the benign char-
acter of my reflexivities is represented by a compositional semantic dependence
relation. In particular, these conventions dictate that sometimes compositional
dependences of this kind are (more or less stipulatively) ‘restricted’ so as to
facilitate unambiguous compositional derivation in the presence of some kinds
of virtuous semantic circularities. But I also mentioned that there was a second
component of the conventions. To see this, note that there are hypothetical lan-
guages in which uses analogous to (4) purposefully and compositionally default
to a third status, like a truth-value gap, or even to falsehood. Those languages will
agree with our use of ‘true’ that utterances like (4) only compositionally semanti-
cally depend on Sid’s other utterances. So obviously the difference between ‘true’
in our language and in these other languages depends on a further conventional
feature: how the truth value of (4) is to be settled on the basis of its restricted compo-
sitional semantic dependences. In our language, (4) is true if all its compositional
semantic dependences are true, false if one of them is false. In the hypothetical
language just envisioned, by contrast, an analogous utterance could be false even
if all its compositional semantic dependences are true. The difference between the
languages is just in these stipulated conventions for evaluating (4) on the basis of
its compositional dependences.
Thus, compositional conventions for the use of ‘true’ seem to encode two
elements: relations of compositional semantic dependence, and a method for
Semantic s for Semanti c s 243
assigning truth values to predications of ‘true’ on the basis of those, perhaps con-
ventionally restricted, compositional dependences.
Note that these two conventions are all that is needed. Indeed, the two kinds of
conventions are individually necessary and jointly sufficient to cope with default-
ers and semantic generalities. Dependence relations are necessary, since without
the information that defaulters and generalities can be excluded from consider-
ation when evaluating themselves—that is, without the information that they’re
at least not compositionally semantically self-dependent—a model of speaker
competence with ‘true’ is sent on an unending loop in evaluating the statuses
of the relevant utterances. Evaluation methods are also necessary, because even
with information about which restricted set of semantic properties are needed for
compositional purposes, owing to restricted dependences, there is still latitude in
how that restricted information is to be applied.
Joint information about both dependence relations and evaluation methods
is sufficient, because together they entail the conditionals that bridge the truth
values assigned to dependences and the truth values of our problematic utter-
ances. The joint information entails that if the rest of Sid’s utterances are true,
Sid’s (4) is true. It entails that if there are no counterexamples to the gener-
ality in the first disjunct of (8) besides (8) (and perhaps some other interde-
pendences), then that disjunct is false. Those are the very pieces of conditional
information that we could not retrieve, but needed to retrieve, on the truth-
extensionalist view. And with those pieces of information, the truth values of
the relevant utterances will be entailed by independent information that speak-
ers can acquire—the information that the antecedents of the relevant condi-
tionals are true.
So, as promised, reflecting on the conventions required by defaulters and gen-
eralities has clarified the work that a revamped semantic value for ‘true’ needs to
perform. The thesis that the semantic value for ‘true’ does it is the following.
A little metaphorically, the semantic value for ‘true’, directly or indirectly, encodes
two things by convention: paths through utterances along which truth values can
be compositionally assigned, and a procedure, or rule, for assigning truth values
based on assignments earlier on those paths.
The ‘proceduralism’ in ‘Truth-Proceduralism’ signals that the meaning of ‘true’
is not just an extension, or intension, but a more liberalized semantic value—some
244 Reflections on the Liar
1992, p. 227). Calling relations are, in effect, a version of what I called a semantic
dependence relation. The semantics consists in stipulating these relations, along
with rules for assigning truth values along them. It is thus a truth-proceduralist
semantics in the most immediate sense. I will have a little more to say about its
structure in section 9.6.36
But for now I want to note that my constraints above on semantic values
allow for other theories, which we can call ‘implicitly truth-proceduralist’. Such
views don’t explicitly detail compositional semantic dependence relations and
rules for assignment in giving a semantics for ‘true’. Instead, their semantics
induces the elements (a) and (b). A broad class of theories which may do this
include axiomatic theories. An axiomatic theory of truth gives rules for settling
the values of utterances (as determined by derivability from the axioms along
with empirical information, say). And those rules give an implicit characteriza-
tion of a semantic dependence relation (as sentences one may yet have to prove,
to establish the truth of the sentence one is presently trying to settle). Such
a theory can explain defaulters and generalities in the way suggested above,
perhaps simply treating the generalities as axioms, and directly stipulating sep-
arate axioms governing when defaulters may default. Such theories are struc-
turally poised to overcome the particular obstacles for Truth-Extensionalism
detailed here, precisely because they may enhance their model of the under-
standing of ‘true’ arbitrarily with added axioms governing its behavior. I do
want to flag that, although the theories have this virtue, I think they face seri-
ous further challenges in satisfying the goals of a compositional theory if infer-
ential relations are pervasively taken as explanatorily fundamental, instead of
being derived from something more like a traditional recursive assignment of
truth conditions. Still, there is no space to discuss these issues here. It suffices
for now to note that axiomatic theories are fully compatible with the particular
arguments I’ve given so far.
As I conceded, the discussion in this section has proceeded at a high level of
abstraction. This was necessary to gain an appreciation for the form shared by
compositional theories that avoid the danger of generating compositional circu-
larities. Still, even with only this abstract form in hand, we are poised to draw
some important lessons that the transition to procedural semantics will involve.
In particular, we learn a lesson about how to think not only about the virtuous
circularities of foundational semantics, but the vicious circularity of the Liar.
What’s written on line 1 is paradoxical for familiar reasons. One might wish to call
it untrue on those grounds. This is precisely what happens in line 2. If we can con-
sistently pronounce paradoxical utterances untrue, one might expect that we are
able to do so because new tokens of the same type are somehow immune to the
defects of the original paradoxical utterances. Versions of something roughly like
this idea have been noted and exploited in different ways by a number of authors,
including Parsons (1974), Burge (1979), Gaifman (1992, 2000), Koons (1992),
Simmons (1993), and Glanzberg (2001, 2004).
Any approach to paradox that treats what is uttered on line 1 as semantically
problematic, and what is uttered on line 2 as simply true, obviously requires an
asymmetry in the semantics of lines 1 and 2—an asymmetry which traces some-
how to the circumstances of their production, since they appear to be tokens of the
same semantic type (perhaps up to hidden indexical elements). What is intriguing
is that once one adopts a procedural semantics, as I’ve claimed we must, we will
have independently motivated exploitable compositional asymmetries between
lines 1 and 2 of a very special kind.
To understand why, it is helpful to return to our simple case of defaulting.
Consider a more stilted sentence that Sid might have uttered.
Suppose Sid utters (13) on March 15, a day on which he additionally produces
only some respectable number of uncontroversially true statements. Then Sid’s
utterance, us, seems true for now familiar reasons. Suppose, at some point, Pia
248 Reflections on the Liar
shares in Sid’s assessment and utters (13) as well. Again, if the circumstances are
as I mentioned, this new utterance up will count as true.
Sid’s utterance us and Pia’s up appear to be tokens of the same sentence type. They
also have the same truth values. But what is interesting about them is that, from
a compositional perspective, the truth of each utterance has different grounds.
Sid’s utterance requires a defaulting mechanism to engage. Were it not for the
conventions of English, Sid’s utterance might have been defective. But Pia’s utter-
ance requires no such mechanism. It does not ascribe a truth value to itself in the
way that Sid’s utterance does in the circumstance I have described, and because of
this, it can be compositionally sensitive to a broader distribution of truth values
than Sid’s utterance. Sid’s utterance us must not require information about itself
while its truth value is being settled if it is to be conventionally truth-valued—it
must not be compositionally semantically self-dependent. However Pia’s utter-
ance may, should, and seemingly is compositionally responsive to that very same
utterance us (just consider if Sid had said different things at that moment). So
the operation of the truth predicate’s semantic value in these two utterances is
different: it responds to different bodies of semantic information, and does so
in different ways. Put in more technical terms: the two utterances have different
compositional semantic dependences, despite being two tokens of the same type.
And as such, a procedural semantics modeling their behavior ‘makes room’ for
assignment rules for ‘true’ that distinguish between the utterances and assign
them different truth values. Of course, a language which associated ‘true’ with a
different procedural semantic value on which us and ui receive distinct truth-value
assignments would be bizarre, and entirely unlike English. And it does not seem
like it would have ‘true’ express a viable truth-concept. What is important is that
the shift to a procedural semantics independently opens up space for a difference
in truth-value allotments to the two tokens of the same type, owing to the seman-
tics of the word ‘true’.
For related reasons, we have empirical evidence that English already associates
the word ‘true’ with a semantic value of a kind that could permit the tokens on
lines 1 and 2 to bear different truth values. These utterances, too, are naturally
treated as having a kind of asymmetry in the structure of their compositional
semantic dependences: the first has reflexive dependences, while the second does
not. A procedural theory may associate ‘true’ with rules of evaluation along com-
positional dependences that end up treating utterances with unresolvable reflex-
ive dependences like line 1 as less than true, with the nonreflexive utterance of
line 2 as being true, and responsive to this defect.
But we need to be careful about the nature of the differential assignment that
is opened up by a procedural theory. The potential sensitivity is, again, not a form
of context sensitivity like that found in ordinary indexical pronouns or grad-
able adjectives. Up to the fact that they may mediate differential assignments of
truth values to tokens of the same type, the two kinds of sensitivity have little in
common.
Semantic s for Semanti c s 249
with the arguments I’ve given here, are the beginning of a conversation about
how such resolutions might be developed. The main lesson we’ve learned is that
a special class of approaches to paradox, with a highly unusual form of expres-
sive flexibility, actually only make use of linguistic resources in understanding the
semantic values of semantic terms like ‘true’ that every adequate compositional
theory must accommodate.
I have hopes that a broadly truth-proceduralist framework may provide us with
the best resources to understand virtuous and vicious semantic circularity alike.
But even if this hope is ill-founded, the goals of foundational and compositional
semantic theorizing ensure that procedural semantic values will become an inte-
gral part of our final story about what it is to understand a language that is able to
talk about its own semantic properties.
Acknowledgments
For helpful comments and advice I’m grateful to Bradley Amour-Garb, Michael
Glanzberg, Anil Gupta, Doug Patterson, David Ripley, James Woodbridge, audi-
ence members at the “Truth at Work” conference in Paris, and several anonymous
reviewers.
Notes
1. For a mere sampling of only recent work on the semantics of epistemic modals, for example,
see Anand and Hacquard (2013), Dorr and Hawthorne (2013), von Fintel and Gillies (2007,
2008), Hacquard (2006, 2010), Hacquard and Wellwood (2012), MacFarlane (2011), Moss
(2015), Willer (2013), Yalcin (2007, 2010, 2011, 2012).
2. This is not to say that there is no recent work on truth of relevance to compositional theo-
rizing. There have been discussions of the relationship of truth-conditional semantics to
deflationism, for example in Horisk et al. (2000), Collins (2002), Patterson (2005), and
Burgess (2011). But such investigations tend to focus on how truth is employed in giving
our semantic theories, not in how ‘true’ figures as an object of study within them. There are
also some more direct, linguistically driven treatments of the semantics of ‘true’ itself—
a recent case being Moltmann (2015). But there, the critical issue of semantic circularity
tends to drop out of the picture. A topic closer to my theoretical interests here is that of how
to accommodate the threat of paradox within something like a Davidsonian truth-theoretic
semantics. But the focus on vicious, as opposed to virtuous, circularity in such investiga-
tions tends to lead to pessimistic outcomes: broadly hierarchical treatments that seem to
preclude circularity, rather than accommodate it (Lycan 2012), treatments that take our
interpretive truth theories to simply be inconsistent (Lepore and Ludwig 2007), or treat-
ments that abandon broadly truth-theoretic approaches to meaning altogether (Pietroski,
this volume). As will eventually become clear, an aspiration of this chapter is to consider
both virtuous and vicious circularity together, and to exploit compositional lessons about
the former to illuminate the latter. The eventual hope is that this methodology will help us
safeguard interesting forms of semantic circularity in a broadly truth-conditional setting.
It is also worth flagging that all work on logics of truth can potentially be construed as
indirectly constraining our compositional theories. The trouble is that it is uncommon for
Semantic s for Semanti c s 251
like a fixed point. If there is nothing like the relevant fixed-point construction to begin with,
that is a concern for the objector in the dialectic.
34. Again, contingent on the methodological assumption that any such language exists to
begin with.
35. I grant that the conventions in question might operate for a ‘language as a whole’, rather
than being encoded at the level of individual lexical items, especially for a case like English.
I’ll continue to talk as if the conventions are encoded in the word ‘true’ for now, since it
is at least conceptually possible that defaulting conventions differ from semantic word to
semantic word.
36. As noted by Gaifman (2000, p. 85), these kinds of formalisms have natural antecedents in
‘operational semantics’ for programming languages (see Winskel 1993, chap. 2). Gaifman
often describes these kinds of semantics as ‘noncompositional’ (e.g., Gaifman 2000, p. 83),
which may make them seem inappropriate for the explanatory purposes I am interested in.
But Gaifman’s sense of ‘noncompositional’ does not, or does not obviously, mark something
as being at odds with the working assumptions I gave in section 9.2. For example, Gaifman
acknowledges the importance of accommodating productivity facts of the kind I’ve claimed
characteristically drive compositional investigation (Gaifman 1992, p. 236). There are some
terminological issues to iron out here in how to use the term ‘compositional’. I am ambiva-
lent about whether there is a fruitful sense in which a procedural semantics counts as being
noncompositional. But I will not pursue this terminological question here.
37. Gaifman, who as noted in section 9.5 operates within a truth-proceduralist framework,
early on stressed differences like those I’ve been harping on between the ‘token-sensitivity’
of truth-proceduralist frameworks and ordinary forms of context sensitivity. See, for exam-
ple, the extended discussions in Gaifman (1992, secs. 3, 5).
38. In the most recent list of theorists whose views were contrasted with truth-proceduralist
approaches, Glanzberg, and Parsons, whom he draws on, were omitted. This is because
Glanzberg’s theory and perhaps Parsons’s, unlike those others, may share one of the two
virtues of the truth-proceduralist views I’ve just described. In particular, the ‘extraordinary’
kind of context sensitivity Glanzberg posits in quantifier domains to account for shifts in
truth-value allotments to utterances containing semantic predicates may allow Glanzberg
to avoid fragmenting the concept of truth. Or, at least, it is much less clear that Glanzberg
is somehow committed to fragmenting truth in his theory. Still, a sensitivity in quanti-
fier domains, even of the special type Glanzberg posits, is unhelpful in accounting for the
benign circularities I’ve made my focus. From that perspective, Glanzberg’s theory, like the
others, is not poised to account for the data without taking on the shift to a procedural
semantics.
References
Anand, P. and V. Hacquard. 2013. “Epistemics and Attitudes.” Semantics and Pragmatics 6
(8): 1–59.
Beall, J. 2009. Spandrels of Truth. New York: Oxford University Press.
Burge, T. 1979. “Semantical Paradox.” Journal of Philosophy 76: 169–198.
Burgess, A. 2011. “Mainstream Semantics + Deflationary Truth.” Linguistics and Philosophy 34
(5): 397–410.
Chihara, C. 1979. “The Semantic Paradoxes: A Diagnostic Investigation.” Philosophical Review 88
(4): 590–618.
Chomsky, N. 1965. Aspects of the Theory of Syntax. Cambridge, MA: MIT Press.
———. 1986. Knowledge of Language: Its Nature, Origin, and Use. New York: Praeger.
Collins, J. 2002. “Truth or Meaning? A Question of Priority.” Philosophy and Phenomenological
Research 65 (3): 497–536.
Dorr, C., and J. Hawthorne. 2013. “Embedding Epistemic Modals.” Mind 122 (488): 867–913.
Semantic s for Semanti c s 255
10.1. Introduction
We aim to investigate the question of when it is reasonable to replace an inconsis-
tent concept. By ‘replace’ we do not mean ‘eliminate’. Instead, we are interested in
the question of when it makes sense to introduce a new concept or concepts that
are designed to fill at least some of the roles played by the concept discovered to be
inconsistent. It might turn out that the concept in question is still used in certain
situations even by those who recognize that it is inconsistent.
The main application of our inquiry is to the concept of truth and the so-called
inconsistency approaches to the paradoxes that affect truth (e.g., the Liar, Curry,
and Yablo). These approaches entail that truth is an inconsistent concept and that
the paradoxes are symptoms of this inconsistency. Our question is this: if truth
is an inconsistent concept, then does it need to be replaced? Or more generally,
when is the cure worse than the disease?1
257
258 Reflections on the Liar
factive. Among them are proposals by Paul Boghossian (1997), Gillian Russell
(2008), and Juhl and Loomis (2009). One recent account by David Chalmers is
factive as well, but the factivity is tacked on to the primary account, which need
not entail factivity. Chalmers calls his notion discursive analyticity. Here is his
definition:
(3) A dispute over S is broadly verbal iff for some expression T in S, the
parties disagree about the meaning of T, and the dispute over S
arises wholly in virtue of this disagreement regarding T. (Chalmers
2011, p. 522)
Those with Quinean sympathies will probably be unhappy with Chalmers’s appeal
to meanings in this definition, despite the fact that Chalmers claims his notion of
analyticity would be amenable to the Quinean (2011, pp. 561–562).
Instead, we might try something a bit more noncommittal like John Burgess’s
pragmatic notion of analyticity. He illustrates his view by considering the law of
excluded middle as an example in the following passage:
Burgess’s notion is going to have some odd consequences, like if during a chat
session the ‘z’ button on my keyboard suddenly stops working, then it will be
helpful to stop using the term ‘zipper’ in a disagreement over whether Elias Howe
invented the zipper. However, it seems odd to say that ‘Elias Howe invented the
zipper’ is constitutive of ‘zipper’. There are exceptions in the other direction as
well. When we talk to Graham Priest about the paradoxes affecting truth, we
know he accepts that some things are both true and not true. That violates a con-
stitutive principle for us—one for ‘and’ and ‘not’. Priest knows at least one of us
does not accept T-In (e.g., if p then <p> is true) and T-Out (if <p> is true, then p),
which he takes to be constitutive of ‘true’. We all keep track of these differences
when we have conversations and it does not cause any problems as long as we
have not been at the pub too long. So would it be helpful to stop using ‘true’ or
to add ‘dialetheic’ to Priest’s truth claims? Not for us, and we think Priest feels
the same way. Does that make T-In and T-Out any less constitutive for ‘true’? No.
Nevertheless, Burgess’s notion of pragmatic analyticity will do for our purposes
here.2
10.3. Cases
Our purpose is to sketch the options for what to do with inconsistent concepts
and, in particular, when they are up for replacement. So we need some cases. In
this section, we sketch four examples of (what we take to be) inconsistent con-
cepts. Again, those are concepts whose constitutive principles are inconsistent,
either with each other or with otherwise uncontroversial facts.
We will not argue in detail that these four concepts are inconsistent. For one
thing, that would take us too far afield. For another, given the vagueness and
context sensitivity of ‘constitutive principle’, we realize that there are other inter-
pretive options available, and that there may be contexts in which some of these
concepts are not inconsistent. The distraction would not further the present
agenda. We ask readers to go along with our classification, at least for the sake of
argument.
R ev is ing Incons istent C onc e pt s 261
10.3.1. Truth
Our primary example, of course, is the concept of truth. There are a number of
choices for its constitutive principles, and we need not settle on a single batch of
them. The usual candidates are the T-scheme,
T<p> ≡ p,
one instance for each appropriate sentence p. Instead, one might take the intro-
duction and elimination rules:
10.3.2. Set/Membership
As noted by Kurt Gödel (1944, 1964), and a host of others, there are at least two
different conceptions of set. According to one of them, a set as a collection of
previously given objects. We might speak of sets of US senators or sets of natural
numbers. In this sense, ‘set’ is a sort of context-sensitive expression, perhaps with
an elided constituent.4 ‘Set’ would be short for ‘set of X’s’, where context would
indicate what goes in for X.
As Gödel noted, this conception of set can be iterated. We can speak of sets
of sets of senators and sets of sets of natural numbers. And sets of sets of sets
of senators, and so on. And, of course, the iteration can be carried into the
262 Reflections on the Liar
transfinite. So let us call this the iterative conception. The X’s, upon which we
start the iteration, are sometimes called urelements. In some more or less stan-
dard interpretations of pure Zermelo Fraenkel set theory, there are no urele-
ments. We begin the iteration on nothing, producing the empty set; and we go
on from there. Gödel remarks that the iterative notion of set has not been shown
to be inconsistent, nor do the usual arguments affect it, at least not directly. It is
captured by the axioms of Zermelo or perhaps Zermelo-Fraenkel set theory, with
or without urelements.
The other notion of set, which we can call the logical conception, takes a set to
be the extension of a property, concept, propositional function, or predicate. For
convenience, we’ll stick to properties. The idea is that each logical set is tied to at
least one property. The main principle for individuating logical sets is, of course,
extensionality: sets are identical just in case their corresponding properties are
coextensive. If we add to this a principle that every property has an extension,
then we have Gottlob Frege’s Basic Law V, and inconsistency.
To wax metaphysical, a logical set is tied to a property, etc., or to a class of prop-
erties, while an iterative set is constituted by its members. So the two notions
may have different modal profiles, but such issues will not be pursued here.
Mathematics is generally taken to be thoroughly extensional.
For purposes of this chapter, we simply stipulate that the logical notion of set
is inconsistent, with Basic Law V, or something similar to that, as a constitutive
principle. This might be challenged, but we won’t take up such challenges here.
The Secretary Liberation Club was an enormous success. As the number of suc-
cessful, snobby clubs grew, so did the number of disenfranchised secretaries.
Eventually, the Secretary Liberation Club got so large that the members decided
to hire their own secretary, an efficient man named Pat. Pat got to know the ins
and outs of the Secretary Liberation Club, and liked what they were doing; and he
developed friendships with many of the members. In time, Pat applied for mem-
bership in the Secretary Liberation Club (you didn’t see that coming, did you?).
After pondering things for a few minutes, the membership committee realized
that they had a potential problem. They asked if, per chance, Pat worked as sec-
retary for another club, hoping it was a snobby one. Alas, Pat served as secretary
only for the Secretary Liberation Club. They then noticed that a member of the
Secretary Liberation Club was about to take maternity leave from her job, and
encouraged Pat to seek employment with that club—or perhaps with another
snobby club. Pat asked if they were somehow displeased with his work, and were
hinting that his services were not wanted any more. They quickly reassured him
that, indeed, they were most pleased with his service to the Secretary Liberation
Club, and were thinking that he might want to take this opportunity to supple-
ment his income. Pat thanked them for their concern, but told them that, as a
single parent, he does not want to spend any more time away from his children.
Thus the Secretary Liberation Club was thrown into a crisis.
264 Reflections on the Liar
valuable to its members, presumably; and we assume there are no moral issues on
the downside. The inconsistent phrase is part of the project of administering the
affairs of the club. Without it, there is no Secretary Liberation Club. So we take it
that this example is part of a valuable project.
Infinitesimals played an important—essential—role in the development of the
calculus and real analysis, from about the middle of the seventeenth century until
well into the nineteenth century, and beyond if we take physics and engineering
into account. So we have a clear case of the inconsistent terms playing a role in a
valuable project, an intellectual project in this case.
Set theory is a branch of mathematics, in its own right. That presumably counts
as a valuable project on its own. We think it safe to say that, at least nowadays,
it is agreed that a legitimate branch of mathematics must be consistent. And, of
course, it is no accident that contemporary set theory concerns the iterative con-
ception of set, not the inconsistent logical conception.
We focus here on the role of sets in two other projects. The logical notion
played a key role in logicist accounts of mathematics. This is quite explicit in Frege
(1893), where what we call “logical sets” are instances of what he called “courses of
values.” We take it that logicism is, or was, at least prima facie a valuable project,
at least for some people.
Our second project is also foundational, but in a somewhat different sense
than logicism. Serious issues concerning the coherence, and perhaps consistency,
of certain theories occasionally arise in the normal practice of mathematics, and
the community needs a way to deal with those. Mark Wilson illustrates the his-
torical development and acceptance of a space-time with an “affine” structure on
the temporal slices:
The claims that Maddy makes on behalf of set theory can be disputed, but we will
take them for granted here. Providing this sort of mathematical foundation is
indeed a valuable project.
if they ever come to it. The situation is essentially the same even after they hire
Pat, especially if it is far-fetched to think that Pat will ever want to join the club.
