Reasons, Justification, and Defeat. Jessica Brown and Mona Simion (Ed.)

OUP CORRECTED AUTOPAGE PROOFS – FINAL, 22/2/2021, SPi
Reasons, Justification, and Defeat

Copyright © 2021. Oxford University Press USA - OSO. All rights reserved.
Reasons, Justification, and Defeat, edited by Jessica Brown, and Mona Simion, Oxford University Press USA - OSO, 2021.
Reasons, Justification,
and Defeat
Edited by
JESSICA BROWN AND MONA SIMION
1
3
Great Clarendon Street, Oxford, OX2 6DP,
United Kingdom
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 in certain other countries
© the several contributors 2021
The moral rights of the authors have been asserted
First Edition published in 2021
Impression: 1
All rights reserved. No part of this publication may be reproduced, stored in
a retrieval system, or transmitted, in any form or by any means, without the
prior permission in writing of Oxford University Press, or as expressly permitted
by law, by licence or under terms agreed with the appropriate reprographics
rights organization. Enquiries concerning reproduction outside the scope of the
above should be sent to the Rights Department, Oxford University Press, at the
address above
You must not circulate this work in any other form
and you must impose this same condition on any acquirer
Published in the United States of America by Oxford University Press
198 Madison Avenue, New York, NY 10016, United States of America
British Library Cataloguing in Publication Data
Data available
Library of Congress Control Number: 2021931162
ISBN 978–0–19–884720–5
Printed and bound in the UK by

TJ Books Limited
Links to third party websites are provided by Oxford in good faith and
for information only. Oxford disclaims any responsibility for the materials
contained in any third party website referenced in this work.
Contents
List of Contributors vii
1. Introduction 1
Jessica Brown and Mona Simion
2. The Normativity of Knowledge and the Scope and
Sources of Defeat 18
Sanford C. Goldberg
3. The Structure of Defeat: Pollock’s Evidentialism, Lackey’s
Framework, and Prospects for Reliabilism 39
Peter J. Graham and Jack C. Lyons
4. Losing Knowledge by Thinking about Thinking 69
Jennifer Nagel
5. Dispositional Evaluations and Defeat 93
Maria Lasonen-Aarnio
6. Suspension, Higher-Order Evidence, and Defeat 116
Errol Lord and Kurt Sylvan
7. Reasons for Reliabilism 146

Bob Beddor
8. Knowledge, Action, and Defeasibility 177
Carlotta Pavese
9. Undercutting Defeat: When it Happens and Some
Implications for Epistemology 201
Matthew McGrath
10. Defeaters as Indicators of Ignorance 223
Julien Dutant and Clayton Littlejohn
11. Competing Reasons 247
Justin Snedegar
12. Perceptual Reasons and Defeat 269
Mark Schroeder
Index 285
List of Contributors
Bob Beddor, National University of Singapore

Jessica Brown, University of St Andrews
Julien Dutant, King’s College London
Sanford C. Goldberg, Northwestern University
Peter Graham, University of California Riverside
Maria Lasonen-Aarnio, University of Helsinki
Clayton Littlejohn, King’s College London
Errol Lord, University of Pennsylvania
Jack Lyons, University of Glasgow
Matthew McGrath, Rutgers University
Jennifer Nagel, University of Toronto
Carlotta Pavese, Cornell University
Mark Schroeder, University of Southern California
Mona Simion, University of Glasgow
Justin Snedegar, University of St Andrews
Kurt Sylvan, University of Southampton
1
Introduction
Jessica Brown and Mona Simion
1.1. Introduction
The notion of defeat has been central to epistemology, practical reasoning, and
ethics. Within epistemology, it is standardly assumed that a subject who
knows that p, or justifiably believes that p can lose this knowledge or justified
belief by acquiring a so-called ‘defeater’, whether evidence that not-p, evidence
that the process which produced her belief is unreliable, or evidence that she
has likely misevaluated her evidence. Within ethics and practical reasoning, it
is widely accepted that a subject may initially have a reason to do something,
although this reason is later defeated by her acquisition of further information.
The notion of defeat has been central to a wide range of different philo-
sophical debates, including, but not limited to:
(1) The nature of justification and knowledge: since knowledge and

justification are taken by many to be defeasible, the extent to which

one account or another of the nature of knowledge/justification is able
to account for/accommodate defeat constitutes an important ground
for assessing its theoretical credentials (see e.g. Sudduth 2018).
(2) Internalism versus externalism: several epistemologists worry that
epistemic externalism has a hard time accommodating psychological
defeat; at the same time, conversely, if justification supervenes on
mental states alone, as per internalism, it seems mysterious that it
could be defeated by normative defeaters lying outside of the cognizer’s
ken (see e.g. Pappas 2017).
(3) Epistemic norms and reasons: for accommodating the phenomenon of
defeat, debates on epistemic norms and epistemic reasons owe us, at a
minimum, an account of epistemic normative overriding, as well as an
account of reasons against belief (see e.g. Simion 2020)
(4) Evolutionary debunking arguments: the nature of defeat has
repercussions for the debate about the metaethical implications of
093/oso/
2     
evolutionary explanations of morality. It is argued that some, if not all,

human moral beliefs are the product of evolutionary forces: we were
selected for having useful rather than true moral beliefs. Learning about
the evolutionary origin of our moral beliefs undercuts their prima facie
justification, or so the challenge goes (see e.g. Silva 2016).
(5) Evidence and higher-order evidence: since evidence is widely taken to
be defeasible, a plausible account of evidence should come with a
corresponding plausible account of defeat. For instance, one important
desideratum for any such account is that it explains the defeating power
of higher-order evidence, namely of evidence that one’s first-order
beliefs are the output of a flawed process (see e.g. Brown 2018).
(6) Closure and transmission: one popular solution to alleged failures of
closure principles for knowledge and transmission principles for war-
rant is known as ‘the defeat solution’: roughly, the thought goes, closure
and transmission hold prima facie, and the intuitions of failure are to
be explained in terms of psychological defeat. This solution, of course,
hangs on the assumption that there is such a thing to begin with: i.e.
that psychological defeat is a genuine epistemic category (see e.g. Pryor
2004).
(7) Disagreement: one way to characterize the debate between conciliatory
and steadfast views of disagreement is as centred around the question:
can the testimony of one’s peer carry defeating power? Steadfast-ism
answers ‘no’, conciliationism answers ‘yes’. The correct account of the
nature of defeat can help settle the issue (see e.g. Frances and Matheson
2019).
(8) Reductionism versus anti-reductionism about testimony: say that a
suspect for murder S tells you that she did not do it. According to both
of the main views in the epistemology of testimony, you are not
justified to believe S. According to reductionism, that’s because you
always need positive, non-testimonial reasons to believe what you are
being told. In contrast, according to anti-reductionism, you are prima
facie justified to believe S, but your justification is defeated. The correct
nature of the nature of defeat will likely go a long way in the direction of
settling the issue (see e.g. Green 2008).
It is useful to categorize contributions to the defeat literature as falling under

one of the following two broad categories: (i) the nature and extent of defeat
and (ii) kinds of defeaters. In what follows, we will first briefly run through the
state of the art thus categorized (Sections 1.2 and 1.3). Last, we give an
 3
overview of the volume’s chapters and explain how they build on the state of
the art (Section 1.4).
1.2. The Nature and Extent of Defeat
1.2.1 Defeaters as Reasons
The first and now considered the classic view on the nature of defeat in
epistemology is due to Pollock (1986). According to this view, D is a defeater
of E’s support for p for S if and only if (i) E is a reason to believe p for S, and
(ii) E&D is not a reason to believe p for S (henceforth Pollock’s view).
The account has a lot going for it; first, it promises to cut across normative
domains, in virtue of being framed in terms of reasons; after all, epistemolo-
gists hardly enjoy exclusivity on reasons. With Pollock’s view in play, it is easy
to see how we could generalize it to cover different targets (e.g. actions) and
types (e.g. moral, prudential) of normativity. Second, Pollock’s view makes
good on the intuitive thought that defeaters are actualizers of the possibility of
a positive normative status to be overridden or undercut; what the view says,
in a nutshell, is that defeaters are the kind of things that render a permissible
belief impermissible.
More recently, though, Pollock’s view has come under heavy attack. First, as
Chandler (2013) points out, Pollock’s view is problematic in that it cannot
accommodate what can be called justifying defeaters. As the name suggests, a

justifying defeater JD of E’s support for p for S does double duty: it prevents E
from being a reason to believe p for S but, at the same time, it gives S a new
reason to believe p. So, JD is a defeater for S, but it does not fulfil clause (ii) of
Pollock’s view, as E&JD still is a reason for S to believe p.
Second, the view has been found wanting on prior plausibility in virtue of
the very fact that it’s stated in terms of reasons. Pollock’s view fits snugly with
evidentialism about justification: the epistemic status of a belief is determined
by the reasons for and against believing. As such, on Pollock’s view, there is no
defeat without reasons. With the increased popularity of externalist, process-
based accounts of justification, however, conceiving of defeat in terms
of reasons is not very helpful. One reason for this is that, on standard
forms of reliabilism (process reliabilism, virtue reliabilism, proper functional-
ism), reasons are neither necessary nor sufficient for justification. If
reasons don’t have any justificatory power, however, it is mysterious why
they might have defeating powers. Furthermore, notable defenders of
4     
reliabilism deny that reasons are substantive epistemic normative categories to

being with (e.g. Lyons 2009; Kornblith 2015).
1.2.2 Defeaters as Reliable Processes
Reliabilist theories of justification have been extremely popular in the last

three decades and come in a variety of forms, the gist of the view is that a belief
is justified if and only if formed via a (normally) reliable procedure, or ability.
Reliabilism is a theory of prima facie justification. As such, in line with
normative theories in general, it needs a theory of defeat in order to hold
water. The standard reliabilist account of defeat comes from Alvin Goldman:
The alternative reliable process account (ARP): S’s belief is defeated if there
are reliable (or conditionally reliable) belief-forming processes available to S
such that, if S had used those processes in addition to the process actually used,
S wouldn’t have held the belief in question (Goldman 1979).
One can see how ARP is an elegant reliabilist translation of the Pollockian
thought that defeat is the kind of normative entity that, when taken in conjunction
with the extant epistemic support for the relevant belief, fails to render it justified.
Bob Beddor (2015) is the locus classicus for criticism of ARP; if Beddor is
right, ARP is both too weak and too strong. Against ARP’s sufficiency direc-
tion, Beddor offers the following case:
Thinking about Unger: Harry sees a tree in front of him at t. Consequently, he

comes to believe the proposition TREE: 〈There is a tree in front of me〉 at t. Now,
Harry happens to be very good at forming beliefs about what Peter Unger’s 1975
time-slice would advise one to believe in any situation. Call this cognitive process
his Unger Predictor [ . . . ]. What’s more, [ . . . ] whenever it occurs to Harry that
Unger would advise him (Harry) to suspend judgement about p, this causes Harry
to [ . . . ] suspend judgement about p. So, if Harry had used his Unger Predictor, he
would have come to [ . . . ] suspend judgement regarding TREE.
What this cases shows is that ARP is too weak: contra ARP, for my belief
that p to be defeated, it is not enough that I would change my mind about p in
a counterfactual world due to employing some reliable process. What this
objection identifies is that ARP is normatively too weak: just because I would
change my mind in world W, it does not follow that I should change my mind
in world W: defeat is a normative notion.
 5
Here’s Beddor’s case against ARP’s necessity direction:
Job opening: Masha tells Clarence that her department will have a job
opening in the fall. Clarence believes Masha; assuming that Masha is usually
reliable, Clarence’s belief counts as prima facie justified. Sometime later,
Clarence speaks with the head of Masha’s department, Victor, who informs
him that the job search was cancelled due to budget constraints. Now suppose
that Clarence harbours a deep-seated hatred of Victor that causes him to
disbelieve everything that Victor says, what’s more, no amount of rational
reflection would rid Clarence of this inveterate distrust. Consequently, he
continues to believe that there will be a job opening in the fall.
This case shows that ARP is also too strong: just because, in all counterfac-
tual worlds, I would irrationally and stubbornly hold on to my belief, it does
not follow that I should do so. Once again, ARP is not normative enough to do
the job it is supposed to do.
1.2.3 Defeat Scepticism
We have seen that, within epistemology, it is standardly assumed that a subject

who knows that p, or justifiably believes that p, can lose this knowledge or justified
belief by acquiring a defeater. Within ethics and practical reasoning, it is widely
accepted that a subject may initially have a reason to do something although this
reason is later defeated by their acquisition of further information.
However, the traditional conception of defeat has recently come under attack.
Some have argued that the notion of defeat is problematically motivated
(Lasonen-Aarnio 2010, 2012, 2014; Hawthorne and Srinivasan 2013).
According to a second strand of attack, the notion of revisionary defeat is
inconsistent with important desiderata on an account of knowledge or justifica-
tion such as naturalism, externalism, evidentialism, or a rule-based account
(Greco 2010; Lasonen-Aarnio 2010, 2012, forthcoming, Beddor 2015; Baker-
Hytch and Benton 2015; Weatherson, 2019). A further worry concerns how the
defeat of a proposition’s status as evidence is compatible with a Bayesian frame-
work (e.g. Weisberg 2009; Pryor 2013; Greco 2017). Furthermore, those who
adopt a steadfast view on peer disagreement can be understood as denying the
defeating effect of evidence that a peer disagrees with one (e.g. Kelly 2005).
In the light of these attacks, some have suggested that we should diagnose
the intuition that there is something wrong with someone who continues to
6     
believe that p in the face of defeating evidence not by the suggestion that the
relevant belief loses its status as justified or as constituting knowledge but
instead in some other way, e.g. by some variety of error theory.
According to Williamson (2009: 315), for instance, when we intuit that
one’s knowledge gets defeated, we are, in fact, confusing between whether one
knows and whether it’s probable on one’s evidence that one knows. Another
popular suggestion is that continuing to believe that p in such circumstances is
to exhibit an epistemic disposition likely to lead to trouble in the long run
(Lasonen-Aarnio 2010; Hawthorne and Srinivasan 2013). According to this
view, in cases of alleged defeat, our intuitions fail to distinguish between
impermissibility and mere blameworthiness generated by irresponsible
behaviour.
Against defeat scepticism, in recent work, one of us (Brown 2018: Ch. 5) has
argued that it involves an unnoticed and unacceptably high cost: once we
explicitly distinguish between contributory (i.e. synchronic) and revisionary
(i.e. diachronic) defeat, it is hard to argue against revisionary defeat. That is
because it is deeply implausible to deny the contributory notion of defeat—i.e.
roughly, evidence against p that is part of the initial body of evidence. For, to
do so is effectively to deny that when evidence affects justification, it is one’s
overall evidence that matters and not merely a part of it. But, once we accept
the contributory notion of defeat, it is hard to deny the revisionary given the
plausible epistemic symmetry between cases of contributory and revisionary
defeat.
1.2.4 Higher-Order Defeat
Higher-order evidence is evidence about what evidence one possesses or what

conclusions one’s evidence supports. One interesting question involving
higher-order evidence is: how should our beliefs respond to our beliefs about
our beliefs when we do have higher-order evidence (Feldman 2005;
Christensen 2010; Kelly 2005, 2010). A related issue in recent epistemology
is explaining the defeating power of higher-order evidence, namely of evidence
that one’s first-order beliefs are the output of a flawed process (see for instance
Christensen 2007, 2010; Kelly 2010; Lasonen-Aarnio 2014). Suppose, for
illustration, that you are a pilot who calculates whether you have enough
fuel to reach the closest airport. Upon calculating, you reach the conclusion
that you have more than enough fuel to get to an airport fifty miles further
than the one in your initial plan. Suppose you then glance at the altimeter to
 7
see that you’re at 10,500 feet and remember that hypoxia is a risk at altitudes of
10,000 feet and higher. You now have evidence that you might have hypoxia,
and thus you have evidence that you might have miscalculated. Are you now
justified in believing that you can get to the more distant airport? Are you
justified in believing that your evidence supports that claim? (Christensen
2010).
Answering ‘yes’ to both questions is intuitively extremely problematic: it
licenses an extremely form of dogmatism. Answering ‘no’ to the first question
and ‘yes’ to the second doesn’t seem very promising either: even if you can’t
actually bring yourself to believe F, being justified in believing your evidence
supports F prima facie justifies you in believing F. Answering ‘no’ to both
questions is the traditional way to go for champions of defeat, and probably
the most widely spread view in the literature. For one, this is the answer that
suggests itself on both the main views on the nature of defeat (the reasons-
based and the alternative reliable process-based accounts).
One highly debated view in the literature combines defeat scepticism with
answering ‘yes’ to the first section and ‘no’ to the second question. On this
view, then, your second-order evidence fails to defeat your first-order justifi-
cation: you are still justified to hold your first-order belief that you’ll reach the
second airport, but you are no longer justified in believing that your evidence
supports your claim. This view has become known as the ‘level-splitting view’
(Lasonen-Aarnio 2010, 2014; Wedgwood 2011; Coates 2012). Thus, on this
‘level-splitting view’, the pilot is justified to believe that they’re going to reach
the second airport and that their evidence does not support this conclusion
(henceforth Rational Akrasia).
Feldman (2005: 110–11) argues that it is impossible for this belief to be both
true and reasonable since the second conjunct undermines the reasonableness
of the first conjunct (see also e.g. Elga 2007; Christensen 2007; 2010; Horovitz
2014; Titelbaum 2015). Furthermore, the level-splitting view has also been
argued to licence problematic theoretical and practical reasoning (e.g. Brown
2018; Horovitz 2014). For instance, if Rational Akrasia obtains, if the pilot
believes that she has enough fuel to reach the second airport, then it seems that
she can exploit the claim that she has enough fuel to reach the second airport
in her practical reasoning. For example, she might reason that since she has
enough fuel to reach the second airport, she should not bother to land on the
first airport. Nonetheless, while ignoring the first airport on the grounds that
she has enough fuel to reach the second one, she ought to also admit that it’s
unlikely on her evidence that she’ll reach the second airport! But acting in this
way seems entirely unreasonable.
8     
1.3. Kinds of Defeaters
1.3.1 By Mechanism
How does defeat work? Pollock’s distinction between rebutting (also known as
overriding) and undercutting (also known as undermining) remains the
classic answer to this question.
According to Pollock (1986), there are two ways in which defeaters act:
roughly, they either speak against the content of one’s belief, or against the
credentials of its formation procedure. To see how this works, say that my
friend Mary tells me that TRAIN: The train leaves at eight sharp, and thereby
I come to justifiably believe that TRAIN. Now, say that my other good friend,
Alice, sharply disagrees: according to Alice, the train leaves at 8.30. The
thought is that, after receiving Alice’s testimony, I am no longer (fully)
justified in believing TRAIN, in virtue of acquiring evidence against TRAIN
being the case. In this case, Alice’s testimony constitutes itself in a rebutting
defeater for my believing TRAIN.
In contrast, undercutting defeaters speak against the credentials of the
source of my belief. Consider for instance, an alternative scenario in which
Alice tells me that Mary is a compulsive liar about trains’ schedules. After
receiving her testimony, I am no longer (fully) justified in believing TRAIN
based on Mary’s assertion.
On Pollock’s view, a rebutting defeater, then, for a belief that p of S is a
reason for S for either believing non-p or for believing some proposition q,
where q is incompatible with p. In that, on Pollock’s view, rebutting defeaters
work by overriding the reasons in favour of believing that p. In contrast,
undercutting defeaters don’t speak against p, but rather against the credentials
of the belief formation procedure employed in forming the belief that p. An
undercutting defeater for a belief that p of S is a reason for S attacks the
connection between S’s ground for believing p and p.
One worry that arises about Pollock’s distinction is whether it is plausible
that it is an exclusive one, as the account seems to suggest. Pryor (2013), for
instance, gives an example meant to illustrate that a rebutting defeater can, in
virtue of its rebutting normative power, also constitute itself in an undercut-
ting defeater: to see this, suppose S’s justification for believing p from evidence
E is rebutted, as S acquires new evidence E* that more strongly supports q,
which is incompatible with p. When this happens, Pryor argues, S is left with
the question of why p is false even if E is true. Pryor’s suggestion is that the best
answer available to S is that, in the present context, E is not a reliable indicator
 9
of the truth of P. Hence E* also acts as an undercutting defeater, since it

supplies S with some evidence against the reliability of E in supporting p.
In contrast, Sturgeon (2014) argues that undercutting and rebutting defea-
ters exemplify two essentially different kinds of defeat. According to Sturgeon,
for instance, in contrast to rebutting defeat, undercutting defeat is not fully
independent: it acts in tandem with the subject’s belief that the source under
attack is the one that is actually responsible for generating the belief in
question.
1.3.2 By Normative Status
An important question addressed in the literature on defeat concerns the

normative status of defeaters. The normative status question concerns the
issue of whether defeaters need to be justified themselves. Lackey’s (1999)
distinction between doxastic (aka merely psychological) and normative defea-
ters is central to this debate. In Lackey’s view, doxastic defeaters are merely
psychological defeaters: they are beliefs of S, not necessarily true or justified,
that speak in favour of the claim that S’s belief that p is false or is based on an
unreliable source. Normative defeaters are (good) reasons for entertaining
doxastic defeaters. According to Lackey, then, the answer to the normative
status question is ‘no’: defeaters need not be justified themselves.
Many epistemologists agree with Lackey, and it is fair to say that Lackey’s
distinction—which itself entails that defeaters need not have positive epistemic
status—has become a classic in the epistemological literature of the last years.
According to Alvin Plantinga (2000: 364–65), for instance, irrational and
unwarranted beliefs can defeat beliefs that have impeccable epistemic status.
Suppose I believe that I’m made of flesh, blood, and bone. I then come to
believe, due to some cognitive disorder, that my head is made of blown glass.
According to Plantinga, given that I come to hold this second belief I now have
a defeater for the prior belief, even if the defeater was formed by way of
cognitive malfunction. According to Plantinga, this case is one where I have
a ‘rationality defeater’: given that I acquire the second belief, it is no longer
rational for me to hold the first one.
Contra Plantinga, Alston (2002) argues that only beliefs with positive
epistemic status can defeat beliefs that have positive epistemic status, and,
furthermore a belief D can defeat belief A only if D has greater warrant
than A. One of us (Simion 2020) has also argued that the existence of
merely psychological defeat implies the existence of merely psychological
10     
justification, i.e. it implies justification internalism. If this is so, externalists

about justification should be especially worried about the category of mere
psychological defeat.
1.3.3 By Psychological Status
Another important question for taxonomizing defeat concerns whether defea-

ters need to enjoy psychological status or not, or, more clearly put, whether
only mental states (e.g. beliefs) can do defeating work, or, to the contrary, facts
outside our skull can also have defeating influence over the normative status of
our beliefs.
It is fair to say that it is widely acknowledged that, in the case of knowledge
defeaters, the answer to the psychological status question is ‘no’: facts out there
in the world can act upon our knowledgeable beliefs and thereby lower their
epistemic status. The classic case to illustrate this point is Gilbert Harman’s
(1973: 143–44) assassination case. Suppose that a political leader has been
assassinated. A reporter who is a witness to the assassination dictates details of
the event to his news agency so that the story may be included in the day’s final
edition of the paper. Jill picks up the paper and reads the story and believes
that the political leader has been assassinated. However, before Jill picks up the
newspaper and reads the story, loyalists to the political leader declare on
nationwide television that the bullet actually struck and killed someone in
the political leader’s entourage. Jill reads the true story in the paper but misses
the false report on television. Many people believe with Harman that Jill
doesn’t know that the political leader has been assassinated. There are, how-
ever, several different views on the market as to what explains this datum.
According to Swinburne (2001) and Pollock (1986), the explanation is social:
the fact that Jill does not have knowledge is the consequence of there being a
true proposition (suggestive of a defect in justification) that is widely believed
in Jill’s society. Others think that the explanation lies with the easy availability
of the relevant defeating information (Sudduth 2018).
We have seen that epistemologies that incorporate doxastic defeaters typ-
ically take them to defeat justification (Alston 1989: 238–9; Bergmann 2006:
155–6) or some species of rationality (Plantinga 2000: 357–66). When it comes
to non-doxastic defeaters, however, things look slightly different: while non-
doxastic defeaters are widely taken to have an impact on knowledge, whether
mere facts out there in the world can defeat justification is a more controver-
sial issue. The epistemic internalist will have to answer ‘no’: after all, by
 11
internalism’s lights, justification supervenes exclusively on facts internal to the

cognizer’s mind. Epistemic externalists, however, can afford to have mixed
views about the issue. Sandy Goldberg (2016), for instance, argues that
evidence one does not have, but that one should have had, can defeat one’s
justification. He thinks this comes about in virtue of social facts, i.e. in virtue of
the reasonable expectations of others grounded in norms internal to a social
practice. According to Lackey (2008), too, mere normative (non-doxastic)
defeat can defeat justification in cases where the believer has negligently
ignored important counter evidence, in virtue of epistemic responsibility
constituting a crucial component of epistemic justification.
1.4. Summary of the Volume
This volume brings together recent work to re-examine the very notion of
defeat, and its place in epistemology, and in normativity theory at large.
In Chapter 2, The Normativity of Knowledge and the Scope and Sources
of Defeat, Sanford C. Goldberg investigates the nature and scope of defeat
with an aim to argue in support of the category of normative defeat, as well
as in favour of there being a social dimension to knowledge. To this effect,
the chapter appeals to a prior grasp of the normativity of knowledge itself—
its role in entitling a subject to confidence and in authorizing others to
believe on the strength of one’s epistemic standing—to shed light on the
nature and scope of defeat. Goldberg’s strategy is to focus on cases in which

an otherwise epistemically well-positioned subject fails to enjoy these nor-
mative standings, and to argue that the best explanation is the presence of
normative defeaters.
Chapter 3, The Structure of Defeat: Pollock’s Evidentialism, Lackey’s
Framework, and Prospects for Reliabilism, by Peter Graham and Jack Lyons,
investigates the structure of defeat. It has two main aims; the first is mostly
critical: it argues that several classical distinctions that the literature on defeat
endorses are problematic. In particular, according to Graham and Lyons, the
traditional categorization of defeaters as doxastic and normative is mistaken:
first, unjustified beliefs can’t defeat, therefore there are no such things as
merely doxastic defeaters; second, the reason why ignored evidence can defeat
is different from the rationale traditionally taken to support the existence of
normative defeat. The second aim of the chapter is to develop a novel,
reliabilism-friendly view of the nature of defeat. On this view, having a
defeater for a belief that p is a matter of either having warrant to believe
12     
not-p, or else having warrant to believe that their warrant for believing that p
are inadequate (where warrant is understood in non-evidentialist terms).
Chapter 4, Losing Knowledge by Thinking about Thinking, by Jennifer
Nagel, puts forth a novel defense of infallibilism in epistemology against the
threat coming from the phenomenon of defeat. Defeat cases are often taken to
show that even the most securely based judegment can be rationally under-
mined by misleading evidence. Jennifer Nagel argues that defeat cases really
involve not an exposure of weakness in the basis of a judgement, but a shift in
that basis. For example, when threatening doubts are raised about whether
conditions are favourable for perception, one shifts from a basis of unreflective
perceptual judgement to a basis of conscious inference. In these cases, the basis
of one’s knowledge is lost, rather than rationally undermined.
Chapter 5, Dispositional Evaluations and Defeat, by Maria Lasonen-Aarnio,
argues that what explains the intuition of impermissibility in putative cases of
defeat resides not in the impermissibility of the target beliefs themselves, but
rather in the believer’s criticizability for manifesting bad epistemic disposi-
tions to believe. Subjects who retain their beliefs in the face of higher-order
evidence that those very beliefs are outputs of flawed cognitive processes are at
least very often criticizable. This, however, on Lasonen-Aarnio’s account, is
not because such higher-order evidence defeats various epistemic statuses such
as justification and knowledge. Instead, she argues that they manifest disposi-
tions that are bad relative to a range of candidate epistemic successes such as
true belief and knowledge. In particular, being disposed to only give up belief
in response to higher-order evidence when that evidence is not misleading

would require subjects to have dispositions that discriminate between cases in
which their original cognitive processes is fine, and cases in which they merely
seemed to be fine. But, according to Lasonen-Aarnio, such dispositions are not
normally humanly feasible.
In Chapter 6, Suspension, Higher-Order Evidence, and Defeat, Lord and
Sylvan focus on the epistemic effect of higher-order evidence. The chapter
makes two main claims: a negative one and a positive one. Their negative
claim is that extreme views about the issue—claiming either that higher-order
evidence has trumping effects, or that it has none whatsoever—are mistaken.
The positive view in turn has two parts. The first part defends the idea that
higher-order evidence provides direct reasons for suspending judgment that
typically leave evidential support relations on the first order intact: instead of
destroying these relations, the reasons for suspension defeat or compete with
the epistemic reasons for belief generated by these relations. Secondly, the
Lord and Sylvan framework purports to explain how this defeat is possible by
 13
showing how these distinctive reasons for suspension of judgment flow from
the constitution of suspension of judgment.
In Chapter 7, Reasons for Reliabilism, Bob Beddor motivates and develops
a synthesis between two leading approaches to justification which are
typically developed in isolation from each other; the first one comes from
the reliabilist tradition, which maintains that a belief is justified provided that
it is reliably formed, while the second one comes from the ‘Reasons First’
tradition, which claims that a belief is justified provided that it is based on
reasons that support it. On the view proposed by Beddor, justification is
understood in terms of an agent’s reasons for belief, which are in turn analysed
along reliabilist lines: an agent’s reasons for belief are the states that serve as
inputs to their reliable processes. Beddor argues that this synthesis allows each
tradition to profit from the other’s explanatory resources. In particular, it
enables reliabilists to explain epistemic defeat without abandoning their nat-
uralistic ambitions.
Chapter 8, Knowledge, Action, and Defeasibility by Carlotta Pavese, reviews
some motivations for a ‘knowledge-centred psychology’—a psychology where
knowledge enters centre stage in an explanation of intentional action,—it
outlines a novel argument for the claim that knowledge is required for
intentional action, and discusses some of its consequences, in particular for
the debate on the defeasibility of know-how. As Pavese argues, a knowledge-
centred psychology motivates the intellectualist view that know-how is a
species of know-that. In its more extreme form, this view is committed to an
epistemologically substantial claim—i.e. that the epistemic profile of know-

how is the same as that of propositional knowledge. If that is correct, one
corollary of intellectualism is that the defeasibility of know-how patterns with
that of knowledge. Against recent challenges, Pavese argues that this predic-
tion is burned out, for know-how and knowledge are indeed defeated exactly
when one’s ability to intentionally act is defeated.
Matthew McGrath’s Chapter 9, Undercutting Defeat: When it Happens and
Some Implications for Epistemology, investigates recent scepticism about the
Pollockian operative commonality between rebutting and undercutting defeat.
In particular, the chapter looks into the plausibility of the claim that the
mechanism of undercutting defeat, in contrast to that of rebutting defeat,
occurs only in conjunction with certain higher-order contributions, i.e. with
beliefs about the basis on which one does or would believe. McGrath argues
that, in the case of defeat of inferential justification, undercutting defeat is a
genuine phenomenon and takes roughly the shape Pollock suggests, not
needing help from higher-order beliefs or justifications. However, according
14     
to McGrath, for noninferential justification, the Pollockian account is in

trouble. This difference, it is argued, has important implications for epistem-
ology: for one, what seems to follow is that there is less noninferential
perceptual or testimonial justification than is commonly thought.
Chapter 10, Defeaters as Indicators of Ignorance, by Clayton Littlejohn and
Julien Dutant develops a novel account of the nature of rationality defeat.
According to their view, defeaters are indicators of ignorance, evidence that
we’re not in a position to know some target proposition. When the evidence
that we’re not in a position to know is sufficiently strong and the probability
that we can know is too low, it is not rational to believe. Littlejohn and Dutant
argue that their account retains all the virtues of the more familiar approaches
that characterize defeat in terms of its connection to reasons to believe
or to confirmation but provides a better approach to higher-order defeat.
Furthermore, it is argued, the view provides a unified normative framework,
one that gives a unifying explanation of the toxicity of different defeaters that
is grounded in a framework that either recognizes knowledge as the norm of
belief or identifies knowledge as the fundamental epistemic good that full
belief can realize.
In Chapter 11, Competing Reasons, Justin Snedegar considers different
ways that reasons, bearing on our options, can compete with one another to
determine the overall normative status of those options. Two key claims
defended in the chapter are (i) that the theory of this competition must include
a distinct role for reasons against, in addition to reasons for, and (ii) that the
theory must allow for comparative verdicts about which options are more
strongly supported than others, rather than simply which options are required
or permitted. Snedegar rejects a simple and familiar balancing account of the
competition, as well as an account that understands the competition in terms
of giving and answering criticisms of the options, and he introduces a new
account that incorporates a distinct role for reasons against.
In Chapter 12, Perceptual Reasons and Defeat, Mark Schroeder focuses on
the defeasibility of perceptual evidence. If something looks red to you, it is
reasonable to believe that it is red, but if you realize you are wearing rose-
tinted glasses, it may not be reasonable at all to believe this, unless you have
some independent source of evidence. Schroeder compares several models for
how to understand this phenomenon. These models differ in their answers to
two questions: what evidence we get about the external world through per-
ception, and what our having that evidence consists in. According to
Schroeder, some models have the advantage of fitting seamlessly into general
accounts of non-monotonic inference but carry with them a commitment to a
 15
restricted space of possible options in general epistemology. Their challenge is

to extract an adequate treatment of objective defeat from their elegant treat-
ment of subjective defeat. Other models lead to commitments about the
differences in what explains reasonable belief in good and bad cases. Their
challenge is to extract an adequate treatment of subjective defeat from their
elegant treatment of objective defeat. According to Schroeder, his favourite,
non-factive content model shares neither of these problems, and it offers
parallel elegant treatments of both objective and subjective defeat.
References
Alston, W. (1989). Epistemic Justification. Ithaca: Cornell University Press.

Alston, W. (2002). Plantinga, naturalism, and defeat. In: James Beilby (ed.)
Naturalism Defeated? Essays on Plantinga’s Evolutionary Argument against
Naturalism. Ithaca: Cornell University Press: 176–203.
Baker-Hytch, M. and Benton, M. (2015). Defeatism defeated. Philosophical
Perspectives, 29(1): 40–66.
Beddor, B. (2015). Process reliabilism’s troubles with defeat. The Philosophical
Quarterly, 65(259): 145–59.
Bergmann, M. (2006). Justification without Awareness. New York: Oxford
University Press.
Brown, J. (2018). Fallibilism: Evidence and Knowledge. Oxford: Oxford University
Press.
Chandler, J. (2013). Defeat reconsidered. Analysis, 73(1): 49–51.
Christensen, D. (2007). Epistemology of disagreement: The good news.
Philosophical Review, 116(2): 187–217.
Christensen, D. (2010). Higher-order evidence. Philosophy and Phenomenological
Research, 81(1): 185–215.
Coates, A. (2012). Rational epistemic akrasia. American Philosophical Quarterly,
1(2): 113–24.
Elga, A. (2007). Reflection and disagreement. Noûs, 41(3): 478–502.
Feldman, R. (2005). Respecting the evidence. Philosophical Perspectives, 19:
95–119.
Frances, B. and Matheson, J. (2019). Disagreement. In: Edward N. Zalta (ed.) The
Stanford Encyclopedia of Philosophy (Winter 2019 Edition). Stanford: Stanford
University Press. https://plato.stanford.edu/archives/win2019/entries/disagree
ment/.
16     
Goldberg, S. (2016). On the epistemic significance of evidence you should have

had. Episteme, 13(4): 449–70.
Goldman, A. (1979). What is justified belief? In: G. S. Pappas (ed.) Justification
and Knowledge. Dordrecht: Reidel: 1–25; reprinted in A. I. Goldman (2012).
Reliabilism and Contemporary Epistemology. New York: Oxford University
Press: 29–49.
Greco, D. (2017). Cognitive mobile homes. Mind, 126(501): 93–121.
Greco, J. (2010). Knowledge as an Achievement. Cambridge: Cambridge University
Press.
Green, C. R. (2008). The Epistemology of Testimony. Internet Encyclopedia of
Philosophy. http://www.iep.utm.edu/ep-testi/.
Harman, G. (1973). Thought. Princeton, NJ: Princeton University Press.
Hawthorne, J. and Srinivasan, A. (2013). Disagreement without transparency:
Some bleak thoughts. In: D. Christensen and J. Lackey (eds.) The
Epistemology of Disagreement: New Essays. Oxford: Oxford University Press:
9–30.
Horowitz, S. (2014). Epistemic Akrasia. Nous, 48(4): 718–44.
Kelly, T. (2005). The epistemic significance of disagreement. In: J. Hawthorne and
T. S. Gendler (eds.) Oxford Studies in Epistemology. Vol. 1. Oxford: Oxford
University Press: 167–96.
Kelly, T. (2010). Peer disagreement and higher-order evidence. In: R. Feldman and
T. A. Warfield (eds.) Disagreement. Oxford: Oxford University Press: 111–74.
Kornblith, H. (2015). The role of reasons in epistemology. Episteme, 12(2):
225–39.
Lackey, J. (1999). Testimonial knowledge and transmission. The Philosophical
Quarterly, 49: 471–90.
Lackey, J. (2008). Learning from Words. Oxford: Oxford University Press.
Lasonen-Aarnio, M. (2010). Unreasonable knowledge. Philosophical Perspectives,
24: 1–21.
Lasonen-Aarnio, M. (2012). Higher-order evidence and the limits of defeat.
Philosophy and Phenomenological Research, LXXXVIII: 314–45.
Lasonen-Aarnio, M. (2014). The dogmatism puzzle. Australasian Journal of
Philosophy, 92(3): 417–32.
Lasonen-Aarnio, M. (2020). Enkrasia or evidentialism. Philosophical Studies, 177:
597–632.
Lyons, J. (2009). Perception and Basic Beliefs: Zombies, Modules, and the Problem
of the External World. Oxford: Oxford University Press.
 17
Pappas, G. (2017). Internalist vs. externalist conceptions of epistemic justification.

In: Edward N. Zalta (ed.) The Stanford Encyclopedia of Philosophy (Fall 2017
edn). https://plato.stanford.edu/archives/fall2017/entries/justep-intext/.
Plantinga, A. (2000). Warranted Christian Belief. New York: Oxford University
Press.
Pollock, J. (1986). Contemporary Theories of Knowledge. Savage, MD: Rowman
and Littlefield.
Pryor, J. (2004). What’s wrong with Moore’s argument. Philosophical Issues, 14:
349–78.
Pryor, J. (2013). Problems for credulism. In: C. Tucker (ed.) Seemings and
Justification. Oxford: Oxford University Press: 89–131.
Simion, M. (2020). Epistemic Norms and Epistemic Functions. Unpublished Book
Manuscript.
Sturgeon, S. (2014). Pollock on defeasible reasons. Philosophical Studies, 169(1):
105–18.
Sudduth, M. (2018). Defeaters in Epistemology. Internet Encyclopaedia of
Philosophy. https://www.iep.utm.edu/ep-defea/#SH1b.
Swinburne, R. (2001). Epistemic Justification. Oxford: Clarendon Press.
Titelbaum, M. (2015). Rationality’s fixed point (or: in defence of right reason). In:
S.T. Gendler and J. Hawthorne (eds.), Oxford Studies in Epistemology. Vol. 5.
Oxford: Oxford University Press: 253–94.
Weatherson, B. (2019). Normative Externalism. Oxford: Oxford University Press.
Wedgwood, R. (2011). Justified inference. Synthese, 189(2): 1–23.
Weisberg, J. (2009). Commutativity or holism? A dilemma for conditionalizers.

British Journal for the Philosophy of Science, 60: 793–812.
Williamson, T. (2009). Replies. In: P. Greenough and D. Pritchard (eds.),
Williamson on Knowledge. Oxford: Oxford University Press: 279–384.
2
The Normativity of Knowledge and the
Scope and Sources of Defeat
Sanford C. Goldberg
2.1.
Knowledge is a normative standing. Perhaps this much is uncontroversial.

What this amounts to, however, is not. Minimally, to say that knowledge is a
normative standing is to say that there are evaluative standards that must be
satisfied if one is to count as having knowledge. Those standards include truth
as well as the epistemic standards governing doxastic states. However, wide-
spread agreement on this much does not preclude disagreement over the
normativity of knowledge itself.
A good deal of the disagreement over the normativity of knowledge con-
cerns the terms of the debate. More specifically, there is controversy over
which term(s), if any, to treat as basic. Should we treat the normativity of
knowledge as basic, and understand normativity of all other epistemic notions

(justification; rationality; evidence) in terms of it? Such is part and parcel of the
knowledge-first programme in epistemology.¹ Opposing this are those who
deny the priority of knowledge in an account of epistemic normativity. Other
notions which are candidates for foundational status in such an account are
justification, rationality, accuracy, and coherence.² Alternatively, there will be
those who deny that there is any foundational notion; all we have are a variety
of “epistemic desiderata.”³
However, even after we settle the matter of the terms of the debate, there
remains another source of disagreement: the nature of the evaluative standards
of epistemology themselves. To take a particularly clear example, consider
justification. Suppose that we propose to account for the normativity of
¹ Williamson (2000), to whom we owe this approach, appears to deny that there is much illumin-
ation to be had regarding standards other than knowledge.
² For a discussion, see Goldberg (2018).
³ The term and the claim are Alston’s; see Alston (1993, 2005).
       19
knowledge at least partly in terms of the normativity of justification, and so we

embrace the project of providing an illuminating account of the evaluative
standards for justified belief. Even so, we can still disagree over what those
standards are. Do they require reliability in the processes through which one
formed and sustained one’s belief? Good reasons or adequate evidence?
Coherence with one’s standing corpus of belief? Responsibility in belief-
formation and belief-maintenance?
In this paper, I want to side-step these difficult questions, and instead
identify two claims that we should want as part of our account of the
normativity of knowledge. I formulate the claims as follows:
CONFIDENCE
If S knows that p, then S is entitled to be confident that p.⁴
AUTHORIZATION
If S knows that p, then S is in an epistemic position from which to authorize
others to believe that p.
As I say, I regard the truth of both CONFIDENCE and AUTHORIZATION as

desiderata for accounts of the normativity of knowledge: if your favored
account fails to be consistent with either of these, so much the worse for
your account. (Although I won’t insist on it here, I am tempted to think that an
adequate account must not only cohere with these claims but be able to explain
why they are true as well.)
I should note both CONFIDENCE and AUTHORIZATION are relatively

weak claims.
For one thing, they state sufficient conditions, not necessary conditions, on
their consequences. CONFIDENCE tells us that knowledge of p is sufficient for
being entitled to confidence that p, not that it is necessary. (For all that
CONFIDENCE tells us, there are epistemic standings that are less demanding
than knowledge, but which are consistent with being entitled to confidence.)
AUTHORIZATION tells us that knowledge of p is sufficient for being in a
position to authorize others to believe that p, not that it is necessary. For all
that AUTHORIZATION tells us, there are epistemic standings that are less
demanding than knowledge which are consistent with being in a position
from which to authorize others to believe that p. I should add, too, that
⁴ Can’t one know that p, despite having misleading evidence as to one’s epistemic position with
respect to p? Such a view appears to be endorsed by several authors in the higher-order evidence debate:
Williamson (2014); Larsonen-Aarnio (2010, 2014); Baker-Hytch and Benton (2015), and others. I will
reply to this thought in the next section.
20  . 
AUTHORIZATION is a claim about S qua speaker; it is silent on what

conditions there are if an audience is to acquire this authorization from
S (when S testifies in a way that expresses her knowledge).
The fact that AUTHORIZATION is silent on the conditions on an audi-
ence’s acquiring an authorization from a speaker is important for two further
reasons, both of which point to its relative weakness as a claim. First, this
silence renders AUTHORIZATION neutral on the flash-point in the epistem-
ology of testimony: AUTHORIZATION has no implications that favor either
reductionism (according to which an audience requires positive reasons to
regard a testimony as credible prior to being entitled to form a belief through
accepting the testimony) or anti-reductionism (which holds that it suffices that
the audience lack reasons for doubting the credibility of the testimony). But
second, AUTHORIZATION’s silence on the conditions on an audience’s acquir-
ing an authorization from a speaker enables proponents of AUTHORIZATION
to allow for cases in which, while the speaker was in a position to authorize
(because she had knowledge-sufficient evidence for her claim), the audience was
not in a position to acquire that authorization from her—he was not in a position
to acquire the knowledge that her testimony expresses—e.g., because he had a
defeater that the speaker herself lacked.⁵
There is one final way in which both CONFIDENCE and AUTHORIZATION
are relatively weak claims: they are silent on the source of the entitlement and the
authorization. All that they say is that if one knows, then one is entitled to
confidence and one is in an epistemic position from which to authorize others
to believe. Neither claim weighs in on whether it is knowledge itself that grounds

the entitlement and the authorization, or whether instead it is the satisfaction of
evaluative epistemic standards which grounds the entitlement and the authoriza-
tion (and which serves as a condition on knowledge).
I will not be offering any detailed defense of the claim that CONFIDENCE
and AUTHORIZATION are desiderata for an account of the normativity
of knowledge. But I can say a few brief things. First, both of these claims are
in keeping with various views in the literature about the normativity of other
related phenomena: assertion and action,⁶ for example. Take the recent dis-
cussion of the norm of assertion. Many people hold that knowledge
warrants assertion;⁷ they will not bat an eye at endorsing AUTHORIZATION.
⁵ I thank Jessica Brown for indicating the need to make this point.
⁶ Not everyone is agreed that knowledge is sufficient to warrant action: see e.g., Brown (2008a,
2008b) and Reed (2010). In at least some of these cases we might come to doubt whether knowledge is
sufficient to warrant assertion as well (for which see also the references in fn 9).
⁷ For an overview of the various positions, see Goldberg (2015a, 2015b).
       21
I note, too, that since AUTHORIZATION states a sufficient condition (not a

necessary condition) on being in a position to authorize others to believe,
anyone who thinks that assertion is warranted by something which itself is
entailed by knowledge—justification, say—will not bat an eye at endorsing
AUTHORIZATION.⁸ Others hold that knowledge warrants belief.⁹ Anyone
who does should not bat an eye at endorsing CONFIDENCE. But so too
anyone who thinks that what warrants belief is some status that is entailed by
knowledge—justification, say—should endorse CONFIDENCE. Again, I would
guess that this is most of the field of those who think about the norm of (or
epistemic standards on) belief.
In what follows, I want to use both CONFIDENCE and AUTHORIZATION
to reflect on the nature and scope of epistemic defeat.
2.2.
How might CONFIDENCE and AUTHORIZATION be used so as to cast

light on the nature and scope of defeat? While neither of these principles
mention defeat, they can be connected to the phenomenon of defeat through
their connection with knowledge. To see this, note that they can be used to
draw conclusions asserting a failure or lack of knowledge. Here it will be
helpful to deal, not with CONFIDENCE and AUTHORIZATION themselves,
but with their inverse, which I will call the defeat corollaries:¹⁰
DEFEATCONF
If S is not entitled to be confident that p, then S does not know that p.
DEFEATAUTH
If S is not in a position to authorize others to believe (with confidence) that p,
then S does not know that p.
My reason for calling these the ‘defeat corollaries,’ despite the fact that they
nowhere mention defeat, is straightforward: the presence of defeaters is
⁸ This covers most of the field of those who endorse some norm of assertion. Most, but not all: some
endorse a truth norm of assertion (Weiner (2005); Whiting (2013)); others endorse a variable-standard
norm of assertion and allow that there are cases in which the standard can be satisfied even though the
speaker is not justified in believing what she asserts (McKinnon (2013, 2016) and Gerken (2014)). And
I should acknowledge, too, the arguments supporting the idea that there are cases in which knowledge
itself is not sufficient to warrant assertion (see Lackey (2011)).
⁹ See e.g., Williamson (2002).
¹⁰ The following discussion has been greatly improved by comments from Jessica Brown.
22  . 
typically what offers the best explanation for the truth of the antecedents of
DEFEATCONF and DEFEATAUTH. In particular, given a subject who otherwise
has a good deal of evidence that p but who nevertheless fails to be entitled to be
confident that p (or to be in a position to authorize others to believe that p),
the best explanation for our subject’s failure to satisfy these antecedents is that
there are relevant defeaters in play. In these cases, the subject’s failure to
occupy the normative standing that knowledge brings in its wake—and, by
extension, the subject’s failure to know—indicate the presence of a relevant
defeater. In this way these corollaries are useful when our interest turns to the
presence of defeaters.
Before proceeding to use the corollaries in this way, it is important to
highlight the two distinct types of case in which the corollaries might be
applied. As noted, both DEFEATCONF and DEFEATAUTH can be used to
address whether a subject S knows that p: if S fails to be entitled to be confident
that p, or alternatively if S fails to be in a position from which to authorize
others to believe that p, then (by the lights of these corollaries) S fails to know.
There are two distinct types of case to which these might be applied, corres-
ponding to the two types of scenario in which we might be interested in the
question whether S knows that p. The first type of scenario—what I will call
the knowledge-prevention scenario—involves a context in which S has relevant
evidence but has yet to acquire the belief that p in the first place. Here the
question of whether S knows that p is tantamount to the question whether her
newly acquired belief that p amounts to knowledge; and when the corollaries
deliver a negative verdict, the result, in effect, is that defeaters prevented the
acquisition of knowledge. The second type of scenario—what I will call the
knowledge-defeating scenario—arises for a subject who currently knows that
p, but who gets misleading evidence at some point down the road. Here the
question is whether S’s belief that p retains its status as knowledge even after
the acquisition of misleading counter-evidence; and when the corollaries
deliver a negative verdict, the result, in effect, is that defeaters undermined
or defeated the knowledge itself.
It would seem that we can apply DEFEATCONF and/or DEFEATAUTH to
either sort of scenario.
Take a scenario of the first (knowledge-prevention) type. Suppose that,
prior to forming the belief that p, one’s total evidence, including whatever
misleading evidence one has, is such that (i) one isn’t entitled to be confident
that p and (ii) one isn’t in a position to authorize others to believe that p. For
example suppose that, before Lilly enters a room, she is told by an apparently
reliable informant that the room’s walls are cleverly illuminated by colored
       23
lighting, and then she walks in and encounters what appears to be a red wall. It
is plausible that if under these conditions (and without further ado) she were
to come to believe that the wall is red, her belief would not constitute
knowledge that the wall is red. In this way the defeat corollaries can be used
to bring out that (misleading) evidence can prevent one from acquiring
knowledge in the first place—even in cases in which, bracketing the misleading
evidence, one has what would otherwise be knowledge-sufficient evidence.
But what of scenarios of the second (knowledge-defeating) type, in which
one already has the knowledge that p but then encounters further relevant
counterevidence (albeit of a misleading variety)? For example, suppose that
Sally saw (what appeared to be) the red wall before getting the (apparently
trustworthy) testimony that she was in a room whose walls were cleverly
illuminated by colored lights. And suppose that the wall really is red (no
lighting is involved). It is plausible to think that in this case, while she
originally acquired knowledge of the wall’s being red (prior to getting the
testimony), she loses her knowledge on getting the testimony. The hypothesis
of lost knowledge can be further supported on the assumption that scenarios
of the first (knowledge-preventing) type are possible. Given this assumption,
one can deny the possibility of scenarios of the second (knowledge-defeating)
type only at the cost of having to insist that the order in which one acquires
evidence matters to whether one knows. The proof of this is that for any case of
the second (knowledge-defeating) type, we can construct a corresponding case
of the first (knowledge-preventing) type in which the misleading evidence
came prior to the other evidence bearing on the proposition that p, thereby
preventing the acquisition of knowledge that p in the first place. So, if we deny
the possibility of knowledge-defeating cases while allowing for the possibility
of knowledge-preventing cases, then in effect we are committed to thinking
that the order of evidence acquisition is epistemically significant in this way.
But this is implausible. Consequently, anyone who grants the possibility of
scenarios of the first (knowledge-preventing) type ought to grant the possibil-
ity of scenarios of the second (knowledge-defeating) type.¹¹
Even so, I acknowledge that the claim in the second (knowledge-defeating) type
of scenario, to the effect that knowledge itself is defeasible, is controversial¹²—more
controversial, anyway, than is the claim in the first (knowledge-preventing) type
¹¹ Compare Brown (2018: Ch. 5), where she presents a slightly different argument for the same
conclusion—namely, that if defeaters can prevent one from knowing then defeaters can undermine
one’s prior knowledge as well.
¹² See e.g., Lasonen-Aarnio (2010, 2014); Williamson (2014); Schoenfield (2015); Baker-Hytch and
Benton (2015).
24  . 
of scenario (that misleading evidence can prevent one from acquiring knowledge
in the first place). In order to minimize the degree to which I antagonize those
who reject the possibility of the knowledge-defeating type of scenario, in what
follows, for the most part I will be restricting my attention to scenarios of the
knowledge-prevention type.¹³ This will enable me to make the points I wish to
make without having to take a stand on the somewhat controversial question
whether knowledge itself is defeasible.
2.3.
Before proceeding to use the defeat corollaries to discern the nature and scope
of the phenomenon of defeat, we would do well first to acquire some confi-
dence that both of the defeat corollaries themselves are acceptable. So, let us
apply them to cases that everyone should agree are cases of a lack of knowledge
owing to defeat, in the hope that they deliver this verdict.
PERCEPTION
Percy has been told (by a speaker he has good reason to trust) that he is about
to enter a room in which many of the walls are expertly illuminated by colored
lights to give the impression of being painted. Percy himself has no further
basis for questioning this testimony. Percy then goes into the room and
perceives what appears to be a red wall. (He has no basis for confirming
whether this is the effect of red lighting.) Verdict: he is not in a position to

know (through perception) that the wall is (painted) red.
DISCERNMENT
Dalia is looking at what she initially takes to be deer tracks in the mud.
However, she soon recalls that she is in a part of the state in which there are
several different animals that make very similar tracks, and she has observed
previously that she is not particularly good at discerning between these
different types of track. Verdict: she is not in a position to know (through
perceptual discernment) that the tracks are those of a deer.
MEMORY
Malia seems to remember that she was twelve the first time she got on a plane.
However, she knows that she is currently on a medication one of whose known
¹³ This said, I will occasionally present cases of the second (knowledge-defeating) type as well.
Sometimes I will do so because I am describing cases others have presented in these terms (as a threat to
previously existing knowledge); on other occasions I will do so for reasons of ease of exposition.
       25
side effects is to create false memory impressions. Malia has no basis for ruling
out the possibility that her present memory impression is false. Verdict: she is
not in a position to know (through memory) that she was twelve the first time
she got on a plane.
I take it that a ‘no knowledge’ verdict is obvious in these cases, and that this
verdict reflects the presence of the defeater. Happily, the defeat corollaries
deliver this result. In none of these cases is the subject entitled to be confident,
and in none of them is the subject in a position to authorize others to believe
(with confidence). Thus: Percy is not entitled to be confident that he is looking
at a red wall, nor is he in a position to authorize others to believe that the wall
is red; Dalia is not entitled to be confident that she is looking at deer tracks, nor
is she in a position to authorize others to believe that the tracks are of deer;
Malia is not entitled to be confident that she was twelve the first time she got
on a plane, nor is she in a position to authorize others to believe that she was
twelve the first time she got on a plane. These facts, together with
DEFEATCONF and DEFEATAUTH, can be used to derive the ‘no knowledge’
verdicts. And, if we ask why the antecedents of these corollaries hold, the best
explanation is that the subject has a relevant defeater. These are cases, then, in
which the presence of a defeater undermines the subject’s occupying the
normative standing associated with knowledge, and so (we can conclude)
fails to know.
Of course, it is not particularly interesting or noteworthy to learn that the
corollaries deliver the desired results in these cases. But that, after all, is
precisely the point: my aim in these examples is to confirm the validity of
the corollaries, not to use them to get interesting results. Now, having been
given some reason to trust their validity, we can go on to employ the defeat
corollaries in cases in which the presence (or absence) of defeaters is not so
obvious. To this end, I will focus on cases in which an otherwise epistemically
well-positioned subject fails to be entitled to confidence—alternatively, in
which such a subject fails to be in a position to authorize others to believe
(with confidence). The upshot will be that in such cases we have good reasons
to postulate the presence of relevant defeaters, as the best explanation for the
subject’s diminished epistemic standing (despite her otherwise strong epi-
stemic position). That we can get this result in cases in which the result itself
is not obvious is proof positive of the utility of the corollaries in helping us
discern the nature and scope of defeat.
Let us now move on to some of these cases.
26  . 
2.4.
There has been a good deal of discussion of certain cases of would-be higher-
order defeat.¹⁴ These are cases in which a given subject has what in fact is
conclusive first-order evidence that p, but also has evidence—unbeknownst to
her, misleading evidence—that she is not in a position to reliably assess her
first-order evidence. One example, based on a case due to David Christensen,
is the following:
LOGIC
Lonny is highly competent at logic. Prior to taking a logic exam, she is told by a
person she regards as sincere and competent that there is a good chance that,
unbeknownst to her, she was recently administered a drug whose side effect is
that it renders one highly unreliable in discerning which would-be proofs are
actual proofs, even in very simple cases. (One effect of the drug is to make even
simple invalid inferences seem perfectly and straightforwardly valid.) While
under these conditions, she examines what in fact is (and what strikes her as) a
fairly straightforward proof of the proposition that p.
Does Lonny know that p, or does the testimony she received serve as a (higher-
order) defeater? While many people who have addressed this matter agree
with Christensen that Sally does not know that p,¹⁵ others do not.¹⁶
My own sympathies are with Christensen. I come at the issue by way of two
alternative questions. They are:
After reviewing the would-be proof (having previously gotten the testi-
mony), is Lonny entitled to be confident that p?
After reviewing the would-be proof (having previously gotten the testi-
mony), is she in a position from which to authorize others to believe that p?
It seems clear to me that the verdict for both of these is “no.” And either of
these verdicts, together with the relevant defeat corollary, can be used to draw
the conclusion that Lonny doesn’t know that p.
¹⁴ Christensen (2010); Lasonen-Aarnio (2010, 2014); Horowitz (2014); Horowitz and Sliwa (2015);
Schoenfield (2015, 2018); Brown (2018); Skipper (2018); Ye (2018).
¹⁵ See e.g., Horowitz (2014); Brown (2018); Ye (2018); Goldberg and Matheson (forthcoming). (For
his part, Christensen is less interested in the question whether Sally knows that p, than he is in
modeling how the higher-order evidence affects Sally’s warranted credence in p.)
¹⁶ See e.g., Lasonen-Aarnio (2010, 2014); Schoenfield (2015); Baker-Hytch and Benton (2015).
       27
It is perhaps no surprise that the defeat corollaries can be brought to bear on

the sort of higher-order defeat illuminated by Christensen’s work. (While
Christensen himself does not formulate the corollaries themselves, he does
appear to use considerations in this vicinity in defense of his analysis.) But,
I believe that the corollaries can be used in contexts in which the hypothesis that
defeaters are present will be much more controversial. More specifically, the
corollaries provide some support for the claim that when others have legitimate
social expectations concerning one’s epistemic condition and one fails to live up
to these expectations, this can generate defeaters—even in cases in which one
lacks the defeating evidence itself. This claim is a conclusion I draw in Goldberg
(2018); in what follows I use the defeat corollaries to bring this conclusion out.
To this end, I offer several sorts of case in which others have legitimate
expectations of a subject, where the subject fails to fulfill these expectations,
and where this fact (among others) bears on the question of knowledge. I begin
with two cases that have been previously discussed by others (albeit not in
precisely these terms).
UNOPENED LETTER (from Pollock)

Elodie believes on good evidence that Clancy is currently in San Francisco.
However, Clancy wants to get Elodie to believe that she (Clancy) is in
Providence, Rhode Island, and so she arranges to provide Elodie with evidence
to this effect. So, Clancy writes a letter to that effect and has the letter mailed to
Elodie from a friend in Providence (to get the date and postmark right). One
day, as Elodie reflects on Clancy’s being in San Francisco, the letter from
Clancy is sitting in a pile of mail that is right in front of her, unopened, on the
desk. In fact, Clancy is in San Francisco.
NEWSPAPER (from Harman)
Ying Yue reads the morning paper to learn that the president has been
assassinated, and that there has been a coup. In fact, this is precisely what
happened. But later in the day the generals now in power want to fool
everyone into thinking that no such thing happened. So, they order the
afternoon edition of the newspaper to issue a retraction. The afternoon edition
goes out across the community and is read by virtually everyone else.
However, Ying Yue never sees the afternoon edition (and never overhears
others speaking of its contents). As a result, she continues to believe that the
president has been assassinated and that there has been a coup.
Both Pollock (explicitly) and Harman (implicitly) rely on the idea that there is
a social expectation of the subject in the case in question to avail herself of the
28  . 
relevant evidence: in Elodie’s case, there is a legitimate “social expectation”

that we read the letters that we receive, whereas in Ying Yue’s case, there is a
legitimate social expectation that we stay abreast of the very significant break-
ing news in our community when it is prevalent. The failure to satisfy this
expectation, together with the fact that had they done so they would no longer
have been justified in believing as they do, defeats the subjects’ knowledge in
both cases.
We can reinforce this result through application of the defeat corollaries:
applying them to the cases at hand, we get the result that there is no knowledge
in either case. To bring this out, we address two questions raised by these
corollaries regarding the subjects of these cases:
In a situation in which a subject (Elodie; Ying Yue) violates a legitimate social

expectation in connection with her belief that p, where this expectation
involves getting further evidence that would have defeated her justification
to believe, is the subject herself entitled to be confident that p?
In a situation in which a subject (Elodie; Ying Yue) violates a legitimate social
expectation in connection with her belief that p, where this expectation
involves getting further evidence that would have defeated her justification
to believe, is the subject in a position from which to authorize others to
believe that p?
I submit that in each case (UNOPENED LETTER and NEWSPAPER) the

answer to both questions is negative. This is diagnostic of the defeat of the

would-be knowledge that the two subjects would otherwise have had. Here, as
previously, the hypothesis alleging the presence of defeaters is the best explan-
ation of the fact that a subject who otherwise is well-positioned epistemically
fails to be entitled to confidence, and fails to be in a position from which to
authorize others to believe with confidence.
This diagnostic enables us to address an otherwise perplexing aspect of
these cases. In effect, the claim on the table is that evidence one doesn’t have—
the contents of a letter (or of a prevalent afternoon edition of the newspaper)
that one didn’t read—can nevertheless defeat one’s knowledge, even when one
is oblivious to the very existence of that evidence. But how can this be? How
can evidence that a subject herself lacks play any epistemic role, let alone a
defeating role, at all? It won’t do to respond to this question by insisting
that the defeat works only when the subject herself is aware that there
is further evidence out there (which she fails to have); for, while the
       29
evidence-of-evidence-is-evidence view might account for some cases, it does

not account for the cases as Pollock and Harman described them. Happily, the
defeat corollaries give us another way to address this seemingly perplexing
element of these cases: when the evidence one doesn’t have is evidence that
one should have had, where this reflects others’ legitimate expectations of one,
then one’s failure to have that evidence affects—or, at any rate, can affect
(more on which in a moment)—one’s right to be confident, as well as whether
one is in a position from which to authorize others to believe. Insofar as
knowledge is a status that underwrites the right to confidence and the position
from which to authorize others, it should come as no surprise, then, that in
these sorts of cases evidence one doesn’t have can prevent one from knowing:
such evidence bears against one’s entitlement to be confident and against one’s
being in a position to authorize others to believe.
We can develop this analysis, and illustrate its plausibility, in both of these
cases.
Starting with UNOPENED LETTER, the key claim (made by Pollock
himself) is that there is a “social expectation” that one reads the letters
addressed to one (Pollock 1986). We might think of this expectation as
spelling out our sense of when and where and under what circumstances
one is expected to do so—though there may be some indeterminacy on
these matters. Even so, there are clear cases, and UNOPENED LETTER was
meant by Pollock to be such a case. When this expectation is unmet, and the
result is that one fails to have some crucial evidence which one would have had
if one had met the expectation, we are in a position in which knowledge is at

risk. On my analysis, Elodie’s knowledge that Clancy is in San Francisco would
be defeated just in case the result of expanding Elodie’s current belief set so as
to include the information contained in the letter would be a belief set in which
Elodie isn’t justified in believing that Clancy is in San Francisco. The hypoth-
esis of defeat is the best explanation for why Elodie fails to be entitled to
confidence of Clancy’s whereabouts (or to authorize others to believe with
confidence), despite the fact that Elodie has good evidence for thinking that
Clancy is in San Francisco.
A similar analysis is plausible in connection with NEWSPAPER. (Here I am
going a bit beyond what Harman himself said, but I think this analysis is in
keeping with the spirit of his example.) There is a social expectation that citizens
keep abreast of the significant breaking news in their communities (especially
when its dissemination is prevalent). Again, we might think of this expectation as
spelling out our sense of when, where, how, and under what circumstances one is
expected to do so—though there may be some indeterminacy on these matters.
30  . 
Even so, there are clear cases, and NEWSPAPER was meant by Harman to be
such a case. When this expectation is unmet, and the result is that one fails to
have some crucial evidence which one would have had if one had met the
expectation, we are in a position in which knowledge is at risk. On my analysis,
Ying Yue’s knowledge that there has been a coup would be defeated just in case
the result of expanding Ying Yue’s current belief set so as to include the infor-
mation in question—that there has been an afternoon edition of the newspaper
publishing a retraction—would be a belief set in which Ying Yue isn’t justified in
believing that there has been a coup. The hypothesis of defeat is the best
explanation for why Ying Yue, despite being in an otherwise strong epistemic
position with respect to the existence of the coup from having read the morning
paper, nevertheless fails to be entitled to confidence regarding the coup (or to
authorize others to believe with confidence).
If these analyses are correct, then the defeat corollaries have a significance
that may not be obvious at first glance. They can be used to make clear that the
normativity of knowledge includes a social dimension, and that this social
dimension can render evidence one doesn’t have epistemically significant with
respect to one’s epistemic standing on some proposition.¹⁷ As noted previ-
ously, such an idea might initially seem curious: how can evidence one doesn’t
possess have any epistemic significance whatsoever, let alone prevent one from
acquiring knowledge (or defeat knowledge one already has)? The answer that is
suggested, in part, by the defeat corollaries is that this prospect merely reflects
the social dimension of knowledge’s normativity: to know that p requires a
justified belief that p held in circumstances in which there is no evidence one

should have—that is, no evidence which it is (legitimately) socially expected that
one would have—which, if one had it, would undermine one’s entitlement to
confidence and would prevent one from authorizing others to believe. When
one’s justified belief that p meets this condition, I will call it socially robust.
2.5.
I just argued that the defeat corollaries can be used to support the idea that
there is a kind of normative (social) defeat. These are cases in which the
acquisition of knowledge is prevented (or existing knowledge is defeated)
owing to the subject’s failure to satisfy legitimate social expectations as to
her epistemic conduct or condition, in contexts in which there is adequate
¹⁷ This point has been emphasized in the ethics literature by Moody-Adams (1994).
       31
counterevidence that she would have acquired had she satisfied those
expectations. Both UNOPENED LETTER and NEWSPAPER were cases of
this sort; and the application of the defeat corollaries enabled us to discern this
sort of “normative (social) defeat.” However, my analysis may strike some as
unconvincing. It might be wondered, for example, whether, far from estab-
lishing a kind of normative (social) defeat, these cases are better regarded as
undermining the defeat corollaries themselves. In this section, I argue that this
reaction is wrongheaded.
What might prompt one to jettison the defeat corollaries themselves, rather
than having to countenance the possibility of a kind of normative (social)
defeat that is supported by those corollaries?
The obvious way to try to motivate this reaction would be to insist on the
idea (i) that unpossessed evidence can never have any epistemic significance
for the subject who lacks the evidence in question. For this idea is incompatible
with the combination of two further views, (ii) that both of the defeat
corollaries are true, and (iii) that the subjects in UNOPENED LETTER and
NEWSPAPER are such that, given the facts of the case, neither one is entitled
to be confident or to authorize others to believe with confidence. Since the
combination of (i)–(iii) form an inconsistent triad, (at least) one of these must
be jettisoned. So those who insist on (i) will need to reject either (ii) or (iii).
Before considering the move to reject (ii)—the defeat corollaries themselves—
let us first consider the proposal to reject (iii). According to this proposal, the
subjects in UNOPENED LETTER and NEWSPAPER are entitled to confidence
in their beliefs, and they are in a position to authorize others to believe with
confidence—despite whatever impression to the contrary that is generated by
our reflection on the evidence not in their possession. Indeed, it is easy to see
how one might think to combine a rejection of (iii) with a commitment to (i).
According to (i), unpossessed evidence is epistemically insignificant. If we
combine this with a rejection of (iii), we reach the conclusion that the only
evidence that is relevant to the question whether the subject is entitled to
confidence—and the only evidence relevant to whether the subject is in a
position to authorize others to be confident—is evidence in the subject’s pos-
session. Those who endorse this approach will argue that if we restrict ourselves
to the evidence in the subject’s possession in the previously mentioned cases
(UNOPENED LETTER and NEWSPAPER), the subject is entitled to be confi-
dent, and is in a position to authorize others to be confident.¹⁸
¹⁸ See e.g. Feldman (2000) for a defense of a view in this spirit.
32  . 
Such a view inherits the burden of explaining away any intuitions to the
contrary. I can think of two ways to do so. One would be to trace such
intuitions to a conflation of the subjects’ epistemic position before and after
they learn of the facts. After the subjects learn of the facts—after Elodie reads
the letter; after Ying Yue acquires and reads a copy of the afternoon paper
containing the retraction—they are not entitled to confidence in their original
beliefs, nor are they in a position to authorize others to be confident. But this is
consistent with the claim that prior to learning of these facts both are entitled
to confidence and in a position to authorize others to be confident. The other
way to try to explain away the intuition that (iii) holds is to argue that the
intuition rests on conflating the knowledgeable third-person perspective with
the subjects’ own perspectives. It is true that anyone who is aware of the
relevant facts¹⁹—of the contents of the unopened letter, of the existence of the
afternoon news retraction—is not entitled to confidence, and is not in a
position to authorize others to be confident, on the matter at hand. But this
is consistent with saying that nevertheless both Elodie and Ying Yue are so
entitled and so positioned (prior to being apprized of those facts).
I find both of these defensive maneuvers implausible. However, to explain
why, I will be appealing to (ii)—the defeat corollaries themselves. This suggests
that those who want to insist on (i)—i.e., that unpossessed evidence is epi-
stemically insignificant—would do best to reject both (ii) and (iii). Conversely,
if (ii) is plausible, so is (iii), in which case we ought to jettison (i). (That is, we
ought to allow that unpossessed evidence can have epistemic significance for
the person who lacks it.)

My contention is that an insistence of the doctrine asserting the epistemic
insignificance of unpossessed evidence has untoward implications. While
I will appeal to the defeat corollaries themselves to bring these implications
out, my hope is that the status of the corollaries, far from being undermined, is
further confirmed in the case that follows.
PEDIATRICIAN
Patricia is a pediatrician whose practice is more than two decades old. One
day a patient of hers, Luis, comes in with a relatively rare form of asthma.
Patricia remembers reading about this condition when she was in medical
school (more than two decades ago), and she distinctly remembers the
then-prescribed treatment protocol. At the same time, the American
Academy of Pediatrics (AAP) recommends that pediatricians stay on top of
¹⁹ And who lacks a defeater-defeater.
       33
the developments for treatment protocol of all rare conditions; accordingly, it

recommends that pediatricians do a thorough literature review every 18
months to confirm that their views regarding such conditions are consistent
with currently prescribed best practices. Through no fault of her own, Patricia
never became apprized of the AAP’s recommendation. Happily, on most
matters she does literature reviews on a regular basis anyway, and so (without
explicitly aiming to do so) she keeps her practice in line with the AAP
recommendation. Still, she failed to do one on this occasion, and so does not
check to confirm that the old protocol is still being recommended.
Unfortunately for her, a series of studies in recent years support the hypothesis
that a different treatment protocol for the relevant sort of asthma is more
reliable, and on this basis the AAP changed its recommendation to the new
treatment. However, unbeknownst to the AAP community a large, the studies
that support this recommendation failed to address an obscure confounding
factor (one which the community was not in a position to appreciate at the
time). The result is that, in fact, the most reliable treatment remains the one
Patricia had learned about in medical school.
In what follows, I want to argue for two things. First, despite the fact that her
belief as to the best treatment protocol is true, Patricia is not entitled to be
confident, nor is it the case that she is in a position in which to authorize others
to believe with confidence. As a result, Patricia does not know that the best way
to treat Luis is X (where ‘X’ designates what she learned in medical school),
where—and this is my second point—her failure to know reflects Patricia’s

failure to live up to the data-collecting responsibilities that were properly
expected of her as a pediatrician.
First, it is not the case that Patricia is entitled to be confident in what she
regards as the best way to treat Luis’s asthma, nor is it the case that she is in a
position in which to authorize others to be confident in the claim that her
proposed treatment is best. It is common knowledge that as a pediatrician,
Patricia is someone on whom others rely for matters of the health of children.
What is more, as a pediatrician, Patricia has professional responsibilities,
among which are to conform in her practice to the recommendations of the
AAP. When others rely on her in matters of their children’s health, they rely
on her to live up to her professional responsibilities (even if they themselves
are in no position to articulate those responsibilities). On this score, it does not
matter that Patricia is non-culpably ignorant of (one of) her professional
responsibilities, and so is non-culpably ignorant of the AAP’s current treat-
ment recommendation. Even so, it is still her responsibility as a pediatrician to
34  . 
live up to the standards set forth by the AAP (the relevant regulatory entity), as
it is the legitimate expectation of those who rely on pediatricians that they do
so.²⁰ But then it would seem that her failure to do so, when the failure pertains
to evidence she should have had, affects her normative standing—it affects her
entitlement to confidence, her being in a position from which to authorize
others to believe. In short, because she is normatively deficient in matters of
her profession, and because this deficiency highlights a flaw—her lacking
evidence that, given her professional duties, she ought to have had, and
which bear against her belief—she is thereby normatively downgraded in the
ways noted previously: she lacks entitlement to confidence, and is not in a
position from which to authorize others to believe. Both antecedents of the
defeat corollaries hold.
Some might think to question this by insisting that Patricia’s failure to
possess the relevant evidence is merely a professional failure—a matter of
medical negligence, as it were—which as such does not constitute the basis for
any epistemic downgrade of her attitudes. But this reaction would be wrong-
headed. It is here, of course, that the defeat corollaries earn their keep: these
make clear that what many might try to explain away as a merely professional
shortcoming is in fact an epistemic shortcoming—one that deprives Patricia of
knowledge. The evidence that Patricia fails to know just is that she is not
entitled to be confident and that she is not in a position from which to
authorize others to believe. Of course, her belief is true; and it is supported
by what she learned in medical school. But given her ignorance of the evidence
she ought to have had, and the fact that this evidence counters her belief as to
the best treatment, she is not entitled to be confident, nor is she in a position
from which to authorize others to believe with confidence.
Might one agree that Patricia fails to know, but argue that this case is best seen
as a Gettier case?²¹ Like Gettier cases, PEDIATRICIAN involves additional
“nearby” evidence which, if the subject had that evidence, she would not be
justified in believing as she does. But the assimilation of this case into the
category of Gettier cases is troubled by what is distinctive in PEDIATRICIAN:
like the subjects in UNOPENED LETTER and NEWSPAPER, Patricia is prop-
erly expected to have had the evidence in question. This is no part of standard
Gettier cases: one is not standardly expected to be aware of when one’s
officemates sell their Ford, to be sensitive to the prospect of entering Fake
Barn County, to realize that sheepdogs look a lot like sheep from a distance, or
²⁰ See Goldberg (2017, 2018) for a discussion.

²¹ I thank Jessica Brown for suggesting that I consider this option.
       35
to be sensitive to the prospect that one’s clock has stopped a short time ago.
The unpossessed evidence in PEDIATRICIAN, by contrast, is evidence that
others properly expect the subject to have had—it was evidence she should
have had. In light of this it would be curious indeed to explain the ‘no
knowledge’ verdict as a matter of (Gettierizing) epistemic luck (rather than a
matter of insufficient evidence).
I submit that the points I am making with PEDIATRICIAN are general.
They hold not only for pediatricians, but for anyone who is a member of a
professional organization whose standards capture good epistemic practice
and on whose members others rely for information (in the expectation that
they conform to those standards): such professionals ought to conform in their
practice to the standards articulated by their professional organization. In
addition, these points hold in cases involving experts as well as people who
play one or another role in a social or work group, as well as in cases involving
participation in other social practices, whether between partners at work,
family members at home, friends, or what-have-you.²² The point is simply
this: when a person who is legitimately expected to have certain evidence (or to
follow a certain protocol in inquiry etc.) fails to do so, where her failing to do
so results in her failing to have relevant counterevidence that she ought to have
had, the person is not entitled to confidence, nor is she in a position from
which to authorize others to believe (with confidence). This is seen not only in
PEDIATRICIAN but also in both of the unpossessed evidence scenarios in
UNOPENED LETTER and NEWSPAPER as well. In sum, the general lesson is
that legitimate social expectations give rise to the prospect of a kind of

(normative and social) defeat. What is interesting about this kind of defeat is
that it can bear against the epistemic standing of a subject’s belief even when
the subject is ignorant of the defeating consideration itself.
2.6.
The time has come to draw the various pieces of the picture into a more
cohesive whole. My general conclusion is that the normativity of knowledge
has more far-reaching implications than is typically assumed. In this paper,
I have tried to capture some of these implications in terms of the two defeater
conditionals themselves; these can be applied to certain cases to help us
discern a normative (social) type of defeat. Failure to recognize this type of
²² An early example of this was given in Gibbons (2006).
36  . 
defeat, I submit, is tantamount to a failure to have in one’s sight a state that has
the sort of normativity we associate with knowledge.
I believe that this picture of the normativity of knowledge is far-reaching.
While we may well be able to characterize a condition in fully naturalistic
terms, such that non-human animals can be in that condition, my strong
suspicion is that such a condition would not answer to the sort of thing we
expect human knowledge to be. The defeat conditionals articulate the basis for
this suspicion. One lesson to be learned, if I am right is that even if (following
Ernie Sosa) we designate such a naturalistic condition as a kind of (mere)
‘animal knowledge,’ there is a great chasm separating such a condition
and what we ought to recognize as distinctly and paradigmatically human
knowledge.²³ The chasm is greater than many appear to suppose, since it
involves moving to a domain rife with the normative social expectations we
have of one another as we move about the world acquiring information in our
characteristically socially dependent ways.²⁴
References
Alston, W. (1993). Epistemic desiderata. Philosophy and Phenomenological

Research, 53(3): 527–51.
Alston, W. (2005). Beyond Justification: Dimensions of Epistemic Evaluation.
Ithaca: Cornell University Press.
Baker-Hytch, M. and Benton, M. (2015). Defeatism defeated. Philosophical

Brown, J. (2008a). Subject-sensitive invariantism and the knowledge norm for
practical reasoning. Noûs, 42: 167–89.
Brown, J. (2008b). Knowledge and practical reason. Philosophy Compass, 3(6):
1135–52.
Brown, J. (2018). Fallibilism: Evidence and Knowledge. Oxford: Oxford University
Press.
Research, 81(1): 185–215.
²³ See Goldberg and Matheson (Forthcoming), where we reflect on this further.