Why would he want to? Our conclusion is that there is only trouble—pressure
to revise the bylaws—when Pat does apply for membership or perhaps when it
becomes clear that he is going to apply.
We do not wish to get embroiled in delicate matters concerning the history of
infinitesimals, and their eventual replacement in modern analysis. It is a long and
complex story. We will rely on the sketch in Kitcher (1983, chap. 10).5 It seems that
logical critiques of infinitesimals in the Analyst did have some impact in England,
among followers of Newton, but it had virtually none on mathematicians in the
Continent. Those mathematicians felt little or no need to focus their talents on
the task of finding a replacement. They felt they had better things to do. For about
two centuries, mathematicians were able to follow their intuitions, and get many
interesting and valuable results within the infinitesimal calculus. Moreover, these
results displayed a strong mutual coherence with each other, and most could be
verified using more traditional methods. Suppose, for example, that someone
proved, using infinitesimals, that a certain sequence converges to π/4. They were
then able to verify the accuracy of the result to any desired approximation by cal-
culating members of the sequence.
It is not at all clear that anyone could give a batch of rules that exactly
codified the practice of using infinitesimals. The practitioners just seemed
to know what to do and, more important, what not to do with them. They
developed fine intuitions concerning what could be done with infinitesimals
and what could not be done, and were able to impart those intuitions to their
students.
This started to change sometime during the nineteenth century, when some
important results came in conflict with other results. In retrospect—after the
calculus was put on a rigorous foundation—we can see what went wrong. For
example, one result would be correct (by contemporary lights) if ‘convergence’
meant pointwise convergence, and another, seemingly conflicting result required
convergence to be uniform convergence. According to Kitcher, it was situations like
this, internal to the normal practice of mathematics, that motivated the pioneer-
ing work to make analysis more “rigorous”, giving birth to the ε-δ definitions we
use today—and to the banishing of infinitesimals, at least for a while.
The lesson of this episode (at least as we have characterized it) is that incon-
sistent terms need not undermine an otherwise successful and productive intel-
lectual project. The project can go on so long as the practitioners have a good feel
for what they can and cannot do with the potentially troublesome terms. And this
particular project went on splendidly for some 200 or 300 years, engaging some of
the finest mathematical minds ever. Until the trouble arose, internally, the project
was not regarded as broken, and was in no need of fixing.
To take one more batch of examples, some philosophers argue that all
vague words express inconsistent concepts in the foregoing sense (Eklund
R ev is ing Incons istent C onc e pt s 269
2005; Tappenden 1993; and Kamp 1981). The idea, it seems, is that a tolerance
principle—something like the main premise of a sorites argument—is constitu-
tive of the term in question. Even if we assume that such views are correct, the
inconsistency does not prevent vague terms from being applied in everyday life.
Ordinary speakers just seem to know, tacitly, not to run too far down a sorites
series.
One crucial difference between the mathematical cases and the ordinary ones,
concerning vagueness, and perhaps the Secretary Liberation Club, is that math-
ematics (and science, to the extent that it relies on mathematics) thrives on long
chains of deductive reasoning. So inconsistencies are more dangerous. If a mathe-
matical concept is inconsistent, it seems quite likely that, eventually, in the course
of a long argument, or several interlocking long arguments, the practitioners will
lose track of the adjustments required to avoid trouble, or else their intuitions for
what they can and cannot do will eventually fail. Nevertheless, the enterprise can
get a good run with inconsistent concepts.6
In the previous section, we noted two projects involving the notion of
set: Fregean logicism and mathematical foundation. Scholars differ on what the
goals of logicism were. One of Frege’s stated goals was to establish that arith-
metic is analytic—that its truths could be derived from basic logical laws and
definitions. Of course, it turned out that one of those supposed basic logical laws,
the fifth, is inconsistent. Frege, at least, took the news of this inconsistency as
undermining the program, writing to Bertrand Russell, “Your discovery of the
contradiction caused me the greatest surprise and, I would almost say, consterna-
tion, since it has shaken the basis on which I intended to build arithmetic” (van
Heijenoort 1967, pp. 126–128).
It seems clear that an inconsistent term cannot play a role in a Fregean logi-
cism, at least not if the target theory, arithmetic in this case, is based on classi-
cal logic. One of the stated goals of Frege’s logic was to provide a framework for
derivations that are absolutely valid and free from gaps. The purpose of that is
to establish that the theorems of arithmetic do not invoke anything like Kantian
intuition. We cannot rely on tacit rules of thumb or our instincts concerning what
can and cannot be done, for if we do, something like Kantian intuition may have
snuck in. To serve Frege’s purposes, the rules of derivation must be explicitly
stated, in full generality. If any sentence at all can be derived in the system, it is
useless.
It might be added that Frege took himself to be continuing the program of
making analysis rigorous, the same program that led to the banishing of infini-
tesimals. Inconsistent concepts have no role to play.
Consider next the role of sets in the mathematical foundational program
sketched above. According to Maddy (1997, p. 26), set theory provides “a court
of final appeal for questions of mathematical existence and proof: if you want
to know if there is a mathematical object of a certain sort, you ask (ultimately)
if there is a set-theoretic surrogate of that sort; if you want to know if a given
270 Reflections on the Liar
Concerning logicism, Frege, at least, did not know how to avoid the problem.
He tried an alternate definition, a small change to Basic Law V. There is no need to
decide whether the proposal invoked a replacement notion, since it, too, proved
to be inconsistent.
It is interesting that Russell did not find the paradox as troubling to logicism as
Frege did. He thought that Frege’s definitions of the individual natural numbers
were essentially correct. The paradox is avoided by invoking the language of rami-
fied type theory (Whitehead and Russell 1910). The notion employed is, arguably,
a version of the logical notion of set, since a set is still thought of as the extension
of a predicate (or an attribute). Consistency is maintained by carefully delimiting
what counts as a property.7
Consider now the mathematical foundational program sketched above.
Suppose that someone tried to develop this using the logical notion of set. The
inconsistency would undermine the program, for the foregoing reasons. One
could try to run the program with the notion of set that comes from ramified
type theory. But then the theory is too weak to serve the foundational goals—for
much the same reason that Whitehead’s and Russell’s logicism failed its goals.
We’d need an axiom of infinity, and something like the an axiom of reducibility
to undermine the effects of ramification. In effect, we’d invoke a version of the
iterative notion of set.
The first change would make Pat ineligible, and the second would make him
eligible.
272 Reflections on the Liar
10.5. Truth
Like we said, we’re going to take it for granted that truth is an inconsistent con-
cept and that its constitutive principles are the instances of the T-schema for an
expressively rich language like English. Now we want to consider whether it satis-
fies the conditions for replacement.
Truth predicates also allow one to indirectly endorse or reject propositions one
cannot directly assert or deny. For example, even if one does not know, or forgot,
what the Poincaré conjecture is, one can assert ‘the Poincaré conjecture is true’
and thereby endorse it. It allows us to simultaneously endorse lots of propositions
that might even be inconvenient or even impossible to assert one by one; e.g., the
theorems of Peano Arithmetic are true.
A third expressive role is as a device of generalization. Let’s say someone
thinks that if most physicists believe that the speed of light is invariant, then the
speed of light is invariant. Moreover, they claim that if most physicists believe
that there is dark matter, then there is dark matter. In fact, the point is perfectly
general—it has nothing to do with these two particular claims. We can state
the general point by using quantifiers and the truth predicate: if most physi-
cists believe something, then it is true. Without the truth predicate, we would
be unable to formulate something that generalizes over all these instances. It
should be clear that there is some overlap between the generalizing function
and the endorsement function, but they are distinct (in the example about what
physicists believe, the truth predicate is not used to endorse anything because it
occurs in the consequent of a conditional).
These expressive roles are clearly valuable and figure in all sorts of important
projects. Moreover, there is general consensus that truth plays these roles.
At this point, however, inflationists and deflationists about truth part
ways—the latter claim that these expressive roles are truth’s only legitimate
functions. Once they have been specified, we have said all that can be said
about the nature of truth.9 Inflationists, on the other hand, think that we can
provide an analysis or reductive explanation of truth and this vindicates its
explanatory role.10
Truth has been a remarkably popular explanans. Meaning, validity, proof,
inquiry, belief, assertion, necessity, knowledge, analyticity, predication, and ref-
erence have all been tied to truth via various proposed explanations. If one thinks
that truth is a legitimate explanans and that providing some kind of explana-
tion of these other concepts is worthwhile, then all these are legitimate projects
involving truth. Let us consider three of them.
Plato might not have been the first to suggest that knowledge is justified true
belief, but the analysis that later philosophers found in his writings has been tre-
mendously influential (Plato 1976, 97d–98a). Even those epistemologists writ-
ing in response to Gettier’s famous purported counterexamples debate about the
role justification in an analysis or explanation of knowledge—the vast majority of
them take it for granted that truth is a necessary condition for a belief to count
as knowledge.11
Since Tarski, the following is a common reductive explanation of the concept
of logical consequence: a sentence f is a logical consequence of a set of sentence G
iff in every model in which all the members of G are true, f is true.12 This account
appeals to truth in a model, which is, arguably, distinct from truth. Nevertheless,
274 Reflections on the Liar
one might adopt a similar view according to which a sentence f is a logical con-
sequence of a set of sentences G iff necessarily, if all the members of G are true,
then f is true. There is, of course, a vast literature on explanations of logical conse-
quence. What matters for our purposes is that one popular way of explaining one
of our most important concepts appeals to truth.
The final example is the one we use as our primary example in the rest of the
chapter: meaning. When considering philosophical discussions of meaning, it
is common to distinguish between theories of meaning and meaning theories.
A theory of meaning offers an analysis or explanation of meaning; one popular
family of theories of meaning have in common the claim that the meaning of a
sentence is its truth conditions.13
A meaning theory specifies, for some particular language, the meanings of
each of its sentences (e.g., by specifying the conditions under which they would
be true). The connection between theories of meaning and meaning theories is
fairly obvious: a meaning theory specifies for the sentences of particular lan-
guages whatever a theory of meaning says meanings are (e.g., truth conditions,
inferential roles, verification conditions, context change potentials). Very often
linguists and philosophers of language propose a fragment of a meaning theory
that applies only to certain linguistic expressions of a particular language; this
is often called “providing a semantics” for the expressions in question. There are
all kinds of theories of meaning and meaning theories based on them, but truth-
conditional versions already mentioned are popular.14 We take these projects and
others in which truth figures as an explanatory device seriously and think they
constitute valuable projects.
involved in the debate about the nature of truth have for decades treated the para-
doxes like harmless puzzles that have no conceivable impact on their projects (see
Horwich 1998 for an example). So the idea that those interested in pursuing the
projects described above can just ignore the paradoxes associated with truth has
a long history and has been something like the received view. However, in the
last decade or so there seems to be a growing consensus that many of the popular
approaches to the paradoxes are not neutral with respect to theories of the nature
of truth. This attitude has contributed to greater interaction between the two dis-
cussions of truth in the analytic tradition.
One reason for thinking that the paradoxes are relatively harmless for proj-
ects involving truth was the popularity of what we might call monster-barring
approaches to the paradoxes.18 These approaches claim to find some problem
with the purportedly paradoxical sentences, propositions, assertions, or other
items involved in the paradoxes. For example, one holds that the liar sentence
is meaningless and so poses no threat whatsoever. Although some still advocate
these kinds of approaches, they have been on the decline for a long time. And for
good reason. We have plenty of theories of meaning, propositions, etc. and none
of them provide any kind of independent support for a monster-barring view.
Accordingly, we do not find this way of avoiding the problems posed by the para-
doxes promising.
Another way to try to sidestep the problems in question might be to stipulate
that one’s semantic theory need only specify the meanings of sentences that are
commonly used in conversation. The problem here is that it is impossible to tell
which sentences might turn out to be useful in some conversational context. We
mentioned above that the paradoxes almost never have a deleterious impact on
everyday conversations. But the point is not that paradoxical sentences never get
asserted. Rather, if they do happen to get asserted, the participants in the conver-
sation probably won’t know that they are paradoxical and so just continue with
the conversation as if nothing was wrong. Thus, this strategy for avoiding the
problems posed by the paradoxes does not seem very plausible.
instead think of it as defining a new concept that might be used to replace the
concept of truth. However, replacements like this are not going to do all the work
we expect a truth predicate to do because they do not fully satisfy the functions of
endorsement and rejection. Whether these kinds of replacements can do enough
of this work to be worthwhile is a delicate issue we do not take up here.
Perhaps we could look for multiple replacements instead. One could think of
the concepts in the Tarskian hierarchy as replacements for truth. In fact, this is
a plausible way to read Tarski’s own proposal. However, since none of them obey
T-In, they won’t do all of truth’s work either.
Vann McGee proposes two concepts that he thinks will do the work of truth.
One is a vague truth-like concept and the other is a notion of definiteness. McGee
proves some nice results; for example, he can express the theory of the two con-
cepts in one of its object languages without contradiction (see McGee 1991).
Although his truth predicate fails to obey T-In or T-Out, one might be able to use
the definite truth predicate for endorsements and rejections. We have no idea
how his concepts might be used in something like a truth-conditional semantics
or any of the other projects described above.
One of us advocates a pair of replacements as well. One concept obeys T-In but
not T-Out and the other obeys T-Out but not T-In. Thus one serves as a device of
endorsement and the other as a device of rejection. The theory of these concepts
can be expressed in its object language without any worry about contradictions
and the theory is also compatible with classical logic. Finally, the two replace-
ments can do truth’s explanatory work in the projects discussed here—most
notably in formal semantics (see Scharp 2013).
Deciding between various replacements is a complicated business best left for
some other occasion. Instead, we can conclude by recapping the situation with
truth. Deflationists have little reason to replace truth as long as they stick to the
claim that truth predicates should not be used as explanans. Those who do think
that truth plays an explanatory role in a valuable project like truth-conditional
semantics might want to replace truth if they think of a semantics for an entire
natural language as a worthwhile goal and do not see much hope for monster-
barring strategies for avoiding the paradoxes.
Notes
1. We borrow this medical analogy from Chihara (1979, p. 618) who writes: “So, in the end, it
may be wiser to live with the illness than to undergo the kind of surgery needed to remove
all paradox-producing elements.”
2. Perhaps Burgess could avoid this problem by appealing to the fact that his notion of ana-
lyticity is supposed to be context dependent. Either way, we will not pursue this matter
further here.
3. It is common to mention the fact that coreferential singular terms can be substituted in a
truth attribution salva veritate, but this principle seems to be constitutive for any predicate
and is not specific to truth.
278 Reflections on the Liar
4. Shapiro (1991) misleadingly calls a version of this notion the “logical” conception of set, and
uses it to interpret second-order languages.
5. Even if Kitcher’s account is completely inaccurate, we suggest that things could have evolved
in that way. That is enough for present purposes. See also Wilson (1993).
6. Oliver Heaviside once said: “Shall I refuse my dinner because I do not fully understand the
process of digestion.” Another, related Heaviside quip is, “Logic can be patient, for it is eter-
nal.” See http://www.azquotes.com/author/21473-Oliver_Heaviside.
7. Russell’s “no class” theory does away with sets altogether. Certain constructions on attri-
butes are invoked instead.
8. The founding principle of the logical notion of set is (something like) Frege’s Basic Law V,
saying that two concepts have the same extension just in case they are coextensive. The
founding principle of the number operator, now known as Hume’s Principle, a statement
that two concepts have the same number just in case they are equinumerous.
9. See McGee (1993), Field (2001), and Horwich (1998) for formulations of deflationism.
10. See David (1994), Tennant (1997), Kitcher (2002), and Lynch (2009) for various kinds of
inflationism.
11. Gettier (1963). One might think of a knowledge-first view as an exception since it denies that
knowledge has any reductive explanation much less one in terms of truth; see Williamson
(2000).
12. Tarski (1936). See also Etchemendy (1990) and Beall and Restall (2006).
13. Davidson (1967). See Williams (1999) for an alternative interpretation of Davidson’s views
on meaning.
14. See Heim and Kratzer (1998) and Chierchia and McConnell-Ginet (2000) for examples.
15. This led Stephen Yablo to quip (in conversation) “Sure it works in practice, but does it work
in theory?”
16. Kripke (1975) makes much of this observation.
17. Classical logic, intuitionistic logic, and stronger relevance logics (e.g., R) are enough to
derive the contradiction.
18. The term ‘monster-barring’ comes from Lakatos (1976).
19. See Shapiro (2010) for discussion of descriptive vs. revisionary readings of theories of truth.
References
Beall, J. C., and Greg Restall. 2002. Logical Pluralism. Oxford: Oxford University Press.
Berkeley, George, 1734. The Analyst. In William Ewald, ed., From Kant to Hilbert: A Source Book in
the Foundations of Mathematics, vol. 1, 60–92. Oxford: Oxford University Press. 1996.
Boghossian, Paul. 1997. “Analyticity.” In Bob Hale and Crispin Wright, eds., A Companion to the
Philosophy of Language, 331–368. Oxford: Blackwell.
Brown, Bryson, and Graham Priest. 2004. “Chunk and Permeate, a Paraconsistent Inference
Strategy. Part I: The Infinitesimal Calculus.” Journal of Philosophical Logic 33: 379–388.
Burgess, John. 2004. “Quine, Analyticity and Philosophy of Mathematics.” Philosophical Quarterly
54: 38–55.
Chalmers, David. 2011. “Verbal Disputes.” Philosophical Review 120 (4): 515–566.
Chierchia, Gennaro, and Sally McConnell-Ginet. 2000. Meaning and Grammar. 2nd ed. Cambridge,
MA: MIT Press.
Chihara, Charles. 1979. “The Semantic Paradoxes: A Diagnostic Investigation.” Philosophical
Review 88: 590–618.
David, Marian. 1994. Correspondence and Disquotation: An Essay on the Nature of Truth.
Oxford: Oxford University Press.
Davidson, Donald. 1967. “Truth and Meaning.” In Inquiries into Truth and Interpretation.
Oxford: Oxford University Press.
R ev is ing Incons istent C onc e pt s 279
van Eijck, Eric, and Albert Visser. 2016. “Dynamic Semantics.” Stanford Encyclopedia of
Philosophy. Ed. Edward N. Zalta. Winter 2016 ed.
van Heijenoort, Jean, ed. 1967. From Frege To Gödel: A Source Book in Mathematical Logic,
1879–1931. Cambridge, MA: Harvard University Press.
Whitehead, Alfred North, and Bertrand Russell. 1910. Principia Mathematica. Vol. 1.
Cambridge: Cambridge University Press.
Williams, Michael. 1999. “Meaning and Deflationary Truth.” Journal of Philosophy 96: 545–564.
Williamson, Timothy. 2000. Knowledge and Its Limits. Oxford: Oxford University Press.
Wilson, Mark. 1993. “There’s a Hole and a Bucket, Dear Leibniz.” Midwest Studies in Philosophy
18: 202–241.
11
where “P” stands for any sentence and “<P>” stands for a name of that sentence,
the liar paradox arises. And it is only after the truth theorist incorporates some
device (like a Tarskian hierarchy or a Kripkean grounding process) for blocking the
paradox that he is entitled to continue developing his theory.
In this chapter I would like to turn the tables on this traditional conception of
the relation between the liar paradox and the theory of truth. I would like to show
that the theorist of truth need not worry about the liar paradox in developing her
theory, that if she focuses deeply enough on the “material” adequacy of her theory
(i.e., whether it adequately accounts for the nature of truth) as distinct from its
“formal” adequacy (i.e., whether it is a consistent theory), the liar challenge is
unlikely to arise for her theory at all.1
In approaching the liar challenge in this way, we are treating the theory of
truth like most other theories. Consider a physical, psychological, geometrical,
or moral theory. Normally, the theorist in any of these fields is aware that his
theory must satisfy certain norms of formal adequacy. But his focus in construct-
ing his theory is not on this matter. His focus is on the material adequacy of his
theory, on its success in giving a correct, comprehensive, and explanatory account
of the physical structure of the world, human psychology, the geometry of space,
the grounds of human morality, etc. Once his theory is completed (or reaches a
281
282 Reflections on the Liar
temporary state of completion), the theorist has to verify, to the best of his abil-
ity, its formal consistency. Of course, if it turns out that his theory is inconsistent,
this is a serious matter and the theorist must re-examine the theory: its mate-
rial principles, its logical-linguistic framework, its background assumptions, etc.
But under normal circumstances, the need to guard against inconsistency does
not dominate, or even noticeably affect, the construction of our theories. Our
theories are constructed based on material considerations, and the consistency
check is just that, a check. In contrast, the theorist of truth, on the approach
I am challenging, cannot go about his business—the construction of a correct,
comprehensive, and informative theory of truth—in this way. Here, the issue of
consistency is a major concern right from the beginning, and he is not entitled to
proceed in developing the material content of his theory before taking care of its
formal adequacy, or so the common wisdom says.
My claim in this chapter is that the task of constructing a theory of truth is
methodologically more similar to that of constructing a physical, psychological,
geometrical, or moral theory than it is commonly thought to be. More specifically,
if the theorist of truth builds a materially good theory (choosing some—any—
reasonable logical-linguistic framework for working in), then the liar challenge is
unlikely to arise at all for her theory.
The key to this turnaround is methodological, and one way to explain it is
by saying that according to the prevalent conception of the theory of truth,
constructing a theory is like constructing a logical inference. But in fact con-
structing a theory is very different from constructing a logical inference. Let me
explain. If you think of your theory of truth as a logical inference, where each
of its principles is an independent premise and the question is whether these
premises entail a contradiction, then you conclude that if the equivalence prin-
ciple, E, leads to a paradox, then the equivalence principle together with any
other material principle of truth, M, will form a conjunction that also leads to
that paradox:
If this were correct, it would justify the requirement that we take care of the liar
paradox as soon as we accept the equivalence principle. But this view of a theory
is wrong. A theory—say, a philosophical theory—is not like a set of premises in a
logical inference, and the process of building a philosophical theory is not that of
creating a conjunction of independent principles. When you build a philosophical
theory, each added principle constrains, and is constrained by, the other prin-
ciples. That is, it affects, or can in principle affect, their scope and content, hence
their consequences, and is similarly affected by them. As a result, if “☒” is the
operation of placing two principles, P1 and P2, together in a theory, then:
The difference between (2) and (1) is due to the difference between ☒ and &. The
process of constructing a theory is such that the addition of a new principle to
the theory generally involves the updating of all its principles, old and new. We
may say that P1☒P2 is more like [(P1↾P2) & (P2↾P1)], where “X↾Y” stands for “X as
updated by Y”, than like P1&P2. And the fact that for an arbitrary consequence C
explains why
5. It also disconnects the liar paradox and the hierarchical conception of truth
from the question of natural language versus artificial language.
6. It draws new lessons from the existence of non-Tarskian solutions to the liar
paradox.
7. It points to an additional advantage of a substantivist over a deflationist theory
of truth besides the straightforward advantages of greater depth, informa-
tiveness, and explanatory power. Investigating the nature of truth beyond the
minimalist equivalence principle, it shows, might save us from problems—
including formal problems—that arise for deflationist or, more generally,
“bare-bones” theories of truth.
I should emphasize, however, that saying we do not need to consider the liar para-
dox prior to the construction of a material theory of truth does not minimize
the value of facing the liar paradox on its own. On the contrary, confronting the
liar paradox on its own, i.e., independently of the development of a materially
adequate theory of truth that might block it, has led to immense progress in our
understanding of such central topics of philosophical inquiry as language, logic,
semantics, definition, circularity, self-reference, vagueness, contextuality, revi-
sion, truth itself, and more.
the theorist of nature. This challenge has been widely discussed by philosophers
and scientists under the rubric of “the disunity of science” and, not surprisingly,
many of its lessons are independent of the empirical nature of natural science.
In Sher (2004) I suggested that philosophers ought to adopt the same common-
sense guidelines for dealing with this challenge as those devised by scientists. The
gist of this idea is concisely captured by Dyson when he says that “every theory
needs for its healthy growth a creative balance between unifiers and disunifiers”
(1988, p. 47).
In the case of truth, my suggestion is that we think of a substantive theory
as a body of substantive principles—some more general, others more particular,
some simpler, others more complex, some manifesting greater inner unity, others
greater diversity, but all interconnected. This strategy frees us to search for, rather
than legislate in advance (e.g., in the form of a definition by fiat), the material
principles of truth, including principles that involve a significant adjustment, or
updating, of other principles.
(ii) Foundational-Holistic Methodology. Another challenge facing the theo-
rist of truth, along with theorists of other philosophical subject matters, is the
“foundational” challenge. Philosophical theories often deal with very basic subject
matters that call for a foundational treatment. But foundational studies in phi-
losophy have come upon serious difficulties and many philosophers view them as
doomed to failure. In Sher (2010 and 2013a) I have suggested that this is due to
the fact that the foundational project is commonly associated with a self-defeating
traditional methodology, the so-called foundationalist methodology. The founda-
tionalist methodology sets strict, unsatisfiable requirements on the foundational
project, including a strict ordering of all areas of knowledge. To conduct a foun-
dational study of an area of knowledge X we must limit ourselves to cognitive
resources generated in areas lower than X. This means that it is impossible to gen-
erate tools for a foundational study of the most basic areas of knowledge, those
lying at the base of the ordering. The foundational study of such areas is bound to
involve either circularity or resources generated higher-up in the ordering, and as
such it violates a fundamental principle of the foundationalist methodology. This
has led many philosophers to shift from the foundationalist to the holistic meth-
odology, but holism is commonly believed to involve a complete renouncement of
the foundational (and not just the foundationalist) project or else acquiescence to
a very limited (e.g., a coherentist or a narrowly naturalistic) foundational project.
In the above-mentioned works I argue that holism is perfectly compatible
with a robust, noncoherentist, non-narrowly-naturalistic foundational project.
The foundational project, rightly conceived, is the project of substantively and
critically studying the main principles of some basic subject matter, with special
emphasis on explanation and justification. Now, if we think of this project as one
that is pursued in ways appropriate for humans (rather than, say, for gods), then
we are seeking not absolute certainty or instantaneous knowledge (achieved by a
flash of intuition), but progress achieved by probing inquiry, carefully monitored
Tr uth and Trans cend e nc e 287
limit on the complexity of thoughts, on whether they address their subject matter
literally or nonliterally, directly or indirectly. Given our broad conception of world,
thoughts themselves are part of the world, and therefore the category of imma-
nent thoughts includes thoughts directed at thoughts. The category of immanent
thought is very broad, but not all thoughts are immanent. In particular, thoughts
that do not intend, or do not succeed, in saying something about the world or in
attributing a property (relation) to some things in the world are not immanent.
To be immanent is to genuinely attribute a property to something in the world,
something of the kind that the property in question can be attributed to, and
this is not a trivial thing. Normally, the category of immanence encompasses all
statements and theories of all genuine fields of knowledge as well as large parts
of everyday discourse. But depending on the type of thought involved (and/or the
type of objects and properties involved) immanence might pose certain specific
requirements, as we will see in the case of truth-thoughts—thoughts that attribute
truth-properties to thoughts—below. Desirably, the boundaries of immanence
are delineated in a systematic manner. But even before an adequate systemati-
zation is available, we can use our judgment to decide, in many particular cases,
whether a given thought is immanent.10 A nontrivial presystematic candidate for
a nonimmanent thought is “The number 2 is laughing”.11
While immanence is a basic condition for truth, immanence by itself does not
suffice to yield truth. To focus on the world, to say something about it, is not yet
to approach it through the prism of truth.
Transcendence. A second condition for the emergence of truth as a fundamental
standard for human cognition is transcendence. By “transcendence” I mean the
ability, and actual practice, of moving outside a given thought in order to reflect upon
it, examine it, say something about it, ask and answer questions about it, set norms or
standards for it, challenge it, attribute properties to it, and so on. The transcendence
required for truth is transcendence of immanent thoughts. Henceforth, we will
understand by “transcendence” transcendence of this kind.
Transcendence has fallen into disrepute lately. To say that truth is transcen-
dent, it is claimed, is tantamount to saying that we have a “God’s-eye view” on
language and the world. But “transcendence” does not need to have this connota-
tion, and in our own use of this word it does not. Transcendence, as we under-
stand it here, is not something mysterious or superhuman; rather, it is something
quite simple and commonplace. Transcending an immanent thought, or a region
of immanent thoughts, is casting a reflective look at it from a standpoint that
holds it in view yet is located within the purview of human thought. In accordance
with our foundational-holistic methodology, our conception of transcendence is
holistic. Transcending a thought is finding a standpoint anywhere on Neurath’s
boat from which we can see it. Transcendence of this kind is also dynamic. We
transcend sociology in order to view it from a philosophical standpoint, and we
transcend philosophy in order to view it from a sociological standpoint. What
counts as appropriate transcendence varies according to task and circumstance.
290 Reflections on the Liar
If our task is to describe the syntax of a given theory, then, as Gödel (1931) has
shown, if the theory is sufficiently rich, we can view its syntax from within it.
I.e., a theory itself can provide an adequate standpoint for examining some of its
features. But if our task is different, we may need a standpoint outside it. While
the principle of transcendence, like the principle of immanence, calls for a sys-
tematization, here too we are able to judge in many particular cases whether an
appropriate transcendence is achieved prior to a full systematization.
Although transcending an immanent thought or a domain of immanent
thoughts is merely a human act, it is a cognitively powerful act. Gödel’s complete-
ness and incompleteness theorems, Church’s thesis, Turing’s proof of the unsolv-
ability of the halting problem, the Löwenheim-Skolem theorem, Lindström’s
theorems, all testify to power of the transcendent mode in one of its forms—
ascent to a metalanguage.
Transcendence is central to numerous fields of knowledge: psychology, soci-
ology, many areas of philosophy, metalogic, metamathematics, etc. On the one
hand, the very idea of metalogic or of philosophy (sociology, psychology) of, say,
knowledge, requires transcendence; on the other, transcendence enables us to
develop tools that are needed in these disciplines.
It is important to recognize that transcendence does not conflict with imma-
nence. On the contrary, most transcendent thoughts are immanent. Their tar-
get is something in the world—human thoughts; they attribute properties to
their target thoughts, relate them to other thoughts as well as to things other
than thoughts, and so on. As such, they are genuinely immanent. There are many
types of transcendent thoughts, and one of these is the truth-thought—a thought
that attributes a truth-property to some thought(s). To be an admissible truth-
thought, i.e., a truth-thought that is admissible as a truth-bearer, a truth-thought
must be both immanent and transcendent, and its object(s)—the thought(s)
to which it attributes a truth property—must also be immanent. I will call this
interplay between immanence and transcendence the ‘immanence-transcendence
complementarity’.
Immanence and transcendence by themselves, however, are still not sufficient
for truth. By transcending an immanent discourse either to a higher level of dis-
course or to a standpoint on Neurath’s boat that has it in view, we can do many
things that are not related to truth: we can ask questions of a variety of kinds
about these thoughts (e.g., are they mathematical thoughts), attribute properties
and relations of multiple types to them (e.g., being thoughts about something
funny), set standards (norms) of different sorts for them (e.g., clarity), doubt,
challenge, justify, refute them on diverse grounds, enjoy them or be disgusted
by them, etc. Truth arises when we ask questions of a special kind about imma-
nent thoughts and set standards (norms) of a special kind for such thoughts. This
requires a third mode of thought.
Normativity. A third condition for the emergence of truth as a fundamental
standard for human cognition is the ability to engage in normative activities and
Tr uth and Trans cend e nc e 291
(QT) Is the world the way the immanent thought X says it is? Do the
objects that X talks about have the properties, or stand in the
relations, that it attributes to them?
. The question of truth arises for all and only immanent thoughts.
A
B. Truth is a normative-transcendent standard for immanent thoughts. (Truth
and falsehood (nontruth) are normative-transcendent properties of immanent
thoughts.)
C. Truth-thoughts are immanent, transcendent, and normative.
The immanence thesis has significant ramifications for other principles of truth,
including the equivalence principle and, through it, the liar paradox. Since the
equivalence principle and the liar paradox are commonly viewed as pertaining
to small units of thought—sentence-like units rather than whole-theory-like
units—from now on I will identify “thought” with “sentence-like thought” or, for
short, “sentence”.
P is true iff the world is as P says it is, i.e., on one formulation, “<P> is true iff
P”. In this way, the immanence thesis supports, or gives rise to, the equivalence
principle, E. But the immanence thesis, and in particular two of its subprinciples,
immanence and transcendence (with significant help from normativity), also
restrict E by setting certain constraints on it. Two such constraints are:
(Im) For E to hold for any given sentence P, P must be a bona fide
immanent sentence.
(Trans) If P is a truth-sentence, then for E to hold for P, P must be
appropriately transcendent. (The truth predicate of P must be
appropriately transcendent.)
Accordingly, our theory of truth sanctions only those instances of E that satisfy
the two material conditions, Im and Trans. Three instances of E that are sanc-
tioned by these conditions are:
E1 says that the immanent sentence “Snow is white” satisfies our (transcendent)
standard of truth, or has the (normative-transcendent) property of being true,
iff the world is as the sentence “Snow is white” says it is, i.e., under one formula-
tion, iff snow is white. E2 says that the immanent truth-sentence “ ‘Snow is white’
is true” satisfies our (transcendent) standard of truth, or has the (normative-
transcendent) property of being true, iff the world is as the immanent truth-
sentence “ ‘Snow is white’ is true” says it is, i.e., under one formulation, iff “Snow
is white” is true. And E3 says that the immanent truth-sentence “ ‘Snow is white’
is false” satisfies our (transcendent) standard of truth, or has the (normative-
transcendent) property of being true, iff the world is as the truth-sentence “ ‘Snow
is white’ is false” says it is, i.e., under one formulation, iff “Snow is white” is false.
All these instances of the equivalence principle satisfy the requirements Im and
Trans: “Snow is white”, “ ‘Snow is white’ is true”, and “ ‘Snow is white’ is false” are
appropriately immanent, and the last two are appropriately transcendent.
Now, it follows from the material constraints Im and Trans that EP ‒ E as applied
to P ‒ is an inadmissible instance of E in at least two circumstances:
I.e., all instances of E that satisfy either one of the material conditions (a) or
(b) are blocked by our substantivist theory of truth.
294 Reflections on the Liar
where “P” stands for any sentence satisfying Im and (where applicable) Trans, and
“<P>” stands for a name of this sentence.
where “L” abbreviates “ ‘L is false’ ”. We arrive at the liar paradox by replacing “ ‘L
is false’ ” with “L”, getting:
Our theory, however, blocks LP by ruling, based on the material conditions Im and
Trans, that EL is an inadmissible instance of E or, alternatively, that the liar sentence,
(L) L is false.
is an inadmissible truth-bearer.
Before explaining why L/EL fails to satisfy Im and Trans, let us set the context
for our explanation by a few preparatory and background comments:
(i) In discussing whether L satisfies the Im and Trans conditions we will not
appeal to any mechanical (or semimechanical) test. This is connected with the
fact that at this stage we are not committed to any specific systematization
of the immanence and transcendence principles. Systematization has many
benefits and is an important goal of theorizing; but it often comes with a
price: an ad hoc treatment of outliers.16
Tr uth and Trans cend e nc e 295
(ii) Given our cognitive-epistemic perspective, our goal is not to rule out only
(or specifically) paradoxical sentences as admissible truth-bearers; our goal
is to rule out any sentence that from a cognitive perspective is not a genuine
truth-bearer. This distinguishes us from theorists like Kripke who aim at rul-
ing out only paradoxical sentences. For example, for us it might be just as
important to reject
(T) T is true
as to reject L as an admissible truth-bearer.17
(iii) Since the liar paradox is commonly thought to enter theories of truth through
some instance of E, in this chapter I focus on E as a “gate” to the Liar. But it
is worthwhile to note that from a cognitivist-epistemic perspective E is less
central to theories of truth than it is often taken to be. The connection of
immanent sentences to the world (through their truth conditions) is very
central, but E itself offers just one particular way of representing this connec-
tion and there could in principle be alternative representations. For example,
in the case of the immanent sentence “Snow is white” it is possible to repre-
sent its connection to the world (through its truth conditions) by a non-E-
biconditional such as
“Snow is white” is true iff snow reflects light of all huescompletely and
diffusely.
And for some purposes (e.g., informativeness) this representation might be
superior to the more common E-representation:
“Snow is white” is true iff snow is white.
Thus one possible way of blocking the paradox could be to remove E from the
theory altogether and introduce a different equivalence principle in its place.
Rejecting E altogether, however, would be ad hoc, since most instances of E are
true. For that reason I prefer to approach the task by limiting the scope of E.
(iv) Given that L is a self-referring sentence and that this trait is sometimes
associated with its paradoxicality, the question naturally arises whether
immanence and transcendence ban (either individually or together) all self-
reference. The answer is no. For example, the sentence
(S) S is short,
is self-referring, yet in spite of this it is not banned by either Im or Trans.
Since S is not a truth-sentence, it is not subject to Trans. It is subject to Im
but it has no difficulty satisfying Im, since it attributes a property applicable
to syntactic entities (being short) to a bona fide syntactic entity in the world,
namely, to itself as a syntactic entity, or to the syntactic facet of itself (as a
sentence).
296 Reflections on the Liar
We are now ready to turn to the liar sentence, L, and the corresponding E-
instance, EL. L says of itself that it is false, i.e., it attributes a truth-property
to itself. As such it is a truth-sentence and therefore subject to both Im and
Trans. The question is whether L satisfies these requirements. To satisfy Trans,
the operation of truth predication in L must be appropriately transcendent to
the object it is predicated of, namely L itself. I.e., in light of the conditions set
on transcendence by the normativity of truth (see last paragraph of subsection
of normativity) L minus “is false” must be immanent. But once we remove “is
false” from L, no immanent sentence is left. So L fails to satisfy Trans: there is
no appropriately immanent object for L to transcend.19 It follows that L is an
inadmissible truth-bearer according to our theory’ hence EL is an inadmissible
instance of E and the liar paradox (in its classical form) is blocked.
We have shown that the liar paradox can be blocked by material principles
arrived at in the course of developing a substantive theory of truth. To do that
we have focused on one, albeit classical, form of the paradox. This is sufficient to
Tr uth and Trans cend e nc e 297
make our main point, but the question naturally arises whether the same material
principles are capable of blocking other forms of the paradox as well.
Let us examine two other liar sentences, leading to other forms of the liar para-
dox. First, let us consider a type of Liar exemplified by
To show that (7) is not an admissible truth-bearer we note that for (7) to be an
admissible truth-bearer, it must satisfy Trans, and this requires both that (8) be
an admissible truth-bearer and that (7) be appropriately transcendent to (8). But
for (8) to be an admissible truth-bearer, it has to be appropriately transcendent
to (7), and this is impossible. Transcendence is an antisymmetric relation, hence
(7) and (8) cannot mutually transcend each other. As in the case of L, (7) fails to
satisfy Im.
The next “liar sentence” we consider is only contingently a liar sentence. Let
me explain. So far we have considered sentences that give rise to the liar paradox
due to their own semantic structure and/or the semantic structure of other sen-
tences they are relevantly connected to. Kripke (1975) shows that the liar paradox
can arise due to circumstances that hold in the world, so that a sentence with
a given semantic structure can be paradoxical under some circumstances and
non-paradoxical under others. Are liar sentences of this kind ruled out by Im and
Trans? And if they are, are they ruled out in all circumstances or only in problem-
atic circumstances? The answer is that they are ruled out in problematic circum-
stances (of the kind considered by Kripke) but not in all circumstances.
Kripke’s example of a contingently liar sentence is:
Constellation of Circumstances #1:
(i) One of Nixon’s assertions is
(10) Everything Jones says about Watergate is true (Kripke 1975, p. 691);
(ii) Aside from this sentence, ‘Nixon’s assertions about Watergate are evenly bal-
anced between the true and the false’ (Kripke 1975, p. 691), and
(iii) Jones made exactly one assertion about Watergate (9).
298 Reflections on the Liar
Clearly, under this constellation of circumstances (9) does not yield a paradox, nor
is it rejected by Im or Trans.
With this we have completed the task of showing that the liar paradox can be
blocked by substantivist theories of truth due to their material principles. It fol-
lows that philosophers need not impose liar-motivated constraints on their theo-
ries of truth in advance, constraints that might unnecessarily limit their options
in constructing their theories. They may wait until a later stage in the construction
of their theory to decide which constraints (among all the alternatives offered by
different solutions to the paradox) are most suitable for their theory, if the need
for such constraints arises at all.
Our next task is to explore how changing our attitude toward the liar paradox
in this way affects our perspective on existent solutions to the paradox and on
some of the factors commonly associated with these solutions.
basis for the construction of a hierarchy of languages and is the lowest element
in this hierarchy. The next language is a so-called metalanguage, ML0 or L1, and
we arrive at it by transcending L0. L1 is fully equipped for (truth-)transcendence: it
has resources for talking about L0, resources for talking about that part or aspect
of the world that L0 is directed at, and resources for setting a standard of truth
for sentences of L0. Furthermore, L1 is itself an immanent language, and as such
has its own metalanguage, L2 (= ML1 = MML0) to which it stands as an object
language. Each metalanguage has a truth predicate applicable to its predecessor
in the hierarchy, and as a result, truth itself is represented by a hierarchy of predi-
cates: T1, T2, T3, …, which apply to sentences of L0, L1, L2, … , respectively.
As a systematization of the immanence- transcendence complementarity
Tarski’s hierarchy has both advantages and disadvantages. Among its advantages
are a clear, sharp, unambiguous, and formally rigorous structure, as well as one
that has proved extremely fruitful in multiple fields of knowledge. Let me briefly
reflect on its fruitfulness. It is a remarkable fact that while Tarski used the object-
language–metalanguage duality as a formal solution to a formal problem concern-
ing the largely material notion of truth, this duality, partly under his influence,
has been thoroughly integrated into our conception of knowledge and has proved
extremely profitable in a number of fields, in particular metalogic and metamath-
ematics. My present suggestion is that one of the reasons this duality has been a
major force in advancing human knowledge is that it captures the cognitively fun-
damental and highly fruitful complementarity of immanence and transcendence.
Among the weaknesses of the Tarskian hierarchy, from our perspective, is a rigid
and inflexible structure, as reflected in, e.g., the exclusion of many self-referring
sentences satisfying Im and (if applicable) Trans. Its rigidity is also reflected in
some artificial aspects of its treatment of truth, for example, the existence of
multiple truth predicates, each limited to a particular language in the hierarchy,
rather than a single predicate, common to the entire hierarchy. Another weakness
is the narrow range of languages that Tarski’s hierarchy encompasses, namely,
its limitation to so-called “formalized languages of the deductive sciences”. These
languages by no means exhaust the full range of cognitively efficacious languages
licensed by the (presystematic) principles of immanence and transcendence.
A different model of the immanence-transcendence complementarity is sug-
gested by Kripke’s solution to the liar paradox.
Kripke’s Internal Hierarchy as a Systematization of the Immanence-Transcendence
Complementarity. In Kripke’s model we have only one language and only one truth
predicate. The idea of immanence is represented by the concept of groundedness
and the idea of transcendence by a stage by stage determination of the extension
(and counterextension) of the truth predicate. Kripke informally describes the
intuition underlying his concept of groundedness as follows:
It has long been recognized that some of the intuitive trouble with
Liar sentences is shared with such sentences as [T21] which, though
300 Reflections on the Liar
The grounded sentences in Kripke’s system are all immanent in our sense, and
most immanent sentences in our sense are included in the class of Kripke’s
grounded sentences.22
The idea of transcendence is represented by Kripke’s conception of the (single)
truth predicate as defined in stages, or its extension as determined in stages. In
Stage 1 the truth predicate, Tr, is assigned an extension (and a counterextension)23
in the set of all grounded/immanent sentences that contain no truth predicate (or
semantic predicates more generally). In Stage 2 the extension (and counterexten-
sion) of Tr is extended to all grounded/immanent sentences whose truth value is
determined in Stage 1. And so on. In this way, the assignment of a truth value to
each immanent sentence in Kripke’s system is made from a standpoint transcen-
dent to it, albeit internal to the (single) Kripkean language, and Kripke’s system
can be viewed as representing the immanence-transcendence complementarity by
an internal hierarchy. Things change, however, once we get beyond the “ ‘minimal’
or ‘smallest’ fixed point” (Kripke 1975, p. 705), i.e., beyond the point in which
the hierarchy is limited to grounded/immanent sentences. Beyond this point
Kripke deals with sentences like T, which are neither grounded nor paradoxical,
and shows how a conventional assignment of truth values to such sentences can
proceed. Since these sentences are nonimmanent (ungrounded), they do not fall
under our immanence and transcendence principles, and as such are not admis-
sible truth-bearers according to our theory.24
One important feature of Kripke’s system is its allowing a sentence to “seek
its own level” (1975, p. 696) in the internal hierarchy. Whether or not a given
sentence has a place in the internal hierarchy, and if it does, what this place is,
Kripke points out, may be determined by empirical circumstances surrounding
the sentence. It is in this context that he introduces the sentence numbered (9) in
the present chapter, which, he rightly says, is an admissible truth-bearer in some
circumstances, inadmissible (and paradoxical) in others.
Among the advantages of Kripke’s hierarchy over Tarski’s from our perspective
are its considerable flexibility and highly dynamic nature. These are reflected in its
ability to recognize the admissibility of sentences that are excluded from Tarski’s
Tr uth and Trans cend e nc e 301
Natural Language and the Cognitive Perspective on Truth. Many of the objections
to Tarski’s hierarchy in the philosophical literature have to do with natural lan-
guage. In a way, Tarski himself invited these objections. In setting the ground for
his theory of truth he described his goal as constructing a philosophical definition
of the concept of truth, and his explanation suggested that it is “the meaning of
the term ‘true sentence’ in colloquial language” (Tarski 1933, p. 152) that such
a definition aims at capturing. Yet Tarski emphasized that a “thorough analysis
of the meaning current in everyday life of the term ‘true’ is not intended here”
(1933, p. 153), and he proceeded to question the very consistency of natural
302 Reflections on the Liar
language on the ground that it leads to paradox: Due to the occurrence of the liar
paradox in natural language, “the very possibility of a consistent use of the expression
‘true sentence’ which is in harmony with the laws of logic and the spirit of everyday lan-
guage seems very questionable, and consequently the same doubt attaches to the pos-
sibility of constructing a correct definition of this expression” (1933, p. 165). Finally,
he formulated his goal as that of defining truth for a narrow class of highly spe-
cialized artificial languages: the “formalized languages” of “the deductive sciences”
(1933, p. 166). It is not surprising, therefore, that many philosophers questioned
Tarski’s approach to the construction of a theory of truth, starting with his sweep-
ing conclusion about the inconsistency of natural language and ending with his
narrow focus on artificial logical languages.26
Leaving the controversy over Tarski’s approach to truth aside, let me briefly
explain how natural language is viewed from the perspective of the theory pre-
sented here—both from its broader substantivist perspective and from its nar-
rower, though still very broad, cognitivist perspective. From a substantivist
perspective the diverse, multifaceted, and sometimes conflicting uses of “truth”
and its cognates in natural language(s) reflect the extraordinarily breadth and
diversity of truth, and an account of the breadth and diversity of truth and their
manifestations in natural language is one of the tasks of a substantive theory
(family of theories) of truth. But it is not the only task, or necessarily the most
important philosophical task, of the theory of truth.
Indeed, there are several complicating factors in focusing the philosophical
study of truth on natural language. Among these are (i) the fact that natural lan-
guage is a natural phenomenon, and (ii) the fact that natural language is a mul-
tipurpose tool. These facts suggest that the use of “truth” in natural language is
shaped by multiple, sometimes conflicting and often accidental determinants, and
that as a result natural language is less than an optimal source of understanding
the role of truth in specific human endeavors, including cognition or knowledge.
This is the point at which we, given our interests, distance ourselves from natural
language; but it is also the point at which we distance ourselves from Tarski’s for-
malized languages. We are investigating truth from a specific, yet relatively broad
perspective, the cognitivist-epistemic perspective, and this means that on the one
hand we cannot limit ourselves to the behavior of truth in formalized languages,
but on the other hand we are not concerned with all its manifestations in natural
language. Moreover, our interest in truth is partly normative. This means that
as far as language is concerned we are interested in a model of language that is
recognizably a model of human language, but one that is geared toward effective
cognition of the world rather than toward other tasks. Thus, from our perspective,
a study of the behavior of the term ‘true’ in natural language is valuable in under-
standing some aspects of truth, but for understanding the role of truth in human
knowledge and cognition natural language is secondary.