²⁴ I would like to express my gratitude to Jessica Brown for her thoughtful and incisive comments on
an earlier draft of this paper; and to Mona Simion, Jennifer Lackey, and Baron Reed, with whom I have
discussed these and related topics at great length. None of them are responsible for the shortcomings of
this paper!
       37
Feldman, R. (2000). The ethics of belief. Philosophy and Phenomenological

Research, 60(3): 667–95.
Gerken, M. (2014). Same, same but different: The epistemic norms of assertion,
action and practical reasoning. Philosophical Studies, 168(3): 725–44.
Gibbons, J. (2006). Access externalism. Mind, 115(457): 19–39.
Goldberg, S. (2015a). Assertion: On the Philosophical Significance of Assertoric
Speech. Oxford: Oxford University Press.
Goldberg, S. (2015b). Recent work on assertion. American Philosophical
Quarterly, 52(4): 365–80.
Goldberg, S. (2017). Should have known. Synthese, 194(8): 2863–94.
Goldberg, S. (2018). To the Best of Our Knowledge: Social Expectations and
Epistemic Normativity. Oxford: Oxford University Press.
Goldberg, S. and Matheson, J. (Forthcoming). The impossibility of mere animal
knowledge for reflective subjects. Erkenntnis.
Horowitz, S. (2014). Epistemic akrasia. Noûs, 48(4): 718–44.
Horowitz, S. and Sliwa, P. (2015). Respecting all the evidence. Philosophical
Studies, 172(11) 2835–58.
Lackey, J. (1999). Testimonial knowledge and transmission. Philosophical
Lackey, J. (2011). Assertion and isolated second-hand knowledge.” In: J. Brown
and H. Cappelen (eds.) Assertion: New Philosophical Essays. Oxford: Oxford
Lasonen-Aarnio, M. (2010). Unreasonable knowledge. Philosophical Perspectives,
24(1): 1–21.
Lasonen-Aarnio, M.( 2014). Higher-order evidence and the limits of defeat.
Philosophy and Phenomenological Research, 88(2): 314–45.
McKinnon, R. (2013). The supportive reasons norm of assertion. American
Philosophical Quarterly, 50(2): 121–35.
McKinnon, R. (2016). The Norms of Assertion: Truth, Lies, and Warrant. New
York: Palgrave MacMillan.
Moody-Adams, M. (1994). Culture, responsibility, and affected ignorance. Ethics,
104(2): 291–309.
Pollock, J. (1986). Contemporary Theories of Knowledge. Maryland: Rowman and
Littlefield.
Reed, B. (2010). A defense of stable invariantism. Noûs, 44: 224–44.
Schoenfield, M. (2015). A dilemma for calibrationism. Philosophy and
Phenomenological Research, 91: 425–55.
38  . 
Schoenfield, M. (2018). An accuracy-based approach to higher order evidence.

Skipper, M. (2018). Reconciling enkrasia and higher-order defeat. Erkenntnis, 84
(6): 1369–86.
Weiner, M. (2005). Must we know what we say? The Philosophical Review, 114(2):
227–51.
Whiting, D. (2013). Stick to the facts: On the norms of assertion. Erkenntnis, 78
(4): 847–67.
Williamson, T. (2014). Very improbable knowing. Erkenntnis, 79(5): 971–99.
Williamson, T. (2000). Knowledge and Its Limits. Oxford: Oxford University
Press.
Williamson, T. (2002). Knowledge and Its Limits. Oxford: Oxford University
Press.
Ye, R. (2018). Higher-order defeat and intellectual responsibility. Synthese, 1–21.
3
The Structure of Defeat
Pollock’s Evidentialism, Lackey’s Framework,
and Prospects for Reliabilism
Peter J. Graham and Jack C. Lyons
Epistemic justification is very often defeasible: the addition of new information

can weaken or destroy what was a perfectly adequate justification. Normally, if
an object looks red to you, then you’re justified in believing it’s red. Often such
justification is good enough to yield knowledge. But if it looks red and you find
out it’s illuminated with red light, then—absent further information—you’re
no longer justified in believing it’s red. If you persist in believing so, you don’t
know it. You might have prima facie justification, but you lack ultima facie
justification: your initial justification is defeated.
Most discussions of defeat occur within either an evidentialist or a respon-
sibilist framework. This is problematic. First, it is far from obvious what the
resulting theories of defeat have to do with each other and whether they can
even be rendered compatible. Second, both approaches seem to fit poorly with

a reliabilist epistemology—a fact that is of much concern to us, since we
endorse broadly reliabilist views.
In this chapter we first discuss evidentialist and responsibilist accounts of
defeat before offering a reliabilist view. In Section 3.1, we discuss Pollock’s
seminal evidentialist views on reasons and justification, defeat and defeaters.
We then discuss Jennifer Lackey’s responsibilist-inspired account in Sections
3.2–3.4, focusing on her influential claim that all defeaters are either “doxastic”
or “normative.”¹ Despite its influence, we argue her distinction is entirely
unfounded. “Doxastic” and “normative” defeaters in Lackey’s sense simply
do not exist. We then turn to the reliabilist treatment of defeat in Section 3.5.
¹ Green (n.d.); Grundmann (2011); Palermos (2011); Matheson (2011); Carter (2015), (2018);
Carter and Navarro (2017); Goldberg (2018); Goldberg and Matheson (2020); Pritchard (2018).
3
40  .    . 
We think defeat fits naturally into a reliabilist epistemology, in a way that is

neither responsibilist nor evidentialist.²
3.1.
Pollock provided the seminal treatment of defeat and defeaters.³ Pollock first
defines (epistemic) reasons (to believe) in terms of justified belief.
A state M of a person S is a reason for S to believe P if and only if it is logically

possible for S to become justified in believing P by believing it on the basis of
being in state M.⁴
Justified belief is then Pollock’s explanatory primitive. Other evidentialists take

the notion of evidence or reasons as basic. Evidence is their explanatory
primitive.
What do reasons have to do with defeat? The technical terms “epistemic
defeat” and “defeater” then get their sense from the defeasibility of reasons and
justification. Suppose Carol perceptually experiences an object as red and as
having a particular shape. This is a reason to believe that is an apple.⁵ But then
she picks it up and feels its texture and weight. It feels more like a polished
rock than an apple. Perhaps it is an objêt d’art made of stone, cleverly made to
look just like an apple. Carol’s new experience is a reason to believe both that is
a stone and that is not an apple. Her new experience defeats her old experience
as a reason to believe that is an apple. Her old experience combined with her
new experience is not a reason to believe that is an apple, but instead a reason
to believe it is a stone, cleverly shaped to look like an apple. Given her new
experience, the rational thing for Carol is to change her mind, if she hasn’t
already.
Given one set of relevant reasons, you can have a good reason to believe
p. Given another, more inclusive set, you might no longer have a good reason
² Please note that our topic here is justificatory defeat. The term ‘defeater’ is sometimes used in
discussions of “defeasibility” responses to the Gettier problem, to refer to an unknown fact that would
defeat justification if known (or justifiedly believed). These “knowledge defeaters”—also sometimes
called ‘factual’ defeaters—are not our concern here.
³ Pollock (1967), (1968), (1970), (1971), (1979), (1986), (1987), (1994), (2001), (2006).
⁴ Pollock and Cruz (1999): 195.
⁵ We ignore here debates about the contents of perceptual states and perceptual beliefs, whether they
involve indexicals, singular terms, or existential generalizations. Our use of the demonstrative is for
convenience’s sake and isn’t intended to take a stand on this issue.
    41
to believe p. Some reasons to believe p are defeated by other reasons. Those

other reasons are defeaters. Those defeaters, in turn, might be defeated by
further reasons. Those are defeater-defeaters. A reason that can be defeated is a
defeasible reason. What you are ultimately justified in believing thus depends
on your total evidence, not just part of it.⁶
Pollock and Cruz (1999) formulate the notion of a defeater as follows:
If M is a reason for S to believe p, a state M* is a defeater for this reason iff it

is logically possible for S to be in the combined state consisting of being in
both state M and state M* at the same time, and this combined state is not a
reason for S to believe p.⁷
M and M* range over the kinds of mental states that can serve as reasons for
beliefs. Defeaters are then just reasons functioning in a destructive rather than
constructive mode.⁸
For Pollock, all defeaters are either rebutters or undercutters:
If M is a defeasible reason for S to believe p, M* is a rebutting defeater for

this reason iff M* is a defeater (for M as a reason for S to believe p) and M* is
a reason for S to believe ~p.⁹
If M is a defeasible reason for S to believe p, M* is an undercutting defeater
for this reason iff M* is a defeater (for M as a reason for S to believe p) and
M* is a reason for S to doubt or deny that he or she would not be in state
M unless p were true.¹⁰
For Pollock, beliefs—doxastic states—are not the only possible defeaters, for
he denies “doxasticism” in epistemology—the view that only beliefs can serve
as reasons. M and M* range over a variety of mental state types, including
perceptual experiences, sensations, feelings, moods, and memories, as well as
doxastic states. Pollock’s view is an evidentialist one: the justificational status
of a belief is determined by the evidence—the reasons—for and against
believing. Pollock’s evidentialism entails that you can’t have defeat without
⁶ Authors often describe the phenomenon of defeat in temporal terms. First you have a good reason
to believe that p. Next, you acquire new information that defeats that reason. Many of our examples
proceed just like this. But nothing about defeat requires this temporal ordering.
⁷ Pollock and Cruz (1999): 195.
⁸ Similar formulations occur in Chisholm (1977): 72–3, (1989): 53.
⁹ Pollock and Cruz (1999): 196. ¹⁰ Pollock and Cruz (1999): 196–7.
42  .    . 
reasons/evidence; the only things that can produce defeat are things that can
produce justification.
Reasons aren’t restricted to beliefs for Pollock, but they are restricted to things
that you might base a belief on. Prima facie doxastic justification is limited to
things that you have based your belief on. Ultima facie justification—even
ultima facie doxastic justification—is importantly different, for it also depends
on other reasons you possess but which haven’t positively or negatively causally
influenced that belief. Pollock is surprisingly inexplicit about ultima facie
justification, but it is fairly clear that he thinks that a belief is ultima facie
justified just in case it is prima facie justified, and the agent doesn’t possess any
ultimately undefeated defeaters for it. To possess a (for example, rebutting)
defeater for the belief that p is to have a reason to believe ~p, i.e., to be prima
facie propositionally justified in believing ~p. This makes ultima facie (doxastic)
justification depend in part on propositional justification in a way that prima
facie (doxastic) justification doesn’t. Thus, while having merely propositional
justification can’t increase your doxastic justification, it can diminish it.
Though rightly influential, there are a few idiosyncratic features of his
formulations worth pointing out.
First, in Pollock’s classic formulations (though this is rectified in Pollock
(2001)), the definition of defeat doesn’t admit of degrees and thus doesn’t
allow for partial defeat: cases where a reason reduces justification without
eliminating it. Second, this problem is exacerbated by the oddly conjunctive
framing of the definitions for rebutting and undercutting defeat: a rebutting
defeater is something that’s both a defeater and a reason to believe the

negation of the target proposition. If we weaken the definition of defeat to
allow for partial defeat, then the first conjunct will be entailed by the second
and is thus unnecessary. If we don’t allow for partial defeat, then the conjunc-
tive definitions of undercutting and rebutting defeat are too narrow.
Third, only reasons/evidence can be defeaters. This is very much a deliberate
choice on Pollock’s part. Later, we will see it come under attack from two quite
different quarters, responsibilist and reliabilist. It is worth highlighting for now.
Fourth, somewhat surprisingly, only reasons can be defeated. Of course, this
has to be true in the case of undercutting defeat, but it’s not obvious why
rebutting defeat attacks reasons for the belief that p rather than the belief ’s
prima facie justification. It seems more natural to think of rebutting defeaters
as working foremostly on prima facie justification and only derivatively on
reasons.
For these reasons, it’s worthwhile to simplify and reformulate the basic
Pollockian outlook. Here’s a theory that’s squarely Pollockian in spirit:
    43
d is a rebutting defeater for S’s belief that p iff d is a reason for S to believe
not-p.
d is an undercutting defeater for S’s belief that p iff d is a reason for S to
believe that her reasons for believing p are inadequate.
Since we haven’t defined undercutting and rebutting defeat in terms of the

more general notion, we can do it the other way around:
d is a defeater for S’s belief that p iff d is a rebutting or an undercutting

defeater for S’s belief that p.
Finally:
S’s prima facie justification for believing p is defeated iff S possesses a

defeater for that justification, and that defeater is not ultimately defeated.
A prima facie justified belief is ultima facie justified iff it’s not defeated.
These definitions make clear the evidentialist presumption that defeat is
simply a matter of having reasons/evidence, and that ultima facie justification
is thereby a matter of what the balance of the evidence—what the whole of the
evidence—supports. They allow defeat to come in degrees, like reasons more
generally. On the evidentialist view, the basis for defeat is the same as the basis
for justification: the subject’s evidence.
3.2.
Jennifer Lackey has argued that Pollock’s distinction between rebutting and
undercutting defeaters requires supplementation with another distinction
between “doxastic” and “normative” defeaters.
According to Lackey, a doxastic defeater is “ . . . a doubt or belief that is had
by S that indicates that S’s belief that p is either false or unreliably formed or
sustained.”¹¹ By ‘indicates’ she means ‘obviously entails.’ In particular:
¹¹ Lackey (2006b): 438. See also Lackey (1999): 483, (2003): 707, (2005a): 164, (2005b): 638, (2008):
44, 253, (2011a): 317. Lackey uses ‘psychological defeater’ and ‘doxastic defeater’ interchangeably. The
former serves to emphasize that in addition to beliefs, some doubts can function the same way as
doxastic defeaters, but since they are not narrowly classified as beliefs, we need a broader term. Hence
‘psychological.’ Nothing will hang on this point in our chapter. What about perceptual experiences, and
other non-doxastic states? Though in two places she included experiences along with beliefs and doubts
in her list of psychological defeaters (Lackey (2006a): 87, (2006c): 4), otherwise experiences do not
44  .    . 
• A doxastic rebutting defeater for a subject’s belief that p is a further belief

of the subject that q, where that belief, given its content and the contents
of other beliefs held by the subject, very obviously entails ~p. A doxastic
rebutting defeater creates an incoherent set of first-order beliefs. For
example: S believes {p, q, if q then ~p}.
• A doxastic undercutting defeater for a subject’s belief that p is the
subject’s further belief that my grounds do not make it likely that p (or
something functionally equivalent, e.g., the source of my belief that p is
not reliable). A doxastic undercutting defeater also creates an incoherent
set of beliefs: S believes {p, my grounds do not make it likely that p}. The
incoherence here does not lie solely in the contents, but in the subject’s
attitudes towards those contents, like Moore-paradoxical thoughts
generally.
Must doxastic defeaters themselves be justified to do their defeating work?

Not at all. Lackey repeatedly emphasizes that doxastic defeaters need not be
true, justified, or rationally held to do their defeating work.¹² Doxastic defea-
ters are justified or unjustified beliefs that indicate that a subject’s belief that p
is either false or unreliably formed or sustained.
That’s her account of doxastic defeaters. There are also instances of defeat
where the agent doesn’t have a doxastic defeater, hence the need for another
category of defeaters: normative defeaters. There are two very different uses of
the phrase ‘normative defeater’ in the literature, Lackey’s use and another we
will explain shortly. We begin with Lackey’s use.

According to Lackey, a normative defeater involves three parts. A proposition
q is a normative defeater for a subject’s belief that p in Lackey’s sense iff:
(1) S has evidence—in the form of experiences, beliefs, etc.—that justifies

believing q for S.
(2) q would be a doxastic defeater for S’s belief p if S were to believe q.
(3) S ought to believe q.¹³
appear in her formulations (e.g., Lackey (1999), (2003), (2005a), (2005b), (2008): 44, 157, 204, (2011a):
317, (2014a): 292, (2014b): 75). The more common formulation states her considered view: psycho-
logical (doxastic) defeaters for Lackey are necessarily doxastic states (propositional attitudes) like
beliefs or doubts, never non-doxastic experiences. Non-doxastic experiences, instead, ground norma-
tive defeaters, as we will see. That’s how experiences enter her “ontology of defeat.”
¹² Lackey (2005b): 647, cf. Lackey (2003): 707, (2005a): 164, (2006a): 87, (2008): 44–5, (2011a): 317.
¹³ Lackey (1999): 475, (2003): 707, (2005b): 638, (2005a); 164, (2006a): 87–8, (2006c): 4, (2008): 45,
157, (2011a): 317.
    45
Condition (1) explains why evidence grounds normative defeaters. Condition

(2) explains why they defeat the subject’s belief that p; it explains why the evidence
is counterevidence for it explains why normative defeaters are “counterparts” of
doxastic defeaters.¹⁴ Condition (3) explains why Lackey calls them “normative”
defeaters, for epistemic duty requires the subject to believe the normative
defeater. (3) Can look redundant given (1), but it’s actually much stronger. To
say that you’re justified in believing p is to say that you’re epistemically permitted to
believe p. To say that you ought to do something, however, is to say that you’re doing
something wrong if you don’t do it; you have a duty or obligation to do it. Clearly
one could have epistemic permission to believe p without having an epistemic
obligation to believe it. We will explain why Lackey includes (3) in Section 3.4.
What is the other sense of the phrase ‘normative defeater’? On the other use,
a normative defeater is:
1. A worldly piece of evidence (like a letter hidden under the doormat).

2. That the subject is not aware of and so does not “possess” (the subject
has no idea about the letter).
3. That if the subject were aware of, then the subject would possess a
defeater for one of her beliefs (if she read the letter, she’d now have a
defeater for a belief).
4. That the subject should be aware of, that the subject should possess (for
some reason she should check under her doormat and read the letter).
A normative defeater for a subject’s belief in this sense is a potential defeater

that the subject does not actually possess but should.¹⁵ This is not Lackey’s
sense. Lackey’s idea of normative defeat arises from beliefs the subject should
have but does not, given evidence the subject already possesses.¹⁶ This other
use of ‘normative defeaters’ is not what she has in mind.¹⁷
¹⁴ Lackey (2006a): 8.
¹⁵ For this use of the phrase ‘normative defeater,’ see Lyons (2009); Grundmann (2011); Goldberg
(2018); Pritchard (2018).
¹⁶ Lackey addresses evidence one does not possess but should in only two of her articles, and there
she is using ‘normative defeater’ only to address issues of knowledge, not of justification. Lackey
(2005b), (2014a).
¹⁷ A passage in a recent paper (Lackey (2018): 161) might lead one to think she now has this other
sense in mind in addition to her sense as defined in the text. She gives an example where she says a
subject’s belief is defeated, for the subject lacks evidence that she should have had. She then favorably
cites (Goldberg (2018), who had this other sense in mind. But the example makes it clear that this other
sense is not what she has in mind. For, in her example, the “evidence that the subject should have had”
simply consists in a justified belief that the subject should have formed, given other evidence (in this
case, another justified belief) that the subject already possesses. So, the evidence the subject should have
had is just the proposition that the subject should believe, given her other evidence she already
possesses.
46  .    . 
An example from Lackey of undercutting normative defeater should help

make her use of the phrase ‘normative defeater’ clear:
Alice is incorrectly told by an otherwise reliable optometrist that her vision is

nearly completely unreliable, yet she refuses to accept his diagnosis, without
having any rational basis for doing so. So, one might say, even though the
optometrist’s report is false, Alice should accept his diagnosis, given all of the
evidence that she has available to her, and thus she has a normative defeater
for her visual beliefs.¹⁸ Nevertheless, as she is walking out of the doctor’s
office, [Alice] sees a car accident on Michigan Avenue, forms the corres-
ponding true belief that there was such an accident. She rejects the report
[that she has unreliable vision] for no rational reason.¹⁹
Alice has good evidence from the doctor’s diagnosis that her vision is unreli-
able: she knows he told her that her vision is unreliable, and she believes he is a
reliable informant on matters optometrical. Given this evidence she possesses,
believing the proposition my vision is unreliable is justified for her. Given
Lackey’s definition of doxastic defeaters, this proposition would be a doxastic
defeater if she were to believe it. Furthermore, Lackey insists, she ought to
believe it. That’s why it is a normative defeater. The proposition my vision is
unreliable is then Alice’s normative defeater for her belief there was a car
accident.²⁰
Lackey’s example involves a belief as the evidential basis for the normative
defeater. This belief must be justified, given her definition of normative defea-
ters, for unjustified beliefs are not a part of one’s evidence that justifies further
beliefs. Non-doxastic experiences can serve as the evidential bases for normative
defeaters as well, for experiences too are a part of one’s total evidence.
How does Lackey understand the basis of defeat, for both doxastic and norma-
tive defeaters? “The underlying thought” behind both kinds of defeaters, she says,
. . . is that certain kinds of counterbeliefs [doxastic defeaters] and counter-

evidence [the grounds for normative defeaters] contribute epistemically
unacceptable irrationality to doxastic systems and, accordingly, that justifi-
cation and knowledge can be defeated or undermined by their presence.²¹
¹⁸ Lackey (1999): 487. ¹⁹ Lackey (2008): 63. ²⁰ Lackey (2008): 64.

²¹ Lackey (2005a): 164, (2006a): 87, (2008): 45, (2014a): 292, (2014b): 75.
    47
On Lackey’s view, irrationality—the failure to conform to the requirements of

rationality—is then the source of defeat. We will elaborate on this idea as we
critically engage Lackey’s framework. Both categories, we shall argue, are problem-
atic. There is no such thing as a doxastic or a normative defeater in Lackey’s sense.²²
3.3.
We begin with doxastic defeaters.

Although Lackey’s examples of doxastic defeaters throughout her corpus
involve justified doxastic defeaters, central to Lackey’s concept of doxastic
defeaters is the insistence that they can be unjustified, or even irrational.²³ We
will argue that that is a serious problem.
To test the intuitive plausibility of this view, we should imagine cases
involving unjustified doxastic defeaters. Here’s a recipe:
(1) S believes that p on very strong total evidence.

(2) S comes to believe q on absolutely no evidence, in a completely
unreliable way. If any belief is unjustified, this one is. Let the belief
occur through wishful thinking, emotional attachment, a brain tumor,
mental illness, radiation, etc.
(3) S’s background beliefs include if q then ~p.
(4) The subject now has an incoherent set of beliefs: S believes {p, q, if q
then ~p}.
(5) By Lackey’s definition, S’s belief that q is then a doxastic defeater for S’s
belief that p. Absent a defeater-defeater for q, then on Lackey’s view S’s
belief that p is no longer justified.²⁴ For Lackey, S should stop believing p.
(6) But that is counter-intuitive, given that S has very strong evidence for p,
and no evidence against it. (The only consideration that speaks against
believing p is q, but she has, by stipulation, no evidence for q and shouldn’t
believe it.)
²² Lackey also frequently says, in addition to doxastic and normative defeaters, that there is a third
category of so-called “factual” defeaters. Following Lackey, we’ve seen some philosophers to even say of
defeaters that there is a “traditional” three-way taxonomy of “psychological” (doxastic), “normative,”
and “factual” defeaters. Firstly, this is not the traditional view. Secondly, we deny the existence of so-
called “factual” defeaters, for we reject defeasibility analyses of knowledge.
²³ Lackey (2005b): 647, cf. Lackey (2003): 707, (2005a): 164, (2006a): 87, (2008): 44–5, (2011a): 317.
²⁴ Lackey must allow the possibility of cases where such defeater-defeaters are lacking, or otherwise
the possibility of unjustified defeaters is an idle wheel, as they would always be defeated by a defeater-
defeater.
48  .    . 
We think that in such cases the subject should not believe q but should
continue believing p. p is based on strong evidence, formed in a reliable way. q
is neither. How could a belief that is well-justified on strong evidence have the
force of its evidence defeated by an unjustified belief, based on no evidence
at all?
A number of other authors have endorsed—sometimes offhandedly, some-
times deliberately—the idea of unjustified defeaters, i.e., the idea that an
unjustified belief can nevertheless defeat the justification of a prima facie
justified belief.²⁵ But Lackey’s discussion of defeat disguises this very contro-
versial claim as an innocuous terminological point. It’s nothing of the sort.
Notice first that this view of doxastic defeaters is incompatible with any
roughly evidentialist view. On an evidentialist view, nothing can be relevant
to justification but evidence, and unjustified beliefs aren’t evidence.²⁶ On the
evidentialist view, defeat occurs when one’s partial evidence supports believing
p but one’s total evidence (or even just a more inclusive body of partial
evidence) does not. A defeater is then just another piece of evidence. On the
Pollockian theory of defeat, nothing can provide defeat that can’t also provide
justification, but unjustified beliefs can’t provide justification. Hence, an
unjustified belief cannot be a defeater.
Lackey needs an argument that there might be unjustified doxastic defea-
ters. Without such an argument, her entire category of doxastic defeaters is
unmotivated, for the evidentialist already recognizes that justified beliefs can
play the role of defeaters, either rebutting or undercutting. Without an argu-
ment, there’s no need for Lackey’s broader category of doxastic defeaters.

Does she have an argument? She does not articulate one. But here is a
suggestive passage:
[It] is important to notice that what makes an undefeated doxastic defeater

epistemically problematic is that it is held in conjunction with another belief.
The defeater itself need not be true, justifiedly believed, or rationally believed
to have the power to defeat other beliefs precisely because it need not be true,
justifiedly believed, or rationally believed to render it irrational to hold
certain other beliefs.²⁷
²⁵ For example, Goldman (1986) and Bergman (2006), among others.

²⁶ It is therefore surprising that an evidentialist like Matheson (2011) cites Lackey approvingly in
connection with doxastic defeat.
²⁷ Lackey (2005b): 647, emphasis added. See also Lackey (2003): 707, (2005a): 164, (2006a): 87,
(2008): 44–5, (2011a): 317, (2014b): 75.
    49
Doxastic defeaters thereby contribute “epistemically unacceptable irrationality

to doxastic systems.”²⁸ That’s the idea we flagged at the end of the last section.
She thus seems to be thinking something like the following:
(1) Rationality requires coherence amongst one’s propositional attitudes

(assumption).
(2) When one holds a particular belief that p despite also holding a
doxastic defeater for that belief—either a rebutter or undercutter,
whether justified or unjustified—then one holds an incoherent set of
propositional attitudes (definition of doxastic defeater).
(3) Hence, if one holds a belief that p despite a doxastic defeater (justified
or unjustified), then one is being epistemically irrational (from 1, 2).
(4) If one is being irrational partly in virtue of holding a belief that p, then
that belief that p is not justified (assumption).
(5) Hence, if one holds a belief that p despite a doxastic defeater, whether
justified or unjustified, then that belief that p is not justified (from 3, 4).
(6) If a belief that p was justified, but is no longer justified, due to the
possession of a defeater, then its justification is defeated (assumption).
(7) Hence, if one holds a belief that p despite a doxastic defeater, whether
justified or unjustified, then its justification is defeated (from 5, 6).
This argument looks valid. And it makes plain a connection between

doxastic defeaters, irrationality, and defeat, while allowing that doxastic defea-
ters may be unjustified.

Justified or unjustified, doxastic defeaters induce incoherence; incoherence
induces irrationality; irrationality induces defeat. That’s why unjustified dox-
astic defeaters defeat.
But is the argument sound? No. (4) Is false. Notice that (3) says that if a
subject holds a belief that p, then acquires a belief that q, where q is a doxastic
defeater (possibly unjustified), but the subject continues to believe that p, then
the subject has entered an irrational state, the state of holding an incoherent
set of propositional attitudes. But (4) then condemns one of the beliefs in the
set in particular as unjustified—the belief that p that the subject started with.
And if rationality requires that we not believe propositions that are not
justified for us, then rationality requires that the subject give up his belief
that p. That’s the route the subject is required to take to exit his current state of
irrationality. So, (4) says if the subject starts with a justified belief that p, then
²⁸ Lackey (2005b): 639, 647, cf. (2005a): 164, (2006c): 4, (2008): 253, (2014b): 75.
50  .    . 
acquires an unjustified belief that q, where that belief that q combined with the
belief that p then puts the subject into an irrational overall state, then the
subject is required to give up on the justified belief that p to exit the irrational
state.
This, we think, is obviously false. That is not what rationality—even coher-
ence rationality—requires of the subject. What coherence rationality requires
is that the subject give up at least one of the beliefs in the incoherent set, but
not any one belief in particular. The subject could exit irrationality by giving
up q, or some other auxiliary belief. Rationality requires some accommodation.
It does not require giving up the original belief in particular. That’s just the
lesson of the Quine–Duhem thesis. It seems perverse to require that the subject
abandon her justified belief that p when she acquires an unjustified belief that
q, as if two wrongs made a right. One might rather think—and this is just the
point we’ve made already—that if anything, rationality would have required
the subject not to form the unjustified belief that q in the first place. Now that
she has, the best option is to retract it.
Not only is this the intuitively obvious response, and the response that
evidentialism requires, it’s the response that a coherence/responsibilist theory
would endorse anyway. The coherence theory never said to get rid of every
belief that’s implicated in a local incoherence; it says to get rid of (or modify)
the less-well-connected beliefs to maximize coherence among the better-
connected beliefs. We want a large, comprehensive, and coherent belief set,
and the way to get that isn’t to just toss out everything that’s involved in any
kind of conflict, but to selectively pin the blame for the conflict on whichever
beliefs are less coherent with the rest of the set, independent of the conflict in
question. The proponent of unjustified defeaters would have us do the oppos-
ite of what the coherence theory recommends: they would have us abandon
the well-connected beliefs to accommodate the poorly connected ones.
There’s another kind of argument one sometimes sees in defense of unjus-
tified defeaters. The proponent will say something along the lines of “given
that she believes her senses are unreliable, S should stop believing that p.”²⁹
This would support a modus ponens argument for the desired conclusion (that
S should stop believing p) if ‘given that’ functioned like ‘if ’: i.e., if the starting
point were “if S believes her senses are unreliable, S should stop believing that
p.” But that’s far from obvious; in fact, it’s the very controversy we’re trying to
resolve. More likely, the plausibility of the ‘given that’ claim derives from its
²⁹ For example, see Bergmann (2006): 165–7. For some discussion of the issue, see Frances (2014):
79–84.
    51
introduction of a scope ambiguity between something like ‘if Bel(p) then

Ought(~Bel(q))’ and ‘Ought(if Bel(p) then ~Bel(q)).’³⁰ The argument requires
the former disambiguation, while it’s only the latter that’s either intuitively
obvious or supported by coherence considerations.
This kind of ‘given that’ talk is quite tempting. It’s natural to say that given
that you believe that this tank of piranhas is a hologram, you ought to be
willing to put your hand in it. It doesn’t follow that it’s rational for you to put
your hand in the tank—just because you do (irrationally) believe it’s a holo-
gram. It’s only conditionally rational, rational relative to a belief that turns out
to be itself irrational. Relative to that crazy belief, that action wouldn’t be crazy.
But since the belief in fact is crazy, so the action is too. The action is rational
relative to the rationality of the belief, not relative to its mere existence. You
don’t get to run a modus ponens-type inference in these cases.
Finally, Lackey frequently makes reference to undefeated doxastic defeaters:
“Millicent acquires an undefeated defeater for her visual powers via another
person’s testimony, and hence fails to have the knowledge in question . . . .
Millicent’s hearers [are justified] but Millicent is not, because she has the
undefeated defeater.”³¹ Such repeated qualification suggests that even Lackey
herself thinks that a defeated doxastic defeater doesn’t kill the prima facie
justification of the belief it’s a defeater for. But why should that be? If defeaters
don’t need to be justified, they shouldn’t need to be undefeated to do their
defeating work. She thinks they can be irrational and still defeat, after all.
Could it really matter whether a defeater is unjustified because defeated, or
unjustified because the agent pulled it out of a hat? This suggests a tacit
recognition on Lackey’s part that unjustified beliefs can’t serve as defeaters
after all.
Lackey’s right that incoherence is a problem, and also right that so-called
doxastic defeaters, as she has defined them, introduce incoherence. And we
agree that a subject is required to make some adjustments. But we just don’t
think incoherence as such is the same as epistemic defeat. Defeaters defeat
propositional justification by introducing counterevidence into the subject’s
total evidence, not by merely introducing incoherence at the level of the
subject’s propositional attitudes.
³⁰ See Broome (2013) for similar comments. There is a lively and expanding discussion, partly
growing out of Broome’s work, on questions surrounding the so-called “enkratic” principle, and cases
where a subject’s total evidence might justify the first-order belief that p and the higher-order belief that
one’s total evidence does not justify the belief that p. See for example Worsnip (2018) and Lasonen-
Arnio (2020), among others.
³¹ See Lackey (1999): 485; emphases added; see also the most recent block quotation.
52  .    . 
This means that Lackey is not only wrong about unjustified doxastic
defeaters, but also about why justified doxastic defeaters defeat. They don’t
defeat because they introduce incoherence; they defeat because they negatively
change the subject’s total evidence.
Lackey’s talk about doxastic defeaters is not a harmless and neutral bit of
taxonomy. The idea of unjustified defeaters is flatly incompatible with evi-
dentialism, which holds that only reasons or evidence can provide or play the
role of defeaters; unjustified beliefs cannot play that role. The idea that
“doxastic” defeaters (whether justified or unjustified) defeat in virtue of inco-
herence simply misunderstands the nature of defeat. There is no rationale to
introduce Lackey’s category of doxastic defeaters.
3.4.
We now turn to Lackey’s category of normative defeaters. Recall that for

Lackey a normative defeater for a subject’s belief that p is a proposition that
q where (1) S has evidence that justifies believing q for S; (2) q would be the
content of a doxastic defeater for S’s belief that p if S were to believe q; and (3)
S ought to believe q. Just as we see no rationale for Lackey’s category of
doxastic defeaters, we also see no rationale for her category of normative
defeaters, and not only because normative defeaters are understood in terms
of doxastic defeaters.
Before turning to criticism, there’s an important difference between her two

categories. Lackey’s concept of doxastic defeaters is supposed to describe a
certain phenomenon—defeat by belief, whether justified or not—that simply
doesn’t exist. The phenomenon that normative defeaters are supposed to
describe, on the other hand—defeat that results from ignored evidence—is
certainly a real thing. It’s Lackey’s explication of this phenomenon, in terms of
duties to adopt doxastic defeaters, that is flawed, as we’ll argue.
There are three problems with Lackey’s theory of normative defeat.
First, it involves unnecessary complications to treat the cases of defeat it
is designed to cover, while simultaneously relying on dubious category of
doxastic defeaters. Normative defeaters, according to Lackey, are grounded
in counterevidence possessed by the subject. But they do not consist in that
counterevidence. Instead, Lackey makes that counterevidence relevant
only through the circuitous route of doxastic defeaters that one should have,
beliefs that you would have had, had you not been flouting your epistemic
duties.
    53
Why is this complication unnecessary? Look at things from the evidential-

ist’s perspective: the counterevidence alone already defeats the subject’s prop-
ositional justification. If the subject maintains the original belief in the face of
such counterevidence, then the subject’s belief is not doxastically justified, for
it is no longer propositionally justified. That’s all there is to it. Univocal
evidence supports believing p; equivocal evidence (evidence that contains
sufficient counterevidence) does not. There’s no need to postulate the obliga-
tion to believe the proposition that the counterevidence supports.
Recall Alice. Alice sees a car crash. Ordinarily that is sufficient evidence to
justifiably believe there was a car crash. But Alice was also told by a reliable
optometrist that her vision has been significantly impaired. The evidence from
the testimony of her senses and the evidence from the testimony of her
optometrist combined fail to justify believing there was a car crash. The
combined evidence is equivocal and thus does not provide sufficient overall
reason. Thus, when she does so believe, her belief is not doxastically justified.
The testimony of her senses is defeated by the testimony from her optometrist.
That’s it. There is no need to introduce an obligation to believe the proposition
the evidence supports to explain why the evidence defeats.
Second, it’s not very plausible that defeat will typically involve obligations to
believe anything at all, let alone to believe doxastic defeaters. Epistemic reasons
typically produce permission, not obligation. And yet, any time an agent has
permission to believe ~p, she thereby has a defeater for believing p. Lackey’s reliance
on propositions you ought to believe will leave us with far too few defeaters.³²
Third, there’s no chance that standard cases of defeat will even involve
permissions to believe doxastic defeaters, let alone obligations. In other words,
Lackey’s view makes counterintuitive predictions about standard, straightfor-
ward cases of defeat. Suppose David is indoors and Edwina calls from outside
and says it is raining. David believes her and is justified in this belief. But then
Fiona calls from outside and tells David that it is not raining. David has no
reason to trust Edwina over Fiona or vice versa, but he ignores Fiona’s
testimony and continues to believe it is raining. The evidentialist would say
that David’s belief that it’s raining is no longer propositionally justified, now
that his evidence is equivocal and thus isn’t doxastically justified. Edwina’s
³² Could Lackey get by with mere permissions, redefining normative defeaters as propositions you
are permitted to believe (rather than ought to believe), which would be doxastic defeaters if you did? We
doubt it. While there’s something plausible about the transfer of obligation (if you’re obligated to A,
and doing A would obligate you to B, then you’re obligated to B) this idea loses all plausibility when one
of those obligations is replaced by a permission (if you’re permitted to A and doing A would obligate
you to B, you’re obligated to B).
54  .    . 
testimony gives him a reason to believe it’s raining, but Fiona’s testimony gives
David a rebutting defeater.
What would Lackey say about this case? David has a rebutting defeater, but
he doesn’t believe a rebutting defeater, so on Lackey’s view he must have a
normative rebutting defeater. But this means that he ought to believe a
rebutting doxastic defeater for the belief that it’s raining. That is, he ought to
believe that it’s not raining. But clearly this is wrong. In the face of such
conflicting testimony, David shouldn’t believe anything about the weather. Not
only is he not obligated to believe Fiona’s testimony, he’s not even permitted to
believe it in the circumstances. This kind of case is easily and naturally handled
by a Pollockian theory of defeat, while Lackey’s theory fails pretty decisively.
Why, then, did Lackey ever think there had to be normative defeaters? Her
idea, as we’ve already noted for doxastic defeaters, seems to be that they follow
from the requirements of rationality themselves. If a subject holds a belief
despite a normative defeater, the subject is being irrational and so their belief is
not justified. Here’s a relevant passage:
. . . a subject who holds a belief in the face of a normative defeater is being

epistemically irrational and/or irresponsible. For instance, in discussing a classic
case of normative defeat [the case of Samantha, who believes the President is in
NYC but has massive counterevidence that she ignores that he is in Washington
DC], Laurence BonJour says that such a subject is “ . . . being thoroughly
irrational and irresponsible” in disregarding the evidence in question.³³
BonJour goes on to say that the subject’s “irrationality and irresponsibility

prevent[s] [Samantha’s] belief from being epistemically justified.” Along with
BonJour, Lackey also approvingly cites similar passages in Michael Williams
(1999: 22–3). Here is a relevant passage from Williams:
[We] can ask whether, in forming a certain belief, I have negligently ignored
important counterevidence . . . . beliefs reached [this way] would be irre-
sponsibly held and would not count [as justified for] epistemic responsibility
is the primary sense of justification . . . an essential component of rationality.
So-called normative defeat must then occur when a subject ignores or disre-
gards relevant counterevidence they possess, for the subject is being irrational
and/or irresponsible in so doing.
³³ Lackey (2005a): 171–2. Italics in the original. Bold added for emphasis. Lackey cites BonJour
(1985): 39. See also Lackey (2005a): 178 for another example.
    55
If the subject is being irrational, then the subject must be violating a

requirement of rationality. That’s the “underlying thought,” after all, at least
according to Lackey, when a subject’s belief is defeated. What then is the
requirement of rationality that Lackey must have in mind? Clearly, it must be
the requirement not to ignore one’s counterevidence.
With this requirement in mind, here is an argument for normative defea-
ters, in Lackey’s sense:
(1) A subject’s belief that p is doxastically justified only if it is not irration-

ally held. Rationality is necessary for justification (assumption).
(2) A subject’s belief that p is irrationally held if the subject ignores her
counterevidence; rationality requires acknowledging counterevidence
(requirement of rationality).
(3) A subject ignores her counterevidence unless she believes the proposi-
tions that the relevant body of counterevidence supports. (An account
of what it is to acknowledge counterevidence).
(4) Hence, if a subject who believes p does not believe the propositions that
her body of counterevidence, relevant to p, supports, then her belief
that p is irrationally held (from 2, 3).
(5) Hence, if a subject who believes p does not believe the propositions that
her relevant body of counterevidence supports, then her belief is not
doxastically justified (from 1, 4).
(6) Hence, if a subject believes p in the face of normative defeaters (as
defined by Lackey), then her belief is irrationality held (from 4) and so

not justified (from 5).
The first three premises seem to validly entail the conclusions, and so it
looks like we have an argument that makes explicit connections between
rationality, justification, defeat, and so-called normative defeaters. We’ve
thereby made sense of Lackey’s view that when a subject “ignores” or “disre-
gards” counterevidence, that entails “epistemically unacceptable irrationality”
that defeats the subject’s justification for her belief. So, if we are looking for an
argument for why there must be normative defeaters, we have found one.
The argument, however, does not work. Without taking a stand on the first
or second premise, we think it should be clear to everyone that the third
premise is false. One way to regard or acknowledge counterevidence is to
believe what it is evidence for. That is certainly true. But it is not the only way
to regard or acknowledge counterevidence. When subjects like Alice, Carol,
and David have counterevidence, that evidence defeats their prima facie
56  .    . 
justification for their belief that p, so that their total evidence no longer
propositionally justifies believing that p. How might they acknowledge or
show due regard to their counterevidence? They might do it—and this is
what the evidentialist recommends—by giving up their belief that p. And
that is because their belief that p is no longer propositionally justified on
their total evidence. Alice could and should acknowledge the force of her
counterevidence by not believing there was a car crash. Carol could and should
give due regard to her counterevidence by giving up her belief that is an apple.
David could and should respect his counterevidence from Fred’s testimony by
suspending judgment as to the outside weather.
Insofar as our characters violate a requirement of rationality if they con-
tinue to believe despite their counterevidence, it is not that they are failing to
believe the propositions that their counterevidence supports. It is rather that
they are believing a proposition that their total evidence does not support. The
requirement of rationality at work in all of these cases—a requirement that
supports the requirement to acknowledge one’s counterevidence—it is the all-
too-familiar requirement not to believe what one’s total evidence does not
support. Lackey’s view of normative defeat misunderstands the phenomenon
of defeat and defeaters for it misunderstands not only how counterevidence
defeats, but also what it takes to acknowledge one’s counterevidence. Defeaters
qua counterevidence do not defeat by introducing obligations to believe
propositions supported by the defeating evidence. Defeaters qua counterevi-
dence defeat by changing what propositional attitudes a subject’s total evi-
dence supports taking. The evidentialist has a handy treatment of all of these
cases without introducing a requirement to believe—or even a permission to
believe—what the counterevidence supports. Just as there are no doxastic
defeaters, there are no normative defeaters.³⁴
Internalist epistemologists tend to be motivated by one or both of two
distinct principles. The first is that your beliefs should always fit your evidence.
The second is that when you believe something, that imposes on you an
obligation to adjust your other beliefs accordingly if doing so is necessary to
maintain coherence. We have seen the first—evidentialist—principle used to
articulate a relatively promising theory of epistemic defeat. We have seen an
attempt to use the second—responsibilist—principle to understand defeat, but
that effort seems to us much less promising.
³⁴ It is worth noting that none of the responsibilists Lackey cites—BonJour, Chisholm, McDowell,
Williams—ever suggest that the only way to acknowledge one’s counterevidence is to believe the
propositions one’s counterevidence justifies believing. As far as we can tell, these philosophers do not
acknowledge her category of normative defeaters.
    57
In particular, the doxastic/normative defeat framework should be avoided.

Talk about doxastic defeaters gives a free pass to the implausible view that
unjustified beliefs can defeat beliefs that have otherwise very strong prima facie
justification. Normative defeat is closer to what the rest of us would recognize
as defeat—defeat by counterevidence. But if we’re going to understand defeat
in terms of reasons/evidence, it should be in terms of reasons permitting or
failing to permit the belief whose justification is at issue (as Pollock does), not
in terms of reasons producing duties to believe competing propositions.
3.5.
So far, we’ve been looking at attempts to understand defeat in evidentialist and

responsibilist frameworks. But these are not the only frameworks for under-
standing defeat. We think that the work we’ve done points the way toward a
distinctively reliabilist account.
This is important. Reliabilism is primarily a theory of prima facie justifica-
tion and as such, it needs a theory of defeat. If the only way to understand
defeat is in evidentialist or responsibilist terms, this seems to present a
problem, with the reliabilist invoking reliability in her treatment of prima
facie justification, but then pivoting to the quite alien concepts of evidence or
responsibility in her treatment of defeat and thus ultima facie justification.
Although extant versions of reliabilism have perhaps failed so far to offer
adequate theories of defeat, we think this is not a problem intrinsic to

reliabilism. The various forms of reliabilism do have sufficient resources for
illuminating the nature of epistemic defeat, without deviating from the basic
spirit of reliabilism. We argued above that Pollock’s evidentialist framework
is much more promising than Lackey’s responsibilist one. This is good news
for reliabilism, for although neither framework sits comfortably with reliabi-
lism without modification, Lackey’s framework is much more deeply inhos-
pitable to reliabilism. For example, it’s hard to see a rationale in terms of
reliability for demanding that an agent ignore her highly reliable visual
processes, simply because she has the false and irrational belief that those
processes are unreliable.³⁵ In contrast to this deep and probably constitutional
³⁵ One could, of course, try to claim that there’s a reliable process of responding to higher order
defeaters, or something similar. We don’t think this is at all promising, in part because this is a poor
candidate for a cognitive process in the needed sense (see Lyons (2019)) but also because even if it were,
it would seem to be the kind of belief-dependent process that can only transmit or preserve justification,
not generate or create it (see Goldman (1979)).
58  .    . 
incompatibility between responsibilism and reliabilism, the conflict between

evidentialism and reliabilism might be relatively superficial—at least in the
present context. We can get a reliabilist-friendly account of defeat—
admittedly a sketchy and preliminary one—by taking the Pollockian structure
articulated above and purging it of any reliance on reasons or evidence. Let’s
address the why of this before the how.
Although several authors have developed evidentialist versions of reliabilism,³⁶
reliabilism is typically taken to be a competitor to evidentialism. One reason for
this is that standard forms of reliabilism state sufficient conditions for justifica-
tion (e.g., process reliability) that don’t entail that the agent has any reasons, or
evidence. Additionally, some reliabilists think reasons by their very nature are too
internalistic and will deny that any justification is a matter of reasons. More
modestly, some beliefs, especially so-called “IIPM beliefs” (introspection, intu-
ition, perception and/or memory beliefs) might be justified in a way that doesn’t
have anything to do with being based on reasons. Recall that reasons and
evidence, on the broadly Pollockian view, are things on which a belief is based;
on the nonevidentialist view, however, we “just know” the deliverances of
perception, memory, and so on, without basing them on anything. We won’t
argue against evidentialism here, but we do want to show how a reliabilist theory
of defeat can avoid evidentialism.
Suppose you have it on good authority—and justifiedly believe—that the
previous tenants left no furniture in the apartment. Later, when you open the
door, you see furniture. This gives you a defeater for your belief that there’s no
furniture in the apartment. If we’re assuming a nonevidentialist reliabilism

where perception gives justification without the involvement of reasons, we’ll
need a way to describe the situation without invoking reasons. We’ll say
instead that perception gives you a warrant for believing that there’s furniture
in the room: it gives you prima facie propositional justification for that belief,
whether you end up believing that or not. Warrants are like reasons, but
without any assumption that there’s some other mental state, introspectible
or otherwise, on which the belief is based. Having a warrant for believing p is
thus compatible with having reasons for p, but if there is justification that
doesn’t derive from reasons, it’s also compatible with lacking any reasons for
p. Thus, the concept of having a warrant is a more general and inclusive
concept than that of having a reason. Any time you have a reason you have a
warrant, but you might have a warrant without having a reason.
³⁶ See for example Alston (1988); Comesaña (2010); Tang (2016); and Miller (2019).
    59
This allows us to articulate a general view about defeat that is very much in
the Pollockian spirit. We can use Pollock’s view about the structure of defeat
and simply drop the assumption that all warrants involve reasons. We simply
take what Pollock says in terms of reasons and say it instead in terms of
warrants:
• S has a rebutting defeater for her belief that p iff she has a warrant to
believe not-p.
• S has an undercutting defeater for her belief that p iff she has a warrant to
believe that her warrants for believing p are inadequate.
• S has a defeater for her belief that p iff she has a rebutting or an
undercutting defeater for her belief that p.³⁷
Having a warrant to believe p, we take it, is nothing more than having prima
facie propositional justification to believe p. Thus, all the reliabilist needs to do
to have a theory of defeat is to develop a theory of propositional justification.
This is non-trivial, since reliabilism is first of all a theory of doxastic justifica-
tion, but it’s far from hopeless.³⁸ Pollock himself offers what’s clearly a theory
of propositional justification, where propositional justification is derived from
doxastic justification: a reason, for Pollock, is something that could give you
doxastic justification.
Representing a more pessimistic view, Bob Beddor argues in a recent paper
that the standard reliabilist accounts of defeat fail, and suggests that reliabilists
appeal to reasons to correct this deficiency.³⁹ While we want to concede that

his argument is effective against a standard extant reliabilist theory of defeat,
we will insist that it poses no insuperable problem for reliabilism more
generally. And it does nothing to show that defeat needs to be understood in
terms of reasons.
A (perhaps the) standard reliabilist theory of defeat is Goldman’s Alternate
Reliable Process theory (ARP), according to which S’s prima facie justified
belief that p at t is defeated just in case S has another reliable process available
³⁷ For the purposes of this paper, we are assuming that all defeaters are rebutting or undercutting,
which means that you only get defeat where you have warrants, thus that there are no “normative
defeaters” in that other, not-Lackey, sense mentioned above.
³⁸ We’re arguing here that reliabilism needs a theory of propositional justification because it needs it
for a theory of defeat. This stands in contrast both with the views of Kornblith (2017), who thinks that
doxastic justification is the only one of the two that matters, and of Feldman and Conee (1985), who
claim to base their theory of doxastic justification on their theory of propositional justification.
³⁹ See Beddor (2015), and also his Chapter 7 in this volume, which is similar in important ways to
what we’re offering here.
60  .    . 
to her at t, which if used in addition to or instead of the one actually used,

would have resulted in S’s not believing that p.⁴⁰ ARP formulates defeat in
terms of counterfactual non-belief rather than counterfactual disbelief, largely
because, as we saw previously, the proper response to defeaters is often
agnosticism, rather than doxastic reversal.
Against ARP, Beddor offers the following cases:
S sees a tree but has a reliable skeptic’s-advice-predicting mechanism, which

S normally trusts when she uses it. The skeptic would have advised not
believing there’s a tree, and in this case, had S used this mechanism,
S wouldn’t have believed there was a tree. Nevertheless, S’s visual prima
facie justification is, intuitively, undefeated by the availability of this
mechanism.⁴¹
This case aims to show that, contrary to ARP, the agent’s having an available
reliable process that would have led to non-belief, is insufficient for defeat.
Here’s Beddor’s other case:
S hears a defeater for her belief that p, from a reliable source, but one for
whom S holds an inveterate and irrational distrust. S would never take this
person’s testimony to heart, so she persists in believing p. Intuitively S’s belief
is defeated, even though her distrust means that the process of taking the
testimony on board is not available to her.⁴²
This case aims to show that, contrary to ARP, the agent’s having an available
reliable process that would have led to non-belief is unnecessary for defeat.
There are serious objections one could level against the kind of thing Beddor
counts as a process and what he takes availability to be. We won’t pursue these
objections, for we think that Beddor’s cases reveal a genuine flaw with ARP,
whether or not they refute ARP as they’re currently stated.
Beddor thinks that the problem with ARP is that it omits consideration of
reasons. We think the problem with ARP is that it is formulated in terms of
what the agent would have believed, or not believed, had she used another
process. Roughly, an adequate theory of defeat needs to be formulated in terms
of what the agent should have believed or not believed had she used another
⁴⁰ Goldman (1979); Lyons (2009). To be at all plausible, this needs to be interpreted as claiming that
the belief is prima facie defeated, not that it is ultimately defeated.
⁴¹ Beddor (2015): 150–1. ⁴² Beddor (2015): 155.
    61
process. More precisely, since we’ve seen how badly that fared for Lackey’s
normative defeaters, what the agent would or would not have had (prima
facie) warrant to believe, had she used another process. Counterfactual non-
belief only entails defeat if all the relevant components of the counterfactual
situation go normatively correctly, but this isn’t guaranteed merely by stipu-
lating that the process is reliable. This is because the counterfactual situation
will typically involve all kinds of factors beyond the immediate consequences
of the use of the alternate process. That alternate reliable process might feed
into an unreliable one, so that the agent forms a justified belief and then
concludes something unjustified from it. Or the justified belief might be one to
which the agent would respond in an irrational way that’s in no way sanc-
tioned by reliabilism. Beddor’s cases trade on just these kinds of possibilities,
on the gap between what the agent would have believed (/not believed), and
what the agent should have believed (/not believed), or is licensed in believing
(/not believing). ARP is flawed because it focuses on the way alternative
available processes yield counterfactual non-belief, when it ought instead to
be focused on the way alternative available processes yield warrants.
Does this appeal to warrants, and to what the agent does or doesn’t rightly
believe in the counterfactual situation, amount to a significant concession to
Beddor, or to the evidentialist, or to the opponent of reliabilism? Not at all!
Reliabilism is a normative epistemological theory, after all. Its raison d’être is
to give an account of warrants, of what beliefs are prima facie normatively
appropriate. If it can’t do that, then defeat is the least of reliabilism’s problems.
Here, then, is the sort of thing a reliabilist can and should say about
possession of warrants and thus about propositional justification:
S has (prima facie) warrant for believing p at t iff a cognitive process that
satisfies the general theoretical requirements for prima facie (doxastic)
justification (a) is available to S, and (b) if used at t, taking as inputs only
states that S is already in, does or would likely produce p as output.
Again, this is a sketch and is intended primarily for the purpose of illustrating
that reliabilism has promising resources for accommodating defeat. Nevertheless,
several comments about the proposal are in order.
Notice that the formulation is very general, leaving open what the require-
ments are for prima facie justification. This is for two reasons. First, it allows
the present theory of warrant possession (and consequently of defeat) to be
easily imported into reliabilist theories that differ in other respects. Second, it
makes quite explicit the fact that the present theory of defeat is making use of
62  .    . 
old theoretical apparatus, not importing new machinery that might not fit the
general reliabilist ethos. It will therefore inherit the virtues (and vices) of that
apparatus. Plug in a naturalistic and reductive and nonevidentialist theory of
prima facie doxastic justification, and you’ll get a naturalistic and reductive
and nonevidentialist theory of propositional justification, and thus of defeat,
and thus of ultima facie doxastic justification.
Clause (b) restricts the relevant outputs of the available processes to those
that would result from the system’s taking as inputs only states S is already in.
This is perhaps already entailed by a reasonable individuation of cognitive
processes (see Lyons (2019)), but it’s worth making it explicit. We don’t want
the fact that S has an unopened encyclopedia nearby to give S warrant for
everything written in the encyclopedia.
Next, the formulation mentions the process producing “p as output,” rather
than producing the “belief that p.” We trust that it is intelligible to say that a
belief-forming process might be producing an output, even when that output
isn’t a belief. A visual process might output the proposition that the lines are of
unequal length, even though the viewer is convinced that it’s an illusion and
thus does not believe they’re of unequal length. More generally, we want to say
that having a process available can give you warrant (/propositional justifica-
tion) for some proposition, even though you don’t believe that proposition,
and perhaps never would.⁴³
Finally, the formulation is vague in some crucial respects. In particular, it
relies on the availability of a process that does or would likely produce p. Some
might object that the vagueness connected to availability and likelihood here
are unacceptable. On the contrary, these might just be unavoidable features of
propositional justification, on any reasonable theory. If I justifiedly believe p
and that p implies q, then I’m propositionally justified in believing q; and all
this gives me a defeater for my belief that ~q. But not merely because my
justified beliefs entail q—they entail every necessary truth, many of which I’m
not propositionally justified in believing and which don’t provide defeaters.
Why? Because these consequences aren’t obvious to me, where obviousness is
cashed out by the reliabilist in terms of process availability. Since obviousness
comes in degrees, availability will have to as well, if this move is to have any
hope of working. For similar reasons, we can’t demand of an available process
⁴³ Could a belief-forming process produce as outputs states that aren’t at all belief-like? If an
alternate reliable process produced the desire that p, it doesn’t seem like this would be a defeater for
the belief that ~p (thanks to Beddor, personal communication). Maybe we need to restrict the outputs
to states with assertoric force or which are in some other way sufficiently belief-like. We won’t try to
sort out the details here.
    63
that it would lead to an output that p. For those logical consequences that are
not screamingly obvious but yet obvious enough to count as propositionally
justified, it’s plausible that the use of the relevant process wouldn’t guarantee
the output of that consequence but would only make it quite likely.
Suppose Chuck believes that he’s never owned a weapon, although if he
thought about it for a while, he’d remember that his parents bought him some
Chinese throwing stars for a birthday back in the 1980s. Is Chuck’s belief that
he’s never owned a weapon defeated? It depends, we think, on how hard it
would be for Chuck to recall this: how long and hard Chuck would have to
probe memory, how robust or fragile the cueing conditions would need to be
in order to retrieve the memory, how likely a memory search would be to
trigger recall, etc. Given that defeat comes in degrees, we might even claim—
with some plausibility—that Chuck’s degree of ultima facie justification is
affected by these considerations: if three seconds of thinking about it, under
pretty much any conditions, would fetch the relevant memory, then Chuck’s
prima facie justification for believing he’s never owned a weapon is signifi-
cantly reduced; if only a great deal of focused concentration, when primed
with ninja movies, would result in retrieval of this memory, then Chuck’s
initial justification is only slightly reduced. In this way, what Goldman (1986)
calls power (a ratio of truth to queries, where reliability is a ratio of truths to
falsehoods) might be directly relevant to propositional justification and defeat,
even if it’s not directly relevant to positive, prima facie justification.⁴⁴
Some people don’t like the idea that possession of reasons, or warrants, or
evidence, might come in degrees. In order to avoid this result, Feldman (1988)
goes so far as to claim that the evidence a person possesses at a given time is
restricted to those conscious mental states she’s in at that time. The worry for
this proposal is that it would seriously diminish the body of evidence anyone
possesses at a given time. How many items can be in consciousness at once?
Serious estimates range from one (Dehaene 2014) to 7 +/-2 (Miller 1956), with
most people thinking it’s about 4. Not enough, it seems. We saw above that
Pollock largely elides this important issue; although he articulates a very
detailed theory of ultimate defeat, none of it applies directly to questions of
⁴⁴ These thoughts seriously complicate the treatment of defeater-defeaters, the full treatment of
which we reserve for another occasion. If a regular defeater is a “first-level” defeater, and a defeater-
defeater is a “second-level” defeater, and so on, then higher-level defeaters will tend to be less available
than lower-level defeaters, due to the cognitive difficulty involved in thinking of them and appreciating
their significance, with the result that higher-level defeaters will tend, roughly, to asymptote to
irrelevance. This contrasts with the Pollockian view (see Beddor, Ch. 7, this volume), according to
which higher-level defeaters destroy justification just in case there is an odd, rather than even, number
of levels.
64  .    . 
an agent’s actual epistemic status, unless we add the claim that an agent’s belief
is (prima facie) defeated iff she is in possession of a reason that serves as a
defeater for that belief.
Propositional justification was going to come in degrees anyway, because
reasons and warrants vary in their strength. But strength, availability, and
likeliness of output are orthogonal dimensions of variation: a piece of evi-
dence, for example, can be weakly supportive but fully obvious/available,
strongly supportive but far from obvious/available, etc. The fact that the
current proposed reliabilist theory of defeat reveals and highlights some of
the vagaries surrounding propositional justification is no criticism of that
proposal, not if these vagaries were really already a part of propositional
justification.
This proposal is surely in need of further refinement, and different reliabi-
lists may want to tweak it in different ways. A full treatment of this issue
demands a paper of its own. The point we wanted to make with it here was
this: a roughly Pollockian view about defeat can be smoothly accommodated
in a reliabilist epistemology, without the reliabilist having to embrace any
evidentialist or responsibilist elements that are in any way foreign to the
naturalistic spirit of reliabilism. A belief is prima facie defeated, roughly,
when the agent has available to her a “good” cognitive process (the kind that
would yield prima facie doxastic justification) that would output the content
that that belief is false or unwarranted. The proposal bears some obvious
similarities to the well-known ARP (which we take to be a selling point),
although it fixes some serious problems by trading warrants talk for talk of
counterfactual non-belief. At the same time, this shift from reasons to war-
rants serves to highlight what was right about Pollock’s seminal work on
defeat. Although Pollock himself focused on reasons, rather than warrants
more generally, there’s nothing stopping the nonevidentialist from taking on
board the vast majority of the Pollockian view. Having defeaters is having
prima facie propositional justification and having prima facie propositional
justification is having warrants. Having warrants is something that can be
understood in whatever general epistemological framework—evidentialist,
reliabilist, perhaps even responsibilist—you like.
The recent literature has been dominated by two influential discussions of
defeat. One of these has been healthy and productive; the other has clouded
some relevant issues and taken as uncontroversially true something that we
think no one should even believe, but certainly no one should simply assume:
namely, that unjustified beliefs can defeat justified beliefs. The good strand of
the standard view of defeat is the evidentialist strand, but it’s not good because
    65
it’s evidentialist. It’s good because it sees defeat as proceeding (only) from
warrants, which is true. Evidentialists make the further claim that only evi-
dence provides warrants. Those of us who deny that can still find a great deal
of common ground with the evidentialist theory of defeat.⁴⁵
References
Alston, William. (1988). An internalist externalism. Synthese, 74: 265–83.

Beddor, Bob. (2015). Process reliabilism’s troubles with defeat. The Philosophical
Beddor, Bob. (Ch. 7, this volume) Reasons for reliabilism.
Bergmann, Michael. (2006). Justification without Awareness. Oxford: Oxford
University Press.
BonJour, Laurence. (1985). The Structure of Empirical Knowledge. Cambridge:
Harvard University Press.
Broome, John. (2013). Rationality Through Reasoning. Hoboken, NJ: Wiley-
Blackwell.
Carter, J. Adam. (2015). Group knowledge and epistemic defeat. Ergo, 2: 711–35.
Carter, J. Adam. (2018). Meta-epistemic defeat. Synthese, 195: 2877–96.
Carter, J. Adam and Jesus Navarro. (2017). The defeasibility of knowledge-how.
Philosophy and Phenomenological Research, 95: 662–85.
Chisholm, Roderick. (1977). Theory of Knowledge. 2nd Ed. Englewood Cliffs, NJ:
Prentice Hall.
Chisholm, Roderick. (1989). Theory of Knowledge. 3rd Ed. Englewood Cliffs, NJ:
Prentice Hall.
Comesaña, Juan. (2010). Evidentialist reliabilism. Nous, 44: 571–600.
Dehaene, Stanislas. (2014). Consciousness and the Brain. New York: Penguin.
Feldman, Richard. (1988). Having evidence. In: Philosophical Analysis. D. Austin
(ed.) Dordrecht: Kluwer: 83–104.
Feldman, Richard and Earl Conee. (1985). Evidentialism. Philosophical Studies,
48: 15–34.
⁴⁵ Thanks to Bob Beddor, Adam Carter, Paul Silva, Mona Simion, and especially Jessica Brown for
comments on an earlier draft that led to improvements. Thanks also to Jonathan Kvanvig and Duncan
Pritchard for helpful conversations on these topics. We also thank audiences at the 2019 Bled
Philosophical Conference and at the 2019 Social Epistemology Network Event at Underwood
International College in South Korea for beneficial comments and questions. In particular, we recall
especially helpful comments from Lizzie Fricker, Chris Kelp, Mark Kaplan, Hilary Kornblith, and Brent
Madison.
66  .    . 
Frances, Bryan. (2014). Disagreement. Cambridge: Polity Press.

Goldberg, Sanford. (2018). To the Best of Our Knowledge. Oxford: Oxford
University Press.
Goldberg, Sanford and Jon Matheson. (2020). The impossibility of mere animal
knowledge for reflective subjects. Erkenntnis, 85: 829–40.
Goldman, Alvin. (1979). What is Justified Belief? In: G. S. Pappas (ed.)
Justification and Knowledge. Dordrecht: Springer: 1–23.
Goldman, Alvin. (1986). Epistemology and Cognition. Cambridge: Harvard
University Press.
Green, Christopher. (n.d.). Epistemology of Testimony Internet Encyclopedia of
Philosophy, http://iep.utm.edu/ep-testi/.
Grundmann, Thomas. (2011). Defeasibility theory. In: The Routledge Companion
to Epistemology. S. Bernecker and D. Pritchard (eds.) New York: Routledge:
156–66.
Kornblith, Hilary. (2017). Doxastic justification is fundamental. Philosophical
Topics, 45: 63–80.
Lackey, Jennifer. (1999). Testimonial knowledge and transmission. The
Philosophical Quarterly, 49: 471–90.
Lackey, Jennifer. (2003). A minimal expression of non-reductionism in the epis-
temology of testimony. Nous, 37: 706–23.
Lackey, Jennifer. (2005a). Testimony and the infant/child objection. Philosophical
Studies, 126: 163–90.
Lackey, Jennifer. (2005b). Memory as a generative epistemic source. Philosophy
and Phenomenological Research, 70: 636–58.

Lackey, Jennifer. (2006a). Learning from words. Philosophy and Phenomenological
Research, 73: 77–101.
Lackey, Jennifer. (2006b). Knowing from testimony. Philosophy Compass, 1/5:
432–48.
Lackey, Jennifer. (2006c). Introduction. The Epistemology of Testimony. J. Lackey
and E. Sosa (eds.) Oxford: Oxford University Press: 1–21.
Lackey, Jennifer. (2008). Learning from Words. Oxford: Oxford University Press.
Lackey, Jennifer. (2011a). Testimonial knowledge. In: The Routledge Companion
to Epistemology. S. Bernecker and D. Pritchard (eds.) New York: Routledge:
316–25.
Lackey, Jennifer. (2011b). Testimony: Acquiring knowledge from others. In: Social
Epistemology: Essential Readings. A. Goldman and D. Whitcomb (eds.) Oxford:
Oxford University Press: 71–91.
Lackey, Jennifer. (2014a). Socially extended knowledge. Philosophical Issues, 24:
282–98.
    67
Lackey, Jennifer. (2014b). A deflationary account of group testimony. In: Essays in

Collective Epistemology. J. Lackey (ed.) Oxford: Oxford University Press: 64–96.
Lackey, Jennifer. (2018). Credibility and the distribution of epistemic goods. In:
Believing in Accord with the Evidence: New Essays on Evidentialism, Synthese
Library 398. K. McCain (ed.) Cham, Switzerland: Springer: 145–68.
Lasonen-Arnio, Maria. (2020) Enkrasia or evidentialism? Philosophical Studies
177: 597–632.
Lyons, Jack. (2009). Perception and Basic Beliefs: Zombies, Modules, and the
Problem of the External World. Oxford: Oxford University Press.
Lyons, Jack. (2019). Algorithm and parameters: Solving the generality problem for
reliabilism. The Philosophical Review 128: 463–509.
Matheson, Jonathan. (2011). Epistemological problems concerning skeptical the-
ism. Faith & Philosophy, 28: 323–31.
Miller, Emelia. (2019). Liars, tigers, and the bearers of bad news, oh my! Towards a
reasons account of defeat. The Philosophical Quarterly, 69: 82–99.
Miller, George. (1956). The magical number seven, plus or minus two: Some limits
on our capacity for processing information. Psychological Review, 63: 81–97.
Palermos, Spyridon Orestis. (2011). Dualism in the epistemology of testimony and
the ability intuition. Philosophia, 39: 597–613.
Pollock, John. (1967). Criteria and our knowledge of the material world. The
Philosophical Review, 76: 28–60.
Pollock, John. (1968). What is an epistemological problem? American Philosophical
Pollock, John. (1970). The structure of epistemic justification. Studies in the

Theory of Knowledge, 4: 62–78.
Pollock, John. (1971). Perceptual knowledge. Philosophical Review, 80: 287–319.
Pollock, John. (1974). Knowledge and Justification. Princeton: Princeton
University Press.
Pollock, John. (1979). A plethora of epistemological theories. In: Justification and
Knowledge. G. S. Pappas (ed.) Dordrecht: Springer: 93–115.
Pollock, John. (1986). Contemporary Theories of Knowledge. Lanham, MD:
Rowman and Littlefield.
Pollock, John. (1987). Defeasible reasoning. Cognitive Science, 11: 481–518.
Pollock, John. (1994). Justification and defeat. Artificial Intelligence, 67: 377–407.
Pollock, John. (2001). Defeasible reasoning with variable degrees of justification.
Artificial Intelligence, 133: 233–82.
Pollock, John. (2006). John Pollock. In: Epistemology: 5 Questions. V. F. Hendricks
and D. Pritchard (eds.) Copenhagen: Automatic Press/VIP.
68  .    . 
Pollock John and Joseph Cruz. (1999). Contemporary Theories of Knowledge.

Lanham, MD: Rowman and Littlefield.
Tang, Weng Hong. (2016). Reliability theories of justified credence. Mind, 497:
63–94.
Michael Williams. (1999). Problems of Knowledge. Oxford: Oxford University
Press.
Worsnip, Alex. (2018). The conflict of evidence and coherence. Philosophy and
4
Losing Knowledge by Thinking about
Thinking
Jennifer Nagel
Knowledge can be real without being armor-plated. We gain and lose knowledge
in many ways in the normal course of life. We learn and then naturally forget,
within minutes or hours, many details of what we have perceived for ourselves
and heard from others. We can also lose knowledge without losing belief; for
example, when the world changes in unseen and unexpected ways behind our
backs. In addition to passive processes of forgetting and becoming outdated,
we can actively switch to suspension of judgment on points where we once had
knowledge. This last type of switching is often routine and unremarkable; for
example, when one decides to suspend belief upon recognizing that one’s
grounds or basis for judgment on a given point may well have weakened
over time. Upon reflection, I have doubts about the accuracy of my mental
map of the city where I used to live two decades ago, a city I used to know quite
well. The quality of my judgments about Albuquerque has suffered from

erosion on two fronts: the city’s businesses, streets, and bus routes have
doubtless changed over time, and my memories have been shifting and fading.
Given this natural deterioration of my basis, it is entirely reasonable—and
epistemologically uncontroversial—for me to suspend judgment on any finer
points of that city’s layout. My current state of suspension casts no shade on
the epistemic status of the judgments I routinely made in this area twenty
years ago.
Other cases of switching into suspension seem potentially more troubling.
To some, these cases suggest a certain tension between rationality and know-
ledge, or some rational fragility in the bases of the judgments which constitute
knowledge. Known in the literature as “defeat cases,” these scenarios are
sometimes taken to show that it is possible to have knowledge of a proposition
on a certain basis, and then find oneself rationally required to switch to
suspension of judgment, despite continuing to have exactly one’s original
basis for judgment. Others have argued that even if these cases show that it
/9780198847205.003.0004
70  
is in some sense unreasonable to cling to the originally knowledge-sustaining

basis of one’s judgment, one can nonetheless retain knowledge by doing so.
This chapter argues that defeat cases have been misinterpreted by both
sides. Knowledge is indeed lost in defeat cases, but it is not lost while one’s
original justifying basis is retained: a deep feature of these cases is that they
inevitably involve epistemically significant—although sometimes hard-to-
spot—shifts in the basis of one’s judgment of the original first-order question.
These shifts are produced in various ways, including the production of self-
conscious worries about how one is thinking. The new basis that arises after
the shift will not sustain knowledge of the key proposition, but it is not the
same as the basis one had before the self-conscious worries or other defeating
conditions arose. Close examination of these cases can help us gain a sharper
sense of what actually constitutes the basis of a judgment, and why exactly the
bases that underpin knowledgeable judgments are strictly immune to rational
undermining, while nevertheless being vulnerable to loss through what I will
call basis displacement. The fact that knowledge of a given proposition can be
lost does not show any weakness in its original basis: just as you might come to
forget ever having felt the pain of some particular toothache, where this
forgetting casts no shadows over the epistemic credentials of your originally
vivid knowledge of the pain, so also you can lose knowledge through the
displacement of your original basis for judgment, notwithstanding the optimal
epistemic quality of that original basis.
Section 4.1 outlines my infallibilist approach to justification and the more
popular fallibilist alternative. Section 4.2 aims to explain the appeal of fallibil-
ism, and Section 4.3 examines defeat cases in the light of my theory of
justification.
4.1. Justification and the Basis of a Judgment
Despite immense debate over just what is meant by ‘justification’ in epistem-

ology, there is widespread agreement that it denotes a positive normative
status, and that this status applies at least to all states of knowledge. My
approach here will be to start with the justification of what is known, before
moving outward to more controversial cases.
What is the justifying basis of a person’s knowledge of a fact? We might start
by observing that facts (or true propositions) can be known in a variety of
ways. It is possible to know a fact—say, where some object is currently
located—through several distinct sensory modalities, through appropriate
      71
testimony from a trusted knower, through sound inference from premises

themselves known through observation or testimony, or through some com-
bination of these. In addition to the outer-world channels of perception,
inference, and testimony, we also have interoceptive awareness of facts about
inner phenomenal states such as thirst and dizziness. The range of ways in
which propositions may be believed includes all these ways of knowing—given
that knowing entails believing—plus a wide variety of epistemically worse
ways of coming to make a judgment. One might come to believe a proposition
on the basis of an illusion, deceptive testimony, or unsound inference, or in
some other way, perhaps even as a result of a novel medical intervention. One
might wonder what unites the more restricted set of ways in which proposi-
tions can be known, as opposed to just believed. If we take knowing to be
distinguished from believing by its special relationship with the truth—it is an
essential feature of knowing that only what is true can be known—then ways
of knowing are likewise most naturally distinguished from the broader class of
ways of believing in virtue of their relationship to the truth. In my view, a way
of judging that p is a way of knowing only if the truth of p is essential to this
way of judging.¹ On this way of individuating “ways of knowing,” inference is
not a way of knowing, but sound inference from known premises is. Likewise,
perception counts as a way of knowing only if it is understood in success terms,
as involving sensory systems that have an appropriate causal sensitivity to a
range of facts, where some fact within this range causes a corresponding
judgment through the normal operation of these systems. Although “percep-
tion” is sometimes used in a broader sense that encompasses any kind of

sensory activation, I will use it in the more restricted success sense here.
Similarly, testimony will denote acceptance of an honest knower’s word, rather
than the absorption of any report. With this restricted understanding of ways
of knowing in hand, my proposal here is that the justifying basis of a person’s
knowledge of a fact—what makes that attitude epistemically right, indeed
optimal—is nothing other than the way it is known. While my initial focus
is on the justifying bases of knowledge, I do not mean to imply that judgments
falling short of knowledge can have no justifying bases whatsoever; their
weaker, derivative level of justification is a function of how closely the ways
in which they were formed resemble ways of knowing (further details will
emerge in due course).
¹ Note that this is a necessary, and not a sufficient condition; it may be essential to guessing-
correctly-on-a-Tuesday that the relevant proposition is true, but this will not be enough to qualify it as a
way of knowing.
72  
This basic approach to the relationship between knowledge and justification

has old roots in the history of philosophy, most vividly in the Classical
Indian Nyāya tradition. In the Nyāya Sutra (c. 200 CE), Gautama characterizes
perception as “inerrant” (1.1.4), and stipulates that it must involve a causal
connection between the perceived object and the sensory faculty (Dasti
and Phillips 2017). Other ways of knowing (pramānas, : in Sanskrit) such
as testimony and inference, are likewise described in the tradition as “fault-
less,” “unfailing,” and “non-deviating” (Matilal 1986: 135). Less-than-sound
inference is labelled “pseudo-inference,” hallucination and the like are
“pseudo-perception,” and misinformation is not testimony but “pseudo-
testimony” (Dasti and Phillips 2010; Dasti 2012). Contemporary non-skeptical
epistemologists will readily agree that the good cases here have some justifying
basis; what is more controversial is to identify the justifying basis of a known
fact with the way it is known.
This understanding of the justifying basis of what is known is frankly
incompatible with the currently popular fallibilist approach to knowledge,
according to which there is typically a significant gap between one’s knowledge
of a proposition and one’s justifying basis, so that what is known can have the
same justifying basis as what is merely believed, or even falsely believed. Here
is Baron Reed’s gloss on the core idea of fallibilism:
Roughly stated, the basic idea is that the subject can know something even
though it could have been false. This is not the same as saying that the subject
can know something that is false – it is very widely accepted by philosophers

that, if a belief counts as knowledge, it is true. Rather, the claim is that a belief
held with a particular epistemic grounding can be knowledge even though
the subject could have held that belief with the same grounding in circum-
stances where the belief is false (and, of course, in those circumstances the
belief would not count as knowledge). (Reed 2012: 585)
According to fallibilists, false beliefs can have precisely the same epistemic
grounding as instances of knowledge, so epistemic grounding in cases where p
is known cannot be understood in a way which makes the truth of p essential
to one’s epistemic grounding for p.
Not all formulations of fallibilism are given in terms of epistemic grounding;
some speak of evidence or justification. Another standard formulation of
fallibilism is rendered by Reed as follows: “A person S knows that p in a
fallibilist way just in case S knows that p on the basis of some justification j
and yet S’s belief that p on the basis of j could have been false (or mistaken or
      73
in error).”² (Reed 2012: 586) The example he offers to bolster his point will
bring out the difference between our views of justifying bases:
For example, I know that the Cubs beat the Dodgers the last time they
played; I know it because my brother told me what happened, and he is
usually reliable about this sort of thing. But, if he had misread the box score
in the newspaper, I still would have believed him. In that case, my belief
would have been held with the same justification, but the belief would have
been false. Because this case is possible, the knowledge I actually have (when
my brother did not misread the box score) is fallible. (Reed 2012: 586)
For simplicity, I will understand the subject of these cases as hearing from his
brother just that the Cubs won, without any further elaboration about his
having read the paper as opposed to having watched the game live or on
television. I will also understand the subject to be taking his brother’s word for
it, in both cases, as opposed to forming his belief on the basis of explicit
reasoning from the premises, “My brother said that the Cubs won,” and “My
brother is usually reliable about this sort of thing.” We generally accept trusted
testimony without such calculated reasoning about its likelihood of being
correct, and if we were to restrict ourselves to such explicit reasoning, it is
not clear that we would gain as much as we do from the reports of others. If
explicit reasoning from exactly the premises just given yields knowledge, it
seems only to yield knowledge of a weaker conclusion like, “It is highly likely
that the Cubs beat the Dodgers,” as opposed to knowledge of the stronger
conclusion that the Cubs won. I take Reed’s use of the stronger conclusion in
this example to indicate that he also accepts trusted testimony as a path to
knowledge that is interestingly different from the explicitly reasoned calcula-
tion that could deliver knowledge of the weaker claim. Indeed, the weaker
claim (that it was likely the Cubs had won, given the report of a typically
reliable person), could be true, and arguably even known, even in the case of
the misread loss, so, if the contrast between the cases is to illustrate that the
very proposition known might have been false, we should not weaken it.
I agree with fallibilists like Reed that knowledge can be transmitted through
testimony, even third-hand, as from a brother’s report of a newspaper story.
However, the view I will defend distinguishes different justifying bases for the
² In the face of concerns about knowledge of necessary truths, this provisional formulation is later
amended to “S knows that p in a fallible way just in case S knows that p on the basis of some
justification j and yet S’s belief that p on the basis of j could have failed to be knowledge” (Reed 2012:
587). This amended formulation remains incompatible with my account.
74  
good case and the bad: in the good case, I take the justifying basis to consist in
the entire conduit from the Cubs’ actual victory to the subject’s awareness of
that victory, incorporating the successful perceptual and inferential processes
giving rise to knowledge on the part of the reporter, and the safe transmission
of this knowledge to the brother, who reports what he came to know by
reading the paper to the appropriately trusting subject. It is because the
subject has learned in such a way from a knowledgeable source that his
judgement is epistemically in the clear, normatively appropriate, and fully
justified. In the bad case, there is no such chain of transmission between the
outcome of the game (presumably a loss) and the subject’s formation of a
belief in a victory. Misreading a report is not a way of knowing, nor is trusting
someone who is mistaken. There are some commonalities between the good
and bad cases: for example, in both, the subject trusts someone who is usually
reliable. Thanks to the partial similarities between the knower and the subject
of the bad case, even the bad case subject has some measure of justification (on
which more later), but the justifying bases of the good and the bad cases are
not the same.
On the common fallibilist understanding of these cases, by contrast, the
subjects of the good and bad cases can have precisely the same justification.³
One way to make sense of this would be by identifying justification with what
is available to the subject to report as the basis of his judgment that p, if asked
how he knows that p. What matters here is not the actual history of the
subject’s judgment, but only those aspects of the apparent history which are
readily available to the subject for report. This subjective impression of one’s
basis will be called the subjective basis, for short, although we should be careful
to note that the subjective basis is not a basis for the judgment in the causal
³ I say “can have” rather than “always have,” to admit a greater variety of types of fallibilism. For
some forms of internalist fallibilism, the subjects of good and bad cases like the ones just given will
always have exactly the same justification; there is no gap between what one’s justifying basis is and
what it seems to be (e.g., Smithies 2019). But even for externalist fallibilists like Goldman (1979), who
would assign sharply different levels of justification to the perceiver and the hallucinator, running
distinct cognitive processes with different levels of reliability, there are at least some phenomenally
indistinguishable pairs of good and bad cases where the same process is at work across the divide. For
example, if the same retinal stimulation is produced by a convex object lit form above and a concave
object lit from below, generally resulting in a perceptual impression of convexity, given our learned
background expectations of light from overhead, then the same reliable but not infallible process is at
work in the person who sees the normally lit convex object as convex, and the person who sees the
concave object as convex in the presence of subtle lighting from below. These good and bad cases are
equally justified, in Goldman’s later fallibilist reliabilism (although not in his earlier infallibilist causal
view), given his restriction on the individuation of belief-forming processes to what happens “within
the organism’s nervous system” (Goldman 1979: 12). Although for simplicity I focus largely on
internalist forms of fallibilism here, I believe there are ways of extending the arguments of this chapter
to positions like this kind of reliabilism.
      75
sense of that word, even if it represents some elements in the causal basis.
Between this subjective fallibilism and my objective infallibilism lie a spectrum
of views, including externalist forms of fallibilism such as process reliabilism;
for simplicity, I will focus mainly on the extreme ends of the spectrum in what
follows.
We have a limited view of the causal processes behind our judgments,
even in cases where we gain knowledge. The subject who comes to know
about the victory can trust his brother without being able to report on the
exact track record supporting his sense that his brother is reliable
about “this sort of thing” (and indeed without being able to explain the
exact scope of “this sort of thing”), although there may well be some
intelligent appreciation of the brother’s track record underlying this trust.
Consciousness presents products of cognition, rather than the underlying
processes, let alone the distal features of the processes that secure our
grip on external truths. It is because what is available to consciousness is
so restricted that the bad case (pseudo-testimony) can be mistaken for
the good (testimony): from the subject’s first-person perspective, I can
equally respond to a challenge in either case by saying something like,
“my brother said that the Cubs won, and he is usually reliable about this
sort of thing.”
In my view, the similarity of reports produced in the good and bad cases is
evidence of the disparity between subjective and objective bases: one might
reasonably say the same thing about one’s basis in a case of error and in a
case of knowledge, and in either case, one has limited first-person access to
what actually makes one’s judgment normatively appropriate. The fact that
one’s judgment is normatively appropriate in the good case means that one
does have suitable first-person access to the key first-order fact (for
example, that the Cubs won), but it is possible to know that the Cubs won
without knowing the full story behind how it is that you know this fact,
down to the relevant characteristics of the reporter who saw the game,
and so forth. Despite failing to capture the full story of how we are
normatively in the clear, subjective impressions of our ways of thinking
nevertheless play an interesting role in our cognitive economy, and are not
simply epiphenomenal to the course of our thinking about first-order issues
over time. In particular, conscious representations of the bases of our
judgments play a vital role in our cognitive flexibility, our capacity to
change how we are thinking about first-order issues, whether such changes
are prompted by anomalies we encounter as individuals, or by challenges we
receive from others.
76  
4.2. The Function of Subjective Representations of Basis
The appeal of fallibilism can be better understood if we undertake a deeper

investigation of the character and function of conscious representations of the
bases of our judgments. One intriguing feature of these conscious representa-
tions is their pervasive availability. While our attention is ordinarily directed at
first-order questions in the world—one feels elation at the Cubs’ victory, one
notices a large object in the room—these first-order questions are for self-
conscious creatures like us always settled with some accompanying represen-
tation of how they are settled (through testimony, through visual or haptic
perception, and so on). There are consciously available differences between
seeing or touching an object, and between judging that the Cubs won on the
basis of having seen them play, having inferred that they must have won, or
having heard about it via testimony. It is a good question why we have this
kind of metacognitive awareness, and what relation it ultimately bears to the
normative status of our judgments.
To answer these questions, we can begin by reviewing some evidence
bearing on the general availability of subjective representations of basis. In
roughly a quarter of the world’s languages, ordinary declarative judgments
bear obligatory markings of the distinctions between major ways of knowing
such as first-hand perception, testimony, and inference (Aikhenvald 2004),
but the source monitoring functions that underpin those grammatical
markings are taken to operate in typical adult speakers of all languages
(Johnson, Hashtroudi et al. 1993; Papafragou, Li et al. 2007). In earlier work,

I emphasized the social function of source monitoring: if we have some indica-
tion of where our judgments are coming from, we can do a better job of
sharing our knowledge with others, and where needed, selectively telling
them what they don’t already know (Nagel 2015). Source monitoring also
enables us to share our perceptual knowledge of features of our current context
by directing others’ perceptual attention appropriately (“look over there!”),
sparking their own independent way of knowing the fact in question. Another
function of source monitoring is connected to our individual cognitive flexi-
bility: consciously available representations of the bases of our judgments
guide us in our active epistemic exploration of the world (stepping closer,
turning the lights on, moving our hands, or gaze over the object), and in
rethinking matters when something we have judged to be the case turns out to
be problematic. Realizing I have taken someone else’s word for a disputed
claim, I can try to verify it with my own eyes; if what I seem to see strikes me as
implausible I can ask others if they also see it. These tasks of sharing and
      77
rethinking can be combined: in epistemically cooperative communities, the

work of sorting testimony from pseudo-testimony (and so on) is socially
distributed. Others are often better positioned to tell whether you are seeing
or hallucinating, or whether your source is trustworthy. So, even in languages
such as English, where source marking is lexical and optional, we frequently
communicate source, marking what is inferred with a construction like
“I gather that . . . ” or what is derived from testimony with “I hear that . . . .”
These constructions can be used to invite those who may be better positioned
epistemically to corroborate or correct our judgments (Pomerantz 1980).
The social role of subjective bases does something to explain the natural
appeal of fallibilist views of justification. Even if there is something norma-
tively undesirable about cases of pseudo-perception or pseudo-testimony
(other things being equal, surely these are not the conditions we hope to
inhabit), the subjects who find themselves in these conditions will behave
just like their good case counterparts when their epistemic status is challenged.
They will tell similar (partial) stories about how they arrived at their judg-
ments, including to themselves, if they become worried about their own
epistemic standing and attempt to introspect. As far as their active control
of their epistemic agency is concerned, subjects are guided by their subjective
bases, so theories that take justification to be a matter of the proper exercise of
epistemic agency will place considerable weight on subjective bases.
Explicitly inferential reasoning deserves special attention in this account.
Reasoning might seem to be an exception to my contention that the subjective
impression of one’s basis is only a partial representation of what makes the

judgment normatively appropriate. In particular, when one is conscious of the
premises of a deductively sound argument, and then derives the conclusion,
one might think that the subjective and objective bases of the judgment
coincide, and that in this type of case, the subjective basis fully captures
what causes the judgment and makes it normatively appropriate. However,
even here, there are many normatively relevant features of the inferential
judgment that are not presented to consciousness. If one is to attain knowledge
of the conclusion, it matters not only which premises manifest themselves in
consciousness (say, by being rehearsed in inner speech), but also the epistemic
standing of these premises (which will typically depend on factors not present
in consciousness), and the soundness of the rules of inference sparked by the
recitation of those premises to produce the conclusion. Explicit reasoning, like
all reflective cognition, consists in operations performed on consciously avail-
able material, but most of the workings of reflective cognition are themselves
unconscious (Evans 2009). We can fail to notice this because typical humans
78  
share common capacities to deploy common rules of inference, including for

example simple rules of syllogistic reasoning (Mercier 2011, Mercier and
Sperber 2011), so, in cases where the epistemic standing of one’s premises is
not in question, one can share inferential knowledge of a conclusion by
reciting one’s premises in a way that will equally trigger the subconscious
operation of the same relevant rules of inference in oneself and one’s audience,
to secure the transition to one’s conclusion. Thanks to the typical common-
alities in our rational capacities, explicit reasoning is a powerful device for the
sharing of knowledge: if my epistemic standing on a given point is in initially
in doubt, but I can produce a sound demonstration of that point from
premises known to my audience, reliant on rules of inference they also deploy,
then what I transmit to the audience is not only knowledge of the conclusion,
but also the activation of a way of knowing it. They do not need to trust me, or
see me as knowing, to gain knowledge of the conclusion we arrive at together.
This is one reason why we are so naturally inclined towards explicit argumen-
tation when we encounter resistance from an audience or doubts about our
epistemic standing.
In the face of doubts about our epistemic standing on a given point, we can
engage in explicit reasoning, even if our original judgment on this point was
not made on the basis of reasoning. If I have seen a highly unusual event
outside the deictic sphere of our current conversation, I cannot share my
knowledge of the event by directing my audience’s perceptual attention to it. If
I attempt to transmit my knowledge of the unusual event through simple
testimony, I face the risk of being dismissed as failing to know, given that what
I am reporting will sound unlikely. A better strategy will be to present
whatever evidence I can share to support my claim that the event occurred,
pointing out its causal traces, and so forth, even though my own awareness of
the event was secured through first-hand perception and not inference. Where
a truth is itself known through explicit reasoning, there is little difference
between the strength of one’s dialectical position and the strength of one’s
epistemic position; in other cases, dialectical strength is used as a test for the
presence of knowledge, and there may well be a gap between whether one can
pass this social test for knowledge and whether one actually knows.
There are internalists who make inference essential to justification, seeing
even perception as essentially involving inference from prior evidence. For
example, Stephen Schiffer describes the formation of a perceptual belief (say,
that there is a red cube before him, abbreviated ‘Cube’) in a favorable case
(‘Good’) as follows: ‘in Good I come to be justified in believing Cube by
inferring it from the fact that I am having such-and-such sensory experiences
      79
as of Cube, and I cannot become justified in believing Cube in Good other

than by inferring it from that evidence’ (Schiffer 2009: 198). This position is
problematic in several ways. One might worry that such an inference would
not actually be a very good one: from the fact that I am having sensory
experiences “as of ” the presence of a cube, it does not actually follow that
there is a cube in front of me. Such experiences might be caused by hallucin-
ations or holograms as well as cubes. One might worry about one’s entitlement
to the “as of ” locution, which seems to presuppose some prior acquaintance of
the experiences that correlate with the real presence of cube, where knowledge
of this correlation will be hard to secure. And lastly, one might worry about the
reality of such inferences: if we take it to be a condition on “evidence” that it is
evident in the sense of being present to consciousness, ruling out accounts in
which the relevant inference is subpersonal, then it seems that perceptual beliefs
on Schiffer’s account will only very rarely count as epistemically justified. How
often are we first conscious of having a certain type of “as-if” seeming, and then
only subsequently aware of the outer object, through inference?
By contrast, externalists like Timothy Williamson do not insist that justified
belief is ultimately inferential in character, and do not see basing as necessarily
an explicitly inferential relation. Alongside the well-recognized category of
explicitly evidence-based belief, Williamson introduces a category of implicitly
evidence-based belief, where a belief in p is in this category “if it is appropri-
ately causally sensitive to the evidence for p” (Williamson 2000: 191). Evidence
itself, meanwhile, is not a precursor to knowledge but knowledge itself, and
knowledge does not in general need to be based on prior evidence (or

knowledge). Like many other externalists, he emphasizes the naturally unre-
flective, and non-inferential character of perception: “When we acquire new
evidence in perception, we do not first acquire unknown evidence and then
somehow base knowledge on it later. Rather, acquiring new evidence is
acquiring new knowledge. That knowledge need not itself be based on further
evidence, nor is it evidence for itself in some non-trivial way. But it is evidence
for or against potential answers to questions to which we do not yet know the
answer” (Williamson 2014: 4).⁴ On this approach, which fits and indeed
⁴ One might wonder whether this more recent formulation of Williamson’s view is a strengthening
of the earlier view, according to which a perceptual belief that p could be implicitly evidence-based by
being appropriately causally sensitive to the perceptual evidence for p. On the old position, one could
have perceptual evidence for p without knowing that p, for example if one had misleading evidence
about the source of light. However, even in the new position, which highlights situations in which we
just gain the perceptual evidence that p, thereby coming to know that p, we can still distinguish a
separate notion of “perceptual evidence for p,” where the knowledge we gain speaks in favour of p
without entailing knowledge of p. This kind of knowledge might be especially important in cases where
80  
inspires the view I am defending in this chapter, we are not initially aware of
visual appearances as of the cube, from which some belief about the cube is
then inferred; in the good case, what we are initially conscious of is the cube.
Our first moment of evidence is that there is a cube before us, in the good case.
If we turn introspective, directing attention inward, we are aware that we have
specifically visual experience “as of a cube,” but this is a metacognitive
byproduct of the original unreflective judgment, and not a prior condition of
it. It is true that hallucination would generate a similar metacognitive bypro-
duct, but false that hallucination would make us equally justified. In the bad
case, all we are conscious of is that there is something which appears to be a
cube, but it is part of the badness of the bad case that we will typically judge
that there is a cube present. If we wish to make a judgment which will be
legitimate in both cases, we can restrain ourselves to the more cautious
judgment that a cube seems to be present, but a general strategy of restricting
oneself to such judgments would entirely undermine the value of our world-
sensitive perceptual mechanisms. It is a feature of unreflective cognition that it
does not involve reflection on how we ought to be thinking: in unreflective
cognition such as ordinary perceptual judgment we are focused on the world,
and not on the question of how we should now be making judgments.
Because metacognitive impressions of our judgments play a key role in
defending our epistemic status under challenge, it is not easy to give up the
idea that instances of knowledge always require the backing of prior evidence.
In her defense of fallibilism, Jessica Brown observes that there is something
quite counter-intuitive about positions like the one I am defending here. She
defines both fallibilism and infallibilism in terms of a relation obtaining
between evidence and judgment: what infallibilists hold and fallibilists deny
is that, “knowing that p requires evidence which entails that p” (Brown 2018: 4).
She observes that Williamson’s equation of knowledge with evidence (in
Williamson 2000) gives rise to an unusual form of non-skeptical infallibilism:
because any known proposition is automatically included in one’s evidence
set, the fact that every proposition entails itself will ensure that no proposition
is held without entailing evidence (Brown 2018: 32). She then argues that it is
inappropriate to cite a proposition as evidence for itself (Brown 2018: Ch. 3).
Where the function of evidence is to lend support to a contested propos-
ition, citing that very proposition will beg the question at issue, whether the
we have misleading evidence speaking against p which makes us think inferentially about the question
of whether p, rather than forming a naïve perceptual belief. Thanks to Carlotta Pavese for valuable
discussion on this point.
      81
contest arises from some audience who could have doubts about whether the
speaker knows that p, or whether it arises from the speaker’s own doubts about
her epistemic status with respect to p. So, Brown is right that one’s knowledge
that p is generally not self-supporting in a way which will satisfy those who
have raised doubts about one’s knowledge that p. However, if we can separate
the task of proving one’s epistemic status from the conditions of initially
entering that status, we do not need to insist that every item of knowledge
be backed by prior evidence. If our sensory systems give us appropriate causal
sensitivity to some features of reality, for example, then when these systems
make us aware of those features, we can gain knowledge without prior
supporting evidence.
Brown advances two independent lines of argument against the kind of
strategy I advance at the outset of this chapter, in which ways of knowing
constitute our sources of justification. If I insist that a way of judging that p
only yields full justification when the truth of p is essential to one’s judgment
that p, Brown objects both that our natural ways of individuating belief-
forming processes do not involve this kind of factivity, and also that an
insistence on it will rule out some plausible instances of knowledge, most
notably knowledge gained through inductive inference (Brown 2018: Ch. 2).
On the first point, she grants that English-speakers do use the factive locution
“see that” for successful cases of visual perception, and even that we have some
extended use of it for the detection of mental states (e.g., seeing that another
person is in pain), but she correctly observes that there is no directly parallel
construction for testimony in English, and it is generally strained to apply “see

that” here.⁵ Brown notes that one might say something like “I saw she was
telling me the truth” on some special occasion, but this is not our ordinary
practice in reporting second-hand knowledge, and it would be outright pecu-
liar to say “I saw that Wikipedia was telling me the truth” on having consulted
that source on some minor point of trivia (Brown 2018: 33).
If we take our task to be one of individuating belief-forming processes,
where it is understood from the outset that a belief is an attitude which can
attach to either true or false contents, and where the aim is to predict or
explain the subject’s behavior or claims independently of their environment,
then it is not surprising that the processes we come up with will be of types
that make no essential reference to the relation between the subject and the
environment. It is certainly possible to individuate belief-forming processes on
⁵ However, it is worth noting that constructions of the form “S was told + interrogative” are factive;
as are constructions like “S learned from X that p.” Thanks to Carlotta Pavese for this observation.
82  
the basis of the metacognitive byproducts they produce in us, and we may be
particularly tempted to do this if we approach epistemological questions from
the first-person perspective: from the inside, there is a conspicuous common-
ality between states of seeing and states of hallucinating. Still, anyone
impressed by Jessica Brown’s concerns about what is reflected in natural
languages might worry that similar concerns also arise when we individuate
belief-forming mechanisms from the inside. Do English-speakers commonly
refer to a way of forming beliefs that is neutral between seeing and seeming to
see? The question only becomes more pressing when we look across languages:
for example, the most comprehensive survey of grammaticalized evidentials
(Aikhenvald 2004) reports no language which has an evidential marking that
is neutral between seeing and visually hallucinating.
However, we do not have to frame the problem from the outset as one of
individuating ways of forming beliefs. Another route in to the problem would
start by distinguishing ways of learning about the world; this path can make it
seem more natural to focus on the distinctions between first-hand perception,
second-hand learning through testimony, and sound inference. Even if similar
problems arise on the side of belief-forming mechanisms individuated from
the first-person perspective, it remains a good question why these factive
modes of learning are not all conveniently and transparently lexicalized across
languages, especially given our reasons to believe that speakers of all languages
are generally monitoring the relevant distinctions. One reason may be that our
conversational contributions do not typically assert their epistemic character;
even in languages where assertions must be evidentially marked, the evidential

marking is not part of the at-issue content of the assertion (Simons 2007;
Murray 2017).
When we do make epistemic standing explicit, some ways of learning carry
more weight in common dialectical contexts than others: I can reasonably
expect others to defer to me on commonplace events of which I have first-
hand experience (how do you know that Bill broke the window?—I saw him
do it), but it is less persuasive to respond to challenges by citing sources whose
epistemic credentials are more complex, and are equally open to the challenger
(such as Wikipedia). Still, there is some evidence that we naturally think of
testimony in a way which recognizes the distinction between it and pseudo-
testimony: for example, note that in his initial description of the baseball game
case, Baron Reed describes himself as knowing the outcome, “because my
brother told me what happened.” This way of capturing the basis of the
judgment is not in fact neutral between the good case and the bad: “he told
me what happened” is factive, where “he told me that the Cubs won” is not. In
      83
the case where the Cubs lost, it is not true that my brother who said otherwise
told me what happened. But Reed is not wrong to reach for the stronger
construction in the context of a discussion of the testimonial transmission of
knowledge: the tell-wh construction has a good claim to be our more natural
way of thinking of what is happening when we speak with those we trust.⁶
Likewise, when we use an epistemic modal like “must” to indicate inference,
we are not necessarily indicating any diminished commitment to the main
claim we are making, or admitting the possibility of error in our inference
(Von Fintel and Gillies 2010).⁷ In conditions where the epistemic credentials
of our claims are under challenge, it may be dialectically appropriate for us to
retreat to more neutral formulations concerning what we have been told, or
what we gather may well be the case, but this does not mean that our initial
way of thinking about the origins of our judgments is similarly cautious.
Brown’s concern about inductive inference is harder to address. It may seem
odd to suggest that the progression from observing any finite subset of ravens
to concluding that all ravens are black could be a process which securely locks
us onto the truth, so that the truth of this conclusion is essential to the process.
I will have to insist upon individuating processes so that a different process is
carried out when one samples ravens in a possible world containing a few
unsampled albino specimens, versus a world in which all ravens are black, and
one is subject to no risk. These processes might look the same to the raven-
counter, but one of them involves epistemically important luck in avoiding the
white ravens, where the other was relevantly safe. In my view, the justifying
basis of a judgment is not strictly a function of the choices for which one may
be responsible, or even those together with the subpersonal processes involved
in cognition; the basis includes environmental conditions as well. In this larger
process, the evidence afforded by individual observations plays a crucial role in
legitimizing the ultimate judgment, but it does not exhaust the normative
conditions responsible for what one comes to learn through induction. Proper
treatment of this issue lies beyond the scope of the present work, but I hope to
have given some sense of the limited role I am assigning to consciously
available evidence in my overall account of justification.
Defeat cases pose an especially interesting challenge to my account, exactly
because they seem to leave the basis of a judgment fixed, while toying with the
⁶ I am grateful to Patrick Shirreff and Carlotta Pavese for discussion on this point.
⁷ This point is controversial, as Carlotta Pavese has reminded me. For a contrary view of ‘must,’ see
Daniel Lassiter (2017).
84  
subject’s representation of that basis. The next section takes a fresh look at
these cases.
4.3. Defeat Cases, and Shifts of Basis
Roger White has a sharp statement of the problem that defeat cases pose the
type of view I am defending in this chapter. White is concerned to defend a
‘Cartesian’ conception of evidence, in which ‘the evidence I gain from visual
experience consists in information about how things visually appear to me’
(White 2014: 301). On this conception, one has precisely the same evidence in
seeing or hallucinating. White contrasts his preferred view with ‘Evidence
Externalism.’ according to which ‘the evidence I gain from visual experience
goes beyond information about how things visually appear to me and may
include facts about how things are in my environment’ (White 2014: 301). The
evidence externalist denies that the subject enjoys the same evidence in the
Good case of seeing one’s hand and the Bad Case of hallucinating a hand: the
Good case includes ‘a kind of direct epistemic access to the presence of a hand
before me.’ White then lays out a case in which a person (we will call him
“Hank”) who is actually looking at his hand and is given highly persuasive but
misleading evidence that he is merely hallucinating. The Cartesian view of
evidence has an easy explanation of why Hank seems to lose his justification to
believe that he is looking at his hand, under the pressure of the misleading
evidence. The ordinarily strong relationship between the visual appearance of

a hand and the presence of a hand is screened off by the apparently plausible
misleading testimony about the drug. If the evidence externalist agrees that the
misleading testimony renders Hank unable to have a justified belief that there
is a hand before him, then she will need to explain how the subject’s ‘direct
epistemic access’ to outer world facts has disappeared. White argues that there
is nothing available from the externalist’s depiction of the scenario to explain
this shift:
First, why is it that misleading evidence that I’m hallucinating robs me of this
extra evidence? Nothing appears to have changed about my perceptual state
and connection to the world. Suppose I even go on dogmatically thinking
here’s a hand despite the mounting evidence that I’m hallucinating. I am at
least still seeing my hand. My eyes are still receiving accurate information
from the world and my belief forming mechanisms are operating as usual in
forming a true belief. Isn’t that enough to have direct perceptual access to the
      85
layout of my environment? I also happen to have some misleading evidence

that this is not so which I am blithely ignoring. But it is hard to see how my
possession of such evidence changes my perceptual state in such a way to rob
me of my direct perceptual access to the world. (White 2014: 314–5)
White is right to observe that nothing in Hank’s eyes or environment need

change as the misleading reports about drugging are taken on board. What is
controversial here is his claim that Hank’s ‘belief forming mechanisms are
operating as usual in forming a true belief ’ as he dogmatically affirms here is a
hand, notwithstanding the apparent evidence that he is not in a good position
to judge this question. By the description of the case, Hank is not engaging just
the same belief-forming mechanisms on the question of his hand’s presence
before and after he registers the misleading evidence. In the original and naïve
perceptual judgment, Hank is immediately conscious of his hand. The infalli-
bilist account of the basis of this belief can cite such factors as the proper
functioning of Hank’s visual capacities and their safely true production of a
noninferential judgment. To generate the impression that Hank’s belief is
defeated, note that it is important that he has a properly first-personal
present-tense recognition of the bearing of his misleading reports about his
situation on the judgment he is asked to make. Because Hank’s new evidence
opens the question of how he should think about whether his hand is really
before him or not, he is switched from naïve to reflective cognition on that
embedded question at the same time.⁸ To understand the misleading reports,
Hank must consciously entertain the possibility that he himself has ingested a
drug which produces illusions as of hands; to make matters worse, he has also
been given a persuasive argument that this possibility is actual. If we imagine
him as ‘blithely ignoring’ this possibility and dogmatically judging here’s a
hand, this deliberate evasion of evidence is part of the complete belief-forming
process now sustaining his belief (for another version of this argument, see
Van Wietmarschen (2013: 413–15). If this process now involves thinking
dogmatically and disregarding evidence, Hank is no longer making a judg-
ment in a manner which generally leads safely to the truth: sound naïve
thinking has been replaced by questionable reflective cognition.
⁸ Against this kind of move, Ru Ye urges that “how one’s belief is based is a purely psychological
process, a process about how the belief is formed or maintained, and gaining a piece of evidence needs
not change this psychological process” Ye (2018: 7). I would urge that gaining evidence is also a
psychological process, and one which forces changes in the related psychological process of answering a
suitably related question. Ye maintains that defeat is ultimately best understood as a matter of
intellectual responsibility; this may be right, but one could have an account of responsibility which
includes the specification of responsible processes.
86  
Maria Lasonen-Aarnio offers a rival analysis of these cases. She distin-

guishes between the basis and the method of a judgment, where one’s basis
is the belief-forming process responsible for a belief, and one’s method is the
epistemic rule that one follows. An example illustrates the distinction:
Assume, for instance, that Suzy comes to believe that Tuesday is a rainy day
based on looking outside in the morning. She then immerses herself in a
book, ceasing to consciously entertain the belief, and forgetting all about the
rain. At noon she takes a look outside again and undergoes a new mental
process producing a belief that Tuesday is a rainy day. In this case, she may
well believe that Tuesday is a rainy day by using the same method as before,
but not on the same basis as before. (Lasonen-Aarnio 2010: 5)
In Lasonen-Aarnio’s view, the stubborn subject who retains his belief in the
face of misleading evidence is being unreasonable: the disposition that he is
showing in insisting upon this method of belief-formation does not generally
lead safely to truth, if we consider its application across a wide variety of cases.
However, Lasonen-Aarnio argues, in this particular case the belief is still well-
based and safely true: by stipulation our subject is exercising perception in
favorable conditions, and his method of believing on the basis of perception is
at no risk of delivering error in these conditions. The method is a good one,
given the circumstances, although the subject’s adoption of it shows a dispos-
ition which would in most circumstances be problematic (crime does not
generally pay, but in this case the subject gets away with something). The
subject’s dogmatic belief actually constitutes knowledge, on Lasonen-Aarnio’s
account, and any intuitions to the contrary are to be explained by appeal to a
general heuristic of taking unreasonable beliefs to fall short of knowledge, as
they typically do.
Lasonen-Aarnio describes the stubborn subject as adopting the method of
believing ‘on the basis of perception.’ If the basis of a belief is the whole belief-
forming process responsible for its production, then arguably this moment of
‘adoption’ should itself constitute part of the basis of the stubborn subject’s
belief, a part which distinguishes this basis from that of the naïve subject.
Taking a more restricted view of the basis as naïve perception, the problem is
rather that this basis is no longer an option for the subject, given the stipula-
tions of the case. It is certainly possible for the dogmatic subject to believe
the same proposition that he initially believed on the basis of perception, to
re-endorse the verdict delivered to him by sight, after this verdict has been
brought into question. However, if perceptual judgment is by its nature
      87
noninferential and unreflective, the stipulated presence of conscious contents

that offer rival answers to the first-order question—what is before you is
perhaps not a hand, but a mere hallucination—is enough to ensure that our
subject is not judging unreflectively anymore: he now has to step through the
minefield of these contents to make up his mind. Whether he likes it or not,
the basis of his judgment now includes factors other than the straightforward
causal sensitivity to features of the environment that constitutes ordinary
visual perception. We could perhaps imagine a subject with exceptional
powers of mental self-control, able to induce a selective temporary amnesia
in himself to banish all the misleading evidence from his consciousness. But
the agent who has somehow managed to forget all his defeaters and now looks
innocently at his hand is undefeated; at the moment of judgment he is not
unreasonable but naïve, and once again able to know. Accepting a proposition
unreflectively is not the same psychological process as accepting it in the face
of recognized defeaters.
Becoming reflective on a question ordinarily settled unreflectively does not
automatically trigger epistemic self-sabotage. It is sometimes possible to won-
der whether the lighting is tricky, for example, and collect more evidence to
reassure oneself rationally that it is not. In such a case, one is no longer simply
‘believing on the basis of perception’; one is believing on the basis of conscious
reflection which takes a succession of inputs, including perceptual inputs. As
long as conscious reflection consists in sound reasoning, however, it can still
support knowledgeable judgment. However, when reflection is coupled with
scrutiny of the background assumptions of unreflective thought, in conditions

where that scrutiny cannot be rationally addressed with the materials available
to reflection, we are in trouble. One advantage of understanding defeat cases
this way, rather than Lasonen-Aarnio’s, is that we can accept the common
intuition that defeated subjects lack knowledge; we do not need to concoct an
error theory.
Bob Beddor poses an interesting case, in which a subject (Consuela) sees a
red vase, but is told by a “usually reliable informant” that it is a white vase
illuminated by red light; however, Beddor stipulates, “Consuela (unjustifiably)
regards her interlocutor as completely unreliable; hence, his testimony doesn’t
causally affect her credence [in the proposition that there is a red vase before
her]” (Beddor 2014: 147–8). Beddor argues that this type of case will block
efforts like mine that aim to deal with defeat through the individuation of
cognitive processes. This case is somewhat more pared-down than White’s:
Hank was given highly persuasive testimony, so it would be harder for him to
come across as rational in discounting the odd possibility of a hallucinogenic
88  
drug out-of-hand. Consuela is simply told, falsely as it turns out, that there is a
red light shining on the vase. If she understands this report, even if she does
not accept it, then it is a good question whether Beddor can simply stipulate
that the report has no causal impact on Consuela’s credal state. To understand
the report is to contemplate an explanation of one’s current percepts at odds
with the judgment delivered by naïve perception, at which point one would be
already engaged in reflective cognition; in my view, understanding the inter-
locutor will already suffice for basis displacement. The interlocutor does not
need Consuela to accept that the lighting is abnormal in order to raise the
question of whether the lighting is abnormal, a question that cannot be settled
by naïve perception, but one which bears on the legitimacy of trusting her
initial impression of the color of the vase. If Consuela closes the question about
the lighting with some unjustified reasoning about the interlocutor’s unreli-
ability, then again, her belief formation is compromised. Perhaps there is a way
of understanding Consuela’s attitude to the interlocutor as so dismissive that
she does not even consider the question of whether there actually is something
wrong with the lighting; on this way of reading the case, I would venture, her
unreasonable attitude to this misleading interlocutor could insulate her from
defeat.⁹
So far, I have considered cases involving a switch from unreflective to
reflective cognition. Even within the realm of reflective cognition, we can
generate similar problems by shifting to higher-order reflection. Reflective
cognition itself involves sequentially presented contents, where the global
broadcast of these contents in consciousness enables them to trigger a variety

of modular subpersonal processes, including processes responsible for explicit
reasoning (for a review, see Carruthers 2015). In higher-order reflection, it is
possible to raise the worry that these underlying operations are going awry, so
that one’s ordinary capacity to reason well has been hijacked. David
Christensen (2010) details a case in which one has just solved a simple logic
puzzle, but then is given what seems like excellent evidence that one is under
the influence of a drug which makes one generate inaccurate answers to logic
puzzles of this kind, while leaving one with ordinary feelings of logical
competence, and an intact capacity to understand the puzzles’ initial
⁹ The insulation can only happen under rather special circumstances: perhaps the interlocutor is
Consuela’s little brother who teases her with suggestions she finds too irritating to consider seriously.
Other ways of being dismissive do not leave the epistemic standing of her belief unscathed: if she is
weirdly dogmatic about the color of the vase, for example, then Consuela’s grip on the relevant first-
order truths about the world is no longer safe. Thanks to Mona Simion for helpful criticism on this
point.
      89
parameters. Rethinking the puzzle, Christensen observes, it seems the rational

thing is to lower one’s confidence that the apparently correct response is right.
If the evidence about the drug is misleading, one does still have entailing
grounds for one’s conclusion in the well-understood initial parameters of the
puzzle. However, the misled subject will be compromised in her ability to
exploit those grounds. Ordinarily, rational reflection on the parameters of a
logical puzzle enables the judgment that a certain answer is correct, where this
answer comes to mind in a manner that seems compelling, thanks to under-
lying cognitive processes that execute a valid rule of inference such as modus
ponens. These rules are executed in us by processes which are themselves
unconscious, like the processes supporting syntactic judgments, although
realized in different brain regions (Monti and Osherson 2012). For the subject
who is conscious of the misleading evidence about the effect of the drug—an
effect that produces false but apparently compelling solutions to logical
puzzles—the question of the correct answer to the puzzle has been re-opened,
and the fact that her initial answer to the puzzle seems compelling is now a
warning sign against it. Where she could initially accept the answer on the
basis of logical capacities which safely deliver the truth, where those capacities
were simply exercised on the parameters of the puzzle, she is now challenged
to reflectively endorse that answer while taking on board not only the param-
eters, but also her new apparent evidence against the hypothesis that the
answer which seems compelling is right. The basis of her judgment has a
new level of psychological complexity, and working in this complex manner
with materials available to her, it is no longer possible for our subject to naively
endorse the initial and correct answer to the puzzle, and it is no longer rational
for her to reflectively endorse this answer either. Because her original reason-
ing was sound, the basis of her original judgment has not been rationally
undermined: one could not rationally demonstrate that this flawless basis was
at fault.
There is a certain elegance in understanding defeat as involving a shift in the
basis of one’s judgment: this strategy saves externalists from needing to add a
special internalist patch to their theory of knowledge. The addition of special
patches to cover defeat cases is a longstanding practice in broadly externalist
analyses of knowledge, and a practice which has long been seen as problematic
for these approaches. In the classical Indian Nyāya tradition, for example,
knowledge was analyzed as a ‘truth-hitting cognitive episode,’ born of a
faultless causal or justificatory source, qualified with the extra condition that
the resulting judgment must be ‘non-dubious in the sense that no reasonable
ground for doubting its truth has appeared’ (Matilal 1986: 135). Bimal Matilal
90  
notes that the 11th-century skeptic Śrīharśa was already concerned that the
‘no-defeater’ clause seem oddly redundant, given the condition on the quality
of one’s source (ibid). Contemporary Anglo-American externalists still add
similar ‘no defeater’ clauses to their basic externalism, analyzing knowledge as
something like reliably produced true belief in the absence of defeaters (e.g.,
Goldman 1986). Because the other conditions on his analysis of knowledge are
often weaker, they are not always open to Śrīharśa’s worry about redundancy,
but the motivations of this type of hybrid position remain difficult to explain.
Laurence BonJour raises the worry that any theory of this sort will be an
‘untenable halfway house’ between internalism and externalism (BonJour and
Sosa 2003: 32). When externalists consider what makes a belief normatively
appropriate, the defining feature of their position is that they do not insist that
the subject should always have conscious access to positive justifying grounds.
It is curious, then, for externalists to add a requirement excluding the con-
scious availability of negative evidence. Why should conscious access matter
on one side of the ledger and not the other?
If the misleading apparent evidence in defeat cases has a direct impact on
our need for conscious materials, then there is an explanation of the asym-
metry: by raising concerns about the assumptions of less reflective forms of
judgment, defeat cases shift us upwards into more reflective ways of thinking,
in which we will need consciously available materials to proceed rationally and
make safely true judgments. Psychologically realistic externalists do not have
to see the conscious recruitment of reasons as epiphenomenal in our belief-
forming processes: operations on conscious contents are distinctive causal

processes too. Not only in defeat cases, but across the board, externalists can
insist that subjects need to have appropriate consciously available input for
reflective cognition, given that this type of thinking consists in sequential
operations on what is consciously available. The externalist can consistently
maintain that consciously available input is not needed for our unreflective
forms of judgment such as perception. Maintaining psychological realism,
these forms of judgment do not consist in operations on consciously available
material; they simply generate consciously available results. Knowledge is
generated both reflectively and unreflectively following uniform epistemic
principles: no special criteria are needed to explain when we need reflectively
accessible evidence, or to manage defeat cases.¹⁰
¹⁰ For very helpful written comments on this chapter, I am grateful to Carlotta Pavese and to Mona
Simion. For discussion, I am grateful to David Barnett, Maria Lasonen-Aarnio, Sergio Tenenbaum, and
Jonathan Weisberg. I would also like to thank the Social Sciences and Humanities Research Council of
Canada for funding my research.
      91
References
Aikhenvald, A. Y. (2004). Evidentiality. Oxford: Oxford University Press.

Beddor, B. (2014). Process reliabilism’s troubles with defeat. The Philosophical
Quarterly, 65(259): 145–59.
BonJour, L. and E. Sosa. (2003). Epistemic Justification: Internalism vs. Externalism,
Foundations vs. Virtues. Malden, MA: Blackwell.
Brown, J. A. (2018). Fallibilism: Evidence and Knowledge. Oxford: Oxford
University Press.
Carruthers, P. (2015). The Centered Mind: What the Science of Working Memory
Shows us about the Nature of Human Thought. Oxford: Oxford University
Press.
Christensen, D. (2010). Higher order evidence. Philosophy and Phenomenological
Research, 81(1): 185–215.
Dasti, M. and S. Phillips. (2017). The Nyaya-sutra: Selections with Early
Commentaries. Indianapolis, IN: Hackett Publishing.
Dasti, M. and S. Phillips. (2010). Pramana are factive: A response to Jonardon
Ganeri. Philosophy East and West, 60(4): 535–40.
Dasti, M. (2012). Parasitism and disjunctivism in Nyāya epistemology. Philosophy
East and West, 62(1): 1–15.
Evans, J. (2009). Introspection, confabulation, and dual-process theory.
Behavioral and Brain Sciences, 32(02): 142–3.
Goldman, A. (1979). What is justified belief? In: Justification and Knowledge.
G. S. Pappas (ed.) Dordrecht, Holland: D. Riedel: 1–23.

Goldman, A. (1986). Epistemology and Cognition. Cambridge, MA: Harvard
University Press.
Johnson, M. K., S. Hashtroudi, and D. S. Lindsay. (1993). Source monitoring.
Psychological Bulletin, 114(1): 3–28.
Lasonen-Aarnio, M. L. (2010). Unreasonable knowledge. Philosophical
Lassiter, D. (2017). Graded Modality: Qualitative and Quantitative Perspectives.
Oxford: Oxford University Press.
Matilal, B. K. (1986). Perception: An Essay on Classical Indian Theories of
Knowledge. Oxford: Oxford University Press.
Mercier, H. (2011). On the universality of argumentative reasoning. Journal of
Cognition and Culture, 11(1–2): 85–113.
Mercier, H. and D. Sperber. (2011). Why do humans reason? Arguments for an
argumentative theory. Behavioral and Brain Sciences, 34(02): 57–74.
92  
Monti, M. M. and D. N. Osherson. (2012). Logic, language and the brain. Brain
Research, 1428: 33–42.
Murray, S. E. (2017). The Semantics of Evidentials. Oxford: Oxford University
Press.
Nagel, J. (2015). The social value of reasoning in epistemic justification. Episteme,
12(2): 297–308.
Papafragou, A., P. Li, Y. Choi, and C. Han. (2007). Evidentiality in language and
cognition. Cognition, 103(2): 253–99.
Pomerantz, A. (1980). Telling my side: “Limited access” as a “fishing” device.
Sociological Inquiry, 50(3–4): 186–98.
Reed, B. (2012). Fallibilism. Philosophy Compass, 7(9): 585–96.
Schiffer, S. (2009). Evidence = knowledge: Williamson’s solution to skepticism. In:
Williamson on Knowledge. P. Greenough and D. Pritchard (eds.) Oxford:
Oxford University Press: 183–202.
Simons, M. (2007). Observations on embedding verbs, evidentiality, and presup-
position. Lingua, 117(6): 1034–56.
Smithies, D. (2019). The Epistemic Role of Consciousness. Oxford: Oxford
University Press.
Van Wietmarschen, H. (2013). Peer disagreement, evidence, and well-
groundedness. Philosophical Review, 122(3): 395–425.
Von Fintel, K. and A. S. Gillies (2010). Must . . . stay . . . strong! Natural Language
Semantics, 18(4): 351–83.
White, R. (2014). What is my evidence that here is a hand? In: Scepticism and
Perceptual Justification. D. Dodd and E. Zardini (eds.) Oxford: Oxford

Williamson, T. (2000). Knowledge and its Limits. New York, Oxford University
Press.
Williamson, T. (2014). Knowledge first. In: Contemporary Debates in
Epistemology. 2nd Edition. M. Steup, J. Turri, and E. Sosa (eds.) New York:
Wiley: 1–9.
Ye, R. (2018). Higher-order defeat and intellectual responsibility. Synthese, 197:
5435–55.
5
Dispositional Evaluations and Defeat
Maria Lasonen-Aarnio
5.1. Success and Good Dispositions
Consider the following example, which many would diagnose as involving

defeat by so-called higher-order evidence:
Lobelia and evidence about hypoxia

It’s just the first day, but Lobelia is getting bored on her vacation in the
Pyrenees, waiting in the hut while her friends climb a mountain. To pass the
time, she starts working through some real forensic cases for practice. Her task is
to identify the culprit in a case, given a range of evidence (laboratory evidence,
evidence about the crime scene, etc.). The task is far from straightforward. Lobelia
competently evaluates a case, and forms the opinion that Penelope committed the
crime (p). In fact, she comes to know this. Shortly thereafter one of her friends, a
doctor specialized in expedition medicine, comes in, examines Lobelia, and tells
her that given the altitude and various symptoms she exhibits, Lobelia is suffering
from hypoxia. Such hypoxia impairs one’s reasoning while making it seem
perfectly fine—it is likely, her friend warns Lobelia, that her belief in p is the
output of a flawed process. This time, however, Lobelia’s friend is mistaken:
Lobelia’s original reasoning was impeccable.
Despite being an expert, on this occasion Lobelia’s friend is deeply confused:

though Lobelia does not know this, people just don’t normally develop
hypoxia at her altitude. Let us assume that despite the warnings, Lobelia
retains her belief in p. Many epistemologists think that a variety of familiar
epistemic statuses are defeated in such cases: Lobelia’s belief in p is no longer
rational, no longer justified, and no longer constitutes knowledge.
Yet, as many have argued, familiar stories of these statuses fail to yield
defeat. Of all proposed conditions on knowledge apart from truth and belief,
safety is perhaps the most widely endorsed. But it is far from clear that
093/oso/9780198847205.003.0005
94  -
Lobelia’s belief is no longer safe from error.¹ Nor is it easy to see why it is no
longer produced by a reliable process.² Perhaps more promising would be to
argue for knowledge defeat via justification defeat. A popular way of thinking
about justification is in terms of evidential support. But, it is far from clear
whether Lobelia’s total evidence ceases to support p (e.g., by ceasing to make p
likely): a peculiar feature of higher-order evidence is that it often seems to have
no bearing whatsoever on the relevant first-order propositions.³ And wholly
independently of particular conditions on knowledge, consider the following
verdict: “Lobelia can just see, based on the evidence, that Penelope did it!” To
me this sounds like a perfectly natural thing to say. But, if seeing that is a way
of knowing that, then Lobelia still knows p.⁴ Further, I have argued that there
are principled reasons to think that correct epistemic norms cannot allow for a
systematic phenomenon of defeat by higher-order evidence,⁵ and that there is
a serious tension within views that aim to accommodate it.⁶
My aim here is not to argue that there is no possible theory able to
adequately accommodate and explain defeat by higher-order evidence. The
project of this paper is more positive: to explore a new kind of strategy for
explaining the seeming badness of retaining a belief in the face of evidence that
it is flawed.⁷ Though my main focus will be on defeat, this paper is an
implementation and application of a broader framework. What I see myself
as doing is providing the most promising account of a kind of evaluation
indispensable for making sense of cases ranging from ones involving victims of
massive deceit to so-called Jackson cases—and most importantly for my present
purposes, cases involving higher-order evidence like Lobelia’s.⁸ As I see things,

cases structurally similar to those that originally prompt the search for
more “subjective” or “internalist” kinds of evaluations and norms—evaluations
¹ See Lasonen-Aarnio (2010); Baker-Hytch and Benton (2015).

² See Beddor (2015) for a discussion of process reliabilism.
³ David Christensen (2010) argues that it is peculiar to higher-order evidence that conditionalizing
on it leaves the degree to which one’ s evidence supports the relevant proposition p intact. Indeed, that
is why he thinks that one must “bracket” part of one’s evidence. See Schoenfield (2018) for a discussion
of whether conditionalizing on a self-locating proposition might help. See also Baker-Hytch and
Benton (2015).
⁴ I think this is also a natural thing to say at least in some putative cases of ordinary undercutting
defeat. If Lobelia is looking at a red object, but then acquires misleading evidence about trick lighting, it
seems natural to say that she can still see that the object is red.
⁵ Lasonen-Aarnio (2014). ⁶ Lasonen-Aarnio (2019).
⁷ This diagnosis is continuous with one that I sketch in my 2010 paper, Unreasonable knowledge.
I here outline the approach in more detail and refine it, applying it, in particular, to putative cases of
defeat by higher-order evidence.
⁸ For so-called Jackson cases—and more generally, cases that have been invoked to argue for so-
called subjective oughts—see Lasonen-Aarnio (Forthcoming B). For the New Evil Demon problem, see
Lasonen-Aarnio (Forthcoming A).
    95
normally expressed by epistemologists using the ideology of justification and

rationality—keep cropping up for a wide range of views, and there are prin-
cipled reasons to think that they will keep cropping up.⁹ The inability of various
views of justification to deal with higher-order defeat is, I think, yet another
symptom of this general malaise.
The reason we want to negatively evaluate Lobelia, I will argue, is that she
manifests dispositions that are bad relative to a range of candidate epistemic
successes such as true belief and knowledge. My hypothesis is that, for pretty
much any success, manifesting good dispositions is neither necessary nor
sufficient for success: even knowing is compatible with manifesting disposi-
tions that lead one astray across a range of relevant counterfactual cases. This
general hypothesis, however, is too big a claim to defend in this paper. Rather,
I will argue that given considerations about humanly feasible ways to be,
subjects such as Lobelia who obstinately retain beliefs in putative cases of
defeat by higher-order evidence are naturally construed as manifesting prob-
lematic dispositions. This, I will argue, is ultimately why Lobelia is
criticizable.¹⁰
5.2. Ways and Dispositions
The kinds of evaluations I will be concerned with evaluate doxastic states

(choices, actions, etc.) in a manner that is sensitive to the way in which those
doxastic states are formed and retained (the choices made, the actions per-
formed, etc.). Instead of thinking of these ways in terms of methods, processes,
or bases, I will let the dispositions that manifest themselves as one’s ϕ’ing
(coming to believe something, retaining a belief, making a choice) do the work
of identifying these ways. I will simply assume—like I think everyone should—
that there are dispositions of a wide variety of different kinds, and that
inanimate objects and rational subjects alike are constantly manifesting
them.¹¹
⁹ See Lasonen-Aarnio (2014, 2019, Forthcoming B).

¹⁰ Someone might take manifesting good dispositions to be a condition on knowledge. I won’t here
take on the task of arguing against such views, though see Lasonen-Aarnio (2010).
¹¹ Given a realist background metaphysics of dispositions, thinking about ways of forming and
revising beliefs in terms of dispositions makes progress with the generality problem: there simply is a
fact of the matter regarding what dispositions a subject has and manifests. On my view, the goodness of
a disposition depends on the values of its manifestations across relevant counterfactual cases, and what
might be seen as a version of the problem re-arises in connection with what counts as relevant.
I endorse the context-sensitivity of dispositional evaluations: I think it is a good outcome that our
96  -
Metaphysicians (e.g., Lewis 1986) often draw a distinction between a sparse

and abundant notion of a property; correspondingly, we can distinguish
between a sparse and abundant notion of a disposition.¹² Being formed in
the same or very similar ways should make for a real, objective similarity
between two beliefs. A project aimed at drawing evaluative conclusions about
doxastic states based on what dispositions they are manifestations of is best
served by focusing on sparse dispositions. I will assume that we can find sparse
dispositions at every level of reality, not just the microphysical level. The kinds
of dispositions of interest when evaluating beliefs and choices occupy the same
level of reality as these objects of evaluation—namely, the psychological level,
broadly construed.¹³
My starting point will be that when an action, doxastic transition, or
doxastic state can in some sense be attributed to an agent, it is the manifest-
ation of some of these dispositions. (Sometimes things more or less happen to
us: if, for instance, my mind comes to be controlled by some external force
causing me to be in a neural state that makes for believing p, then, believing p
is not something properly attributable to me.) The claim that, in standard
cases, our actions and beliefs are manifestations of our dispositions, should not
be conflated with the false claims that actions attributable to an agent always
arise out of habit, or that we only ever do things that we are disposed to do. We
sometimes act out of character but are still criticizable in a way that I want to
be able to capture. Consider, for instance, a subject who uncharacteristically
snaps at another person due to being tired and hungry. Even if she is
focus as evaluators can shift what counts as relevant. Further, I don’t think there is any view in
epistemology that can escape all versions of the generality problem.
¹² Abundant properties come cheap. As a general rule, there is one corresponding to every coherent,
non-paradoxical predicate of a language. They can be wildly extrinsic and unnatural, and any two
things share an indefinite number of them. They are closed under various logical operations, satisfying
various comprehension principles. For instance, if there is the property of being an F, there is also the
property of being either an F or a G, and the property of not being an F. By contrast, from the fact that
F and G are sparse properties, we cannot trivially infer that having either F or G is a sparse property;
that having F and G is a sparse property, or that not having F is a sparse property. For arguments that
we should distinguish between fundamental properties pertaining to a fundamental level of nature
(assuming there to be one) and sparse properties, see Schaffer (2004).
¹³ Corresponding to these different levels of reality are different levels of explanation. Dennett (1969:
93) famously distinguished between personal and subpersonal levels of explanation, taking both to be
psychological explanations. Dennett’s concern was with explanations of behavior, but such a distinc-
tion could be extended to explanations of belief. Following Dennett, we could distinguish between
personal and subpersonal dispositions. Dispositions that work at the personal level are dispositions of
(entire) persons. By contrast, subpersonal dispositions are dispositions of subsystems or parts of
persons. There is much controversy about whether subpersonal explanations are really psychological
in the first place (for a good overview, see Drayson, 2014). I will remain neutral on these issues. The
dispositions I am interested in are psychological, whether or not this means that they must be
dispositions of persons.
    97
uncharacteristically irritable, she might still be disposed snap at others in

circumscribed conditions, when very tired and hungry, or perhaps she is
disposed to become irritable (acquiring a new disposition) in these conditions.
But further, doing something can originate in and be a manifestation of one’s
dispositions even in the absence of such more local or circumscribed disposi-
tions: one might do it because of the way various specific situational factors
interact with one’s dispositions.¹⁴ For instance, a combination of certain
company, an uncommonly boisterous mood, and some recent life-changing
news might activate various dispositions a subject has (to try new things, to do
things that are putatively good for her health, etc.), while masking others
(having to do with avoiding discomfort). In these circumstances, jumping into
an icy lake can be a manifestation of the subject’s dispositions, even if she is
disposed to avoid cold water. In this way, even doing things that one is not
disposed to do can be a manifestation of one’s dispositions.
These points are related to an important way in which my dispositional
evaluations differ from virtue-theoretic ones appealing to the notions of virtue
and competence. Even if we think of competences and virtues dispositionally,
the associated dispositions are general, complex, and entrenched, impinging
on a very wide range of different situations. In the case of moral virtue, for
instance, what is at issue is character traits bearing on how one responds to
and weighs reasons, how one feels, acts, deliberates, and reacts; what one
notices and attends to, desires, and values. Subjects can on particular occasions
act excellently, in ways that an evaluative perspective oriented on good
dispositions should be able to pick up on, without manifesting anything like

full-blown virtue or competence. An unkind person might have a soft spot,
responding on this occasion in a commendable way to the needs of those
around her. Similarly for a person who is only in the process of learning to be
kind—she might only unreliably be able to manifest dispositions to act kindly.
Even when people act out of character, or fail to manifest full-blown compe-
tence or virtue, we can ask how good the dispositions they manifest on this
occasion are.¹⁵
Dispositional evaluations share important features with evaluations stand-
ardly expressed by epistemologists using the ideology of justification or
rationality and differ from evaluations having to do with mere blamelessness
¹⁴ Cf Sher (2006: 23), and the range of cases he discusses.

¹⁵ Though I am not offering a theory of blameworthiness, for similar reasons, I am very sympathetic
to Sher’s (2006: Ch. 2) criticism of a Humean theory on which one can only be blameworthy for doing
something bad if it originated from a vice or defect in one’s character, thought of as some more general
disposition to act in bad ways. See also Hurka’s (2006) criticism of contemporary virtue ethics.
98  -
or excuses. First, the evaluative perspective I outline forthwith can be directed

at an agent’s acts in a broad sense of the term: for instance, one’s coming to
believe a proposition—and hence, one’s belief—can be positively or negatively
evaluated depending on whether it is a manifestation of good dispositions
(which is not to say that a person cannot be criticized for manifesting bad
dispositions). By contrast, blame is a reaction to a person on the basis of what
they have done, or on the basis of their character traits; it is persons who are
blameworthy or blameless.¹⁶ Second, manifesting good dispositions is not
merely a matter of a negative evaluation being inappropriate, in the way that
blamelessness is. I share a common dissatisfaction with appeal to blameless-
ness and excuses in defense of more externalist or objectivist norms, which is
that the tool is too blunt to carve out distinctions we want to make: our
reactions to a range of cases seem to involve a kind of positive evaluation
distinct from a mere failure to blame an agent.¹⁷ Blamelessness is disunified in
the sense that there are multiple different ways of being blameless, and
manifesting good dispositions is but one. Further, just as being positively
evaluable from the dispositional perspective is not a matter of mere blame-
lessness, being negatively evaluable is not a matter of blameworthiness. Agents
who manifest problematic dispositions are often, but not always, blame-
worthy. For instance, had Lobelia grown up in a community that instilled in
her bad dogmatist habits under the guise of epistemic virtue, or had she been
cognitively impaired in a way that excused her dogmatic habits, she might be
blameless, but she would still be manifesting problematic dispositions.¹⁸
Let me now outline in more detail how I think of good dispositions.
¹⁶ Cf Sher (2006: 7). Just as we may blame the hurricane for the low voter turnout, we may blame
one’s way of forming a belief for the falsity of the belief, but this kind of blaming is standardly taken to
be is essentially different from e.g., moral blame directed at persons (cf. Sher 2006: ix)—it is not
normative in a broad sense of the term.
¹⁷ This is a widespread criticism; for a recent elaboration in epistemology, see Simion, Kelp, and
Ghijsen (2016).
¹⁸ What about praise? The objects of praise appear to be broader than those of blame. We often
praise ways of acting or believing: we might praise Mahler for his orchestration, but we can also praise
the orchestration itself, just as we can praise the technique displayed by an expert archer’s shot. Indeed,
by praising the orchestration in a Mahler symphony, we bestow praise on the symphony itself, and by
praising the technique displayed by a shot we praise the shot. One reason I am cautious, however, of
expressing my view in terms of praise is that where failing to merit blame is not positive enough,
meriting praise seems excessively positive. A composer’s orchestration might not be praiseworthy, but
not criticizable or negatively evaluable, for it might simply be perfectly fine. By contrast, subjects who
fail to manifest good dispositions in the sense I will outline below are criticizable. Relatedly, it is not
clear, for instance, whether it is appropriate to praise the forming of an ordinary perceptual belief in
good conditions, or other routine acts that require little expertise or effort. Nevertheless, such beliefs are
positively evaluable, being paradigmatic examples of beliefs labelled justified or rational.
    99
5.3. Dispositional Evaluations
We form beliefs, normally retain them for some period of time, and often
retrieve them from memory. One might, for instance, originally form a belief
about a particular event through an exemplary perceptual process, but the
belief about the event retrieved from memory might involve misremembering
or confabulation. Conversely, one’s memory about a past event might corres-
pond to one’s past experience, but the experience itself may have involved
misperception that manifests bad dispositions, such as cognitive biases. Hence,
our focus should not be exclusively on belief-formation, on acts of coming to
believe. Especially important in connection with discussions of defeat will be
ways of retaining beliefs in light of new evidence. Despite the good credentials
of one’s original belief, the dispositions at play in its retention might be
problematic. Indeed, this will be my diagnosis of what happens in many
putative cases of defeat. Hence, for a belief to be evaluated positively from
the dispositional perspective, the formation of the belief must manifest good
dispositions, and similarly for its retention and possible retrieval from
memory.
Assume that a subject ϕ’s (e.g., comes to believe a proposition p, retains
belief in p, or retrieves a belief in p from memory), and that her ϕ’ing is the
manifestation of some disposition D.¹⁹ What makes D good? Goodness is
always relative to some success: the dispositions at play when one comes to
believe that one’s lottery ticket will lose just based on the odds may be good if
epistemic success is a matter of believing truly, or believing what is likely on

one’s evidence, but not if it is a matter of knowing. The goodness of a
disposition will depend on how successful its manifestations are across rele-
vant counterfactual cases in which it manifests itself—more precisely, on the
values of its manifestations in counterfactual cases, weighted according to
relevance. Given a weighting of counterfactual cases in which a disposition
manifests itself according to relevance, and a value function, we can determine
an overall score for the disposition. We then need to say how the overall
goodness of a disposition is related to its score. On the account I favor,
manifesting good dispositions is a matter of manifesting the best feasible
alternative dispositions, or at least ones that compare favorably enough with
the best feasible alternatives. One’s ϕ’ing is to be evaluated positively just in
¹⁹ For the purposes of stating my view, I will assume for simplicity that there is just one disposition.
It is, however, more realistic to assume that one’s ϕ ’ing is the joint manifestation of multiple
dispositions. In that case, the relevant cases will be ones in which all of these dispositions jointly
manifest themselves.
100  -
case it is a manifestation of the best (or close enough to the best) feasible
alternative dispositions. I will take up the different components of the account
in turn. I will then say how the role played by feasibility in my account allows
for taking shortcuts in making dispositional evaluations.
Relevance: I think that dispositional evaluations are highly malleable and
context-dependent: what counts as relevant can easily shift depending on our
focus and context as evaluators. Nevertheless, there are some general struc-
tural points to be made. The first thing to emphasize is that relevance should
not be understood in terms of any relation that a case trivially bears to itself
(to a maximal degree). In particular, relevance is not a matter of relevant
similarity: counterfactual cases are not weighted according to how similar they
are to the actual case. Similarly, relevance is not a matter of what could easily
have occurred. This is because the case a subject is actually in might be deviant
or abnormal and as a result, it—and cases very much like it—may be irrelevant
when evaluating the goodness of a disposition. I won’t here offer a theory of
what such normality consists in, but even lacking a precise account, we have
some initial grasp on what sorts of cases count as normal. Various circum-
stances involving highly misleading evidence, for instance, are abnormal: it is
abnormal to encounter a real-looking fake tree among real ones, to experience
an intricately crafted perceptual illusion with no hints that one is experiencing
it, and so on. The dispositions manifested by a subject who comes to form a
belief on the basis of such an illusion might be good, for across somewhat
normal cases they manifest as knowledge-constituting (or otherwise epistem-
ically successful) belief.²⁰

Here is a toy model that at least comes close to how I think about relevance.
As evaluators, we consider some features of a case to be idiosyncratic, while
holding others fixed. We can think of the features held fix as determining a
contextually determined type of case. Relevant cases are somewhat normal or
typical instances of the type. The relevance weighting might be binary, each
normal case instantiating the relevant type getting the same, positive weight.
Or, relevance might come in degrees, a case being assigned more weight the
more normal or typical it is. In the case of Lobelia, for instance, a very natural
type to focus on is Lobelia receiving a certain kind of higher-order evidence
regarding her own cognitive functioning from a certain kind of expert. We
²⁰ In my treatment of the new evil demon problem, I have argued that what counts as normal often
depends on the kind of world or case that a subject is anchored to—she might in fact be in a case that is
abnormal for her (see Lasonen-Aarnio, Forthcoming A). Hence, “alien cognizers” living in environ-
ments that are abnormal by our lights can be evaluated positively from the dispositional perspective, for
their dispositions fare well in counterfactual cases that are not too abnormal for them.
    101
then consider other counterfactual cases falling under this type in which the
same dispositions that Lobelia actually manifests manifest themselves.
Scores: the score of a disposition, representing how well the disposition does
across counterfactual cases, is a function of the weighting of these cases
according to relevance, and a value function. I will assume that the score of
a disposition is a weighted average of the values of its manifestations in
counterfactual cases, the weightings being by relevance. (I think this is a
natural view, though it is not the only possible one.)
The nature of the value function will, of course, depend on what the relevant
success under consideration is. For some successes all that matters is whether a
manifestation is a success or failure—coming close to succeeding doesn’t get
assigned any extra value (e.g., penalty shots in football). In other cases, success
itself comes in degrees. Consider, for instance, throwing darts, where one gets
more points the closer the dart lands to the central region. And even if
knowledge, for instance, itself is an all-or-nothing matter, one might think
that a value function should assign less value to a belief that is false than to a
belief that is merely true, but doesn’t constitute knowledge. I return to these
issues below.
It is worth noting that even fixing the success being focused on, the
framework sketched is compatible with the value function shifting depending
on the situation under consideration—being sensitive, for instance, to what is
practically at stake for a given subject. Compare a subject in a low-stakes
situation in which the practical cost of holding a false belief is not great, and a
subject in a high-stakes situation in which false belief has dire practical

consequences. The serious cost of false belief might incline us, as evaluators,
to assign greater disvalue to false belief. The result of this may be that whereas
disposition D₁ gets a higher score than D₂ when considering the low-stakes
situation, this order is reversed when considering the high-stakes situation. As
a result, though D₁ might be better for a low-stakes subject, D₂ will be better
for the high-stakes subject. This is one of the ways in which the framework can
accommodate a stakes-sensitive kind of epistemic evaluation.
Goodness and feasible alternative dispositions: how does the score of a
disposition determine its overall goodness? There are various approaches one
could take here. On one possible view, goodness is a matter of having a score
that is high enough in absolute terms. According the approach I favor, dis-
positional goodness is a matter of doing the best one feasibly can, or at least
coming sufficiently close: whether one manifests good dispositions depends on
what feasible alternatives there are. If D₁ and D₂ are both feasible alternatives
in one’s situation, then D₁ is better D₂ than just in case it has a higher score.
102  -
Just how close a disposition must be to the best feasible alternatives might
depend on context.
There are various constraints, of different strengths, on how we could be
disposed in the first place, and on what dispositions could be manifested in
specific circumstances. Some are logical, some nomological, some metaphys-
ical. Of particular importance below will be feasibility constraints flowing from
broad features of our cognitive design: as a general rule, feasibility is con-
strained by our cognitive architecture, allowing for some idealizing.²¹ Hence,
feasible alternative dispositions are dispositions that could, in a relevant sense
of ‘could’ determined by context but constrained by our cognitive architecture,
be manifested in the situation under consideration. An alternative disposition
is one that would manifest itself as a relevant doxastic state, choice, or action in
one’s situation. For instance, relevant alternative dispositions in the Lobelia
case manifest in the situation at hand as some doxastic state regarding
proposition p.²² (It is feasible in Lobelia’s situation to manifest a disposition
to scratch one’s head, but this disposition is not an alternative to the doxastic
dispositions Lobelia manifests when retaining her belief in p.)
Of particular importance below will be considerations of feasibility regard-
ing dispositions that discriminate between various cases. For instance, the
rationale for safety precautions that we routinely take is that we cannot be
disposed to only take those precautions in situations in which they are not
redundant. We cannot be disposed to only use safety belts on car rides that end
in crashes, or to buy insurance only for trips that end up involving some sort of
calamity. But there are other, perhaps more interesting cases, in which our
dispositions cannot pick up on certain features of the case we are in, even if the
good and bad cases differ in some respects “from the inside”. For instance,
climbers are taught to check their harnesses and knots every time before
relying on them. The aim is to be almost automatically disposed to go through
certain checks in every situation. It would of course be more efficient to only
²¹ Consider the literature on heuristics and biases—for instance, the systematic patterns of fallacious
probabilistic reasoning exhibited by humans. Psychologists tend to think that such errors flow from
features of our cognitive design, which consists of different systems tacked together (e.g., what is known
as System 1 and System 2). If a subject commits the so-called bankteller fallacy, thinking a conjunction
to be likelier than its individual conjunct, then even if she manifests some dispositions typical of human
beings, she is not manifesting good dispositions: there is a better available disposition—indeed, we are
sometimes able to pick up on such errors ourselves.
²² Assume that the subject in fact manifests disposition D₁. Then, in the actual case she manifests D₁,
and no other disposition. Nevertheless, we can ask whether some other disposition that could have been
manifested is better. This is why I talk about situations: when asking what the best feasible dispositions
are in one’s situation, we keep fixed other features about the actual case, but not facts about what
dispositions (if any) one manifests. My account also assumes counterfactual facts regarding what
doxastic state a given disposition would have manifested as in one’s situation.
    103
double check when one has made some sort of error, but the problem is that in
situations in which one has made an error, and in which double checking
would reveal (or would at least be likely to reveal) the error, the error has
already gone undetected until the procedure of double checking. These are, of
course, often precisely the kinds of situations in which one’s cognitive skills are
somehow compromised due to fatigue, altitude, dehydration, etc. In order to
be disposed to double check only when one has committed an error, one
would have to be sensitive to having committed the error. But how could one
be thus sensitive when the error was undetected in the first place, often due to
compromised cognitive skills?
Cases involving higher-order evidence that one’s doxastic state is flawed are,
I will argue, similar in many respects. The thought will be that given natural
assumptions about human psychology, it is plausible that in a wide range of
somewhat paradigmatic cases involving putative higher-order defeaters, a
subject cannot be disposed to retain belief in light of higher-order evidence
pointing to some sort of cognitive error only when no such error has in fact
occurred. It is not humanly feasible to have retention-dispositions that dis-
criminate between cases in which the higher-order evidence is misleading and
cases in which it is not. As a result, the dogmatic subject’s dispositions
manifest as retaining belief whether or not the higher-order evidence is
misleading. Her dispositions are problematic, for they will lead her to retain
a botched belief—a belief that fails to constitute knowledge, and that fails to be
appropriately based on one’s reasons or evidence—across many relevant
counterfactual cases. This is why she is negatively evaluable: there are signifi-
cantly better feasible dispositions.
According to the evaluative standard I have sketched, then, a doxastic state
is positively evaluable just in case it is a manifestation of dispositions that
compare favorably with feasible alternatives. A subject is criticizable if she
manifests dispositions that are is worse (or significantly worse) than feasible
alternatives. I think that evaluations that thus take into account feasibility
constraints arising, in particular, from a subject’s cognitive architecture are the
most promising for capturing a more subjective dimension of evaluation
across both the theoretical and practical realms—a dimension often picked
out by talk of justification, or of a subjective sense of ‘ought’.²³
That is not to say, however, that how good a disposition is in absolute terms
is of no evaluative significance. Assume that a subject succeeds at something
(for instance, at goaling in football, or at knowing). We can now ask: given the
²³ See Lasonen-Aarnio (2019).
104  -
dispositions manifested, would she have succeeded across a sufficiently wide

range of cases involving somewhat normal variations of the idiosyncratic-
seeming features of her situation? If so, the relevant dispositions have a kind of
non-accidental connection to success: the subject’s success does not essentially
hinge on idiosyncrasies of her situation.²⁴ The non-accidentality of a success
seems important, in particular, for being praiseworthy or creditworthy for
succeeding.²⁵ Hence, not all cases of manifesting the best feasible dispositions
are on a par.
I claimed at the outset that we often deploy the kind of dispositional
evaluative perspective I have outlined. But at this point the reader might
worry whether this is so, for occupying the perspective requires making
judgments we rarely make or are in a position to make. For instance, it
seems to require tracking exactly which dispositions a belief manifests. But
surely, the thought goes, we have very little access to the fine details of another
subject’s psychological processes. Moreover, the evaluations involve engaging
in complicated calculations involving indefinitely many counterfactual cases.
In response to these worries, let me point to various shortcuts one can deploy
to make dispositional evaluations.
First, we often have plenty of knowledge about how good the dispositions
manifested are, even if we are unable to identify the exact dispositions at play.
If, for instance, a belief was formed as a result of ordinary perception, we don’t
need to be able to identify the exact dispositions, many of which might be sub-
personal, to be able to judge that the dispositions involved are good: we have,
after all, plenty of knowledge about how ordinary perception fares across
counterfactual cases.
Moreover, it is important to see that considerations of feasibility alone can
help determine that a subject’s belief manifests bad dispositions. Consider the
following simple case. You observe a subject not putting on a seatbelt. You are
not sure exactly what dispositions she manifests, and just how circumscribed
they are. (Perhaps, for instance, the subject is only disposed to forget when
very distracted.) Nevertheless, you know that certain possible ways to be—
certain possible dispositions to have—are just not ways that ordinary human
²⁴ Virtue epistemologists often talk about success being creditable to a subject: a subject succeeding
in virtue of, or because of her competence. Knowledge, then, is belief that is true in virtue of being
competent. As I am thinking about things, a belief can constitute knowledge, and be evaluated
positively from the dispositional perspective, even if the fact that it constitutes knowledge is largely
explained by somewhat invariant features of one’s situation. I think this is plausibly true in many cases
of testimony—indeed, testimony has been raised as a problem for virtue epistemologists (e.g., Lackey
2007).
²⁵ See Lasonen-Aarnio (Forthcoming B).
    105
beings could be. It is not feasible to be disposed to not put on a seatbelt only
when the ride won’t end in a crash. So, whatever the dispositions in fact being
manifested are, it is reasonable to assume that they cannot discriminate
between cases in which the safety belt will prove useful and cases in which it
won’t. But dispositions that are thus indiscriminate between good and bad
cases are problematic; it would be better to take the safety precaution and be
systematically disposed to put on a seatbelt. Note also that this judgment does
not depend on actually calculating the precise scores of the no-seatbelt and
seat-belt dispositions, and then comparing these. We simply know that when
crashes do happen, the outcomes for subjects wearing seatbelts tend to be
much better, and that there are no significant harms to wearing seatbelts.
To sum up: we often have a good idea of what kinds of dispositions are
humanly feasible. We often have a good idea of what kinds of humanly feasible
dispositions a subject’s belief (choice, action) could be a manifestation of. If all
of the feasible dispositions are dominated by (significantly) better alternatives,
then this is all we need to know in order to correctly negatively evaluate her
belief (choice, action) from the dispositional perspective.
Let me now finally apply dispositional evaluations to Lobelia’s case.
5.4. Putative Defeat and Dispositional Discrimination
Consider again the example Lobelia and evidence about hypoxia presented in
the very beginning of this paper. Assume that despite the testimony of her
friend, Lobelia retains her original belief. In what follows, I will evaluate
Lobelia’s dispositions relative to the success of knowing, but parallel points
could be made given other kinds of epistemic success, such as true belief.
(Indeed, the reasoning in the section that follows is neutral between veritism
and gnosticism.)
What dispositions does Lobelia’s belief manifest? Some of these are no
doubt excellent: her belief was originally formed in an exemplary way. But
as pointed out previously, we must also evaluate the dispositions at play when
Lobelia retains her belief in light of expert testimony that it was formed by a
flawed process. In fact, we don’t need to know exactly what dispositions
Lobelia manifests, for considerations of feasibility provide a shortcut. Begin
with the following question: could Lobelia be disposed to retain belief in the
face of similar evidence (e.g., expert testimony) about her cognitive malfunc-
tioning only when such evidence is misleading? It is plausible that she could
not, if she is an ordinary human being. Consider, in particular, counterfactual
106  -
cases in which Lobelia has similar higher-order evidence that is not mislead-
ing, cases in which she holds a belief formed through a flawed process that
nevertheless seems to her to be perfectly fine. Being disposed to only give up
belief in response to such higher-order evidence when that evidence is not
misleading would require Lobelia to have dispositions that discriminate
between cases in which her original cognitive process is fine, and cases in
which it merely seemed to be fine. But such dispositions are not humanly
feasible.
Someone might object that by retaining her belief, Lobelia is acting in a way
that is highly atypical of human subjects. In fact, most of us might be
psychologically incapable of retaining belief in Lobelia’s situation, and for
this reason Lobelia’s case might seem puzzling. But a natural response to
such puzzlement is to start looking for explanations of what might make it
feasible for Lobelia to retain her belief. Perhaps Lobelia has a dogmatic strain
resulting from being overconfident in her own epistemic abilities. Or perhaps
she has a high need for closure, tending to seize on the first piece of evidence
available, and to then freeze her opinion, being more generally indifferent to
subsequent evidence. The fact that Lobelia’s obstinacy is atypical seems to only
strengthen my case, for all such explanations involve problematic dispositions.
Hence, whatever the precise dispositions manifested by Lobelia are, it is
natural to take her to manifest a kind of obstinacy that is indiscriminate in
problematic ways—obstinacy that manifests itself as retaining a belief formed
by a flawed cognitive process in cases in which her higher-order evidence is
not misleading.
Though I cannot defend these further claims here, on my view Lobelia can
still know that Penelope committed the crime. (After all, she can just see that
this is what the evidence points to—there is nothing whatsoever wrong with
her cognitive faculties!)²⁶ Further, perhaps she could even come to know that
the testimony of her friend is on this occasion misleading. She might, for
instance, reason that if she was hypoxic, it would be very unlikely for her to
have come to truly believe p, and yet, she does truly believe p. If Lobelia knows
that her higher-order evidence is misleading, then there is a perfectly good
sense in which she can discriminate the case she is in from those in which she
²⁶ Hence, when it comes to judgments about knowledge, my view is error theoretic: many epistem-
ologists have misdiagnosed cases like Lobelia’s as ones involving knowledge defeat, falsely inferring
failure to know from the manifestation of problematic dispositions. I can here only gesture toward an
error-theoretic explanation, but the rough idea is that in paradigm cases knowledge is accompanied by
good dispositions—indeed, in paradigm cases, good dispositions play an important role in explaining
knowledge.
    107
is in fact hypoxic, or the victim of some other cognitive defect. Call such
discrimination epistemic. This is compatible, I think, with Lobelia not being
able to discriminate her case and cases in which she is hypoxic in a disposi-
tional sense. It is not feasible, I have argued, for Lobelia to only be disposed to
ignore the kind of higher-order evidence she has in cases in which it is
misleading, and her original cognitive process was in perfect order. In retain-
ing her knowledge, Lobelia is manifesting dispositions that fail to discriminate
between cases in which her cognitive process was flawed and cases in which it
was not. As a result, the very dispositions she manifests lead her to retain belief
formed by a botched process across a wide range of counterfactual cases.
Epistemic access or discrimination does not entail dispositional doxastic
discrimination, or vice versa.
It is worth emphasizing that the mere fact that our doxastic dispositions
indiscriminately issue false beliefs in some counterfactual cases may be of no
significance, for these cases might not be relevant. For example, assume that
I come to believe the there is a tree just outside the window. Now consider
counterfactual cases in which I am looking at a very realistic-looking tree
replica instead. My dispositions cannot discriminate between cases involving
real trees and ones involving very realistic fakes. Nevertheless, my belief may
still be dispositionally good, manifesting the best (or close enough to the best)
feasible dispositions, for recherché cases involving highly realistic tree-replicas
planted outside windows are just not relevant. What is different about putative
cases of defeat by higher-order evidence is that in such cases subjects have
fairly strong evidence that their cognitive abilities are malfunctioning or, more
generally, that a doxastic state they are in fails by some normative standard. It
is very natural to evaluate Lobelia’s belief, now retained in light of the higher-
order evidence, by looking at other, somewhat normal cases involving such
evidence. (Indeed, epistemologists tend to talk about cases involving higher-
order evidence as forming one epistemically salient kind.) There is nothing
abnormal about such higher-order evidence not being misleading. In fact, if
anything, it is more normal for expert testimony not to be misleading than for
it to be misleading.
In sum, I have argued that by retaining her belief despite the expert
testimony that it is the output of a flawed process, Lobelia manifests disposi-
tions that fail to discriminate between cases in which her higher-order evi-
dence is misleading and cases in which it is not. As a result, these dispositions
manifest as retaining a belief that is in fact the result of a flawed cognitive
process across a wide range of relevant counterfactual cases—namely, cases in
108  -
which the higher-order evidence is not misleading. I will now argue that there
are better feasible alternative dispositions.
5.5. A Better Feasible Disposition
There are two salient kinds of dispositions that a subject in Lobelia’s situation
could manifest. First, she could retain her true belief (and, on my view,
knowledge), thereby manifesting dispositions that would have her retain
flawed beliefs in other relevant cases involving higher-order evidence.
Second, she could give up her belief by suspending judgment, manifesting
dispositions that in other relevant cases manifest as giving up belief,
whether or not the belief was formed by a flawed process. For ease of
exposition, in what follows I will simply speak of a “retention-disposition”
and a “suspension-disposition”. Many epistemologists have urged that a
subject in Lobelia’s situation ought to give up her belief. I will now argue
that given plausible assumptions about the value of knowledge/true belief and
the disvalue of false belief, the best feasible dispositions manifest in Lobelia’s
situation as suspending judgment. This is so even if we concede, as I think we
should, that Lobelia can retain knowledge.
We will evaluate the retention-disposition by only looking at cases in which
Lobelia retains her belief and evaluate the suspension-disposition by only
looking at cases in which Lobelia suspends judgment. Nevertheless, the rele-
vant counterfactual cases are otherwise similar: for instance, those assigned
some significant weight, or considered relevant, are ones in which Lobelia gets
higher-order evidence that is similar to the evidence that she acquires in the
actual case. Whichever disposition we are evaluating, we can divide the
relevant counterfactual cases into three classes:
(1) Lobelia’s initial belief is the result of a good doxastic process, p is true,
and the higher-order evidence is misleading.
(2) Lobelia’s initial belief is the result of a flawed process, but
nevertheless true.
(3) Lobelia’s initial belief was false (whether as a result of a good or bad
doxastic process).
Instead of assigning each individual case a weight, for each disposition, we

can just divide all of the counterfactual cases into these three classes according
to which kind they fall under, and assign these classes weights (I will assume
    109
that these weights sum up to 1). Assuming each relevant case to have the same
weight, we can think of the weight assigned to the first class of cases, for
instance, as corresponding to the proportion of counterfactual cases in which
Lobelia’s higher-order evidence is misleading. Moreover, it is plausible that
these weights should be the same whether or not we are evaluating the
retention-disposition or the suspension-disposition: the weighting of counter-
factual cases should depend on issues like how normal it is for expert testi-
mony of the sort that Lobelia receives to be misleading.
In order to score the two dispositions, we need to determine the values of
believing p (Bp) and not believing p (~ Bp) in cases falling into each of the
three classes. I have assumed that if Lobelia is able to retain her belief, then the
higher-order evidence need not destroy her knowledge. I will assume that if
Lobelia retains belief in cases that fall under (1), she continues to know (this
will just make it more difficult, not easier, to argue that the suspension-
disposition is better).²⁷ In the following table, ঋ denotes the value of knowing,
ঔ the value of truly believing, and Àআ the disvalue of falsely believing. Not
believing (for instance, suspending judgment), I will assume, has a neutral
value 0 across the board.
Kind of case U (Bp) U (~Bp)