The Immanence-Transcendence Complementarity and Bivalence. It is common
to contrast Tarski’s and Kripke’s hierarchies not just along the formal-language
Tr uth and Trans cend e nc e 303
versus natural-language dimension, but also along the bivalence versus triva-
lence (or more generally nonbivalence) dimension. From my perspective, these
contrasts are only partly warranted. In particular, the viability of Tarski’s and
Kripke’s hierarchies as models of the cognitive immanence-transcendence com-
plementarity has little to do with these contrasts. Focusing on the bivalence-
trivalence dichotomy, let me emphasize, first, that the immanence-transcendence
complementarity is perfectly compatible both with classical logic and with non-
classical logic, both with classical set theory and with nonclassical set theory.
There is nothing in my account of immanence or transcendence to favor one logic
(or set theory) over the other. If the world, or some aspects of the world, exhibit
a tripartite rather than a bipartite property structures, then immanence will be
better represented by a nonclassical logic (and/or a nonclassical set theory) than
by a classical one. The account of the formal structure of properties will have a
trivalent rather than a bivalent operation of complementation, so that given a
domain D and a property P, complementation will partition D into three rather
than two regions. Transcendence would then require a standpoint from which
we can view the world in its tripartite structure, and the determination of truth
values would be made based on this view of the world, hence involve three rather
than two truth values. There is thus no intrinsic sense in which Tarski’s notions
of object language and metalanguage, or my notions of immanence and tran-
scendence, or my material conditions of Im and Trans, are tied up with, or limited
to, classical logic or bivalence. Significant structural differences will still exist
between Kripke-like and Tarski-like models of the immanence-transcendence
complementarity, but the differences between formalized and natural language
on the one hand and bivalence versus trivalence on the other are not central
to them.
Notes
1. The contrast between “material” and “formal” is intended to invoke the similar contrast in
Tarski (1933), but my notion of “material adequacy” is not limited to “extensional correct-
ness,” as Tarski’s notion is sometimes thought to be. “Material,” in the sense intended here,
has the connotation of “internal to the subject matter of the theory,” whereas “formal” has
the connotation of “having to do with formal matters that are extraneous to the subject
matter of the theory.” (When the subject matter of a theory is itself formal in the sense of
dealing with, say, logical or set theoretical issues, the contrast between “material” and “for-
mal” in the present sense is reduced or altogether disappears.)
2. Here and below the arrows represent metalogical consequence relations, or more broadly,
monotonic consequence relations.
3. For various aspects of this theory, including further discussion of issues that I will only
briefly gloss over below, see Sher (1999a, 2004, 2012, 2013b, and 2015).
4. I intentionally construe potential truth-bearers very broadly.
5. “Cognitive” and “epistemic,” or “cognition” and “knowledge,” are used here as cognate
notions, with “cognition” being the broader, weaker, and less specific of the two. While
“knowledge” connotes success in cognition, “cognition” connotes mainly an attempt at
acquiring knowledge (information).
304 Reflections on the Liar
References
Dyson, F. 1988. Infinite in All Directions. New York: Harper & Row.
Frege, G. 1892. “On Sense & Nominatum.” Translated by H. Feigl in Readings in Philosophical
Analysis, ed. H. Feigl and W. Sellars, 85–102. New York: Appleton-Century-Crofts, 1949.
Gödel, K. 1931. “On Formally Undecidable Propositions of Principia Mathematica and Related
Systems.” In Collected Works, Volume 1. Eds. S. Feferman et al., 145–195. New York: Oxford
University Press, 1986.
James, W. 1890. “The Stream of Thought.” The Principles of Psychology, Vol. 1: 224– 290.
New York: Cosimo, 2007.
Kripke, S. 1975. “Outline of a Theory of Truth.” Journal of Philosophy 72: 690–716.
Quine, W. V. 1970. Philosophy of Logic. New York: McGraw Hill. 2nd ed. Cambridge, MA: Harvard
University Press, 1986.
———. 1981. “Things and Their Place in Theories.” In Theories and Things, 1–23. Cambridge,
MA: Harvard University Press.
———. 1986. “Reply to Harold N. Lee.” In L. E. Hahn and P. A. Schilpp, eds., The Philosophy of W.V.
Quine, 315–318. La Salle, IL: Open Court.
———. 1995. “Reactions.” In P. Leonardi and M. Santambrogio, eds., On Quine: New Essays. 347–
361. Cambridge: Cambridge University Press.
Sher, G. 1999a. “On the Possibility of a Substantive Theory of Truth.” Synthese 117: 133–172.
———. 1999b. “What Is Tarski’s Theory of Truth?” Topoi 18: 149–166.
———. 2004. “In Search of a Substantive Theory of Truth.” Journal of Philosophy 101: 5–36.
———. 2010. “Epistemic Friction: Reflections on Knowledge, Truth, and Logic.” Erkenntnis
72: 151–176.
———. 2012. “Truth & Knowledge in Logic & Mathematics.” In M. Peliš and V. Punčochář, eds.,
Logica Yearbook 2011, 289–304. London: College Publications, King’s College.
———. 2013a. “The Foundational Problem of Logic.” Bulletin of Symbolic Logic 19: 145–198.
306 Reflections on the Liar
——— . 2013b. “Forms of Correspondence: The Intricate Route from Thought to Reality.”
In N. J. L. L. Pedersen and C. D. Wright, Truth and Pluralism: Current Debates, 157–179.
New York: Oxford University Press.
——— . 2015. “Truth as Composite Correspondence.” In T. Achourioti, H. Galinon, H.
Fujimoto, and J. Martinez-Fernández, eds., Unifying the Philosophy of Truth, 191–210.
Dordrecht: Springer.
Siewert, C. 2016. “Consciousness and Intentionality.” Stanford Encyclopedia of Philosophy. Ed.
Edward N. Zalta. Fall 2016 ed.
Tarski, A. 1933. “The Concept of Truth in Formalized Languages.” Translated by J. H. Woodger in
Tarski, Logic, Semantics, Metamathematics, ed. J. H. Woodger and J. Corcoran, 152–278. 2nd
ed. Indianapolis: Hackett, 1983.
Wittgenstein, L. 1921. Tractatus Logico-Philosophicus. Trans. D. F. Pears and B. F. McGuinness.
London: Routledge & Kegan Paul, 1961.
12
Approaches to truth and the liar paradox seem invariably to face a dilemma.
Either they must eventually appeal to some sort of hierarchy, or else they must
say that some apparently perfectly coherent concept is in fact incoherent. But
each option would seem to involve severe expressive restrictions, and would thus
seem unsatisfactory. The aim of this chapter is a new approach, which avoids
the dilemma and the expressive restrictions. Previous approaches often try to
solve the Liar by appealing to some sort of new semantic value for the truth
predicate to take; but I will argue that approaches of this sort inevitably face
the dilemma in question. In contrast, the approach that I will propose will stick
to classical semantic values, but will allow that the compositional rules associ-
ated with these have exceptions. One might worry that it will be impossible to
develop such an approach rigorously and systematically—dispensing, as it does,
with exceptionless compositional rules. But I will explain how such a develop-
ment is in fact possible.
The structure of the chapter is as follows. In section 12.1 I will explain how
existing approaches tend to face the dilemma in question, and why neither horn
would seem satisfactory. In section 12.2 I will outline the proposed approach and
explain why—if it could be adequately developed—it would seem to promise a
way out of the dilemma. In section 12.3 I will explain how the approach can be
adequately developed. In section 12.4 I will consider objections. And in section
12.5 I will discuss the implications for logic, if the approach is correct.
307
308 Reflections on the Liar
The aim of the approach of Kripke (1975) is to give an account of how lan-
guages can contain their own truth predicates—despite the existence of (for
example) sentences that say of themselves that they are not true. The basic idea
is to give the truth predicate a new sort of semantic value. Thus, rather than a
simple extension, the truth predicate is assigned a pair of disjoint sets: an exten-
sion together with an ‘antiextension’, where the idea is that the predicate is true
of the things in its extension, false of the things in its antiextension, and neither
true nor false of everything else. This new sort of semantic value for the truth
predicate necessitates an extended approach to the logical operators (i.e., the con-
nectives and the quantifiers). For suppose that T is the truth predicate, and that
Tc is neither true nor false. What then about ¬Tc, for example? Is this sentence
true, false, or neither? In fact, Kripke’s general approach works for a number of
different ‘evaluation schemes’ (i.e., a number of different extended approaches to
the logical operators). But the approach that he focuses on—and which commen-
tators have followed him in focusing on—is the ‘strong Kleene’ scheme; so I will
similarly focus on this scheme here.2 Thus, use u (for ‘undefined’) for the value
taken by sentences that are neither true nor false. Then, according to the strong
Kleene scheme, ¬ is interpreted so as to correspond to the function that sends t
to f, f to t, and u to itself: so ¬A is true if A is false, false if A is true, and neither if
A is neither.3
This understanding of ¬ allows languages to contain their own truth
predicates—despite the presence of liar sentences. For, on this approach, if T is a
truth predicate for a language, then its extension will be the set of true sentences
of the language. Thus, suppose that T is a truth predicate for a language that itself
contains both T and a sentence ¬Tc, where c denotes this very sentence (which
I will call B). B clearly cannot be either true or false: if it was true, then (by the
strong Kleene understanding of ¬) Tc would be false; i.e., the denotation of c (= B)
would be in the antiextension (and not the extension) of T; but B being true then
contradicts the hypothesis that T is a truth predicate for the language; similarly, if
B is false then (by the strong Kleene treatment of ¬ again) it follows that B is in the
extension of T, and this again contradicts the hypothesis that T is a truth predi-
cate for the language (since, by hypothesis, B is false, and so not true). But, on the
other hand, we do not get any sort of contradiction if we instead say that B is nei-
ther true nor false: in this case all we get from the strong Kleene understanding of
¬ is that Tc is also neither true nor false, and thus that the denotation of c (= B) is
in neither the extension nor the antiextension of T; but this of course in no way
contradicts T being a truth predicate for the language. Thus, this understanding
of negation seems to allow for languages that contain their own truth predicates,
even if they also contain liar sentences.4
To see the limitations of this approach, however, one need only consider a
simple alternative to the strong Kleene understanding of negation. For, once one
allows predicates to have pairs of sets as their semantic values, there is then
more than one natural treatment of negation. Thus, strong Kleene negation
Tr uth , Hierarchy, and Inc ohe re nc e 309
surely essential to philosophy that we be able to express such things, and so this
horn of the dilemma would seem to be unsatisfactory.
What about the second option, of saying that exclusion negation is inco-
herent? The problem is that this option also seems to involve severe expres-
sive restrictions. For even if this option allows for general truth predicates, it
does so only at the cost of a restriction on what one can say using them. For
although one could (on this option) perhaps express the general claim ‘Nothing
is both true and false’, there will be other very natural claims that one will not
be able to express. For example, the version of the claim ‘Everything is either
true or not true’ that uses exclusion negation (and I hope that I have already
made clear just how natural a version of this claim this is). Thus this option
also seems to lead to severe expressive restrictions, and would thus also seem
to be unsatisfactory. Hence, overall, the problem that the dilemma raises for
Kripke’s approach.7
And a similar dilemma (and hence a similar problem) would seem to be faced by
a wide range of other approaches to truth. For reasons of space, I will not discuss
any of these in detail. But, in the remainder of this section, I will indicate how a
number of these face a version of the dilemma.
Thus, consider first the ‘revision theory’ of truth (see, e.g., Gupta 1982;
Herzberger 1982; and Gupta and Belnap 1993). On this approach, truth pred-
icates have ‘revision rules’ as their semantic values: i.e., functions from ‘hypo-
thetical’ extensions to ‘revised’ extensions. As in the case of Kripke’s approach,
however, there will—on this approach—be natural forms of negation that lan-
guages cannot contain in addition to their own truth predicates, for example,
a form of negation ~ such that ~Tc is true (under a hypothetical extension for T)
iff the denotation of c does not belong to any revised extension for T.8 The upshot
is that the revision theory faces a dilemma similar to that faced by Kripke’s
approach, and where each option would seem to be similarly unsatisfactory.
Two more recent approaches are those of Maudlin (2004) and Field (2008).
Maudlin’s approach is closely related to Kripke’s,9 and so exclusion negation results
in a similar dilemma for this approach. However, a significant part of Maudlin
(2004) is devoted to defending the incoherence horn; specifically, to arguing that
there are independent reasons for thinking that exclusion negation is incoherent
(see pp. 26–67). Put simply, the basic idea is that no legitimate connective can
send u to either t or f, because if a sentence A receives u, then the ‘facts’ do not
determine whether A is the case; but this ‘determinacy failure’ cannot then deter-
mine that another sentence (e.g., ~A) is true (say)—because, essentially, truth or
falsity can only be determined by ‘facts’, and not by the mere absence of these.
Intuitively, however, it is hard to be completely convinced by this: for language
would seem to be a pretty flexible tool, and it is hard to see anything truly inco-
herent in the idea of one sentence being made true simply by another failing to be
so made. I would suggest, then, that even given Maudlin’s defense, his approach
would seem to lead to an unsatisfactory pair of options.
Tr uth , Hierarchy, and Inc ohe re nc e 311
The approach of Field (2008) does not propose a new semantic value for the
truth predicate to take.10 But it does something sufficiently similar that it seems
to face a version of the dilemma similar to those discussed above. Thus, a cor-
nerstone of this approach is the rejection of the law of excluded middle for sen-
tences of the form Tc: i.e., the rejection of (some) sentences of the form Tc ∨ ¬Tc.
Intuitively, it would thus seem that a sentence can fail to be true (in the sense of
T) but still not be ‘untrue’ in the sense of ¬T (as in cases in which the relevant
instance of excluded middle is rejected). But surely we can also introduce a form
of negation ~ such that a sentence ~Tc will be true precisely when the denotation
of c simply fails to be true (in the sense of T). However, Field’s approach cannot
handle languages that contain both such an operator and their own truth predi-
cates, and so one would seem to once again to face a version of the dilemma.11
Thus, a range of approaches would seem to face the ‘hierarchy-incoherence
dilemma’: either they must eventually appeal to some sort of hierarchy of lan-
guages, or else they must say that some simple, and apparently perfectly coher-
ent, concept is in fact incoherent. But, as we have seen, each option would seem to
involve severe expressive restrictions, and so to be unsatisfactory. The aim of the
rest of the chapter is thus an approach that avoids the dilemma.
sentence ¬Ga is true (false) if the semantic value of a does not (does) belong to the
semantic value of G; and a sentence ∃x(Gx ∧ Hx) is true (false) if some (no) object
in the domain is in the semantic value of G and the semantic value of H; and so on.
The proposal, then, is that in cases such as the Liar these rules have exceptions.
However, a very natural worry to have at this point is that it will be impos-
sible to develop any such approach rigorously and systematically. For (one might
think) surely any such development will require exceptionless rules—whereas the
whole idea behind the proposed approach is to dispense with these. I will argue
below, however, that we can in fact give a rigorous and systematic such develop-
ment. Before doing that, however, I will explain why, if such a development were
possible—i.e., if this sort of an approach could be adequately developed—then
that would seem to promise a way out of the dilemma.
In fact, though, it is actually pretty clear why this is so. To see the point, recall
Kripke’s approach, for example. We forced this approach into the dilemma by con-
sidering exclusion negation ~ and a liar sentence ~Tc. In particular, the hypothesis
that T is a truth predicate for a language that contains this sentence led to contra-
diction. But: the derivation of this contradiction made essential use of the compo-
sitional rules associated with the semantic values of ~, T, and c; for example, of the
fact that if the denotation of c is not in the extension of T, then ~Tc will be true.
Thus if we could develop an approach on which the compositional rules fail in such
cases, then this general sort of argument would be blocked. And so the sort of
approach that I have described would seem to offer a way out of the dilemma—if
it could be adequately developed.
However, as I noted, there does seem to be a prima facie concern about the
possibility of such a development. Thus, I now move on to making the case that
this really is possible. In fact, I will argue that we are already in possession of
formal theories of truth that—although they were designed with very different
aims in mind—can be seen to yield natural models of the sort of languages that
I have described (i.e., languages with classical semantic values, but where the
compositional rules associated with these have exceptions).12 In particular, the
formal theories that I have in mind were developed to respect what is sometimes
called the ‘Chrysippus intuition’, which is as follows. Suppose that at time t Zeno
says, ‘What Zeno says at t is not true’. Then, according to this ‘intuition’, Zeno’s
utterance is neither true nor false. But—the intuition continues—if Chrysippus,
having recognized this fact, says, ‘What Zeno says at time t is not true’, then
Chrysippus’s utterance is true—despite the fact that he has used exactly the same
words that Zeno did, arranged in exactly the same way. That is the content of the
‘intuition’.
I will discuss specific formal theories based on this intuition in a moment—
in particular, using one such theory as an illustration, I will explain how these
theories yield natural models of languages whose semantic values are classical,
but where the compositional rules associated with these have exceptions. But
I will first explain in general terms why such formal theories would seem to be
Tr uth , Hierarchy, and Inc ohe re nc e 313
the natural place to look for models of such languages. Thus, on any such theory,
it will be possible for there to be two tokens of the form ¬Tc, where c names the
first of these, and where the first token is neither true nor false, while the second
is simply true (i.e., the first token corresponds to Zeno’s utterance, while the sec-
ond corresponds to Chrysippus’s). But then, from this fact alone, it would seem
to follow that—whatever the compositional rules associated with the language
are—they are going to have exceptions. For here we have two tokens, made up of
the same words, arranged in the same way, but differing in truth value.13 Further,
given that the compositional rules are going to allow exceptions in this way, then
one would expect the semantic values associated with T (for example) to be of the
familiar classical sort: for if one is going to treat the liar sentence as an exception
to the compositional rules, then there would seem to be little motivation for also
introducing a new sort of semantic value for T to take.
Thus, formal theories based on the Chrysippus intuition would seem to be the
natural place to look for models of the sort of language that I have proposed. The
proponents of these theories do not make the point that their theories yield such
models (nor do they seem to see their theories as having any particular relevance
for anything like the hierarchy-incoherence dilemma). But in the next section
I will argue that these theories do in fact yield just the sort of models that we
are after.
with truth and falsity for tokens, whereas the version stated here is concerned
rather with truth and falsity for sentences (i.e., types rather than tokens). A full
treatment of truth (and falsity) must ultimately encompass both sentences and
tokens; and a full treatment of the ‘Chrysippus’ phenomenon in particular must
distinguish different tokens of the same sentence (since the utterances of Zeno
and Chrysippus above are two tokens of the same sentence). However, for the
purposes of this chapter it will not be essential to do this, and so I will keep things
simple by focusing exclusively on sentences. Further, one will still get instances of
the Chrysippus-phenomenon, even within this simplified framework: for exam-
ple, in the case of a pair of sentences (i.e., types, not tokens) of the form ¬Ta and
¬Tb, where a and b are distinct names, each denoting ¬Ta; in such a case, the
formal theory will give the result that ¬Ta is neither true nor false (like Zeno’s
utterance), while ¬Tb is simply true (like Chrysippus’s). Thus, even within this
framework, we will get instances of the phenomenon in question.
The procedure for assigning sentences values operates via three rules: one for
assigning the ‘standard’ values t and f, and two ‘failure’ rules for assigning n. The
first of these rules (the ‘standard values rule’) is straightforward, and as follows.
At a typical point in the procedure, one has assigned some sentences a value, while
others are yet to receive one. Thus, one has already partially constructed the final
semantic values of T and F: for one has determined of the t-sentences that they
will end up in the extension of T, and of the f- and n-sentences that they will
not; and similarly for the extension of F (but with t and f switched). Thus, for
many sentences of the language, one will already have done enough to determine
what the standard compositional rules will say about them (i.e., what they will say
about them, given the final semantic values for T and F, together, of course, with
the semantic values of the expressions of the initial language). For example, if A is
Tc, and c denotes a sentence that one has already determined will end up in the
extension of T, then one has already done enough to determine that the composi-
tional rules will say that A is true. The standard values rule simply says that, in such
a case, if a sentence has not already been assigned a value, then one may assign it
the value that the compositional rules will say that it should get.16 It is the ante-
cedent clause of this rule (requiring that the sentence is yet to receive a value) that
ultimately allows for exceptions to the standard compositional rules: for in the
case of a sentence that has been given the value n, the procedure does not allow
one to reassign it a standard value via this rule, and so in the final interpreted
language such a sentence will be an exception to the compositional rules (because
they would always deliver t or f).
I now move on to describing the failure rules (i.e., the rules for assigning n to
sentences). Thus, on the way of thinking about things proposed in this chapter,
these are the rules that determine where exactly the exceptions to the composi-
tional rules will occur. The basic idea behind these rules is that one should assign
n to a sentence if it could never receive a standard value in the ‘proper’ way (i.e.,
via an application of the standard values rule). More specifically, the idea is that
Tr uth , Hierarchy, and Inc ohe re nc e 315
these sentences are identifiable by the sort of ‘referential networks’ they belong
to. Thus, in the case of a liar sentence ¬Tc (where c denotes this very sentence),
this sentence belongs to a ‘loop’: if one tries to work out whether it is true, one
is sent to consider the denotation of c (i.e., the sentence itself). Because of this
the sentence could never receive a value via the standard values rule: for, before
it could, one would have to have determined whether the denotation of c (i.e., the
sentence itself) is in the extension of T; but this will only have been determined
once the sentence has been assigned a value. Similar examples of loops are {Ta},
where a names the sentence Ta; {Tb, Fd}, where b denotes Fd and d denotes Tb;
and {Te → 0 = 1}, where e denotes Te → 0 = 1. Thus, the first failure rule assigns n
to sentences that belong to such loops.17
It is this rule (in combination with the standard values rule) that allows the
final language to respect the Chrysippus intuition. For suppose that c denotes
¬Tc. Then {¬Tc} will be a loop, and so it will be assigned n via an application of
this rule. But now suppose that b is a distinct name of this sentence. Then, once
¬Tc has been assigned n, ¬Tb can be assigned t via the standard values rule (for
one has now done enough to determine that the standard compositional rules
will say that ¬Tb is true, since in assigning ¬Tc n, one has determined that it will
be outside the extension of T; further, the standard values rule can be applied to
¬Tb, because it will not yet have been assigned a value—which is of course what
distinguishes it from ¬Tc, which has already been assigned one).
The second failure rule assigns n to sentences that belong to a different sort
of referential network: namely, ‘groundless’ sets. Thus, an example of such a set
is {Ta1, Ta2, …, Tan, …} where, for each i, ai denotes Tai+1. As in the case of loops,
it is clear that the members of such a set could never receive a standard value in
the normal way (i.e., via the standard values rule). Thus, the second failure rule
assigns n to the members of such groundless sets; and so, as with the members
of loops, these sentences will, in the final language, be exceptions to the compo-
sitional rules.
I will describe these rules more precisely in a moment, but I will first say
something about the end result. Thus, it will be easy to see that by repeatedly
applying these three rules one will eventually be able to assign t, f, or n to each
sentence of the language.18 The end result of this process can be naturally thought
of as follows. We started with a standard interpreted language L, and we have
now constructed classical semantic values for the two 1-place predicate letters
that were added to it. Specifically, the extension of T is the set of sentences we
have assigned t to, while the extension of F is the set of sentences that we have
assigned f to. Further, the sentences of our new language are all either true, false,
or neither (i.e., depending on which they value they received). In the case of sen-
tences that are true or false, it is easy to see that these will have the truth value
that is determined by the standard compositional rules: for such a sentence must
have been assigned t or f via an application of the standard values rule; and this
rule assigns a sentence a value only if it has already been determined that the
316 Reflections on the Liar
standard compositional rules will—on the basis of the final extensions of T and
F—deliver that value for the sentence.19 On the other hand (as I have noted), in
the case of a sentence that is neither true nor false in the final language (such as
a liar sentence), this will be an exception to these standard compositional rules
(for these rules always say that a sentence is either true or false). Finally, this lan-
guage is of course also one that contains its own truth and falsity predicates: for
the true sentences of the language are those that were assigned t, and these are
precisely the contents of the extension of T (and, similarly, the false sentences of
the language are precisely the contents of F). Thus, the language is of just the sort
that we hoped for: for it contains its own truth predicate, and it is a language with
classical semantic values, but where the compositional rules associated with these
have exceptions. We saw above that the general approach proposed here would
seem to offer the natural way out of the hierarchy-incoherence dilemma. What
we are now seeing is that—contrary to an initially natural worry—this approach
can in fact be rigorously developed. It would seem, therefore, that we are making
a strong case for the proposed approach.