(1) HOE misleading, p ঋ 0
(2) HOE not misleading, p ঔ 0
(3) ~p Àআ 0
There are different ways of thinking about the value of true belief versus the
value of knowledge versus the disvalue of false belief. According to veritists,
true belief is valuable, and there is no added value to knowing; knowing
inherits its value from the value of true belief.
Veritist evaluations : ঋ ¼ ঔ>0> À আ
According to gnosticists, knowledge is valuable, and more valuable than true

belief. Gnosticists have various options when it comes to the value of true
belief.²⁸ They nevertheless all agree that knowledge is more valuable than mere
true belief, and that mere true belief is no worse than false belief:
²⁷ One might object that there are some cases in which p is true, and the HOE is misleading, but
Lobelia nevertheless does not know p, due to being in some sort of Gettier case. I agree but relaxing this
idealization would only make a suspension-disposition better, not worse.
²⁸ See Dutant and Fitelson (Unpublished manuscript).
110  -
Gnosticist evaluations : ঋ > 0 > À আ; ঋ > ঔ Àআ
All of this, I take it, is rather uncontroversial. I will also make the following
assumption: ঋ < আ. That is, the disvalue of false belief is greater than the value
of true belief or knowledge.²⁹
Whether one is a veritist or a gnosticist, it follows from these assumptions
that if the weight assigned to (3) is at least .5, then not believing does better
than retaining belief across the relevant counterfactual cases. Assume for
simplicity that each individual case is assigned the same non-zero weight.
We can then rephrase the point as follows: if at least half of the relevant
counterfactual cases are ones in which Lobelia’s initial belief is false, then
suspending judgment is a better overall strategy—that is, the suspension-
disposition is better. And it is plausible that Lobelia’s initial belief is false in
at least half of the relevant counterfactual cases, cases in which she gets
evidence akin to the testimony of an expert that her cognitive process is
flawed. Assume, for instance, that in 90 per cent of relevant counterfactual
cases, Lobelia’s higher-order evidence is not misleading: she is in fact hypoxic
(or suffers from some other defect). In those cases, Lobelia’s dispositions will
manifest as retaining a flawed belief. In none of those cases does the belief
constitute knowledge (presumably, even true beliefs that are outputs of flawed
cognitive processes don’t constitute knowledge). Moreover, Lobelia’s belief
will be false in the vast majority of these cases, for flawed processes rarely
happen to produce true beliefs.
One consequence of this approach is that if Lobelia’s higher-order evidence
was weaker, retaining belief might manifest good dispositions. Assume, for
instance, that Lobelia is told that at an earlier time, prior to forming her initial
belief, there was an objective chance of 10 percent that she would be hypoxic
by the time she formed this belief. Depending on the value of knowledge/true
belief, and the disvalue of false belief, in this case the best feasible dispositions
might manifest as retaining belief. I think this is a very good result. After all,
we should almost always assign some credence to the hypothesis that any
given belief was formed by a flawed process. Nevertheless, this shouldn’t lead
us to an unending process of second-guessing ourselves or calibrating our
opinions.
Before wrapping up, I want to distinguish my way of evaluating dispositions
from a more instrumentalist one, and to say how my view is not susceptible to
various objections leveled against consequentialist views in epistemology.
²⁹ See Dutant and Fitelson (Manuscript) for discussion.
    111
5.6. Contrast with Instrumentalist Views
On the broad approach outlined, doxastic states are evaluated by evaluating

the dispositions that these states are manifestations of, dispositions being
evaluated by looking at the values of their manifestations across relevant
counterfactual cases. Consider a success S, and a disposition D manifested
by a subject in a given situation. Here are two importantly different ways of
evaluating the goodness of D, relative to S.³⁰ Take manifestations of D across
relevant counterfactual cases. First, we could ask whether manifestations of
D tend to be successful—or at least as successful as those of feasible alternative
dispositions. (This is a rough and ready way of putting things: I intend it to be
compatible with assigning values to these manifestations depending on how
close to success they come.) Second, we could evaluate manifestations of D by
asking whether they tend to produce the relevant success, whether they tend to
be good means to success, or whether they tend to have the relevant success as
a downstream consequence. It should be clear that I have been assigning
values to dispositions in accordance with the first approach.
I think it would be wrong to restrict epistemic evaluations to dispositions
that manifest as forming, revising, retaining, or retrieving doxastic states. For
instance, engaging one’s rational brain before reasoning about difficult, emo-
tionally laden topics and consulting a variety of sources are good things to do
from the epistemic perspective. This is a point I return to later. Nevertheless,
I have focused on a specific kind of epistemic evaluation of doxastic states.
Some dispositions are good when evaluated from the more instrumental
perspective focused on the consequences of their manifestations, but bad
when evaluated from the perspective focused on the manifestations them-
selves. To take an example from Horowitz (2019), consider an arachnophobic
subject who is almost bound to form botched beliefs about a wide range of
topics whenever she believes that there are spiders nearby. She has plenty of
evidence about the proximity of spiders, but she manages to believe that there
are in fact no spiders nearby. For such a subject, a disposition to believe, no
matter what evidence she has, that there are no spiders nearby might be an
instrumentally good disposition, since it allows her to form other good beliefs
(beliefs that are true, proportioned to the evidence, that constitute knowledge,
etc.). Nevertheless, such a disposition is not good in the sense that I have
³⁰ Earlier statements of my view did not clearly distinguish between these, though it was always the
first kind of view I had in mind. For instance, in Lasonen-Aarnio (2010), I talk about “adopting or being
guided by policies, thought of as sets of rules, that have certain consequences for knowledge.”
112  -
outlined, for it does not tend to have successful manifestations: across a range
of counterfactual cases including the actual one, it manifests as belief that is
false, that fails to be proportioned to the evidence, and that fails to constitute
knowledge.
It should be clear, then, that the view outlined is not susceptible to an
objection often levelled against more consequentialist views—namely, that it
allows for problematic kinds of tradeoffs.³¹ Relatedly, one might object that
evaluating dispositions in terms of the epistemic consequences of their mani-
festations takes us out of the traditional epistemic realm: though failure to get
enough sleep might have bad epistemic consequences, norms telling one to get
enough sleep are not epistemic, but practical ones.³² But of course, the
manifestations of a dispositions to stay up too long at night are not doxastic
states, and hence, not the sorts of things that could constitute epistemic
successes like knowing.
Whether we call dispositions that are good in the more direct way outlined
knowledge-conducive is a terminological choice. Similarly, whether the evalu-
ations I have outlined are consequentialist may be a terminological issue: there
might be a perfectly good sense in which a disposition has good consequences
in virtue of its manifestations tending to be successful. What matters is that
there is an important difference between two ways in which a disposition
could be said to be conducive to success, or to have good consequences for
success. The kinds of dispositional evaluations I have outlined evaluate dis-
positions by looking at how successful their manifestations are across relevant
counterfactual cases, compared with feasible alternatives.³³

Nevertheless, I think it is a virtue of focusing on dispositions that while we
can simultaneously make sense of the kinds of more narrow epistemic evalu-
ations that epistemologists have often been exclusively concerned with, the
kind of framework outlined can be tweaked to make sense of epistemic
³¹ For a discussion of consequentialist views in epistemology and how they allow for tradeoffs, see
Berker (2013).
³² Cf. Horowitz (2019).
³³ Though this should be very clear by now, it is also worth emphasizing that my account does not
evaluate dispositions in terms of subjective expectations. Consider the success of having accurate
credences, where a credence in a proposition p is more successful the closer it is to 1 if p is true, and
the closer it is to 0 if p is false. Schoenfield (2018) proposes a way of evaluating doxastic plans by asking
what the (subjective) expected accuracy of making the plan—of trying to update in accordance with a
given procedure—is. (Similarly, we could evaluate doxastic dispositions by asking what the expected
accuracy is of credal states produced by those dispositions.) Though our accounts have some structural
similarities, dispositional evaluations are more objective: a subject’s opinions about the goodness of a
disposition might themselves be inaccurate, or she might lack any such opinions. Note also that
Schoenfield’s results (like many results concerning the expected accuracy of update procedures) also
rely on assumptions about evidence that I would reject, like partitionality.
    113
evaluations in a broader sense. Indeed, we can evaluate a wide range of

dispositions relevant to successful inquiry. As an example, consider a very
narrowly focused subject whose everyday habits don’t expose him to much
evidence. There isn’t anything wrong, as such, with his sources, it’s just that
they are not very plentiful or diverse: he tends to read a single newspaper,
restricting his attention to certain topics within it. His beliefs are reasonable in
light of the available evidence—indeed, many are even true, and constitute
knowledge. Or consider an even more worrying case involving limited evi-
dence. As a result of subconscious biases, an employer only superficially
glances over a job candidate’s CV. He fails to gain relevant evidence about
her—for instance, he never learns about a highly pertinent degree. However,
given the evidence he has, he reasonably decides that she is not a suitable
candidate: among other things, he knows that she has a(nother) degree in an
area not very relevant for the job. The employer’s beliefs might be perfectly
proportioned to the reasons and evidence he has.³⁴ A broader evaluative
perspective focused on dispositions allows criticizing the evidence-seeking
dispositions of these subjects. For instance, one might point out that they
are not as conducive, in the instrumental sense, to knowledge as alternatives.
Or we might instead focus on the success of having a representative, suffi-
ciently comprehensive body of evidence. The evidence-seeking dispositions of
these subject are not very good (in the more direct, non-instrumental sense)
when it comes to this success.
5.7. Conclusions
I have outlined an evaluative perspective that explains why Lobelia, as well as

other subjects who acquire seemingly defeating higher-order evidence, are
criticizable for retaining their beliefs. I argued that, given plausible assump-
tions, it is not feasible to manifest dispositions that discriminate between cases
in which the kind of higher-order evidence Lobelia has is misleading and cases
in which it is not. Whatever dispositions Lobelia manifests, her obstinacy is
indiscriminate in problematic ways—it manifests itself across a range of
relevant cases involving similar evidence as retaining flawed beliefs. There
are better feasible dispositions, dispositions that would manifest in her situ-
ation as suspending judgment. This is so given a range of veritist and gnosticist
³⁴ See Miracchi (2019) and Flores and Woodard (manuscript) for a more detailed discussion of such
cases.
114  -
views and hence, given a range of different ways of assigning epistemic values
to doxastic states.³⁵
References
Baker-Hytch, Max and Benton, Matthew A. (2015). Defeatism defeated.

Philosophical Perspectives, 29: 40–66.
Beddor, Bob. (2015). Process reliabilism’s troubles with defeat. The Philosophical
Quarterly, 65(259): 145–59.
Berker, Selim. (2013). Epistemic teleology and the separateness of propositions.
Philosophical Review, 122(3): 337–93.
Christensen, David. (2010). Higher-order evidence. Philosophy and Phenomenological
Research, 81(1): 185–215.
Dennett, Daniel C. (1969). Content and Consciousness. Routledge & Kegan Paul.
Dutant, Julien, and Fitelson, Branden. Unpublished Manuscript. Knowledge-
Centered Epistemic Utility Theory.
Drayson, Zoe. (2014). The personal/subpersonal distinction. Philosophy Compass,
9(5): 338–46.
Flores, Carolina and Woodard, Elise. Manuscript. Easy evidence and epistemic
upgrades: In defense of a duty to gather evidence.
Horowitz, Sophie. (2019). Predictably misleading evidence. In: M. Skipper and
A. Steglich-Petersen (eds.) Higher-Order Evidence: New Essays. Oxford: Oxford
University Press.
Hurka, Thomas. (2006). Virtuous act, virtuous dispositions. Analysis, 66(1):
69–76.
Lackey, Jennifer. (2007). Why we don’t deserve credit for everything we know.
Synthese, 158(3): 345–61.
Lasonen-Aarnio, Maria. (2010). Unreasonable knowledge. Philosophical
Lasonen-Aarnio, Maria. (2014). Higher-order evidence and the limits of defeat.
Lasonen-Aarnio, Maria. (2019). Higher-order defeat and evincibility. In:
M. Skipper and A. Steglich-Petersen (eds.) Higher-Order Evidence: New Essay.
³⁵ I am grateful to Jessica Brown, Julien Dutant, Giada Frantantonio, Jaakko Hirvelä, Aleks Knoks,
and Niall Paterson for helpful comments on earlier drafts. Many thanks to Chris Kelp for detailed
comments on a closely related talk. Many thanks to Daniel Drucker, Jeremy Goodman, John
Hawthorne, and Tim Williamson for discussions.
    115
Lasonen-Aarnio, Maria. (Forthcoming A). Competent failure and victims of

deceit. In: Dutant and Dorsch (eds.) The New Evil Demon. Oxford: Oxford
University Press.
Lasonen-Aarnio, Maria. (Forthcoming B). Perspectives and good dispositions.
Philosophy and Phenomenological Research.
Lewis, David. (1986). On the Plurality of Worlds. Oxford: Basil Blackwell.
Miracchi, Lisa. (2019). When evidence isn’t enough: Suspension, evidentialism,
and knowledge-first virtue epistemology. Episteme, 16(4): 413–37.
Schoenfield, Miriam. (2018). An accuracy-based approach to higher order evi-
dence. Philosophy and Phenomenological Research, 96(3): 690–715.
Schaffer, Jonathan. (2004). Two conceptions of sparse properties. Pacific
Sher, George. (2006). In Praise of Blame. Oxford: Oxford University Press.
Simion, Mona, Kelp, Christoph, and Ghijsen, Harmen. (2016). Norms of Belief.
Philosophical Issues. A supplement to Nous, 26(1): 375–92.
6
Suspension, Higher-Order
Evidence, and Defeat
Errol Lord and Kurt Sylvan
6.1. Introduction
In regulating one’s intellectual life, one should pay heed to the evidence. Often
this is a matter of considering which facts confirm or disconfirm some
hypothesis one is considering. Although this is hard enough, it turns out
that reckoning with the evidence would be much easier if this were the only
task. For besides evidence that confirms or disconfirms a hypothesis one is
considering (first-order evidence), there is also evidence that bears on our
rationality in inquiring into the hypothesis (higher-order evidence). So, for
example, you might know that you are prone to make mistakes when reason-
ing about the significance of certain statistical information; or you might know
that, given the fact that you’ve been driving all night, you are in less than ideal
condition for doing mental math; or you might know that, despite your having
conscientiously reached a conclusion on a difficult subject, someone you take
to be cleverer disagrees. All these facts about one’s competence, condition, or
situation bear on whether one’s stance is rational, but it is unclear at best how
they bear on truth-value of the hypothesis one is considering.¹
As it happens, this basic contrast between first-order evidence and higher-
order evidence raises a host of puzzles and problems. These difficulties have
led many philosophers to adopt some extreme views in two related debates.
The first debate is about the rational impact of disagreement. The second is
about whether it is possible for there to be rational epistemic akrasia (e.g., cases
where one rationally has a doxastic attitude D even though one believes one
shouldn’t have D).
¹ There are possible cases in which higher-order evidence alters first-order relations of evidential
confirmation. But we will be arguing that the distinctive significance of higher-order evidence HOE does
not consist in this kind of impact.
, - ,   117
When it comes to disagreement, extreme views abound. On one extreme, it

has become popular to think that the higher-order evidence provided by peer
disagreement always has a dramatic impact; on this view, higher-order evi-
dence provided by peer disagreement always requires us to converge in some
way with those with whom we disagree. On the other extreme lies the view that
the higher-order evidence provided by disagreement has no rational impact
whatsoever. According to this view, we should always just go with the first-
order evidence and hence stick to our guns.
Extreme views also have considerable popularity in the other debate. Some
think nothing prevents higher-order evidence from giving rise to rational
epistemic akrasia. According to this view, there are not the right sorts of
systematic connections between first-order evidence and higher-order evi-
dence to prevent such uncomfortable situations. Although this view has its
opponents, little has been done by them to explain why avoiding epistemic
akrasia is required. It is hard to believe that avoiding akrasia is a fundamental
obligation: as the vast literature on the normativity of rationality suggests,²
coherence requirements of this kind are dubious as absolute requirements, and
even dubious as principles there is any reason to follow, barring some further
story about what this reason is. Without a vindication of the normativity of
coherence requirements, it is not easy to feel comfortable ruling out the
extreme view.
We think such extreme views must be wrong. But we want a framework that
explains why they are wrong, and partly for this reason we also aren’t satisfied
with existing moderate approaches.³ Hence, our central task is to develop a

principled moderate framework for accommodating the normative signifi-
cance of higher-order evidence.
The work in the paper can be divided into two parts: negative and positive.
Our negative claim is that the gridlock between extreme views and the existing
moderate views owes to mistaken background assumptions about the rela-
tionship between first-order relations of evidential favoring, reasons for belief,
and reasons for suspension. We think, however, that more work needs to be
done to provide a systematic explanation of why these background
² For overviews, see Way (2010) and Lord (Forthcoming A).

³ As we will acknowledge again in the next section, some principled moderate stories have been
told—e.g., the accuracy-based epistemic consequentialist story of Schoenfield (2018). But we think
these stories are overly committal on foundational normative questions on which we think the best
general view about the impact of higher-order evidence would be non-committal. One of us is also an
epistemic non-consequentialist (see Sylvan (2020)), and for this reason cannot accept Schoenfield’s
story as the fundamental story.
118     
assumptions fail in the cases that interest us. For this reason, we also have a
positive project.
The positive view in turn has two parts. The first part is the idea that higher-
order evidence provides direct reasons for suspending judgment that typically
leave evidential support relations on the first order intact: instead of destroying
these relations, the reasons for suspension defeat or compete with the epi-
stemic reasons for belief generated by these relations. Secondly, and more
importantly, our framework explains how this defeat is possible by showing
how these distinctive reasons for suspension of judgment flow from the
constitution of suspension of judgment. As a result, our framework is embed-
ded within an account of suspension of judgment that shows how new insights
about its nature lead to a different picture of its rational profile. This frame-
work provides a compelling basis for more moderate positions about disagree-
ment and epistemic akrasia, which show the puzzles about these topics to rest
on more fundamental mistakes about suspension and the relationship between
reasons for suspension, reasons for belief, and evidence.
Our framework for explaining the reasons for suspension that derive from
higher-order evidence might be described as a sort of constitutivist framework.
But we would want to stress two things before having the reader embrace this
description. Firstly, we’re neutral on whether to think about the constitutivism
as a constitutivism of aims (leading to teleology) or a constitutivism of norms
(leading to a non-teleological view). Secondly, to be neutral in a further way,
we are here understanding constitutivism differently than it is normally
understood, following Markovits (2014)’s understanding of Kantianism

about reasons. Markovits took the Kantian view to be consistent with non-
naturalism, and to be a substantive view according to which reasons are
ultimately given by but perhaps not reductively explained by facts about the
nature of agency. Although we happen to like reductive constitutivist views
(see Lord and Sylvan (2019)), we here don’t use ‘constitutivism’ to refer to
such views. Hence, we take this paper to be neutral on the obviously quite
controversial question of whether a constitutivist reduction of normativity is
possible.
With these aims and ideas in view, here is our plan. In the next section, we
will provide some initial motivation for thinking that a framework like ours is
needed by looking at the gridlock that has emerged between principled but
extreme views and moderate but unsatisfying or unprincipled views about
epistemic akrasia and disagreement. The gridlock owes, we’ll see, to over-
looked assumptions about the relationship between first-order relations of
evidential favoring, reasons for belief, and reasons for suspension. But a
, - ,   119
systematic framework like ours is needed to explain how these assumptions

could fail. Having made these points, we turn to our positive work in
Section 6.3, seeking to provide a more fundamental motivation for the claim
that higher-order evidence provides reasons to suspend judgment. We will do
this by showing both how this claim fits into a more general view of the nature
and rational profile of suspension, and by showing that pure evidentialists
about epistemic rationality have a hard time accounting for the full range of
higher-order evidence. In Section 6.4, we explain how our view accounts for
the full range of higher-order evidence, and why it is preferable to existing
views. We then draw things to a close in Section 6.5 by taking a step back to
consider two ways in which our view can be incorporated into more funda-
mental frameworks in the ethics of belief.
6.2. Extremism, Moderation, and a New Third Way
6.2.1 Extremism and Moderation
To appreciate why some have been led to extreme views about the epistemic-
ally appropriate response to disagreement and other sources of higher-order
evidence, it will help to reflect on examples. Consider a now-classic example
that we take verbatim from Christensen (2010: 187):
Drugs. ‘I am asked to be a subject in an experiment. Subjects are given a drug,

and then asked to draw conclusions about simple logical puzzles. The drug
has been shown to degrade people’s performance in just this kind of task
quite sharply. [ . . . ] I accept the offer, and, after sipping a coffee while
reading the consent form, I tell them I’m ready to begin. Before giving me
any pills, they give me a practice question:
Suppose all bulls are fierce and Ferdinand is not a fierce bull. Which of the
following must be true? (a) Ferdinand is fierce; (b) Ferdinand is not fierce;
(c) Ferdinand is a bull; (d) Ferdinand is not a bull.
I become extremely confident that the answer is that only (d) must be true.
But then I’m told that the coffee they gave me actually was laced with the
drug. My confidence that the answer is ‘only (d)’ drops dramatically.’
On the one hand, it is tempting to think that the subject’s reaction in this case
is rational, and that it would be irrational to maintain confidence that ‘only
120     
(d)’ is the correct answer. This thought is tempting because it is tempting to

think that Christensen should not be confident that he has sufficient reason to
believe this conclusion, and hence should not believe it. On the other hand,
this conclusion does follow from the premises, and we can imagine that the
subject sees that this is the case before learning about the coffee. Hence, it
appears that ‘only (d)’ is the conclusion that there is conclusive reason to
believe, given straightforwardly apprehensible logical truths. It is hard to think
the subject ceases to possess this reason just owing to the information about
the drug. From these apparent facts, we already have a strong intuitive tension,
which can be more officially seen with the help of two plausible principles:
Conclusive Reason: If there is a conclusive reason R to believe that p, and

S possesses this reason, then it cannot be irrational for S to believe p on the
basis of R.
Level-Bridging: If there is a strong undefeated reason to believe that one
lacks sufficient reason to believe that p, then one is rationally required not to
believe p.
From the Conclusive Reason principle and the thought that there is conclusive
reason to believe simple logical consequences of rational beliefs, we get the
conclusion that the subject in Christensen’s example cannot be irrational to
believe that the answer is ‘only (d).’ But from the Level-Bridging principle, we
appear to get the opposite conclusion.
Something, it seems, must go. Three kinds of extreme views have emerged
to explain what should go. On the one hand, some accept Conclusive Reason
but reject Level-Bridging. This option is defended by Coates (2012). He agrees
that if there is a strong undefeated reason to believe that one lacks sufficient
reason to believe that p, then one ought to believe that one lacks sufficient
reason to believe that p. But he denies that if one rationally believes that one
lacks sufficient reason to believe that p, it cannot still be rational for one to
believe that p. Hence, he thinks that rational epistemic akrasia is possible. This
idea is, we think, worth avoiding if possible.
On the other hand, some try to accept both principles but deny that they
conflict by denying that one can have a strong undefeated reason to believe a
false proposition about what the reasons recommend. Titelbaum (2015) and
Way and Whiting (2016) in effect defend this option. Although we respect the
arguments given by these authors, we also think that it would be nice to avoid
the view that mistakes about what the reasons recommend are impossible.
Such mistakes seem possible.
, - ,   121
Finally, some accept the Level-Bridging Principle but reject Conclusive

Reason. Worsnip (2018) recommends an intriguing version of this option.
He follows philosophers of practical reason like Scanlon (1998), Kolodny
(2005), and Broome (2013) in distinguishing between what reasons require
and what rationality requires. Defenders of this option would then recom-
mend replacing Conclusive Reason with a principle that swaps the ‘it cannot
be irrational’ for ‘it cannot be contrary to reason.’ Although we agree that we
should distinguish between what reasons simpliciter require and what formal
rationality requires (and we take this to be all Scanlon and Kolodny believe),
we think it is extremely counterintuitive to deny that possessed reasons and
rationality can issue conflicting requirements. For this reason, we also see
Worsnip’s tack as a last resort.
In the narrower literature on the higher-order evidence provided by dis-
agreement, we find a different batch of extreme views. On the one hand, there
are those who hold the equal weight view, which maintains that whenever you
disagree with epistemic peers,⁴ you are rationally required to split the differ-
ence between your credence and theirs. Translating to ‘coarse-grained’ or
outright attitudes (our focus here),⁵ the equal weight view predicts that
whenever you disagree with an epistemic peer, you are rationally required to
suspend judgment. On the other extreme there are what we’ll call extreme
steadfasters. Extreme steadfasters maintain that you are never required to
change your mind in the face of peer disagreement.
Perhaps unsurprisingly, the most popular strategies for defending these
positions rely on the idea that one sort of evidence is irrelevant. Equal weight
theorists argue that when you disagree, you are required to set aside the first-
order evidence you initially heeded.⁶ Once you bracket this evidence, though,
there isn’t anything left to support belief or disbelief. Thus, suspension is
required.
On the other side, there are theorists arguing that the higher-order evidence
is not relevant at all. According to them, the evidence says what it says, and we
should just follow it. Most of the arguments for this position contend that
⁴ Various definitions of epistemic peerhood are provided in the literature. The rough idea is that
your epistemic peers are those with whom you share evidence and who are just as good as you at
processing that evidence.
⁵ It is important to note that the view we develop below is, in the first instance, about outright
attitudes. There is then a hard and interesting question about the relationship between these attitudes
and credences. We don’t know how to resolve that question. But we would want to resolve it before
saying anything directly about credences. Thanks to Jade Fletcher for pressing for clarification here.
⁶ This is often done by appealing to a principle like David Christensen’s Independence principle (see
Christensen (2007), Christensen (2011), Christensen (2009); and also Elga 2007; Cohen (2013); for
criticism, see Kelly (2013); Sosa (2013); Lackey (2010); Lord (2014)).
122     
giving higher-order evidence rational power creates more problems than it is

worth. Weatherson (2019), for example, argues that thinking that higher-
order evidence generates first-order reasons gives rise to vicious regresses,
which are avoidable only by ‘screening off ’ the rational impact of higher-order
evidence. He uses this view to support a steadfast line on disagreement.
Each extreme view about the rational impact of disagreement is a member
of a wider class of views. The equal weight view is a form of conciliationism, a
more general view which holds that the fact that some peer disagrees is always
a reason to reform one’s opinion. Conciliationism on its own says nothing
about how much one should reform one’s opinion. The equal weight view
takes an extreme line on this. On the other hand, the extreme steadfast view
is a member of a wider class of steadfast views, which all maintain that at
least sometimes it is rational to maintain one’s original belief in the face of
disagreement.⁷ Steadfasting per se is compatible with disagreement having
massive rational effects in many cases. It just denies that it always has these
effects.
Steadfasting is compatible with conciliationism. It is compatible with the
claim that at least sometimes one can maintain one’s original view that facts
about disagreement always provide reasons against belief. And, indeed, this
combination of views is downright commonsensical. It is, we think, a moder-
ate position worth fighting for, and we intend to fight for it. The problem,
though, is that it is unclear how to vindicate it. Extant versions of moderate
conciliationism and steadfasting have been light on the details,⁸ with the
notable exception of Schoenfield’s consequentialist conciliationism (which

we think is overly committal on foundational normative questions). They
fail to offer views about the mechanics of the rational impact of disagreement
that vindicate common sense. One of our main goals is to provide a framework
for delivering these mechanics.
6.2.2 Making Room for Moderation
To shed light on the under-explored area of logical space we would like to

inhabit, we need to consider an insufficiently targeted background assumption
lurking behind our earlier formulation of the akrasia puzzle. Here it will help
⁷ Most extant opponents of the equal weight view (and other extreme conciliationist views) are in
fact moderate steadfasters rather than extreme steadfasters, with Tom Kelly being the exemplar.
⁸ See Lord (2014) for discussion.
, - ,   123
to consider a further kind of case and a specific version of the reasoning that
has appeared in the literature. Worsnip (2018)’s discussion is, we think,
instructive. Worsnip suggests that in cases where we can rationally believe
false propositions about what the evidence supports, it can be rational to hold
doxastic attitudes that are not supported by the epistemic reasons that this
evidence provides. Now, to get to Worsnip’s conclusion in such cases, we will
need the following assumption:
Evidential Confirmation-Reasons Principle: Necessarily, if the evidence

confirms attitude A, then the epistemic reasons favor attitude A.
From this assumption and our two earlier principles, we can infer that in such
cases, it could be rational to hold doxastic attitudes which are not supported by
the epistemic reasons. But if this assumption is false, we cannot make this
inference. We could give a different diagnosis: while the evidence bears a
strong probabilistic relation to a certain conclusion, and favors that conclusion
more than others, there is not yet sufficient reason to jump to that conclusion.
This diagnosis would avoid a tension between our two principles. It would
restore the connection between levels at the cost of weakening the link between
strength of evidential confirmation and overall strength of reasons.
Is this option worth considering? We think so. Reflecting on another hidden
assumption in the discussion of our first example gives strong reason to
consider it. Notice that in Drugs, we cannot generate a tension between the
Conclusive Reason and Level-Bridging principles without the help of the

following assumption:
Entailment-Reasons Principle: Necessarily, if the facts entail that p and one

is aware of them, then one’s epistemic reasons decisively favor believing p.
This principle is far from obvious, as the literature querying the relationship
between logic and reasons for belief already suggests.⁹ Admittedly, the existing
literature mainly suggests that there cannot necessarily be objective reason to
believe whatever is entailed by one’s beliefs, since these beliefs may be unjus-
tified. But as both of us have more recently stressed in developing accounts of
possessed normative reasons,¹⁰ it is also not plausible that one possesses a
reason to believe p simply in virtue of having a justified belief which entails
that p: one must also be sensitive to the relevant ground of the reason-for
⁹ For the locus classicus, see Harman (1984). ¹⁰ See Sylvan (2015); Lord (2018).
124     
relation (e.g., some relevant logical relation or evidential relation). What we

want to suggest here is that the example of higher-order evidence gives a
further reason for doubting this principle.¹¹
We can now state the more general doctrine that we take to be indispensable
for generating a tension between the Conclusive Reason and Level-Bridging
principles:
Logico-Evidentialism about Epistemic Reasons: If evidence E entails or

strongly confirms p, then there is always a correspondingly strong reason to
believe p.
Of course, a story does need to be told about why this doctrine fails in the cases
that interest us. It may, after all, remain tempting to think that this principle at
least generates default reasons that would need to be defeated by competing
reasons, and we need to know more about the mechanics of this interaction.
As Christensen and others have noted, it isn’t enough here to appeal to the
notion of undercutting defeat, since undercutting defeaters are meant to sever
evidence-for relations, and these relations are not severed in the examples that
concern us.¹²
Here, however, we think some further existing literature serves as a useful
guide. Note that we are not the first theorists to reject this doctrine. Defenders
of pragmatic encroachment have already rejected it. Schroeder (2012)’s story is
instructive. He suggests that even if all epistemic reasons for belief are gener-
ated by the logical and probabilistic relations between the evidence and further
propositions, it does not follow that all reasons against belief have the same
foundation. And if not, then it cannot be obvious that there is sufficient
epistemic reason to believe even straightforward consequences of one’s evi-
dence. For even if there is an impeccable reason for belief, one cannot follow it
if there is an impeccable reason against. Schroeder suggests that stakes gener-
ate such reasons against belief, yielding a deeper explanation of how pragmatic
encroachment works.¹³
¹¹ Gonzalez de Prado Salas (2020)) argues that higher-order evidence casts doubt on the Entailment-
Reasons Principle precisely because it makes it the case that one fails to possess the first-order evidence,
by making one insensitive to the grounds of the reason. This does happen, but we don’t think it is a fully
general explanation.
¹² See, for example, Christensen (2010); Lasonen-Aarnio (2014); and Pryor (2013).
¹³ As a matter of detail, Schroeder does not maintain that the stakes directly provide reasons to
suspend. Instead, he holds that the stakes provide reasons against belief, and that sometimes those
reasons are strong enough that the balance of the reasons for and against belief requires suspension. EL
argues against this in favor of the view that we need to posit direct reasons to suspend in Lord
(Forthcoming B). On pragmatic encroachment, EL then defends similar conclusions to Schroeder.
, - ,   125
We think that higher-order evidence plays a similar role to the role

Schroeder thinks the stakes play.¹⁴ In one respect, we are on easier footing
than Schroeder, since it is more obvious why the reasons here should be
properly epistemic.¹⁵ But again, to get an account that is preferable to extant
views, more will need to be said about why such reasons for suspension exist.
Thankfully, ideas already on the shelf provide a guide. Many epistemolo-
gists and philosophers of normativity are attracted to the idea that epistemic
reasons for belief are generated by standards derived from the essential nature
of belief.¹⁶ A common example of such a story is the attempt to ground
epistemic reasons for belief by appealing to the essential aim or standard of
correctness for belief. Given Schroeder’s observations, it would be natural to
tell a similar story about how the nature of suspension might be able to
generate properly epistemic but non-evidential reasons for belief. One might
claim that suspending judgment essentially aims to protect us from the risk of
being wrong, where the relevant notion of risk isn’t purely evidence-
determined. This is why, one might suggest, it is possible to suspend judgment
in a high-stakes case on the basis of the stakes attaching to acting on the belief.
The naturalness of extending a constitutivist account of epistemic reasons for
belief to reasons for suspension provides a strategic basis for our project. The
resulting derivation of reasons for suspension from the nature of suspension
will take place at a high level. It could then be interpreted differently by
theorists with different larger normative frameworks; epistemic non-
consequentialists, for example, might develop the story about the details in
one way, and consequentialists in another way. But while the account leaves
some theoretical options open in this respect, it is clearly a principled and
explanatory account.
We suggest that this kind of account provides the best general explanation
of why higher-order evidence generates defeaters. We will turn in the next
section to consider the nature of suspension of judgment and how this leads to
higher-order defeaters, and then return in the section after to explain in more
detail how the account resolves the gridlock and fits in the literature on
Similar moves are made within a contrastivist framework in Snedegar (2017: Ch. 6). We want to be
neutral in this paper on whether contrastivism is true. KS indicates some sympathy for contrastivism in
Fogal and Sylvan (2016). We might be willing to embrace it if it is needed to explain the weight of some
reasons for withholding, as Snedegar suggests.
¹⁴ But again, we think that the higher-order evidence provides direct reasons to suspend, not just
reasons against belief.
¹⁵ See Lord (Forthcoming B) for a more systematic argument that they are properly epistemic.
¹⁶ See, for example, Velleman (2000); Sosa (2007); Sosa (2010); Sosa (2015); Shah (2003);
Wedgwood (2002); Lord and Sylvan (2019).
126     
disagreement and higher-order evidence. We will conclude with some direc-

tions for further research, noting two importantly different avenues left open:
one avenue unites the story with Schroeder’s account of pragmatic encroach-
ment, while a different avenue remains within a non-pragmatic, purely
accuracy-based epistemology.
6.3. Higher-Order Evidence and Reasons to Suspend
In this section we will provide motivation for thinking that higher-order

evidence generates reasons to suspend. This motivation has two parts.
Firstly, continuing the train of thought started in the last section, we will
argue further against the idea that we can fully explain epistemic rationality by
appealing to facts that confirm or disconfirm a hypothesis. We will do this by
arguing that such views cannot account for the rational impact of all higher-
order evidence. This argument will provide independent motivation for deny-
ing the Logico-Evidentialist background assumptions at the heart of the
debates canvassed in the previous section. It will also motivate the claim that
higher-order evidence provides direct reasons to suspend judgment.
The second ambition of this section is to argue that a plausible general
account of the nature of suspension and reasons to suspend implies that
higher-order evidence provides direct reasons to suspend. We will sketch
this general account and provide some motivation for it, but we won’t fully
defend it. Our goal is merely to show that we have a plausible story about how
higher-order evidence provides direct reasons to suspend.
6.3.1 When the Evidence Gives Out
Those attracted to broadly evidentialist views—i.e., nearly all the main parti-
cipants in our two debates—might doubt the need to use direct reasons to
suspend to understand the rational impact of higher-order evidence. Why
posit direct reasons to suspend if you can get by just with the evidence?
There are several reasons why we need reasons to suspend which aren’t
determined just by evidential confirmation relations and logical relations
between evidence and attitudes.¹⁷ One of these reasons is particularly germane
here: there is no plausible straightforward evidentialist explanation of the full
¹⁷ See Lord (Manuscript) for a more comprehensive survey.
, - ,   127
range of cases involving higher-order evidence. This is because some higher-

order evidence is merely higher-order evidence—i.e., it is evidence that only
bears on your rational standing vis-à-vis some question. It doesn’t speak to the
first-order question at all. This fact precludes any straightforward evidentialist
account of the rational role of the higher-order evidence.¹⁸
Let’s unpack the thought. Consider Drugs again:¹⁹
Drugs. ‘I am asked to be a subject in an experiment. Subjects are given a drug,

and then asked to draw conclusions about simple logical puzzles. The drug
has been shown to degrade people’s performance in just this kind of task
quite sharply. [ . . . ] I accept the offer, and, after sipping a coffee while
reading the consent form, I tell them I’m ready to begin. Before giving me
any pills, they give me a practice question:
Suppose all bulls are fierce and Ferdinand is not a fierce bull. Which of the
following must be true? (a) Ferdinand is fierce; (b) Ferdinand is not fierce; (c)
Ferdinand is a bull; (d) Ferdinand is not a bull.
I become extremely confident that the answer is that only (d) must be true.
But then I’m told that the coffee they gave me actually was laced with the
drug. My confidence that the answer is ‘only (d)’ drops dramatically.’
Plausibly, the fact that the coffee is laced with the drug has some impact on what
the subject epistemically ought to do. Indeed, it is plausible that he is forbidden
from believing that (d) is the right answer. He ought to suspend judgment.
Furthermore, this intuition remains even if we emphasize that his reasoning was
logically impeccable, and he had proof that (d) is the correct answer.
How might the evidentialist explain this result? One sort of evidentialist
simply can’t. This character accepts what we will call Strong Evidentialism:
Strong Evidentialism: The rational doxastic attitude toward p is solely

determined by the evidence for p and for ~p.
The balance of the evidence for and against whether (d) is the correct answer is
clear. Again, we could even imagine that the subject has a proof that (d) is
right. So, if all we have is the first-order evidence, we can’t explain why he
¹⁸ In one way or another, this fact has been pointed out frequently in the literature. See Christensen
(2010); Schoenfield (2015a); Schoenfield (2015b). Related but more general issues are discussed by
Pryor (2013).
¹⁹ See Lord (MS); Lord (Forthcoming B) for further discussion of a similar case that is inspired by
Elga (Ms) and Schoenfield (2015a).
128     
should suspend judgment. Of course, not all evidentialists are committed to

Strong Evidentialism. The natural place to turn is Weak Evidentialism:
Weak Evidentialism: The rational doxastic attitude for S to adopt is solely

determined by S’s total evidence.
This view is the sort of view that both conciliationists and steadfasters have in
mind. But here’s the problem: it is wholly unclear how the ‘total evidence’
makes it the case that I should suspend judgment. This result cannot simply be
explained by probability relations. The interaction between first and higher-
order evidence doesn’t affect the probability relations between the first-order
evidence and proposition at issue. Maybe it’s fine to say in these cases that the
total evidence ‘points toward’ suspension. But we want some explanation of
why it points toward suspension.²⁰
Reasons to suspend offer resources to explain how it is that the higher-order
evidence has the requisite effect. It isn’t because it evidentially interacts with the
first-order evidence to yield sufficiently low confirmation or make the needle
on some metaphorical evidential dial move in some direction. Rather, the
higher-order evidence directly provides a reason to suspend, which balances
against the reasons to believe. In this case, it makes sense for me to suspend.
Given what I know about my possible impairment, this is the proper response.
Thus, the reason to suspend is intuitively weightier than the reasons to believe.
Opponents will now likely complain that we haven’t provided an adequate
explanation. For we haven’t given a deeper story about why the reason to
suspend is weightier. That’s true, but it wasn’t what we were after. We don’t
think we need to give a general normative theory of weight; this could only be
descried through the correct fundamental normative theory, a matter on which
we wish to remain neutral here. Rather, what we want is a structural explan-
ation of how suspension could be required, and direct reasons to suspend
provide that explanation.²¹ The evidentialist lacks a clear explanation at the
same level, since it is not clear what they can invoke to do the explanatory work.
²⁰ Christensen (2010) suggests that we understand the evidence just in terms of the facts that bear on
the rationality of belief. EL argues in (MS) that this view trivializes the debate between evidentialists and
pragmatists. Better, we think, to conceive of the evidence as the facts that confirm or disconfirm
hypotheses. But doing this necessitates positing reasons to suspend.
²¹ There may be more of a puzzle in cases where one has very strong evidence for p and equally
strong evidence against p. Plausibly, these evidential relations generate correspondingly strong reasons,
and the reason to suspend must be sufficiently weighty to outweigh them both if we want to explain
why one ought to suspend here. Here it may be necessary, as Snedegar (2017: Ch. 6) suggests, to
combine the kind of account we are offering with a contrastivist picture of the reason-relation.
, - ,   129
They cannot exploit the aspect of evidence one usually exploits to do the
work—i.e., to the amount of confirmation or evidential support. By appealing
to direct reasons to suspend, we get additional structural resources to explain
how suspension could be required despite the confirmational facts.
This fact yields strong initial motivation for appealing to reasons for
suspension in the context of theorizing about the rational power of higher-
order evidence. In our view, this motivation is decisive. We can’t get by simply
by with the evidence for and against p. But even if we retreat to Weak
Evidentialism, we need a tool to help explain how mere higher-order evidence
can require suspension. Reasons to suspend furnish this tool.
In the next subsection we will further bolster this point by showing how the
existence of direct reasons to suspend in higher-order evidence cases flows
naturally from a plausible account of suspension of judgment, before turning
to argue that reasons to suspend provide the right tool to prop up a principled
moderate account in the two debates discussed at the outset.
6.3.2 The Nature of Suspension and What it Tells us

about the Rational Profile of Suspension
Whether or not one is a functionalist about mental states, it should seem

undeniable that function is a helpful tool for differentiating mental states.
With that thought in mind, we will be arguing for a view about suspension
which yields direct reasons for suspension on the grounds that it best captures
key functional desiderata. In fact, we will argue that there are two different
varieties of suspension, and that there are interesting and underappreciated
direct reasons for both varieties.
Let’s start by thinking about the function of belief and disbelief. A helpful
tool in this project is an intellectual agenda, which one could represent by a set
of questions on which one takes a stand or is seeking to take a stand. Belief and
disbelief are ways of determining the status of a question on your agenda.
Determining is a way of taking a settled stance on some issue. When you
determine the status of some claim via belief or disbelief, you settle the
question for yourself, and thereby add some belief to your doxastic corpus.
A functional upshot of this is that when you determine the status of a claim in
this way, the claim is available to you for rational purposes.
Suspending judgment is another way of taking a settled stance. However,
unlike belief and disbelief, when you suspend you do not add any beliefs to your
doxastic corpus. To use Scott Sturgeon’s apt phrase, when you suspend about
130     
whether p, you adopt a settled neutrality when it comes to p.²² While you don’t
add any belief to your doxastic corpus, you do adopt some attitude that leaves
the issue on your agenda. This thought captures the datum that suspension is an
honest-to-goodness attitude, not merely the absence of belief and disbelief.
What is the functional upshot of adopting a state of settled neutrality? Or, to
put the question in normative terms, why should one ever adopt such an
attitude when one could just remain unopinionated about p—i.e., not form a
settled stance at all? To answer this question, we need to further investigate the
contours of settled neutrality. Cutting to the chase, we think there are two
important forms of settled neutrality.²³ On the one hand, you can form an
interrogative attitude about whether p.²⁴ As etymology and syntax suggest,
interrogative attitudes seem to be mental states directed at questions.²⁵
Paradigm examples include curiosity and wonder. When you are curious
whether p, you have an attitude directed at the question of whether p is the
case: the question is what piques your interest. In paradigmatic cases of
curiosity, you don’t believe or disbelieve, but you do have some sort of attitude
toward the question of whether p.
The functional role of this kind of state is to dispose you to determine
whether p. For example, it grounds dispositions to be on the lookout for
evidence salient to whether p, to perform actions that will help you settle
whether p, to draw your attention to other knowledge germane to p, and so
on. Hopefully it is obvious why this state often has a rational point. Frequently
we lack the resources to rationally determine the answer to some question we
ought to keep on the agenda. For example, in many cases where one’s evidence
is evenly split, it makes sense to keep the question of whether p on one’s agenda
even though neither belief nor disbelief is rational. It would be unfortunate if
this evidential situation forced us to drop the issue off the agenda completely.
Indeed, it is precisely in these sorts of situations that we need a state that
disposes us to settle whether p. Interrogative attitudes fit the bill.
Thus, interrogative attitudes allow us to take a settled stance without adding
some claim (or its negation) to our doxastic corpus. They do this because they
are directed at questions. We can thus take a settled stance about whether p,
and this stance can keep the issue of whether p on the agenda without
determining the answer. Furthermore, by adopting an interrogative attitude,
²² See Sturgeon (2010). ²³ For much more, see Lord (MS); Lord (Forthcoming B).
²⁴ See Friedman (2013a), Friedman (2013b), Friedman (2017).
²⁵ It is controversial whether the fundamental objects of interrogative attitudes are questions. But as
we have previously indicated, the surface appearances certainly make them appear to be question-
directed: verbs for interrogative attitudes take questions rather than that-clauses as complements, and
the word ‘interrogative’ is derived from the Latin word for questioning. For deeper arguments,
however, see Friedman (2013a).
, - ,   131
we adopt an attitude that disposes us to settle whether p. Often this attitude is

the rational one to adopt.
Interrogative attitudes are useful in situations where we need to keep an
issue on the agenda because settling the issue is important. This is why it
makes sense to form an interrogative attitude rather than drop the question in
cases where one’s evidence doesn’t justify belief or disbelief. On the other end
of the spectrum, there are anti-interrogative attitudes. Anti-interrogative atti-
tudes allow us to take a settled stance towards an issue in a way that disposes us
not to settle the question in the future. Staunch agnosticism in the religious
case provides a good example of such an attitude in action: one settles to avoid
the question of whether God exists because one doesn’t believe one will ever
have the evidence needed to settle it. Hence, while forming an interrogative
attitude is a way of taking a settled stance that disposes one to settle the issue,
and deciding to be unopinionated is a way of dropping the issue, forming an
anti-interrogative attitude is a way of burying the issue.
Staunch agnosticism aside, it may seem less obvious why forming an anti-
interrogative attitude might make rational sense in ordinary cases. To get a
better sense of why, notice that there are situations where it makes sense to
adopt a settled neutrality even though it doesn’t make sense to form an
interrogative attitude. For example, consider the question of whether Karl
Marx had an odd or even number of food particles in his beard when he died.
We take it you are like us and have no evidence bearing on this question.
Furthermore, you are highly unlikely to ever get such evidence.
Dropping the issue is one option. But if we were to adopt an attitude

regarding the issue, it doesn’t look like any of the options considered so far
fit. Belief and disbelief are out. Interrogative attitudes dispose us to determine
truth-values, but there is little point in having an attitude like that in this case.
It is plausible that we know that we will never have sufficient evidence to
rationally determine the truth-value. In virtue of this last fact, though, it makes
sense to form an attitude that insulates you from the question. We don’t
merely want to forget about it. We want to take steps to make sure it doesn’t
bother us again. This is what anti-interrogative attitudes do. They dispose you
to not be on the lookout for evidence, to not pick up on salient background
knowledge, to not perform actions that would uncover evidence, and so on.
Such attitudes make sense in cases where a stand is needed towards a claim
that is unlikely to be rationally settled.²⁶
²⁶ In other work (Lord Forthcoming B), EL argues that anti-interrogative attitudes are also called for
in cases where you owe it to yourself (e.g., cases of so-called grit a la Morton and Paul (2018)) or others
(e.g., cases where you owe it to your friends to think well of them) to become insulated from evidence.
132     
Now we have enough to start to see the contours of reasons to suspend.

They will be reasons to form either interrogative or anti-interrogative atti-
tudes. Both sorts of attitudes are forms of settled neutrality. The crucial
difference between them is that interrogative attitudes open us up to the
evidence relevant to some question, whereas anti-interrogative attitudes
insulate us from such evidence. Reasons to suspend are reasons for such
attitudes.
As we saw in the last section, there is good reason to think that higher-order
evidence provides reasons to suspend, and this is plausible even before inves-
tigating suspension and its normative profile. After all, what we want to
explain in the most interesting cases involving higher-order evidence is the
fact that one ought to suspend judgment on the basis of the higher-order
evidence. A natural explanation of this fact is that the higher-order evidence
directly provides reasons to suspend that are decisive in the relevant cases.
The plausibility of this claim is reinforced by the positive account given
here. Interrogative attitudes are especially relevant. In many of the interesting
cases involving higher-order evidence, it is plausible that one should form an
interrogative attitude. That is, one should take a stand about the relevant
question, but only insofar as one adopts a state that disposes one to determine
the answer to the question. This is because in most of the relevant cases, the
question is important, and it is also important for one to be disposed to
determine its answer. We can confirm this by thinking about Drugs. There
it is plausible that the subject should form an interrogative attitude upon
learning about the laced coffee. After all, he was asked to form a view about
the question. So, he should be on the lookout for ways to determine the
answer. But given what seems to him to be his likely impairment, it is too
risky to rely on the first-order evidence. Forming an interrogative attitude is
the thing to do.
6.4. A Wondrous Resolution
Having sketched our framework, the first thing that we want to note is that it is
on track to vindicate conciliationism of some sort.²⁷ Of course, the claim that
We put these cases to one side because of the controversy surrounding them. For other reasons to adopt
anti-interrogative attitudes, see Rosenkranz (2007).
²⁷ It’s not for nothing that we say, ‘on track.’ For there are important exceptions. For example, it is
not plausible that there will be a reason against belief every time there is disagreement. For there can be
undercutting defeaters of the reason to suspend provided by the fact about disagreement. A second
, - ,   133
higher-order evidence generates reasons for suspension is strictly speaking

compatible with the further claim that these reasons are never sufficient, since
reasons are one thing and sufficient reasons another. And so full-fledged
conciliationism may not appear to be automatically vindicated. But we think
our framework gives novel reasons for rejecting extreme steadfast views that
deny that there is ever sufficient reason to suspend in the cases at issue.
To see why, let’s first consider extreme steadfast views that seek to accept
both of the principles that initially seemed to generate our puzzle while also
holding on to Logico-Evidentialism about Epistemic Reasons. These views
seek to avoid this implication by holding that one cannot rationally be misled
about what the evidence supports. These views imply, we believe, that it
cannot be rational even to wonder whether one might have been wrong in
thinking that one’s evidence supports p if it does support p. They imply not
only that Christensen’s confidence shouldn’t drop, but also that he shouldn’t
even be curious about whether his evidence supports p. This opposition to
curiosity and wonder strikes us as worse than obstinacy or dogmatism: it
smacks of anti-intellectualism and illiberalism. Only in cases where it is just
obvious that p might it seem right to claim that wondering whether p involves
a failure of rationality. Yet it is not obvious from the subject’s perspective that
the evidence gives sufficient reason for belief in the Drugs case.²⁸
Here is a reductio argument for thinking that the steadfast view in question
invites this implication. Suppose what is plausible for the sake of argument: the
subject in Christensen’s example should wonder whether the answer he first
embraced is, after all, the right one. More specifically, he should wonder
whether he has sufficient evidence for this conclusion. If our view is right, it
follows that he should suspend judgment, since suspension of one sort just is
the holding of some interrogative attitude. Hence if some interrogative attitude
point to stress here is that it matters here how you define conciliationism. We don’t think our view
necessarily provides a vindication of conciliationism about confidence or credence. It’s an open
question how reasons to suspend impact our credences. So, it might be that facts about disagreement
don’t always provide reasons to lower our confidence, at least on some ways of understanding what it
means to lower one’s confidence. So, what we think is vindicated is conciliationism about belief: facts
about disagreement provide reasons against belief. If facts about disagreement provide higher-order
evidence and higher-order evidence provides reasons to suspend, then facts about disagreement will
provide reasons to suspend. These are reasons that compete with the reasons to believe, and thus are
reasons against believing.
²⁸ Perhaps there is some purely objective sense in which the underlying logical truths here are
obvious, but this sense is not the relevant sense for understanding rationality. Indeed, as far as we can
see, no logical truth L is so obvious from an objective point of view that it couldn’t make sense to
wonder whether L is true. Otherwise wondering about the possible correctness of various non-classical
logics would be necessarily irrational. But it just doesn’t seem right to claim that one becomes irrational
merely by taking non-classical logics seriously in meta-logical inquiry. Graham Priest may be wrong,
but he is not irrational; similarly, for people who seriously engage with his work.
134     
is required, one kind of suspension is required. Yet the following more specific
principle seems no less plausible than our original level-bridging principle:
Further Level-Bridging Principle: If one ought rationally to suspend judg-

ment about whether one has sufficient reason to believe that p, then one ought
rationally not to believe that p.
Since the subject ought rationally to wonder whether he has sufficient reason
to believe p in Christensen’s case, it follows that he ought rationally to
suspend on whether his evidence supports p. Given the bridging principle,
it follows that the subject ought rationally not to believe that the answer is
‘only (d).’ Yet he intuitively continues to possess conclusive evidence for this
conclusion. So, if Logico-Evidentialism is preserved, the tension between
Level-Bridging principles and Conclusive Reason is preserved. But this is
contrary to the steadfast view at issue. Hence—barring revisions—the view
implies that Christensen’s subject would be irrational to wonder about
whether his attitude was rational.
We say that it isn’t irrational to wonder in such cases. We offer a simpler
solution: reject the background assumption of Logico-Evidentialism. This
allows us to keep our principles and the data without contempt for curiosity.
And again, rejecting Logico-Evidentialism is not an extreme option. It is what
the nature of suspension of judgment together with the systematic links
between reasons for belief and reasons for suspension recommend.
Other views encourage the same contempt for curiosity. Consider Worsnip
again. Although Worsnip doesn’t try to resolve the puzzle by keeping all the
principles, he will embrace modified versions of these principles which will
still condemn curiosity unless Logico-Evidentialism is rejected. In particular,
although Worsnip denies that reasons and rationality go together, he doesn’t
deny that there is some other normative status that co-travels with the reasons.
One such status we can call the ‘should’ of all-things-considered reason. Here is
a modified principle that Worsnip would accept:
Modified Conclusive Reason: if there is a conclusive reason R to believe that

p, and S possesses this reason, then S should (given the reasons) believe that p.
Yet assuming Logico-Evidentialism, this principle still conflicts with an

equally plausible modified version of the second principle:
, - ,   135
Modified Level-Bridging Principle: if there is a strong undefeated reason to

believe that one lacks sufficient reason to believe that p, then one shouldn’t
(given the reasons) believe that p.
The only way out for Worsnip is to agree with Titelbaum about the ‘should’ of
all-things-considered reason and deny that one can have strong undefeated
reason to believe false normative propositions of the relevant sort. But this also
requires gratuitous condemnation of intellectual curiosity and wonderment.
Our view again gives a happier resolution.
Although we think extreme steadfast views are too quick to condemn
wonderment, we don’t recommend jumping to the opposite extreme.
Sometimes it is irrational to wonder whether p. If it is plain that p from the
perspective you occupy, it is less clearly acceptable to wonder whether p. For
example, if it is clear from your perspective that p (e.g., if you can easily see
that p), wondering and wondering whether p would seem to manifest inatten-
tion or some other lapse of reason. If one avoids such lapses, one can easily go
ahead and get the knowledge from the plain facts. And, if that knowledge is
obtained, one would do oneself an injustice by allowing the disagreement to
obscure one’s plain view of the facts. One deserves better.
There is a spectrum of cases between the case of wondering whether p when it
is evident that p and wondering whether you selected the right answer in a
Drugs-like case. At what point does wondering become sensible? Common sense
suggests that it is later than extreme conciliationists say, but earlier than extreme
steadfast views say. This was why the steadfast conciliationism introduced in
Section 6.2 was commonsensical. Now, we can see how our view can vindicate
common sense. Higher-order evidence provides reasons to suspend; hence,
conciliationism. Sometimes those reasons to suspend are not weighty enough
to defeat the reasons to believe provided by the evidence; hence, steadfasting.
6.5. The Substantive Options for a Finer Resolution
We now have principled structural reasons for accepting a more moderate

view about the normative significance of higher-order evidence and disagree-
ment. We do not, however, yet have anything like an exact formula for
determining just how strong the reasons for suspension will be in some case,
or a systematic normative theory which spits out such a formula. Although it is
beyond the scope of this paper to develop these things, our framework affords
a new way of organizing the options.
136     
The options can all be viewed as answers to the question of what funda-
mentally determines whether it is epistemically appropriate to inquire into
whether p, where inquiry is understood as seeking to help determine whether
p.²⁹ These options are constrained by obvious truths. It isn’t epistemically
appropriate to inquire into whether p in our sense if it is already clear that p or
already clear that ~p, at least where clarity entails knowledge or being in a
position to know (as we assume). It also isn’t epistemically appropriate to
inquire into whether p in our sense if it is clear that it couldn’t be determined
whether p.³⁰ These obvious truths give us minimal necessary conditions on
epistemically appropriate inquiry: S’s inquiry into whether p is epistemically
appropriate only if (a) the truth-value of p isn’t evident to S, and (b) it isn’t
evident to S that S couldn’t help to determine whether p by her inquiry.
What else is necessary? Strong liberalism would say nothing else is neces-
sary. Unfortunately, this view may seem to give a thumbs-up to foolhardy
inquiry. If it is unlikely that S could even help to determine whether p, it may
seem a fool’s errand for S to inquire into whether p with any earnestness.³¹
Barring rejection of this intuition, it seems we need other conditions on
epistemically appropriate inquiry.
We can now see some important large-scale divisions. On the one hand, we
can make a broad distinction between consequentialist views and non-
²⁹ We say ‘help determine’ to indicate that one’s contribution to inquiry may be part of a larger
inquiry—say, the inquiry of a research team, a social group tasked with answering certain questions
(e.g., political figures interested in whether certain acts would be constitutional), or even a whole
academic discipline or intellectual tradition. It may be clear that one couldn’t alone determine whether
p in one’s life even if it is clear one could partly help some relevant group could determine whether
p. Here it may make sense to inquire. But if it is evident that one couldn’t even help advance an inquiry,
it does seem to us foolhardy to inquire into whether p. (Thanks to Justin Snedegar for making us clarify
our beliefs here.) It is worth stressing that there is an important contrasting notion of reflection on an
issue that doesn’t assume that the issue could ever be rationally decided by an individual or a group. It
doesn’t, we assume, make sense to seek to determine whether p if it is obvious that p can’t be
determined. Does it follow that it makes no sense to think about p? This seems extreme and indeed
anti-philosophical: even if we know we can’t settle whether p, it may be worth thinking about what it
would take to settle whether p, and what would need to be true for p to be true. Even if no intuitive
evidence could ever decide between whether natural properties are tropes or universals, it doesn’t
follow that the nature of naturalness isn’t worth contemplating. For this reason, we might call the wider
notion of inquiry contemplation. It can even make sense to contemplate what it would be for a
proposition to turn out true even if one knows it is false: one might deny that gods are actual but
still find theology a worthy study in speculative metaphysics. We leave the constitutive norms of
contemplation (if there are any) for other theorists to determine. It does, however, seem plausible to us
that ideas can be worth contemplating even if we know we will never be able to determine whether they
are right. (Indeed, ideas can be worth contemplating even if they are obviously false: consider Leibniz’s
delightful ontology of monads, for example.)
³⁰ In this case, the only sort of attitude that is rational is an anti-interrogative attitude (although,
sometimes one should form no attitude at all).
³¹ Again, in these cases, it is tempting to think one should form an anti-interrogative attitude. These
reasons aren’t as strong as the reasons to form anti-interrogative attitudes when we know p is
undeterminable, but they might be very strong.
, - ,   137
consequentialist views. Consequentialist views work under the assumption

that the fundamental norm of inquiry is productivity: inquiry ought to effi-
ciently yield results. It is important to emphasize ‘fundamental’ here, since
other theorists needn’t deny that productivity has some derivative significance.
At a minimum, they merely deny that a person’s intellectual economy should
be governed by a principle of ‘growth for growth’s sake.’ The distinction is
helpfully illustrated by contrasting the value of research from the researcher’s
point of view and the value of research from the point of view of a university
driven by managers concerned with its profile in the ‘knowledge economy.’
The sole fundamental intellectual value of such a manager is research output
as such. This fact is suggestive: unless the manager is genuinely confused about
the structure of intellectual value, it is tempting to think that they don’t really
care about the research at all, but rather about some extra-intellectual source
of value, such as prestige, power, or money. Although researchers can become
alienated cogs by internalizing managerial ideology, their natural interest is
not in output for output’s sake. Still, an unalienated researcher needn’t deny
value to productivity: they just treat productivity as an upshot of genuine
passion for the issues, and as having value as an indicator or manifestation of
passion.
A true consequentialist in our sense will view things the other way around:
what is fundamentally important is productivity, and since passion is required
for it, one should be passionate. Such a consequentialist may recommend the
ideology of pure research, but only because of the bad effects of alienation
through obsession with productivity for its own sake.³²

From the objective point of view, this consequentialist may then adopt two
sorts of views. A purist consequentialist will say that the constitutive norms of
inquiry are constrained by purely intellectual value. Her third condition on
epistemically proper inquiry might then be something like: inquiry into p is
epistemically proper only if it is likely enough that p can be determined in
good time, compatibly with a research program that maximizes production of
intellectual value. Inquiry into matters that are very unlikely to be determined
will often be spurned on the grounds of opportunity cost: the effort would be
better spent in areas more likely to yield fruit, and so no research program
should permit wasting time on such matters. In such cases, anti-interrogative
attitudes or even dropping of the issue are needed to move one into more
³² Compare Railton (1984)’s discussion of the alienation objection to ethical consequentialism; it is a

sad irony that Railton’s idea would here be used to prop up the epistemic analogue of capitalism (or at
least the epistemic analog of neoliberal managerial ideology). Recently, Singer (2018) explicitly extends
the Railtonian view to epistemology.
138     
fertile fields. Such a consequentialist may, however, believe that some matters
have great enough intellectual importance that even if they are not particularly
likely to be determined in good time, spending more time on them is worth
the risk.
Impurist consequentialism allows further values that are not purely intel-
lectual to encroach on the significance of some issue. Such a consequentialist
might instead propose a third condition on epistemically proper inquiry along
the following sorts of lines: inquiry into whether p is epistemically proper only
if it is likely enough that p can be determined in good time, compatibly with a
plan for conduct that maximizes the production of practically important
knowledge.
These two kinds of consequentialism have different implications about the
conditions under which higher-order evidence can defeat belief. To see why,
let’s first consider more officially what our framework would say about when
suspension of judgment is rational. According to our framework, suspending
judgment about whether p is rational if and only if it is rational to have either an
interrogative attitude toward whether p or an anti-interrogative attitude toward
whether p. More briefly, given the foregoing points: suspending judgment
on whether p is rational if and only if it is epistemically proper to inquire into
whether p, assuming the issue shouldn’t just be tabled. For simplicity, we can
bracket issues that should be tabled by restricting attention to live issues (to
borrow some language from William James (1896). It is then epistemically
proper to suspend judgment on some live issue if and only if inquiry into the
issue is epistemically proper. Higher-order evidence will be sufficient to defeat

first-order evidence supporting some belief if it makes it fitting to (re)inquire
into the issue. The wider significance of the inquiry is what matters on this view.
Let’s consider cases where one hasn’t already formed a belief. And let’s
imagine that one possesses evidence for p which is, in fact, sufficient to
establish that p. Let’s also imagine that one possesses higher-order evidence
which raises some doubt about the fact that the evidence is sufficient to
establish that p. How serious must that doubt be to prevent the evidence
from giving sufficient reason for belief? According to the consequentialist
proposals we’ve been considering, the answer is determined by the importance
of the issue, since such importance will determine the strength of the reason to
inquire. Purist views will deny that the case for scrupulous inquiry is deter-
mined by the practical significance of the issue. Intellectual significance alone
can justify inquiring longer than might otherwise seem justified. Still, there is a
sort of encroachment here by considerations beyond strength of evidential
confirmation.
, - ,   139
These implications fit with some important intuitions. The intuitions in the

case of impurist consequentialism are the familiar ones behind pragmatic
encroachment: as practical importance increases, any reasons for hesitancy
about whether one has got the data right are magnified, and first-order
evidential reasons for belief are defeated. But there is also an overlooked
intuition with a structurally similar basis that arguably isn’t pragmatist:
namely, when the intellectual importance of some issue is great, any reasons
for hesitancy about whether one has got the data right are magnified, and first-
order evidential reasons for belief are defeated. This idea may help to explain
why higher-prestige journals rightly have lower rates of acceptance: more
work is needed to quell doubt when the discovery would be significant enough
for publication in such a journal.
Our reflections also shed light on an insufficiently noticed attraction of
epistemic consequentialism. Although it can be difficult to believe that the
consequences of forming a belief should bear on whether there is reason of the
right kind for that belief, it is much easier to believe that they bear on whether
there is reason of the right kind to inquire into some issue. We can, it seems,
respond in a way that isn’t roundabout to incentives for and against inquiry
that rest on the importance of the discovery. If this is true, then it is easier to
understand how the sufficiency of some epistemic reason for belief may
depend on possible consequences of getting it right or getting it wrong.
Other theorists have already used the framework of epistemic utility theory to
provide a systematic and principled account of when higher-order evidence
defeats, as we noted earlier. What we have just said suggests a deeper potential
reason for signing up to their story: the nature of suspension of judgment
together with the systematic connections between the weight of reasons for belief
and the weight of reasons for suspension provide independent reasons to think
that certain sorts of consequences might decide the sufficiency of the evidence.
Nonetheless, we think it would be too quick to assume that the nature of
suspension of judgment and data about higher-order defeat force a conse-
quentialist treatment; indeed, one of us is a diehard Kantian. We will conclude
with some thoughts about the compatibility of non-consequentialism with
our framework, and how a non-consequentialist might offer a different story
about higher-order defeat and explain away the apparent evidence for
consequentialism.
Epistemic non-consequentialists will deny that fundamental epistemic val-
ues such as accuracy are merely ‘to be promoted,’³³ and for this reason oppose
³³ See, e.g., Sylvan (2018); Sylvan (2020); Littlejohn (2011); Gibbons (2013).
140     
the extension of the ideology of sheer productivity to epistemology. Some of

these theorists follow Kant and suggest that fundamental epistemic value is, at
bottom, to be respected rather than promoted. Now, this kind of picture is
well-positioned to explain the significance of higher-order evidence. Even if
the evidence entails that p, if you have strong reason to think you’re in a state
of mind in which you’re an unreliable judge of what the evidence supports, it
would seem reckless to just go ahead and believe that p, other things being
equal. These non-consequentialist views will hence fit well with common sense
and avoid the pitfalls of extreme steadfast views.
There is, however, some worry that such views might license a kind of
stalling. For notice that the constraint of respect for truth is primarily a
negative constraint. It rules out believing when doing so would be reckless
or negligent by the lights of the value of accuracy. But it does not seem to
encourage believing when it wouldn’t be reckless or negligent. Of course, if one
denies that there are positive epistemic duties, one might be happy with this
result: believing is never required. Still, an epistemic version of Hamlet does
seem open to a distinctively epistemic sort of criticism. Worrying too much
about an issue may seem excessive, and not merely for practical reasons:
paranoia is an epistemic vice. It would, for example, seem silly for the subject
in Christensen’s case to suspend with an interrogative attitude for too long
after rethinking the issue multiple times and coming to see each time that the
evidence entails ‘only (d)’ to be the correct answer.³⁴
If epistemic non-consequentialists lacked resources to explain these intu-
itions, it would represent a significant cost of their approach, though perhaps

one that could be partly surmounted by casting doubt on whether the criti-
cisms are genuinely epistemic. Thankfully, however, there are other, more
direct responses available to epistemic non-consequentialists. One of these
responses is able to capture what seems attractive about some of the conse-
quentialist options described previously, while also avoiding some so far
unmentioned bad features of epistemic consequentialism.
One option turns on a familiar point about many non-consequentialist
views in ethics. Non-consequentialists needn’t deny that there is reason to
bring about good states of affairs, and indeed needn’t even deny that there is
an underived reason to promote the good. At a minimum, they only must
insist that the project of promoting the good take place within the constraints
³⁴ Perhaps the subject should adopt an anti-interrogative attitude instead, but this also seems a bit
extreme: more plausibly, one can and should just go with what repeated witnessing of truth makes
plain.
, - ,   141
imposed by respect. Provided that it wouldn’t be disrespectful of truth to

believe p, there might then be a promotion-based reason to believe p on
such a view. A view of this kind would be reminiscent of a modest deontology
that accepts side constraints but recommends doing as much good as possible
within these constraints.
Such a view may, however, not be the best kind of non-consequentialist
view. This is because it would inherit some of the bad predictions of epistemic
consequentialism discussed by Berker (2013).³⁵ Views of this kind would
invite serious inquiry into p in cases where such inquiry would generate
unrelated epistemic payoffs, even when it is extremely probable that p and
there are no real reasons to doubt that p. Indeed, it is hard to see why views of
this kind wouldn’t recommend inquiry into p when it would have such
unrelated benefits even when it is evident that p. To be sure, the consequen-
tialist views mentioned previously were supposed to be constrained: it is only
fitting to inquire into p in cases in which it is not already evident whether
p. But this constraint lacks a straightforwardly consequentialist motivation,
and so the views discussed earlier were already in one way not fully conse-
quentialist (though their additional condition was specified only in terms of
consequences).
Accordingly, there seems to be good reason to pursue a more purely
non-consequentialist view. How, though, can such a view generate even
proposition-relative reasons for closing inquiry in the aim of thereby learning
a truth? Here it is instructive to remember that Kant centrally claimed in his
later work—i.e., the Metaphysics of Morals in contrast to the Groundwork—

that personhood demands not only respect but also love, and used this to
explain positive duties of beneficence, directed at particular individuals. One
might similarly claim that truth demands something more like love in addition
to respect, and use this requirement to explain why there is something
epistemically objectionable about thinkers who are overly chary in forming
beliefs, as well as about thinkers who are simply uninquisitive. Such love would
be favored on a proposition-by-proposition basis, in much the way that love
for personhood is shown through love for particular persons (though uncon-
ditionally, as persons, in the case of the morally required sort of love).
Of course, some have had doubts about love of truth and its relevance to
epistemology (e.g., Sosa (2000)). But while there is no requirement to love
truth, the intuitions considered previously do suggest that there is some
³⁵ For other kinds of objections that we think would generalize, see Sylvan (2020).
142     
properly epistemic reason to love truth. At some point, we should all learn to
stop querying and embrace the truth.³⁶
6.6. Conclusion
We must leave the task of deciding between these options for future work: it is a
task worthy of many papers and books. What we would like to emphasize
in conclusion is that our wondrous structural resolution of the puzzles indi-
cates that wider controversies—most centrally, consequentialism vs. non-
consequentialism and purism vs. impurism—must be settled in order for us
to know exactly how weighty the higher-order evidence will be in a given case.
As far as the general rationale for accepting higher-order defeat goes, we don’t
think much more can be said beyond what we’ve said earlier. This fact is
heartening for those of us who are more interested in theory than application:
for what we’ve seen suggests that some of the apparent gridlock in the literature
on disagreement and higher-order evidence owes simply to the fact that these
topics are applied topics, and could never be fully resolved without the help of a
systematic normative theory (barring epistemic particularism).
References
Berker, S. (2013). Epistemic teleology and the separateness of propositions.


Broome, J. (2013). Rationality through Reasoning. Oxford: Blackwell.
Christensen, D. (2007). Epistemology of disagreement: The good news.
Christensen, D. (2009). Disagreement as evidence: The epistemology of contro-
versy. Philosophy Compass, 4: 756–67.
Research, 81: 185–214.
Christensen, D. (2011). Disagreement, question-begging, and epistemic self-
criticism. Philosophers’ Imprint, 11: 6.
Coates, A. (2012). Rational epistemic akrasia. American Philosophical Quarterly,
49: 113–24.
³⁶ Thanks to Jade Fletcher, Mona Simion, Justin Snedegar, and Daniel Whiting for detailed
comments, and to audiences at the University of Southampton and New York University for questions
which shaped the paper.
, - ,   143
Cohen, S. (2013). A defense of the (almost) equal weight view. In: J. Lackey and
D. Christensen (eds.) The Epistemology of Disagreement: New Essays. Oxford:
Oxford University Press.
Elga, A. (2007). Reflection and disagreement. Nous, 41: 478–502.
Elga, A. Ms. Lucky to Be Rational. Unpublished paper (written 2008), Princeton
University.
Fogal, D. and Sylvan, K. (2016). Contextualism about epistemic reasons. In:
J. Ichikawa (ed.) The Routledge Handbook of Epistemic Contextualism.
London: Routledge.
Friedman, J. (2013a). Question-directed attitudes. Philosophical Perspectives, 27:
145–74.
Friedman, J. (2013b). Suspended judgment. Philosophical Studies, 162: 165–81.
Friedman, J. (2017). Why suspend judging? Nous, 51(2): 302–26.
Gibbons, J. (2013). The Norm of Belief. Oxford: Oxford University Press.
Gonzalez de Prado Salas, J. (2020). Dispossessing defeat. Philosophy and
Harman, G. (1984). Logic and reasoning. Synthese, 60: 107–27.
James, W. (1896). The will to believe. Reprinted in The Will to Believe and Other
Essays in Popular Philosophy (1956). New York: Dover.
Kelly, T. (2013). Disagreement and the burdens of judgment. In: D. Christensen
and J. Lackey (eds.) The Epistemology of Disagreement: New Essays. Oxford:
Kolodny, N. (2005). Why be rational? Mind, 114: 509–63.
Lackey, J. (2010). A justificationist view of disagreement’s epistemic significance.