I must now finish the task of stating the failure rules of the procedure. For this
we need the notion of ‘calling directly’, which is as follows. Thus, let S be some
stage in the procedure of assigning values to sentences (i.e., S can be thought of
as a partial function from the sentences of the language into {t,f,n}). Say that
a sentence A is ‘determined’ at this stage if one has already done enough to fix
which standard value the compositional rules will say that A should take;20 and
say that A is ‘undetermined’ at S otherwise. A calls directly B at S if there are
A1, … , An that are undetermined at S such that A1 = A; for each i < n, Ai+1 is either
an immediate sentential component21 or an instance of Ai; and An is of the form
Tc or Fc for some closed term c denoting B. For example, it is easy to see that if
E is ¬Ta, where a denotes E, and E has not yet been assigned a value at S, then E
will call itself directly at S. Similarly, if G is Td → 0 = 1, where d denotes G, and
G has not received a value at S, then G will also call itself directly at S. Next, say
that A calls B at S if there are A1, … , An such that A1 = A; for each i < n, Ai calls
directly Ai+1 at S; and An = B.
The rules are then stated in terms of this notion of calling. A set of sentences X
is a loop at S if each member of X calls each member of X at S, and no member of X
calls any nonmember of X at S. For example, it is easy to see that if E is once again
¬Ta, where a denotes E, and E is yet to receive a value at S, then {E} will be a loop
at S; similarly, if G is Td → 0 = 1, where d denotes G, and G has not received a value
at S, then {G} will also be a loop at S. The first failure rule (the loop rule) is as follows:
if X is a loop at S, then one may simultaneously assign n to each member of X. This
rule will thus assign n to liar sentences, truth tellers, the members of ‘liar cycles’,
sentences such as Tb → 0 = 1 (where b denotes this sentence), and so on.
The second failure rule, concerned rather with ‘groundless’ sets, is as follows.
A set of sentences X is groundless at S if (i) every member of X calls some member
of X at S; (ii) no member of X calls a nonmember of X at S; and (iii) no nonempty
Tr uth , Hierarchy, and Inc ohe re nc e 317
subset of X is a loop at S. For example, if (for each i) Ai = Tai+1, where ai+1 denotes
Ai+1, and S is undefined on each Ai, then {A1, A2, …, An, …} will be groundless at
S. A groundless set at S is complete at S if it contains every sentence that calls a
member of it at S. The second failure rule (the groundless sets rule) is as follows: if
X is a complete groundless set at S, then one may simultaneously assign n to each
member of X.
That completes the description of the formal procedure. I hope to have made
clear that the end result would seem to be a natural model of the sort of lan-
guage that we need to understand, if we are to avoid the hierarchy-incoherence
dilemma, and the expressive restrictions that it leads to.
In the rest of the chapter I will (in section 12.4) consider some objections to
the proposed approach, and (in 12.5) consider what the implications for logic are,
if the proposed approach is correct.
this scheme that restricts attention to total interpretations in which the exten-
sions of T and F must be disjoint (a very natural restriction!) will deliver the
result that ¬∃x(Tx ∧ Fx) will come out as true (i.e., it will be assigned t in the
models that use this scheme). Indeed, more generally, the more that one refines
the supervaluationist scheme, the more such generalizations will come out as
true. Thus, although (O1) does point to a real problem with the models of the
last section, it would seem to be one that can be overcome.24
The second objection I will consider is as follows.
In fact, I am inclined to give the objector the benefit of the doubt here. That is, it
seems at least plausible that, given enough tinkering, one will eventually be able
to come up with some notion of ‘similarity’ that will do what she wants it to: i.e.,
Tr uth , Hierarchy, and Inc ohe re nc e 319
yield a sentence E* along the lines of her E, but such that there is no natural
way to express the fact that all sentences ‘similar’ to E* are untrue (because any
sentence expressing this fact would itself be ‘similar’ to E*, and so would say of
itself that it is untrue). And—if such a notion of ‘similarity’ can be found—there
would indeed seem to be an expressive restriction that the proposed approach
leads to.
However, this expressive restriction would seem to be fundamentally differ-
ent from those discussed in section 12.1. For, consider again Kripke’s approach,
for example. In that case either one cannot express the general notion of truth
(if one takes the hierarchy horn), or one cannot express some extremely sim-
ple, and apparently perfectly coherent, logical notion (i.e., exclusion negation,
if one takes the incoherence horn). Either way, the expressive restrictions
entailed will prohibit natural claims that have nothing in particular to do
with paradoxes (for example, ‘Everything is either true or not’, using a general
notion of truth and the natural but prohibited form of negation). This would
seem a much more problematic restriction than that raised by (O2). In the lat-
ter case, all we have are certain ‘strong’ paradoxes that we cannot say certain
things about without ourselves saying something paradoxical (i.e., something
which is neither true nor false). This is a real expressive restriction. But not
being able to say certain things about certain paradoxical sentences seems far
less objectionable than not being able to express important general claims that
do not seem to have anything in particular to do with paradoxes. Thus, (O2)
does not establish a problem with the proposed approach comparable to those
raised in section 12.1.
A (d ) , d = e A (e).
320 Reflections on the Liar
That is, on the proposed approach, there are exceptions (in this sense) to this clas-
sically valid argument-scheme.
Further, other examples are easy to come by. To illustrate, consider again the
models of section 12.3.26 Thus, consider again a liar sentence ¬Tc (call this B).
Once this sentence has been assigned n via the loop rule, 0 = 0 ∧ B (for example)
will be assigned t via the standard values rule. Conjunction elimination (in par-
ticular, P ∧ Q ⊨ Q) will thus also have exceptions on this approach. Conjunction
introduction will as well: for let a denote the sentence 0 = 0 ∧ ¬Ta (for example;
call this sentence E). Then {E} will be a loop, and so it will be assigned n. But once
it has been, ¬Ta (i.e., the second conjunct of E) will be assigned t (as, of course,
will 0 = 0 be). Each conjunct will thus be true, even though the conjunction is not
(giving an exception to P, Q ⊨ P ∧ Q). Therefore, just as there are exceptions to
the standard compositional rules, so there are exceptions to almost any logical
rule one can think of.27
Just as the proposed approach is to be sharply distinguished from those
on which the truth predicate receives a new sort of semantic value, so it is to
be distinguished from those on which a new logic is proposed.28 For it is not
that, on the proposed approach, although classical logical rules have exceptions,
there is some alternative logic whose rules will always preserve truth. Rather,
as I have said, almost any logical rule one can think of will have exceptions on
this approach.29 The situation is thus as follows. Classical logic is correct in the
sense that if A1, … , An ⊨ An+1 is classically valid, and An+1 receives a standard
value, then: if A1, … , An are all true, so is An+1. But in cases in which the conclu-
sion fails to be true or false, even classically valid argument-schemes can have
exceptions—just as (on the proposed approach) the classical compositional rules
will have exceptions. There is not some alternative logic whose rules do not have
this feature. Rather, on the proposed approach, logical rules, like compositional
ones, admit of exceptions.
Indeed, it is hardly surprising that, having set out to avoid the hierarchy-
incoherence dilemma, we have ended up with an approach that does not propose
a new logic. For the most natural way of generating a new logic is by introducing
some new (i.e., nonclassical) sort of semantic value. But we saw that any approach
that does this would seem to face the hierarchy-incoherence dilemma. Thus, we
instead developed an approach with familiar classical semantic values, but where
the compositional rules associated with these have exceptions. It was to be
expected that this would result in ‘classical logic with exceptions’, rather than an
alternative logic with exceptionless rules.
I hope, then, to have gone some way toward developing an approach to truth
and the liar paradox that avoids the hierarchy-incoherence dilemma, and the
expressive restrictions that it leads to.
Tr uth , Hierarchy, and Inc ohe re nc e 321
Acknowledgments
I would like to thank Brad Armour-Garb, Andrew Bacon, George Bealer, Jc Beall,
David Chalmers, Adam Elga, Hartry Field, Eric Guindon, Elizabeth Harman, John
Morrison, Agustín Rayo, Zoltán Gendler Szabó, and members of a seminar at Yale
for comments and discussion.
Notes
1. For discussions of issues closely related to the dilemma that I will raise see, for example,
Kripke (1975, pp. 714–15), Maudlin (2004, pp. 26–67), and Field (2008, pp. 309–324).
2. However, everything that I say could easily be modified so as to apply to any of the alterna-
tive versions of the approach. For these see Kripke (1975, pp. 711–712).
3. The scheme handles other operators as follows. A conjunction is true if both of the conjuncts
are, false if one of them is false, and neither otherwise. A sentence of the form ∀xA(x) is true
if A(x) is true of everything in the domain, false if it is false of something in the domain, and
neither otherwise. Other operators are defined out of negation, conjunction and universal
quantification in the usual way.
4. For further details of the approach see Kripke (1975).
5. Throughout I will assume that the languages discussed contain self-reference, in the sense
that for any formula of the language A(x), there will be a closed term of the language b
denoting the sentence A(b). The reason for doing this is that any viable approach to truth
must work even in the presence of self-reference: for even if one was to prohibit ‘direct’
self-reference (e.g., using ‘Jack’ to name the sentence ‘Jack is short’), still there could be no
hope of prohibiting ‘indirect’ self-reference (e.g., sentences such as ‘Every sentence uttered
by the lottery winner is short’ uttered by the lottery winner). It is then merely a harmless
simplification to assume that all of the languages discussed already contain self-reference,
and that in the sense described.
6. More precisely, the argument that showed that (on this approach) languages cannot contain
both exclusion negation and their own truth predicates can easily be modified to show that
(on this approach) no language can contain both exclusion negation and a general truth
predicate.
7. I should note that Kripke himself seems clearly to prefer the former horn (i.e., ‘hierarchy’
rather than ‘incoherence’) (1975, pp. 714–715).
8. To spell out the argument, let r be the revision rule for T. Then, for a hypothetical extension
X, r(X) is the set of sentences that are true, under the hypothesis that X is the extension of T.
But then consider a sentence ~Tb, where b denotes this very sentence (call it A). Then for any
X, A ∈ r(X) iff for any Y, A ∉ r(Y) (by the definition of ~). So A ∉ r(X). But then A ∈ r(X) after
all: and so we have a contradiction. Thus, as claimed, the approach cannot handle languages
that contains both ~ and their own truth predicates.
9. Although there are significant differences between the two approaches. For example,
Maudlin (2004) contains a distinctive account of the permissibility of assertions (where
permissibility radically diverges from truth; see especially pp. 95–104 and 141–177).
10. A component of this approach is a model-theoretic semantics for languages that con-
tain their own truth predicates. However, this is claimed merely to establish the consis-
tency of the theory, and not to provide a ‘model’ (in the informal sense) of a new sort of
semantic value.
11. The argument can be spelt out as follows. If ~ is introduced as in the text, then it would seem
that it must satisfy the following rules of inference (here ⊥ can be taken either to be a logical
constant, or a given contradiction, such as 0 = 1; and ⊨ stands for logical consequence):
322 Reflections on the Liar
(1) A, ~A ⊨ ⊥, and
(2) if A ⊨ ⊥ then ⊨ ~A.
For, suppose for the purposes of illustration that A is of the form Tc. Then it is surely impos-
sible for the denotation of c to both be true in the sense of T, and also to fail to be; but
then (1) must hold, given how we have introduced ~. Similarly, for (2), suppose that Tc ⊨ ⊥
holds; then the denotation of c must surely fail to be true in the sense of T; i.e., ~Tc must
hold, as required. Thus, given how we have introduced ~, it would seem to be self-evident
that rules (1) and (2) must hold for it.
However, if so, then it follows that no language can contain both its own truth pred-
icate and ~, on Field’s approach. For consider a sentence ~Ta where a denotes this
very sentence. Then Ta ⊨ ⊥ (because, on Field’s approach, one has Tb ⫤ ⊨ B, where
B is the denotation of b; and so Ta ⊨ ~Ta; but then we have Ta ⊨ ⊥ by (1)). But then
⊨ ~Ta (by (2)). And so we have ⊨ Ta and thus ⊨ ⊥ (by (1)): contradiction. It would
seem therefore that Field’s approach faces a version of the dilemma similar to those
discussed above.
12. I should stress that when I say that these languages have ‘classical semantic values’ I mean
that the semantic values of atomic expressions (such as the truth predicate) are always
classical. I do not mean that the semantic values of sentences are always classical: for the
semantic value of a sentence is its truth value, and the classical truth values are just t and
f; whereas in the sort of languages that I have described some sentences (such as liar sen-
tences) will be neither true nor false; and these one can very naturally think of as having
nonclassical semantic values.
13. An alternative possibility would be to try to respect the intuition via some sort of context-
sensitivity. For example, the first token could be interpreted as on Kripke’s approach, while
the second could be interpreted as containing some sort of ‘higher level’ truth predicate.
This would not seem to be the most natural approach, however (for it seems that Zeno
and Chrysippus mean the same thing by ‘true’, and similarly for the other words they use).
Further, this is not the approach that the theories based on the intuition that I discuss
below take, and so I will ignore the possibility of this approach below.
14. For simplicity, I will assume that every member of the domain of L is denoted by some
closed term of L.
15. I should say that in stating this version of Gaifman’s theory I will make various minor
changes to his notation and terminology (in part so as to fit with that used elsewhere in the
chapter).
16. More precisely, the rule is as follows. A given stage of the evaluation procedure corresponds
to a partial function from sentences into {t,f,n}, which in turn corresponds to a ‘partial
interpretation’ of the language. I.e., an interpretation in which T and F are assigned pairs of
extensions and antiextensions: so the extension of T will contain the t-sentences, and the
antiextension will contain the f- and n-sentences, together with everything in the domain
that is not a sentence of the language (and similarly for F, but with t and f switched). The
standard values rule then says that if a sentence A is ‘satisfied’ by this partial interpretation,
and A is yet to receive a value, then one may assign A t; and, similarly, if ¬A is ‘satisfied’,
and A is yet to receive a value, then one may assign A f. But—of course—there are various
ways of understanding ‘satisfies’ here. The simplest is via the strong Kleene scheme, and so
it is this version of the rule that I will use in this statement of the theory. But alternatives
in terms of some version of the supervaluationist scheme are also natural here (see section
12.4).
It is important to note that the use of partial interpretations here is simply an instru-
mental device: it is not that in the final language T or F will be interpreted by pairs of sets;
rather, their semantic values will be single sets (i.e., extensions). Partial interpretations are
simply a natural tool to use in the construction of these extensions (since at a given stage
we have only partially constructed these!).
Tr uth , Hierarchy, and Inc ohe re nc e 323
17. I will state the rule more precisely below, but note that it is only sentences that say some-
thing semantic about themselves that count as belonging to loops (for example). Thus, e.g.,
even if a denotes a = a, {a = a} will not count as a loop in the relevant sense.
18. Although it is less easy to see, this assignment of values is also unique (i.e., it does not
depend on the order in which one applies the rules to particular sentences). For a theorem
to this effect see Gaifman (2000).
19. More precisely, the point is as follows. The standard values rule assigns t (f) to a sentence
A if A (¬A) is satisfied by a given partial interpretation. But the final semantic values extend
the semantic values of this partial interpretation. Thus it follows that the standard compo-
sitional rules, using the final semantic values of T and F, will deliver the same result for A as
the standard values rule did. (This last point follows from the ‘monotonicity’ of the strong
Kleene evaluation scheme, together with the fact that the strong Kleene scheme reduces to
the standard compositional rules in the case of a ‘total’ interpretation.)
20. More precisely, S is determined if either A or ¬A is satisfied by the partial interpretation of
our language that corresponds to S (see note 16). As with the standard values rule, a number
of different understandings of ‘satisfied’ are possible here, but in this statement I will stick
with the strong Kleene understanding.
21. Thus, the sole immediate sentential component of ¬E is E, the sole immediate sentential
components of E ∧ G are E and G, and so on.
22. For, under the strong Kleene scheme, this sentence will be satisfied by a partial interpre-
tation only if every sentence of the language is in either the antiextension of T, or the
antiextension of F. This sentence will thus be eligible to be assigned t (via the standard
values rule) only once every sentence has been assigned a value—which of course means
that this sentence will never be assigned t. (It will also never be assigned f; rather, it will be
assigned n.)
23. A partial interpretation P satisfies a sentence A under the basic supervaluationist scheme
iff every ‘total’ interpretation extending P satisfies A. Alternative versions of the scheme
restrict attention only to those total interpretations satisfying some further condition.
24. The contrast with, for example, Kripke’s approach should be noted. For that approach also
works with a number of different evaluation schemes (including versions of the super-
valuationist scheme). However, moving to these schemes does not, in this case, help with
the problem: for it is easy to see that exclusion negation will still lead to the hierarchy-
incoherence dilemma (the argument of section 12.1 goes through essentially unchanged).
25. A similar objection could be raised in terms of sentences that say of members of ground-
less sets that they are untrue. For, in the models presented in section 12.3, such sentences
will themselves be neither true nor false; and this would thus constitute another sort of
expressive restriction faced by the proposed approach (at least as developed in section 12.3).
This version of the objection could be responded to similarly to the way in which I respond
to (O2) in the text. The reason I focus on (O2), rather than on this alternative objection,
is that there are versions of the proposed approach in which some sentences that say of
members of groundless sets that they are untrue are themselves true. (E.g., versions of the
proposed approach based on versions of Gaifman’s theory with this feature: see Gaifman
(2000, pp. 113–118).) In contrast, the expressive restrictions gestured at in (O2) do not
seem avoidable.
26. That is, I will once again for the purposes of illustration use the formal theory stated in
section 12.3 (i.e., the version of the theory of Gaifman (2000) stated there). However, it is
plausible that any natural development of the basic idea behind the proposed approach will
share the features I am about to describe, at least in essentials. In particular, everything I
will say about the models of section 12.3 will apply to the alternatives that use a version of
the supervaluationist scheme. (Gaifman, in part because he does not emphasize the version
of his theory stated in section 12.3, does not point out that his theory is ‘logically exception-
alist’ in the sense discussed in the text; but it would seem to be one of the most interesting
and important aspects of this theory.)
324 Reflections on the Liar
27. ‘Almost’ because there will be some rules without exceptions: e.g., P ⊨ P and P, ¬P ⊨ Q
(and ⊨ P ∨ ¬P on supervaluationist versions of the theory).
28. Approaches on which a new logic is proposed include the version of Kripke’s approach
discussed in section 12.1 (i.e., the strong Kleene version), and the approaches of Maudlin
(2004) and Field (2008).
29. Thus, although rules of certain special sorts—e.g., those in which the conclusion is among
the premises, or in which the premises can never all be true—will not have exceptions on
this approach (see note 27), the system comprised of only these rules appears too minimal
to count as an ‘alternative logic’.
References
Field, H. 2008. Saving Truth from Paradox. Oxford: Oxford University Press.
Gaifman, H. 2000. “Pointers to Propositions.” In A. Chapuis and A. Gupta, eds., Circularity,
Definition, and Truth, 79–121. New Delhi: Indian Council of Philosophical Research.
Gupta, A. 1982. “Truth and Paradox.” Journal of Philosophical Logic 11: 1–60.
Gupta, A., and N. Belnap. 1993. The Revision Theory of Truth. Cambridge, MA: MIT Press.
Herzberger, H. G. 1982. “Notes on Naive Semantics.” Journal of Philosophical Logic 11: 61–102.
Kripke, S. 1975. “Outline of a Theory of Truth.” Journal of Philosophy 72: 690–716.
Maudlin, T. 2004. Truth and Paradox: Solving the Riddles. Oxford: Clarendon Press.
Skyrms, B. 1984. “Intensional Aspects of Self-Reference.” In R. L. Martin, ed., Recent Essays on
Truth and the Liar Paradox, 119–131. Oxford: Clarendon Press.
13
325
326 Reflections on the Liar
the speakers of L have agreed an explicit semantic theory of L, any more than
the speakers of a natural language have agreed an explicit semantic theory of it.
Rather, as with a natural language, the speakers of L share ordinary linguistic
competence in L, using it as a public language, relying on the publicly available,
informally explained meanings of its sentences. They acquire such competence in
the first place by the direct method of immersion in the language, supplemented
with occasional ad hoc explanations of specific points, usually given by already
competent speakers. As always, this shared competence permits, and indeed
enables, deep disagreements between speakers to be expressed in the common
language.
We are going to define a relation of logical consequence between sets of sentences
of L and sentences of L, but first we need to sketch in some more background.
For simplicity, we assume that all sentences of L are declarative. If L has
context-sensitive expressions, we provisionally fix a preferred context—ideally,
one that leaves quantifiers unrestricted. Taking nondeclaratives and indexicality
into account would complicate the arguments below without seriously imped-
ing them. For concreteness, we further assume that L has at least the expres-
sive power of a first-order language with identity: L has negation, conjunction,
disjunction, universal and existential quantifiers, an identity predicate, and an
appropriate variety of other predicates. We assume that open or closed formulas
of L can be evaluated as true or otherwise on an assignment of values to variables;
after all, L is an interpreted language, not a mere formalism.
The constants of L are the atomic nonvariable expressions of L, of whatever
type. From among them we select a set of logical constants. In Tarski’s original
paper, he left it open where and how the distinction between logical and nonlogi-
cal constants is to be drawn, and indeed whether it can be drawn nonarbitrarily.
In a later paper, he proposed that the logical constants are those invariant under
permutations of individuals (Tarski 1986). If a once-and-for-all criterion of logi-
cality is wanted, something along those lines may be the best we can do. But for
present purposes a once-and-for-all criterion is not wanted. Rather, the choice
of logical constants is pragmatic. In effect, one is investigating which general
structural principles the logical constants satisfy. Varying the extension of ‘logi-
cal constant’ amounts to varying what one is investigating the general structural
features of. For example, in modal logic, if one is interested in the specific struc-
tural features of metaphysical necessity and metaphysical possibility, one may
make the operators □ and ◊ logical constants (typically, in addition to the stan-
dard ones) with the respective intended interpretations (as in Williamson 2013).
Alternatively, one may be interested in the structural features general to all oper-
ators, and make □ and ◊ nonlogical. Both choices are legitimate. Indeed, both
investigations need to be carried out sooner or later. Of course, some choices of
logical constant will send the investigation outside the pragmatically appropri-
ate domain of logic. For instance, if one counts ‘mass’ and ‘energy’ as ‘logical’
constants, one will be doing physics rather than logic in any distinctive sense.
Semantic Parad ox e s and Abduc tive Me thod ol og y 327
Experience and good judgment are needed to determine which choices deter-
mine promising inquiries.
In what follows, we assume that a set of logical constants has been given, and
that it includes at least the usual suspects: negation, conjunction, disjunction, the
universal and existential quantifiers, and the identity predicate. Others will be
added as required.
For each nonlogical constant e of L we extend the language with a new vari-
able ve of the same semantic type as e, where ve = vf only if e = f. For each sen-
tence α of L, let αv be the result of substituting ve for e throughout α for each
nonlogical constant e of L. For each set Γ of sentences of L let Γ v = {α v : α ∈Γ }.
Some of the new variables will be of higher type, for instance those replacing
atomic predicates and sentences, and will need to be assigned values appropri-
ate to their type. If the context restricted the values of the variables to a contex-
tually relevant domain, the assignments would need to meet those restrictions,
but for present purposes we can work on the simplifying assumption that the
quantifiers are unrestricted. Although in this chapter we will continue speak-
ing in terms of ‘values’ and ‘assignments’, a more rigorous treatment would be
articulated in part by quantifiers of suitably higher types in the metalanguage,
in order to avoid Russellian paradoxes with unrestricted quantification.1 These
complications do not affect the broad methodological issues which concern this
chapter.
We can now define logical consequence. A sentence α of L is a logical conse-
quence of a set Γ of sentences of L (Γ ⊨ α) if and only if αv is true on every assign-
ment on which every member of Γv is true. As a special case, α is logically true (⊨ α)
if and only if αv is a logical consequence of the empty set, in other words, αv is true
on every assignment. We can think of αv as representing the logical form of α,
and the pair <Γv, αv> as representing the logical form of the argument from Γ to α,
subject to the stipulation that logical form is invariant under permutations of the
variables of a given type.