In: A. Haddock and D. Pritchard (eds.) Social Epistemology. Oxford: Oxford
University Press.
Lasonen-Aarnio, M. (2014). Higher-order evidence and the limits of defeat.
Littlejohn, C. (2011). Reasons and belief ’s justification. In: A. Reisner and
A. Steglich-Petersen (eds.) Reasons for Belief. Cambridge: Cambridge
University Press.
Lord, E. (2014). From independence to conciliationism: An obituary. Australasian
Journal of Philosophy, 92: 1–13.
Lord, E. (2018). The Importance of Being Rational. Oxford: Oxford University
Press.
Lord, E. (2020a). The normativity of rationality. In: R. Chang. and K. Sylvan (eds.)
The Routledge Handbook of Practical Reason. London: Routledge.
Lord, E. (2020b). Reasons to suspend, rationality’s competition, and the reach of
the epistemic. In: G. Ernst and S. Schmidt (eds.) The Ethics of Belief and Beyond:
Understanding Mental Normativity. London: Routledge.
144     
Lord, E. and Sylvan, K. (2019). Reasons: Wrong, right, normative, fundamental.

Journal of Ethics and Social Philosophy, 15: 43–74.
Lord, E. (2020). Reasons to suspend and the failure of evidentialism. Unpublished
draft paper.
Markovits, J. (2014). Moral Reason. Oxford: Oxford University Press.
Morton, J. and Paul, S. (2018). Ethics, 129: 175–203.
Pryor, J. (2013). Problems for credulism. In: C. Tucker (ed.) Seemings and
Justification: New Essays on Dogmatism and Phenomenal Conservatism.
Railton, P. (1984). Alienation, consequentialism, and the demands of morality.
Philosophy & Public Affairs, 13: 134–71.
Rosenkranz, S. (2007). Agnosticism as a third stance. Mind, 116: 55–104.
Scanlon, T. (1998). What We Owe to Each Other. Harvard: Harvard University
Press.
Schoenfield, M. (2015a). Bridging rationality and accuracy. Journal of Philosophy,
112: 633–57.
Schoenfield, M. (2015b). A dilemma for calibrationism. Philosophy and
Schoenfield, M. (2018). An accuracy-based approach to higher-order evidence.
Schroeder, M. (2012). Stakes, withholding, and pragmatic encroachment on
knowledge. Philosophical Studies, 160: 165–85.
Shah, N. (2003). How truth governs belief. Philosophical Review, 112: 447–82.
Singer, D. J. (2018). How to be an epistemic consequentialist. Philosophy

Quarterly, 68: 580–602.
Snedegar, J. (2017). Contrastive Reasons. Oxford: Oxford University Press.
Sosa, E. (2000). For the love of truth? In: A. Fairweather and L. Zagzebski (eds.)
Virtue Epistemology: Essays on Epistemic Virtue and Responsibility. Oxford:
Sosa, E. (2007). A Virtue Epistemology. Oxford: Oxford University Press.
Sosa, E. (2010). Knowing Full Well. Princeton: Princeton University Press.
Sosa, E. (2013). The epistemology of disagreement. In: J. Lackey and
D. Christensen (eds.) The Epistemology of Disagreement: New Essays. Oxford:
Sosa, E. (2015). Judgment and Agency. Oxford: Oxford University Press.
Sturgeon, S. (2010). Confidence and Coarse-Grained Attitudes. In: T. S. Gendler
and J. Hawthorne (eds.) Oxford Studies in Epistemology. Oxford: Oxford
, - ,   145
Sylvan, K. (2015). What apparent reasons appear to be. Philosophical Studies, 172:
587–606.
Sylvan, K. (2018). Veritism unswamped. Mind, 127: 381–435.
Sylvan, K. (2020 ). An epistemic non-consequentialism. Philosophical Review 129:
1–51.
Titelbaum, M. (2015). Rationality’s fixed point (or: In defense of right reason).
Oxford Studies in Epistemology, 5: 253–94.
Velleman, D. (2000). On the aim of belief. In: D. Velleman (ed.) The Possibility of
Practical Reason. Oxford: Oxford University Press.
Way, J. (2010). The normativity of rationality. Philosophy Compass, 5: 1057–68.
Way, J. and Whiting, D. (2016). If you justifiably believe you ought to ϕ, you ought
to ϕ. Philosophical Studies, 173: 1873–95.
Weatherson, B. (2019). Normative Externalism. Oxford: Oxford University Press.
Wedgwood, R. (2002). Practical reason and desire. Australasian Journal of
Philosophy, 80: 345–58.
Worsnip, A. (2018). The conflict of evidence and coherence. Philosophy and
7
Reasons for Reliabilism
Bob Beddor
7.1. Two Approaches to Justification
What determines whether a belief is justified?

One answer comes from the reliabilist tradition:
Reliabilist Answer: whether a belief is justified depends on whether it is

reliably formed.¹
Reliabilism holds considerable appeal. First, it captures the intuition that there
is an important connection between justification and truth: a reliable process is
standardly defined as a process that produces a high ratio of true to false
beliefs. Second, it explains how justification reduces to non-epistemic proper-
ties. Justification is explained in terms of reliability, which is explained in
terms of truth and falsity.
Reliabilism holds considerable appeal. First, it captures the intuition that

there is an important connection between justification and truth: a reliable
process is standardly defined as a process that produces a high ratio of true to
false beliefs. Second, it explains how justification reduces to non-epistemic
properties. Justification is explained in terms of reliability, which is explained
in terms of truth. Reliabilism thus offers a way of locating epistemic properties
within a naturalistic worldview.²
Despite its appeal, reliabilism faces significant challenges. Many of these are
by now well-known, and have elicited replies from reliabilists, which in turn
have elicited counter replies.³ In this paper, I will bypass this well-worn terrain
¹ The locus classicus of reliabilism about justification is Goldman (1979). For further development
and defense, see Goldman (1986, 2012); Kornblith (2002); Lyons (2009).
² The reductive goals of reliabilism are clearly announced by Goldman, who writes: “I want a theory
of justified belief to specify in non-epistemic terms when a belief is justified” (1979: 90). As Kim (1988)
notes, one motivation for seeking a reductive account is that epistemic properties supervene on natural
properties. A reductive account offers to explain this supervenience.
³ For an overview of the major challenges and replies, see Goldman and Beddor (2015).
05.003.0007
   147
to focus on a challenge that has received comparatively little attention—one

that attacks the heart of reliabilism’s reductive aspirations.
The challenge arises from the fact that a belief can be reliably formed even
though the believer has good reason to think the belief is false, or that it was
unreliably formed. When this happens, the belief is defeated. In order to
account for such cases, reliabilists need to provide a theory of defeat. And,
in order for this account to be faithful to reliabilism’s reductive ambitions, it
had better not use any epistemic terms in the analysans.
Providing such an account is no easy task. The standard reliabilist approach
says that a belief is defeated when there is an alternative reliable process that
would have led the believer to abandon the belief, had it been used. But, as
we’ll see shortly, this approach faces serious problems. This raises the worry
that reliabilists are unable to explain a central facet of justification.
Faced with this difficulty, it is tempting to look beyond the reliabilist
tradition for help. Another prominent account of justification comes from
the ‘Reasons First’ tradition. This tradition offers a different answer to our
starting question:
Reasons First Answer: whether a belief is justified depends on whether it is

supported by adequate reasons.
An advantage of the Reasons First framework is that it provides a promising

account of defeat. A belief is prima facie justified if it is supported by the agent’s
prima facie reasons. Defeat occurs when this support is undermined by the
acquisition of further reasons. This approach to defeat has been elaborated in
great detail by John Pollock, who develops a rigorous system for computing the
defeat statuses of an agent’s beliefs on the basis of their prima facie reasons.⁴
Despite these advantages, the Reasons First framework faces difficulties of
its own. As standardly developed, the Reasons First framework lacks the
explanatory benefits that make reliabilism attractive. It leaves unexplained
the intuitive connection between justification and truth, and it does not reduce
epistemic properties to non-epistemic properties. After all, Reasons Firsters
take the notion of a reason to believe as a primitive. But this is clearly an
epistemic notion.
Given this tradeoff, we might hope for a theory that combines the attractions
of both traditions. This paper develops one such theory: “Reasons Reliabilism.”
⁴ See Pollock (1987, 1992, 1994, 1995, 2001). For a useful overview of the development of Pollock’s
framework, see Prakken and Horty (2012).
148  
The theory follows the Reasons First tradition in explaining defeat in terms of
the notion of a reason to believe, which is taken to be the most fundamental
normative notion. But, unlike most Reasons First views, my theory does not
leave this notion unanalyzed. Rather, it goes on to explain this notion in
reliabilist terms. Simplifying slightly, I identify an agent’s reasons for belief
with the states that serve as potential inputs to their reliable processes. What
emerges is a more reasonable form of reliabilism, one that preserves the best of
both traditions. In particular, it provides an elegant treatment of defeat while
remaining faithful to reliabilism’s reductive project.
Of course, I am not the first to advocate an ‘impure’ or ‘hybrid’ version of
reliabilism. In recent years a number of other authors have proposed syntheses
of reliabilism and evidentialism.⁵ However, I argue that extant evidentialist-
reliabilist hybrids lack one of the chief advantages of the view advocated here:
namely, its ability to explain defeat. Much like the ‘pure’ Reasons First
framework, extant hybrids struggle to provide a theory that is both reductive
and predictive. They also problematically single out a privileged class of
beliefs—those entailed by the agent’s evidence—as immune to defeat. The
synthesis developed here fares better on both counts.
7.2. The Classic Reliabilist Account of Defeat
7.2.1 Why Reliabilists Need an Account of Defeat

In order to introduce reliabilism’s difficulties with defeat, it will be helpful to

start with a simple version of reliabilism:
Simple Reliabilism: an agent’s belief is justified iff it is formed by a reliable

belief-forming process.
Next, consider a stock example of defeat:
Seeing Red: Lori is gazing at a wall, which appears red. Consequently, she
comes to believe : The wall is red. Just then, a generally reliable
⁵ See Alston (1988); Henderson et al. (2007); Comesaña (2010, 2018); Goldman (2011); Tang
(2016); Pettigrew (2018); Miller (2019). Most of these syntheses have been motivated by considerations
other than defeat, though Miller (2019) is an important exception.
   149
acquaintance, Sal, mentions to Lori that there are hidden red lights angled
towards the wall.⁶
According to Simple Reliabilism, Lori’s belief in  remains justified even

after she receives Sal’s testimony. After all, her belief is formed via vision, and
we can stipulate that she has excellent eyesight. But, intuitively, Sal’s testimony
defeats Lori’s justification for believing .⁷
In view of such cases, most reliabilists conclude that Simple Reliabilism is at
best an adequate account of prima facie justification. In order for a belief to be
ultima facie justified (that is, justified full-stop) it is not enough for it to be
reliably formed. It also needs to satisfy a ‘No Defeaters’ condition.⁸
However, introducing a ‘No Defeaters’ condition raises a difficult question.
Defeat is clearly an epistemic notion. In order to fulfill their reductive ambi-
tions, reliabilists need to explain this notion in non-epistemic terms. Can this
be done?
7.2.2 The ARP Account of Defeat
Reliabilists have not left this question unanswered. The standard reliabilist
strategy is to explain defeat in terms of counterfactuals about what the agent
would have believed, were they to have used some alternative reliable process.⁹
More precisely:
⁶ For discussion of this sort of case, see a.o., Chisholm (1966); Pollock (1995); Lasonen-Aarnio
(2010a).
⁷ Perhaps, some may suggest, once Lori receives Sal’s testimony, her belief in  is no longer the
result of vision alone. Rather, it’s the result of a complex process: using vision while disregarding
testimony that vision is locally unreliable. Arguably, this complex process is unreliable. However, I think
there is reason to be skeptical of this ‘typing maneuver.’ In order for it to handle all cases of defeat, we
would need to take on board a substantive commitment: in every case of defeat, there is some way w of
typing the agent’s belief-forming process on which it comes out unreliable. And in order to ensure that
this approach is not ad hoc, we’d need to give some independent motivation for typing the belief-
forming process using w. For further development of this concern, see Beddor (2015a): 147–8, where
I argue that this typing maneuver stands in tension with promising solutions to the generality problem.
For related criticisms of the typing maneuver, see Lasonen-Aarnio (2010a): 4–7; Baker-Hytch and
Benton (2015): 45–7.
⁸ Adding a ‘No Defeaters’ condition may also help the reliabilist deal with other challenges. Take
Bonjour’s (1985) famous example of Norman the clairvoyant. Norman has a reliable clairvoyant faculty
but has no reason to suspect that he does. One day his clairvoyance causes him to believe that the
president is in New York. Bonjour contends that Norman’s belief is not justified, even though it is
reliably formed. As Goldman notes (1986: 112), one strategy for handling this case is to maintain that
Norman’s belief is prima facie justified but defeated.
⁹ This proposal dates back to Goldman (1979), and has been recently defended by Lyons (2009,
2016). Close cousins are defended in Grundmann (2009) and Bedke (2010).
150  
Alternative Reliable Process (ARP) account: an agent A’s belief B is

defeated iff there is some alternative reliable (or conditionally reliable) process
available to A which, if it had been used in addition to the process actually
used, would have resulted in A’s not holding B.
At first blush, ARP looks promising. It appears to explain defeat in entirely

naturalistic terms. It also seems to deliver the right verdicts in many cases.
Take Seeing red: since Sal is stipulated to be generally reliable, deferring to Sal’s
testimony is a reliable process. And if Lori had used this process in addition to
vision, she wouldn’t have continued to believe .
These advantages notwithstanding, ARP faces three serious challenges.
7.3. Difficulties for the Classic Reliabilist Account
7.3.1 Defeater Defeaters
An initial difficulty for ARP—noted in passing by Lyons (2009): 124—is that it

yields the wrong results when defeaters are themselves defeated. An example:
Two Testimony Seeing Red: as before, Lori believes the wall is red, based on
its appearance. And, as before, Sal comes along and mentions that the wall is
illuminated by red lights. But now another reliable acquaintance, Anne, comes
along and provides compelling—though ultimately misleading—testimony

that Sal is a compulsive liar.
According to ARP, Lori’s belief in  is defeated, even after receiving the
evidence of Sal’s mendacity. After all, Anne’s testimony is misleading, and so
trusting Sal’s testimony continues to be a reliable process. And this process
remains available to Lori. (Lori could, after all, simply disregard Anne’s
testimony.) But this is the wrong verdict. Intuitively, Anne’s testimony defeats
the defeater provided by Sal’s testimony. In doing so, it reinstates Lori’s
justification for believing .
7.3.2 Hidden Circularity
A second concern for ARP was raised by Fumerton (1988) but has not
received much attention in the subsequent literature. The worry is that ARP,
   151
when properly unpacked, smuggles the notion of ultima facie justification into
the analysis of defeat. As a result, the reliabilist account of justification fails to
be reductive; worse still, it is circular. Fleshing out this worry requires some
stage setting.
Many reliabilists opt for a theory that distinguishes between inferential and
non-inferential beliefs. To motivate this complication, consider an inferential
process such as deducing the consequences of what one already believes. This
process is not reliable or unreliable simpliciter; it is only conditionally reliable.
In order to handle beliefs formed through conditionally reliable processes,
Goldman (1979) officially formulates reliabilism as a recursive theory:
Recursive Reliabilism
Base Clause: if (i) A’s belief B results from a belief-independent process that
is unconditionally reliable, and (ii) B is undefeated, then B is ultima facie
justified.
Recursive Clause: if (i) A’s belief B results from a belief-dependent process
that is conditionally reliable, (ii) the inputs to this process were ultima facie
justified, and (iii) B is undefeated, then B is ultima facie justified.¹⁰
The recursive clause requires that the conditionally reliable processes operate
on ultima facie justified beliefs. Despite this, the theory is still reductive. After
all, we can use the theory to explain what it is for these input beliefs to be
ultima facie justified. Either these inputs are themselves inferential or they are
not. If they are, we appeal once again to the recursive clause; if not, we appeal
to the base clause. Either way, we eventually arrive at some foundational beliefs
whose justificatory status can be explained using the base clause. Now, the only
epistemic notion that appears in the base clause is the notion of being
undefeated. Assuming ARP provides a reductive account of this notion,
Recursive Reliabilism is reductive.
Where, then, lies the problem? Fumerton observes that, according to ARP,
conditionally reliable processes sometimes function as defeaters: a belief is
defeated if there is some conditionally reliable process available to the agent
which, had it been used, would have resulted in the agent no longer holding the
belief. But a conditionally reliable process cannot lead someone to abandon a
belief all on its own; it can only do so if it is fed certain inputs. This raises the
question: what epistemic status do these input beliefs need to have? Fumerton
¹⁰ See also Lyons (2013), who argues that distinguishing between inferential and non-inferential
beliefs can help reliabilists handle the new evil demon problem (Cohen 1984).
152  
suggests that just as Recursive Reliabilism required that the inputs be ultima facie
justified, so too should ARP. And so, ARP really amounts to the following:
ARP unpacked: A’s belief B is defeated iff either:
1. There is some reliable belief-independent process that A could have

used, which would have resulted in A not holding B, or
2. There is some conditionally reliable belief-dependent process that
A could have used to process ultima facie justified inputs, which
would have resulted in A not holding B.
But, if this is the proper way of understanding ARP, then reliabilism’s reduc-
tive project is in trouble. After all, Recursive Reliabilism qualified as reductive
because the base clause purported to tell us what it takes for a foundational
belief to be ultima facie justified without using any epistemic terms in the
analysans. But if we use ARP Unpacked to explain what it is for a belief to be
undefeated, the base clause will itself rely on the notion of ultima facie
justification.
Perhaps, some may reply, this just shows we should reject the assumption
that the inputs to a conditionally reliable process need to be ultima facie
justified. Perhaps the input beliefs need only be prima facie justified; or
perhaps they do not even need this slender epistemic merit.¹¹ To make trouble
for this reply, consider a variant of Seeing Red where Lori unjustifiably
believes that the wall is illuminated by red lights, but continues to believe
 anyway. Clearly, her overall set of beliefs is epistemically defective. But
does this defect render her belief in  unjustified? To answer this, it will help
to consider a slightly different question: Should Lori abandon her belief in
? I think not. After all, she has no good reason to abandon it. Instead, it’s
her belief about the lighting that should get the boot. But if Lori should retain
her belief in , it becomes hard to maintain that this belief is defeated. (If it
is defeated, then she should abandon it!) This intuition motivates the idea that
only ultima facie justified inputs can serve as defeaters. But if we hang on to
this idea, the circularity problem remains.¹²
¹¹ A number of authors have defended the idea that unjustified beliefs can function as defeaters, e.g.
Lackey (1999); Bergmann (2006). Goldman himself flirts with this view in places (1986: 62, 111).
¹² One way out of this problem would be to reinterpret ARP as providing a theory of prima facie
defeat rather than ultima facie defeat. So, in our variant of Seeing Red, Lori’s belief is  is prima facie
defeated, but this defeater is itself defeated by whatever considerations make her belief about the
lighting conditions unjustified. But this response is unavailable to proponents of ARP, since—as we’ve
seen—ARP lacks the resources to handle defeater defeat.
   153
7.3.3 Alternative Processes that One Should Not Use
According to ARP, an agent’s belief is defeated whenever they have an

available reliable process that meets a certain counterfactual condition—
namely, that if they were to use it, they would abandon their belief. But it
seems that an agent can have an available reliable process that meets this
condition without having any good reason to use it. When this happens, the
mere availability of the process does not seem to defeat the belief.
Here’s a case I offered in an earlier paper (Beddor 2015a: 149–50) that
illustrates this point:
Thinking About Unger: Harry sees a tree in front of him; he consequently

believes : There is a tree in front of me. Now, Harry happens to be very
good at forming beliefs about what Peter Unger’s skeptical 1975 time-slice
would advise him to believe in any situation. Call this process his ‘Unger
Predictor’: in any situation, Harry’s Unger Predictor spits out an accurate
belief about what doxastic attitudes Unger’s 1975 time-slice would advise an
agent to adopt in that situation. Moreover, Harry has a high opinion of
Unger’s 1975 time-slice. Were he to realize that Unger would advise him to
suspend judgment on some proposition, this would lead him to suspend
judgment on that claim. So, if Harry had used his Unger Predictor, he
would have come to believe Unger would advise me (Harry) to suspend
judgment regarding . This would, in turn, have caused Harry to suspend
judgment regarding .
According to ARP, Harry’s belief in  is defeated. After all, there is a

reliable process available to him (his Unger Predictor) that would have
resulted in him no longer believing , had it been used. But this seems
wrong. Harry is not, as a matter of fact, using his Unger Predictor; perhaps he
hasn’t used it in many years. As it stands, he has an excellent reason to believe
 (the testimony of his senses). The mere availability of his Unger
Predictor does not seem undermine this reason.
Defenders of ARP may suggest there’s an easy fix. A natural reaction to
Thinking About Unger is that the example exploits a subject matter mismatch.
Harry’s belief is about trees. The Unger Predictor does not produce doxastic
attitudes about trees, but only about Unger’s advice. This suggests a simple
patch to ARP: simply require that in order for an alternative process to defeat
an agent’s belief in p, that process must produce beliefs about p-related
matters.
154  
However, it is doubtful whether this simple fix suffices. Suppose Harry had

used his Unger Predictor. Then some process would have led him to suspend
judgment on . It’s just that this would have been a two-stage process. This
first stage would have been his Unger Predictor; the second stage would have
been a process that implements Unger’s predicted advice. Call this two-stage
process his ‘Unger Emulator,’ Now, this Unger Emulator produces doxastic
attitudes towards all sorts of subjects, including the presence of trees.
Assuming that this process is reliable, then ARP—even once amended—still
delivers the wrong result.
Some might question whether this Unger Emulator process is really reliable.
Sure, it avoids all errors, but only at the cost of avoiding all truths! Perhaps this
is too high a cost; perhaps the right conception of reliability will classify this
process as unreliable.¹³ But even if we concede the point, we can simply tweak
the example (Beddor 2015a: 153–4). Meet Shmunger, whose skepticism is
much more modest. Shmunger has lots of true beliefs about all sorts of
subjects; she is only a skeptic when it comes to trees. We can then run the
case using Shmunger instead of Unger. Simply stipulate that Harry has an
extremely reliable Shmunger Predictor, which is part of a Schmunger
Emulator: were Harry to reflect on what Shmunger would advise, he would
come to believe: Schmunger would advise me to suspend judgment on whether
there is a tree in front of me, which would in turn cause him to suspend
judgment on . Proponents of ARP cannot plead that Harry’s Schmunger
Emulator is unreliable. After all, it systematically produces true beliefs on a
wide array of subjects; it only leads to suspension of judgment on arboreal

matters.¹⁴
¹³ The formal epistemology literature offers a natural place to look for a measure of reliability along
these lines. See the discussions of credal scoring rules in Joyce (1998); Moss (2011); Pettigrew (2016),
among many others.
¹⁴ In my earlier paper, I also offered a counterexample to the necessity of ARP for defeat. In the
proposed counterexample, Clarence reliably forms a belief that p; a reliable interlocutor later tells him
that p is false. But Clarence irrationally disbelieves everything this interlocutor says. Moreover, no
amount of reflection or counseling would ever uproot this deep-seated mistrust. Intuitively, Clarence’s
belief is defeated. But ARP seem unable to deliver this result: there is no process available to Clarence
which would lead him to trust his interlocutor; hence there is no process available to him which, if used,
would lead him to abandon his belief in p (cf. Baker-Hytch and Benton 2015: 53). However, in this
case—unlike Thinking About Unger—it now seems to me that a modification of ARP will suffice. The
key is to reformulate ARP in terms of dispositions rather than counterfactuals: A’s belief that p is
defeated iff there is some alternative reliable process available to A which, when fed A’s current states as
input, is disposed to lead A to cease believing p. Defenders of ARP could then propose that Clarence has
a general testimony-believer process available to him. This is a process that, for any testifier (or, perhaps,
any testifier that he has no good reason to distrust), produces a relatively high credence in their
testimony. This process is generally reliable. Moreover, it is disposed to lead Clarence to cease believing
p, when fed the experience of receiving his interlocutor’s testimony as input. It’s just that this
disposition is masked by Clarence’s mistrust of his interlocutor.
   155
7.3.4 Looking Forward
Taken together, these problems provide compelling grounds for abandoning

ARP. But it would be premature for reliabilists to admit defeat. The rest of this
paper develops a more promising approach, which draws on the resources of
the Reasons First tradition.
Here is the plan for what follows. I start (Section 7.4) by outlining the most
well-developed version of the Reasons First framework to date, which is due to
John Pollock. While Pollock offers a promising formal framework for under-
standing the structure of justification and defeat, I argue that it should not
supplant reliabilism (Section 7.5). Rather, we should seek a theory that com-
bines the structural features of Pollock’s framework with the core reliabilist
strategy for reducing the epistemic to the non-epistemic. Section 7.6 develops
such a theory; Section 7.7 advertises its advantages; and Section 7.8 compares
it to other hybrid views.
7.4. Pollock’s Reasons First Framework
According to the Reasons First tradition, justification is intimately connected

with reasons. This idea has considerable intuitive appeal. ‘A justified belief is
supported by reasons’ has the ring of a platitude; ‘A belief can be justified, even
though all the reasons count against it’ has the ring of a contradiction.
In addition to capturing these intuitions, the Reasons First approach offers a

promising treatment of defeat. For a belief to be prima facie justified is for it to
be based on prima facie reasons that support it. Defeat occurs when the agent
acquires reasons that either count against the belief itself, or against the
support provided by the reasons on which it is based.
This way of understanding defeat has been given a systematic development
by John Pollock in a series of papers spanning over thirty years.¹⁵ Pollock’s
framework has proven influential beyond epistemology, laying the ground-
work for much research in computer science and AI—an influence that attests
to the explanatory fruitfulness of its core ideas. In this section, I offer a
streamlined overview of Pollock’s framework. (Readers uninterested in the
formal nuts-and-bolts should feel free to skim.)
Pollock’s key piece of formal machinery is an inference graph: a labeled
directed graph providing an abstract representation of all the reasons that bear
¹⁵ See the references in fn. 4.
156  
on the justificatory status of an agent’s beliefs. The agent’s reasons—as well as

the conclusions they support—are represented by nodes. Support and defeat
relations are represented by directed edges. An agent need not actually per-
form all of the inferences encoded in their inference graph. Rather, the
inference graph represents all the inferences that are available to them.
For illustration, Fig. 1 provides an inference graph for Seeing Red. Dashed
arrows represent support relations; solid arrows represent defeat relations.
Here Lori’s visual experience provides a prima facie reason in support of the
node, . And Sal’s testimony provides a prima facie reason in support of the
node, The wall is illuminated by red lights (). This in turn supports the node,
Lori’s visual experience doesn’t reliably indicate the truth of  (), which
defeats .
ST Sal’s testimony
α
Lori’s wall illuminated
VE RL
visual experience by red lights
ve doesn’t
wall is red R U reliably indicate R
Fig. 1 Seeing Red
Inferential support often involves multiple steps. Pollock represents multi-

step arguments with inference branches. An inference branch is an ordered
sequence of nodes, each of which is the immediate ancestor of the next. For
example, in Fig. 1 branch α is the directed path from Lori’s visual experience to
. Branch β is the directed path originating in the experience of receiving
Sal’s testimony, leading through , and terminating in .
Using these resources, we can now flesh out the animating idea behind the
Reasons First framework as follows:
Justified Belief as Undefeated Reasoning: an agent’s belief is ultima facie

justified iff it is the result of an ultimately undefeated inference branch.
   157
This formulation invites two questions. First, what does it mean for a belief to
be the result of an inference branch? For starters, the belief must be supported
by the sequence of reasons represented by the inference branch. But this is not
sufficient: the agent must actually have gone through this reasoning and hold
the belief on this basis.
Second, what does it mean for an inference branch to be ultimately
undefeated? This question is harder, and Pollock’s answer proceeds in stages.
The first stage is to give an account of what it means for one reason to be
defeated by another. According to Pollock, there are two species of defeat:
rebutting defeaters and undercutting defeaters. A rebutting defeater for a node
n is a prima facie reason to think that n is false. By contrast, an undercutting
defeater for n targets the inferential connection between n and the reasons that
support it. Seeing Red is like this: Sal’s testimony that the wall is illuminated
by red lights is not itself a reason to believe that the wall is not red, but it is a
reason for thinking that the building’s appearance does not give good grounds
for thinking that the wall is red. While there are different ways of fleshing out
the notion of an undercutting defeater, for our purposes we can define an
undercutting defeater for n as a prima facie reason for thinking that the
considerations that support n do not reliably indicate its truth in the agent’s
present circumstances.¹⁶
If we follow Pollock in assuming that these are the only two species of
defeat,¹⁷ we can venture the following disjunctive definition of when one
inference branch defeats another (cf. Pollock 1992):
Branch Defeat: an inference branch ψ defeats an inference branch χ iff a

node of ψ defeats a node of χ, where a node n defeats a node n0 iff n either
rebuts or undercuts n0 ,
• i.e., either n is a prima facie reason to believe ¬n0 , or n is a prima facie

reason to believe that the immediate ancestors of n0 do not reliable
indicate the truth of n0 in the agent’s present circumstances.
¹⁶ The ‘in the present circumstances’ qualification is important. After all, Sal’s testimony does not
provide a reason for doubting that reddish wall appearances usually indicate the presence of a red wall.
¹⁷ This is a controversial assumption; some have suggested that cases of higher order evidence
constitute a distinct species of defeat (e.g., Christensen 2010). My own inclination is to think that when
higher order evidence functions as a genuine defeater, it does so by indicating that the agent’s actual
(or believed) grounds for their belief do not reliably indicate the truth of this belief. If this is right, then
defeat by higher order evidence is really just a type of undercutting defeat. But this is a debate that will
need to be deferred to another occasion.
158  
This gives us a definition of when one inference branch defeats another. But
what we really want is a definition of when an inference branch is ultimately
undefeated. The simplest option would be to say that an inference branch is
ultimately undefeated just in case there is no inference branch that defeats it.
But this delivers the wrong results in cases of defeater defeat.
Recall Two Testimony Seeing Red, in which Anne testifies that Sal is a
compulsive liar. As in Seeing Red, Sal’s testimony supports the node, The wall
is illuminated by red lights (), which supports the node, Lori’s visual
experience doesn’t reliably indicate the truth of , which undercuts .
Hence the simple account predicts that branch α is ultimately defeated, and
hence that Lori’s belief in  is unjustified. But this is wrong. After all, Anne’s
testimony provides a prima facie reason to believe that Sal is a liar, which
provides a prima facie reason to believe that Sal’s testimony does not reliably
indicate , which undercuts . As noted in Section 7.3.1, it thereby reinstates
Lori’s justification for believing  (see Fig. 2).
For this reason, Pollock opts for a somewhat more complicated account of
what it takes for an inference branch to be ultimately undefeated. Here we will
Anne’s testimony at
β
Sal is a
Sal’s testimony st cl
compulsive liar
α
Lori’s wall illuminated
ve rl uʹ
visual experience by red lights
st doesn’t
reliably indicate rl
ve doesn’t
wall is red r u reliably indicate R
Fig. 2 Two Testimony Seeing Red
   159
follow the treatment in Pollock (1987), who introduces a technical notion of

being in at a level, defined recursively as follows:
In At A Level
1. All inference branches are in at level 0.

2. An inference branch ψ is in at a level n + 1 iff ψ is not defeated by any
inference branch that is in at level n; otherwise, ψ is out at level n + 1.
Next, we use the notion of being in at a level to characterize what it is for an

inference branch to be ultimately undefeated, as follows:
Undefeated Inference Branch: an inference branch ψ is ultimately

undefeated iff there is an m such that for every n m, ψ is in at level n.
To get a feel for this proposal, let’s walk through how it applies to Two
Testimony Seeing Red. While all three inference branches depicted in Fig. 2
are in at level 0, only γ is in at every level, since only it lacks a defeater. Since γ
defeats β, the latter is out at every level 1. And so, while α is out at level 1
(since it is defeated by a branch that is in at level 0), it is back in at level 2, and
remains in at every level thereafter (see Table 1). So, α is ultimately undefeated.
Hence Lori’s belief in  qualifies as ultima facie justified, as desired.
Table 1. Computing Defeat
Level α β γ
0 in in in
1 out out in
2 in out in
n>2 in out in
7.5. Reason to Want More
Pollock’s framework thus offers a promising way of handling some of the cases
that created trouble for ARP—in particular, cases of defeater defeat. Why not
just jettison reliabilism in favor of the Reasons First program?
There are a number of reasons why one might be dissatisfied with Pollock’s
theory as it stands. Some of these are issues of detail which Pollock himself
160  
sought to address in later work. For example, Pollock later refines his character-
ization of what it takes for an inference branch to be ultimately undefeated
refinements that are mainly motivated by self-defeating inferences. He also
complicates the account by adding additional structure in order to represent
the strengths of reasons for belief.¹⁸ Other concerns focus on specific applications
of Pollock’s framework to various philosophical puzzles, such as the lottery
paradox.¹⁹ I will set these worries aside, since they are not directly relevant to
our purposes. Rather, I want to raise two more fundamental concerns.
One concern is that Pollock’s account does not accommodate the intuitions
and impulses that motivate reliabilism. First, it does not capture the intuition
that there is an important connection between justification and truth. Suppose
a belief is based on undefeated reasons that support it. Why should we expect
the belief to be connected with the truth in any interesting way? Pollock’s
framework provides no answer. Second, and more critically, Pollock’s frame-
work does not satisfy the reductive impulse behind reliabilism. Pollock
explains ultima facie justification and defeat in terms of the notion of a
prima facie reason for believing. But this is surely an epistemic notion.
Those who want an analysis of justification in non-epistemic terms will be
left empty-handed.
Of course, some Reasons Firsters might retort that we should never have
hoped for a reductive analysis in the first place. Reductive analyses of other
epistemic phenomena have a spotty track record, the analysis of knowledge
being a case in point. Why think that the prospects for a reductive analysis of
justification will be any better?

But this brings me to my second concern, which that even if we renounce
our reductive hopes, it is natural to expect some account—reductive or not—of
prima facie reasons. After all, without some account, the Reasons First frame-
work will not offer a predictive theory at all. Unless we have some independent
grip on prima facie reasons for belief, we will not be able to apply Pollock’s
framework to particular cases in order to make predictions about whether a
belief is justified or defeated.
Pollock was sensitive to this concern, and in various places he offers
remarks intended to fill this lacuna. For example, he states that perceptual
appearances provide prima facie reasons to believe; so does memory; so does
statistical syllogism; so does deduction and induction (Pollock 1987: 486–90).
¹⁸ See Pollock (1994, 1995) for his refined treatment of self-defeating inference branches. See Pollock
(2001) for discussion of strengths of reasons.
¹⁹ See Lasonen-Aarnio (2010b).
   161
These remarks go some distance towards giving us a grip on the notion of

prima facie reasons.
However, I think there are still grounds for dissatisfaction. As it stands,
Pollock’s remarks look more like a list of various sources of prima facie reasons
than a genuine theory thereof. A genuine theory should be explanatorily
satisfying: it should tell us what perception, memory, and induction have in
common, in virtue of which they furnish prima facie reasons for belief,
whereas, say, wishful thinking and counterinduction do not. By comparison,
reliabilism offers a much more unified and theoretically satisfying account of
the ultimate grounds of justification. According to reliabilism, all the ultimate
sources of justification have one property in common: their reliability.²⁰
For these reasons, we should not simply replace reliabilism with Pollock’s
framework. Instead, we should seek a synthesis that preserves the virtues of
both approaches.
7.6. Reasons Reliabilism
7.6.1 A Reliabilist Account of Reasons
The basic idea behind my proposal is simple. Reliabilists should identify

reasons for belief with the inputs to reliable or conditionally reliable belief-
forming processes.
A more careful statement proceeds recursively. The base clause gives an

account of an agent’s foundational reasons:
Reliable Reasons (Base Clause): if s is a non-doxastic state of an agent A,

and there is a reliable process available to A which, when given s as input,
is disposed to produce a belief in p, then s is a prima facie reason for A to
believe p.
What states play this role? The clearest candidates are perceptual experiences.
For example, Lori’s visual experience of a red-looking wall is a prima facie
reason to believe . Why? Because she has a reliable process that takes the
contents of her perceptual experiences as input and produces a belief in those
²⁰ Cf. Goldman’s criticism of disjunctive accounts of justification (1979: 90).
162  
contents as output. And this process is disposed to produce a belief in ,

when applied to her visual experience of a red-looking wall.²¹
Do states other than perceptual experiences also fit the bill? Perhaps,
depending on one’s views, rational intuitions and seemings may also serve
this foundational role. Perhaps even non-experiential states could play the
part. For our purposes, there is no need to take a stand on this issue.
Next, we add a recursive clause, which gives an account of an agent’s
derivative reasons:
Reliable Reasons (Recursive Clause): if A has a prima facie reason to

believe p, and there is some conditionally reliable process available to
A which, given a belief in p as input, is disposed to produce a belief in q,
then p is a prima facie reason for A to believe q.
To illustrate, suppose Lori is capable of inferring, At least one thing in my

vicinity is red from . Since this inferential process is conditionally reliable,
the recursive clause tells us that  provides a prima facie reason for Lori to
hold this inferential belief.
Round everything out with the customary closure clause (nothing else is a
prima facie reason for A to believe p), and you have a complete reliabilist
theory of reasons. Of course, this theory could—and perhaps should—be
complicated in various ways. For example, in Section 7.5 I mentioned that
Pollock’s final inference graphs include representations of the strengths of an
agent’s reasons. A natural way of modeling this in a reliabilist framework is to

take the strength of an agent’s reasons to correspond to the degrees of
reliability (and conditional reliability) of the relevant processes. Thus, if
there is an extremely reliable process that is disposed to produce a belief in
p, given state s₁ as input, but there is only a somewhat reliable process that is
disposed to produce a belief in p, given state s₂ as input, then s₁ is a stronger
reason to believe p than s₂.
Some might worry that Reliable Reasons fails to do justice to our ordinary
conception of reasons. After all, many proponents of a Reasons First approach
embrace internalism about reasons. They are thus inclined to say that an
²¹ Here the details will depend on how we individuate perceptual processes. Is veridical perception a
different process from non-veridical perception? If yes, then Reliable Reasons predicts that whether
Lori’s experience provides a genuine reason for believing  depends on whether she is hallucinating.
For our purposes, we can refrain from taking a stand on this issue; I intend my reliabilist account of
reasons to be compatible with various positions on how to individuate processes, perceptual processes
included.
   163
agent’s reasons supervene on their non-factive mental states, and that agents
have privileged access to their reasons for belief. From this internalist per-
spective, the account offered here risks mangling the concept of reasons
beyond recognition.
But is the ordinary conception of reasons really internalist? Compare the
dialectic here with debates over the nature of justification. Prior to 70s, most
epistemologists embraced some form of internalism about justification. But
from this the fact alone it does not follow that the folk notion of justification is
internalist, as reliabilists will take pains to insist. Similar remarks apply to
reasons. It is questionable whether the folk concept of reasons is committed—
even implicitly—to technical claims about the supervenience of reasons on our
non-factive mental states.
A more plausible view is that the folk concept of reasons is exhausted by
various platitudes connecting reasons with other epistemic notions—e.g.,
‘A justified belief is supported by reasons’ (Section 7.4). Such platitudes are
consistent with both internalist and externalist construals of the relevant
notions, as well will now see.
7.6.2 From Reasons to Justification
Equipped with Reliable Reasons, reliabilists can embrace Pollock’s framework.

In particular, they can accept Justified Belief as Undefeated Reasoning: for a
belief to be ultima facie justified is for it to be the result of an ultimately

undefeated inference branch. But what this means is now given a reliabilist
interpretation. Let’s take this step by step.
Recall that in order for a belief B to be the result of an inference branch ψ, B
must be supported by the reasoning in ψ, and must be held as a causal
consequence of this reasoning. On the reliabilist interpretation advocated
here, an inference branch is just a chain of reliable or conditionally reliable
processes. If B is a foundational belief, then this chain consists of a single
reliable process applied to some non-doxastic state. If B is an inferential belief,
then the first link in this chain is a conditionally reliable process applied to
some further belief B0 , which is itself the result of a chain of reliable or
conditionally reliable processes.
Next, what does it mean for an inference branch to be ultimately
undefeated, on a reliabilist picture? Here too, reliabilists can accept Pollock’s
account. Following Pollock, reliabilists can explain this in terms of the notion
of being in at a level, which is in turn characterized in terms of when one
164  
inference branch defeats another. And they can go on to analyze what it is for
one inference branch to defeat another in terms of prima facie reasons. But
what it takes for there to be such a reason is now explained in reliabilist terms.
Call the resulting combination of Reliable Reasons and Pollock’s frame-
work, ‘Reasons Reliabilism.’ I now argue that this synthesis preserves the
primary advantages of both frameworks.²²
7.7. Problems Solved
7.7.1 A More Satisfactory Reasons-Based Framework
In Section 7.5, I raised some grounds for dissatisfaction with a ‘pure’ version of
Pollock’s Reasons First framework. The first concern was that a pure Reasons
First framework fails to accommodate the intuitions and commitments that
lend plausibility to reliabilism. The synthesis advocated here fares better in this
respect. First, it captures the intuition that there is an important connection
between justification and truth. On the approach defended here, for an agent
to have a reason to believe p is for them to have a reliable—hence truth-
conducive—process that is disposed to produce a belief in p. Second, our
synthesis remains faithful to reliabilism’s reductive ambitions. Rather than
resting content with a non-reductive analysis of justification in terms of the
notion of a reason to believe, Reliable Reasons shows how this notion can be
cashed out in non-epistemic notions. It thus fulfills the reliabilist goal of

providing a naturalistic account of justification.
The second concern for a pure Reasons First framework was that it lacks an
explanatorily satisfying account of what all reasons have in common. Why is it
that perception, memory, induction, and so on all provide reasons to believe?
²² My proposed synthesis bears comparison to the theory advanced in Graham and Lyons
(Forthcoming)—a proposal which was developed independently of the theory offered here. Graham
and Lyons suggest a theory of defeat built around the notion of prima facie warrants, which they define
in terms of available cognitive processes which would likely produce certain outputs if they were used.
In many respects, their approach is complementary to the account offered here. However, there are
some important differences. First, their theory is structured around the notion of warrants rather than
reasons—indeed, Graham and Lyons reject the need to appeal to reasons in a theory of justification. By
contrast, it is one of the contentions of the present essay that once we naturalize reasons, they pose no
threat to the reliabilist. Second, their account is not explicitly reliabilist: they define prima facie
warrants in terms of available processes that satisfy the requirements of prima facie justification,
without taking a stand on how best to cash out the latter requirements. Third, Graham and Lyons’
analysis of warrants takes a counterfactual form. This raises the worry that it will be subject to versions
of the counterfactual fallacy-style worries that beset ARP (e.g., the sort of worries raised in
Section 7.3.3). By contrast, I’ll be arguing shortly that Reasons Reliabilism avoids these concerns.
   165
Reliable Reasons answers this question. What all reasons have in common is
that they serve as the inputs to reliable (or conditionally reliable) processes.
The upshot: Reasons Reliabilism avoids the main concerns with a pure
Reasons First approach.
7.7.2 A More Satisfactory Treatment of Defeat
I’ll now argue that my proposed synthesis also preserves the main advantage of
Pollock’s framework—specifically, its superior treatment of defeat. To make
this point, let us revisit the difficulties facing ARP and see how Reasons
Reliabilism avoids them.
7.7.2.1 Defeater Defeat

The first difficulty for ARP was that it has trouble with defeater defeat. As we
have seen, Pollock’s definition of an undefeated inference branch is tailor-
made to handle such cases. Since Reasons Reliabilism makes use of Pollock’s
definition of an undefeated inference branch, it can reap the fruits of Pollock’s
labors.
To illustrate, recall Two Testimony Seeing Red. Reasons Reliabilists can
accept the inference graph we sketched for this case (Fig. 2). And they can say
all the things that we said earlier about this inference graph. In particular, they
can say that the inference branch responsible for Lori’s belief in  is
ultimately undefeated, since the only inference branch that defeats it is out
at every level 1.
However, Reasons Reliabilists do not stop there. They supplement this
formal representation of Lori’s reasons with a reliabilist account of where
the nodes come from, and why the various support and defeat links hold.
According to Reliable Reasons, the reason why the experience of receiving
Anne’s testimony is a prima facie reason to believe that Sal is a compulsive liar
is that there is a reliable process (believing reliable interlocutors) available to
Lori that, when fed this experience as input, is disposed to produce a belief that
Sal is a compulsive liar. Similar remarks apply, mutatis mutandis, to the other
nodes depicted in the graph.
7.7.2.2 Circularity Worries

The second difficulty was that ARP turns out to be circular. To recap: the
worry was that the proper way of unpacking ARP will rely on the notion of
ultima facie justification. But ARP is used to articulate the conditions under
166  
which a belief is undefeated—a concept that occurs in the base clause of

Recursive Reliabilism.
Reasons Reliabilism avoids this worry. Following Reasons Firsters, we
define ultima facie justification in terms of the notion of a prima facie reason
for believing. And we then recursively define this notion in terms of the inputs
to various belief-forming processes. Crucially, the base clause of this definition
(Reliable Reasons) does not itself rely on the notion of defeat, or any other
epistemic notion for that matter.²³
7.7.2.3 Alternative Processes that One Should Not Use

The final difficulty for ARP came from cases where an agent has an alternative
reliable process that they have no good reason to use. In Thinking About
Unger, ARP predicts that Harry’s belief in  is defeated merely in virtue of
the fact that he has an available reliable process (his Unger Predictor) which,
were he to use it, would lead him to suspend judgment on whether there’s a
tree in front of him.
The theory advocated here avoids this prediction. To see this, recall that if
Harry’s Unger Predictor were fed Harry’s current states as input, it would
produce as output a belief in the proposition: Unger would advise me to
suspend judgment on . As we saw in Section 7.3.3, this belief could itself
be viewed as the input to a further process, which leads Harry to suspend
judgment regarding . But neither process would produce either a belief in
¬ or a belief that Harry’s current states do not reliably indicate the truth
of . And so, according to Reliable Reasons, Harry doesn’t have either a
rebutting or an undercutting defeater for his belief in .
Perhaps, some may suggest, the problem re-emerges if we amend the case.
Meet Elijah the eliminativist. Elijah thinks trees do not exist, and he wants
others to share this belief. Suppose that Harry has a highly reliable Elijah
Predictor, which is part of an Elijah Emulator: if he were to predict that
Elijah would advise him to believe p in his current situation, this would in
²³ At the same time, Reasons Reliabilism respects Fumerton’s claim that a conditionally reliable
process can only serve as a defeater if it is applied to ultima facie justified inputs. Proof: Suppose A’s
inference graph contains some node n, and suppose that A unjustifiably believes some proposition q
that, when fed into a conditionally reliable process, is disposed to produce a belief in d, where d is either
of the form, ¬n or n’s immediate ancestors do not reliably indicate n. Since A’s belief that q is unjustified,
it follows (from Justified Belief as Undefeated Reasoning) that either (i) A’s belief in q is not the result of
any inference branch, or (ii) it is the result of some inference branch, but one of the branch’s nodes is
ultimately defeated. If (i), then A doesn’t even have a prima facie reason to believe q, and so (by Reliable
Reasons) q is not a prima facie reason to believe d. If (ii), then q is a prima facie reason to believe d, but q
is ultimately defeated. Either way, n is not ultimately defeated.
   167
turn lead him to believe p. And so, if Harry were to apply his current
experiential states to his Elijah Predictor, he would be led to believe ¬.
Does Reasons Reliabilism predict that this gives Harry a rebutting defeater
for his belief in ?
An initial point: given this way of describing the case, it is doubtful whether
Harry’s Elijah Emulator is reliable. After all, it systematically misleads him
about the presence of trees, which hardly bodes well for its reliability!
However, the objector might try to circumvent this point by tempering
Elijah’s eliminativism. Just stipulate that, much like Shmunger before him,
Elijah has entirely correct beliefs about all sorts of topics—astronomy, geog-
raphy, physics, whatever. It is only when it comes to trees that Elijah is an
eliminativist. And so, Harry’s Elijah Emulator is overall reliable, even though it
is unreliable on arboreal matters.
Suppose we grant all of this. Then Reasons Reliabilism does indeed predict
that Harry has a rebutting defeater for his belief in . However, this need
not worry us, provided that this defeater is itself defeated. Consider: why,
exactly, would Harry be unjustified in using his Elijah Emulator in his current
situation? Presumably, because he has good reason to think that his visual
experience of trees reliably correlates with the presence of trees. Where does
this reason come from? Presumably from his past experiences, which support
the generalization that having a visual experience representing x is a reliable
indicator of the presence of x. Plausibly, there is a reliable process available to
him that, given these past experiences as input, is disposed to produce a belief
that his current tree-like experiences reliably indicate  rather than ¬.
If so, these past experiences constitute an undercutting defeater for his rebut-
ting defeater for believing  (see Fig. 3). And so, his belief in  counts as
ultima facie justified, as desired.
Could proponents of ARP co-opt this response? No. After all, the key move
here is to diagnose this variant scenario as a case of defeater defeat. But we’ve
seen that ARP lacks an adequate treatment of defeater defeat.
Harry’s Elijah would

visual experience ve ea pe past experiences
advise believing ¬t
Tree in ea doesn’t
front of Harry
t ¬t u
reliably indicate ¬t
Fig. 3 Thinking about Eliminativism
168  
7.7.3 Capturing the Role of Reasons in Justification
The primary payoff of recasting reliabilism in terms of reasons is that it

provides a satisfactory treatment of defeat. However, a further benefit is also
worth noting.
Historically, reliabilism has had little to say about reasons for belief. While it
has not denied their existence, it has maintained a conspicuous silence about
their nature.²⁴ But clearly there are such reasons, and any complete epistem-
ology should have something to say about them.
Reasons Reliabilism offers a natural way of bringing reasons into the
reliabilist fold. And it does so in a way that underwites the intuitive connec-
tions between reasons and justification.
7.8. Comparison with Evidentialist Hybrids
7.8.1 The Two-Component View
I’ve advocated integrating reliabilism with a reasons-based framework—a

framework that has traditionally been viewed as a rival to reliabilism. In doing
so, I may appear to be joining my voice to a rising chorus. In recent years, a wave
of authors has suggested integrating reliabilism with evidentialism—a view that
has also been viewed as a competitor to reliabilism.²⁵ How, then, does my
synthesis differ from more familiar evidentialist-reliabilist hybrids?

To answer this, it will be helpful to look in some detail at how evidentialist-
reliabilists handle defeat. On a standard evidentialist view, justification and
defeat are explained in terms of evidential support. A belief is prima facie
justified when it is supported by some initial body of evidence e₁. Defeat occurs
when the agent acquires further evidence e₂ which, when combined with e₁, no
longer supports the belief.
There are various ways one could integrate this general approach with
reliabilism. A particularly straightforward strategy is suggested by Goldman
(2011), who proposes that ultima facie justification involves two components:
a reliable process condition and an evidential support condition. That is:
²⁴ This is at least true of process reliabilists in the tradition of Goldman (1979, 1986). The work of
Tyler Burge offers a very different sort of reliabilist approach in which reasons play an important role.
See e.g., the papers collected in Burge (2013).
²⁵ See the references in fn. 5.
   169
Two Component View: A’s belief that p is ultima facie justified iff both:
Reliable Process Condition: A’s belief that p is the result of a reliable belief-
forming process,
Evidential Support Condition: A’s total evidence supports believing p.²⁶,²⁷
It is easy to see, at least in broad brushstrokes, how the Evidential Support

Condition helps with defeat. Take Seeing Red. When Lori first has the visual
experience of a red-appearing wall, her total evidence supports believing ,
hence her belief is ultima facie justified. But once she receives Sal’s testimony,
her total body of evidence expands. This more inclusive body of evidence no
longer supports believing .
As Miller (2019) notes, a view along these lines also avoids many of the
problems facing ARP. Take defeater defeat: when Lori acquires Anne’s testi-
mony in Two Testimony Seeing Red, her total body of evidence changes once
again, and  regains its former level of support. Or take Thinking About
Unger: arguably, Harry’s total evidence supports believing , despite the
availability of his Unger Predictor.
Given these virtues, is there any reason to prefer Reasons Reliabilism to the
Two Component View? While a full adjudication of this issue is beyond
the scope of this paper, I want to briefly raise two reasons for thinking that
the answer is ‘yes.’
7.8.2 First Advantage: Reductive and Predictive
As it stands, the Two Component View is not reductive. After all, the
Evidential Support Condition packages together two epistemic notions:
(i) The notion of an agent’s total evidence,²⁸

(ii) The notion of a body of evidence supporting a belief.
²⁶ The Two Component View could be complicated in a number of ways. For example, we might
hold that the two conditions are not entirely independent. Rather, we might follow Comesaña (2010) in
holding that the reliable process needs to take evidence as input. (According to Comesaña, this helps
with both the generality problem and Bonjour’s case of Norman the clairvoyant.) For my purposes,
I will set such complications aside.
²⁷ Hybrid views that impose some version of an Evidential Support Condition include Tang (2016);
Comesaña (2018); and Miller (2019).
²⁸ Note that the challenge here is not just to give an account of evidence in non-epistemic terms. It’s
to give an account of what it is for an agent to possess evidence in non-epistemic terms. See Beddor
(2015b, 2016: Ch. 1) for discussion of some difficulties on this front.
170  
Perhaps, some might suggest, this just shows that we need to supplement
the Two Component View with a reductive analysis of these notions. To do so,
proponents of the Two Component View could try taking a page from the
Reasons Reliabilists. According to Reliable Reasons, prima facie reasons are
the inputs to reliable and conditionally reliable processes. Why not say the
same about evidence? An agent’s total evidence, on this view, is the total set of
states of the agent that can serve as potential inputs to the reliable processes
available to the agent.
This would give us a reductive analysis of (i). What about (ii)? According
to one common approach, evidential support should be understood in
probabilistic terms: a body of evidence e supports believing p just in case
the probability of p given e is sufficiently high. Putting these two suggestions
together, we get:
Evidential Support Condition (Unpacked): A’s total evidence supports

believing p iff Pr(p|A is in states s₁ . . . sn) >t, where
• s₁ . . . sn are all the states of A that can potentially serve as inputs to A’s
reliable belief-forming processes,
• t is some threshold.
What sort of probability is at issue here? One option would be to define Pr in

epistemic terms: for example, we could say that Pr reflects the credences that
are justified by the evidence. But clearly this is to give up any reductive
ambitions.²⁹ Perhaps, then, we should follow Tang (2016) in taking Pr to
reflect objective probabilities. This would result in a reductive theory, but it
gives rise to a further concern: Is the theory predictive? And do the predictions
vindicate our pretheoretic judgments about defeat?
To flesh out this concern, go back to Seeing Red. On the view under
consideration, to determine whether Lori’s belief is defeated by Sal’s testi-
mony, we check the objective probability of  conditional on Lori’s total
post-Sal-testimony evidence. But how do we check this? Perhaps via intuition,
but it is questionable whether we have clear-cut intuitions about such prob-
abilities. And things only get worse when consider defeater defeat. In Two
Testimony Seeing Red, how do we determine the objective probability of 
conditional on Lori’s post-Sal+Anne-testimony evidence? The worry, then, is
²⁹ See Comesaña (2018), who embraces this consequence.
   171
that while this version of the Two Component View may be consistent with
our intuitions about defeat, it does not yet predict these intuitions.
While this is hardly the final word on the matter, these considerations
highlight the hurdles that arise when we try to develop the Two Component
View in a way that is both fully reductive and predictive.
7.8.3 Second Advantage: No Immunity to Defeat
The second advantage of Reasons Reliabilism over the Two Component View
stems from a structural difference between the two frameworks. One hallmark
of Pollock’s framework is that nothing is exempt from defeat. For any prop-
osition p, it’s easy to concoct a defeater for p—just imagine a reliable source
tells you either ¬p or that your reasons do not support p. By contrast, the
Evidential Support Condition singles out a class of propositions that get a free
pass from defeat. Let me explain.
It’s a familiar observation that whenever some part of an agent’s evidence
entails p, their total evidence also entails p, hence the probability of p condi-
tional on their total evidence is 1. So, probabilistic approaches to defeat predict
that one cannot have a defeater for something entailed by any part of one’s
current evidence:
Limited Indefeasibility: if some subset of A’s evidence entails p at t, then A’s

total evidence supports p at t.
At least two sorts of cases suggest that Limited Indefeasibility runs contrary to
intuition. First, consider cases where an agent has a defeater for a belief in a
necessary truth:
Logical Luck: Tom comes up with a soundproof of a particular logical

theorem L. Sometime later, he is told by his highly accomplished logic
professor that his proof contains a mistake. Tom nonetheless disregards her
testimony, continuing to believe L on the basis of his proof.
Intuitively, the professor’s testimony provides an undercutting defeater for

Tom’s belief in L. But any view that validates Limited Indefeasibility cannot
account for this intuition: since L is a necessary truth, Tom’s total evidence
trivially entails L.
172  
Some might regard necessary truths as a special case, to be dealt with via
independent means.³⁰ Still, a second class of counterexamples remains: cases
where one has a defeater for some proposition that is itself part of one’s evidence.
According to the view of evidence under consideration, one’s evidence consists in
various states that serve as inputs to reliable processes. But why think that these
states enjoy some special exemption from defeat? Consider:
Emotional Introspection: Kilian is happy for his brother, who recently

received a promotion. By introspection, Kilian comes to justifiably believe, I
am happy for my brother (). Later that day, he has a therapy session with
an extremely well-credentialed psychiatrist, who tells him that he is
mistaken: Kilian is actually jealous of his brother; he is simply unwilling to
acknowledge this. While the psychiatrist mounts a compelling argument,
Kilian ignores her, continuing to believe .
Kilian’s total evidence includes his happiness for his brother, since this state
serves as the input to a reliable process (introspection). And this experience
entails .³¹ Limited Indefeasibility thus predicts that his justification for
this belief is undefeated. But this seems wrong. Even though the psychiatrist is
mistaken, her testimony still provides a rebutting defeater for his belief.³²
Reasons Reliabilism fares better here, since it is not committed to Limited
Indefeasibility. Even if p is entailed by one of your reasons, you could still have
a reliable process that is disposed to deliver either a belief that ¬p, or a belief
that your basis for believing p does not reliably indicate its truth. In Logical
Luck there is a reliable process available to Tom (trusting the testimony of
³⁰ For example, some might take a page from Stalnaker’s (1999) strategy for handling the problem of
logical omniscience. According to this proposal, our intuitions about Logical Luck are not really
tracking Tom’s justification for believing L, but rather Tom’s justification for believing some contingent
proposition associated with L—for example, the proposition: SL is true, where SL is some sentence that
expresses L. However, even if this strategy can be made to work, it is a mark in favor of Reasons
Reliabilism that it has no need of such maneuvers.
³¹ It entails it both in the sense that its content (trivially) entails , and in the sense that the fact
that Killian has this experience entails . So regardless of whether we understand evidential
support in terms of probabilities conditional on the contents of an agent’s states or in terms of
probabilities conditional on the fact that the agent is in these states, the counterexample goes
through.
³² Faced with this counterexample, one option would be to try to complicate the Evidential Support
Condition. Perhaps in order for A’s belief that p to satisfy the Evidential Support Condition, we should
also require that every sufficiently similar body of evidence would also support believing p. (Cf. Miller
(2019), who suggests quantifying over ‘partial psychological duplicates’ to handle this sort of case.)
According to this diagnosis, Killian could have been in an internally indistinguishable mental state that
only included the appearance of happiness. Had he been in this state, his total evidence arguably would
not have supported . However, going this route raises a difficult question: can we give a
principled and reductive characterization of when a body of evidence is “sufficiently similar”?
   173
experts) that, when applied to the experience of receiving his professor’s

testimony, is disposed to produce a belief that his proof is not a reliable
guide to the truth about L. Similarly, in Emotional Introspection the same
reliable process is available to Kilian. When applied to the experience of
receiving his psychiatrist’s testimony, this process is disposed to produce a
belief in ¬. For the Reasons Reliabilist, no belief is immune to defeat.
7.8.4 Taking Stock
Reasons Reliabilism has certain affinities with extant hybrids of evidentialism

and reliabilism: both are attempts to meld reliabilism with theoretical frame-
works that are usually associated with internalism. However, there are import-
ant theoretical differences between the two approaches. I’ve given some reason
to think that these differences speak in favor of Reasons Reliabilism. First,
Reasons Reliabilism is both reductive and predictive, whereas it proves diffi-
cult to develop the Two Component View in a way that enjoys both these
virtues. Second, the Two Component View—and probabilistic approaches to
defeat more generally—grant a certain class of propositions a principled
exemption from defeat. Reasons Reliabilism bestows no such favors.
Of course, given the affinities between the two approaches, some may be
inclined to classify Reasons Reliabilism as a type of hybrid view. Should they do
so, I would raise no objection. The important point is that if it is a hybrid view, it
is the most promising one to date, at least when it comes to handling defeat.³³
7.9. Conclusion
For most of their history, the reliabilist tradition and the Reasons First
tradition have been developed in isolation from each other. In this paper,
I’ve argued that an integration of the two approaches proves mutually bene-
ficial. The synthesis developed here avoids reliabilism’s difficulties with defeat,
³³ Another view that bears some resemblance to Reasons Reliabilism is the Reasons First virtue
epistemology developed in Sylvan and Sosa (2018). According to Sylvan and Sosa, facts about what an
agent is justified in believing are determined by facts about what she has sufficient epistemic reason to
believe, which are in turn determined by facts about her competent attractions to assent to various
propositions. While there are a number of similarities between the two approaches, there is also a
crucial difference. Sylvan and Sosa do not offer their Reasons First virtue epistemology as a reductive
approach. Rather, they take the notion of a ‘competent attraction to assent’ to be a normatively loaded
notion—a notion that cannot be reduced to talk of reliable processes.
174  
while still preserving the explanatory advantages that make reliabilism

attractive.
While I have focused on justification, my conclusions also have implications
for the study of knowledge. Recently, some authors sympathetic to externalism
have argued for the surprising conclusion that knowledge is indefeasible.³⁴
One argument for this bold conclusion is that we have no satisfactory exter-
nalist story about how knowledge defeat works.³⁵ The view developed in this
paper shows one way developing such a story. As long as justification is a
necessary condition on knowledge, we can use this account of justification to
explain how knowledge is likewise subject to a defeat condition.³⁶
References
William Alston. (1988). An internalist externalism. Synthese, 74: 265–83.

Max Baker-Hytch and Matthew Benton. (2015). Defeatism defeated. Philosophical
Perspectives, 29: 40–66.
Bob Beddor. (2015a). Process reliabilism’s troubles with defeat. The Philosophical
Quarterly, 65(259): 145–59.
Bob Beddor. (2015b). Evidentialism, circularity, and grounding. Philosophical
Studies, 172: 1847–68.
Bob Beddor. (2016). Reduction in Epistemology. PhD thesis, Rutgers University.
Matthew Bedke. (2010). Developmental process reliabilism: On justification,
defeat, and evidence. Erkenntnis, 73(1): 1–17.
Michael Bergmann. (2006). Justification without Awareness. Oxford: Oxford

University Press.
Laurence Bonjour. (1985). The Structure of Empirical Knowledge. Cambridge, MA:
Harvard University Press.
Tyler Burge. (2013). Cognition Through Understanding. Oxford: Oxford
University Press.
Roderick Chisholm. (1966). Theory of Knowledge. Englewood Cliffs, NJ: Prentice
Hall.
³⁴ See e.g., Lasonen-Aarnio (2010a); Baker-Hytch and Benton (2015).

³⁵ See, in particular, Baker-Hytch and Benton (2015), who use the failure of ARP as a premise in an
argument for the indefeasibility of knowledge.
³⁶ For helpful comments on this material, I am grateful to Simon Goldstein, Peter Graham, Jack
Lyons, Carlotta Pavese, Mona Simion, participants in the NUS Philosophy reading group, and
audiences at Hong Kong University, the Bled Epistemology Conference, the Goldman Retirement
Conference at Rutgers University, and the Knowledge First Epistemology Workshop at Cardiff
University. Special thanks are due to Alvin Goldman for many illuminating conversations about
reliabilism and defeat.
   175
David Christensen. (2010). Higher-order evidence. Philosophy and Phenomenological

Research, 81(1): 185–215.
Stewart Cohen. (1984). Justification and truth. Philosophical Studies, 46: 279–96.
Juan Comesaña. (2010). Evidentialist reliabilism. Noûs, 44(4): 571–600.
Juan Comesaña. (2018). Whither evidentialist reliabilism? In: Kevin McCain (ed.),
Believing in Accordance with the Evidence. New York: Springer: 307–25.
Richard Fumerton. (1988). Foundationalism, conceptual regress, and reliabilism.
Analysis, 48(4): 178–84.
Alvin Goldman. (1979). What is justified belief? In: Pappas (ed.) Justification and
Knowledge. Dordrecht: Reidel.
Alvin Goldman. (1986). Epistemology and Cognition. Cambridge, MA: Harvard
University Press.
Alvin Goldman. (2011). Towards a synthesis of reliabilism and evidentialism. In:
Dougherty (ed.) Evidentialism and its Discontents. New York: Oxford
Alvin Goldman. (2012). Reliabilism and Contemporary Epistemology: Essays. New
York: Oxford University Press.
Alvin Goldman and Bob Beddor. (2015). Reliabilist epistemology. Stanford
Encyclopedia of Philosophy, https://plato.stanford.edu/archives/win2016/
entries/reliabilism/.
Peter Graham and Jack Lyons. The structure of defeat: Pollock’s evidentialism,
Lackey’s distinction, and prospects for reliabilism. In: Simion and Brown (ed.)
Reasons, Justification and Defeat. Oxford: Oxford University Press,
(Forthcoming).
Thomas Grundmann. (2009). Reliabilism and the problem of defeaters. Grazer
Philosophische Studien, 79(1): 65–76.
David Henderson, Terry Horgan, and Matjaž Potrč. (2007). Transglobal
evidentialism—reliabilism. Acta Analytica, 22: 281–300.
James Joyce. (1998). A nonpragmatic vindication of probabilism. Philosophy of
Science, 65(4): 575–603.
Jaegwon Kim. (1988). What is “naturalized epistemology”? Philosophical
Hilary Kornblith. (2002). Knowledge and its Place in Nature. Oxford: Clarendon
Press.
Jennifer Lackey. (1999). Testimonial knowledge and transmission. The Philosophical
Quarterly, 49(197): 471–90.
Maria Lasonen-Aarnio. (2010a). Unreasonable knowledge. Philosophical
176  
Maria Lasonen-Aarnio. (2010b). Is there a viable account of well-founded belief?

Erkenntnis, 72: 205–31.
Jack Lyons. (2009). Perception and Basic Beliefs. Oxford: Oxford University Press.
Jack Lyons. (2013). Should reliabilists be worried about demon worlds? Philosophy
and Phenomenological Research, 86: 1–40.
Jack Lyons. (2016). Goldman on evidence and reliability. In: Kornblith and
McLaughlin (eds.) Goldman and his Critics. Oxford: Blackwell.
Emelia Miller. (2019). Liars, tigers, and bearers of bad news, oh my! Towards a
reasons account of defeat. The Philosophical Quarterly, 69(274): 82–99.
Sarah Moss. (2011). Scoring rules and epistemic compromise. Mind, 120(480):
1053–69.
Richard Pettigrew. (2016). Accuracy and the Laws of Credence. Oxford: Oxford
University Press.
Richard Pettigrew. (2018). What is justified credence? Episteme, 1–18.
John Pollock. (1987). Defeasible reasoning. Cognitive Science, 11: 481–518.
John Pollock. (1992). How to reason defeasibly. Artificial Intelligence, 57: 1–42.
John Pollock. (1994). Justification and defeat. Artificial Intelligence, 67: 377–408.
John Pollock. (1995). Cognitive Carpentry. Cambridge, MA: MIT Press.
John Pollock. (2001). Defeasible reasoning with variable degrees of justification.
Artificial Intelligence, 133(2): 233–82.
Henry Prakken and John Horty. (2012). An appreciation of John Pollock’s work
on the computational study of argument. Argumentation and Computation, 3:
1–19.
Robert Stalnaker. (1999). Context and Content. New York: Oxford University
Press.
Kurt Sylvan and Ernest Sosa. (2018). The place of reasons in epistemology. In: Star
(ed.) The Oxford Handbook of Reasons and Normativity. Oxford: Oxford
University Press.
Weng Hong Tang. (2016). Reliability theories of justified credence. Mind, 125
(497): 63–94.
8
Knowledge, Action, and Defeasibility
Carlotta Pavese
8.1. Introduction
One can intentionally do something only if one knows what one is doing while
they are doing it. For example, one can intentionally kill one’s neighbor by
opening their gas stove overnight only if one knows that the gas is likely to kill
the neighbor in their sleep. One can intentionally sabotage the victory of one’s
rival by putting sleeping drugs in their drink only if one knows that sleeping
drugs will harm the rival’s performance. And so on. In a slogan: Intentional
action is action guided by knowledge.¹
This essay reviews some motivations for a ‘knowledge-centered psychology’—
a psychology where knowledge enters center stage in an explanation of inten-
tional action (Section 8.2). Then it outlines a novel argument for the claim that
knowledge is required for intentional action (Section 8.3) and discusses some of
its consequences for the debate about know-how. Section 8.4 argues that a
knowledge-centered psychology motivates the intellectualist view that know-

how is a species of know-that. In its more extreme form, the view is committed to
an epistemologically substantial claim—i.e., that the epistemic profile of know-
how is the same as that of propositional knowledge. Now, it is widely believed
that know-that can be defeated by undermining and rebutting defeaters (e.g.,
Chisholm 1966; Goldman 1986; Pollock and Cruz 1999; Bergmann 2000). If that
is correct, one corollary of intellectualism is that the defeasibility of know-how
patterns with that of knowledge. A knowledge-centered psychology does predict
that, for it predicts that both know-how and knowledge are defeated when one’s
ability to intentionally act is defeated. In Section 8.5, by replying to a challenge
raised in the recent literature (Carter and Navarro 2018), I argue that this
prediction is actually borne out.
¹ I am grateful for comments to Adam Carter, Clayton Littlejohn, and Mona Simion.
780198847205.003.0008
178  
8.2. Towards a Knowledge-Centered Psychology
A long tradition in the philosophy of mind assigns beliefs a central role in folk
psychological explanations of intentional behavior (e.g., Stich 1978; Fodor
1987; Lewis 1976; Stalnaker 1984; Humberstone 1992). More or less explicitly,
this tradition confines psychological explanations to an explanation of
attempts. Consider the usual example of a psychological explanation, where
one’s belief that there is water in the fridge and one’s desire to drink it together
are supposed to explain one’s attempt to grab a bottle of water from the fridge.
Success happens when the belief is true—when there is indeed water in the
fridge. If one’s belief is true, then one will succeed at finding a bottle; if one’s
belief is false, one will not succeed at finding water. The dominant thought
behind a belief-centered psychology is the idea that, as far as the psychological
explanation of behavior goes, whether the world makes the belief true (e.g.,
whether there is water in the fridge) is irrelevant: what is to be explained is the
fact that one attempted to get water from the fridge, whether or not one has
succeeded. And one’s belief that there is water in the fridge, together with one’s
desire to drink it, suffices to explain one’s attempt, whether or not the belief is
true.²
Let us get a bit clearer about the underlying assumption of a belief-centered
psychology. Let a condition be something that obtains or fails to obtain at a
case, and let a case be a centered possible world or situation—an ordered triple
of a subject, a time, and a location. Some conditions are mental, such as that
I feel pain or believe that it is raining. Other conditions are non-mental, such
as that I have broken my leg. Beliefs are mental conditions and so are
attempts—e.g., my trying to get on the bus is a mental condition.³ Like beliefs,
attempts are non-factive mental conditions: an attempt to ϕ does not entail
successful ϕ-ing, for an attempt might be successful or might fail. As a non-
factive condition, an attempt does not encompass those external aspects of the
world that make for an agent’s success.
The assumption that psychological explanation should be confined to
explaining attempts relies on the idea that actions are decomposable into
mental conditions and non-mental conditions—into attempts, on the one
hand, and into those external conditions that makes for the agent’s
success, on the other. Call this the decomposability assumption and suppose
it is true. If so, we can appreciate one of the main motivations that underlie
a belief-centered psychology. A psychological explanation of intentional
² See in particular Stich (1978). ³ This terminology follows closely Williamson (2000: Ch. 3).
, ,   179
behavior, understood as attempts, does not need to appeal to anything more

than to non-factive mental states—i.e., such as mere beliefs. If when we are
explaining behavior, all we are trying to explain is an attempt (e.g., the
attempt at finding a bottle of water in the fridge) rather than one’s intentional
success (e.g., one’s intentional success at finding the bottle in the fridge) then
all we need is a psychological theory that encompasses non-factive attitudes
such as beliefs and desires (e.g., one’s belief that water is in the bottle and the
desires to drink).
But is the decomposition assumption true? Indeed, there are good reasons to
think that actions are not decomposable into mental and non-mental compo-
nents and that even if attempts were in some sense components of actions, the
mentality of actions would not be exhausted by the mentality of attempts. Here
is an argument for this conclusion. If attempts exhausted the mentality of
actions, then provided that one attempted to ϕ, one’s eventual success at ϕ-ing
would have to be intentional. For on this picture, the intentionality, and hence
the mentality, of an action would be exhausted by its attempt. However, there
are a variety of cases in which one attempts at ϕ-ing, succeeds, and yet fails to
act intentionally. That suggests that the intentionality of actions cannot be
reduced to the intentionality of attempts; and intentionality being a mark of
the mental (e.g., Brentano 1874/1995: 68), that suggests that the mentality of
actions cannot be reduced to the mentality of attempts.
A well-known example in action theory which illustrates how attempts and
intentionality can come apart is Mary the Bomber (cf. Mele and Moser 1994;
Gibbons 2001):
Mary the Bomber: Mary intends to kill her uncle by setting off a bomb by a
bomb in his house and then, after moving a safe distance away, pressing the
large red button on the remote-control device. She does not know much
about how these things work and thinks that pressing the button will cause
the bomb to detonate but has no idea about the details of this process. Her
belief is true and justified and here is what happens. A satellite, launched by
the National Security Agency and designed to prevent bombings of just this
kind, intercepts Mary’s transmission which causes the satellite to send a
warning to the intended victim. But, because of an unfortunate choice of
frequency, this causes the bomb to detonate.
Mary killed her uncle and caused the bomb to detonate and did intend both
things. But she did not do either of these things intentionally. Hence, the
intentionality and mentality of this action is not reducible to the intentionality
180  
and mentality of the attempt. This residual ‘mentality’ of actions calls for an
explanation—and presumably an explanation that reduces the intentionality
of the action to some mental state of the agent. And for this sort of explan-
ation, belief alone cannot suffice.⁴ For intentional action is a factive mental
condition: if one intentionally ϕs, then (trivially) one ϕs. And, if the mental
condition to be explained is factive in the way intentional action is factive, then
its explanation calls for a factive condition. Non-factive (attempts) might be
explainable by non-factive (i.e., beliefs). But factive (intentional actions) ought
to be explained by factive (i.e., knowledge).
Mary the Bomber can be accounted for by a knowledge-centered psych-
ology (Gibbons 2001). Mary does not really know that she can provoke the
explosion by implementing her plan. That is why her success is too coinci-
dental to count as intentional. More generally, the prediction of a knowledge-
centered psychology is that if one’s belief is Gettiered, then one cannot act
intentionally on that belief. This prediction is borne out. To see this, consider
two more examples. The next is also from Gibbons (2001) but slightly revised:⁵
Cindy and the Lottery: Cindy mistakenly believes that someone rigged a
lottery in her favor and that she will be handed the winning ticket at the
ticket store. On this basis, she believes of a particular ticket that is being
handed to her, that if she buys it, she will win. She buys the ticket and wins.
So, her belief that she will win the lottery by buying that ticket is true. It is
even justified. Buying a winning ticket is a perfectly reliable way of winning a
lottery. Still, intuitively, she did not intentionally win the lottery.
Here again, Cindy intends to win the lottery and attempts to do it by buying a
ticket that she believes truly and justifiably to be a winning ticket. This case
differs from the previous one, for here no deviant causal chain plays a role in
explaining her success (i.e., in making her ticket win). And yet, Cindy’s victory
is not intentional: fair lotteries cannot be intentionally won. That suggests,
once again, that the intentionality of actions is not exhausted by the inten-
tionality of attempts. But if so, we need a sort of psychology that differs from a
belief-centered psychology in that it is not only tailored to explaining attempts.
⁴ Cf. also Levy (2013) and Williamson (2017) for similar remarks about the non-decomposability of
actions. Gibbons (2001) also argues that psychological explanations should explain intentional suc-
cesses and not merely attempts.
⁵ Revised in order to overcome an objection raised by Cath (2015) to Gibbons’ original example. See
Pavese (2018) for a detailed discussion of Cath (2015) on Gibbons. Cath’s objection does not extend to
Cindy and the Lottery as presented in the text, nor does it extend to the next example from Pavese
(2019)—i.e., Daniel and the Barn.
, ,   181
As a final example supporting the claim that intentional action requires

more than true belief or even justified true belief, consider the following
variation on Carl Ginet’s fake barn case (Pavese 2019; Beddor and Pavese
2019):⁶
Daniel and the Barn: along the road to Larissa, Daniel is instructed to stop at
any barn that he finds and to await further instructions. However, the road to
Larissa passes through fake-barn county. Daniel passes the first barn-looking
construction but does not stop for he thinks “There is not enough shadow for
me to wait comfortably.” At the second barn-looking construction, Daniel
stops and parks. As it turns out, only the second barn-looking construction
was a real barn.⁷
In Daniel and the Barn, Daniel ends up stopping at a barn and intended to
do so. However, intuitively, he did not intentionally stop at a barn. In fact,
he would have easily stopped at a fake barn, had the shadow been present
there.
In all Mary the Bomber, Cindy and the Lottery, and Daniel and the Barn,
the success of the agent does not count as intentional. This fact can be
explained on a knowledge-centered psychology. If one possesses knowledge,
then one’s belief cannot be lucky (Sosa 1999; Williamson 2000).⁸ In all of these
examples, the subject’s belief turns out to be true by luck and the same sort of
luck undermines the intentionality of their actions.⁹ A knowledge requirement
on intentional action can explain why the luck of their successes can under-
mine the intentionality of Mary’s, Daniel’s, and Cindy’s successes: by under-
mining their knowledge. In contrast, it is unclear that a belief-centered
psychology has the resources to explain why these actions are not intentional.
⁶ Cf. also Goldman (1976).

⁷ Some take fake barn cases to be cases of knowledge (see e.g., Sosa 2007: 31), for they think in those
cases, the success of the belief is clearly due to ability. Many, however, disagree about this diagnosis,
either on the grounds that it is too little intuitive (e.g., Pritchard 2012; Beddor and Pavese 2018) or on
the ground that there is a clear virtue-theoretic rationale for thinking that success in fake barn cases is
not attributable to ability. For the latter style of argument, see Littlejohn (2014).
⁸ Some object to a modal requirement on knowledge. See Beddor and Pavese (2018) for a recent
defense.
⁹ It is important to register that not all luck is epistemically harmful. The kinds of luck that seem
most clearly malignant are the ones that seem to show that it is at least partially accidental that a belief
‘turns out’ to be accurate, correct, and so on. When it is lucky that someone believes but not lucky that
the belief (given the conditions under which its formed) is correct, the luck does not seem to be
malignant. Something similar seems to hold for tryings, attempts, and intentional actions. It seems to
matter that the connection between trying and succeeding is lucky or accidental, not whether it’s lucky
or accidental that the person tried. Thanks to Clayton Littlejohn for the discussion.
182  
But could one act intentionally on the basis of extremely high probabilistic
evidence that however falls short of knowledge?¹⁰ Consider:
Mary, the Bomber, and the Lottery: Mary intentionally plays a trillion ticket
lottery such that the bomb fails to go off and fails to kill her uncles only if she
wins the lottery (and let’s suppose the satellite stuff is left out of the story).
Like in standard lottery cases, arguably Mary does not know that she will lose
the lottery (Williamson 2000). But, the objection goes, it seems as though
Mary could intentionally kill her uncle by simply intentionally playing that
lottery. This is the case even though (like in Gettier cases) she doesn’t know
that playing the lottery will detonate the bomb.
In response, granted, Mary does not and cannot know that playing the
lottery will detonate the bomb. But there is another thing that she does know—
i.e., that it is sufficiently likely that playing the lottery will detonate the bomb,
for she knows that it is sufficiently likely that she will lose the lottery. In fact, as
I will explain later in the essay, it is independently plausible that probabilistic
knowledge of this sort is central to an explanation of intentional action (Pavese
2020). To see this, consider Davidson’s (1971: 50) example:
Carbon Copying: I have a stack of carbon paper in front of me. In order to

save time, I try to sign the top page of the stack with enough force to that
I sign all the copies simultaneously. I do this despite the fact that I do not
believe that I will succeed. In this case, I might succeed in this endeavor while
failing to know that my signature would be legible on the last page because
I do not have sufficient confidence for full belief.
This seems to be a case of intentional action. But it is tempting to say in this

case that I do not have knowledge because of my lack of confidence. This would
be too quick, however. As several people have argued in the recent literature
(Weisberg 2013; Moss 2018), credences can amount to knowledge too.
Although I do not know that I will succeed at signing all the carbon copies,
I still estimate that I will succeed with some probability. And although I do not
know that I will succeed, I still know that it is sufficiently likely that I will. And,
knowing that requires believing that it is sufficiently likely that I will. This belief
can be modeled as a credence of sufficiently high degree that I will sign all the
carbon copies. If that is correct, then Davidson’s (1971) carbon copy
¹⁰ Thanks to Adam Carter for the example.
, ,   183
example—or Mary, the Bomber, and the Lottery, or other similar

examples¹¹—raise no issue for a knowledge requirement on intentional action,
for what guides intentional action in these cases can still be knowledge, albeit
probabilistic.
In conclusion, explanations of intentional action do seem to essentially
appeal to knowledge. These explanations are plausibly psychological in nature,
for they aim at explaining a factive mental condition (intentional action) in
terms of a more primitive factive mental condition (the mental state of
knowledge).¹² As such, they provide support to a knowledge-centered psych-
ology, for they support the view that knowledge plays a central role in
psychological explanations of intentional action.
8.3. Intentional Action does Require Knowledge
The claim that intentional action requires knowledge has been recently chal-
lenged. It is instructive to consider the challenge for it points towards a more
principled argument for thinking that knowledge is required for intentional
action.
Cath (2015: 11) argues that one can have intentional action without know-
ledge, upon considering cases like the following:
Bob the Pilot: Bob wants to learn how to fly in a flight simulator. He is

instructed by Henry. Unbeknownst to Bob, Henry is a malicious imposter

who has inserted a randomizing device in the simulator’s controls and
intends to give all kinds of incorrect advice. Fortunately, by sheer chance,
the randomizing device causes exactly the same results in the simulator as
would have occurred without it, and by incompetence Henry gives exactly
the same advice as a proper instructor would have done. Bob passes the
course with flying colors. He has still not flown a real plane. Bob has a
¹¹ Setiya (2012) puts forward an example of the subject believing their hand to be paralyzed who
tries nonetheless to clench their fingers. This case too can be handled by appealing to probabilistic
knowledge, for although one does not know that one will succeed by intending to succeed, one knows
that one might succeed by intending. Cf. Pavese (2020). For a different sort of diagnosis of Setiya’s case,
cf. Pavese (2018).
¹² Gibbons (2001) emphasizes that the role of knowledge in explaining intentional action provides a
novel argument for the claim that knowledge is a mental state—a claim famously defended by
Williamson (2000) and Nagel (2013). Although both Williamson (2000) and Nagel (2013) provide
arguments for the role of knowledge in explaining behavior, neither focuses specifically on the role of
knowledge in action theory. For a comparison, see Pavese (2019).
184  
justified true belief about how to fly, but that justified true belief does not
amount to knowledge.¹³
Cath uses this example to argue that one can have know-how even though
one’s relevant belief is Gettiered. Cath claims that intuitively Bob can inten-
tionally fly. And for some action ϕ, if one can intentionally ϕ, then one knows
how to ϕ (Williamson and Stanley 2001; Hawley 2003; Setiya 2012; Cath 2015;
Pavese 2018, 2020). Therefore, Bob must know how to fly, even though he
does not know the relevant instructions.
Cath thinks that this diagnosis is supported by comparing the case of Bob
with the case of Joe, who is a near perfect counterpart of Bob except that his
belief is not Gettiered: his simulator operated correctly and did so non-
accidentally; his instructor intentionally gave him the correct advice, and so
on. If Joe were to try to fly a plane in normal circumstances, he would typically
succeed in so doing and his successful actions would be unquestionably
intentional actions. And it is an implicit stipulation of the flight simulator
case that if Bob were to try to fly a plane in normal circumstances then he
would be just as likely to succeed as Joe. Cath contends that not only would
Bob succeed as often as Joe but, like Joe, his actions would appear to have all
the standard kind of properties that are thought to distinguish merely suc-
cessful actions from intentional actions. Hence, according to Cath, it is
plausible that Bob’s successful actions of flying, like Joe’s, would be perfectly
under his control or guidance as he performs them. But is Cath (2015) right in
assuming that, in Bob the Pilot, Bob’s performance is under his control?
Both intuitive and theoretical considerations suggest that the answer to this
question ought to be “No.” Compare Joe and Bob. Note that, strictly speaking,
Joe knows what he is doing while he flies the plane: he knows that he is
following instructions that are conducive to successful flying. In contrast, by
assumption, Bob does not know that. Hence, he does not know what he is
doing while he flies the plane. But consider how unintuitive it is to ascribe
intentional action to one who lacks knowledge of what they are doing:
Awful: Bob intentionally landed the plane, but he did not know that he was
landing the plane by following the given instructions.
¹³ This example is initially due to Stanley and Williamson (2001) who actually give it as an example
showing that know-how is incompatible with epistemic luck. Poston (2009) also discusses it and so
does Stanley (2001: Ch. 8).
, ,   185
A similar intuition has been voiced by several philosophers in the past.

Hampshire claims that “[I]f a man is doing something without knowing that
he is doing it, then it must be true that he is not doing it intentionally”
(Hampshire 1959: 95). Anscombe (1959: Section 8.8) holds that if someone
is ϕ-ing intentionally, she knows without observation that she is ϕ-ing. More
recently, Gibbons claims that “talk of intentional action presupposes a certain
degree of control on the part of the agent. Control, like perception, requires the
right kind of connection between the agent and the facts. An essential ingre-
dient in this kind of connection is knowledge” (2001: 591).
Hence, the intuition that an intentional action’s control depends on one’s
knowledge is very widespread. Of course, no important philosophical claim
should be motivated merely by intuitions, for they are too unstable and
possibly theoretically driven to be conclusive. There is, however, a positive
theoretical argument to the effect that, contra Cath, one’s actions cannot be
intentional unless they are guided by knowledge.¹⁴
Here is the gist of the argument. Cath assumes intentional action ought to
be “under one’s control” and assumes that Bob’s success is under his control.
But as in Mary the Bomber, the belief in question is Gettiered, and so happens
to be true by luck. According to the standard account, a belief is lucky just in
case it could easily fail to be true. If so, both Bob and Mary could easily have
the beliefs they have even if they were false. In this case, an attempt of theirs
based on those beliefs could easily fail. Arguably, then, their success is too
lucky to count under their control. More precisely, here is the argument step
by step:
(a) If one’s action is intentional, then it is under one’s control. (Premise)

(b) If one’s action is (too) lucky, then it is not under one’s control.
(Premise)
(c) If an action is based on a Gettiered belief about how to perform it, then
it is (too) lucky. (Premise)
(d) Hence, if an action is based on a Gettiered belief about how to perform
it, then it is not under one’s control. (From b, c)
(e) Hence, if an action is based on a Gettiered belief about how to perform
it, then that action cannot be intentional. (From a, d)
¹⁴ Greco (2016) considers a different sort of argument for the claim that knowledge is necessary for
explaining action—one that relies on the nature of explanation as counterfactually robust. See also
Pavese (2018).
186  
As rarely happens in philosophy, we have here a deductively valid argument

to the effect that intentional actions ought to be based on knowledge. Is the
argument sound?
Premise (a) relies on rather minimal assumptions about the nature of the
action, widely endorsed in action theory (e.g., Mele and Moser 1994; Gibbons
2001), according to which intentional action is under the agent’s control.
(Because this premise is actually granted by Cath (2015: 10–11), I will assume
it without much argument.)
Premise (b) is the rather uncontroversial claim that, if a certain success is
(too) lucky, then it is not under the control of the agent. This is a common
assumption in action theory, where it is customary to infer from the fact that
an act is too coincidental that it is, therefore, not intentional (e.g., Melse and
Moser 1994: 40), as well as in debates on moral responsibility (Nagel 1979: 59;
Williams 1981: 126, 1993), where moral luck is deemed to be incompatible
with control.¹⁵,¹⁶
The most controversial premise is, I take it, Premise (c)—i.e., the claim that,
if an action is based on a Gettiered belief, then the action is (too) lucky to
qualify as under the agent’s control. Here is an argument for thinking that it is
true. The following is a plausible sufficient condition on a lucky action:
Sufficient: if S succeeds at ϕ-ing at a world @ but S fails to ϕ in many of the

sufficiently close worlds where S tries to ϕ, then S’s ϕ-ing at @ is (too)
lucky.¹⁷
Note that Sufficient does not require that for one’s action at @ to be not-too-
lucky, one succeed at performing it when one tries it in every sufficiently close
world.¹⁸ This would be too demanding a requirement: as Austin (1961: 218)
put it, “a human ability or power or capacity is inherently liable not to produce
success, on occasion, and that for no reason.” If Austin is right, many actions
might be under the control of their agent and even manifest their skills and
¹⁵ It is worth noting that some recent work takes moral luck to be a species of a larger genus of luck,
of which there are other species as well, such as epistemic luck. Such an approach does not build in the
idea that luck is opposed to control. See Pritchard (2006) and Coffman (2015) for similar approaches.
Thanks to Mona Simion for the discussion.
¹⁶ It is, of course, a difficult and partly a context-sensitive matter, what counts as “under one’s
control.” But, for our purposes, we do not need to settle this question, the idea being that if a certain
success counts as too lucky, it is not sufficiently under the control of the agent to count as intentional.
¹⁷ Beddor and Pavese (2018: 6) for a defense of this sort of condition on lucky actions.
¹⁸ Although strictly speaking only attempts might fail to succeed, it will simplify my exposition to
talk as if one’s action might fail to succeed in some worlds where it is tried.
, ,   187
still fail in some close worlds, for no particular reason.¹⁹ Rather, according to
Sufficient, an action is too lucky if it fails in more than just some close worlds
when tried—if it fails in many close worlds (for that is what “many” means:
more than just some).
My argument for Premise (c) is that actions based on Gettiered beliefs are
doomed to satisfy Sufficient, and hence are doomed to be too lucky in the
relevant sense. Here is why. If an action is based on a true belief about how to
bring it about, then it might nonetheless fail when tried in some close worlds.
In fact, even if the action is based on knowledge, it might still fail in some close
worlds. That might be because in those close worlds, something interferes with
the agent’s basing their action on the relevant belief. For example, an archer’s
shot might fail in a counterfactual circumstance where the archer gets dis-
tracted by a passer-by and hence fails to appropriately base their shot on their
knowledge about how to shoot. Or it might fail also because, although their
shot is appropriately based on their knowledge, the world does not cooperate
in the way required for success. For example, the archer’s shot might fail in a
circumstance where a fluke happens, and an unexpected gust of wind inter-
feres. Or it might fail because their hand was for no reason slightly less firm
than they expect it to be a moment ago. In those counterfactual circumstances,
the belief on which the agent’s performance is based might still be true—it is
still true that one must do this and this to shoot the target under certain
conditions (e.g., in non-windy conditions and when one’s hands are firm). It is
just that there the world does not cooperate in the way required for success.
Now, when an agent acts on a Gettiered belief, there will be an additional

reason for why they might fail in some close worlds—i.e., because the belief
they act upon there is false and so does not accurately represent the world.
Gettiered beliefs are generally assumed to be lucky precisely in the sense that
they might fail to be true in some close worlds (Sosa 1999; Williamson 2000;
Pritchard 2005). If so, then an action based on a Gettiered belief will fail in
more close worlds than it would have if it had been based on knowledge. For it
will not just fail in those close worlds where it is not properly based on the
relevant belief or where, even though properly based, the world does not
cooperate. It will also fail in those close counterfactual circumstances where
the basing belief about how to perform the action is false. Hence, it will fail in
more than just some close worlds. Hence, an action based on a Gettiered belief
is doomed to satisfy Sufficient. And if it satisfies Sufficient, then Premise (c)
is true.
¹⁹ Thanks to Clayton Littlejohn for discussion on this point.
188  
This concludes my argument for Premise (c). With Premise (a) to Premise
(c) in play, the conclusion deductively follows actions based on Gettiered
beliefs cannot be intentional. Intentional action ought to be guided by
knowledge.
8.4. From a Knowledge-Centered Psychology to

Intellectualism About Know-How
Hence, a knowledge-centered psychology can be motivated on the basis of

intuitions about cases (Section 8.2). But there are, also, more theoretical
considerations on behalf of a knowledge-centered psychology, ones that rely
on rather minimal assumptions about the nature of Gettiered beliefs, control,
and intentional actions (Section 8.3). This section explores some of the
consequences of a knowledge-centered psychology and details an argument
from a knowledge-centered psychology to an intellectualist view about know-
how.
First, note that the kind of knowledge that, on a knowledge-centered
psychology, explains intentional action is exactly the same kind of knowledge
that, on a broadly intellectualist picture, is required by know-how. To see this
consider the kind of knowledge that would be needed to explain intentional
action. Start with Goldman’s (1970) action theory, according to which one
intentionally ϕs when one has a plan to ϕ, where a plan to ϕ is a belief that
specifies the means to ϕ (cf. also e.g., Audi 1986; Bratman 1987; Ginet 1990;
Harman 1976; Velleman 1989, 2007; Mele and Moser 1994):
(Intentionality/Belief): if s intentionally ϕs, then there are some means

m₁, . . . , mn to ϕ such that s truly believes that m₁, . . . , mn are means for
them to ϕ.
(Intentionality/Knowledge) can be formulated along the same lines:
(Intentionality/Knowledge): if s intentionally ϕs, then there are some means

m₁, . . . , mn to ϕ such that s knows that m₁, . . . , mn are means for them to ϕ.
It is independently plausible that the content of one’s knowledge in

(Intentionality/Knowledge) ought to be spelled out in probabilistic terms.
To see this, start with (Intentionality/Knowledge). It requires of intentionally
ϕ-ing that one have knowledge, of some means, that they are means for one
, ,   189
to ϕ. What does it mean for some means to be means for one to ϕ? Not that,
for some way ψ of ϕ-ing, one will ϕ by ψ-ing: that is far too strong, for one
might intentionally ϕ even though one has some doubts about whether one
will succeed (Goldman 1970; Harman 1976). Recall, for example, Davidson’s
(1971: 50; 1980: 91–4) carbon copy example, discussed earlier.
Should the relevant knowledge be that, for some means ψ of ϕ-ing, one
would in most cases succeed at ϕ-ing by ψ-ing? This is also too strong: one
might intentionally ϕ even though one might fail in most circumstances, as the
baseball player who fails at batting nineteen times out of twenty may none-
theless intentionally bat the one time they succeed. That suggests that the
relevant knowledge is that, for some means ψ of ϕ-ing, one could ϕ by ψ-ing.
But what does it mean that one could ϕ by ψ-ing, if not that one is sufficiently
likely to ϕ by ψ-ing, where what counts as sufficiently likely may vary from task
to task? This gives us:
(Intentionality/Probabilistic Knowledge): if s successfully and intentionally ϕs

at t, then at t s knows, for some means ψ of ϕ-ing, that oneself is sufficiently
likely to ϕ by ψ-ing.
Now, according to standard formulations of intellectualism (Stanley and

Williamson 2001; Stanley 2011; Pavese 2015, 2017), one knows how to ϕ
only if, for some means ψ to ϕ, one knows that ψ is a means for one to ϕ:
(Intellectualism about Know-How) s’s knowing how to ϕ is at least in part of

a matter of knowing, for some means ψ to ϕ, s knows that ψ is a means for
them to ϕ.²⁰
Consider Intellectualism about Know-How. In the original formulation, it is

the view according to which knowing how to ϕ is at least in part a matter of
knowing that certain means are means for one to ϕ. But what does that
mean? We do not want to require, for some means to be means for one to ϕ
that one’s ψ-ing invariably result in one’s successfully ϕ-ing; nor that it result
in one’s successfully ϕ-ing in most cases. That would be too demanding: after
all, Babe Ruth does know how to hit a home run and yet fails at successfully
hitting a home run in many circumstances. In order for ψ to be a way for one
²⁰ I am stating Intellectualism as the view that know-how requires knowledge of the means for
action, rather than as a fully reductive claim. Stating Intellectualism as a fully reductive claim would
require talking about practical modes of presentation, which I cannot discuss here. See Pavese (2015,
2017, 2019a).
190  
to ϕ, all that is required is that one be sufficiently likely to successfully ϕ by

ψ-ing, where what counts as “sufficiently likely” may vary with the task at
hand (and the circumstances under which the task is being performed). This
gets us to:
(Intellectualism about Know-How): s knows how to ϕ only if, for some means
ψ for s to ϕ, s knows that it is sufficiently likely for them to succeed at ϕ-ing
by ψ-ing.
Hence, if one unpacks the clause means for one to ϕ, a plausible upshot is that
both (Intellectualism about Know-How) and (Intentionality/Knowledge)
should be stated through a probabilistic language: they both require knowledge
that one is sufficiently likely to succeed at ϕ-ing through certain means.
Following Moss (2018), one might know that it is sufficiently likely for oneself
to succeed at ϕ-ing by ψ-ing in virtue of possessing a sufficiently high credence
that one will succeed at ϕ-ing by ψ-ing. Because possessing this credence does
not require that one grasp the concepts of likelihood or probability, Pavese
(2020) argues that this rendition of (Intellectualism about Know-How) over-
comes the challenge of over-intellectualization that Setiya (2012) and other
authors have raised against it.
Hence, the knowledge that (Intentionality/Knowledge) requires for inten-
tional action is the same that intellectualists require for know-how. I suggested
that we can use a knowledge-centered psychology to argue for intellectualism
about know-how. How would such an argument go? Start from (Know-How/
Intentionality), endorsed by intellectualists and anti-intellectualists alike (Ryle
1949; Stanley and Williamson 2001; Stanley 2011; Hawley 2003; Hornsby
2004, 2011; Setiya 2012; Pavese 2018):
(Know-How/Intentionality): if s intentionally ϕs, s knows how to ϕ.
Among the motivations behind (Know-How/Intentionality) is the idea that

operations which cannot be performed intentionally, such as digesting, are
ones that one cannot know how to perform (Williamson and Stanley 2001).
Moreover, manifestations of know-how seem to be characteristically inten-
tional: as Ryle (1949) put it, what distinguishes the clumsy person, who falls
and tumbles by accident, and the skillful clown is that the latter, but not the
former, falls and tumbles on purpose.
Further, suppose that (Intentionality/Knowledge) is true and so that the inten-
tionality of an action is to be explained at least in part in terms of propositional
, ,   191
knowledge. Then by (Know-how/Intentionality) and (Intentionality/Knowledge),

we get that, if one intentionally ϕs, one both knows how to ϕ and has propos-
itional knowledge of some means to ϕ:
(Know-how, Intentionality, Knowledge): if s intentionally ϕs, s both knows

how to ϕ and for some means m₁, . . . , mn, knows that means m₁, . . . , mn are
means for then to ϕ.
The intellectualist picture provides the best explanation for why (Know-How,
Intentionality, Knowledge) should hold. According to this explanation, (Know-
How, Intentionality, Knowledge) is true not just out of a coincidental aligning
of propositional knowledge and know-how in intentional action. Rather, its
truth is grounded in the very nature of know-how.
By mostly appealing to a linguistic argument when motivating their views
(Stanley and Williamson 2001; Stanley 2011), intellectualists have sold the
view short.²¹ The chief motivation for the view does not come from linguistics:
it comes from the sort of action theory that a knowledge-centered psychology
recommends.
8.5. The Defeasibility of Know-How
I have argued on behalf of a knowledge-centered psychology by looking at the

role of knowledge in explaining intentional action. Knowledge-centered

psychology naturally goes together with an intellectual picture of know-how
that vindicates the relation between know-how and intentional action.
Now, the intellectualist picture motivated by a knowledge-centered psych-
ology and outlined in the previous section makes a very clear prediction: that
know-how is defeated exactly when knowledge is defeated. Against this pre-
diction, Carter and Navarro (2018) argue that the defeasibility of know-how
does not go together with the defeasibility of knowledge. They use this claim to
argue against the intellectualist claim that know-how consists in a state of
propositional knowledge. We are now in a position to assess the problems with
Carter and Navarro’s (2018) argument, which turns on a failure to appreciate
the relation between knowledge and intentional action.
²¹ For worries concerning the linguistic argument on behalf of intellectualism, see Brown (2013).
192  
Carter and Navarro (2018: 666) propose the following example:
Ana and the Grenade Factory: Ana and Marıa work in a grenade factory
during the Spanish Civil War. They are thoroughly instructed when hired,
with examples and practical explanations. By controlled trial and error, they
learn their job, and both continue working at the factory for years, believing
they are making working grenades. However, one day each comes to realize
that the other is making grenades in an importantly different way, and they
identify the origin of the problem: as it turns out, the instructions were
ambiguous and allowed for two different interpretations. The instructors
were not aware of this, and there is nobody above them now who may say
who is right. Given that the grenades may only be used in battle, which is very
far away, neither Ana nor Marıa knows whose grenades actually work, and so
there is no way to find out who is making them the right way. As a matter of
fact, Ana got the instructions right (she produces grenades in way w, which is
the correct way); she is very successful in producing grenades that later work
perfectly. It is Marıa who got something wrong (she makes them in w’, the
possible interpretation of the instructions that the instructors did not fore-
see), and her grenades are always duds. Unaware of this, both have reasonable
doubts they did not have before, but they have to keep on working.
According to Carter and Navarro (2018), before receiving information about

how her knowledge has been acquired (call this piece of information
‘MISLEADING’), Ana might know, for some means to make grenades, that
it is a means to make grenades; but her knowledge is defeated as soon as the
misleading evidence is acquired. On the other hand, they think that Ana still
knows how to make grenades after receiving MISLEADING. If they were right,
this would be a case where know-how stands undefeated whereas the corres-
ponding knowledge is instead defeated. They conclude (2018: 669):
If know-how really were a case of know-that, we should expect it to be

defeasible by the same kinds of mechanisms by which propositional know-
ledge is defeated. But it is not. In other words: garden variety defeaters of
knowledge-that do defeat the knowledge agents have about the ways in
which they do what they do.²²
²² The argument assumes that knowledge can be easily defeated by high-order evidence. This
assumption is controversial and is not granted by prominent epistemologists (Aarnio 2010, 2014).
Let us play along, however, and see that the putative challenge rests on other false assumptions about
the nature of know-how.
, ,   193
Carter and Navarro’s argument hinges on two claims. The first claim is that
the relevant propositional knowledge is defeated in this case; the second is that
Ana still knows how to make grenades, upon receiving MISLEADING. Let me
grant the former claim and focus on the latter.
Why think Ana still knows how to make grenades, upon receiving
MISLEADING? The intuition that she does is not nearly as robust as they
seem to think. And the only argument Carter and Navarro (2018: 666) provide
in support of their intuition is the following:
The claim that Ana preserves her know-how along all the variations of the
case is supported by the fact that she is still able to make grenades profi-
ciently, and the doubts she acquires do not seem to imperil this ability in any
relevant sense.
Carter and Navarro (2018) are going far too quickly here. Granted, Ana still
preserves some ability that is relevant to grenades-making. What is much less
obvious, and as I argue ultimately incorrect, is to assume that Ana preserves
the sort of ability that goes together with know-how.
As the discussion in the previous section already suggests, know-how does
not just go together with any ability. It goes together with the ability to
intentionally perform a certain task. For example, knowing how to make
risotto does not merely go together with the ability to make risotto but with
the ability to intentionally make risotto. For if one had the ability to make
risotto but lacked the ability to intentionally make it, one would not count as
knowing how to make risotto. This point is well-known in the literature at
least since Hawley (2003) and is accepted by both intellectualists and anti-
intellectualists (Ryle 1949; Setiya 2012; Pavese 2017). For example, the clumsy
person has the ability to fall and tumble, as they reliably do so. But only the
clown has the ability to intentionally do that. As another illustrative example,
Susie may have the ability to irritate Ben, for she would succeed at irritating
him if she tried. But suppose she falsely believes that it is the smell of the
smoke, rather than the noise she makes whenever she smokes, that irritates
Ben. In this case, she does not intentionally irritate Ben: her success is too
coincidental to count as intentional. Because of this, it seems that she does not
know how to irritate Ben. On the bases of similar examples, intellectualists and
anti-intellectualists alike endorse the claim that know-how goes together with
the ability to intentionally perform the task.
Suppose it is true that know-how goes together with the ability to intention-
ally perform a task. If so, it is independently plausible that there is an important
194  
sense in which, upon receiving MISLEADING, Ana does not know any longer
how to make grenades. For upon receiving MISLEADING, when asked to make
grenades, she will be at a loss. Not only that: she will also refuse to teach others
how to make grenades. And she will even stop performing at the workplace,
until she is told that she has been making grenades correctly all along. Suppose
she were forced to reproduce whatever process she initiated before
MISLEADING. She would unknowingly succeed at making grenades. But the
success would be too out of her control to count as intentional. She is still able in
some sense to make grenades but in an important sense she now lacks the ability
to intentionally make grenades. If so, then she also lacks know-how.
If the reader is not yet ready to grant this conclusion, it is because, actually,
things are more complex, and some additional distinctions are called for.
Ascriptions of abilities of the form “s can intentionally perform a task” are
opaque, for as it is well known in action theory, “intentionally” is an inten-
tional operator (cf. Davidson 1971; Goldman 1970). For example, Lois might
intentionally kiss Superman but not intentionally kiss Clark Kent. Because
of the opacity of intentionality reports, it is paramount to distinguish between
(de re ability) and (de dicto ability):
(de re/ability): there is some task t that is in fact the task of making grenades
such that Ann has the ability to intentionally perform t.
(de dicto/ability) Ana has the ability to intentionally make grenades.
While (de re/ability) ascribes Ana a de re ability, (de dicto/ability) ascribes Ana
a de dicto ability.
Now, with this distinction in play, consider again Ana’s situation upon
receiving MISLEADING. (De dicto/ability) is now false: Ana does not have a
de dicto ability any longer. For one to have the relevant de dicto ability, one
needs to be able to make grenades on demand (to be in a situation such that, if
asked to make grenades, Ana would do so). Ana does not have that ability:
were she asked, after receiving MISLEADING, to make grenades, she would
now be at a loss. If know-how goes with the ability to intentionally perform a
task, then, to this distinction between a de dicto ability and a de re ability, there
corresponds the distinction between de re know-how and de dicto know-how:
(de re/KH): there is a task t that happens, unknown to Ana, to be the task of
making grenades, such that Ana still knows how to perform t.
(de dicto/KH): Ana knows how to perform grenades.
, ,   195
(de re/KH) and (de dicto/KH) ascribe different kinds of know-hows—de re

know-how and de dicto know-how—which go along with different disposi-
tions in behavior. One might have de re know-how even if one has no idea that
what one is doing when doing t is making grenades. Suppose, for example, one
is simply instructed to follow a certain procedure but has no idea of its
outcome. In this case, one might have de re know-how without de dicto
know-how. This is plausibly Ana’s quandary: Because Ana still knows how
to execute whatever task she was executing before MISLEADING, she plaus-
ibly still has de re know-how. After all, if she were told at the workplace to do
whatever she was doing before she received MISLEADING, and she obeyed
the order, she would intentionally perform a task, which, unknown to her, is
the task of making grenades. Hence, Ana preserves her de re ability upon
receiving MISLEADING. So, Ana plausibly also retains her de re know-how
upon receiving MISLEADING.
However, Anna does lose de dicto know-how. For her to possess de dicto
know-how, it is not sufficient to possess de re ability; she would need in
addition to have the corresponding de dicto ability, which as we have seen
she lacks. While Ana loses de dicto know-how and de dicto ability upon
receiving MISLEADING, Ana preserves de re know-how and de re ability,
for she still knows how to do whatever it was that she was doing before (which,
as far as she knows, is not accurately making grenades!), and she still preserves
the de re ability to make grenades upon receiving the misleading information.
Crucially, intellectualists can accept all of this. According to intellectualism,
de re know-how only requires de re knowledge while de dicto know-how

requires de dicto knowledge:
(de re/K): there is a task t that is in fact, but unknown to Ana, the task of
making grenades such that Ana knows for some way w that w is a way to
execute t.
(de dicto/K): Ana knows for some way w that w is a way to make grenades.
Upon receiving MISLEADING, (de dicto/K) becomes false: Ana loses the
relevant de dicto knowledge. However, Ana arguably still also preserves the
relevant de re knowledge (ascribed by (de re/K)). MISLEADING only defeats
(if anything) her de dicto knowledge—i.e., the knowledge that the procedure
she was implementing was for making grenades. Her de re knowledge, instead,
is not at all defeated by MISLEADING: Ana continues to have it, as she might
continue to know what procedure she was following before receiving
196  
MISLEADING, when she was intending to make grenades. And so, by intel-
lectualism’s lights, she might continue to know how to make whatever she was
making when she thought (correctly, it turns out!) that she was making
grenades: because Ana retains her de re knowledge, by intellectualism’s lights,
Ana can retain her de re know-how as well as her de re ability.
Let me end by considering two possible responses.
I have argued that Ana might lose her de dicto know-how upon receiving
MISLEADING, while possibly retaining her de re know-how. Could not
Carter and Navarro (2018) reply that Ana does retain he de dicto know-how
and her de dicto ability all along but upon receiving MISLEADING, she simply
cannot act on those, because of her new doubts? Compare: many Olympic
gymnasts know how to do the fancy tricks they do even though many—in the
heat of the competition—have doubts about whether they can do them
successfully.
In response, the analogy with the Olympic gymnasts is misleading. In the
case of Olympic gymnasts, it is plausible that despite their doubts, they still
know how to perform their fancy tricks (de dicto). After all, they can still
intentionally do their fancy tricks (de dicto) outside of the heat of the compe-
tition. Their ability to intentionally act on that knowledge is not lost but only
‘masked.’ This can be explained on the current picture: These athletes retain
their (de dicto) knowledge all along and simply cannot access it in some
circumstances. By contrast, Ana has lost her de dicto knowledge and unless
she regains it, there is no circumstance where she can still intentionally make
grenades (de dicto). In Ana’s case, then it is utterly implausible that her ability
to intentionally act is simply masked.
A second possible response goes as follows. Maybe, Ana does preserve
her de dicto know-how and her de dicto ability upon receiving
MISLEADING. What she lacks is knowledge that she does know how and
that is what explains the lack of de dicto ability. If one embraces this position,
one commits oneself to replacing (Know-how/Intentionality)—a claim that, as
we have seen, both intellectualists and anti-intellectualists agree upon—with a
considerably stronger claim that intentionally ϕ-ing requires knowledge that
one knows how to ϕ. This stronger claim is rather implausible. For one thing,
some non-human animals can certainly intentionally act, though lacking the
concept of know-how. For another, suppose one learns how to make grenades
but has not had the occasion to form a belief one way or another about
whether one has indeed learned. If one is asked to produce one and one
tries, surely one can still intentionally do it (even de dicto!), even though one
does not know that they knew how to make them.
, ,   197
8.6. Conclusions
There is a lot going for a knowledge-centered psychology: it explains our

intuitions in a variety of cases where intentionality of an action seems absent
because of the agent’s lack of knowledge (Section 8.2). The role of knowledge
in explaining intentional action is also demonstrated by more theoretical
considerations showing that an action cannot be under one’s control unless
it is guided by knowledge (Section 8.3). A knowledge centered psychology, in
turn, motivates intellectualism about know-how, for it explains why know-
how and the ability to intentionally act go hand in hand (Section 8.4).
Having motivated a knowledge-centered psychology, I have appealed to it
in a discussion of Carter and Navarro’s (2018) argument to the effect that
know-how differs from knowledge in its pattern of defeasibility Section 8.5).
I have argued that Carter and Navarro’s (2018) challenge fails, for they fail to
show that know-how remains undefeated when knowledge is defeated. Their
alleged challenge turns on the failure to appreciate the relation between
knowledge, know-how, and intentional action. Because of that, they fail to
distinguish between different sorts of abilities that go together with know-how.
Once one appreciates that know-how goes with the ability to act intentionally,
because ascriptions of this sort are opaque, it becomes paramount to distin-
guish between de re abilities and de dicto abilities. With this distinction comes
the corresponding distinction between different sorts of know-hows and
between the different sorts of knowledge that Ana preserves or loses upon
receiving MISLEADING. As we have seen, against Carter and Navarro (2018),

the sort of de re abilities Ana does preserve can be fully accounted for on a
picture on which know-how is knowledge. And those de dicto abilities that she
does lose are also correctly predicted to get lost on the same intellectualist
picture.
Far from coming apart in their pattern of defeasibility, know-how and
knowledge go hand in hand, just as one would expect on the sort of intellec-
tualist picture that is motivated by a knowledge-centered psychology.
References
Aarnio, M. L. (2010). Unreasonable knowledge. Philosophical Perspectives, 24:

1–21.
Audi, R. (1986). Intending, intentional action, and desire. In: Joel Marks (ed.) The
Ways of Desire. Chicago: Precedent: 17–38.
Austin, J. L. (1961). Philosophical Papers. Oxford: Oxford University Press.
198  
Beddor, B. and Pavese, C. (2018). Modal virtue epistemology. Philosophy and

Phenomenological Research, Early View.
Beddor, B. and Pavese, C. (2019). Manuscript. Skill as knowledge.
Bergmann, M. (2000). Deontology and defeat. Philosophy and Phenomenological
Research, 60(1): 87–102.
Bratman, M. (1987). Intention, Plans, and Practical Reason. Center for the Study
of Language and Information.
Brentano, F. (1874/1995). Psychology from an Empirical Standpoint. New York,
NY: Routledge and Kegan Paul.
Brown, J. A. (2013). Knowing-how: Linguistics and cognitive science. Analysis, 73
(2): 220–7.
Carter, A. J. and Navarro, J. (2018). The defeasibility of knowledge-how.
Carter, A. and Pritchard D. (2015). Knowledge-how and epistemic luck. Noûs 49:
440–53.
Cath, Y. (2015). Revisionary intellectualism and Gettier. Philosophical Studies, 172
(1): 7–27.
Chisholm, R. (1966). Theory of Knowledge. Englewood Cliffs, N.J.: Prentice-Hall.
Coffman, E. J. (2015). Luck: Its Nature and Significance for Human Knowledge and
Agency. Princeton: Princeton University Press.
Davidson, D. (1971). Agency. Reprinted in Davidson (1980), pp. 43–61.
Davidson, D. (1980). Essays on Actions and Events. Oxford: Oxford University
Press.
Dretske, F. (1988). Explaining Behavior. MIT Press.

Gibbons, J. (2001). Knowledge in action. Philosophy and Phenomenological
Research, 62(3): 579–600.
Ginet, C. (1990). On Action. Cambridge: Cambridge University Press.
Goldman, A. (1970). A Theory of Human Action. Englewood Cliff, NJ. Prentice
Hall Inc.
Goldman, A. (1976). Discrimination and perceptual knowledge. The Journal of
Philosophy, 73(20): 771–91.
Goldman, A. (1986). Epistemology and Cognition. Cambridge, MA: Harvard
University Press.
Hampshire, S. (1959). Thought and Action. London: Chatto and Windus.
Harman, B. (1976). Practical reasoning. The Review of Metaphysics, 79: 431–63.
Hawley, K. (2003). Success and knowledge-how. American Philosophical
Quarterly, 40(1): 19–31.
, ,   199
Hornsby, J. (2004). Agency and actions. Royal Institute of Philosophy Supplements,

55: 1–23.
Hornsby, J. (2011). Ryle’s knowing-how, and knowing how to act. Knowing how:
Essays on Knowledge, Mind, and Action, 80–100.
Humberstone, I. L. (1992). Direction of fit. Mind, 101(401): 59–83.
Levy, Y. (2013). Intentional action first. Australasian Journal of Philosophy, 91(4):
705–18.
Lewis, D. (1974). Radical interpretation. Synthese, 27: 331–44.
Lewis, D. (1979). Attitudes de dicto and de de. The Philosophical Review, 88(4):
513–43.
Littlejohn, C. (2014). Fake barns and false dilemmas. Episteme,11(4): 369–89.
Mele, A. and Moser, P. (1994). Intentional action. Noûs, 28: 39–48.
Moss, S. (2018). Probabilistic Knowledge. Oxford/NY: Oxford University Press.
Nagel, J. (2013). Knowledge as a mental state. Oxford Studies in Epistemology, 4:
275–310.
Nagel, T. (1979). Mortal Questions. New York: Cambridge University Press.
Pavese, C. (2015). Practical senses. Philosophers’ Imprint, 15(29): 1–25.
Pavese, C. (2017). Know-how and gradability. Philosophical Review, 126(3):
345–83.
Pavese, C. (2018). Know-how, action, and luck. Synthese, 1–23.
Pavese, C. (2019a). The psychological reality of practical representation.
Philosophical Psychology, 32(5): 784–821.
Pavese, C. (2019b). Manuscript. Knowledge and mentality.

Pavese, C. (2020). Probabilistic knowledge in action. Forthcoming in Analysis.
Pritchard, D. (2005). Epistemic Luck. New York: Clarendon Press.
Pritchard, D. (2006). Moral and epistemic luck. Metaphilosophy, 37: 1–25.
Pritchard, D. (2012). Anti-luck virtue epistemology. The Journal of Philosophy,
109(3): 247–79.
Pollock, J. L. and Cruz, J. (1999). Contemporary Theories of Knowledge, 2nd
edition. Lanham, MD: Rowman and Littlefield.
Ryle, G. (1949). The Concept of Mind. Chicago, IL: University of Chicago Press.
Setiya, K. (2012). Know-how. Proceedings of the Aristotelian Society, 112(3):
285–307.
Sosa, E. (1999). How to defeat opposition to Moore. Philosophical Perspectives, 13:
141–53.
Sosa, E. (2007). A Virtue Epistemology: Apt Belief and Reflective Knowledge (Vol. 1).
200  
Stalnaker, R. (1984). Inquiry. The MIT Press.

Stanley, J. (2011). Know How. Oxford: Oxford University Press.
Stanley, J. and Williamson, T. (2001). Knowing how. The Journal of Philosophy, 98
(8): 411–44.
Stich, S. (1978). Autonomous psychology and the belief-desire thesis. The Monist, 61.
Velleman, D. (1989/2007). Practical Reflection. CSLI Publications.
Weisberg, J. (2013). Knowledge in action. Philosophers’ Imprint. Ann Arbor, MI:
Michigan Publishing, University of Michigan Library.
Williams, B. (1981). Moral Luck. Cambridge: Cambridge University Press.
Williams, B. (1993). Postscript. In: Moral Luck. D. Statman (ed.) Albany NY: State
University of New York Press.
Williamson, T. (2000). Knowledge and its Limits. Oxford: Oxford University Press.
Williamson, T. (2017). Acting on knowledge. Knowledge-First: Approaches in
Epistemology and Mind, 163–81.
9
Undercutting Defeat
When it Happens and Some Implications
for Epistemology
Matthew McGrath
It’s commonplace to lose a justification for believing something. It can happen

through simple forgetting. This paper focuses on another way: through defeat.
When you lose a justification through defeat, the ground for your earlier
justification remains but no longer justifies you in believing the target prop-
osition, due to the operation of a “defeater” of that justification. Defeat strictly
speaking encompasses not only loss of a justification but also reduction the
degree of support provided by a justification.
John Pollock has distinguished between two types of defeaters: rebutting
and undercutting.¹ Let’s start with an intuitive gloss. Rebutters defeat by
striking against the target proposition itself, undercutters by striking against
the connection between the ground of a justification and the target propos-
ition. For an example of a rebutter, suppose I obtain some justification from a

Buzzfeed news report for believing that a certain public official suborned
perjury, but later I lose my justification when I learn that The New York
Times and The Washington Post independently reported later that this official
did no such thing. In this case, the information I received the newspapers
strikes against the claim that the official suborned perjury. It defeats the
justification I obtained from Buzzfeed. For an example of undercutting defeat,
we can vary the Buzzfeed example: suppose I obtain the same justification from
the Buzzfeed report but subsequently learn that their reporters relied exclu-
sively on the testimony of a single anonymous source known to have personal
interests at stake. The new information doesn’t strike against the claim that the
official suborned perjury. It defeats, rather, by striking against the connection
between my reason—that Buzzfeed reported that the official suborned
¹ See for instance, Pollock (1970, 1975, 1987); Pollock and Cruz (1999).
202    
perjury—and the proposition I believe—that the official did this.² The category
of undercutters is often thought to represent an important advance in epis-
temology, for it promises to help us explain the defeating power of consider-
ations that don’t bear on the truth of the target proposition, such as evidence
that a testifier is unreliable or evidence that one is hallucinating.
Although there is disagreement about the details, the distinction between
rebutting and undercutting defeat—defeat by striking against the target prop-
osition vs. defeat by striking against the connection between a ground and the
target proposition—is now part of the received wisdom in analytic epistemol-
ogy. Not only that, but many epistemologists think Pollock’s account of how
these defeaters work is roughly on the right track. Recently, however, cracks
have appeared in the consensus. While not questioning the existence of
undercutting defeat, Scott Sturgeon argues that Pollock misconceives it:
undercutting operates differently from rebutting in that it occurs only in
conjunction with certain higher-order contributions, i.e., with beliefs about
(or justifications to believe propositions about) the basis on which one does or
would believe.
I will argue, contrary to Sturgeon, that in the case of inferential justification,
undercutting defeat takes basically the shape Pollock suggests it does, not
needing contributions from higher-order beliefs or justifications. However,
I agree with Sturgeon that for noninferential justification, the Pollockian
account is in trouble. I try to explain why there should be this difference.
My interest, however, is not merely in determining what makes for a good
account of defeat. This difference in types of defeat has important implications

for other parts of epistemology. In a final section, I use the defeat-related
difference between inferential and noninferential justification to argue that
there is less immediate perceptual or testimonial justification than is com-
monly thought.
9.1. Pollock on Defeaters
Let’s start by reviewing the basic elements of Pollock’s mature account of

defeat as developed in his 1999 book with Joseph Cruz. First, there is the
notion of a reason to believe something:
² In principle, the same defeater could be at once a rebutter and an undercutter, if it strikes against
both the target proposition and the connection between the reason and that proposition.
  203
A state M of a person S is a reason for S to believe Q if and only if it is

logically possible for S to become justified in believing Q by believing it on
the basis of being in the state M.
As I understand the terminology here, ‘become justified in believing Q’ ascribes

doxastic justification, i.e., coming to have a justified belief in Q. In terms of this
notion of a reason, Pollock and Cruz define defeaters in general:
If M is a reason for S to believe Q, a state M* is a defeater for this reason if

and only if the combined state consisting of being in both the state M and the
state M* at the same time is not a reason for S to believe Q.
Next, rebutters:
(Rebutters): if M is a defeasible reason for S to believe Q, M* is a rebutting

defeater for this reason if and only if M* is a defeater (for M as a reason for
S to believe Q) and M* is a reason for S to believe ~Q.
For undercutters, they give two definitions, one for doxastic reasons and
another for nondoxastic reasons:
(The doxastic case): if believing P is a defeasible reason for S to believe Q, M*

is an undercutting defeater for this reason if and only if M* is a defeater (for
believing P as a reason for S to believe Q) and M* is a reason for S to doubt or

deny that P would not be true unless Q were true.
(The nondoxastic case): if M is a nondoxastic state that is a defeasible reason
for S to believe Q, M* is an undercutting defeater for this reason if and only if
M* is a defeater (for M as a reason for S to believe Q) and M* is a reason for
S to doubt or deny that he or she would not be in state M unless Q were true.
We can simplify these definitions, even by Pollock’s and Cruz’s own lights.
They go on to tell us that undercutting defeaters are reasons to believe the
appropriate subjunctive conditional is false (Pollock and Cruz 1999: 196).
Their suggestion, I take it, is that it’s because undercutters are reasons to
doubt or disbelieve such conditionals that they manage to be defeaters. This in
turn suggests that the clause that “M* is a defeater for . . . as a reason for S to
believe Q” is redundant in the previous definitions. The same goes for rebut-
ters. Thus, I attribute to them the following simplified definitions of under-
cutting defeat:
204    
Rebutters:
If M is a defeasible reason for S to believe Q, M* is a rebutter for this reason if
and only if M* is a reason for S to believe ~Q.
Undercutters:
(The doxastic case): if believing P is a defeasible reason for S to believe Q, M*
is an undercutter for this reason if and only if M* is a reason for S to doubt or
deny that P would not be true unless Q were true.
(The nondoxastic case): if M is a nondoxastic state that is a defeasible reason
for S to believe Q, M* is an undercutter for this reason if and only if M* is a
reason for S to doubt or deny that he or she would not be in state M unless
Q were true.
So revised, the definitions function as intended only if something’s meeting the

conditions for being a “rebutter” or an “undercutter” guarantees its meeting
the conditions for being a “defeater.” This will be important in the following.
Pollock and Cruz (1999: 197) suggest a simpler replacement for the complex
proposition that it’s not the case that P would not be true unless Q were true,
namely that P does not guarantee Q (in the circumstances). I follow them.
Thus, undercutters in the doxastic case are reasons to think that P does not
guarantee Q, and in the nondoxastic case they are reasons to think that one’s
being in state M does not guarantee Q. This is their attempt to capture the key
intuitive idea of undercutting defeat as defeat that strikes against the connec-
tion between a reason and the target proposition. This striking-against-the-
connection is explained in terms of having reasons to doubt or deny a certain

sort of proposition about that connection.
One might raise a number of questions about Pollock’s account of defeat, or
the parts of it I have described.³ One could ask about the ontology of reasons.
Why should we take reasons to be mental states (e.g., beliefs, experiences) rather
than facts or propositions? One could ask about the normative conditions a
defeater must meet to defeat: must doxastic defeaters be justified? If I form a silly
unjustified belief that I am hallucinating, will that defeat my perceptual justifi-
cation? One could also ask about the sort of conditional in terms of which
Pollock defines undercutters. Is “P doesn’t guarantee Q” the right one? Is that
really equivalent to the negation of the unless-statement “it’s not the case that
P wouldn’t be true unless Q”? Or should we build the account around the
³ I omit exposition of Pollock’s discussion of defeated defeaters, as well as his full account of
justification in terms of status-assignments assigning statuses of “defeated” and “undefeated” to
elements of arguments (see Pollock and Cruz 1999: 197–200).
  205
indicative conditional, or some other conditional? These worries concern, rela-

tively speaking, matters of detail that likely won’t sink the project of defending a
Pollockian account of defeat. One might also worry that Pollock’s account is
incomplete. Are there other kinds of defeaters? And of course one can ask for
clarifications about key notions such as the notion of its being possible to become
justified in believing something based on a combination of states.⁴ My concern
in this paper is whether the Pollockian picture is approximately right as far as it
goes. We might need to change certain details. We might need to supplement it
in a number of ways. My particular question, more precisely, is whether there is a
serious error in the picture when it comes to undercutting defeat.
9.2. Sturgeon against Pollock
Sturgeon argues against Pollock’s account of undercutting defeat, claiming it is

both too strong and too weak. Here I focus on his argument that it is too
weak.⁵ The argument in question is based on a case, the Milk Taster. I have
revised the example to avoid a certain objection.⁶
The Milk Taster:

The milk taster thinks her conclusions about whether a batch of milk is
spoiled are based solely on taste and not at all on smell. However, this is not
true. It turns out that her conclusions are based on smell only and not taste.
Relying on her olfactory experience, she concludes that a particular batch of

milk is ok. She is then given misleading information to the effect that she is
suffering a random olfactory hallucination. She comes to think that her
olfactory experience does not guarantee that the milk is OK.
⁴ As Sturgeon (2014: 108–14) patiently shows.

⁵ Sturgeon’s argument (using the Presupposer case) that it is too strong is convincingly criticized by
Casullo (2018: 2900–1).
⁶ In the original, Sturgeon (2014, 114–15) uses the more realistic example of relying on a complex
gustatory-cum-olfactory experience to believe the milk is OK, while getting evidence of olfactory
hallucination. The subject in his case thinks she is relying on taste alone. Sturgeon says that after a
bit of thinking the subject will deny the relevant unless-statement. She will deny that she wouldn’t have
the complex experience unless the milk was OK. But this seems wrong. The subject has no reason to
think that the olfactory hallucination affects either the milk or whether the taste experience would
covary with the state of the milk. So, she is in a position to reason like this: I am relying on taste
experience T; although I’m having an olfactory hallucination in having olfactory experience O, this is no
reason to doubt that I wouldn’t have T unless the milk was OK; given that O has no bearing on either T or
the state of the milk, I wouldn’t have T+O unless the milk was OK. Of course, strengthening of the
antecedent is not valid for subjunctive conditionals, but in cases where the proposition added to the
antecedent is irrelevant, it may be fine to reason from the original to the strengthened conditional. My
revized case avoids these issues and gives Sturgeon the result he wants.
206    
Sturgeon thinks that the milk taster remains justified in her conclusion about
the milk, in virtue of the same non-doxastic reason that justified her
previously. No defeat occurs. Nevertheless, her belief that she is undergoing
olfactory hallucination meets the (simplified) Pollockian definition of an
undercutter. Putting things in terms Pollock’s framework: although the
olfactory experience o is a reason to believe the milk is OK, and the belief
that one is suffering an olfactory hallucination (H) is a reason to believe that
having o doesn’t guarantee the milk is OK, the combination <o, belief in H>
is still a reason to believe the milk is OK—i.e., one can become justified in
believing that the milk is OK by believing the milk is OK on the basis of
being in this combination of states, and the Milk Taster case is a case in
point.⁷ Sturgeon concludes that Pollock’s account of undercutters is too
weak—it implies that there is an undercutter for a reason in this case
when there is no defeat at all. However, Sturgeon thinks that there is defeat
if we modify the case so that the milk taster has a correct belief about the
source of her belief about the milk. This suggests to him that we need to
revise Pollock’s account of undercutting defeaters by giving an essential role
to beliefs about one’s mental states, in particular to beliefs about the sources
of one’s belief.
More generally, let U be the claim that source S is untrustworthy about
whether P and BOS-P be the claim that your belief in P is (or would be or likely
would be) based on source S. Suppose you believe U and believe P. Sturgeon
proposes that in such a situation:
Your belief in U undercuts your belief that P iff you believe BOS-P.⁸
⁷ As noted earlier, more information is needed about what it is to become justified in believing
something on the basis of being in a certain combination of states. It would be too much to demand
that the combination be used as one’s grounds for the belief. When defeat occurs, it’s not because it
would be wrong to use the combination of states as one’s ground in any intuitive sense. I wouldn’t think
of using the combination of John told me that p and John was insincere in his remark to me on whether p
as my ground to believe p; rather, I might use John told me that p as my ground to believe p despite
being also in the state of believing that John was insincere in his remark to me on whether p. I will read
‘on the basis of being in state M’ in such a way as to allow for such possibilities. To believe p on the basis
of being in state M, then, as I will understand it within Pollock’s framework, is to be in state M and to
believe p using some component state of M (proper or improper) as one’s ground.
⁸ As Casullo (2018) points out, it’s implausible that the mere having of higher-order beliefs does
much epistemological work. What counts is justification to have such beliefs. We could reformulate
Sturgeon’s proposal so that it concerns justification to believe rather than belief. But as I mentioned, my
main concerns don’t hinge on such tweaks, so I will not bother with reformulation. Mellis (2014: 438)
usefully refines Sturgeon’s proposal, bringing in the concept of a justificatory process in addition to the
concept of a source of justification.
  207
If we accept Sturgeon’s understanding of undercutting, we must revise

Pollock’s general definition of a defeater to make it disjunctive. The needed
revision would presumably be something like this:
If M is a reason for S to believe Q, a state M* is a defeater for this reason if

and only if either (i) the combined state of being in M and being in M* is not
a reason for S to believe Q, or (ii) the combined state of being in M, being in
M* and believing that one’s basis for believing Q is or would be M is not a
reason for S to believe Q.
Finally, to handle undercutting defeat of reasons for beliefs one already has, we
would need to adjust the definition of ‘reason’ slightly as well:
A state M of a person S is a reason for S to believe Q if and only if it is

logically possible for S to become or remain justified in believing Q by
believing it on the basis of being in the state M.
Presumably the issue for the milk taster isn’t becoming justified but remaining
justified, and Sturgeon’s thought is that if you have a belief about the identity
of the source of your belief that p, together with a doubt or disbelief that
having that source guarantees P, you can’t remain justified in believing P.
Undercutting for Sturgeon is thus a matter of believing both:
• My belief that P is (or will be) based on source S.

• Source S is (or has a good chance of being) unreliable in the circum-
stances on whether P.
Whether one’s belief actually is or will be based on source S is neither here

nor there, as Sturgeon himself notes (2014: 117). All that matters is what the
believer thinks about her beliefs’ sources. The resulting view implies that if you
think you are using taste and you think that taste is unreliable in the circum-
stances, then even if you aren’t using taste (and are using a source you believe
to be reliable), still your belief is undercut.⁹
⁹ Thus, if Sturgeon is right about undercutting, Pollock’s account is not only too weak; it is also too
strong. An alternative view is that undercutting happens iff Pollock’s conditions are satisfied and one
believes one source is M. On this view, Pollock’s account would not be too strong. I don’t have the space
to examine this view in any detail. But note that even if this view is correct, undercutting doesn’t
happen solely in virtue of having reason to believe something that strikes against the connection
between one’s reason and the target proposition—higher-order contributions are also part of the
explanatory story.
208    
There is something odd about Sturgeon’s account of undercutting. The

whole idea of such defeat was that it was to be defeat by virtue of striking
against the connection between an actual ground of justification and the target
proposition. What Sturgeon gives us under the banner of “undercutting
defeat” has nothing to do with the actual ground, as Juan Comesaña
(Forthcoming) notes. Sturgeon-undercutters strike at best against the connec-
tion between one’s perceived basis or ground and the target proposition,
whether or not this perceived basis is one’s real basis.
What Sturgeon is pointing to, I think, is actually a third kind of defeat, what
I’ll call “higher-order” defeat. M* is a higher-order defeater of M as a reason to
believe P iff M* is a reason to think that one’s belief that P is (or would be)
unreliably based. Something can be a higher-order defeater of a reason
M without being a rebutter, i.e., without being a reason to believe not-P, and
without being an “undercutter” in the intuitive sense of something that defeats
by striking against the connection between one’s actual reason and P.¹⁰,¹¹
Casullo (2018: 2904) comes to Pollock’s defense. He questions whether
Sturgeon has put his finger on an asymmetry between undercutters and
rebutters:
Consider a cognizer A. Suppose that source S1 justifies A’s belief that P to

degree d1, that source S2 justifies A’s belief that not-P to degree d2, and that
d1 = d2. (S1 and S2 need not be different sources). Is this sufficient for A’s
belief that not-P to defeat A’s justification for the belief that P? There are
three possible responses:
1. Weak Internalist: No, A must believe that d1 = d2.

2. Strong Internalist: No, A must believe that d1 = d2 and that belief must be
justified.
3. Externalist: Yes.
¹⁰ We saw an example of this previously. You falsely think you are relying on taste, which you think
(and even know) to be unreliable, but in fact you’re relying on olfaction: you rely on an olfactory reason
O. Then your belief that you are using the unreliable source of taste is a higher-order defeater of O as a
reason to believe P, but it is not a reason to believe not-P (and so not a rebutter) and it does not it strike
against the connection between O and P (and so is not an undercutter).
¹¹ Note that higher-order defeat does not boil down to the same thing as “defeat involving defeaters
with higher-order content.” Perhaps some defeaters with higher-order content, such as I have strong
evidence against p are rebutters, and perhaps others can be undercutters. “Higher-order defeat,” as I am
stipulatively using this terminology, refers to defeat by virtue of being a reason to believe that one’s
belief is unreliably based.
  209
Since Sturgeon sees no role for higher-order beliefs or justifications in the case
of rebutting defeat, Casullo thinks Sturgeon must give the externalist answer.
Casullo’s claim is then that there is no reason to think that if the externalist
position is correct here it wouldn’t also be correct for cases of undercutting
defeat. He concludes that the two sorts of defeat work in the same way—they
both work with—or without—the help of higher-order beliefs/justifications.
Casullo, as I read him, doesn’t strictly take a stand on whether rebutting
defeat needs higher-order help. But it is implausible to think that it does. If
I have evidence for P, and I then gain equally good evidence for not-P, I don’t
need to have higher-order beliefs or justifications about these matters in order
for the different justificatory forces to weigh up in one direction or another.
This is intuitive as it stands. But we can add as well that if higher-order
contributions were necessary, worries would arise about how creatures that
don’t grasp these higher-order propositions could have rebutting defeaters
(e.g., young children or higher animals).¹²
There is also reason to worry that if rebutting defeat requires higher-order
contribution, the same would hold for a single source’s justificatory power. We
could ask: in order for a source to justify a subject’s belief to a certain degree,
must the person believe that the belief is justified (by that source) to that
degree? It is difficult to see why the facts about how justificatory forces weigh
up should require beliefs about those forces while the existence of a single
justificatory force doesn’t. But it is implausible to think that in order for a
source to provide justification for a belief the person would have to believe (or
be justified in believing) some proposition about its doing so.

I conclude that rebutting defeat does not require higher-order contribution.
This doesn’t of course show Sturgeon is correct to distinguish the way rebutting
and undercutting defeat do their work. For, as we saw with help from
Comesaña, what Sturgeon describes is not really undercutting defeat at all—
at least not in the intuitive striking-against-the-connection sense; it is what
I have called “higher-order defeat,” which naturally requires higher-order
contribution.
¹² Using a distinction from Audi (1993), we can distinguish two ways that a justificatory fact—such
as my having justification from a piece of evidence or my justification being defeated—could depend on
higher-order beliefs (or justifications). One way is negative dependence: the justifications obtain only if
you do not have certain negative higher-order beliefs (justifications), such as a belief that your basis is a
poor one. But this is not the kind of dependence at issue. At issue is positive dependence: in order for
justificatory facts to obtain one must have a higher-order beliefs (justifications) about those facts.
210    
9.3. In Defense of Pollockian Undercutting Defeat

(in the Inferential Case)
We have seen that what Sturgeon says about the Milk Taster case is plausible:
Pollock’s conditions for undercutting are met but there is no defeat. I agree
with Sturgeon that the failure is not a minor one: there are no defeaters of
nondoxastic reasons that do their work by attacking the connection between
one’s actual reason and the proposition believed. The closest things are defea-
ters that work by giving one a reason to think one’s belief is unreliably based.
But Sturgeon draws a much broader conclusion: there are no defeaters—even
for doxastic reasons—that work in this attacking-the-connection way. I now
argue that this broader conclusion is false: we find just such defeat in the case of
inferential justification, i.e., of defeat of doxastic reasons.
Consider an example of undercutting defeat described by Sturgeon (similar
to one discussed by Pollock and Cruz):
The Polling Case

A pollster surveys 1000 voters in Texas at random, asking whether they will
vote Republican or Democrat in the next election. The pollster believes
G(oing) = Roughly 87 percent of Texans are going to vote Republican.
On the basis of believing
T(ell) = 87 percent of respondents tell the pollster they will vote Republican.
Suppose the pollster then comes to believe
U = Respondents decided their answer by coin flip.
Belief in U fits the (simplified) Pollockian definition of an undercutter. Is there

defeat? On Pollock’s own definition of defeat, it appears the answer is yes: it
seems impossible to become (or remain) justified by believing in G based on
the combination of <belief in T, belief in U>. Moreover, this fits with the
intuitive thought that there is defeat in this case. Notice that higher-order
beliefs don’t come into it. The pollster might even lack such a belief.¹³
¹³ What if higher-order beliefs are present? Suppose, despite believing respondents decided their
answers by coin flip (U), and despite believing that roughly 87 percent of Texans are going to vote
Republican (G) on the basis of the belief this is what 87 percent of the Texan respondents told you (T),
you think you believe G on the basis of some other belief, H, which you think makes G likely true. In
such a case could you become or remain justified in believing G based on <belief in T, belief in U>?
I think not. These false beliefs about your basis even if justified do not seem to restore your actual basis
as a reason to believe G by somehow defeating your defeater (U) for that reason. Intuitively, you’re on
the hook, epistemically, for the basis you have; and you can’t get off the hook by having beliefs (even
justified beliefs) about having some other basis.
  211
Why is it not possible to become justified in believing G based on this

combination? Having a plausible explanation will bolster the previously men-
tioned intuitive judgments. I want to explore two accounts.
The first appeals to the well-known taking condition on inference. This is
the condition that one must take the premise(s) to support the conclusion in
order to infer the conclusion from the premise(s). Once the pollster comes to
think that the respondents were determining their answers by coinflip (U), it
seems the pollster is no longer entitled to a number of crucial claims about the
connection between T and G, such as:
• If T, then G
◦ If 87 percent of Texas respondents told me they will vote Republican,
then roughly that percentage of Texans are going to do so.
• It wouldn’t be that T unless G
◦ It wouldn’t be that 87 percent of Texas respondents told me they will
vote Republican unless roughly that percentage of Texans are going to.
• T indicates in the circumstances that G
◦ 87 percent of Texas respondents telling me they will vote Republican
indicates in the circumstances that roughly that percentage of Texas
are going to do so.
In fact, the pollster gains reason to believe these are false. Notice that each of
these conditionals seems to be a plausible way of spelling out what T guaran-
teeing G amounts to, and similarly for T supports G.¹⁴ Assume, then, that
the taking condition is a genuine condition on inference. Then there is an
explanation available for why you cannot become justified in believing G based
on <believing T, believing U>, as follows:
This combination provides at best inferential justification from T to believe

G. So, in coming to be justified in believing G based on <believing T,
believing U> you must be epistemically appropriate to infer G from T,
¹⁴ Isn’t guaranteeing much stronger than mere support? As the terms are used in ordinary parlance,
yes. I use ‘support’ here only because it is the standard term used for the taking condition on inference.
However, its ordinary meaning is not sufficiently strong for the purposes of giving an account of
inference. I might take the presence of some clouds in the sky to support rain, but I certainly won’t infer
it will rain based just on such slim support. By contrast, if I take the particular clouds to establish,
ensure, imply, indicate—guarantee—rain then I will make the inference. If I don’t so take them, I may
well not make the inference. There are complexities here about whether ‘guarantee’ is too strong, but
I don’t think they matter to our discussion, since the undercutters at issue are reasons not merely to
deny that there is a guarantee but also to deny that there is strong support. To keep things simple, then,
I use ‘support’ and ‘guarantee’ interchangeably.
212    
despite believing U; but for your inference to be epistemically appropriate,

since inference in this case requires taking T to support G, your taking T to
support G must be epistemically appropriate, too, which it is not, because
U gives you reason to doubt that T guarantees or supports G.¹⁵
Believing U spoils the epistemic appropriateness of taking T to support G,

which prevents one from becoming justified in believing G based on <believ-
ing T, believing U>.
This argument relies on the assumption that for an inference to be epistem-
ically appropriate, the relevant taking must be epistemically appropriate.
Suppose you have good reason not to take P to support Q, say because you
have reason to think that P does not support Q. Then if inferring involves
taking, it seems you should not make the inference.
But is the taking condition a plausible condition on inference? There are
prima facie reasons to accept it. It helps to explain, as Hlöbil (2014) notes, why
“P, so Q, but P does not support Q” seems incoherent. In affirming the second
conjunct you are explicitly denying something which you are taking to be the
case insofar as you are making the inference expressed in the first conjunct.
The taking condition would also help to explain, as Boghossian argues, how
inference differs from mere association: we can be responsible for the infer-
ence itself and not merely for the output belief—and we can be directly
responsible, rather than merely responsible for it in virtue of being responsible
for doing something or omitting doing something which could foreseeably
have led to the inference, as is the case for associations. The taking condition is
not perhaps the only way to explain these facts, but it is one way.
But the taking condition brings with it a host of problems as well. What is
taking? Is it believing? If so, then presumably for inference to transmit
justification from premises to the conclusion, the taking state must itself be
justified, which raises at least the threat of a regress.¹⁶ If taking isn’t believing,
¹⁵ One could alternatively recast the previous explanation in terms of remaining justified in believing
G on the basis of the combination <believing T, believing U>. In place of the language of “being
appropriate to infer G from T” one could substitute “being appropriate to continue to base inferentially
one’s belief that G on one’s belief that T.”
¹⁶ However, one might claim that both this linking belief and its justification are derivative from the
disposition to make the inference and the transmissibility of justification available through the inference.
For an example of someone who argues that the linking belief is derivative from the reasoning, consider
Broome (Forthcoming): “my account of reasoning entails that, if you reason from premises to conclu-
sion, you implicitly have the linking belief. You believe the conclusion as a result of a reasoning process
starting from your believing the premises. This shows you implicitly believe that the conclusion follows
from the premises. This implicit belief is constituted by your disposition to derive the conclusion from
the premises by a process of reasoning.” If one accepts this view of linking beliefs, it’s natural to claim
that their justification derives from the justificational properties of the available inference.
  213
what is it? And why should taking P to support Q be epistemically inappro-

priate if you have good reason to think P doesn’t support Q? The taking state
presumably cannot be a seeming, for instance, since it is not in general
epistemically inappropriate for it to seem to you that something is the case
when you have good reason to think it isn’t the case.
Given such worries about the taking condition, it might seem I’m hitching my
wagon to a controversial position about inference. Fortunately, a similar, second
explanation doesn’t presuppose the taking condition. In essence, we can just
omit the taking step in the previous explanation. Why is it not possible to
become justified in believing G based on <belief in T, belief in U>? The answer is:
This combination provides at best inferential justification from T to believe

G. So, in coming to be justified in believing G based on <believing T,
believing U> you must be epistemically appropriate to infer G from T,
despite believing U, which you are not, because U gives you reason to
doubt that T guarantees G.
Continuing to assume that ‘guaranteeing’ and ‘supporting’ are interchangeable

in this context, the crucial claim is that it cannot be epistemically appropriate
to infer G from T because U gives you a reason to doubt that T supports G.
This is explanation does not depend on the taking condition. The fact that
inferring from the premise T to conclusion G would be epistemically inappro-
priate if one has reason to doubt that the premise wouldn’t be true unless the
conclusion was true—or if you prefer that T supports G—is plausible on its
own without assuming the taking condition. In fact, its plausibility provides at
least some reason to accept the taking condition, insofar as the taking condi-
tion would help explain why inference would be epistemically inappropriate in
such situations.
A key background assumption in these two explanations is that inference from
premise(s) to a conclusion is something that itself can come up for assessment as
epistemically appropriate or not, ex ante and ex post. This seems plausible.¹⁷ You
can be epistemically inappropriate to infer even when you know the premises,
and epistemically appropriate to infer—to draw the inference—even if your
beliefs in the premises are epistemically inappropriate or unjustified. Once
you’ve inferred, if you weren’t epistemically appropriate to infer, we can criticize
your actual inference as epistemically inappropriate (ex post).
¹⁷ See also Balcerak Jackson and Balcerak Jackson (2013: 115). They focus on normative evaluation
of inferences people draw, and so on ex post rather than ex ante epistemic appropriateness.
214    
I conclude that Pollock-style undercutting defeat can occur inferential cases,

with no need for higher-order contributions.
9.4. Undercutting Defeat for Noninferential Justification?
The conclusion of the previous section raises a question. Given that there is
genuine undercutting defeat in the Polling case and in other cases of inferential
justification, why wouldn’t the same hold for cases of noninferential justifica-
tion such as the Milk Taster case? Why, in particular, wouldn’t the explan-
ations discussed in the last section apply just as well to non-inferential cases?
This section aims to answer these questions.
Is there a taking condition on experience-to-belief transitions (of the sort
that can result in non-inferential justified belief)? This might seem implausible
on its face (how could a taking be required for noninferential transitions;
wouldn’t that make them inferential?¹⁸), but it’s worth looking into the details.
If a taking condition holds for these transitions, then either the taking state
must be explanatory of the transition or it needn’t be. Let’s look at both
possibilities in turn.
A taking condition for experience-to-belief transitions cannot be explana-
tory of these transitions. Consider Boghossian (2014) on inference. He thinks
that it’s because you take your premises to support your conclusion that you
“draw” your conclusion. But this clearly depends on your already believing the
premises. If you take P to support Q but don’t believe P, you will not be in a
position to infer Q. In the same way, your taking some fact about your
experience to support Q, or even your taking your experience itself to support
Q, can’t by itself explain why you come to believe Q. Takings are relevantly like
commitments to conditionals, and for a commitment to a conditional to
explain belief in the consequent, one needs a commitment to the antecedent,
so that the content of this state together and that of the taking state “lock
together” to explain the transition to the belief in Q. Thus, one must be
committed to that which one takes to support Q—the fact about the experi-
ence or the experience itself. Plausibly, a belief is needed for commitment to
¹⁸ Balcerak Jackson and Balcerak Jackson (2013) claim that where there is no gap between the
contents of states involved in a transition, it is implausible to posit inference. One might think that in
genuinely perceptual justification, the transition is from an experience with the content P to a belief
that P.
  215
the antecedent. I must believe this fact about the experience obtains or that the
experience itself exists.¹⁹
What if a taking state is required for the relevant transitions but isn’t
explanatory of them? John Broome (Forthcoming) is sympathetic to the
view that if one reasons from beliefs in P1, . . . , Pn to a belief in Q then one
has a disposition to reason from the former to the latter, and that this
disposition is an implicit belief that Q follows from P1 . . . Pn. This implicit
“linking belief” doesn’t help explain inference; rather, the nature of inference
guarantees that one has such a belief. Could something like this hold for
experience-to-belief transitions? I don’t think so. The relevant disposition
exercised in a noninferential experience-to-belief transition doesn’t seem, by
itself, to constitute such a linking belief.
Why not? It’s instructive to consider why Broome himself denies that every
instance of reasoning requires a linking belief. He gives the example of
reasoning involving transitions between intentions, such as instrumental
reasoning in which one starts with an intention for an end and a belief
about the means to that end and concludes with an intention concerning the
means. Unlike the case of inference involving only beliefs, in which the
contents for the linking belief would not go significantly beyond the contents
of the beliefs involved, adding only whatever content is needed for the
conditional, the contents of the linking beliefs corresponding to the instru-
mental reasoning with intentions would contain new material going far
beyond the contents of the intentions/beliefs involved and beyond the con-
tents needed for a conditional; the linking belief would have to be about
intentions—it would have to be a belief that if I had such and such intention,
then . . . . The same goes for experience-to-belief transitions; the linking belief,
here too, would need to be about experiences. But it isn’t plausible that your
having the disposition to make experience-to-belief transitions ensures that
you have a linking belief with such a rich content. Not only that, but as
Broome notes, it’s implausible to think that young children have such linking
beliefs, but not implausible to think that they engage in the relevant instru-
mental reasoning with intentions. The same goes for the experience-to-belief
transitions.
So, the explanation based on the taking condition, even if successful in the
case of inferential justification, doesn’t carry over to explain how there can be
¹⁹ It’s seeming that a fact obtains or that an experience exists is not enough. Seemings do not ensure
the needed commitment of the subject. The point is familiar: if it seems to me that one rod being longer
than another, I need not be committed to its being longer.
216    
undercutting defeat in the case of noninferential justification from experience.

But, perhaps the second explanation, which omits reference to a taking
condition, does carry over.
For the second explanation to extend to the noninferential case, we would
have to be able to see how having reason to believe that being in experiential
state E doesn’t guarantee Q could, without higher-order help, make it epi-
stemically inappropriate for you to make a noninferential transition from
being in E to believing Q. Of course, with the right higher-order help, we
can explain the inappropriateness of forming a belief. Supposing you knew
that if you believe Q, it would be on the basis of experience E, you could reason
like so:²⁰ there is a good chance that I could have E without Q being the case;
now, if I believe Q it would be on the basis of E; and so, if I believe Q there is a
good chance I would be basing my belief on a source that is unreliable in the
circumstances. And if you (justifiably) believed the conclusion of this reason-
ing, this would explain why it would be epistemically inappropriate for you to
form that belief. All this is with the higher-order help. But without it, how can
we devise any reasoning, available to the subject, that concludes with some-
thing belief in which, or justified belief in which, makes the formation of the
belief epistemically inappropriate? The conclusion there is a good chance that
I could have experience E without Q being the case doesn’t seem by itself to
make forming a belief in Q (based on E) epistemically inappropriate, for one
might have no idea that one would be relying on E. The best we can get is this:
If I rely on E in believing Q, my belief that Q has a good chance of being

formed in a way that is unreliable in the circumstances.
But this by itself doesn’t make the transition from experience E to believing
Q epistemically inappropriate. Whereas in the inferential case a reason to
doubt that P wouldn’t be true unless Q seems by itself to make the transition
from the belief that P to the belief that Q inappropriate, a reason to doubt that
having experience E supports Q at most makes it inappropriate to transition
from the belief that you have E to the belief that Q, but doesn’t make it
inappropriate to transition from the experience E itself to the belief that
Q. The Milk Taster case exemplifies this contrast well. If the milk taster
believed she was basing her belief on smell, it would be inappropriate for her
²⁰ I assume here that a doubt could appear in thought as an affirmation of a statement about chance.
One might question this. However, these issues could be bypassed by thinking of a case in which the
subject believes (justifiably) that it’s not true that she wouldn’t have E unless Q. We could then just load
this negation into the reasoning.
  217
to keep the belief; but if not, and especially if she believes (justifiably) that she
isn’t basing her belief on smell, it doesn’t seem inappropriate.
One might balk here, and indeed balk at Sturgeon’s (and my) judgments
about the Milk Taster case. Isn’t there something epistemically inappropriate
about basing a belief on a source while believing that there is a good chance
that that source is unreliable, even if you don’t have believe it’s your source?
But just what is inappropriate about doing this? It’s true that an ideal cognitive
agent would do no such thing, but that is because an ideal cognitive agent
would know its sources (Mellis 2014). It doesn’t follow that we lack justifica-
tion from the fact that we are non-ideal in this way. It might be claimed that
the real problem is that you’re conducting yourself in such a way as to
guarantee that your resulting belief fails to be knowledge if your belief about
your source’s unreliability were true and you still believed on the basis of that
source. If the milk taster’s belief that olfaction is unreliable in the circum-
stances were true, and if she still based her belief about the milk on olfaction,
it’s true that her belief wouldn’t be knowledge. But it’s hard to see what the
bearing of these facts is on whether the milk taster is in fact justified in
believing the milk is ok.
Thus, while there may be senses in which the milk taster’s belief is epistem-
ically non-ideal and senses in which the belief would be problematic if the
belief about the unreliability of olfaction were true, I don’t think these defects
are ones a theory of defeat needs to capture, because they do not eliminate or
reduce justification.
9.5. A Tool for Distinguishing Inferential from

Non-inferential Justification
The conclusions from the previous two sections imply that there is no such
thing as Pollock-style undercutting defeat of noninferential justification, only
of inferential justification. The nearest thing to genuine Pollockian attack-the-
connection defeat of a noninferential justification is what I’ve called higher-
order defeat: defeat in virtue of having reason to think that one’s belief is
unreliably based, a kind of defeat that requires beliefs about one’s bases or
sources and so about oneself and one’s mind. If these conclusions are correct,
we might use them to help us determine whether a justification in a certain
case is inferential or not. This section considers the implications for perceptual
and testimonial justification.
218    
First, perception. Suppose there can be no undercutters for noninferential

justification, only higher-order defeaters. And suppose you have noninferen-
tial justification from experience to believe that a wall is red. Then, if you learn
that the lighting is abnormal, and this defeats your justification, it must be
because this belief, together with a belief about the basis for your belief about
the color of the wall, give you a reason to think that your belief is (or would be)
unreliably formed. That is, the defeat occurs only because you have reason to
think something about your own mind, about your belief and its formation. Or
suppose you have noninferential justification from experience to believe that
the thing before you is an eggplant, and then you learn that there are look-alike
fake eggplants about. This, too, would serve as a defeater for your justification
only if, together with your belief about the source of your belief, it gives you
reason to think your belief is or would be unreliably formed. This is surprising.
It is surprising that defeat in such a case should require beliefs about oneself.
Where you have a certain perceptual justification for believing P, borrowing
terminology from Pryor (2004), let’s say that a non-perceiving possibility for
you for P is a possibility in which you have this same justification but you are
not be in a position to perceive that P. Some such possibilities are objective:
they are conditions of the world around that preclude you from perceiving that
P solely by affecting the stimuli available to you. Objective non-perceiving
possibilities include the abnormal lighting or the abundance of look-alike
fakes. It is implausible that evidence for objective non-perceiving possibilities
should require evidence about oneself. This is not so implausible in the case of
defeat through learning of subjective non-perceiving possibilities such as

hallucination or some defect of one’s eyes. (I don’t claim these categories
exhaust non-perceiving possibilities.) It wouldn’t be surprising to find out
that the way evidence of the obtaining of these possibilities defeat is by helping
give you reason to believe your belief is or would be unreliably formed.
To further confirm these claims, consider what it would take to reply
adequately to a persistent challenger, when the two of you are standing in
front of something that looks like an eggplant:
: “Why don’t you conclude it’s an eggplant?”