Logical consequence in the sense of ⊨ obeys the standard structural rules for a
consequence relation. That is, the following hold for all sentences α and β of L and
all sets Γ and Δ of sentences of L:
Assumption {α} ⊨ α
Monotonicity (Thinning) If Γ α then Γ ∪ ∆ α
Cut If Γ α and ∆ ∪ {α } β then Γ ∪ ∆ β
substitute variables for nonlogical constants and constants for logical or nonlogi-
cal constants, within the same type; thus a uniform substitution may map α to αv
or vice versa. As usual, sΓ = {sα: α∈Γ}. Then:
Someone might ask what logical consequence in the sense of ⊨ has to do with
logic. Notoriously, Tarski’s account does not impose any modal or epistemic con-
straints, yet many philosophers assume that principles of logic should be neces-
sary or a priori or analytic. They often present themselves as speaking on behalf
of some intuitive pretheoretic conception of logic. But it has already been empha-
sized that we are not trying to be faithful to any such conception. Rather, we are
trying to construct a definition that will best serve the purposes of theoretical
inquiry in this general area. We want to define logical consequence rather than
treating it as primitive, because any pretheoretic conception associated with the
phrase ‘logical consequence’ will be too inchoate, too primitive in the adverse
sense, to be adequately constrained. Tarski’s austerely clean and clear definition
is perfectly suited to a discipline as fundamental as logic. In abstracting from the
specific meanings of the nonlogical constants, it enables us to recognize the pat-
terns formed by the logical constants, which pick out the field of our interest.
Once we have that dimension of generality, adding a second dimension of neces-
sity, a priority, or analyticity needlessly complicates the picture, mixing together
questions that our fundamental terminology should hold carefully apart so that
it can represent their interrelations perspicuously.
The wisdom of not adding such accretions to logical consequence is confirmed
when we come to the semantic paradoxes. Suppose that we are confronted with
a liar-like derivation of an absurd conclusion. We want a diagnosis which tells us
where the derivation goes wrong. Roughly speaking, we want to know which step
goes from true to false, or at least to untrue (this is only rough, in part because
‘true’ is itself one of the contested terms in the semantic paradoxes). If we are
told that the conclusion of some step is not a logical consequence of its immedi-
ate premises, because the connection is contingent, or a posteriori, or synthetic,
but the step turns out still to be materially truth-preserving, then we have not
yet been given the needed diagnosis, because we have not been told where to stop
going along with derivation. Of course, we want to know whether the step is mate-
rially good, but that is just an instance of the banal general need to know the
results of a science in order to apply them; it does not stem from any special epis-
temic condition on logical consequence. When we are trying to solve the seman-
tic paradoxes, any special epistemic or metaphysical constraint inserted into the
definition of logical consequence would be a pointless distraction. In answer to
the corresponding worry about the generality condition itself in Tarski’s account,
that generality is the minimal requirement for a theoretically useful relation in
the vicinity.
Semantic Parad ox e s and Abduc tive Me thod ol og y 329
ask which universal generalizations really hold. Such an inquiry resembles an early
stage of mathematics when there is still widespread disagreement over which first
principles we may rely on in proving mathematical results. The investigation has a
more philosophical flavor. Even here, for suitable choices of the logical constants,
no discipline is better fitted than logic to evaluate the universal generalizations at
issue, given their extreme generality.
Of course, the term ‘logic’ covers a wide variety of legitimate inquiries. The
diverse branches of model theory, proof theory, set theory, and recursion the-
ory all count as logic, and these days there are more logicians in departments
of mathematics or computer science than in departments of philosophy, as a
glance at a logic journal will indicate. Those more technical branches of logic
are pursued by purely mathematical methods of an established sort. Moreover,
insofar as they investigate logical consequence at all, they do so from a metalin-
guistic standpoint: their results about the object-language hold independently
of its intended interpretation. By contrast, although the sort of nonmetalin-
guistic inquiry just sketched is logic, it is not primarily metalogic (see also
Williamson 2014). While it in no way supplants those more technical parts of
logic, it has its own fundamental significance. It is, in one good sense of the
term, philosophical logic.
Then, using the standard structural rules above, one can easily show that {α1, …,
αn} Ⱶ β if and only if Ⱶ α1 → (α2 → (… (αn → β) …). Thus, at least for finite sets
of premises, each consequence is encoded in a corresponding theorem, so in a
way one loses nothing by concentrating on logical truth. Even for infinite sets of
premises, Ⱶ may be compact, like the standard consequence relation for first-order
logic, in the sense that Γ Ⱶ α if and only if Γ* Ⱶ α for some finite subset Γ* of Γ, in
Semantic Parad ox e s and Abduc tive Me thod ol og y 333
which case the consequence relation is still reducible to its theorems. Of course,
Ⱶ may be noncompact, like the standard consequence relation for second-order
logic. In that case, we may still be able to reduce logical consequence to logical
truth in an extension of the language with an infinitary conjunction operator ∧.
For suppose that every set of sentences Γ has a conjunction ∧Γ, a single sentence,
where for every sentence α, Γ Ⱶ α if and only if ∧Γ Ⱶ α. Then Γ Ⱶ α if and only if
Ⱶ ∧ Γ → α, as required.
However, for some nonclassical logics such reductions are unavailable. Cases
in point are the weak and strong Kleene logics. They are based on three-valued
tables; we may label the values ‘True’, ‘False’, and ‘Neutral’. True is the only
designated value: Γ Ⱶ α if and only if whenever every member of Γ is assigned
True, so is α. For the strong Kleene tables, if # is an n-place connective then
#(α1, … , αn) is assigned Neutral if every one of α1, … , αn is assigned Neutral.
For the weak Kleene tables, #(α1, …, αn) is assigned Neutral if at least one of α1,
… , αn is assigned Neutral. Either way, any sentence built up by connectives out
of atomic sentences is assigned Neutral on the line of the three-valued table
on which every atomic sentence is assigned Neutral.3 Thus no such formula is
assigned True on every line: the logic has no theorems at all. In particular, no
conditional → is definable in it for which (→I) holds, since that would require
p → p to be a theorem, because {p} Ⱶ p by Assumption. Nevertheless, neither
the strong nor even the weak Kleene logic trivializes logical consequence. For
example, the usual introduction and elimination rules for conjunction still
hold: {α & β} Ⱶ α, {α & β} Ⱶ β, {α, β} Ⱶ α & β. For such logics, we cannot afford
to focus exclusively on logical truth. We especially cannot afford to do so when
comparing nonclassical logics proposed in response to the semantic paradoxes,
since one of the most salient is strong Kleene logic. If we just look at its theo-
rems, it seems powerless; if we look at its consequence relation, it seems mod-
erately powerful.
Even for logics with the full set of classical theorems, theoremhood may be a
poor guide to the consequence relation when →E or →I fails. Here is an extreme
example of a kind opposite to the Kleene logics. For a propositional language, let
Γ Ⱶ α if and only if α is either a classical tautology or a member of Γ. Thus Ⱶ yields
the same set of theorems as classical logic. One can easily check that Ⱶ obeys the
structural rules for consequence relations. Nevertheless, it is radically impover-
ished. In particular, it lacks both →E and →I. For example, where p and q are
distinct atomic sentences, we have {p → q} Ⱶ p → q and {p} Ⱶ p but not {p → q, p}
Ⱶ q, so →E fails, and we have {p, q} Ⱶ q but not {q} Ⱶ p → q, so →I fails. If we just
look at its theorems, the logic seems very powerful; if we look at its consequence
relation, it seems almost powerless.
We need our methodology for comparing logics to be capable of taking seri-
ously logics that lack a connective → obeying →I and →E. Clearly, then, it must
take account of the full consequence relation directly, not just of its precipitate in
theorems.
334 Reflections on the Liar
methodology is the best science provides, and we should use it.5 In particular, we
should use it when comparing the theories generated from a given set of premises
by rival consequence relations.
The idea of applying an abductive methodology to logic and mathematics is far
from new. According to Bertrand Russell, what he calls ‘mathematical philosophy’
The role of such an inquiry is ‘to take us backward to the logical foundations of
the things that we are inclined to take for granted in mathematics’ (Russell 1919,
pp. 1–2).
We cannot attempt a general discussion of abductive methodology here.
However, some specific comments on the criteria may be useful, concerning their
application to logic.
First comes fit with the evidence. At a minimum, one might think, it requires
consistency with the evidence. But that raises the worry that the operative stan-
dard of consistency is set by some transcendental background logic, thereby
undermining the ideal of fair process in comparing alternative logics. Another
worry is that a standard of consistency is unfair to dialetheist treatments of the
semantic paradoxes, which embrace the contradictions and restrict the logic to
block their trivializing consequences.6 We can answer the first worry by using
the logics under test to set their own standard of consistency. Moreover, we can
answer the second worry too by treating consistency as avoidance of trivialization
rather than avoidance of contradiction. To be more precise, let E be the relevant
evidence, expressed in a set of sentences of L. Then a consequence relation Ⱶ is
consistent with E if and only if for not every sentence α of L does it hold that E Ⱶ
α. That criterion should be acceptable even to dialetheists, who are almost as keen
as everyone else to avoid trivialization, as well as to classical logicians and others
who accept that everything follows from a contradiction.
Such consistency with the evidence does not exhaust fit with the evidence.
A theory to which the evidence is simply irrelevant is consistent with the evidence.
A theory fits the evidence better if the evidence verifies some of its predictions as
well as falsifying none of them. Evidence here is not confined to observations. We
may use anything we know as evidence (Williamson 2000). For example, in the
case of propositional modal logic, we may know that the coin could have come up
heads, and could have not come up heads, but could not have both come up heads
and not done so, and on that basis eliminate this proposed law:
(◊p ∨ ◊q)
↔ ◊(p ∨ q)
In that sense, we can verify some predictions of the law by using our pretheoretic
ability to evaluate particular modal claims. The law even goes some way toward
unifying and explaining its instances, by bringing them under an illuminating
generalization.
The criterion of strength also requires clarification in the context of logic. In
one standard logical sense, a theory T is stronger than a theory T* if and only if T
entails T* but T* does not entail T: every theorem of T* is a theorem of T, but not
every theorem of T is a theorem of T*. Similarly, one might call a consequence rela-
tion Ⱶ stronger than a consequence relation Ⱶ* if and only if whenever Ⱶ* holds,
so does Ⱶ, but Ⱶ sometimes holds when Ⱶ* does not. Since we are concerned with
theories in a given interpreted language L, and consequence relations for L, we do
not consider ‘translations’ between languages, or nonhomophonic ‘translations’
of L into itself, in making comparisons of strength.7 However, relative strength
of the logical sort just explained is rarely at issue in the abductive comparison
of scientific theories. Typically, we are comparing internally consistent theories
that are inconsistent with each other (on any reasonable standard of consistency).
But if T is stronger in the sense above than T*, so T entails T*, yet T is also incon-
sistent with T*, then T is internally inconsistent. For abductive purposes, we
need a looser sense in which one internally consistent theory may be stronger
than another with which it is inconsistent, because the former is more specific
or informative than the latter. For instance, ‘The time is between 3.14 and 3.16’
is more specific than ‘The time is between 4.00 and 12.00’, even though they are
inconsistent with each other. To take an extreme case, let T be a consistent scien-
tific theory axiomatized by a conjunction of universal generalizations with many
interesting consequences, and let T* be the consistent theory axiomatized by the
negation of that conjunction. T* is inconsistent with T and concerns the same
subject matter; if the probability of T on our evidence is less than 0.5, then T* is
more probable than T on our evidence. Nevertheless, T* would typically not even
be treated as a rival theory to T, because it is too uninformative: it says that there
is a counterexample somewhere or other to one of the universal generalizations in
T, but nothing more. One role for the informal scientific standard of strength is to
provide a minimal threshold of informativeness below which theories do not even
come up for serious abductive evaluation. We want scientific theories to inform
us about their subject matters; weak theories do too little of that to give us what
we want. Furthermore, strength contributes to explanatory power in the broad
sense sketched above, the capacity to bring our miscellaneous information under
generalizations that unify it in illuminating ways.8
If T is stronger than T* in the strict logical sense, then T is also stronger than
T* in the looser scientific sense, but the converse fails. Both senses are applicable
Semantic Parad ox e s and Abduc tive Me thod ol og y 337
However, two differences may be noted. First, Quine bases his argument on
‘the maxim of minimum mutilation’. By contrast, I argued above that abductive
comparisons of the relevant kind need not make such appeals to a principle of
conservativism. Second, the quotation comes from Philosophy of Logic, in which
Semantic Parad ox e s and Abduc tive Me thod ol og y 341
Quine notoriously claims of the deviant logician: ‘when he tries to deny the
doctrine he only changes the subject’ (1970, p. 81). Here Quine takes disputes
over alternative logics to be verbal. By contrast, I have argued that they typically
involve genuine nonverbal disagreement (a view perhaps closer to that of Quine
1951).9
But wait: the friends of disquotation are not finished yet. Like other oppo-
nents of classical logic, they argue that it fails only in exceptional cases, and can be
recovered in nonexceptional ones. When we need it, we can have it. For example,
many proponents of a nonclassical approach to the semantic paradoxes opt for
the strong Kleene logic K3. True is the only designated value. Liar-like sentences
are supposed to be neither True nor False. The law of excluded middle fails in K3,
because both ¬p and p ∨ ¬p are Neutral when p is Neutral. However, adding p ∨ ¬p
to K3 for each atomic sentence p restores classical logic, since it is True when and
only when p is True or False, and the Kleene tables yield the standard bivalent
outputs whenever all the inputs are bivalent. Thus the Kleenean recovers classical
logic for nonparadoxical sentences by adding appropriate instances of excluded
middle. This goes for Hartry Field’s approach (2008), and for many other nonclas-
sical logics.
Of course, it can be far from obvious which sentences are paradoxical.
Semantic paradoxes can be contingent on all sorts of circumstance (Kripke
1975). Even if we are working in a language for physics, with no overtly seman-
tic vocabulary, a reductionist about semantics may worry that the property
of Truth is expressed by some complex predicate or other of the language for
physics, so that confining our sentences to that language is no guarantee of
nonparadoxicality.
There is a more general concern. The retreat to K3 invalidates vast swathes of
ordinary mathematical reasoning, since mathematicians freely reason in ways
that depend on the law of excluded middle. Such restrictions on mathematics in
turn restrict its applications to natural science. The natural scientists might over-
come the restrictions by postulating instances of excluded middle as needed.10
But then their explanations invoke those auxiliary assumptions, which reduces
their explanatory value; elegant explanations get as much as possible out of as
little as possible. The point is not that the auxiliary assumptions exceed the clas-
sical logician’s commitments: they do not, because the classical logician is anyway
committed to the unrestricted law of excluded middle. Rather, the point is that
the auxiliary assumptions are ad hoc for the Kleenean in a way they are not for
the classical logician, who derives them all from the simple, elegant, general prin-
ciples of classical logic. The Kleenean can give them no such general explanation.
The best the Kleenean can do is derive them from a metalinguistic principle such
as that all instances of excluded middle in the language of physics are true; but it
would be bizarre to claim that ordinary physical explanations which happen to
need excluded middle actually involve such metalinguistic considerations. Thus
the Kleenean strategy pays a heavy abductive cost across a vast range of ordinary
342 Reflections on the Liar
13.5. Complications
The foregoing argument oversimplifies in various ways. Although they compli-
cate the overall picture, they change its overall outline little. Still, they deserve
mention.
Semantic Parad ox e s and Abduc tive Me thod ol og y 343
disquotation must be restricted anyway. But that still leaves disquotation with
respect to the theorist’s own context. The danger is that the semantic paradoxes
force the theorist of crashes into locating some of those crashes for the terms of
the crash theory itself in the very context of the theorizing, which looks like some
sort of defeat for the crash theory itself. Of course, once we are willing to theorize
about L in a metalanguage for L that contains a predicate ‘true’ for L not in L itself,
we can have a disquotational principle for that predicate over all sentences of L,
but such a potentially hierarchical approach falls short of the originally envisaged
form of disquotation. And presumably the term ‘crash’ will have to get the hier-
archical treatment too. But we can still hope for a simple, strong version of the
hierarchical approach that fits our evidence.
Many ways of handling the semantic paradoxes are consistent with classical
logic. There is no need to try to rank them here. But when they are ranked, the
usual abductive methodology will of course apply.
The methodological case made here for maintaining classical logic even in the
face of the semantic paradoxes is a very natural one. Why have such obvious con-
siderations been so widely neglected? Presumably, the explanation lies in the ten-
dency to discuss the paradoxes in isolation. A narrow focus has many advantages,
but from time to time it is worth looking at the bigger picture.15
Notes
1. Although Tarski (1936) is often cited as the origin of the model-theoretic conception of logi-
cal consequence, its assignments are not accompanied by a specification of a domain in the
model-theoretic sense. He seems to have envisaged the quantifiers as ranging unrestrict-
edly over the relevant type in a type-theoretic hierarchy. Thus first-order quantifiers range
over absolutely all individuals. See Williamson (2003 and 2013, pp. 221–261) for more
discussion.
2. We spare the reader the further constraints needed to avoid clashes of variables. For
instance, ∀x Rxx should not count as a substitution instance of ∀x Rxy.
3. Since L is an already interpreted language, the status of the three-valued semantics is that
of a theory about the semantics of L, comparable to that of a theory about the semantics of
English or some other natural language. It is no mere stipulation.
4. We require the sentences in Γ to be jointly well-confirmed, not just individually so, to avoid
problems with lottery and preface paradoxes.
5. For a general discussion and assessment of abductive methodology see Lipton (2004).
6. A seminal work here is of course Priest (1987).
7. The reason for the inverted commas is that such mappings typically fail to preserve meaning.
8. Proof-theoretic criteria of strength are also too coarse-grained for present purposes.
9. See Cozzo (1998, pp. 271–279) for an interesting development of Quine’s view in a similar
spirit to that of this chapter (although I regard the commitment to evidence-transcendent
truth he discusses as an interesting outcome rather than a cost of classical logic with known
auxiliary assumptions).
10. Halbach (2011, p. 293) points out that the use of K3 for treating the semantic paradoxes
involves an essential loss of the principle of transfinite induction up to the ordinal ε0, and a
corresponding loss of combinatorial principles of arithmetic, although that would presum-
ably not be crippling for most natural science.
11. See the appeal to complexity considerations in the critique of Field (2008) by Welch (2011).
Semantic Parad ox e s and Abduc tive Me thod ol og y 345
12. See Tarski (1935). Williamson (1998) postulates a less systematic sort of linguistic variation
provoked by the semantic paradoxes.
13. See Smiley (1993) for a sketch of such an approach.
14. Of course, as well as classical logic and disquotation, we also need some elementary syntax
or the equivalent to derive the paradoxes by diagonalization or the like. Above, the equation
λ = ‘¬True(λ)’ was treated as part of our evidence. Alternatively, one could treat it as up for
abductive assessment too. That would only reinforce the main moral of the chapter, that
classical logic is the very last thing to revise.
15. Previous versions of this material were presented at a conference in honor of Michael
Dummett at the Institute of History and Philosophy of Sciences and Techniques in Paris
and a seminar at Oxford University, and earlier in some of the 2013 Hägerstrom Lectures
at Uppsala University. It has benefited greatly from questions and comments on those
occasions.
References
Burge, Tyler. 1979. “Semantical Paradox.” Journal of Philosophy 76: 169–198. Reprinted with post-
script in Robert Martin, ed., Recent Essays on Truth and the Liar Paradox (Oxford: Clarendon
Press, 1984), 83–117.
Cozzo, Cesare. 1998. “Epistemic Truth and Excluded Middle.” Theoria 64: 243–282.
Dummett, Michael. 1981. Frege: Philosophy of Language. London: Duckworth.
Field, Hartry. 2008. Saving Truth from Paradox. Oxford: Oxford University Press.
Halbach, Volker. 2011. Axiomatic Theories of Truth. Cambridge: Cambridge University Press.
Kripke, Saul. 1975. “Outline of a Theory of Truth.” Journal of Philosophy 72: 690–716.
Lipton, Peter. 2004. Inference to the Best Explanation. 2nd ed. London: Routledge.
Parsons, Charles. 1974. “The Liar Paradox.” Journal of Philosophical Logic 3: 381–412. Reprinted
with postscript in Robert Martin, ed., Recent Essays on Truth and the Liar Paradox
(Oxford: Clarendon Press, 1984), 9–45.
Priest, Graham. 1987. In Contradiction: A Study of the Transconsistent. The Hague: Martinus
Nijhoff. 2nd ed. Oxford: Oxford University Press, 2006.
Putnam, Hilary. 1969. “Is Logic Empirical?” In R. Cohen and M. Wartofsky, eds., Boston Studies in
the Philosophy of Science, 5:216–241. Dordrecht: Reidel.
———. 2012. “The Curious Story of Quantum Logic.” In Putnam, Philosophy in an Age of
Science: Physics, Mathematics, and Skepticism, ed. M. De Caro and D. Macarthur, 162–177.
Cambridge, MA: Harvard University Press.
Quine, Willard van Orman. 1951. “Two Dogmas of Empiricism.” Philosophical Review 60: 20–43.
———. 1970. Philosophy of Logic. Englewood Cliffs, NJ: Prentice Hall.
Russell, Bertrand. 1919. Introduction to Mathematical Philosophy. London: George Allen & Unwin.
Smiley, Timothy. 1993. “Can Contradictions Be True?” Proceedings of the Aristotelian Society,
suppl. 67: 17–33.
Tarski, Alfred. 1935. “Der Wahrheitsbegriff in den formalisierten Sprachen.” Studia Philosophica
Commentarii Societatis Philosophicae Polonorum 1: 261–405. Translated by J. H. Woodger as
“The Concept of Truth in Formalized Languages” in Tarski, Logic, Semantics, Metamathematics,
ed. J. H. Woodger and J. Corcoran, 152–278, 2nd ed. (Indianapolis: Hackett, 1983).
———. 1936. “O pojciu wynikania logicznego.” Przegląd Filozoficzny 39: 58–68. Translated by
J. H. Woodger as “On the Concept of Logical Consequence” in Tarski, Logic, Semantics,
Metamathematics, ed. J. H. Woodger and J. Corcoran, 2nd ed. (Indianapolis: Hackett, 1983),
409–420.
———. 1986. “What Are Logical Notions?” History and Philosophy of Logic 7: 145–154.
Welch, Philip. 2011. “Truth, Logical Validity, and Determinateness: A Commentary on Field’s
Saving Truth from Paradox.” Review of Symbolic Logic 4: 348–359.
Williamson, Timothy. 1994. Vagueness. London: Routledge.
346 Reflections on the Liar
14.1. Introduction
14.1.1. Background: Three Families of Conceptions
Generally, traditional inflationary conceptions of truth commence with two the-
ses: namely, substantivism and monism. According to substantivism, truth is a
property with a substantial nature or underlying realizer, ρ. According to monism,
there is exactly one way of being true. For instance, some traditional conceptions
take the nature of truth to consist in an identity relation, others a kind of bijective
correspondence mapping, and others just some kind of weak morphism; some
posit coherence, and still others posit limit convergence or other such properties.
But while they differ over just which nature or essence truth consists in, tradi-
tional inflationary conceptions are united in claiming that there is just one such
realizer property ρ constitutive of it.
By contrast, traditional deflationary conceptions of truth reject substantivism.
While there are many frequently cited reasons why, a relevant one for present pur-
poses is that the plausibility of traditional inflationary conceptions seems to vary
by region or sector of discourse D: D1, …, Dn, and this variance seemingly indicates
a failure of generalization or limitation of their theoretical scope—a failure which
is then used to cast doubt on all traditional inflationary conceptions, tout court.
After all, if there is no underlying substantive property ρ that is shared by all the
truths and which is such that truths are true because they are ρ, then the tradi-
tional inflationist’s project of uncovering the sole substantive essence or nature of
truth appears unachievable. Traditional deflationary conceptions typically accept
monism, though—Quine’s and Field’s disquotationalism and Horwich’s minimal-
ism being familiar examples. That is, they concur that there is but one way of
being true, but deny that it is a property consisting in or realized by any more
basic property ρ.
Pluralist conceptions of truth invert the deflationary response: instead of
denying substantivism and accepting monism, they accept substantivism but
deny monism.1 On the intuitive assumption that language is indeed divisible into
347
348 Reflections on the Liar
nonidentical regions or sectors (see Christian, 1975), pluralists assert that truth
is a substantive property, but one that consists in different ρ1, …, ρn in different
D1, …, Dn.2 For example, truths in commercial advertising might consist in one
way of being true, while truths in gastronomy consist in a different property and
jurisprudential or arithmetical truths consist in yet others still. So, pluralists dif-
fer from traditional inflationists in denying that truth consists in a single prop-
erty across all domains of discourse, but differ from deflationists in contending
that a plurality of properties are constitutive of truth and are together necessary
and sufficient for explaining why the truths are true.
14.1.2. Agenda
The alethic paradoxes, such as the naïve Liar,
or the strengthened liar,
requires being assigned to the right discourse and necessitates having the right
alethic properties. Consequently, pluralists may be able just to decline to assert
the dialethicist’s claim that λ is both true and false, and can decline to assert the
analethicist’s claim that λ is neither true nor false; and they can decline to assert
that λ is either true or false, and that λ is true, and that λ is false. Moreover, to
so decline, pluralists need not positively assert that λ is meaningless. Rather, for
pluralists, λ may be meaningful but undecidable.
status, most are specifically linguistic; among those, the vast majority are also
conventional. A conventional linguistic unit is a symbolic structure ε with unit
status, which is prototypically produced by the human vocal apparatus, and which
is recognized and shared by a significant number of members of a given speech
community (Langacker 1987: 62). Rather than a state device with production
rules for strings, the grammar G of a language L is understood to be a countably
open-ended inventory of conventional linguistic units (GL: ⟦ε1⟧, ⟦ε2⟧, …), which
serve to structure a cognizer’s experience of the world for expressive purposes—
no more and no less.
A conventional linguistic unit may be simplex (e.g., symbols such as β or ⊥ or
free or bound morphemes like pre- or set or -ness) or complex. The formation of
complex expressions requires of cognizers a constructive mental effort, in which
at least two component expressions εi and εj are integrated and assembled into a
composite symbolic structure εk. Such an assembly—namely, a construction—may
be a specific expression (lexical) or a schema thereof (grammatical).
With these preliminaries, note, first, that speakers use expressions to commu-
nicate, much as laborers might use a hammer to pound; hammers themselves do
no pounding, much as expressions themselves do no communicating. Second, the
evocative powers of any linguistic expression ε—i.e., the power of its phonologi-
cal form (orthographic structure, etc.) to evoke or give voice to certain associated
semantic structures—are what determines its effective use. Propositions them-
selves cannot strictly ‘say’ or state anything; for the result of decoupling any ε
from either its constituent semantic (sem) or phonological (phon) structures is
either sensible silence or nonsensical noise, and, quite literally, no sensible silence
is capable of ‘saying’ anything. So propositions do not say or express; and what-
ever does say or express cannot be a proposition.6
Consequently, in explaining how the alethic paradoxes arise, we will not refor-
mulate our claims in terms of propositions in the standard way—that way lies
misconception. But more to the point, for us to speak of propositions as truth-
bearers at all, we must be able not only to identify the thing that does the bearing,
but to individuate any such bearer from any other; and so we must have a way,
upon inquiry, of determining not only whether a given sentence σ expresses a
proposition, but also, if so, which proposition it expresses. But answering any
such inquiries will necessitate the production of an answer, which can only just be
another expression ε; and this is precisely not at all what was asked for when ask-
ing for a proposition to be specified. So unless propositions get a free pass vis-a-
vis identity and individuation conditions, then because they can be specified only
in the course of symbolization, then propositions cannot be the proper bearer of
truth values. Hence, for present purposes, I will presume that the relevant class
of constructions with which to discuss λ are those sentences having judgments
in the alethic mode as their semantic poles (see Sher & Wright 2007). The subse-
quent dismissal of positivists’ meaningless strategy will therefore not be formu-
lated in terms of whether or not some proposition is expressed by λ.
P luralism and the Liar 351
The weak syntactical constraints are usually taken to be capture and release.
Specifically, and using corner quotes ‘⌜’ and ‘⌝’ as a metalinguistic operation that
exchanges a declarative sentence σ for its structural-descriptive name or Gödel
encoding, a predicate T is said to play the role unrestricted capture just when
(4) σ T ( σ )
holds for all relevant well-formed sentences σ. Conversely, T is said to play the role
unrestricted release just when
(5) T ( σ ) σ
holds for all σ. Then T is then said to be a truth predicate only if T is a capture/
release predicate, such that
(6) T ( σ ) σ
352 Reflections on the Liar
holds, again, for all such σ.7 Beyond transparency, embedding, and capture/
release, various other discourse-inspecific platitudes and constraints have
been proposed, and the idea is that they provide en masse an initial mea-
sure for pinning down the concepts necessary for establishing T as a truth
predicate.
The second contention is that different sectors of discourse, including those
for the good, the funny, and the physical, differ along various parameters. These
parameters typically concern things like factivity and mind-independence and
norms of belief-formation or endorsement; for some discourses are fully realist,
goes the thought, while others are belied by surrealist ‘as if’ operators and while
still others may be quasi-realist or error-theoretic. So certain additional platitudes
and constraints may be needed for particular sectors of discourse, which T must
also satisfy in order to continue to count as a truth predicate in them. In principle
discourses can be weighted and ordinally ranked ( D1 Di Dn ) by the number and
kinds of parameters associated with them.
The third contention— and here is the discourse pluralist’s enabling
thought—is that a truth predicate T is always a predicate for a particular dis-
course Di ⊂ , that is, TDi . Compliance with different platitudes subserving each
region or sector of discourse is then said to result in numerically distinct truth
predicates TD1, , TDn . Given further principles that map numerically distinct
truth predicates to different ways of being true, a corollary to the discourse plu-
ralist’s proposal may then be that different sectors of discourse are regulated by
different truth properties.
(7) There are many truth predicates TD , , TD , one of which applies to all
1 n
true σ.
So, one of the truth predicates among TD1, , TDn applies to all true sentences,
and so is a generic or universal truth predicate T applicable in any/all regions
or sectors of discourse . On the stronger version of this thesis, strong linguistic
pluralism (SLP),
all true σ.
P luralism and the Liar 353
So, not only is there more than one truth predicate, but no single predicate applies
across the board; and each TDi is correctly applied only to true sentences in Di.
Between WLP and SLP, there may be room to develop a view under which some
truth predicates apply to multiple domains without applying to all. Obviously,
both WLP and SLP are incompatible with the stronger counterpart of linguistic
monism (SLM),
(9) There is exactly one truth predicate T D, which applies to all true σ.
(10) Wet cats are funny. / This cat is wet. // This cat is funny.
(12) (∀σ ) ( T ( σ ) TD ( σ ) , ∨ , ∨ TD ( σ ))
1 n
354 Reflections on the Liar
and second, that, being unable to prevent its subsequent introduction into L, dis-
course pluralists are thereby committed to it.9 Consequently, pace Pedersen’s lin-
guistic liberalism thesis, the only viable conception of discourse pluralism turns
out to be WLP.
But now suppose that the expression in question is ¬ T D( σ ). Then, on the
assumption that L provides for PA, Cotnoir observes that the diagonalization
lemma can be used to demonstrate that
and upon release in (5), we get the liar equivalence λ ↔ ¬λ . Cotnoir therefore
concludes that discourse pluralists should avoid WLP and its posited universal
disjunctive truth predicate in (7); by implication, they also should reject the thesis
of linguistic liberalism that generates (12).
As Cotnoir also observes, though, SLP itself is not yet out of the woods either.
The strong linguistic pluralist who denies T D still has the discourse-relative capture/
release predicates posited in (8), i.e., TD1 ,,TDn . For SLP, each TDi need only sat-
isfy (6) for all σ in Di. Yet, given the liar sentence λ Di for truth predicate TD1 ,
where λ Di is the Gödel encoding of ¬TDi (x ), the diagonalization lemma can be
again used to construct λ Di : ¬TDi ( λ Di ) (Cotnoir 2013b: 341). Consequently,
restricting truth predicates to sectors of discourse does not allow advocates of
SLP to preclude
P luralism and the Liar 355
which again, upon release in (5), again gives us the generic form of the liar
λ ↔ ¬λ. So, as Cotnoir nicely demonstrates, discourse pluralists who endorse any
kind of linguistic pluralism about truth predication seem to incur paradox and
inconsistency.
Then, as before, diagonalization from the open expression ¬ T (x) allows a Gödel
encoding of a universal liar λ ω in any/all discourses such as to derive the paradox:
So Cotnoir demonstrates how infinitary disjunction for any/all TD ( σ) can yield,
i
the class whose members comprise things like standing professions (archaeol-
ogy, economics, journalism, etc.), technical trades or vocational disciplines and
subdisciplines (graphic web-design, gastroenterology, cosmetics, etc.), or subject
matters (political anarchism, comedy, fluid dynamics, etc.). As there are only so
many of these, our presumption should be in favor of being a large but count-
ably finite class—contrary to the first assumption. This is not the only way to
conceive of the pluralist’s construct discourse, of course (see Christian 1975;
Wyatt 2013). However, the extent to which discourse pluralists deviate from this
presumption is the extent to which they lose their grip on the motivating ratio-
nale for introducing pluralism in the first place, which was just the thought that
some discourses—ethics, jurisprudence, and aesthetics being exemplars—are
less than fully realist. Consequently, discourse pluralists can retain that moti-
vating rationale but not the first assumption; and if the postulation, and subse-
quent rejection, of infinitary disjunction requires just such an assumption, then
Cotnoir’s treatment of the Liar may be one that discourse pluralists must consider
parting with.
But we may also part with discourse pluralism; for pluralists about truth are
not required to set up as discourse pluralists in the way that Cotnoir’s results
and solution may require. (In later sections, I will rehearse determination plural-
ism as a viable alternative.) So when we consider the assumptions from which
his treatment derives, it becomes worth noting that the second assumption—the
endorsement of many truth predicates T1, …, Tn—cannot be a general prescrip-
tion. For instance, the assumption will fail to hold for any alethic pluralist about
truth or conceptual pluralist about truth who is committed to SLM.
For such theorists, true is lexically polysemous: a single expression with mul-
tiple related senses. But one need not be a pluralist of any kind to make this claim;
as many other philosophers (Brentano, Stout, Russell, Wittgenstein, Næss, Tarski,
Austin, etc.) have observed, true is lexically polysemous de facto.10 And as many
linguists have observed, lexical polysemy is the norm—especially for important
terms in frequent usage.
Abductively, the lexical polysemy of true is exactly what we would expect were
SLM to be correct. But this would seem to be inconsistent with the third assump-
tion, linguistic pluralism about truth predication, and consistent with the empiri-
cal facts about the grammar G of each language L, which appear to have but one
default truth predicate (er sand in Danish, is waar in Dutch, is true in English, etc.).
In that sense, pluralists should endorse Pedersen’s claim that SLP was never a
real contender (2006: 107), although the reason why is not that linguistic liberal-
ism is an ‘independently plausible and correct’ thesis which subsequently forces
a disjunctive T -construction. On the one hand, denying this third assumption
raises the question of whether Cotnoir’s solution will require some adjustments;
for if there is exactly one default truth predicate in GL, then the concern over
the disjunctive introduction of a universal or generic disjunctive T D from the
idle wheels T1, …, Tn, and the need to go on to reject infinitary disjunction, are
P luralism and the Liar 357
undercut. On the other hand, denying this third assumption does not entail that
Cotnoir’s commitment to strong pluralism must go by the boards. For again, while
our presumption should be in favor of strong monism about truth predication
(SLM), pluralists can be alethic pluralists about truth or conceptual pluralists
about truth, or both. Of course, it remains to be seen whether pluralists about
truth or truth who are strong monists about truth predication incur other grave
problems, and in particular, whether those problems impact pluralists’ ability to
diagnose and treat the alethic paradoxes.
2010: 97). It is unclear what a philosophical truth predicate could be; for phi-
losophy is not a language with its own grammar, and so not a language with a
grammar that includes a philosophical truth predicate as one of its conventional
linguistic units.13 Perhaps what was intended is the plausible claim that philoso-
phy is its own discourse Dϕ, such that the ‘kind of truth predicate in which phi-
losophers are interested’ is just the truth predicate of their home discourse, Tϕ.
But by discourse pluralists’ own relativization of truth predicates to discourses,
TDφ cannot then apply to sentences from other discourses—an awkward result,
especially where discourse pluralism is supposed to deliver a conception of truth
predication spanning multiple sectors of discourse. Confining the scope of their
claims to the discourse of philosophy would also be to accept that there is exactly
one (relevant) truth predicate—namely, TDφ ; but this would be to accept (a variant
of) SLM, which is presumably not what Pedersen or Cotnoir have in mind.14
It might be that that truth predicates philosophers are interested in, pace
Pedersen, are Tarskian truth predicates. As is well known, the Tarskian hierarchi-
cal approach relativizes truth predicates to formally regimented languages and
orders them into a countably infinite set {L0 < L1 < L2 < }. Then, in order to
avoid the incursion of a λ-sentence, this approach ramifies lower-order schemes
into successively higher-order (meta)languages by uniform substitution, such
that no order contains its own truth predicate:
(19a) Lm :
(b) L2 : is true - in-L1 is true - in-L2 ↔ ( is true - in-L1 ↔ )
(c) L1 : is true - in-L1 ↔
(d) L0 :
the idea being that these conditionals underwrite explanatory claims. So in some
Di, it is because σ corresponds to certain facts that σ is true, but not conversely
(Edwards 2013: 118). So the nature of each discourse is associated with an alethic
property the possession of which is sufficient to determine the possession of its
overlying truth property. Then, supposing that each discourse is associated with
exactly one property, the determination pluralist will form up biconditionals of
the form,
Presumably, it will not be that each discourse is associated with exactly one unique
property, since the exactly one property associated with a discourse might also
be associated with another: while psychopharmacology might be associated with
being homomorphic to a fact, so too might oncology or ontology.
One point to notice is that, rather than commencing with the discourse plu-
ralist’s linguistic focus on truth predication, the simple determination pluralist
instead focuses alethically on truth properties. And presumably, she can pro-
ceed with the naturalized linguistic monism previously urged in the interlude in
§14.3: there is exactly one default predicate is true which designates a property
that is differentially determined relative to sector of discourse. But the overarch-
ing thrust, however, remains the same: as Putnam put it, ‘[o]n the one hand, to
regard an assertion or a belief or a thought as true or false is to regard it as being
right or wrong; on the other hand, just what sort of rightness or wrongness is in
question varies enormously with the sort of discourse’ (1994: 515).
This brief discussion of determination pluralism is intended merely as an
exemplification of how pluralists can both avoid discourse pluralism and endorse
SLM, while also refocusing from truth predication to truth properties. Alethic dis-
junctivism might be equally instructive. It remains to be seen, however, whether
determination pluralists or disjunctivists incur inconsistency and paradox, and,
if so, whether they are able to diagnose or treat it. In that regard, perhaps the
P luralism and the Liar 361
true sentence
Max
Min
true sentence
time
Max tr σ
Min
time time
which differ both in the extent of their meaningfulness and, more obviously, their
degree of grammatical acceptability (marked by ‘*’).
Competent speakers of English have little trouble navigating the symbolic
assemblies in (27), despite their being self-referential. This is to be expected: the
phenomenon of linguistic self-reference is generally unproblematic. And while the
relationship between the general phenomenon of self-reference and diagnoses of
the paradoxes remains an open question, these remarks inevitably carry over to
the naïve truth-teller,
That the constructions in (29) are conceptualizable can also be seen using con-
structions such as
and the same point can be alternatively made using constructions such as
(31a) exical items, such as the ones exemplified here, are inherently
L
meaningful.
(b) Each of the lexical items comprising this sentence is inherently
meaningful; so too are their symbolic assemblies.
(c) ‘The sentence comprised of these very words is meaningful’ is
meaningful.
So, each component ⟦ε⟧ in (31) is a meaningful lexical unit in GL, and the overall
configuration of these component semantic structures is inherited as the seman-
tic pole of each composite assembly. Here too, our semantic theory predicts that
each of the constructions in (31) differ in conceptualization; yet, to so differ, the
variants in (31) must have semantic structures with which to do.
Competent speakers naturally comprehend the self- referential con-
structions in (31) no less than those in (29) and (30). As Armour-Garb &
Woodbridge (2013: 848) observe, linguistic comprehension of constructions
requires knowledge of their meanings; but having knowledge of their mean-
ings requires that such sentences have meanings. So unless the appearance of
competent speakers comprehending self-referential constructions turns out
to be mere appearance, then the truth-tellers in (29) are meaningful, ultima
facie.15 But that the truth-tellers in (29) are manifestly meaningful would seem
to bode poorly for the positivist’s meaningless strategy. Further, a theory of
semantics and symbolization, if to address itself to the paradoxes at all, should
be uniformly applied to both the Liar and truth-teller. Since (29) does not
364 Reflections on the Liar
substantively differ from the Liars in (1) and (2), the Liars in (1) and (2)
should each be treated as meaningful. And so it would appear that the
meaningfulness of these sentences implies that pluralists should not be
positivists—at least, not positivists who advance the meaningless strategy as a
treatment of λ.
It is worth pausing to address two serious worries, in case advocates of the
meaningless strategy find this to be too quick. The first worry is that this rea-
soning is (nonviciously) circular: it assumes that the truth-teller is meaningful
in order to show that the Liar is meaningful. In response, our recognition of
the meaningfulness of the truth-teller has the status of defeasible empirical
observation—not an assumption—which is both intuitively and experimen-
tally confirmable, and which is accounted for using the semantic theory at
hand. So suppose that λ is meaningful. Then the second worry is that taking λ
to be both meaningful and undecidable then results in an unstable combina-
tion.16 The idea is that we must antecedently accept all instances of the follow-
ing schema,
Taking the Liar to be one such σ, the meaningfulness of liar sentences implies
that they have truth-conditions, but having truth-conditions implies that
those sentences are decidable. So if λ is undecidable, then it must be meaning-
less; but this was just what the positivist’s original strategy was, and so the
motivation to treat λ as undecidable undercuts itself. In response, the schema
that generates the worry issues from controversial theses and semantic theo-
ries that we are not required to accede to or that have already been rejected
here, such as thesis that sentences are meaningful only if they express proposi-
tions. (Recall that conceptualization is a function of both the content evoked
and the construal imposed; but since the same truth-conditions can be differ-
entially construed and differences in construal yield differences in conceptual-
ization, linguistic meaning cannot be exhaustively reduced to truth-conditions
(Langacker 1987).)17
14.4.2. Positivism, Redux
We might assert that the predicate is meaningless is not meaningless, and our
assertion would be prima facie true. At the end of §1.2, it was suggested that the
meaningless strategy is not something that pluralists should be attracted to, for
much the same reason. Yet, while the foregoing considerations justify denying
that λ is meaningless, the argument of this chapter does not require it.18 So we
may be agnostic about the ultimate truth of the matter.
However, the general positivist approach still has something to offer; for it
might instead by articulated as the thesis that λ is not so much both true and
P luralism and the Liar 365
false, or neither true nor false, as it is just undecidable: we decline it any kind of
truth-valuation.
In the case of the naïve Liar in (1), the grammatical subject this sentence is
a grounded nominal; it does not function logically as a name and so does not
name a sentence or otherwise self-refer to a truth-evaluable sentence. So it is
not apt for truth-predication if truth-predication requires that an expression
be Gödel-encoded. We can predicate properties and relations of it, such as the
predicate has two words (i.e., as in ‘this sentence’ has two words); but truth and
falsity are not among them. This is part of the reason why revenge truth-tellers
and Liars, e.g.,
are effective; for the grammatical subject of (34), This sentence is false, does func-
tion logically as a name σ, such that T ( σ ) is well-formed.
For the alethic pluralist, though, the issue needs to be the assignment of the
sentence to region of discourse. So, upon release, which discourse does the named
sentence belong to? One response is that the subjects this sentence is true and this
sentence is false in (33) and (34), and (1a) and (29a), cross-cut all regions or sectors
of discourse (as the pluralist understands them). That is, there is no (uncontrived,
nongerrymandered) discourse D ⊂ that aggregates only the truths, or only the
falsehoods, or both, or some such; and so neither naïve, nor strengthened, nor
revenge liars will reside in any region or sector of discourse. For the determina-
tion pluralist, for example, this means that sentences which do not reside in any D
thereby cannot satisfy the conditionals in (20)–(24), and thus cannot satisfy the
subsequent biconditionals, either. To be sure, it may very well be that λ possesses
some ρ. But it will never be that, in, e.g., psychopharmacological (geological, arith-
metical, moral, method acting, etc.) discourse, if λ possesses ρ, then λ possesses
truth; for if λ is homeless then there will be no D such that, in D, one finds λ, much
less λ possessing ρ. The suggestion is that, while not necessarily meaningless, λ
should be characterized as undecidable.
Decidability requires that there exists some exact procedure
A ∈{ A1, A2, , An } for a well-defined input/output function f: I → O, in which
inputs i ∈ I are mapped to outputs f (i ) ∈O, and which determines in poly-
nomial time whether or not a given string belongs to a given discourse. Of
course, it is well known that not every function f is computable (Turing 1936),
and many f that are computable in principle are not computable in practice
(Garey & Johnson 1979).
Results from computability theory suggest that the problem of deciding how
to bin certain sentences may have Turing-undecidable equivalents of the halting
problem. In part, this is because of difficulties inherent in the task of providing
identity and individuation conditions on sectors of discourse—or as, some have
366 Reflections on the Liar
(35a) For any given sentence σ and any given discourse D, is σ ∈D?
(b) For any given σ and a set of discourses , does there exist a D ⊂
such that σ ∈D?
(c) For any given σ and a set of discourses , can we can return some
D such that σ ∈D?
and then provide the requisite logical and conceptual analyses of the different
senses of the term negation (choice, exclusion, etc.). However, recall that negation
enjoys a definition in terms of falsity,
(b)
¬ σ : = σ is false
the circularity of which makes this option less desirable for pluralists. Following
Parsons (1983: 245; see also Scharp 2010), pluralists might opt for antiexten-
sional definitions of natural-language falsity predicates ( σ is false:=σ is in the
antiextension of truth), although it is unclear how to prevent this solution from
just reducing to the previous view of falsity as the negation of truth and the nega-
tion of truth in terms of falsity. Subsequently, pluralists might instead consider
restricting their claims to atomic sentences, so as to claim that falsity is just the
truth of negation,
and then to handle the truth of negation as with any/all other truth-functional
compounds (see also Edwards 2008). A further option—one perhaps more conso-
nant with the shift from linguistic to alethic pluralism advocated here—would be
to conceive of falsity as the absence of the alethically potent property ρ in some
specific Di. Of course, given the (Russellian) requisite of parity between truth and
falsity, nothing in principle prevents alethic pluralists from developing a more
metaphysically extravagant view of a constitutive property, the possession of
which suffices to determine whether a given discursive construction possesses the
property of being false. However, if the positivistic suggestion explored here—i.e.,
treating liar constructions as being Turing-undecidable for determination plural-
ists—holds any interest, pluralists might instead suppose that falsity results from
the mismatch of alethically potent properties and regions of discourse. So in dis-
course Di, let σ possess truth just in case σ is ρi per biconditionals of the form in
(25). Then σ will be false just in case either σ belongs to discourse Di but does not
possess ρi, or possesses ρi but belongs to Dj. In that case, σ is false in part because
it belongs to a domain; but λ is not false, given that there is no discourse D to
which it belongs, and so no Di in which it belongs but fails to possess ρi.
Positing a (tractably computable) assignment function should come with a
commitment to an exact algorithm for computing it. The thought that λ might
not be so much meaningless as undecidable depends on the thought that we can
think of the designatum of assignment algorithmically—i.e., as an effective deci-
sion procedure—when we think of a sentence’s being true as requiring that the
sentence be assigned to some sector or region of discourse. If there is no effec-
tive decision procedure that runs in nonexponential/polynomial time and halts
with output λ in Di, then λ cannot possess the relevant domain-specific property
ρ sufficient for rendering λ true. Perhaps the function computed can be made to
368 Reflections on the Liar
be fixed-parameter tractable and the number of strings being input are countably
infinite; something like a greedy algorithm that uses locally optimal choices to
approximate a globally optimal solution might then make tolerably clear enough
why liar sentences never get assigned to a discourse. Nonetheless, some further
tinkering with the suggestion is called for.19 For intuitively, there will be many
of the wrong kinds of sentences that seem to go the way of the Liar. Standard
examples include mixed compounds such as foie gras is edible and foie gras is mor-
ally repugnant, which presumably belong to none of the determination plural-
ist’s sectors of discourse even if its atomic sentences each belong to one (see also
Edwards 2008; Cotnoir 2009). Must the response then be the same, i.e., that foie
gras is edible and foie gras is morally repugnant is Turing-undecidable? That would
be counterintuitive. Consequently, it seems that σ in the determination pluralist’s
biconditionals must be restricted to atomic sentences, with the truth-values of
compounds being decided by applying the usual truth-functional methods. But
even some atomic sentences, such as some truths are long-winded, may also present
challenges, given that, presumably, they reside in no proper sector of discourse.