: “Because there are look-alike non-eggplants all about.”
: “What does there being look-alike non-eggplants all about have
to do with it?”
: “Because if there are, then my experience would be unreliable, and since
if I believed it would be on the basis of experience, my belief would be
unreliable.”
  219
It might be surprising that you start talking about yourself in the last line, but
it’s these facts about oneself that articulate the source of your justification for
not concluding it’s an eggplant. Contrast this final reply with a more
natural one:
: “Because if there are look-alike fakes around, then, even though this
looks like an eggplant, it could easily be one of the fakes.”
Similarly, in reply to a challenge about why you don’t conclude the wall is red
in a case of strange lighting, you might say:
: “Because if the lighting conditions are abnormal, then, even though the
wall looks red, it could easily be some other color.”
These more natural final replies don’t appeal to one’s own mental states. The
proposition that the thing looks like an eggplant does not entail anything
about my mind, or arguably about other people’s minds. It is a fact about
eggplants that can survive my demise and that of the rest of humanity.
Is there a way to block the “mentalization” of such defeaters for perceptual
justification? The classical foundationalist takes one’s justification in the
relevant cases to be inferential but still takes the justification to come from
equally mentalized premises, about which experiences one has. But there is
another sort of inferential approach available, suggested by the more natural
replies to the previous Challenger character. Suppose the structure of my

perceptual justification for believing this is an F (e.g., this is an eggplant) is
as follows: I’m inferentially justified in believing this is an F in virtue of being
justified in believing it looks like an F.²¹ If such an account is correct,
information about rampant look-alikes can defeat my inferential justification
for thinking a thing is an F without the assistance of higher-order contribu-
tions. The information about look-alike fakes can defeat in the way that
genuine undercutters do, by being a reason to doubt a connection holds
between one consideration—this thing has a certain look—and another—
²¹ There would need to be a story of how the visual experience contributes to the justification of the
belief about the thing’s looking like an F. Are the latter beliefs noninferentially justified, or are they
themselves inferentially justified? Eventually, we must get down to beliefs that are not inferentially
justified. For them, undercutting defeat would be impossible. I am sympathetic to the view that the
noninferentially justified beliefs are beliefs about things having certain looks, and I suspect that such
beliefs aren’t subject to undercutting defeat, only higher-order and rebutting defeat of an object justifies
me in believing it looks like an F, and this, perhaps together with some background information or a
default assumption, inferentially justifies me in believing that it is an eggplant.
220    
this thing is an eggplant. Such a doubt makes it epistemically inappropriate to

make the inference. A similar explanation would hold for how information
about lighting conditions defeats. Information about lighting condition spoils
the connection between a thing’s having a certain look that distinctive of red
and a thing’s being red, making the inference epistemically inappropriate.
There are various worries one might raise about such an account.²² What
are these “objective looks”? Isn’t the resulting view of perceptual justification
too intellectualized? Relatedly, do we really go through these reasoning pro-
cesses when we see an eggplant? I will not try to answer these objections here.
I only note that this view has one point in favor of it against the standard view
according to which one’s perceptual justification for believing this is an F is
noninferential. On that view, if the conclusions of this paper are right, there is
no such thing as simple attack-the-connection defeat—Pollock-style under-
cutting defeat—of one’s perceptual justification. Even objective-condition
defeaters about lookalike fakes or lighting conditions only defeat because
they join up with beliefs about one’s beliefs’ sources to generate a higher-
order defeater. And that is implausible.
An easier case, though, is testimony. One familiar epistemological theory of
testimony holds that being told that p provides noninferential justification for
believing that p (Burge 1993; Graham 2006). If this view is correct, then
information that one’s testifier is insincere or incompetent cannot by itself
defeat one’s testimonial justification; rather one must build up a higher-order
defeater such as “if I believe p, it would be on the basis of this testimony, which
is unreliably in the circumstances, and therefore my belief would also be

unreliable in the circumstances.” But again, this seems wrongly to assimilate
objective and subjective non-learning possibilities. Non-learning possibilities
for me for P are possibilities in which I have the same P-related testimonial
justification that I actually have but in which I’m not in a position to learn that
P through this testimony. Non-learning possibilities such as insincerity or
incompetence of a testifier are objective ones, whereas possibilities such as my
now suffering a momentary bout of receptive aphasia are subjective ones. The
subjective ones do seem to defeat only because they contribute to a higher-
order defeater, but this doesn’t seem true of the objective ones. Again, it
doesn’t seem that adequate replies to challenges invoking objective non-
learning possibilities must ultimately appeal to considerations about oneself
and one’s beliefs. It seems adequate to point out, e.g., that if the testifier is being
insincere then they could easily have told me that P despite not-P.
²² For some development of this sort of account, see McGrath (2017, 2018).
  221
We can avoid these consequences if we take testimonial justification to be

inferential. For then we can appeal to Pollockian undercutting. The informa-
tion that the testifier is insincere gives you a reason to doubt that the testifier
word guarantees the truth, and this is a reason not to infer its being so from the
testifier’s saying it.
9.6. Conclusion
Contrary to Sturgeon, I’ve argued that Pollock-style undercutting defeat—

defeat via attacking-the-connection between a reason and the proposition
believed—is possible in cases of inferential justification. But, I’ve agreed with
him that there is no such thing in the case of noninferential justification, and
that the nearest cousin to it is what I’ve called higher-order defeat, i.e., defeat
by having reasons to believe the source of one’s belief is unreliable. Finally, I’ve
explored how we can use this difference between the ways inferential and
noninferential justification can be defeated to make progress on questions on
whether a certain sort of justification is inferential or not, my test cases being
perceptual and testimonial justification.²³
References
Audi, R. (1993). The Structure of Justification. Cambridge: Cambridge University

Press.
Balcerak Jackson, Magdalen and Brendan Balcerak Jackson. (2013). Reasoning as
a source of justification. Philosophical Studies, 164: 113–26.
Boghossian, Paul. (2014). What is inference? Philosophical Studies, 169: 1–18.
Broome, John. (Forthcoming). Linking beliefs. In: Balcerak Jackson, Magdalen
and Brendan Balcerak Jackson (eds.) Reasoning. Oxford: Oxford University
Press.
Burge, Tyler. (1993). Content preservation. Philosophical Review, 102(4): 457–88.
Casullo, Albert. (2018). Pollock and Sturgeon on defeaters. Synthese, 195:
2897–906.
Comesaña, Juan. (Forthcoming). Empirical justification and defeasibility. Synthese,
²³ I am grateful for comments from an audience at the University of Pavia, especially Tommaso
Piazza, James Pryor, and Crispin Wright. For comments on written drafts, I thank Jessica Brown, Al
Casullo, Juan Comesaña, and Jack Lyons.
222    
Graham, Peter J. (2006). Testimonial justification: Inferential or noninferential?

Hlöbil, Ulf. (2014). Against Boghossian, Wright and Broome on inference.
Philosophical Studies, 167: 419–29.
McGrath, Matthew. (2017). Knowing what things look like. Philosophical Review,
126(1): 1–41.
McGrath, Matthew. (2018). Looks and perceptual justification. Philosophy and
Phenomenological Research, 96(1): 110–33.
Mellis, Giacomo. (2014). Understanding undermining defeat. Philosophical
Studies, 170: 433–42.
Pollock, John. (1970). The structure of epistemic justification. American
Pollock, John. (1974). Knowledge and Justification. Princeton: Princeton
University Press.
Pollock, John. (1987). Defeasible reasons. Cognitive Science, 11: 481–518.
Pollock, John and Cruz, Joseph. (1999). Contemporary Theories of Knowledge. 2nd
ed. Lanham, MD: Rowman and Littlefield.
Pryor, James. (2004). What’s wrong with Moore’s argument. Philosophical
Sturgeon, Scott. (2014). Pollock on defeasible reasons. Philosophical Studies, 169:
105–18.
10
Defeaters as Indicators of Ignorance
Julien Dutant and Clayton Littlejohn
10.1. Introduction
Consider three desiderata for a theory of defeat.
1. The theory should be extensionally adequate. We figuratively scoured

the globe to give you this list of the different types of defeaters proposed in the
literature:
1. Opposing Defeaters
2. Undercutting Defeaters
3. Reason-Defeating Defeaters
4. Collective Defeaters
5. Pragmatic Defeaters
6. Identification Defeaters
7. Higher-Order Defeaters
8. Negative Self-Appraisal Defeaters
We can think of the different items as standing for the different mechanisms
by which something threatens rationality. A good theory of defeat should tell
us which of the things on the list, if any, are genuine defeaters.¹
2. Monistic views are prima facie preferable to pluralistic views. Monists
and pluralists agree that there are different kinds of defeaters, but the monist
¹ In addition, there are discussions about whether defeaters are pieces of evidence, psychological
states or contents, things that we should have known about or believed, or facts. As interesting as these
discussions are, they don’t directly address questions about how defeaters defeat. And that is our focus.
For good discussions of normative defeat, see Goldberg (2017); Lackey (2006); Madison (MS). For
discussions of psychological defeat, see Bergmann (2006); Lackey (2006).
224     
thinks that there’s some unifying explanation of their rational toxicity.²

Monistic views have greater explanatory power.
3. It is prima facie preferable that our theory of defeat can be embedded in a
larger normative framework.³ We want a theory of what makes rational belief
rational that explains why defeaters threaten rationality. As Grundmann
(2011) observes, many theories of rational belief include non-redundant no-
defeat clauses. Such views would fall short of this ideal.
Here is our plan for what follows. In Section 10.1, we give a brief presen-
tation of some of Pollock’s seminal ideas about defeat and introduce an
alternative view that seems to satisfy our second and third desiderata. This
account characterizes rational belief in terms of confirmation and then char-
acterizes defeat in terms of things that have a negative impact on confirmation.
We shall argue that this approach cannot handle certain kinds of higher-order
defeat and that it mischaracterizes some epistemically beneficial factors as
defeaters. It doesn’t satisfy our first desiderata. In Section 10.2, we explain
where defeat should be located in a normative theory and argue that the first
view ran into trouble because it is a truth-centred view. In Section 10.3, we
propose a knowledge-centric view as an alternative, one that builds on the
strengths of the first truth-centred approach but improves upon it because it
handles the cases that cause trouble for the truth-centred approach.
10.2. Defeaters as De-Confirmers
Recall Pollock and Cruz’s characterization of defeat:
 : if p is a reason for S to believe q, r is a defeater for this

reason iff (p & r) is not a reason for S to believe q.
(Pollock and Cruz 1999: 37)
They use the notion of a reason in their characterization of defeat. On their

view, p is a reason to believe q iff it is possible for this thinker to rationally
² Recall Spohn’s (2002) complaint that some theories (e.g. Pollock’s) were normatively deficient
because they offered laundry lists of (alleged) defeaters and no unified explanation as to why these
different defeat relations were toxic.
³ See Williams (1965: 107) for an interesting argument that the reasons we have to believe in
pluralism in the practical domain do not carry over to the epistemic domain. There might be reasons to
prefer a pluralistic view to a monistic one, but the existence of such reasons doesn’t change the facts
about what’s prima facie preferable to what.
     225
believe q by believing it on the basis of p. A reason, so understood, is defeasibly

sufficient (i.e. something that is sufficient for ex ante rationality in the absence
of defeaters). They also say that defeaters are themselves reasons. For example,
in the case of opposing defeat, they say that when p is a reason for S to believe
q, r is a rebutting defeater when r is a reason for S to believe not-q.
We think that this reliance on defeasibly sufficient reasons is problematic.⁴
If we wanted to give accounts of partial and full defeat, we wouldn’t want to
deploy the concept of a defeasibly sufficient reason in characterizing partial
defeat. If a thinker acquires some evidence against p, that might have some
negative impact on the rational status of the thinker’s belief even if relevant
evidence is not a defeasibly sufficient reason for believing p’s negation. Partial
defeat doesn’t require defeasibly sufficient reasons. Full defeat doesn’t require
them, either. Suppose that it’s rational to believe p only if the total evidence
provides sufficiently strong support for believing p. It’s possible for even very
weak evidence to bring the degree of support below the threshold. Something
can function like a full defeater for a belief in p without being a defeasibly
sufficient reason to believe ~p.
We might try to improve upon this by introducing some other notion of a
reason, such as evidential reasons:
: e is a reason to believe h iff e confirms h, that is, P(h|e) >

P(h).⁵
Evidence is better suited to account for partial defeat. To characterize a full

defeater, further normative machinery is needed. We need to introduce a
normative standard so that we know when evidence would be sufficient:
: it is rational for S believe q iff P(q) is sufficiently high.⁶
⁴ We should note that much of the literature on the defeat of rationality or justification is concerned
with the way in which defeaters defeat status or standing. On Pollock and Cruz’s characterization,
however, defeaters defeat reasons. The connection between defeating reasons and status is not obvious
without appeal to further assumptions about reasons and rationality. Because one of us doesn’t believe
that all beliefs need to be supported by reasons to be rational, it’s not clear to us how to connect the
defeat of reasons to the defeat of status. For arguments that beliefs needn’t be based on evidence or
reasons to be knowledge or to be rationally held, see Anscombe (1962); Littlejohn (2017); McGinn
(2012). For discussion of a more reason-centric approach, see Lord (2018).
⁵ Let’s assume that the relevant notion of probability is some kind of subjective probability or
evidential probability that captures the confirmation relation in question. Maria Lasonen-Aarnio
rightly pointed out that Pollock’s approach is not probabilistic. We are thinking of this as a way of
trying to build on some of his ideas in a probabilistic approach.
⁶ See Dorst (2019); Easwaran (2016); Sturgeon (2008) for defences of .
226     
If we incorporate  into our account of rational belief, we can give

an account of partial and full defeat in terms of confirmation. Any defeater
(partial or full) has to have a negative impact on ex ante or ex post rationality.
Plausibly, it must be in the confirmation game. We can distinguish between
different kinds of defeaters by thinking about the different ways that they
negatively impact confirmation. In keeping with , we can deter-
mine whether a partial defeater constitutes a full defeater by thinking about the
degree of support that remains after we have taken account of the partial
defeater’s impact.
We can now characterize defeat as follows:
-: defeaters are de-confirmers.⁷
A de-confirmer negatively impacts the confirming support the evidence pro-

vides. As with confirmation and disconfirmation, de-confirmation is a matter
of degree. A de-confirmer/defeater only has to confirm to some degree to be a
partial defeater. A partial defeater constitutes a full defeater when it ensures
that the total evidence doesn’t provide sufficiently strong support.
- satisfies our second and third desiderata. The view is
monistic. We know what it is that makes a defeater a defeater (i.e. it is a de-
confirmer). The view can be embedded in a larger normative framework.
Although we can include a no-defeat clause in , it would be
redundant. As an added bonus, the normative framework that supports this
view of defeat is independently quite plausible. Most epistemologists currently

accept this theory of the epistemic good:
: the fundamental epistemic good is true belief and false belief is the
fundamental epistemic evil.⁸
Rational believers have the twin goals of acquiring true beliefs and avoiding
false beliefs. They’re often uncertain about whether they achieve these goals by
forming a belief and so they need to strike a balance between unreasonable
aversion to risk and recklessness. They rationally ought to believe in ways that
maximize expected epistemic value. This gives us the argument for .⁹
⁷ Chandler (2013) and Kotzen (2019) provide sophisticated formal characterizations of different
defeaters in the spirit of -.
⁸ See Goldman (1999); Lynch (2004); Sylvan (2018) for defences of veritism.
⁹ See Easwaran and Fitelson (2015); Dorst (2019) for details.
     227
Not all epistemologists want to characterize rational belief in terms of a

connection to epistemic value. Some prefer to characterize rational belief in
terms of a connection to epistemic norms. This would be in keeping with a
truth-centric approach:
 : we ought to believe p if we can believe it and therein believe

the truth. We ought not to believe p if we cannot believe it and therein believe
the truth.
If we believed in such objective norms, we wouldn’t think that rational belief

required conforming to such norms. If we are rationally uncertain, say, about
whether, say, this coin will land heads when flipped, there’s a sense in which
we ought to suspend. One way to characterize this more subjective normative
notion is by providing a theory of what would be rational to do give uncer-
tainty about our objective obligations. One approach begins by ‘consequentia-
lizing’ the truth-centric theory of what we objectively ought to believe (e.g. by
giving a cardinal ranking of our options that represents the weights of the
reasons associated with these norms). We could characterize what we subject-
ively ought to believe or what we could rationally believe in terms of the
minimization of expected objective wrongfulness.¹⁰ This gives us a new rationale
for . The reason we ought (in a sense) believe in accordance with
 is that by so believing we rationally respond to the two normative
pressures that we’re under.
10.2.1 Back to the List
It is easy to see how - accommodates opposing and under-

cutting defeaters. In cases of opposing defeat, a thinker acquires a defeater that
suggests that the thinker’s belief is false (e.g. the thinker believes that it is
Sunday but sees neighbors rushing to work). In cases of undercutting defeat, a
thinker acquires a defeater that suggests that her grounds for believing do not
support her belief (e.g. the thinker believes that it rained last night because of
the wetness of the pavement but sees that a sprinkler has been running all
¹⁰ This approach to subjective permissibility is developed in Lazar (Forthcoming); Olsen (2018).

Littlejohn (2020) applies these ideas to the epistemic case by consequentializing knowledge norms and
showing how this gives us a theory of what we ought to believe in objective and subjective senses.
228     
morning).¹¹ If P is your initial probability function and you acquire p as a

reason to believe q, P(q|p) is higher than P(q). When r is an opposing
defeater, it confirms ~q in the context one’s current evidence or beliefs,
namely, P(~q|p&r) > P(~q|p), so we have P(q|p&r)< P(q|p). When r is an
undercutting defeater, it takes q towards its prior probability P(q) in that
context, so we have P(q)≤ P(q|p&r) and P(q|p&r) < P(q|p).
In cases of reasons-defeating defeat, a thinker acquires a defeater that
suggests that the thinker’s supporting reason is false (e.g. you initially believe
because of your perceptual experiences that there is beer in the fridge (b) and
this convinces you that your partner has been shopping (s) but then you
discover later that there was no beer there at all). Here are two ways to think
about these cases. First, we might think that b was incorporated into our body
of evidence, it supported the belief in s, but then a later a discovery kicks b out
of the evidence and thus defeats the thinker’s belief in s. The problem is that in
the probabilistic setting, once b is incorporated into the evidence and treated
as an input to conditionalization, there is no rational process for removing b.
Its probability on the evidence is 1, so no new evidence can lower it. If we want
to give an account of how s’s status is threatened by a reason-defeating defeater
and we are supposed to model this by explaining how learning b boosts s’s
probability and then s’s probability decreases because new evidence knocks b
out of the evidence, this cannot be done.
Perhaps those who believe in reason-defeating defeat don’t think that our
reasons for believing things like s are necessarily part of our evidence. Perhaps
some of our reasons for believing things like s are just the contents of justified
beliefs. Perhaps b is just such a reason in this sense. In that case, we might
imagine that there is some evidence e which supports b and which supports s.
(Perhaps e is a proposition about our perceptual experiences, but the precise
content doesn’t matter provided that e does not entail b or s.) Then we could
say that upon learning e, the probability of b and s increase and allow that it’s
possible that we learn something, d, where d is a reason-defeating defeater
because d lowers b’s probability without affecting the conditional probabilities
of s on b or on ~b.¹²
¹¹ Undercutting (or undermining) is sometimes understood very broadly (e.g., by Brown (2018)
who treats higher-order defeat as a kind of undercutting or undermining defeat), but we want to
understand it narrowly because we think that - cannot accommodate certain kinds of
higher-order defeat.
¹² Let P be your probability after e is acquired. For b to be a reason to believe s it must be that P(s|b) >
P(s|~b). Recall d doesn’t affect the conditional probabilities P(s|b) and P(s|~b), so P(s|b&d) = P(s|b) and
P(s|~b&d) = P(s|~b). Finally, d lowers the probability of b, so P(b|d) < P(b). From this, it follows that d
de-confirms s, namely, P(s|d) < P(s).
     229
Consider collective defeat. In cases of collective defeat, a thinker is said to

acquire a defeater that suggests that some belief in a set involving multiple
beliefs is false without suggesting that any belief in particular is false (e.g. you
memorize the phone book and believe each entry in it only to discover the
errata slip that says that the book contains an error without providing further
information). We are skeptical of some of the things that have been said about
the power of collective defeaters. Some epistemologists have defended views
on which it’s not possible to have a set of justified beliefs known to be
inconsistent (e.g. because that this knowledge of the presence of an unidenti-
fied falsehood defeats the justification for each belief in the set).¹³ We think
that this is not what happens in the case we’ve just described. What might
happen is that reading the errata slip lowers our confidence in some of the
things we believe on the basis of what we read, but it’s clear that it wouldn’t
bring the degree of support for each belief below threshold. One virtue of -
 is that it seems to explain why so-called collective defeaters
might diminish the strength of support without being full defeaters for the
relevant set of beliefs.¹⁴
One further virtue of - is that it can shed light on prin-
ciples of justification associated with collective defeat, such as those that
impose a consistency requirement on justified belief. Suppose that you
were .95 confident that each of the entries in the phone book that you had
committed to memory were correct. And suppose you learn that there was
precisely one error in this book. Under these conditions, it seems that you
acquire a reason to increase your confidence in your beliefs. While fans of

collective defeat might think that this piece of information is rationally toxic, it
might seem to actually be either neutral or beneficial.
- should have no difficulty accommodating pragmatic
defeaters if this is desired.¹⁵ The way that pragmatic defeaters are usually
understood, they are taken to be practical considerations that are not directly
connected to the truth-indicative factors. They defeat by giving us reason not
to believe without stronger evidence than might otherwise be required.
¹³ See Leitgeb (2017); Pollock (1986); Smith (2016); Ryan (1991, 1996) for defences of views that
include this consistency requirement on justified belief.
¹⁴ See Christensen (2004); Easwaran and Fitelson (2015); Makinson (1965); Worsnip (2016) for
arguments that collective defeaters don’t invariably defeat the rational status of our beliefs. In Dutant
and Littlejohn (2020), we argue that views that posit collective defeaters are either overly sceptical or
too externalist (i.e. they deny us too much justification or they pick winners and losers on the basis of
considerations beyond the thinker’s ken).
¹⁵ For arguments that pragmatic considerations can help to determine whether a body of evidence
provides sufficient support for belief, see Fantl and McGrath (2009); Owens (2000); Schroeder (2012).
We think Schroeder is the first to describe these pragmatic considerations as defeaters.
230     
If readers believe that practical considerations (e.g. the stakes or the odds of a
salient bet) can create rational pressure to suspend without acquiring more
evidence than might normally be necessary, - can accom-
modate this kind of pragmatic defeater by positing that such defeaters move
the threshold.¹⁶
While - has a number of virtues, we struggle to see how it
handles further kinds of defeat that are discussed in the literature. Peacocke,
believes there are identification defeaters:
[I] distinguish two kinds of defeasibility, which I will call defeasibility of

identification and defeasibility of grounds. A ground for accepting a propos-
ition can be conclusive even though our entitlement to believe that we have
identified such a ground is defeasible. Identifying something as a conclusive
ground is one thing; its being a conclusive ground is another. My confidence
that something is a proof can be rationally undermined by the report of
mathematicians whose competence I have reason to believe far outstrips my
own. Nonetheless, a proof is a conclusive ground. Here we have defeasibility
in respect of identification but not defeasibility in respect of grounds.
(Peacocke 2004: 30)
At first glance, this might look like a case of undercutting defeat. The expert
tells you that the proof is flawed so you might agree that the putative support
doesn’t actually support the conclusion. If this is a case of undercutting defeat,
- handles it straightforwardly. However, Peacocke intends

this case to be one in which the defeater fails to de-confirm one’s belief. The
idea is that if we have maximal support for the proposition. So, if we have a
defeater, it is not a de-confirmer.
We can see why some might be cool to the idea of so-called identification
defeaters. It doesn’t seem that we need to be able to identify conclusive
grounds as such to be justified on their basis. If so, it’s not obvious that the
considerations that show that we cannot know that some ground is conclusive
are enough to make it irrational to hold the relevant belief. Still, scale and
degree seem to matter. If some considerations show that it’s insufficiently
likely that you have suitable means for settling some question (e.g. because an
expert has shown you that you’ve made a large number of mistakes in your
¹⁶ For discussion of such cases, see Fantl and McGrath (2009); Owens (2000); Weatherson (2005).
For discussion of the ways in which moral considerations might provide pragmatic defeaters, see Basu
and Schroeder (2019); Moss (2018).
     231
reasoning), perhaps these considerations can defeat. If so, we might assimilate

them to the category of higher-order defeaters.
The defeaters described as higher-order defeaters cover a wide variety of
considerations. Some consist of evidence about our evidence (e.g. evidence
that our evidence provides only weak support for our beliefs). Some consist of
evidence that our rational faculties have been compromised (e.g. evidence that
we have been slipped a drug that affects our reasoning, that we suffer from
hypoxia, etc.). Some might consist of evidence that our beliefs would not attain
some desirable status (e.g. evidence that we could not know, evidence that it
would be irrational to believe something, etc.). Some of these cases might be
assimilated to cases of opposing, undercutting, or reason-defeating defeat, but
it doesn’t seem that all can be. Suppose you acquire some misleading evidence
that you have been slipped a drug that interferes with your reasoning abilities.
Suppose there were a simple test that could determine whether the drug was
present in the bloodstream. You simply need to place a strip of paper on your
tongue and check its colour. And suppose that you acquired this evidence that
you were slipped the drug during an important exam. If the test costs nothing
to administer, it seems to make sense to take it before submitting your exam.
Perhaps this is some indication that that the evidence that you’ve been
drugged defeats the justification you have for the beliefs that you would rely
on in answering exam questions. Note that if the evidence is indeed misleading
and you haven’t been drugged, it isn’t clear that this evidence defeated by being
an opposing, undercutting, or reason-defeating defeater.
We can anticipate three responses from fans of -. First,

they might deny the possibility of the case as we described it. They might argue
that there are level-connecting principles according to which higher-order
evidence necessarily is connected to the first-order evidence that either con-
firms or disconfirms our first-order beliefs.¹⁷ We will explain below why we
doubt that there is any true level-connecting principle that will cover the full
range of cases.
Second, someone might simply bite the bullet and deny that the putative
higher-order defeaters defeat. This would allow someone to say that all
genuine defeaters are de-confirmers. This view seems to be a kind of level-
splitting view insofar as it denies that higher-order considerations bear on the
justification or rationality of first-order beliefs. We worry that such views
suggest that certain apparently dogmatic or seemingly reckless attitudes are
¹⁷ See Dorst (2020); Dorst, Fitelson, and Husic (Forthcoming); and Williamson (2019) for discus-
sion of level-connecting principles.
232     
rational. Moreover, we think that the account will struggle with a kind of
higher-order defeat that we call negative self-appraisal defeat.
The third response is to our mind the most promising. The de-confirmers
might try to account for higher-order defeat by suggesting that it threatens
ex post rationality rather than ex ante rationality.¹⁸ The standard view is that
ex post rationality involves a basing requirement (i.e., that we competently
base our beliefs on the support that warrants them). We might account for
higher-order defeat by saying that it impacts ex post rationality on the grounds
that anyone who believes in spite of the presence of these defeaters cannot
meet the basing condition. Perhaps if Agnes believes (falsely but reasonably)
that she has been drugged, it is ex ante rational to believe the conclusion of a
proof but it fails to be ex-post rational because she cannot competently base
her beliefs on the good reasons that she has.
This strategy might help proponents of - accommodate
some higher-order defeat, but we don’t think this covers negative self-
appraisal defeat. A negative self-appraisal defeater is provided by consider-
ations about the epistemic status of some particular belief. Agnes might believe
something about, say, the afterlife. She might then be convinced that she
couldn’t possibly know anything about the afterlife. There seems to be a
rational tension between the belief that we’ll survive our own bodily deaths
and the belief that we couldn’t possibly know whether people can survive their
own bodily deaths.
Negative self-appraisal defeaters differ from the previous higher-order
defeaters in two ways. First, they concern the epistemic status of particular
beliefs (e.g. whether we know, whether we rationally believe, whether it is okay
to believe given the evidence we have, etc.) rather than pertaining directly to
the thinker’s belief forming processes or strength of evidence. Second, they
have a kind of epistemic content missing from more widely discussed cases of
higher-order defeat.
How can - handle cases like these? Suppose that a thinker
believes p and acquires evidence that confirms any of the following:
1. That she couldn’t know p.

2. That it’s irrational for her to believe p.
3. That she doesn’t have evidence that warrants belief in p.
¹⁸ This idea was inspired by Di Paolo (2018); Silva (2017). They do not endorse -,
but we thought this work might help fend off one kind of objection. See also van Wietmarschen (2013).
     233
On its face, it seems that the evidence may raise the probability of (1)–(3)
without lowering that of p itself.¹⁹ So, if negative self-appraisal defeaters defeat,
they might show that there are defeaters that are not de-confirmers.²⁰
Think about lottery cases. Harman nicely captures part of what makes the
lottery case so puzzling:
In the testimony case a person comes to know something when he is told

about it by an eyewitness or when he reads about it in the newspaper. In the
lottery case, a person fails to come to know he will lose a fair lottery, even
though he reasons as follows: “Since there are N tickets, the probability of
losing is (N-1)/N. This probability is very close to one. Therefore, I shall lose
the lottery.” A person can know in the testimony case but not in the lottery
case . . . The contrast between the two cases may seem paradoxical, since
witnesses are sometimes mistaken and newspapers often print things that
are false. For some N, the likelihood that a person will lose the lottery is
higher than the likelihood that the witness has told the truth or that the
newspaper is right. (Bergmann 1968: 166)
Given the weak anti-sceptical assumption that our testimonial beliefs are
rational and Harman’s observation about the chance of error, 
tells us that it is rational to believe lottery propositions. Suppose this is so and
suppose that we know a priori that we cannot know that we’ve lost the lottery.
If so, some Moorean absurdities cross the threshold. -
cannot account for some negative self-appraisal defeaters.

Could the proponents of - maintain that negative self-
appraisal defeaters defeat because they threaten ex post rationality by ensuring
that we don’t satisfy a basing requirement? Probably not. We fail to satisfy the
basing requirement when we cannot competently base a belief on sufficient
¹⁹ See Dorst (2020) and Williamson (2019) for arguments that show that once we allow that we can
be rationally uncertain to even a slight degree about what our evidence supports, we should expect that
there can be bodies of evidence that both make it very probable that p and very probable that it is not
sufficiently probable that p. Given , this would be evidence that would make it rational to
believe p and rational to believe that it is irrational to believe p. In truth-centric accounts, then, it seems
we can only account for the toxicity of (2) by denying  or by accepting a kind of negative
introspection principle that says that where t is the relevant threshold for rational belief (which we
assume is less than 1), P(p)<t only if P(P(p)<t) =1.
²⁰ Bergmann (2006) claims that some beliefs provide negative self-appraisal defeaters even if those
beliefs are not adequately supported. We disagree. We think that he is making the kind of mistake that
people make when they conflate a wide- and narrow-scope requirement. We would agree that a thinker
should not believe p if she believes that she does not know p or if she believes that it’s irrational for her
to believe p, but we don’t think we can detach any interesting normative claim from this even if we add
the further (psychological) claim that the thinker does believe (1), (2), or (3).
234     
reasons. In many cases of negative self-appraisal defeat, we have no reason to

think that we couldn’t meet this condition if  were correct. We are,
after all, perfectly capable of competently judging that Arsenal won (as
reported in the paper) and competently judging that it’s more likely that the
ticket lost than it is that Arsenal won. If so and if  were correct, we
should be able to competently base the belief that some ticket lost on adequate
reasons even while knowing that the belief couldn’t constitute knowledge.
We can use lottery cases to identify a second problem with
-.²¹ Suppose Agnes is thinking about two suspects in two
different crimes, Inge and Agatha. She initially thinks that the evidence of
Inge’s guilt consists entirely of statistical evidence. She initially thinks that the
evidence of Agatha’s guilt consists entirely of eyewitness testimony. She then
realizes that she had the files mixed up. In such a situation, Agnes could
acquire evidence that concerns Inge’s guilt that both lowers the probability of
guilt and makes it rational for her to believe something that she couldn’t have
rationally believed before. According to -, this should be a
case of undercutting defeat,²² but it’s a case in which the old evidence doesn’t
make a belief rational and the new evidence that provides less by way of
confirmation makes the belief rational for the first time. We don’t think this
could be a case of undercutting defeat. Meanwhile, we find that the evidence
that boosts the probability that Agatha is guilty makes it irrational to believe in
her guilt. So, this seems to be a kind of defeater that defeats while confirming,
not one that defeats by de-confirming.
10.3. Rationality, Defeat, and Epistemic Norms
Let’s step back and consider some foundational issues. We have already seen
two different rationales for . If we accept a truth-centred approach
to epistemic matters and either accept   or , there is an
argument to be made that rational thinkers ought to conform to 
because this represents the right way of handling certain risks (e.g. the risk of
²¹ See Smith (2016) and Weatherson (2016) for defence of the idea that non-sceptical accounts of
knowledge have to allow that some of the things we know have lower evidential probability than some
true propositions that we’re not in a position to know. Their defence uses lottery cases.
²² To see this, break down Agnes’s evidence into two components: that the file she thought to be
Inge’s isn’t, and that the file she thought to be Agatha’s is Inge’s. The first is a pure undercutting
defeater of Agnes’s initial reason to believe in Inge’s guilt (namely, the contents of the file she
mistakenly thought was hers). The second is a confirmer of Inge’s guilt, but not as strong a confirmer
as the previous was. Together, they are a mild undercutting defeater, on the de-confirmation account.
     235
believing in ways that are objectively bad or failing to believe in ways that are
objectively good or that of violating some objective norm). If  were
correct, we could easily see how to give a monistic account of defeat and
explain why some familiar kinds of defeaters are rationally toxic without
having to add a clause to our theory of rational belief to handle defeat. That
would mean that - satisfies our second and third desiderata,
but doubt that it satisfies the first, extensional adequacy. Adding a clause to
handle tricky cases (e.g. higher-order defeat, puzzling intuitions about under-
mining defeat in lottery cases) might help with extensional adequacy, but then
the modified view would no longer satisfy our second and third desiderata.
Many epistemologists have raised concerns about  along the
lines of those -previously discussed. Smith (2016), for example, thinks that the
problem has to do with this idea that justification is a matter of managing risk
and so he rejects the risk-minimization approaches to justification (of which
 is an instance) in favour of an alternative view designed to allow
that we might be justified in believing in one case and unjustified in believing
in another even if the risk of forming a false belief is greater in the second kind
of case than the first. The view that we shall propose has a similar feature, but
we don’t want to be too hasty in rejecting the idea that much of rationality is a
matter of properly handling risk.
Before we reject the very idea that there is some connection between
rational belief and the management of risk, we should remember that there
are many different kinds of risk that we might need to manage. If a truth-
centred approach were correct, we should worry about forming false beliefs
and failing to form true ones. If we were veritists, we would describe this as a
concern about managing the risks of forming bad beliefs and failing to form
good ones. If we believed in epistemic norms like truth norm, we would
describe this a concern about managing the risk of norm violation. It seems
strange to us to suggest that rationality shouldn’t be concerned with the
management of these kinds of risk. On its face, the idea that the rational
person’s responses are rational precisely because they manage these kinds of
risks well seems pretty plausible. Perhaps the lesson to draw from the diffi-
culties that arise for  is not that we were mistaken to think that
rationality has a great deal to do with the management of risk but that we were
mistaken about the kinds of risks that need to be managed.
We’re not all veritists. And we don’t all believe that the fundamental
epistemic norms have to do with truth. Some of us prefer gnosticism to
veritism and prefer to think of epistemic norms as concerned with knowledge
and not merely with truth:
236     
: the fundamental epistemic good is knowledge and the funda-

mental epistemic evil is a belief that fails to be knowledge.²³
 : we ought to believe p if we can believe it and therein
come to know. We ought not to believe p if we would not believe it and
therein come to know.
If we take either of these proposals as our starting point, the same arguments
that took us from veritism or truth norms to threshold take us from these
knowledge-centred claims about knowledge’s normative role to this account of
rational belief:
 : it is rational for S to believe q iff the probability that

S is in a position to know q is sufficiently high.²⁴
We can see   as representing a standard that, like

, tells us how to properly manage the risk of failing to promote
the epistemic good or of violating certain objective epistemic norms.²⁵
As with , we can easily extract an account of defeat from
 .²⁶ As with de-confirmation, this account should sat-
isfy our second and third desiderata. On the view that we propose, something
constitutes a partial defeater insofar as it suggests that there is some risk of
believing without knowing and a full defeater if it ensures that the risk is too
great for us to believe given either a proper concern for the epistemic good or
given what it would take to responsibly try to meet our objective epistemic
²³ See Hyman (1999); Littlejohn (2018); Williamson (2000); Unger (1975) on the value of knowledge
for motivations for .
²⁴ See Dutant and Fitelson (MS); and Littlejohn (2020) for discussion and defence of the idea that
rationality is understood in terms of the probability of knowledge. They are primarily concerned with
synchronic wide-scope requirements. We said earlier that we think of probability here as evidential
probability. Some might worry that if we combine this approach with the idea that everything known is
part of our evidence we will narrow the scope of the applicability of the account. (For example, if E=K,
our account says that in cases where you know that you know p, the probability that you know p is 1.
And this would mean that we couldn’t give an account of how belief in p could be defeated.) We aren’t
certain that the narrow scope is a problem and we are not taking any official stand on the relationship
between evidence and knowledge. In particular, we think that there are interesting independent reasons
to look for an account of evidential probability on which it’s possible to know p even if the probability
of p is less than 1. See Bacon (2014) and Brown (2018) for discussion of such reasons.
²⁵ In this way, we think we can follow Ghijsen, Kelp, and Simion (2016) in giving an account of the
subjective and objective dimensions of epistemic assessment without rejecting the idea that the
knowledge norm governs belief. Graham (2010) proposes a similar approach to reconciling an
objectivist approach to obligation with a more subjectivist account of rationality.
²⁶ We think a probabilistic approach to rationality that links rationality to the risk of norm violation
will help objectivists address some of Brown’s (2018) objections to Littlejohn and Williamson’s
proposals about excuses.
     237
obligations. Smith might have been right that rational belief isn’t just a matter
of believing in ways minimize the risk of error, but there are alternative risk
minimization approaches that we should consider. We have just offered one.
10.4. Defeat as an Indication of Ignorance
If it is rational to believe in ways that maximize expected epistemic value, the

view that we get if we assume  is  . If our
monistic theory of defeat ought to be embedded in some larger normative
framework, then we propose embedding it in a framework that incorporates
  as the standard for rational belief, a standard that tells
us how rational thinkers go about trying to deal with their uncertainty about
what it takes to meet the fundamental standards,  . If
rational belief is what   says, we can characterize defeat
as follows:
  : defeaters are indicators of ignorance (i.e. con-

siderations that constitute evidence that the thinker is not in a position to
know the target proposition).²⁷
   can handle opposing defeat, undercutting defeat,

reason-defeating defeat, normative defeat, collective defeat, and pragmatic
defeat in roughly the same way that - did. Being in a

position to know is factive, but knowledge requires more than truth. For this
reason, it will typically be that P(Kp)<P(p). Things that de-confirm will often
serve as an indicator of ignorance, so our account can appropriate much of
what - can offer.
One interesting difference between - and  
 emerges when we think about lottery-type cases. Recall Agatha,
Agnes, and Inge. Once Agnes could tell which belief was based on statistical
evidence and which belief was based on observation, we saw that the prob-
ability for one belief dropped and the belief that wasn’t rational to hold was
²⁷ We hadn’t realized that Gibbons (2013) had proposed that defeaters were evidence that we aren’t
in a position to know. Thanks to Errol Lord for pointing this out. One difference between his approach
and ours is that we introduced a rational standard that should help us decide whether some defeater is
full or merely partial. We also disagree about lottery cases, so he wouldn’t accept our points about
lotteries and undercutting defeat. Although we put Di Paolo’s (2018) work to use earlier in defence of
-, he said that he was sympathetic to a knowledge-centred view.
238     
rational to hold.   and   

explain how this is possible: when the probability of knowing increases even
though the probability of truth decreases, a belief can become rational despite
becoming less likely to be true.
   provides a more straightforward account of
higher-order defeat. It’s hard to see how indications that the oxygen is low,
that the drinks have been drugged, etc. constitute de-confirmers, but it’s not at
all hard to see how they could constitute evidence that the subject is not in a
position to know. Since there seems to be little interesting connection between,
say, something having maximal evidential probability and being something
that the thinker is in a position to know, we don’t have to try to explain
how a thinker who has entailing evidence at one moment can fail to be
rational in her beliefs at a later moment when she is made aware of some
higher-order defeater.
   sheds new light on collective defeat. Whereas
the probability that we know in a lottery case is 0, the probability that we know
that some entry in a book known to contain an error might be quite high.
Optimistically, we think that in a book that is suitably large, if we knew that the
book contained a single error, we could know each entry of the book to be true
save the one that was false. In such a case, it might be that the probability that
each claim in the book is known is quite high (i.e. if the book contains n claims
and n is a suitably large number, the probability that any particular claim in
the book can be known might be 1/n-1). Thus, we get a view on which it would
be possible to believe each claim in the book whilst knowing that one of the
claims in the book is false. Our view differs from truth-centred views in that
they tend to treat lottery and preface cases in the same way. In our view, such
views deliver the wrong result in lottery cases or in preface cases. The
difference in the prospect of coming to know in such cases is the key to
understanding why we’re tempted to think of them as differing in a way that
should matter for our theory of rationality.
   also provides a better treatment of negative self-
appraisal defeat. First, if the conjunction of some first-order belief and the
negative self-appraisal is not knowable (e.g. because the negative self-appraisal
represents the first-order belief as failing to fulfil some necessary condition on
knowledge),   says that the conjunction cannot be
rationally believed because the probability that the conjunction is known is 0.
What about the conjuncts, though? It’s one thing to say that nobody can
rationally believe certain conjunctions that we know cannot be known (e.g.
‘Dogs bark but I don’t know that they do’) and another to say that nobody can
     239
rationally believe each conjunct. Consider some problematic pairs of

propositions:
1. p, I couldn’t know whether p.

2. p, It is irrational for me to believe p.
3. p, My evidence isn’t sufficient to believe p.
Can a thinker be rational in believing both propositions in (1)–(3)?

   tells us that it can only be ex ante rational to
believe both if it can be rational to be sufficiently confident of each proposition
that it is something we can know.
Let’s start with (1). While   does not tell us how
probable it must be that p is something you’re in a position to know for it to be
rational to believe p, it’s natural to suppose that it has to be at least more
probable than not (i.e. if it’s ex ante rational to believe p, P(Kp) > .5). If so, it
cannot also be rational to believe that you cannot know whether p. If it were
rational to believe ~Kp, P(K~Kp) >.5. Since being in a position to know is
factive, P(K~Kp) > .5 only if P(~Kp) > .5. But it cannot be that P(~Kp) >.5 if P
(Kp) >.5. Thus,   tells us that it cannot be ex ante
rational for a thinker to believe both propositions in (1). If we understand
‘having sufficient evidence’ as meeting a condition necessary for knowing,
then this reasoning rules out evidence that makes it ex ante rational to believe
the propositions mentioned in (3).
The difficult case is (2). If ex ante rationality were necessary for being in a
position to know, our treatment of (1) would cover this case.²⁸ Unfortunately
for us, we cannot say that rationality is necessary for knowledge. If we assume
probable knowledge, we should expect that there will be some ‘unreasonable
knowledge’ in the sense that there will be some things that we know that will
not be sufficiently likely to be known to also be rational to believe.²⁹ Still, while
we might want to allow for the possibility of unreasonable knowledge, we
might want to deny that it can be rational to both believe p and believe that this
belief is irrational.
²⁸ Let t be the threshold in  . Suppose knowledge entails rational belief: by
 , P(Kp) ≤ t entails ~Kp. Now suppose it is rational to believe that it is not
rational to believe p. By  , P(P(Kp) ≤ t) > t. Since P(Kp) ≤ t entails ~Kp, P(~Kp) >
t. Assuming t ≥ ½, P(Kp) < t.
²⁹ See Lasonen-Aarnio (2010) for discussion and defence of unreasonable knowledge and Sylvan
(2018b) for defence of the idea that knowledge is a non-normative state that needn’t have a normative
dimension. We think that a non-normative account of knowledge can help address some concerns
about the very idea of defeat raised by Baker-Hytch and Benton (2015).
240     
Whether   prohibits a pair of beliefs like (2) depends

upon whether a high probability of knowledge is compatible with a high
probability (that one knows) that knowledge is improbable. If we take t > ½
to be the threshold of sufficient probability in  , a
sufficient condition to prohibit the pair is:
 t-: if P(Kp)> t then P(P(Kp)>t)≥ 1-t.
For note that P(Kp)> t obtains just if P(Kp)≤ t doesn’t. Hence P(P(Kp)> t)≥ 1-t
is equivalent to P(P(Kp) ≤ t) ≤ t. Given the factivity of being in a position
to know, P(P(Kp)≤ t)≤ t only if P(KP(Kp) ≤ t) ≤ t. Hence given 
t-, if P(Kp)>t then P(KP(Kp)≤ t)≤ t. This would mean that if the
probability that one knows p is above threshold, the probability that one
knows that it is not above threshold is itself below threshold, as desired.
Note that while  t- is an inter-level constraint, it does not
force us to accept strong access constraints like the idea that it is rational to
believe p only if it is rational to believe that one knows that p. It still allows the
combination of P(Kp)>t and P(P(Kp)>t)≤ t and the combination of Kp and P
(Kp)≤ ½.³⁰ Note also that the higher t is, the milder the  t-
requirement is. At t=.9, for instance, it merely requires that if it is highly (.9)
probable that one knows, it is at least somewhat (i.e. .1) likely that it is highly
probable.
While we don’t have a full defence of  t-, we think that it
has some prima facie appeal and is worth exploring. Its initial appeal comes
from the idea that whatever makes it probable that one knows will typically
also make it probable that it is probable that one knows. What confirms to you
that you know p by memory or perception, for instance, is typically the fact
that your perception or memory is particularly clear or detailed, the fact that
you can give additional details pertaining to p, and so on. But those facts are
also evidence that you have evidence that you know, which is to say that they
contribute to make it probable that it is probable for you that you know.
The thought gets further support from an unlikely source: Williamson’s
(2014) unmarked clock models. While these models are meant to illustrate the
failure of various access constraints on knowledge, they turn out to validate
 t-. We can illustrate how this works with an example and we
provide a proof in the Appendix. In Williamson’s models, a parameter (e.g.
time, temperature, etc.) takes different values along a line. Worlds correspond
³⁰ But, see Goodman (2013) for helpful discussion of such constraints.
     241
to points on the line. What you know at a world is whatever holds at worlds
within a margin of error m to the left and to the right of the point on the line.
Your evidential probability is given by conditionalizing a uniform prior on
what you know. When you update in this way, every value outside of the
margins is ruled out and each of the values within the margins are equally
likely. Suppose the actual world (@) is at 50 and the margin for error (m) is 50.
What you know leaves open any world within [0:100]. Let p be some propos-
ition such that Kp holds throughout [30:90]. (For instance, p might be the
proposition that the value of the parameter is between -20 and 140.) Because
Kp occupies 60 percent of the worlds left open by what you know at @, the
probability of Kp at @ is.6. Note that any world between 40 and 80 will also
have within its margin the Kp-segment [30:90]: for 40 leaves open any world in
[-10:90] and 80 leaves open any world in [30:130]. So, at each of these worlds,
P(Kp)≥.6, too. This means that at least 40 per cent of worlds within @’s
margins are Kp-worlds. So, at @, we have an instance of  0.6-
, P(P(Kp)≥.6)≥.4. The reasoning generalizes to any Kp-proposition
that occupies a continuous stretch of 50 per cent of @’s margins.³¹
@±m
0 30 50 90 100
p
Kp
40±m
80±m
P(Kp)≥6
Informally, here is what we have shown. In these models, the features of a

situation that make Kp probable (i.e. that Kp holds in a large proportion of the
worlds left open by one’s evidence) are also features that put an upper limit on
how improbable it can be that Kp is probable (i.e. a sufficient proportion of
the worlds that are left open by one’s evidence are worlds that can see mostly
Kp-worlds). This is why they validate  -. This result depends
upon several features of the models that we’ve worked with, so it remains to be
³¹ Note that propositions of the form Kp have to occupy continuous stretches within a world’s
margin. For if Kp holds at two points within @’s margins, those two points are at most 2m apart, which
means that their own margins connect and p holds throughout the segment between them, which in
turns means that Kp holds at any point between them.
What happens for t below ½? Suppose for instance that Kp holds throughout [80:100]. Then it
occupies 20 per cent of @’s margins, so at @ we have P(Kp) ≥.2 . Now all we can say is that every world
to the right of @ ‘sees’ the whole of [80:100]. So, we have P(P(Kp) ≥.2) ≥.5. More generally, the models
validate an alternative schema: for any t≤½, if P(Kp)> t, (P(Kp)> t)≥.5. This is good news as instances of
 t- with low t are counterintuitive.
242     
seen how principles like  - would fare if we worked with
different models (e.g. ones that do not assume a uniform prior, ones that do
not have a constant margin, etc.). At the very least, we think this gives us some
reason to explore principles like  - further and reason to
hope that such principles will hold in a wide range of cases.
If it is impossible to rationally believe the propositions mentioned in (1)–
(3), our account handles cases of negative self-appraisal defeat quite nicely.
When it is rational for the thinker to believe, say, that she’s not in a position to
know or that it would be irrational for her to believe p, it would not be rational
for her to believe p and we can see why these negative self-appraisals can
function like defeaters. Alternative views that characterize rational belief in
terms of the probability of truth struggle to explain why it would be rationally
mandatory for a thinker who rationally believes that she doesn’t know, say, to
suspend judgment.
If our treatment of negative self-appraisal is correct, we can build on an
interesting proposal about suspension defended by Thomas Raleigh
(Forthcoming). He proposed that suspending is really a matter of believing.
We suspend, on his view, when we believe the evidence doesn’t support a
belief. We think that something very similar might be right. Provided that a
thinker is rational, her rational belief that she’s not in a position to know
whether p should ensure that she’s agnostic about whether p. (Even without a
belief, a sufficiently high degree of confidence should also work.) It is difficult
to explain why this should be in a truth-centric framework since, as we’ve seen,
it’s not clear why, if  is correct, a rational thinker would suspend if
she were to judge that she’s not in a position to know or that her evidence
doesn’t support her belief. Our framework, however, explains why a rational
thinker suspends when she rationally believes that she’s not in a position to
know. We think that it’s a strength of our view that it helps us see what’s
attractive about Raleigh’s proposals about belief and suspension.
10.5. Conclusion
We have proposed a theory of rational full belief according to which it’s rational
to believe when it’s sufficiently probable that the belief would be knowledge.
From this account, we can derive a theory of defeat according to which defeaters
are indicators of ignorance, evidence that if you were to believe, your belief would
fail to constitute knowledge. We think that this account provides a better account
of every kind of (genuine) defeat there is than the leading alternative. If we think
     243
of rationality as something akin to subjective rightness and think of subjective

rightness as being probabilistically related to objective norms, our account of
rational belief and defeat is the account we should want if we think of knowledge
as the fundamental normative standard for belief.³²
Appendix
We show that  t- (for t≥.5) holds in Williamson’s (2014) unmarked clock
models. Worlds are points in the real number line, propositions are regions of the line;|p|is
the size of the region p; for every p, q, the proposition p&q is the intersection of p and q. At
each world w one knows every proposition entailed by the parameter’s value being within a
margin m>0 of w. We let k(w) = [w-m:w+m] and for any p, the proposition that one knows
p, Kp, is the set {w:k(w) ⊆ p}. The evidential probability at w, πw, is a uniform prior
conditionalized on k(w): πw(p)=|p&k(w)|/|k(w)|. Without loss of symmetry we set m=½
which gives us|k(w)|=1 and the simple relation: πw(p)=|p|. The proposition that the
evidential probability of p is above t, P(p)>t, is the set {w: πw (p)>t}. Now let t≥½, @
some world and p such that π@(Kp)>t.
Our aim is to show that π@(P(Kp)>t)≥1-t. Let a be greatest lower bound and b the least
upper bound of Kp&k(@). Since Kp holds at worlds (arbitrary close to) a and b, p holds
throughout the regions]a-m:a+m] and [b-m:b+m[. Since a and b are within k(@),|b-a|<2m
and the regions overlap and p holds throughout]a-m:b+m[. So Kp&k(@) holds throughout]
a:b[, and since π@(Kp)>t,|]a:b[|>t. Now let w be any world within]a+(t-½):b-(t-½)[. (This
interval has positive size since |]a:b[|>t≥½.) Since w>a+(t-½), w-m=w-½>a+t-1, which
given t≤1 entails w-m>a, and by parallel reasoning, w+m<b. Hence ]a:b[ ⊆ k(w). Since]a:b
[ ⊆ Kp and |]a:b[|> t, πw(Kp)>t. So P(Kp)>t holds at any world w within ]a+(t-½):b-(t-½)[.
Now since |]a:b[|>t≥½,|]a+(t-½):b-(t-½)[|>t-2(t-½)=1-t. Therefore π@(P(Kp)>t)>1-t.
References
Adler, Jonathan. (2002). Belief ’s Own Ethics. Cambridge, MA: MIT University Press.
Anscombe, G. E. M. (1962). On sensations of position. Analysis, 22: 55–8.
Bacon, A. (2014). Giving your knowledge half a chance. Philosophical Studies, 171:
373–97.
Baker-Hytch, Max and Benton, Matthew. (2015). Defeatism defeated.
Basu, Rima and Schroeder, Mark. (2019). Doxastic wronging. In: B. Kim and
M. McGrath (eds.) Pragmatic Encroachment in Epistemology. New York:
Routledge: 158–78.
³² For feedback and discussion, we would like to thank Maria Lasonen-Aarnio, Josh Di Paolo,
Branden Fitelson, Errol Lord, Mona Simion, and particularly Kevin Dorst. Kevin was kind enough to
provide proofs that helped us think through the issues in Section 10.4.
244     
Bergmann, Michael. (2006). Justification Without Awareness. Oxford: Oxford

University Press.
Brown, Jessica. (2018). Fallibilism: Evidence and Knowledge. Oxford: Oxford
University Press.
Chandler, Jake. (2013). Defeat reconsidered. Analysis, 73: 49–51.
Christensen, David. (2004). Putting Logic in its Place. Oxford: Oxford University
Press.
Di Paolo, Joshua. (2018). Higher-order defeat is object-independent. Pacific
Dorst, Kevin. (2019). Lockeans maximize expected accuracy. Mind, 128: 175–211.
Dorst, Kevin. (2020). Evidence: A guide for the uncertain. Philosophy and
Dorst, Kevin, Fitelson, Branden, and Husic, Brooke. Manuscript. Evidence of
evidence: A higher-order perspective.
Dutant, Julien and Fitelson, Branden. Manuscript. Knowledge-centered epistemic
utility theory.
Dutant, Julien and Littlejohn, Clayton. (2020). Justification, knowledge, and
normality. Philosophical Studies, 177: 1593–1609.
Easwaran, Kenny. (2016). Dr. Truthlove or: How I learned to stop worrying and
love Bayesian probabilities. Noûs, 50: 816–53.
Easwaran, Kenny and Fitelson, Branden. (2015). Accuracy, coherence, and evi-
dence. In: Tamar Szabo Gendler and John Hawthorne (eds.) Oxford Studies in
Epistemology. Vol. 5. Oxford: Oxford University Press: 61–96.
Fantl, Jeremy and McGrath, Matthew. (2009). Knowledge in an Uncertain World.

Ghijsen, Harmen, Kelp, Christopher, and Simion, Mona. (2016). Norms of belief.
Philosophical Issues, 26: 375–92.
Gibbons, John. (2013). The Norm of Belief. Oxford: Oxford University Press.
Goldberg, Sandy. (2017). Should have known. Synthese, 194: 2863–94.
Goldman, Alvin. (1999). Knowledge in a Social World. Oxford: Oxford University
Press.
Goodman, Jeremy. (2013). Inexact knowledge without improbable knowing.
Inquiry, 56: 30–53.
Graham, Peter A. (2010). In defense of objectivism about moral obligation. Ethics,
121: 88–115.
Grundmann, Thomas. (2011). Defeasibility theory. In: S. Bernecker and
D. Pritchard (eds.) The Routledge Companion to Epistemology. New York:
Routledge: 156–66.
     245
Harman, Gilbert. (1968). Knowledge, inference, and explanation. American

Hyman, John. (1999). How knowledge works. Philosophical Quarterly, 49: 433–51.
Kotzen, Matthew. (2019). A formal account of epistemic defeat. In: Fitelson, B.,
Borges, R., and Braden, C. (eds.) Themes from Klein. Cham, Switzerland:
Springer: 213–34.
Lackey, Jennifer. (2006). Learning from words. Philosophy and Phenomenological
Research, 73: 77–101.
Lasonen-Aarnio, Maria. (2010). Unreasonable knowledge. Philosophical
Lazar, Seth. (2019). Deontological decision theory and the grounds of subjective
permissibility. In: Mark Timmons (ed.) Oxford Studies in Normative Ethics.
Vol. 9. Oxford: Oxford University Press: 204–22.
Leitgeb, Hannes. (2017). The Stability of Belief: How Rational Belief Coheres with
Probability. Oxford: Oxford University Press.
Littlejohn, Clayton. (2017). How and why knowledge is first. In: A. Carter,
E. Gordon, and B. Jarvis (eds.) Knowledge First. Oxford: Oxford University
Press: 19–46.
Littlejohn, Clayton. (2018). The right in the good: A defence of teleological non-
consequentialism. In: K. Ahlstrom-Vij and J. Dunn (eds.) Epistemic
Consequentialism. Oxford: Oxford University Press: 23–48.
Littlejohn, Clayton. (2020). Truth, knowledge, and the standard of proof in
criminal law. Synthese, 197, 5253–5286.
Littlejohn, Clayton. (Forthcoming). On what we should believe (and when (and

why) we should believe what we know we should not believe). In: K. McCain
and S. Stapleford (eds.) Epistemic Duties. New York: Routledge.
Lord, Errol. (2018). epistemic reasons, evidence, and defeaters. In: D. Star (ed.)
Oxford Handbook of Reasons and Normativity. Oxford: Oxford University
Press: 600–31.
Lynch, Michael. (2004). True to Life. Cambridge, MA: MIT University Press.
Makinson, D. C. (1965). The paradox of the preface. Analysis, 25: 205–7.
McGinn, Marie. (2012). Non-inferential knowledge. Proceedings of the
Aristotelian Society, 112: 1–28.
Moss, Sarah. (2018). Moral encroachment. Proceedings of the Aristotelian Society,
118: 177–205.
Nelkin, Dana. (2000). The lottery paradox, knowledge, and rationality. The
Olsen, Kristian. (2018). Subjective rightness and minimizing expected objective
wrongdoing. Pacific Philosophical Quarterly, 99: 417–41.
246     
Owens, David. (2000). Reason Without Freedom. London: Routledge.

Peacocke, Christopher. (2004). The Realm of Reason. Oxford: Oxford University Press.
Pollock, John. (1986). The paradox of the preface. Philosophy of Science, 53: 246–58.
Pollock, John and Cruz, Joseph. 1999. Contemporary Theories of Knowledge.
Lanham, MD: Rowman and Littlefield.
Raleigh, Thomas. (Forthcoming). Suspending is believing. Synthese.
Ryan, Sharon (1991). The preface paradox. Philosophical Studies, 64: 293–307.
Ryan, Sharon. (1996). The epistemic virtues of consistency. Synthese, 109: 121–41.
Schroeder, Mark. (2012). Stakes, withholding, and pragmatic encroachment on
knowledge. Philosophical Studies, 160: 265–85.
Silva, Paul Jr. (2017). How doxastic justification helps us solve the puzzle of
misleading higher-order evidence. Pacific Philosophical Quarterly, 98: 308–28.
Smith, Martin. (2016). Between Probability and Certainty. Oxford: Oxford
University Press.
Spohn, Wolfgang. (2002). A brief comparison of Pollock’s defeasible reasoning
and ranking functions. Synthese, 13: 39–56.
Sturgeon, Scott. (2008). Reason and the grain of belief. Nous, 42: 139–65.
Sylvan, Kurt. (2018). Veritism unswamped. Mind, 127: 381–435.
Sylvan, Kurt. (2018b). Knowledge as a non-normative relation. Philosophy and
Unger, Peter. (1975). Ignorance: A Case for Scepticism. Oxford: Oxford University
Press.
Weatherson, Brian. (2005). Can we do without pragmatic encroachment?

Weatherson, Brian. (2016). Reply to Eaton and Pickavance. Philosophical Studies,
173: 3231–3.
van Wietmarschen, Han. (2013). Peer disagreement, evidence, well-groundedness.
Williams, Bernard. (1965). Ethical consistency. Proceedings of the Aristotelian
Society, Supplementary, 39: 103–38.
Williamson, Timothy. (2000). Knowledge and its Limits. Oxford: Oxford
University Press.
Williamson, Timothy. (2014). Very improbable knowing. Erkenntnis, 79: 971–99.
Williamson, Timothy. (2019). Evidence of evidence in epistemic logic. In:
M. Skipper and A. Steglich-Petersen (eds.) Higher-Order Evidence: New
Essays. Oxford: Oxford University Press: 265–98.
Worsnip, Alex. (2016). Belief, credence, and the preface paradox. Australasian
Journal of Philosophy, 94: 549–62.
11
Competing Reasons
Justin Snedegar
11.1. Competition Between Reasons
One of the most important facts about the normative domain is that some
considerations are contributory, rather than decisive, when it comes to deter-
mining what we ought to, must, or may do.¹ They provide defeasible support
or defeasible opposition to an option. In epistemology, the theory of practical
reasoning, and ethics, it is standard to appeal to contributory or pro tanto
normative reasons to capture this fact. Most of our options have some reasons
in favor and some reasons against; to determine which option to take, we have
to determine how these reasons compare. In particular, on the standard view
reasons compete with one another, and the outcome of this competition
determines what you ought to, must, or may do.² That action A would be
mildly pleasant is a reason to do A, and that an incompatible action B would
prevent a great deal of suffering is a reason to do B. These reasons compete
with one another, since A and B are incompatible. If there are no other
relevant reasons, then what you ought to, must, or may do—A or B—will
depend on the outcome of this competition.
An important task for normative theory, then, is to explain how the compe-
tition between reasons works. A large part of this task will involve investigating
relationships of defeat between reasons, since when reasons compete, at least in
many cases, some reasons will defeat others—that is how they win the compe-
tition. Though the concept of defeat is perhaps more familiar in epistemology,
I will be concerned mostly with the practical domain.
¹ See Ross (1930) for an early and very influential development of this idea; for a recent overview, see
Lord and Maguire (2016).
² I’m using this slightly long-winded phrase, ‘what you ought to, must, or may do,’ because the
differences between these different normative statuses will be important later. To note the difference
between ‘ought’ and ‘must’ here, consider the following, which may be said about some supererogatory
action: “Well, no, I’m not saying you must do it, but you really ought to.”
05.003.0011
248  
Following Pollock (1974), epistemologists have focused primarily on two

distinct kinds of defeat: rebutting defeat and undercutting defeat. One reason
rebuts another when it is a stronger reason for a conflicting option. In the
previous case, for example, it is plausible that the reason for B, that it would
prevent a lot of suffering, rebuts the reason for the conflicting option A, that it
would be mildly pleasant. Importantly, in cases of (mere) rebutting, the
defeated reason is in an important sense still “in force.” It isn’t that there is
no reason for A in this case; after all, it would be mildly pleasant. It’s just that
this reason is (rebuttingly) defeated, or outweighed. If these are the only
reasons in play, then it’s very plausible at least that you ought to do B, and
perhaps that you must or are required to do B, in which case it’s not true that
you may do A, or that A is permissible.
One reason undercuts another, on the other hand, when the first does
undermine the second’s status as a reason—to use a common metaphor, it
attacks the connection between the consideration and the option it would
otherwise support. That you promised to buy me lunch is, or would be, a
reason to do so. But that the promise was coerced or given under duress
plausibly undercuts this reason, so that in fact it is no reason at all to buy me
lunch. So, the first important difference between cases of rebutting and cases of
undercutting is that in the former, but not the latter, the defeated reason is still
in force.³ The second important difference is that in cases of undercutting,
unlike in cases of rebutting, the defeating reason need not, and typically will
not, be a reason for a conflicting option. Note that the fact that your promise
was coerced need not be a reason against buying me lunch, or a reason to do

anything incompatible with buying me lunch, even though it does defeat—in
an undercutting way—the promissory reason you would otherwise have to
buy me lunch.
A third way in which reasons compete, at least in the practical domain, is
captured in Raz’s (1999) notion of an exclusionary reason. There is some
debate about the best way to understand exclusionary reasons, but on a natural
understanding, one reason excludes another when the first is a reason to take
the second out of consideration. To borrow Raz’s example, suppose I promise
my partner that my decision about where to send our daughter to school will
be guided entirely by the quality of the education she would get. In that case,
that Fancy Private School is more expensive than Good Public School just isn’t
³ In fact, this is too simple. There are cases of partial undercutting in which a reason is not fully
undercut, but just weakened. So, a bit more precisely: when a reason is merely rebutted, it retains its full
strength, but when it is undercut, it loses at least some—and perhaps all—of its strength. See Schroeder
(2007): Ch. 5; compare with Dancy (2004): Ch. 3, on attenuators.
  249
something I should take into account in deciding where to send her. It may
seem that my promise is an undercutting defeater of the financial reason to
send our daughter to Good Public School. But in fact there is an important
difference between undercutters, as in the coerced promise previous case, and
exclusionary reasons.⁴ In the undercutting case, the initial reason is no longer
in force at all—it is just not a reason anymore. But, in the exclusionary case,
the reason is in force, it’s just that I have a reason not to take it into account in
my deliberation.⁵ In the undercutting case, there is no reason to be taken into
account. Nevertheless, even though the reason in the exclusionary case is in
force, it will lose the competition—in fact, since it is taken out of consider-
ation, it doesn’t even get to compete.⁶
So, we have seen three different ways in which a reason may be defeated, or
in which reasons can compete. The reason may be rebutted, undercut, or
excluded. Arguably, the most straightforward of these is a rebutting defeat, in
which a reason is defeated by a stronger reason for a conflicting option. One
goal for this paper is to suggest that rebutting defeat is actually not as
straightforward as it seems. In particular, I will argue that we should recognize
two different varieties of rebutting defeat. First, there are cases in which a
reason for one option is defeated by a stronger reason for an incompatible
option. Second, there are cases in which a reason for one option is defeated by
a stronger reason against that same option (or vice versa). Many philosophers
have not recognized this distinction between two kinds of rebutting defeat,
I believe, because they have not recognized a distinction between reasons for
and reasons against. It is a common, though usually implicit, assumption that

reasons against an option are just reasons for incompatible alternatives. I argue
in Snedegar (2018) that this is a mistake. In this paper, I explore the conse-
quences of recognizing the distinction between reasons for and reasons against
for accounts of how reasons compete, focusing on rebutting defeat. The overall
goal is to make some progress in understanding how the competition between
pro tanto reasons determines the overall status of our options.
⁴ See Horty (2012) and Horty and Nair (2018) for the claim that exclusionary reasons are just
undercutters. See Raz (1999): 184; and Whiting (2017): 400, for a rejection of this identification.
⁵ Whiting (2017) makes essentially the same point, but he understands exclusionary reasons as
second-order reasons not to base one’s actions or attitudes (including decisions) on certain ordinary,
first-order reasons. Raz (1999), especially in the postscript, often seems to understand them in this way
as well.
⁶ I am glossing over some complications here. For example, there may be cases in which the
excluded reason is so weighty that it should not be excluded. A full theory of the competition between
reasons, and in particular of how exclusionary reasons function, will need to explain this.
250  
First, I’ll introduce what I take to be the most common way of thinking
about the competition between reasons, which I call balance accounts. Then,
I’ll argue this picture—as so far developed—is inadequate, primarily because it
does not recognize the importance of reasons against. I then consider a very
different way of thinking about the competition between reasons, which I call
criticism-based accounts. This kind of view does recognize the importance of
reasons against but misses out on the importance of the competition between
reasons bearing on competing options. Finally, I sketch a positive view that
incorporates elements of both kinds of view. Open questions remain, but
I argue that it gives us the beginnings of an attractive account of how the
competition between reasons determines the overall status of our options.
11.2. Balance Accounts
According to balance accounts, reasons compete by weighing against each

other in a way that’s broadly analogous to how physical objects on a scale
compete to let us see which is the heaviest.⁷ In fact, discussions of competing
reasons often use a scale as a metaphor.⁸ The reasons are like differently
weighted marbles that go on the pans of a scale, which correspond to the
conflicting options. The pan with the most weight in it once all the reasons
have been taken into account corresponds to the option you are required to
perform. Asking what there is ‘most reason’ to do is naturally understood
as asking about the total amount of support provided by the individual

reasons for each of the options, which involves the accrual of these reasons.⁹
This is probably the most common picture of the competition between
reasons; it is suggested by the platitudinous claim that you are required to
do what you have most reason to do.¹⁰ This gives us the most natural version of
the balance account:¹¹
⁷ A different metaphor comes from Ross (1930), who thinks of the competition between pro tanto
considerations as analogous to the competition between different forces acting on an object. I take the
two physical metaphors to suggest the very similar conceptions of the competition between reasons.
⁸ See, e.g., Broome (2004). ⁹ On accrual, see Nair (2016).
¹⁰ Often authors use ‘ought’ instead of ‘required’ here. I think that most of them have in mind what
I’ll mean by ‘required,’ including that you’d be acting impermissibly by not doing what you “ought” (in
their sense) to do. In any case, this won’t matter much for my discussion of the Balance Account,
though the difference between these overall normative statuses will become important later.
¹¹ Compare Parfit (2011): 32; Schroeder (2015); and Lord (2018): 10–11.
  251
Balance: you are required to A when the reasons for A outweigh the reasons
for any of the alternatives.
The intuitive thought here is that if you have more reason for A than for
anything else, then you are required to A. On this view, we imagine the
deliberating agent facing a range of alternatives. Each alternative will have
some reasons in its favor. We weigh these up and find out where the balance of
reasons lies; that is the option the agent is required to take. This picture relies
heavily on rebutting defeat, of course: the reasons for incompatible options
compete, and the weightier ones rebut the less weighty ones.¹²
This is a very natural picture, but as so far described, it faces problems.
Suppose that pressing either button 1 or button 2 will guarantee that I get a
particular $100 bill. But while pressing button 1 has no cost, pressing button 2
will deliver a painful electrical shock. Assuming these are the only relevant
factors in this case, it is clear that I am required to press button 1 and that I am
required not to press button 2. But note that in this case, my reason for
pressing button 1 is the same as my reason for pressing button 2—doing so
will get me $100. So, I do not have more reason for pressing button 1 than for
pressing button 2, and so, according to Balance, it isn’t true that I am required
to press button 1. What we need, clearly, is to add reasons against: there is a
reason against pressing button 2—the electrical shock—that is not a reason
against pressing button 1, and this explains why pressing 1 is required and why
pressing 2 is impermissible.
This problem may seem fairly obvious, and so you may worry that I’ve been
unfair to defenders of balance accounts. The reason this problem hasn’t been
apparent, I believe, is that defenders have (in most cases, implicitly) made the
assumption mentioned in the introduction, that reasons against an option are
just reasons for alternatives to that option. If this were true, then Balance
would take reasons against into account, since reasons against just are reasons
for alternatives. The obvious thought in this case is that if we understand the
fact that pressing button 2 would give you a shock as a reason for the
alternative to pressing 2—that is, pressing 1—then we get an extra reason
for pressing 1, and so that is what you are required to do.
An alternative, defended by Schroeder (2007), can also avoid this problem.
This is because Schroeder understands the relevant kind of competition to
be the competition between reasons for an option and reasons against that
¹² The picture will be more sophisticated than this, to accommodate undercutting defeat, as well as
what Dancy (2004) calls enablers, disablers, intensifiers, and attenuators, at least.
252  
option. However, Schroeder makes exactly this kind of assumption about

reasons against: he assumes that reasons against an option are simply reasons
for not performing that option.¹³ So his version of the balance account is
not so different from Balance, after all; it’s just Balance with a particular
conception of what the alternatives to A are—namely, not doing A. The
important thing is that both focus just on the competition between reasons
for incompatible options.
In fact, it is important that Schroeder does make this assumption, because
otherwise his account faces a similar problem. Suppose that in the previous
case, the electrical shock that I will receive if I press button 2 is mild enough
that I would be happy to endure it to receive $100. That is, the reason against
pressing button 2 is weaker than the reason for pressing button 2. So. accord-
ing to Schroeder’s view, it looks like I am required to press button 2, which is
false. But if reasons against an option are simply reasons for not performing
that option, we may be able to avoid this problem. In particular, if we can
assume that the reason for pressing button 1—that it will get me $100—is a
reason for not pressing button 2—that is, given Schroeder’s assumption, a
reason against pressing button 2—then we can avoid the result that I am
required to press button 2.¹⁴
So, the assumption that reasons against are simply reasons for alternatives
helps different versions of the balancing account avoid problems. Moreover,
this assumption is a very natural one to come to if you begin with the
balancing metaphor, and then want to accommodate reasons against. We
know how to think about reasons for on this picture: they’re like marbles
that go into the relevant pan. It isn’t immediately obvious how to build in
reasons against, on the other hand. But, if these are just reasons for alterna-
tives, then we know how to accommodate them. Less metaphorically, this
assumption will simplify our theory of the competition between reasons, since
we only have to understand one kind of rebutting defeat: the kind that holds
between reasons for incompatible options. But, as I argue in the next section,
the assumption that reasons against an option are simply reasons for alterna-
tives to that option is false. So, we cannot appeal to this assumption to defend
the balancing account.
¹³ See Schroeder (2015) for relevant discussion, and Nagel (1970): 47, for the same assumption.
¹⁴ See Pendlebury (ms), who also makes this observation.
  253
11.3. Reasons For and Reasons Against
I will begin by introducing a paradigmatic example of a reason against. This is