Must they, too, be undecidable like the Liar?
Other familiar strengthened liar and revenge problems may arise. For analethic
pluralists, if σ is neither true nor false, then σ is not true. As σ’s not being true just
is constitutive of the claim that σ is not either true or false, then σ must be true
and the contradiction is regained. But now, in place of the analethic pluralist’s
claim that σ is neither true nor false, do we not have the same problem for posi-
tivists? If σ is undecidable, then σ is not true. As σ’s not being true just is implied
by the claim that σ is undecidable, then σ is both undecidable and not true. And
so σ is not true, which seems to contradict the claim that σ is undecidable. In
response, positivists or pluralists may try to show that σ is undecidable and not
true is itself undecidable (or undecidable and false, and … off we go.) A related
worry is a revenge variant:
Our positivist may treat (37a) by claiming that strengthened liar sentences like
(37a) are undecidable, and so presumably must accept (37b). But then—so the
thought goes—to accept (37b) is to accept that (37b) is true, from which it follows
that there had better be some discourse D in which (37b) belongs and subsequently
possesses the alethically potent property ρ that is sufficient to determine its truth.
This is a thorny problem. One response is to distinguish between acceptance
and assertion, or some analogue of the familiar distinction between rejection and
denial.20 A more natural and effective response, however, would be to just rise to
the challenge. To that end, let Dk be the discourse for theoretical computer sci-
ence, and observe that some of the sentences in Dk will be about phenomena such
as NP-hardness, the halting problem, oracles, heuristic search, von Neumann
P luralism and the Liar 369
14.4. Conclusion
Using standard diagonalization techniques to recursively construct liar sen-
tences, Cotnoir (2013b) raises a number of problems for the discourse pluralist.
Pluralists should consider rejecting discourse pluralism in favor of a more natu-
ralized approach to truth predication in real languages, and should concentrate
on property and concept versions of pluralism instead. Determination pluralism
is one such version. However, the contention that pluralists about truth should
be monists about truth predication should be supported with an account of plu-
ral truth—as Cotnoir aims to show—that avoids the Liar. Such pluralists may be
interested in a variant on positivism about the Liar in which liar sentences are not
at all meaningless, but are undecidable.21 Being essentially ‘homeless’, determina-
tion pluralists are unable to determine which sector of discourse liar sentences
belong to, which warrants the declination to distribute truth-values over them.
Specifically, pluralists can decline to assert the dialethicist’s claim that λ is both
true and false, and can decline to assert the analethicist’s claim that λ is neither
true nor false; and they can decline to assert that λ is either true or false, and that
λ is true, and that λ is false.
Notes
1. Omitted is discussion of the fourth family of conceptions—deflationary pluralism—that
commences from the denial of both substantive and monism. While this taxon is nonempty,
there are too few representative conceptions that are sufficiently well worked out; excep-
tions include Beall (2013) and possibly Kölbel (2013).
2. As nearly all discussants in this literature have acknowledged, the task of properly war-
ranting and formalizing this assumption proves recalcitrant. Given Lynch’s (2009: 77–80)
requirement that no truth-bearer is a member of more than one sector of discourse, one
might appeal to work in combinatoric mathematics to formalize D as a Sperner family in
which no sector members in any other subclass, i.e., Di D j for all pairs Di , D j ∈ D where
Di ≠ Dj. On disambiguation of the term domain, see also Wyatt (2013).
3. Following Chihara (1979), we may divide these four approaches according to whether they
explain what goes wrong in ordinary language use to give rise to the alethic paradoxes (diag-
nosis), or whether they devise strategies for speakers to avoid serious difficulties (treatment).
370 Reflections on the Liar
(Disquietingly, most of the proffered explanations and strategies neglect real languages and
their speakers, which is perhaps most salient in the Tarskian hierarchical gambit.)
4. For extended discussion of semantic and grammatical preliminaries, as well as a purview
of different cognitive scientists’ theories of the conceptual substrate of sem, see, e.g.,
Langacker (1987).
5. So the lexical expressions familiar to fluent speakers are therefore symbolic structures of
the form [[sem]\[phon]] (simplified as ⟦ε⟧); they function as an integrated whole and can
be employed automatically without having to attend to either their component parts or
organization, and their usage is effortless because the requisite cognitive routines involved
in pairing meaning with form are psychosemantically entrenched.
6. By implication, the relationship between a sentence σ and its sense or proposition is consti-
tution, not expression.
7. Ideally, a fuller summary of discourse pluralism would add grammatical constraints for
quasi-, indirect, mixed, and direct enquotation and disquotation, given that the phenom-
enon of same-saying does not reduce to capture or release.
8. What pluralists have not conceded, however, is the sharp end of Tappolet’s point—her fur-
ther contention that, once pluralists retreat to WLP, the motivation to maintain it gives way
to SLM. The reason why, they argue, is that there is an order of priority making T D explana-
torily dependent on TD1, , TDi . For fuller explanations, see Lynch (2009), Pedersen &
Wright (2013), and Edwards (2011, 2013).
9. Pedersen (2006: 112) contends that the generic or universal disjunctive truth predicate need
not designate a generic truth property, however. This contention does not seem implausible;
for not every predicate designates a property. Presumably no property is designated by a
nullary predicate, for instance, not even a property of 0 adicity. Also, some properties are
not signified by any predicates. For example, it seems that no property is signified by the
predicate is not a property; for suppose that there were some property, ρ, signified by is not a
property. Then ρ is a property, and is not a property is true of ρ. But then the property ρ is not
a property. So ρ is a property and not a property, and so it seems there is no property, ρ, so
signified.
10. This claim was also implied Sainsbury, who observed that ‘[i]t seems descriptively quite
incorrect to suppose that we switch truth predicates in this way’ (1996: 901). Unfortunately,
he missed the opportunity to clarify or elaborate on the observation.
11. Reframing the overarching endeavor for theories of truth in terms of explanation opens
the door to new difficulties (Wright 2016). But if explanation is the overarching theoretical
endeavor, then it is at least crucial to recognize that true’s being polysemous is not so much
an ugly consequence of the pluralist’s explanantia as it is simply part of the explanandum—
i.e., a datum about languages like English that any truth theorist must come to grips with
(Wright 2012: 103 n. 15).
12. While I am inclined to agree that the concept truth is, as Hinzen puts it, ‘an evolved
aspect of a species-specific mode of cognition linked to the evolution of the faculty of
language and one of a number of foundational abstractions that characterize our kind
of mind’ (2013: 219; see also Sher & Wright 2007), neither truth nor its symboliza-
tion found its way into careful and detailed studies of linguistic universals (see, e.g.,
Wierzbicka 1992).
13. Even if Pedersen’s philosophical truth predicate turns out to refer to a constructional schema
instead of specific lexical expression, it will still be a constructional schema in the grammati-
cal inventory of a particular language G .
Li
14. Suppose that the relevant concept of a truth predicate is indeed TDφ , in whatever natural lan-
guage they may use for their theorizing. It will still be the case that the sentences to which
it applies will also be sentences in whatever natural language is being used; and these will
also take a nonrelevant truth predicate, TDχ (e.g., ‘sneeuw is wit’ is waar). But how would the
two work together? Is there some further miscegenated predicate TDφ χ that supersedes them
to overdetermine the truth of the sentence? And to what sector of discourse would any of
these belong, anyway? Ultimately, SLM is simpler and more attractive.
P luralism and the Liar 371
15. It is imperative to be careful about when a self-referential sentence belongs to the class
of paradoxical sentences. For instance, this sentence is not a sentence, and so not a self-
referential sentence; and so ‘This sentence’ is not a sentence is a grammatical truth, but not a
truth-teller of the relevant (paradoxical) sort.
16. Thanks to Brad Armour-Garb for prompting me to address these two worries.
17. Additionally, it may be that some instances of the schema are unacceptable. For example,
consider a meaningful sentence σ that is extremely accurate but not maximally so. In the
presence of (32), the LHS of the biconditional will fall in the extension {False} because
its consequent mispredicates maximal accuracy of meaningful σ. But as Millgram (2009,
chap. 3) nicely demonstrates, it would be misguided to claim that extremely accurate σ on
the RHS must also fall in the extension {False}, and so this instance of the biconditional has
two sides with different extensions/truth-values. Of course, requiring bivalence, or allowing
the concept truth to be analyzed only as correctness, are simple ad hoc escapes; but it is
unclear what could possibly motivate pluralists to play along.
18. To assert that λ is not meaningless is not to say that λ is not semantically defective in some
other way. Hence, in their 2013 paper, Armour-Garb & Woodbridge attempt to show that,
even if we concede this unimpeachable argument from comprehension, λ is nevertheless
meaningless in the sense that it fails to specify what they call M-conditions, which Armour-
Garb & Woodbridge claim to be a component of meaning. Of course, the conceptualization
or content of λ may indeed be defective; but to possess defective content just is to possess
(graded) content of a certain kind. Armour-Garb & Woodbridge’s point, though, seems to be
that, while some sentences like snow is white apparently directly specify their own M-condi-
tions, and truth-predications to such sentences apparently only indirectly specify their own
M-conditions, paradoxical sentences like the strengthened liar in (2) simply fail to specify
any M-conditions (2013: 846). While more needs to be said about what M-conditions are
supposed to be, there may be some resemblance between their direct specification and the
pluralist’s assignment of constructions to discourses.
19. Cotnoir (personal communication) rightly notes that there are potential downsides to
thinking of the determination pluralists’ assignment function as algorithmic in general,
and he is certainly correct that further work must be done to show that the approach can
avoid over-generating. Certainly, it would be awkward for logical or arithmetic truths and
falsehoods not to get assigned to discourses that they intuitively should, and I do not know
whether or to what extent the suggestion generalizes. So the suggestion of an undecidabil-
ity treatment remains exploratory—it may be natural or fruitful to think of tractable input-
output mappings as computable in this way—and is part of a broader class that includes
treatments based on indeterminacy and (un-)groundedness, which determination pluralists
may want to consider as well. Of course, if the problem is that invoking computation to
handle domain assignment introduces new problems associated with domain individuation,
then this might also or instead be more of a problem for pluralists in general, who have yet
to come to a clear consensus on how to formalize the construct domain in ways that map
back on to the idea of discourse.
20. While we might endorse the platitude in (3a) that to assert σ is to present σ as being true,
the same platitude may not hold of acceptance. In some contexts, one might assert σ with-
out accepting that σ (e.g., actors in a play, people aiming to provoke, guardians facilitating
child-rearing, etc.) In other contexts, one might accept σ without asserting that σ (e.g., a
student who neither denies nor rejects global skepticism, or eliminative materialism, etc.
might accept these positions as warranted but not warrantedly assertable).
21. This chapter had to be hastily written in 2014. In the intervening years, I have benefitted
from conversations with Aaron Cotnoir, Doug Edwards, Joe Ulatowski, Nathan Kellen, Brad
Armour-Garb, and several other participants at the Pluralisms workshops at the Universities
of Connecticut and Yonsei. It has also recently come to my attention—too recently, unfor-
tunately, for me to engage here—that Stephen Barker has also suggested an undecidability
treatment of the Liar, with a response from Mark Jago. Although the notion of undecidabil-
ity in Barker’s paper is different, I encourage readers to seek out their exchange.
372 Reflections on the Liar
References
Armour-Garb, Bradley & James Woodbridge. (2013). Semantic Defectiveness and the Liar.
Philosophical Studies 164: 845–863.
———. (2013). Deflated Truth Pluralism. In Nikolaj Pedersen & Cory Wright, eds., Truth and
Pluralism: Current Debates (323–338). New York: Oxford University Press.
Chihara, Charles. (1979). The Semantic Paradoxes: A Diagnostic Investigation. Philosophical
Review 88: 590–618.
Christian, William A. (1975). Domains of Truth. American Philosophical Quarterly 12: 61–68.
Cotnoir, Aaron. (2009). Generic Truth and Mixed Conjunctions: Some Alternatives. Analysis
69: 473–479.
———. (2013a). Validity for Strong Pluralists. Philosophy and Phenomenological Research
83: 563–579.
———. (2013b). Pluralism and Paradox. In Nikolaj Pedersen & Cory Wright, eds., Truth and
Pluralism: Current Debates (339–350). New York: Oxford University Press.
de Saussure, Ferdinand. (1966/2006). Writings in General Linguistics, trans. Carol Sanders,
Matthew Pires, & Peter Figueroa and ed. Simon Bouquet & Rudolf Engler. Oxford: Oxford
University Press.
Edwards, Douglas. (2008). How to Solve the Problem of Mixed Conjunctions. Analysis
68: 143–149.
———. (2011). Simplifying Alethic Pluralism. Southern Journal of Philosophy 49: 28–48.
———. (2012). Pluralist Theories of Truth. Internet Encyclopedia of Philosophy. http://www.iep.
utm.edu/plur-tru/.
———. (2013). Truth, Winning, and Simple Determination Pluralism. In Nikolaj Pedersen &
Cory Wright, eds., Truth and Pluralism: Current Debates (113–122). New York: Oxford
University Press.
Garey Michael & David S. Johnson. (1979). Computers and Intractability: A Guide to the Theory of
NP-Completeness. San Francisco: W. H. Freeman.
Hinzen, Wolfram. (2003). Truth’s Fabric. Mind and Language 18: 194–219.
———. (2013). Naturalizing Pluralism about Truth. In Nikolaj Pedersen & Cory Wright, eds.,
Truth and Pluralism: Current Debates (213–237). New York: Oxford University Press.
Kölbel, Max. (2013). Should We Be Pluralists about Truth? In Nikolaj Pedersen & Cory Wright,
eds., Truth and Pluralism: Current Debates (278–297). New York: Oxford University Press.
Langacker, Ronald. (1987). Foundations of Cognitive Grammar, Vol. 1: Theoretical Prerequisites. Palo
Alto: Stanford University Press.
Lynch, Michael P. (2000). Alethic Pluralism and the Functionalist Theory of Truth. Acta Analytica
15: 195–214.
———. (2009). Truth as One and Many. Oxford: Oxford University Press.
Millgram, Elijah. (2009). Hard Truths. New York: Wiley-Blackwell.
Parsons, Charles. (1983). Mathematics in Philosophy: Selected Essays. Ithaca: Cornell
University Press.
Pedersen, Nikolaj. (2006). What Can the Problem of Mixed Inferences Teach us about Alethic
Pluralism? Monist 89: 103–117.
———. (2010). Stabilizing Alethic Pluralism. Philosophical Quarterly 60: 92–108.
Pedersen, Nikolaj & Cory Wright. (2013). Pluralist Theories of Truth. Stanford Encyclopedia of
Philosophy. http://plato.stanford.edu/entries/truth-pluralist/.
———. (2013). Pluralism about Truth as Alethic Disjunctivism. In Nikolaj Pedersen & Cory
Wright, eds., Truth and Pluralism: Current Debates (87–112). New York: Oxford University
Press.
Putnam, Hilary. (1994). Sense, Nonsense, and the Senses: An Inquiry into the Powers of the
Human Mind. Journal of Philosophy 91: 445–515.
Read, Stephen. (1995). Thinking about Logic: An Introduction to the Philosophy of Logic.
Oxford: Oxford University Press.
P luralism and the Liar 373
Sainsbury, Mark. (1996). Crispin Wright: Truth and Objectivity. Philosophy and Phenomenological
Research 56: 899–904.
Scharp, Kevin. (2010). Falsity. In Cory Wright & Nikolaj Pedersen, eds., New Waves in Truth (126–
136). New York: Palgrave Macmillan.
Sher, Gila & Cory Wright. (2007). Truth as a Normative Modality of Cognitive Acts. In Dirk
Greimann & Geo Siegwart, eds., Truth and Speech Acts: Studies in the Philosophy of Language
(280–306). New York: Routledge.
Tappolet, Christine. (1997). Mixed Inferences: A Problem for Pluralism about Truth Predicates.
Analysis 57: 209–210.
———. (2000). Truth Pluralism and Many-Valued Logic: A Reply to Beall. Philosophical Quarterly
50: 382–385.
Turing, Alan. (1936). On Computable Numbers, with an Application to the Entscheidungsproblem.
Proceedings of the London Mathematical Society, Series 2 42: 230–265.
Wierzbicka, Anna. (1992). Semantics, Culture, and Cognition: Universal Human Concepts in Culture-
Specific Configurations. Oxford: Oxford University Press.
Wright, Cory. (2010). Truth, Ramsification, and the Pluralist’s Revenge. Australasian Journal of
Philosophy 88: 265–283.
———. (2012). Is Pluralism Inherently Unstable? Philosophical Studies 159: 89–105.
———. (2016). Truth, Explanation, Minimalism. Synthese.
Wright, Crispin. (1992). Truth and Objectivity. Cambridge: Harvard University Press.
———. (1995). Truth in Ethics. Ratio 8: 209–226.
Wyatt, Jeremy. (2013). Domains, Plural Truth, and Mixed Atomic Propositions. Philosophical
Studies 166: 255–236.
INDEX
375
376 Ind e x
deflationism, 348 external
deflationist theory of truth, 17, 272, 277, negation operator, 209
284, 348
descriptive, 291 operator, 207
project, 4, 6–7, 20n15 facts, 216, 238
theory, 276 empirical, 216, 228
designation, 349 nonsemantic, 241
determination pluralism, 359–361, 369 semantic, 241
diagnostic project, 4 failure rules, 314–316
diagonalization, 355, 369 false
Diagonal Lemma, 103 LIAR experiment answers, 47, 49–50, 54, 56, 64
dialectically liar sentences as, 14, 26, 160–161, 361
analytic, 259 theory, 179
constitutive, 259 falsehood, 15, 84–87, 104, 197, 242, 365
dialetheists, 42, 207, 233, 257, 342, 348 falsity, 193–195, 216–217, 233–234, 366–367
discourse pluralism, 351–352, 359–360, 369, 370n7 Feferman, Solomon, 13, 202
disjunction, 19, 235–237, 240, 326–327, 339 Field, Hartry, 341
disjunctivists, 360 filtering, 163
dispositions, 174 first-order logic (FOL), 98
disquotation, 83–84, 339–341 fixed-point construction, 226, 234–236
-first viewpoint-arguments, 85 FOL. See first-order logic
disquotational, 152 folk, 40
paradigm, 152 coherence, 41–42, 64
principle, 329, 340, 344 incoherence, 40, 42, 64
schema, 18, 114 reason, 325
disquotationalism, 347 formal
DLCS. See dynamic: lexicon constraint on adequacy, 281
soundness language, 302
D-strategy. See D(avidson)-strategy theories, 215
duck-rabbit, 75–76, 81 formalism, 152, 215
Dummett, Michael, 202, 331 formation rules, 35
dynamic Foster’s problem, 14, 150
conception, 137 foundational
lexicon, 116–123 -holistic methodology, 284, 286–287
liar paradox, and, 134–138 semantics, 213–216, 251n6
logic, and, 129–132 foundationalist methodology, 286
semantics, and, 123–129 Frege, Gottlob, 262, 271
lexicon constraint on soundness (DLCS), Fregean (second-order) logic, 153–154, 269, 271
131–133
Gaifman’s theory, 313
Eklund, Matti, 258 generality, 214, 243, 246, 269
Empedocles, 166 universal, 325–331
Epimenides, 39, 204–205 geology, 359, 366
epistemicism, 138 geometrical theory, 281
error theories, 181, 352 geometry, 184
Etchemendy, John, 41, 77 Euclidean, 265
Euclidean geometry, 265 Kock-L awvere synthetic differential, 272
exceptionless rules, 312 non-Euclidean, 121
existence principles, 265 space, of, 281
experimental Gestalt shifts in the Liar, 13, 71–113
design and methodology, 47 assessment sensitivity, 82–87
philosophy, 12, 43 conceptual multistability of sentences that
explanation, 286 ascribe a semantic value, 74–81
causal, 334 basic case of conceptual multistability, 77
expressively adequate, 7–8 multistability of sentences that self-ascribe a
expressive restrictions, 318–320 semantic value, 77–81
extensionally equivalent theory, 154 perceptual multistability, 74–77
378 Ind e x
Gestalt shifts in the Liar (Contd.) transcendence principle, 17, 283, 287–293,
formalizing the truth operator, 89–96 299–300, 303, 304n19
KT4M (see KT4M as the logic of semantic immanence-transendence complementarity, 290,
modalities) 298–299, 302–303
liar property, solution of the Liar, and the immanent
strengthened Liar, 96–103 sentence, 292, 296
liar property, Gestalt shifts, KT4M, and the thoughts, 288–292
solution, 96–101 incoherence, 307–311, 319
strengthened Liar and revenge, 101–103 incoherent, 5, 13
self-reference and context sensitivity in Liar incompatibility, 79
and truth-teller sentences, 81–82 Incompleteness Theorem, 103
truth predicate and truth operator, 73–74 inconsistency, 4, 9–10, 18, 23, 34–35, 262, 360
undecidable semantic status of liar sentences inconsistent, 5–6, 16, 43, 115, 179
and liar agnosticism, 87–89 concepts, 257–260
Glanzberg, Michael, 202–203 expression, 22–23, 42
Gödel, Kurt, 261–262, 290 term or phrase (C), 264
Gödelized Liar, 73 inconsistent concepts, revising, 15–16, 257–280
Gödel numbering, 292 cases, 260
Grelling expression, 11 (naive) infinitesimal, 262–263
Grelling’s paradox, 11, 23–25, 31, 33, 36n1, set/membership, 261–262
36n3 truth, 261
groundedness, 299 conditions for replacement, 264–272
grounded sentences, 300 overview, 257
groundless sets rule, 315–317 truth, 272–277
indeterminacy, 8, 35, 38n16, 205, 208
hermeneutic treatment, 5–8 inferential roles, 274
Herzberger, Hans G., 13, 72 infinitesimal, 262–263, 265, 268–269, 272
Herzberger’s revision semantics, 77 calculus, 268
heterological, 24–25, 34, 36nn2–3 inflationism, 348
hierarchical, 283 inflationists, 272–273, 348
approach, 344 intensions, 174
hierarchy-incoherence dilemma, 307–311 internal negation operator, 206–207
implications for logic, 319–320 intuitionistic
objections and replies, 317–319 logical systems, 15
human languages, 160, 163, 166–167 propositional logic (IC), 337
generative procedures, as, 166
I-L anguages, as, 171–183 judgments, 45, 181–182, 185, 216, 223
microlanguages, 14 speaker, 216, 218–219, 224
syntax, 153 suspension of, 72
Tarski-style theory of truth, and, 141, 149, justification, 286
151–153, 155–157, 162, 165
truth theories for, 142–148, 169 Kantian intuition, 269
Humean principle, 154 K(aplan)-strategy, 169–170, 176
Humpty Dumpty theory of word meaning, 121 Katz-Fodor-Postal proposal, 182
Kleene
identity, 81, 326 logics, 333, 341
predicate, 326–327, 339 negation, 308
relation, 347 schemes, 232–235, 308, 317, 323n22
I-L anguages and T-Sentences, 14–15, 141–190 knowledge, 40, 273, 303n5
conclusions on, 183–186 Kock-L awvere synthetic differential
human languages as I-L anguages, 171–183 geometry, 272
truth theories Kripke, Saul, 72, 81, 134–135, 295, 298–303
human languages, for, 142–148 Kripkean fixed-point construction, 226, 234–236,
meaning theories, as, 148–157 239, 253n33
immanence principle, 17, 283, 287–293, 299, Kripkean theories, 209, 212, 239
303, 304n8 Kripke-Feferman theory, 13, 72
normativity principle, 17, 283, 287–288, Kripke’s approach, 308, 310, 312, 319, 323n24
291–293 Kripke semantics, 72, 92, 95
Ind e x 379
paradoxes, 12, 18, 32, 185, 193, 215, 223 quantification, 152, 158, 340
assumptions, 2 sentential, 197, 202
inconsistency and, 360 unrestricted, 327
logic of, 212 quantifiers, 275, 308, 339
negation, 319 existential, 200, 326–327
philosophical, 3 generalized, 233, 237, 253n33
procedural accounts of, 246–250 matrix, 236
reflexivity, and, 222 sentential, 202–203
resolution of, 138 universal, 326–327, 329
significance of, 40 quantum
strengthened, 318 logic, 338
two-line, 247 mechanics, 338
paradoxicality, 4–5, 81–82, 101 quasi-realist, 352
Parsons, Charles, 201, 202 quietism, 8
Ind e x 381