borrowed from Greenspan (2005), who also argues (in a different way) for a
sharp distinction between reasons for and reasons against. Then I will consider
various ways of trying to reduce this reason against the option to a reason for
alternatives, and argue that none are satisfactory.¹⁵ This will support the claim
that the reason against relation is distinct from the reason for relation, and
thus the claim that our theory of the competition between reasons should
account for both the competition between reasons for competing options
and the competition between reasons for an option and reasons against that
same option.
Suppose that I am trying to decide which shirt to wear: the blue one, the red
one, or the green one. The blue shirt is boring. This fact is clearly relevant to
the choice at hand. This is because it’s a reason against wearing the blue shirt.
On the view that reasons against are simply reasons for alternatives, its being a
reason against wearing the blue shirt must amount to its being a reason for
alternatives to wearing the blue shirt. The important question, then, is: which
of the alternatives?
Perhaps the most natural answer—which has been endorsed by Nagel
(1970) and, as we saw in the previous section, Schroeder (2007)—is that
reasons against an action are just reasons for not performing that action. So,
to say that the fact that the blue shirt is boring is a reason against wearing
the blue shirt is just to say that this fact is a reason for not wearing the blue
shirt. I believe that this answer is plausible when we think about the choice
the agent is facing as a choice about whether to wear the blue shirt—that is, a
choice between wearing the blue shirt and not wearing the blue shirt. But, in
this case, the choice is more fine-grained than this: it is a choice between
wearing the blue shirt, wearing the red shirt, and wearing the green shirt.
Reasons against wearing the blue shirt, like the fact that it is boring, are
clearly relevant for this choice, and not only the more coarse-grained choice
between wearing the blue shirt and not wearing it, since they count against
wearing the blue shirt. But, as so far stated, the view in question doesn’t
really tell us how it bears on this choice. It must count against wearing the
blue shirt by counting in favor of some alternative option(s). But not
wearing the blue shirt is not one of the options. The same kind of objection
would apply to the view that a reason against wearing the blue shirt is just a
¹⁵ See Snedegar (2018) for more detailed versions of these arguments.
254  
reason for the disjunction of the alternatives, since the disjunction of the
alternatives is not one of the options, either.¹⁶
A second initially plausible answer is that the fact that the blue shirt is
boring is a reason for each of the alternatives. Just as a reason for an option is
or indicates something positive (in some sense) about that option, a reason
against an option is or indicates something negative about that option—
something we’d like to avoid if possible. Wearing the red shirt would let me
avoid what’s bad about wearing the blue shirt, namely its boringness. So,
would wearing the green shirt. But this view is false. Suppose that the green
shirt is even more boring than the blue shirt. Then the fact that the blue shirt is
boring—though it is a reason against wearing the blue shirt—is not a reason
for wearing the green shirt. So, reasons against an option are not simply
reasons for each of the alternatives. Moreover, this case shows that even the
weaker claim that each reason against an option is also a reason for each of the
alternatives (though reasons against may not be reducible to reasons for
alternatives) is also false.¹⁷
A third answer is that what it is to be a reason against an option is to be a
reason for some alternative to that option. The fact that the blue shirt is boring
is a reason against wearing the blue shirt. On this third view, this means that
this fact is a reason for some alternative to wearing the blue shirt. The two
alternatives are wearing the red shirt and wearing the green shirt. Plausibly a
reason for wearing the red shirt—that it’s very comfortable—is a reason
against wearing the blue shirt, if the blue shirt is not as comfortable. But the
fact that the blue shirt is boring doesn’t seem to be a reason for any particular
alternative to wearing the blue shirt. If we suppose that the green shirt is even
more boring than the blue one, as discussed, then it is clearly not a reason for
wearing the green one. Maybe it is a reason for wearing the red one, but the
view in question seems to get the direction of explanation backwards: it isn’t
that it’s a reason against wearing the blue shirt because it’s a reason for wearing
the red one, or that its being a reason against wearing the blue one just
amounts to its being a reason for wearing the red one. Rather, it’s a reason
¹⁶ Appealing to a kind of transmission principle, according to which a reason for not doing A, or for
doing the disjunction of alternatives to A, bears on the choice at hand by being a reason for each way of
not doing A, or of doing the disjunction, makes the view equivalent to the view I discuss in the next
paragraph.
¹⁷ This brings out further objection to the view that reasons against A are reasons for the disjunction
of alternatives to A. Some of the alternatives to A may be even worse than A in the relevant respect (i.e.
the respect which explains the reason against). If so, then doing that alternative is not plausibly a way of
complying with the reason against doing A, but the view that reasons against are just reasons for the
disjunction of alternatives seems to imply that it is, since it is a way of doing that disjunction.
  255
for wearing the red one—if it is—because it’s a reason against wearing the
blue one.¹⁸
So, none of these three natural versions of the idea that reasons against an
option just are reasons for alternatives to that option seem promising. We
could offer more complicated answers, for example, that reasons against an
option are reasons for disjunctions of only some of the alternatives (e.g.,
leaving out the alternatives that are even worse along the relevant dimension).
But answers like this will not be very systematic, and if reasons against an
option are to be identified with or reduced to reasons for alternatives, we might
have hoped for a more systematic account. I think we do better to allow for the
existence of reasons against options that are not just reasons for any of the
alternatives to that option. These reasons against will bear on choices involv-
ing the option. There can also be reasons against an option that are not reasons
for alternatives that are even worse along the relevant dimension. Finally, we
do not need to find any particular alternative that the reasons against one
option are reasons for, in order for them to factor into the competition.
It is also good to step back and think about why we were trying to do
without a distinct category of reasons against in the first place. It is indeed hard
to see how a reason could count against an option other than by counting in
favor of alternatives to that option if we remain wedded to the balancing
metaphor. But it’s well known that this metaphor is oversimplified. Once we
move past that, there seems to be no special problem with reasons against: if
we can understand the idea that a reason for an option counts in favor of it,
there should be no special obstacle to understanding how a reason could count

against an option. Nevertheless, as I said previously, this does mean that our
theory of the competition between reasons will need to recognize at least two
kinds of rebutting defeat or outweighing: that between reasons for conflicting
options, and that between reasons for an option and reasons against that same
option. In the next section, I consider whether a different account of the
competition between reasons that incorporates a distinction between reasons
for and reasons against might be more successful.
11.4. Criticism-Based Accounts
I call the kind of view I consider in this section a criticism-based account. Both
Greenspan (2005, 2007) and Gert (2003, 2004, 2007, 2016) defend views of this
¹⁸ Compare Greenspan (2005) here.
256  
sort. As I’ll explain briefly, the version I describe here is closer to Greenspan’s
version, but I take the most important part of the view (for my purposes) to be
common between Greenspan and Gert.
According to criticism-based accounts, the conception of reasons and
rationality lying behind the balancing account of stacking up reasons for
various options and weighing them against each other is mistaken. Instead,
as Greenspan says, reasons are “invoked primarily to offer or answer criticism
of action,” and rationality is not about doing what your reasons most strongly
support, but rather “avoiding being subject to significant criticism or having a
response to it available.”¹⁹
Note that there are two importantly distinct functions of reasons here, and
correspondingly two distinct ways of acting rationally. First, reasons might
offer, or ground, criticism of an option. Second, reasons might answer such
criticism. You can act rationally either by not acting in ways for which you
could be criticized (or “significantly” criticized), or by having an adequate
response to such criticism. This contrasts with the balanced account, which
understands reasons to have a single underlying function, that of counting in
favor of options. Since one part of the challenge for balance accounts was to
explain two different ways in which reasons could compete, the criticism-
based account seems to be an improvement. In particular, the criticism-based
account includes a distinct role for reasons against—that of grounding (or
perhaps just serving as) criticisms.
The competition between reasons, on this view, works as follows. Reasons
against an option tend to rule it out as a permissible option—thus, they tend to

generate requirements not to act in that way. Reasons for an option, on the
other hand, can help justify you in performing the option even in the face of
the reasons against it. For example, that you risk being killed by running into
the street is a reason against doing so; it would typically be sufficient to
generate a requirement not to do this. But if there is a child stranded in the
street, that is a reason for running into the street that can plausibly justify you
in doing so, even in the face of the requiring reason not to.²⁰ That you risk
being killed by running into the street is a criticism of doing so; that the child is
stranded there answers this criticism, making running into the street rationally
permissible. From a moral perspective, things are reversed: the fact that the
child is stranded in the street counts against staying on the sidewalk and would
make it impermissible. But you have a reason for staying on the sidewalk,
¹⁹ Greenspan (2005): 389. ²⁰ I take this example from Gert (2007): 538.
  257
namely that there are cars hurtling past, that can morally justify you in staying
on the sidewalk.
So far this seems compatible with the balance account, since we could just
understand all of this in terms of the weights of the reasons involved.²¹ What is
distinctive of this view, for the purposes of this paper, is the insistence that
reasons against an option cannot be understood as reasons for alternatives to
the option (including reasons for not performing the option). The central
notion of criticism gives us a basis for a substantive distinction by understand-
ing reasons against an option as grounding criticisms of the option, and
reasons for as answering those criticisms. To criticize one option is not
necessarily to answer a criticism of another option. That the blue shirt is
boring does not necessarily answer any criticism of wearing the red shirt. In
fact, it’s plausible that criticizing one option does not even necessarily provide
any support for alternatives. At the very least, we can understand the argu-
ments against identifying reasons against with reasons for as showing that it is
far from straightforward to identify which alternative a criticism of a given
option would support, in a way that both gives plausible results (by not falsely
claiming that a criticism of A supports B when B is subject to the same kind of
criticism, for example) and explains how reasons for and against can play the
appropriate role in deliberation (by explaining how they actually bear on the
options under consideration). In addition, answering a criticism of one option
is not the same as criticizing an alternative option, so reasons for alternatives
to A cannot just be reasons against A.²² For example, that I like the way the
blue shirt fits may answer the criticism that it’s boring, but it does not
necessarily amount to any criticism of the alternatives. This means that the
opposite kind of reduction to that suggested on behalf of the balancing
account—reducing reasons for alternatives to A to reasons against A, instead
of reducing reasons against A to reasons for alternatives to A—won’t work,
either. Thus, on this view, we have a sharp distinction between reasons for and
reasons against.
Gert (2004, 2007) focuses on a distinction between the requiring and
justifying strengths of reasons. The mechanics of the competition between
²¹ In fact, one of both Greenspan’s and Gert’s main points is that we cannot properly handle these
kinds of cases just by appealing to some single notion of weight. In particular, we can’t explain why you
are permitted to run into the street but not required to do so.
²² Though of course sometimes it can, in a sense. Kurt Sylvan (p.c.) suggests that some criticisms of
one option, A, might be so serious that they suffice to answer or at least justify ignoring criticisms of the
alternative, B. But here it is more natural to say that the reason for B, or the answer to the criticisms B,
are that you can avoid A, or maybe just that the (only) alternative is so terrible. If so then the reason for
B is not identical with but rather derivative of the reason against A.
258  
reasons are similar to that previously sketched: justifying reasons (or reasons
with justifying strength) are used to justify you in performing options that you
would otherwise be required not to perform, due to the requiring reasons
against them. Gert’s examples often involve reasons against options with
requiring strength and reasons for those options with justifying strength. For
example, he holds that in the case involving the child stranded in the street, the
fact that the child is stranded in the street has significant justifying strength in
favor of running into the street—enough to justify you in doing so even in the
face of the strong requiring reason against doing so, but no requiring strength,
since it cannot generate a rational requirement to run into the street (whether
it can generate moral requirements is a different question). But in fact, the
distinction between requiring and justifying and that between reasons for and
against are cross-cutting. Nevertheless, on Gert’s view the primary competition
between reasons is always between reasons for an option and reasons against that
same option, where the two kinds of reasons are playing importantly different
roles. It’s just that sometimes, for Gert, the reasons for play the requiring role
while reasons against play the justifying role, and sometimes the reverse is true.
Thus, in focusing on the competition between reasons for an option and reasons
against it, Gert’s view is relevantly similar to the one sketched in this section,
which is most directly based on Greenspan (2005, 2007).
The criticism-based view, understood in terms of reasons for and reasons
against, gives a plausible account of the competition between reasons for an
option and reasons against it. When the reasons against an option win out, you
are required not to perform that option. When the reasons for an option win
out, you are permitted to perform it. Both Gert and Greenspan emphasize that
this is the most we can say here: we cannot go on to say that you are required
to perform the option, just because the reasons for the option win out over the
reasons against it. This is because both reject a maximizing conception of
rationality. All that rationality requires is not acting as you have overriding
reason against acting or avoiding unanswered criticism. So, your reasons will
generate a requirement for a particular action only by ruling out all of the
alternatives: an option is required if and only if all of the alternatives are
impermissible, while it is permissible.²³
Even granting this non-maximizing conception of rationality, this leaves
out an important element for a normative theory, namely an account of what
you ought to do. Though what you ought to do is not always properly
²³ This last clause is important if we accept the possibility of dilemmas, in which all of our options
are impermissible, and none are required.
  259
distinguished from what you are required to do in moral philosophy, these are
two importantly distinct overall normative statuses. What you are required to
do is plausibly very closely tied to what you can be criticized for not doing. But
what you ought to do is not necessarily tied to criticism in this way.²⁴ Think of
the way we might describe the normative status of some supererogatory
action: “You really ought to help out, though of course you don’t have to.”
On some demanding moral theories, there may not be a substantive distinc-
tion between what you ought to do and what you are required to do, but
I think it is clear that there is at least a conceptual distinction.²⁵
The criticism-based view has trouble accounting for what you ought to do,
in this sense.²⁶ This is because ought is closely connected to what’s best, in the
sense of being the most strongly supported by reasons.²⁷ ‘Best’ and ‘most’ are
comparative terms: you ought to do something when it’s better or more
strongly supported by reasons than the alternatives. The account of the com-
petition between reasons sketched so far on behalf of the criticism-based view
is focused on the competition between the reasons for an option and the
reasons against it. We don’t yet have any way of comparing different alterna-
tives in terms of how well supported they are by the reasons.²⁸
The balance account focuses on the competition between reasons for
competing options but does not have a natural way to account for the
competition between reasons for an option and reasons against it. The
criticism-based account focuses on the latter kind of competition but does
not have a natural way to account for the former. The latter’s trouble account-
ing for what we ought to do, as opposed to what we’re required to do, is a
symptom of this problem.
Suppose that reasons for one option are also reasons against alternatives,
even if we don’t identify reasons against with reasons for alternatives. Then we
have some sort of relationship between the reasons bearing on different
alternatives. But even if this is true, such that reasons for alternatives to an
²⁴ Perhaps some sort of criticism is appropriate in some cases in which you don’t do what you ought
to do, even if you aren’t required to do it; see, e.g., Driver (1992); Macnamara (2013). But the criticism
will at least typically be weaker and not call for apologies, reparations, punishment, and so on.
²⁵ For a sampling of work on this issue, see McNamara (1996a, b); Bedke (2011); Snedegar (2016).
²⁶ Bedke (2011) also argues that Gert’s view fails to explain ought as opposed to requirement, though
in a different way. See also Snedegar (2016).
²⁷ For early work on the connection between ‘ought’ and ‘best,’ see Sloman (1970); Finlay (2010,
2014) also emphasizes the connection between ‘ought’ and ‘best.’
²⁸ Perhaps using ‘ought’ to describe the best action is sometimes too strong, or at least can sound
wrong, for example when the best action is only best by some trivial amount. The objection to the
criticism-based view is essentially just that it leaves us no way to make comparisons between the
options in terms of how strongly supported those options are. Focusing on the overall status expressed
by ‘ought’ is a natural way to bring this out. Thanks to Kurt Sylvan here.
260  
option are factored into the competition between reasons for that option and
reasons against it, this still doesn’t yet solve the main issue raised previously.
The problem is to explain how to compare competing alternatives in terms of
how well they are supported by the reasons. Without this, we won’t be able to
explain how your reasons determine what you ought to do, in addition to what
you must and may do. The reasons bearing on alternatives get factored into
the competition between reasons for and reasons against a given option; but
the outcome of this competition determines what you must or may do. So, we
still don’t have an answer to the question of how your reasons determine what
you ought to do. In the next section, I sketch a positive account that draws on
aspects of both the balance account and the criticism-based account, and argue
that it gives us a more complete picture of how the competition between
reasons determines the overall normative status of our options. But I’ll also
show that open questions and problems remain.
10.5. Toward a Positive Account
To sum up and simplify a bit, criticism-based accounts recognize the import-

ance of considering the reasons against an option in determining its overall
normative status, but do not seem to accommodate the importance of
comparing the option to its alternatives. Balancing accounts, on the other
hand, focus on comparisons between alternatives, but do not have a natural
way of accommodating the importance of reasons against options. If reasons

against an option were simply reasons for alternatives, then each kind of
account may avoid these problems. But as I’ve argued, reasons against cannot
be reduced to or identified with reasons for alternatives. In this section,
I sketch a view that explicitly recognizes the two different kinds of compe-
tition between reasons—the two different kinds of rebutting defeat—that I’ve
focused on in this paper. The view does not require that the two kinds of
competition are radically distinct, such that what it is for a reason for A to
compete with a reason for B is very different, metaphysically speaking, from
what it is for a reason for A to compete with a reason against A. In sketching
the account, I understand both kinds of competition as involving the weights
of the reasons involved. Rather, the difference is in the normative relevance,
in terms of what you ought to, must, or may do, of the outcomes of the two
different kinds of competition. It may be that for the outcomes of the
competitions to play these roles, the two kinds of competition must differ
in some deep way, but I leave that question aside.
  261
The view incorporates elements from both criticism-based accounts and

balancing accounts. I’ll explain the role of each kind of competition in
determining the overall normative status of the options. Along the way
I will flag some important questions and choice points; settling these will
require further work, which will sometimes involve doing substantive norma-
tive theory.
The proposal here is not radically new. I suspect it is largely just a way of
spelling out an orthodox way of thinking about the competition between
reasons.²⁹ What is important is (i) the distinction between the two different
kinds of competition between reasons, and correspondingly, (ii) the different
roles these competitions have in determining overall normative status.³⁰
The discussion will be idealized in familiar ways by assuming that we can
perform something like mathematical functions on and make mathematical
comparisons of the weights of reasons and sets of reasons. I assume (i) that we
can compare the weights of reasons and sets of reasons using something like
the greater than (>), lesser than (<), and equal to (=) relations, (ii) that we can
make sense of the combined weight of reasons (e.g., R1 and R2, taken together,
support the option to a certain degree) and the difference in weights of reasons
(e.g., R1 is weightier than R2 by a certain degree), and (iii) that we can make
sense of the overall weight of the reasons bearing on an option being positive
or negative—that is, that the reasons, taken together, either support the option
or count against it. Importantly, (ii) is not to assume that the weights of
reasons combine in a purely additive way, such that if R1 and R2 are the
only reasons for A, then the combined weight of the reasons for A is simply the
weight of R1 plus the weight of R2, or that if R3 is a reason for A and R4 is a
reason against A, and these are the only reasons in play, then the overall
weight of reasons for A is simply the weight of R3 minus the weight of R4.
Nevertheless, for convenience I will use ‘+’ for the combining weights oper-
ation and ‘ ’ for the difference in weights operation. Issues involving the more
complicated ways that reasons actually seem to combine and weigh against
each other will have to be left aside here.³¹
The first kind of competition is that between reasons for an option and
reasons against it. What we need for this is the combined weight of all the
reasons for A, W(RFA), and the combined weight of all the reasons against A,
²⁹ Compare the generalized weighing framework in Berker (2007), especially: 113–14.

³⁰ Since writing this paper, I have become aware of an article by Euan Metz (2020) that advances a
view with some similarities. I unfortunately lack space to compare and contrast our views here.
³¹ For work on related issues, see Dancy (2004), especially Ch. 3; Bader (2016); Nair (2016); Sher
(2019); Maguire (2016); Maguire and Snedegar (forthcoming).
262  
W(RRA). Then we find the difference in these weights:W(RFA) W(RAA) = W

(RA), which gives us the degree of overall reason for an option. As a first pass at
how the competition between reasons for and reasons against determines the
overall normative status of an option, I propose the following. If W(RA) is
positive (i.e., if W(RFA) > W(RAA)), then A is permissible. If W(RA) is negative
(i.e., if W(RFA) < W(RAA)), then A is impermissible.³² This leaves cases in
which the reasons for and reasons against an option are balanced. I am
inclined to think that such options are permissible, to accommodate the
thought that when there’s no reason at all either for or against A, A is
permissible, but there are likely interesting questions here (compare: “There
was no reason to do that!” as a criticism, vs. “There was no reason not to!” as a
defense). So provisionally, we can say that when W(RFA) W(RAA), A is
permissible.
This takes from the criticism-based approach the idea that the reasons
against an option are particularly relevant in determining whether it is per-
missible. They tend to rule out performing the option, and unless there are
sufficiently strong reasons in favor, the option will be impermissible. Thinking
of the reasons against an option as criticisms of it makes this particularly clear:
if there are criticisms of some option that cannot be answered or compensated
for, then it is plausible that the option is impermissible.
As stated so far, though, this account of permissibility faces a problem.³³ We
can see this by looking at the case I used previously to criticize balance
accounts. You can press either button 1 or button 2. Pressing either button
will get you a particular $100 bill but pressing 2 will also deliver a mild electric
shock. You would be willing to endure this shock to get $100, so it seems that
the reason for pressing 2, that it will get you the $100 bill, is weightier than the
reason against, that you will be shocked. So according to the account, pressing
button 2 is permissible. But this is incorrect: pressing button 2 in this case
would be foolish.
I have tried to explain permissibility (and, shortly, requirement) just in
terms of the reasons for and against the particular option, whereas what you
ought to do will be explained in terms of comparisons among the alternatives.
But the lesson of this case is that even for permissibility and requirement, the
alternatives matter: it is only because of the availability of pressing button 1
that pressing button 2 is impermissible. The task for now is to say exactly how
it is that the alternatives matter for permissibility and requirement.
³² Compare Sher (2019): 2. ³³ Thanks to Jonathan Way here.
  263
One way in which the alternatives can matter is by the reasons bearing on
those alternatives transmitting to the option in question. Sometimes reasons
for one alternative are also reasons against incompatible alternatives. This is
one way to explain the notion of an opportunity cost: among the costs of an
option are the good features of the alternatives that you’ll miss out on. This is
consistent with denying that all reasons against options are nothing over and
above reasons for alternatives. But in this case, given that you will get the $100
either way, the reason for pressing button 1—that it will get you $100—is not a
reason against pressing button 2.
Rather, I think that in cases like this, the alternatives matter by helping to
determine which considerations are or are not reasons for the option in
question. In particular, since you will get the $100 either way, the fact that
you’ll get the $100 by pressing button 2 is not a reason for pressing button 2.
Hence, there is only a reason against pressing button 2 (the shock), and so it is
impermissible.
The surprising consequence of this view is that the $100 also does not give
you reason to press button 1. So, if the money and the shock are the only
relevant considerations, there won’t be any reason to press button 1.
Nevertheless, pressing button 1 will be permissible, since there’s at least as
much reason for it as against it (namely, no reason either way). In fact it will be
required, since the only alternative is impermissible.³⁴ Though it may seem
initially surprising to deny that the fact that you can get $100 by pressing
button 1 (or button 2, for that matter) is a reason for pressing button 1, I think
this is plausible. The crucial point is that you’ll get the $100 no matter what
you do (assuming pressing 1 and pressing 2 are the only options). So, you
cannot, for example, justify yourself in pressing button 1 (instead of button 2)
by pointing out that you’ll get $100. Also, it would be odd to base your decision
to press button 1 (instead of button 2) on the fact that you will get $100 by
pressing button 1. Being able to serve as justifiers or at least sensible bases for
decision are commonly thought to be earmarks, if not conceptual require-
ments, for some consideration’s being a reason.³⁵ So it should not seem so
³⁴ In fact, things will be a bit trickier since it seems that we can make true reasons ascriptions like,
“That you will get shocked by pressing button 2 is a reason to press button 1, instead.” I think that this
is true, but that the reason ascribed is a derivative reason—derivative on the reason against pressing 2.
A full theory here will need to explain when we get these derivative reasons and how their weights do or
do not contribute to the overall status of the options in question. It would seem to be a mistake, for
example, to treat this derivative reason for pressing button 1 as having weight that is independent of
and in addition to the weight of the reason against pressing button 2 on which it is derivative. Compare
Maguire and Snedegar (forthcoming).
³⁵ See, e.g., Schroeder (2007): Ch. 2; Setiya (2014); Gregory (2016); Silverstein (2016); Way (2017);
Snedegar (2019).
264  
surprising that in this case, that you can get $100 by pressing button 1 will not
be a reason to press button 1, since you’ll get the $100 no matter what you do.
If this kind of solution is on the right track, and can be generalized, then we
have the start of an account that tells us when an option is permissible in terms
of the reasons for and against it. We also get an account of when an option is
required: an option is required when all of the other options are
impermissible—that is, when the reasons against each of the other options
outweigh the reasons for them. If more than one option is permissible, then no
single option is required.³⁶ There are clear similarities between this under-
standing of permissibility and requirement and the one we get from the
criticism-based account.
But, as I argued in the previous section, this leaves out an important part of
the normative domain—namely, what we ought to do, as opposed to what
we’re required to do. To quickly review the main points from the previous
section: we ought to do is very plausibly what’s best, in some sense.³⁷ But, to
say that some option is the best is to say that it’s better than the alternatives.
So, to account for what we ought to do, we need to introduce a comparison
among the alternatives.
The overall weight of the reasons bearing on an option, A, which I called W
(RA), is determined by the difference between the weight of the set of reasons
for A and the set of reasons against A. So, we can compare these weights of
overall reason for each of the options. What we ought to do is the option that
ranks at the top of this comparison: you ought to do A when W(RA) > W(RB),
for all the alternatives, B. Whereas the competition between reasons for an
option and reasons against that option are most directly relevant for deter-
mining which options are permissible and required, the competition between
the reasons bearing on competing alternatives is most directly relevant for
determining what you ought to do.³⁸
This picture has some attractive features. First, it explains important overall
or verdictive normative statuses—permissibility, impermissibility, requirement,
³⁶ An open question at this point is whether it is possible for the reasons against to win out for every
option, making every option impermissible. Many people think that such tragic dilemmas are impos-
sible. We could, for example, argue that the fact that an option is the “least bad” is always a decisive
reason to do it.
³⁷ Cf. Sloman (1970); Finlay (2010); von Fintel and Iatridou (2008).
³⁸ As noted in the Introduction, I am focusing on the moral case. In the epistemic domain, it is much
less clear that there is a substantial difference between ought and requirement. One possible explan-
ation for this, which I cannot explore at any length here, is that in the epistemic domain, unlike in the
moral domain, reasons for believing p, or evidence for p, are always reasons against believing, or
evidence against, propositions that are incompatible with p. If so, then it would plausibly turn out that
any time the reasons for believing p outweighed the reasons against believing p, the reasons against
believing any incompatible alternative q would outweigh the reasons for believing q.
  265
and ought—in terms of pro tanto reasons. This is a popular and attractive idea,
but it has proved challenging to provide satisfactory accounts of all of these
notions in terms of reasons.³⁹ Second, the picture captures the correct logical
relationships between permissibility, requirement, and ought. To focus on the
relationship between requirement and ought: if some option A is required, then
it is also what you ought to do. Since A is required, the reasons against each
alternative B outweigh the reasons for B; thus W(RB) will be negative for all the
alternatives, B. Moreover, since A is required, it is the only permissible option, in
which case it is the only option for which W(RA) is positive. So, W(RA) > W(RB)
for each alternative B, and thus A is what you ought to do. But A might be what
you ought to do without being required, since it is compatible with W(RA) > W
(RB) that W(RB) is nevertheless positive—that is, B is permissible. This is just
what we want, because it has been widely noted in both linguistics and philoso-
phy (the former focusing on the modals that express these normative statuses
and the latter focusing on the statuses themselves) that there is a one-way
entailment from requirement to ought.⁴⁰ This captures the commonsense idea
that there are things that you really ought to do, though you are not required to
do so, and are not acting impermissibly if you don’t—think of paradigm cases of
supererogation, for example.
This is just a sketch, and open questions remain, but I do think it gives us
the beginnings of an attractive account. As I have already noted, I don’t take
this view to be any radical departure from an orthodox way of thinking about
how reasons compete. Rather, what is important about it is the incorporation
of the two different kinds of competition between reasons. Both the standard
balance account and the criticism-based account are incomplete because they
focus on just one kind of competition.
References
Bader, R. (2016). Conditions, modifiers, and holism. In: Lord, E. and Maguire, B.
(eds.) Weighing Reasons. Oxford: Oxford University Press: 27–55.
Bedke, M. (2011). Passing the deontic buck. In: Shafer-Landau, R. (ed.) Oxford
Studies in Metaethics. Vol. 6. Oxford: Oxford University Press: 128–52.
Berker, S. (2007). Particular reasons. Ethics, 118(1): 109–39.
³⁹ See Snedegar (2016) for arguments against some recent attempts, including Bedke (2011).
⁴⁰ See, for example, Wertheimer (1972); McNamara (1996b); von Fintel and Iatridou (2008);
Portner (2009); Snedegar (2012, 2016).
266  
Broome, J. (2004). Reasons. In: Wallace, Scheffler, and Smith (eds.) Reason and
Value: Themes from the Moral Philosophy of Joseph Raz. Oxford: Oxford
Dancy, J. (2004). Ethics Without Principles. Oxford: Oxford University Press.
Driver, J. (1992). The suberogatory. Australasian Journal of Philosophy, 70(3):
286–95.
Finlay, S. (2010). What ‘ought’ probably means, and why you can’t detach it.
Synthese, 177: 67–89.
Finlay, S. (2014). Confusion of Tongues. Oxford: Oxford University Press.
von Fintel, K. and Iatridou, S. (2008). How to say ‘ought’ in foreign: The com-
position of weak necessity modals. In: Guéron and Lecarme, (eds.) Time and
Modality. Dordrecht: Springer: 115–41.
Gert, J. (2003). Requiring and justifying: Two dimensions of normative strength.
Erkenntis, 59: 5–36.
Gert, J. (2004). Brute Rationality: Normativity and Human Action. Cambridge:
Cambridge University Press.
Gert, J. (2007). Normative strength and the balance of reasons. Philosophical
Review, 116(4).
Gert, J. (2016). The distinction between requiring and justifying: nothing to fear.
In: Lord, E. and Maguire, B. (eds.) Weighing Reasons. Oxford: Oxford
Greenspan, P. (2005). Asymmetrical practical reasons. In: Reicher, M. and Marek,
J. (eds.) Experience and Analysis: Proceedings of the 27th International
Wittgenstein Symposium. Vienna: Oebv & Hpt: 387–94.

Greenspan, P. (2007). Practical reasons and moral ought. In: Shafer-Landau, R.
(ed.) Oxford Studies in Metaethics. Vol. 2. Oxford: Oxford University Press:
172–94.
Gregory, A. (2016). Normative reasons as good bases. Philosophical Studies 173
(9): 2291–310.
Horty, J. and Nair, S. (2018). The logic of reasons. In: Star, D. (ed.) Oxford
Handbook of Reasons and Normativity. Oxford: Oxford University Press:
67–84.
Horty, J. F. (2012). Reasons as Defaults. Oxford: Oxford University Press.
Lord, E. (2018). The Importance of Being Rational. Oxford: Oxford University
Press.
Lord, E. and Maguire, B. (2016). An opinionated guide to the weight of reasons.
In: Lord, E. and Maguire, B. (eds.) Weighing Reasons. 3–24. Oxford: Oxford
University Press.
  267
Macnamara, C. (2013). Taking demands out of blame. In: Coates, D. J. and

Tognazzini, N. (eds.) Blame: Its Nature and Norms. Oxford: Oxford
Maguire, B. (2016). The value-based theory of reasons. Ergo, 3(9): 233–62.
Maguire, B. and Snedegar, J. (Forthcoming). Normative metaphysics for account-
ants. Philosophical Studies.
McNamara, P. (1996a). Making room for going beyond the call. Mind, 105(419):
415–50.
McNamara, P. (1996b). Must I do what I ought (or will the least I can do do)? In:
Brown, M. A. and Carmo, J. (eds.) Deontic Logic, Agency and Normative
Systems. Berlin: Springer-Verlag: 154–73.
Metz, E. (2020). How to be a deonic buck passer. Philosophical Studies, 177(11):
3193–211.
Nagel, T. (1970). The Possibility of Altruism. Princeton: Princeton University
Press.
Nair, S. (2016). How do reasons accrue? In: Lord, E. and Maguire, B. (eds.)
Weighing Reasons. Oxford: Oxford University Press: 56–73.
Parfit, D. (2011). On What Matters. Oxford: Oxford University Press.
Pendlebury, M. Manuscript. Reasons, ought, must, and may. Unpublished manu-
script, North Carolina State University.
Pollock, J. (1974). Knowledge and Justification. Princeton: Princeton University
Press.
Portner, P. (2009). Modality. Oxford: Oxford University Press.
Raz, J. (1999). Practical Reason and Norms. Oxford: Oxford University Press.
Ross, W. D. (1930). The Right and the Good. Oxford: Oxford University Press.
Schroeder, M. (2007). Slaves of the Passions. Oxford: Oxford University Press.
Schroeder, M. (2015). What makes reasons sufficient? American Philosophical
Quarterly, 52(2): 159–70.
Setiya, K. (2014). What makes reasons sufficient? Philosophical Studies, 167:
221–35.
Sher, I. (2019). Comparative value and the weight of reasons. Economics and
Philosophy, 35(1): 103–58.
Silverstein, M. (2016). Reducing reasons. Journal of Ethics and Social Philosophy,
10(1): 1–22.
Sloman, A. (1970). “Ought” and “better.” Mind, 79(315): 385–94.
Snedegar, J. (2012). Contrastive semantics for deontic modals. In: Blaauw, M. (ed.)
Contrastivism in Philosophy. New York: Routledge: 166–33.
268  
Snedegar, J. (2016). Reasons, oughts, and requirements. In: Shafer-Landau, R.

(ed.) Oxford Studies in Metaethics. Vol. 11. Oxford: Oxford University Press:
155–81.
Snedegar, J. (2017). Contrastive Reasons. Oxford: Oxford University Press.
Snedegar, J. (2018). Reasons for and reasons against. Philosophical Studies, 175:
725–43.
Snedegar, J. (2019). Deliberation, reasons, and alternatives. Pacific Philosophical
Quarterly, 100(3): 682–702.
Way, J. (2017). Reasons as premises of good reasoning. Pacific Philosophical
Quarterly, 98(2): 251–70.
Wertheimer, R. (1972). The Significance of Sense: Meaning, Modality, and
Morality. Ithaca: Cornell University Press.
Whiting, D. (2017). Against second-order reasons. Noûs, 51(2): 398–420.
12
Perceptual Reasons and Defeat
Mark Schroeder
12.1. Background: Reasons and Evidence
I am going to assume in this paper that perceptual experience is a way of

acquiring evidence about the world. I will assume without argument that it is
more reasonable to believe things that are supported by evidence than to
believe without any evidence, or at least that if that is not true, then that is
one of those surprising discoveries that philosophy is often claimed to allow us
to make. I am also going to move back and forth freely between claims about
reasons and claims about evidence. That is because I believe, and argue
elsewhere,¹ that the correct explanation for why it is more reasonable to
believe things that are supported by evidence is that evidence is reason to
believe, and the only kinds of reasons to believe are evidence.
The claim that evidence is reason to believe is supported by the fact that the
same things that are appropriately cited as evidence are also appropriately cited
as reasons to believe, and conversely. ‘Reason’ and ‘evidence’ talk also exhibit
parallel distinctions between objective talk about reasons or evidence that are
out there, to be discovered, and subjective talk about the reasons or evidence
that someone has, or from her perspective. It is also possible to do or to believe
things for reasons, and similarly, it is possible to believe things on the basis of
evidence. These facts, which could be enumerated at much greater length, all
constitute strong evidence that reasons and evidence are closely related.
From an ontological perspective, many kinds of thing can be cited appro-
priately as reasons. For example, the height of the Empire State Building is a
reason not to jump off, Robin Jeshion is a reason to study philosophy of
language at the University of Southern California (USC), and the bruise on
Alyssa’s leg is reason to believe that she has gotten hurt. Heights, people, and
bruises all belong, I take it, to different categories. But it is also true that one
reason not to jump off of the Empire State Building—and not, intuitively, a
¹ Schroeder (2021).
Mark Schroeder, Perceptual Reasons and Defeat In: Reasons, Justification, and Defeat. Edited by: Jessica Brown and Mona
98847205.003.0012
270  
different reason—is that it is so high. Similarly, one reason to study philosophy

of language at USC is that Robin Jeshion teaches there. And one reason to
believe that Alyssa has gotten hurt is that she has a bruise on her leg. So,
although it makes perfect sense to cite many kinds of thing as reasons, in each
case there are adequate paraphrases of what has been reported which attribute
the reason in question using a ‘that’ clause. This may support the conclusion
(I think it does) that strictly speaking, reasons should be identified with what is
expressed by ‘that’ clauses—which I take to be propositions. But whether or
not it supports this strong conclusion about the ontology of reasons, at least it
supports the conclusion that it is a condition of adequacy for any view about
what the reason is for someone to do something that it withstand paraphrase
in terms of a ‘that’ clause.
Finally, the same points go for evidence as for reasons. The height of the
Empire State Building is evidence that it took a long time to build, Robin
Jeshion is evidence that philosophers of language can shed light on rich and
subtle social issues, and Alyssa’s bruise is evidence that she has gotten hurt.
Again, the evidence that it took a long time to build the Empire State Building
is that it is so high. Similarly, the evidence that philosophers of language can
shed light on rich and subtle social issues is that Robin Jeshion has done so in
her work on slurs and derogation. This again illustrates the close parallels
between talk about evidence and talk about reasons, and I’ll rely on this point
in what follows.
12.2. Perceptual Knowledge and Defeat
In this paper I am concerned with the defeasibility of perceptual evidence. But

we should be more careful in describing the phenomena in which we will be
interested, of which there are two. Both phenomena involve cases in which it
appears that someone would know but for some fact. But the kinds of fact but
for which she would know are different in each case.
Objective Defeat: Roberta has a visual experience as of something red in

front of her. But she is wearing rose-colored glasses.
Subjective Defeat: Roberta has a visual experience as of something red in
front of her. But she rationally believes that she is wearing rose-colored glasses.
In Objective Defeat, Roberta cannot know that there is something red in front
of her unless she has some independent source of evidence that that is so. Even
    271
if she does not realize that she is wearing rose-colored glasses, since they make
everything look red, they are incompatible with Roberta having the discrim-
inatory capacity to know that there really is something red in front of her, even
if there is. In Subjective Defeat, it seems on the face of it that Roberta cannot
reasonably believe that there is something red in front of her, since she
rationally believes herself to be in Objective Defeat, and in that case she cannot
know. But it seems that knowledge requires reasonable belief, so if it is not
reasonable for her to believe it, then it seems that she cannot know it, either—
again, unless she has some independent source of evidence, a qualification
which I will henceforth ignore.
There are many pairs of cases that stand to one another as Objective Defeat
and Subjective Defeat stand to one another, in the first of which some worldly
fact is what stands in the way of a subject knowing, and in the second of which
the subject’s rational belief in that same worldly fact is what stands in the way
of her knowing. Least interestingly, there is the pair in which there is in fact
nothing red in front of Roberta, and in which she rationally believes that there
is nothing red in front of her—exhibiting the plausibility of the truth and
rational belief conditions on knowledge. Some pairs consist in the availability
of compelling counterevidence and Roberta’s rational belief in the compelling
counterevidence. According to pragmatic encroachers, there are also pairs in
which Roberta faces high stakes and in which she rationally believes herself to
face high stakes. And many pairs are like the one that I have exhibited—cases
in which the force of Roberta’s evidence appear to be undercut by some further
fact. Though all of these pairs are possible, the views with which I will be
concerned in this paper are all accounts of the paradigmatic kind of defeas-
ibility typically referred to as undercutting defeat—cases in which the strength
or force of one’s evidence has been called into question.
I want to be a little bit careful about how I describe the data about these two
cases, since some of the views that I go on to discuss will disagree about how to
describe them. My own view is that both cases can and should be taken at face
value—that there are two dimensions along which knowledge can be defeated,
both by the facts, independently of what the subject believes about them, and
by the subject’s other beliefs. I will call these two dimensions objective defeat and
subjective defeat, respectively. I also believe that these two dimensions are
independent—cases of objective defeat without subjective defeat are (typically²)
recipes for Gettier cases, and cases of subjective defeat without objective defeat
undermine the justification or, as I have been putting it in this paper, the
² Failures of the truth condition are not, of course, recipes for Gettier cases.
272  
reasonability condition on knowledge. But as we will see later, some theorists are
going to struggle to accommodate subjective defeat. They may still accept that
there is an appearance of defeat of knowledge in cases of subjective defeat, but
explain this appearance in some other way, because something else nearby,
but somewhat different, is true.
12.2.1 The Classical Theory—What
According to what I will call the classical theory of the defeasibility of percep-
tual evidence, it can be properly subsumed to the non-monotonicity of
inference. Non-monotonic inferences are cases where R is a reasonable con-
clusion to draw from P, but not a reasonable conclusion to draw from
P&Q. For example, if you know that Tweety is a bird, it is reasonable to
draw the conclusion that Tweety probably flies, but if you also know that
Tweety is a penguin, then this is not a reasonable conclusion to draw.
Non-monotonicity in inference is possible whenever inferences are rea-
sonable without being entailed. So according to the classical theory, the
evidence about the external world that you get when you have a visual
experience as of something red in front of you does not entail anything
about the external world. It is, for example, that it appears to you that there is
something red in front of you, or that you have a visual experience as of
something red in front of you, that you are appeared to redly, that there is a
red sense-datum, or the like. There are many ways of developing the classical
theory by filling in a more detailed account of the nature of perceptual
evidence, according to the classical theory, but for purposes of this paper,
they share a package of virtues and vices.
The classical theory is, in the first instance, a theory about subjective defeat.
This is because non-monotonicity is a property of reasonable inference, and
only subjective defeat concerns reasonable inference, at least directly. The key
idea of the classical theory is that it is reasonable to infer that there is
something red in front of you from the evidence that it appears that there is
something red in front of you, but not from the evidence that you are wearing
rose-colored glasses and it appears that there is something red in front of you.
But the classical theory can also be extended to offer a treatment of objective
defeat. This was essentially the project of defeasibility analyses of knowledge,
such as those offered by Klein (1970); Lehrer and Paxson (1969); Annis (1973);
Ackerman (1974); Olin (1976); Levy (1977). What all of these theorists noted
was that there seems to be a general pattern whereby knowledge has to be not
    273
only reasonable belief, but belief whose reasonability “stands up” to the facts,
in some way.³ The project explored in their papers was how to say more
precisely and in a way free from counterexamples exactly how knowledge must
be belief whose reasonability “stands up” to the facts, but all of their accounts
agree that this is what goes wrong in cases of objective defeat like Roberta’s
that are not also cases of subjective defeat. In such cases, Roberta may
reasonably believe that there is something red in front of her, and this belief
may be reasonable because she does not realize that she wearing rose-colored
glasses, but it is only reasonable because she is not aware of this. If you add this
proposition to her stock of beliefs, it would no longer be reasonable for her to
believe that there is something red in front of her, and that is why, on this view,
its truth defeats her knowledge.
The strength of the classical theory is the elegance with respect to which it
subsumes the defeasibility of basic perceptual evidence under the defeasibility
of evidence elsewhere, by incorporating it into a general account of non-
monotonic inference. Indeed, the literature on nonmonotonic inference in
general is replete with examples that take for granted that the classical theory is
correct about the defeasibility of perceptual evidence. Its home case is the case
of subjective defeat, and although defeasibility theorists in the 1970s famously
discovered that it was hard to pin down exactly what the “standing up to the
facts” condition on reasonable belief is that is required for knowledge, so it is
intelligible to worry that they will not actually be able to adequately capture all
and only the correct cases of objective defeat, classical theorists at least have a
well-understood strategy for trying to pin these cases down.
12.2.2 Consequences of the Classical Theory
However, the weakness of the classical theory has been regarded by many to be
the key assumption about the nature of perceptual evidence that is required in
order for it to subsume perceptual defeasibility to the monotonicity of infer-
ence more generally. And this is the assumption that perceptual evidence does
not entail anything about the external world. This assumption, in essence, set
all of the central problems of epistemology in the twentieth century, and
shaped the space of possible options.
It is the assumption that perceptual evidence does not entail anything about
the external world from which it follows that there are skeptical scenarios that
³ Compare Schroeder (2015).
274  
are completely consistent with the totality of all perceptual evidence. If

perceptual evidence could entail things about the external world, then no
such skeptical scenarios would be possible, after all. There would still be
scenarios in which we are all brains in vats, or deceived by an evil demon, of
course, but these scenarios would be inconsistent with the totality of our
perceptual evidence, and so the skeptic’s claim that these are compatible
with all of our evidence would be false. So, in a very real sense, classical
theory’s assumption is responsible for a very important aspect of the allure
or threat of skepticism about the external world.
The same gap that is occupied by skeptical scenarios is one that non-
skeptics, of course, would like to close. And most of the rest of twentieth-
century epistemology can be conceived of as responses to how to close this
gap. If a bridge premise is needed to close the gap, for example, then it seems
that that bridge premise must be either empirical or a priori. If it is empirical,
then it must in turn be justified by perceptual evidence, but since by parity of
reasoning, that would again require a bridge premise, this path leads toward
coherentism. And this is, in fact, the most pressing motive toward adopting
coherentism. In contrast, if the bridge premise is a priori, then it turns out that
a priori justification is a pre-requisite to empirical knowledge, which amounts
to a very strong kind of rationalism.
The way to resist both coherentism and rationalism without retreating into
skepticism, therefore, and without giving up on the classical theory, is to hold
that no bridge premise is required, in order to make the inference from percep-
tual evidence to beliefs about the external world, even though these conclusions
are not entailed. This is what dogmatists claim. But in order to avoid skepticism,
dogmatists must claim that inferences from perceptual evidence to conclusions
about the external world are more reasonable than inferences to conclusions
about the intentions of the evil demon, or about the locations of ones and zeros
on the hard-drive of the matrix. Since nothing in the content of the perceptual
evidence breaks this symmetry, something else must be appealed to in order to
explain the break in symmetry, and this is where epistemological externalism
originally came in, in the late 1960s and early 1970s—it looked like every
non-skeptical, non-coherentist, non-rationalist view had to give some causal,
subjunctive, reliabilist, or otherwise fundamentally externalist answer to what
justifies moving from basic perceptual evidence to external world beliefs over
alternative conclusions. This was, as Armstrong (1973) argued for these reasons,
inevitable. And so, externalists argued, everyone needed to accept the adequacy
of such externalist explanations, which obviated the need to explain knowledge
or the reasonability of belief in terms of evidence at all.
    275
I don’t mean to claim here that any of these views—skepticism, coherent-

ism, rationalism, or pure externalism—are false, let alone that they are so
obviously problematic that we need to reject the classical theory of the
defeasibility of perceptual evidence in order to avoid them. But the fact that
this familiar dialectic and set of choices are a consequence of the classical
theory is the most important thing to understand about the classical theory, in
order to come to grips with whether it is a package that is worth accepting. So,
let’s turn to what some of its alternatives might look like.
12.3. Doubly World-Implicating Views
Since the classical view both requires and is enabled by the assumption that
perceptual evidence does not entail anything about the external world, its
alternatives adopt the assumption that perceptual evidence does entail things
about the external world. They assume, as I will put it, that perceptual evidence
is world-implicating. But world-implicating views disagree amongst them-
selves, as we will see, about exactly what sort of content perceptual evidence
is, and also about what sort of condition a subject must satisfy, in order to
possess this sort of evidence.
Recall that classical theorists believe that perceptual evidence does not entail
anything about the external world. But classical theorists also believe that being in
possession of some particular piece of perceptual evidence does not entail anything
about the external world. If your perceptual evidence that there is something red in
front of you is that it appears to you that there is something red in front of you, then
you count as having this perceptual evidence in virtue of it’s being true that it
appears to you that there is something red in front of you. Or perhaps you count as
having this evidence in virtue of believing that it appears to you that there is
something red in front of you. Either way, not only does the content of your
evidence not entail anything about the external world, the condition of your having
that evidence does not entail anything about the external world, either.
The first class of world-implicating views that I want to consider disagrees
with the classical theory on both of these fronts. These views claim not only
that the content of your perceptual evidence is world-implicating, but that the
relationship that you must stand in to that evidence in order for it to be
evidence that can make beliefs reasonable for you is world-implicating as well.
Let us call these views doubly world-implicating.
One example of a doubly world-implicating view of perceptual evidence
comes from Williamson (2000). Williamson holds that your evidence is what
276  
you know (E = K), that knowledge is the most general factive stative attitude,
and that seeing that there is something red in front of you is a factive stative
attitude. So it follows that Williamson holds that when you see that there is
something red in front of you, that makes available to you as part of your
evidence the fact that there is something red in front of you. The content of
your evidence is a proposition about the world outside of your head, so this
view is world-implicating along the content dimension. But the condition of
this being your evidence is also world-implicating—it is that you see that this is
the case. And seeing that something the case is a factive relationship to the
world—a relationship that only holds if there really is something red in front
of you. Let us call this view the factive content view.
Another example of a doubly world-implicating view comes from an
interpretation of some of John McDowell’s views offered by Comesaña and
McGrath (2015). According to this view, which I’ll call the factive attitude
view, in normal cases your evidence that there is something red in front of you
is that you see that there is something red. And you come to have this reason
because it is true—you really do see that there is something red in front of you.
The factive attitude view shares with the factive content view a theory about
the conditions under which you have perceptual evidence about the world but
disagrees about what that evidence is. But like the factive content view, it holds
that both the content of your evidence and the condition of your having that
evidence entail something about the external world. Both entail something
about the external world because seeing that is a factive relation, and so you
cannot see that there is something red in front of you unless there really is
something red in front of you. So, it is doubly world-implicating.
Doubly world-implicating views avoid the consequences of the classical
theory explored in Section 12.2.2. But they do so by undermining the classical
theory’s elegant subsumption of the defeasibility of basic perceptual evidence
to the nonmonotonicity of inference more generally. The inference from ‘there
is something red in front of me’ to ‘there is something red in front of me’ is not
nonmonotonic, and neither is the inference from ‘I see that there is something
red in front of me’ to ‘there is something red in front of me.’ This non-
monotonicity is precisely what avoids the gap between perceptual evidence and
conclusions about the external world that the skeptic exploits and that the
coherentist, rationalist, and externalist all try to close, so there is no avoiding
the conclusion that avoiding the consequences of the classical theory also means
giving up on its elegant treatment of the defeasibility of perceptual evidence.
But doubly world-implicating views—at least, both of the ones that I’ve
described here—come with an alternative treatment of perceptual defeasibility.
    277
And that is because someone who is wearing rose-colored glasses cannot see
that something is red. Seeing that something is red does not only entail that it
really is red—it is also incompatible with having insufficient perceptual
discriminability to discriminate red from other colors. Since rose-colored
glasses interfere with perceptual discriminability, wearing them is incompat-
ible with seeing that something is red. So, if you are wearing rose-colored
glasses, you do not know that something is red, just by having a visual
experience as of it being red.⁴
12.3.1 Consequences of Doubly World-Implicating Views
Recall that the classical theory is most at home in its treatment of subjective
defeat. That is because what it explains, in the first instance, is why certain
inferences are unreasonable, in the presence of further background beliefs. In
contrast, we can now see that doubly world-implicating views are most at
home in their treatment of objective defeat. That is because these views impose
an objective condition on the subject’s relationship to the world, in order for
her to have a reason at all—that she sees that something is the case. And seeing
that something is red is undermined by wearing rose-colored glasses.
But subjective defeat involves background beliefs that might (at least, when
subjective defeat does not overlap with objective defeat) be false. Having a
background belief that you are wearing rose-colored glasses does not prevent
your vision from actually working properly. So, it does not prevent you from
seeing that there is something red in front of you. So, whereas the doubly
world-implicating views get an immediate explanation of why cases of object-
ive defeat are cases in which the subject lacks reasons for her belief (and hence
don’t know), they have no such obvious explanation for why cases of subject-
ive defeat are cases in which agents lack reasons for their beliefs.
⁴ It is worth considering how this strategy might be extended to fake barn cases, which bear many
similarities to paradigmatic cases of undercutting defeat but are slightly more contentious. Fake barn
cases are more contentious because some hold that visual knowledge of a barn is possible even in fake
barn country. But the strategy of doubly world-implicating views can be extended to fake barn cases so
long as either fake barn country undermines both knowing that and seeing that, or it undermines
neither. Fake barn country cases could be an obstacle to this strategy if fake barn country undermines
knowing that without undermining seeing that—a possibility that is opened if we reject, as I do
(Schroeder (Manuscript, ch. 4)), the claim that seeing that entails knowing that. But even if we deny
that seeing that entails knowing that I still believe that we should either accept fake barn country cases
as undermining both, or as undermining neither. The reasons that seeing that does not entail knowing
that are that seeing that something is the case does not entail either belief or reasonable belief that it is
the case, but knowledge does, and this cross-cuts considerations about fake barn cases. Compare
Prichard (2012).
278  
However, defenders of doubly world-implicating views can say something

derivative about cases of subjective defeat. They can say, for example, that
although in cases of subjective defeat, the subject really does know, perhaps it
is reasonable for her to believe that she does not know, or that she probably
does not know. Or they can say, either in addition or as an alternative, that it is
blameworthy for the subject to believe. Alternatively, they can defend the
principle that knowledge is incompatible with reasonably believing yourself
not to know. On this alternative view, since cases of objective defeat are
incompatible with knowledge, and cases of subjective defeat involve reason-
ably believing yourself to be in a case of objective defeat, cases of subjective
defeat will be derivatively incompatible with knowledge. Like the method by
which classical theorists attempt to transfer their account of subjective defeat
over to objective defeat, it is intelligible to doubt whether any of these accounts
will quite work out, but it is very clear that there are a range of overlapping
strategies available here.
More worrisome is the way in which doubly world-relative accounts explain
the defeat of knowledge in the objective case. They explain it by showing that
the subject has no reason at all to believe, in those cases. But reasons seem to be
more directly connected to reasonability than to knowledge. So doubly world-
relative accounts seem to predict not only that someone whose perceptual
belief suffers from objective defeat does not know, but that in the absence of
other evidence, her belief is not reasonable, either. They distinguish, that is,
between ‘good’ and ‘bad’ cases of perceptual belief not only with respect to
whether the subject knows, but also with respect to whether, and why, the
subject’s belief is reasonable.
Now, it is open to doubly world-relative accounts to give an alternative
account of the evidence that is available to a subject who suffers from objective
defeat, in order to explain why her belief is reasonable after all, even though
she lacks the evidence that is possessed by someone in the good case.
A theorist who goes this way is an epistemological disjunctivist—that is, she
tells different stories about the reasonability of perceptual beliefs in the good
and bad cases. Another way of being an epistemological disjunctivist is to deny
that perceptual beliefs are reasonable in the bad case of objective defeat at all
and explain why they seem reasonable by pointing out that the subject does not
know that they are not reasonable.
So doubly world-relative views of the defeasibility of perceptual evidence
have twin vices—they have to explain why knowledge seems to be defeated in
subjective defeat cases, and they have to explain why reasonability seems not to
be defeated in objective defeat cases, and if they do try to explain why beliefs
    279
can be reasonable in objective defeat cases, they must do so without under-

mining their own explanation for why the subject doesn’t know, in such cases.
These are clearly different costs and advantages from the classical theory.
12.4. Singly World-Implicating Views
The classical theory and doubly world-implicating views disagree about two
things. They disagree about whether the content of perceptual evidence can be
world-implicating, and they disagree about whether the condition that a
subject must satisfy, in order to have perceptual evidence, can be world-
implicating. But the features of their view that doubly world-implicating
views cite as their advantage over the classical theory derive from their view
about the content of perceptual evidence, and the features that cast their
treatment of defeasibility into doubt derive from their view about the condition
requires to possess that evidence. The idea behind singly world-implicating
views is to split this difference by adopting the view that the contents of
perceptual evidence are world-implicating but denying that the condition
required in order to possess this evidence is world-implicating.
Let’s take these points one at a time. To see that the advantages of the
doubly world-implicating views turn on what they say about the content of
perceptual evidence, it suffices to recall that the problem with the classical view
was that it required the assumption that skeptical hypotheses can be compat-
ible with the totality of your perceptual evidence. It is necessary and sufficient
to face this problem that the content of perceptual evidence—the proposition
such that your perceptual evidence can be accurately reported as being that
p—entails something about the external world.
In contrast, the problems for the doubly world-implicating views turn on
what they say about the condition of having perceptual evidence. Since they
say that this condition entails something about the external world, it follows
that this condition cannot be shared across good and bad cases, and hence that
the story about why belief is reasonable must be different across the good and
bad cases, since the same evidence is not available to justify it. In contrast, if
the condition under which a thinker counts as having a certain piece of
evidence entails nothing about the external world, then it will cross-cut the
good case/bad case distinction, and so we will be able to tell a non-disjunctive
story about why thinkers are reasonable to believe in the way that they do.
So that is what motivates singly world-implicating views. But there is one major
reason why philosophers have not explored singly world-relative views. And that
280  
is that it is common to assume that evidence must be true. If evidence must be true,
then singly world-implicating views are impossible. For, if evidence entails some-
thing about the external world, and having that evidence entails that it is true, then
the condition of having the evidence must entail something about the external
world as well. So, proponents of singly world-implicating views must reject the
view that evidence (or for that matter, reasons) must be true.
This can sound bizarre, but it is not such an unnatural view. We may
distinguish between objective and subjective reasons, and correspondingly
between objective and subjective evidence. Objective reasons count in favor
of what it is advisable to believe, but subjective reasons count only in favor of
what it is reasonable to believe. Objective reasons must be true, because what it
is advisable to believe (or do) depends on what is true, but subjective reasons
need only be believed (perhaps rationally), because what it is reasonable to
believe (or do) depends only on what you (rationally) believe.
12.4.2 Comparing Singly World-Implicating Views
We can distinguish between two obvious varieties of singly world-implicating

views of the defeasibility of perceptual evidence—one analogue of each of the
doubly world-relative views considered in Section 12.3. Recall that according
to the factive content view, your perceptual evidence when you see that there is
something red in front of you is that there is something red in front of you, and
the condition under which this is your evidence is that you see that there is
something red in front of you. If we relax the condition that the perceptual
condition for having this evidence must be a factive perceptual relation, we are
led to the non-factive content view, according to which your perceptual
evidence is the same—namely, that there is something red in front of you—
but the condition under which this counts as your evidence is that you have a
visual experience as of something red in front of you.
Similarly, recall that according to the factive attitude view, your perceptual
evidence when you see that there is something red in front of you is that you
see that there is something red in front of you, and the condition under which
this is your evidence is that it is true. Again, we can relax the world-implicating
condition of truth that this view requires in order for this to be part of your
evidence, and replace it with some other, more subjective condition, such as
that it seems true to you. Let us call this the non-factive attitude view.
Of all of the views that have been discussed here, the non-factive content
view offers by far the least promising account of the defeasibility of perceptual
    281
evidence. Let us take the case of objective defeat first. In this case, you have a
visual experience as of something red in front of you, but unbeknownst to you,
you are wearing rose-colored glasses. According to the non-factive content
view, you have available as evidence the proposition that there is something
red in front of you. But this evidence is both available to you, and true. So, it is
hard to see why it can’t ground knowledge that there is something red in front
of you, especially because you have no contrary evidence. But it shouldn’t be
possible to know, on the basis of visual evidence, that there is something red in
front of you, if you are wearing rose-colored glasses. The non-factive content
view has no obvious answer to this problem. It loses the elegant answer that the
doubly world-relative views gave to this problem, because it relaxes the
constraint on having evidence too far.
The non-factive content view also fails to adequately treat cases of subjective
defeat. In this case, you have a visual experience as of something red in front of
you, but you also realize that you are wearing rose-colored glasses. According
to the non-factive content view, as with the doubly world-implicating views,
this latter belief does not interfere in any way with your satisfying the
condition required to have world-implicating evidence—in this case, your
evidence being that there is something red in front of you. The doubly
world-implicating views, as we saw, might attempt to try to explain why it is
nevertheless not rational to believe that there is something red in front of you
in a way that was derivative from their treatment of objective defeasibility. But
as we’ve just seen, the non-factive content view does not have a successful
treatment of objective defeasibility. And so, it can’t borrow this strategy either.
In contrast, the non-factive attitude view offers elegant, parallel, treatments
of both objective defeasibilty and subjective defeasibility. It is able to do this, in
effect, because it borrows more from the doubly world-implicating views—it
takes on board their idea that factive perceptual relationships to the environ-
ment play an important role in basic perceptual evidence. The fact that you are
wearing rose-colored glasses defeats your visual evidence that there is some-
thing red in front of you, on this view, because it cannot be true that you see
that there is something red in front of you if you are wearing rose-colored
glasses. This suffices to defeat your knowledge, because knowledge cannot be
based on false lemmas. But it doesn’t defeat the reasonability of your belief,
because reasonability of belief is not defeated by false lemmas. So, this is
exactly the right kind of explanation of the objective defeat to avoid leading
to epistemological disjunctivism.
The non-factive attitude view also offers an elegant treatment of subjective
defeat. In this case, you believe that you are wearing rose-colored glasses, but
282  
part of your evidence about the world is that you see that there is something
red in front of you. But as we’ve been reminded again and again throughout
this paper, these claims cannot both be true. So, either you should give up your
belief that you are wearing rose-colored glasses, or you shouldn’t infer any-
thing from the proposition that you see that there is something red in front of
you. So either your belief that you are wearing rose-colored glasses is not
rational, or, if it is, it is not reasonable to infer anything from this particular
piece of perceptual evidence, and hence it is defeated as a sort of reasonable
belief about the world.
The non-factive attitude view carries other, substantive, commitments,
however. Of the views that I have surveyed, here, it is the only view which is
committed to the possibility that creatures who can have perceptual evidence
about their environments must also be able to represent themselves as seeing
that something is the case. This can seem like a particularly stringent kind of
cognitive demand. But perhaps it is not so implausible. Imagine seeming to see
your friend Bob in front of you, and then learning that what you are in fact
seeing is a hologram of Bob that is triggered by Bob standing in front of you. If
this sounds like a case of an illusion, then we must say what about your
experience makes it illusory. What makes it illusory, a proponent of the
non-factive attitude view may claim, is that it represents itself as a factive
relation to your environment.⁵
12.5. Summing-Up
We’ve now explored four accounts of the defeasibility of perceptual evidence:

the classical account, the two doubly world-implicating views, and the non-
factive attitude view. (The non-factive content view, though it occupies a place
in the logical space occupied by the other views, can hardly be described as an
account of the defeasibility of perceptual evidence, since its problem is pre-
cisely that it has no such account.) Classical views have the advantage of fitting
seamlessly into general accounts of non-monotonic inference but carry with
them a commitment to a restricted space of possible options in general
epistemology. Their challenge is to extract an adequate treatment of objective
defeat from their elegant treatment of subjective defeat. Both doubly world-
implicating views lead to commitments about the differences in what explains
reasonable belief in good and bad cases. Their challenge is to extract an
⁵ This argument extends an argument from Searle (1983).
    283
adequate treatment of subjective defeat from their elegant treatment of object-

ive defeat. The non-factive content view shares neither of these commitments,
and it offers parallel elegant treatments of both objective and subjective defeat.
Are those sufficient reasons to think that it is true? Sufficient, I think, to
wonder, at least in the absence of defeating evidence.⁶
References
Ackerman, Terrence. (1974). Defeasibility modified. Philosophical Studies, 26

(5–6): 431–5.
Annis, David. (1973). Knowledge and defeasibility. Philosophical Studies, 24(3):
199–203.
Armstrong, David. (1973). Belief, Truth, and Knowledge. Cambridge: Cambridge
University Press.
Barker, John A. (1976). What you don’t know won’t hurt you? American
Philosophical Quarterly, 13(4): 303–8.
Comesaña, Juan, and Matthew McGrath. (2015). Perceptual reasons.
Philosophical Studies, 173(4): 991–1006.
Johnsen, Bredo. (1974). Knowledge. Philosophical Studies, 25(4): 273–82.
Klein, Peter. (1971). A proposed definition of propositional knowledge. Journal of
Philosophy, 68(16): 471–82.
Lehrer, Keith, and Thomas Paxson. (1969). Knowledge: Undefeated justified true
belief. Journal of Philosophy, 66: 225–37.

Levy, Stephen. (1977). Defeasibility theories of knowledge. Canadian Journal of
Philosophy, 7(1): 115–23.
Olin, Doris. (1976). Knowledge and defeasible justification. Philosophical Studies,
30(2): 129–36.
Prichard, Duncan. (2012). Epistemological Disjunctivism. Oxford: Oxford
University Press.
Schroeder, Mark. (2015). Knowledge is belief for sufficient (objective and subject-
ive) reason. Oxford Studies in Epistemology, 5: 226–52.
Schroeder, Mark. (2021). Reasons First. Oxford: Oxford University Press.
Searle, John. (1983). Intentionality. Oxford: Oxford University Press.
Williamson, Timothy. (2000). Knowledge and Its Limits. Oxford: Oxford
University Press.
⁶ Special thanks to Jessica Brown, Mona Simion, and Shyam Nair.
Index
For the benefit of digital users, indexed terms that span two pages (e.g., 52–53) may, on
occasion, appear on only one of those pages.
Ackerman 272–3 Brentano 179

action 3, 13–14, 20–1, 51, 95–7, 102, 105, Broome 51n.30, 120, 212n.16,
130–1, 177–97, 247n.2, 249n.5, 215, 250n.8
253–4, 256, 258–9, 259n.28 Brown 1–2, 6–7, 20nn.5,6, 21n.10, 23n.11,
intentional 13–14, Chapter 8 passim 26nn.14,15, 32n.19, 80–3, 191n.21,
Aikhenvald 76–7, 81–2 228n.12, 236nn.25,27
akrasia 7, 116–20, 122–3 Burge 168n.24, 220
Alston 9–11, 18n.3, 58n.36
alternative reliable process theory (ARP) – Carruthers 88–9
see defeat reliabilist account of Carter 39n.1, 65n.45
Annis 272–3 Carter and Navarro 39n.1, 177, 191–3,
Anscombe 225n.5 196–7
Armstrong 274 Casullo 205n.5, 206n.8, 208–9, 221n.23
assertion 8, 20nn.6,8, 82 Cath 180n.5, 183–6
Audi 188, 209n.12 Chandler 3, 226n.8
Austin 186–7 Chisholm 41n.8, 56n.34, 149n.6, 177
Christensen 6–7, 26–7, 26nn.14,15, 88–9,
Bacon 236n.25 94n.3, 119–20, 121n.6, 124, 124n.12,
Bader 261n.30 127n.18, 128n.20, 133–4, 140,
Baker-Hytch and Benton 5, 19n.4, 23n.12, 157n.17

26n.16, 94nn.1,3, 239n.30 closure 2, 106, 162
Balcerak Jackson 213n.17, 214n.18 Coates 7, 120
Basu and Schroeder 230n.17 Coffman 186n.15
Beddor 4–5, 13, 59n.39, 60–1, 60nn.41,42, Cohen 121n.6, 151n.10
62n.43, 63n.44, 65n.45, 87–8, 94n.2, coherence 18–19, 49–51, 56, 117
146, 149n.7, 153–4, 169n.28 Comesaña 58n.36, 148n.5, 169nn.26,27,
Beddor and Pavese 181, 181nn.7,8, 186n.17 170n.29, 208–9, 221n.23
Bedke 149n.9, 259nn.25,26, 264n.38 Comesaña and McGrath 276
Bergmann 10–11, 50n.29, 152n.11, 177, constitutivism 118
223n.2, 233, 233n.21
Berker 112n.31, 141, 261n.29 Dancy 248n.3, 251n.12, 261n.30
blameless 97–8 Dasti and Phillips 72
blameworthy 97n.15, 278 Davidson 182–3, 188–9, 194
Boghossian 212, 214–15 debunking 1–2
BonJour 54n.33, 56n.34, 89–90, 149n.8, defeat
169n.26 doxastic 1–2, 9–12, 43n.11, 204–5,
BonJour and Sosa 89–90 223n.2
Bratman 188 evidentialist Chapter 3, 168
286 
defeat (cont.) epistemic norm 1, 18, 93–4, 227, 234–7

higher-order chapter 1, 26–7, 57n.35, Evans 77–8
94n.7, 103, 107, 113–14, 125–6, 138–9, evidence
142, 157n.17, 208–9, 217–18, 220–1, first-order 26, 116–17, 121, 124n.11,
224, 230–2, 234–5, 238 127–8, 132, 138, 231
normative 1, 4, 8–12, 30–1, 35–6, 39–40, higher-order 2, 6–7, 12–13, 19n.4,
43–5, 47n.22, 52, 54–7, 60–1, 237 26n.15, 93, 94nn.3,7, 100–1, 103,
objective versus subjective 14–15, 220, 105–10, 113–14, chapter 6 passim, 231
chapter 12 passim evidentialism 3–5, 11–12, chapter 3
of reasons 3–4, 40, 57–9, 168, 171, 173, passim, 118–19, 124, 126–9, 128n.21,
204–5, 207, 225n.5, 228 133–4, 148, 168–73
partial versus full 9, 42, 225–6, 229, excuse 97–8, 236n.27
236–7, 237n.28 externalism 1, 3–5, 9–11, 74n.3, 79–80, 84,
psychological – see defeat doxastic 89–90, 97–8, 163, 174, 208–9, 229n.15,
rebutting 8–9, 13, 41–4, 48–9, 53–4, 59, 274–6
59n.37, 157, 166–7, 177, 202n.2,
203–4, 208–9, 224–5, 248n.3, 249, fallibilism 70, 72, 73n.2, 74n.3, 76–7, 80
251–2, 255, 260 Fantl and McGrath 229n.16, 17
reliabilist account of 4–5, 13, 39–40, 42, Feldman 6–7, 63–4
57–64, chapter 7 passim Feldman and Conee 59n.38
responsibilist account of 39–40, 42, 50, Finlay 259n.27, 264n.36
56n.34, 57–8, 64 von Fintel and Iatridou 264n.36, 265n.39
scepticism about 5–7, 13 Flores and Woodard 113n.34
undercutting 1–3, 8–9, 13, 41–4, 46, Frances 50n.29
48–9, 59n.37, 94n.4, 124, 132n.27, Frances and Matheson 2
157n.17, 158, 166–7, 171, chapter 9 Friedman 130n.25, 26
passim, 223, 228n.12, 230–1, 234n.23, Fumerton 150–2, 166n.23
237n.28, 248nn.3,4, 251n.12, 271, 277n.4
Dehaene 63–4 Gerken 21n.8
Di Paolo 223n.1, 232n.19, 237n.28 Gert 255–60

disagreement Ghijsen, Kelp, and Simion 236n.26
conciliatory view of 2, 122, 128, 133, 135 Gibbons 34n.20, 139n.33, 179, 180nn.4,5
equal weight view of 121–2 183n.12, 185–6, 237n.28
steadfast view of 2, 5, 121, 122n.7, 128, Ginet 181, 188
132–5, 139–40 gnosticism 105, 109–10, 113–14,
dispositions to believe 12, 111–12 236n.24, 237
Dorst 223n.1, 225n.7, 226n.10, 231n.18, Goldberg 10–11, 18, 18n.2, 20n.7, 27,
233n.20 31n.18, 39n.1, 45n.15, 17, 223n.2
Dorst, Fitelson, and Husic 231n.18 Goldberg and Matheson 34n.21, 39n.1
Dutant and Fitelson 109n.28, 110n.29, Goldman 4, 48n.25, 57n.35, 60n.40, 63,
236n.25 74n.3, 89–90, 146nn.1,2,3, 148n.5,
Dutant and LittleJohn 223, 229n.15 149nn.8,9, 151, 152n.11, 161n.20, 168,
Drayson 96n.13 168n.24, 174n.36,
Driver 259n.24 177, 181n.6, 188–9, 194, 226n.9
Goldman and Beddor 146n.3
Easwaran 225n.7 Gonzalez de Prado Salas 124n.11
Easwaran and Fitelson 226n.10, 229n.15 Goodman 240n.31
Elga 7, 121n.6, 127n.19 Graham 220, 236n.26
epistemic good 14, 226, 236–7 Graham and Lyons 164n.22, 174n.36
 287
Greco, Daniel 5 Kim 146n.2

Greco, John 5 knowledge-first 18, chapter 8 passim
Green 2, 39n.1 know-how 13–14, chapter 8 passim
Greenspan 253–5 intellectualist view of 13–14, chapter 8
Gregory 263n.34 passim
Grundmann 39n.1, 45n.15, 149n.9, 224 Kolodny 121
Kornblith 3–4, 59n.38, 65n.45, 146n.1
Hampshire 185 Kotzen 226n.8
Harman 10, 123n.9, 188–9, 233
Hawley 184, 190, 193 Lackey 9–12, 21n.8, chapter 3 passim,
Hawthorne and Srinivasan 5–6 104n.24, 121n.6, 152n.11, 223n.2
Henderson, Horgan, and Potrč 148n.5 Lasonen-Aarnio 5–7, 12, 23n.12,
Hlöbil 212 26nn.14,16, 51n.30, 86–7, 90n.10, 93,
Hornsby 190 94nn.1,5–10, 100n.20, 103nn.23,25,
Horowitz 7, 26nn.14,15, 112n.32 111n.30, 124n.12, 149nn.6,7, 160n.19,
Horowitz and Sliwa 7, 26n.14 174n.34, 192n.22, 223n.1, 225n.6,
Horty 249n.4 239n.30
Horty and Nair 249n.4 Lassiter 83n.7
Humberstone 178 Lazar 227n.11
Hurka 97n.15 Lehrer and Paxson 272–3
Hyman 236n.24 Leitgeb 229n.14
level-splitting 7, 231–2
ignorance 14, 34, chapter 10 passim Levy, Stephen 272–3
infallibilism 12, 70, 74n.3, 80, 85 Levy, Y 180n.4
inference 12, 14–15, 26, 51, 70–2, 76–9, Lewis 96, 178
81–3, 88–9, 123, 155–9, 160n.18, LittleJohn 139n.33, 177n.1, 181nn.7,9,
162–5, 166n.23, 211–13, 211nn.14,15, 187n.19, 225n.5, 227n.11, 236n.24,
213n.17, 214n.18, 219–20, 272–9, 236nn.25,27
282–3 Lord 116, 117n.2, 121n.6, 122n.8, 123n.10,
inquiry 35, 112–13, 133n.28, 136n.29, 223n.1, 225n.5, 237n.28, 247n.1, 250n.11
137–9, 141 Lord and Maguire 247n.1
instrumentalism 110–13 Lord and Sylvan 118, 125n.16
intention 13–14, chapter 8 passim, 215, 274 Lynch 226n.9
internalism 1, 9–11, 56, 58, 74n.3, 78–9, Lyons 3–4, 11–12, 39, 45n.15, 57n.35,
89–90, 94–5, 162–3, 173, 208 60n.40, 62, 146n.1, 149n.9, 150,
151n.10, 174n.36
James 138
Johnson, Hashtroudi, and Lindsay 76–7 Macnamara 259n.24
Joyce 154n.13 Maguire 261n.30
justification Maguire and Snedegar 261n.30
doxastic 42, 59n.38, 61–2, 64, 203 Makinson 229n.15
inferential versus noninferential 13, 202, Markovits 118
210–21 Matheson 39n.1, 48n.26
perceptual 204–5, 214n.18, 218–20 Matilal 72, 89–90
propositional 42, 51, 53, 58, 59n.38, 61–4 McNamara 259n.25, 265n.39
testimonial 13, 202, 217, 220–1 McGinn 225n.5
McGrath 201, 220n.22
Kelly 5–7, 121n.6, 122n.7 McKinnon 21n.8
Kelp and Simion 1 Mele and Moser 179, 186, 188
288 
Mellis 206n.8, 217 Pritchard 39n.1, 45n.15, 65n.45, 181n.7,

Mercier 77–8 186n.15, 187, 277n.4
Mercier and Sperber 77–8 Pryor 2, 5, 8–9, 124n.12, 127n.18, 218,
Miller, Emelia 58n.36, 148n.5, 169, 221n.23
169n.27, 172n.32
Miller, George 63–4 Railton 137n.32
Miracchi 113n.34 Raleigh 242
Monti and Osherson 88–9 rationality 9–11, 14, 18, 46–7, 49–51, 54–6,
Morton and Paul 131n.26 69–70, 94–5, 97–8, 116–19, 121, 126,
Moss 154n.13, 182–3, 190, 230n.17 133–4, 223–6, 225n.5, 231–9,
Murray 82 236nn.25–27, 242–3, 256, 258–9
Raz 249nn.4,5
Nagel Jennifer 69, 76–7, 183n.12, 186 reasons
Nagel Thomas 186, 252n.13, 253–4 exclusionary 249nn.4–6
Nair 250n.9, 261n.30 for versus against 1, 3–4, 14, 20–2, 41–3,
naturalism 5, 13, 36, 61–2, 64, 118, 146, 58, 117–19, 122–4, 124nn.13,14, 129,
150, 164 132, 132n.27, 134–5, 138–9, 141,
norm of assertion 21n.8 147–8, 157n.16, 160–3, 160n.18,
norm of belief 14, chapter 10 passim 162n.21, 166, 168–9, 187, 203–4, 207,
normativity 3, 11, chapter 2 passim, 224–5, 228, chapter 11 passim, 277
117–18, 125, chapter 11 passim rebutting 8–9, 13, 41–3, 157, 203–4,
208n.10, 224–5, 248–9, 251–2,
Olsen 227n.11 255, 260
Owens 229n.16, 17 undercutting 8, 41–3, 157–8, 202n.2,
203–4, 207n.9, 208nn.10,11, 211n.14,
Palermos 39n.1 219–21, 234–5, 248n.3, 4
Papafragou, Li, Choi, and Han 76–7 Reed 20n.6, 72, 73n.2, 82–3
Pappas 1 reliabilism 3–4, 11–13, 57–61, 59n.38, 64,
Parfit 250n.11 74n.3, 94n.2, chapter 7 passim
Pavese 177, 180n.5, 181–2, 183nn.11,12, responsibilism 57–8

184, 185n.14, 189–90, 189n.20, 193 Rosenkranz 131n.26
Peacocke 230 Ross 247n.1, 250n.7
Pendlebury 252n.14 Ryan 229n.14
perception 12, 14–15, 24, 58, 70–2, 76–82, Ryle 190, 193
86–8, 90, 99, 104, 161, 162n.21, 164–5,
185, 218, 240, chapter 12 passim Scanlon 121
perceptual evidence 14–15, 79n.4, Schaffer 96n.12
chapter 12 passim Schiffer 78–9
Pettigrew 148n.5, 154n.13 Schoenfield 23n.12, 26nn.14,16, 94n.3,
Plantinga 9–11 112n.33, 117n.3, 122, 127nn.18,19
Pollock 3–4, 8–13, chapter 3 passim, 147, Schroeder 124–6, 124n.13, 229nn.16,17,
147n.4, 149n.6, 155–9, 160n.18, 248n.3, 250n.11, 252n.13, 253–4,
162–5, 171, chapter 9 passim, 224, 263n.34, 269, 269n.1, 273n.3, 277n.4
224n.3, 225n.6, 229n.14, 248 Searle 282n.5
Pollock and Cruz 40n.4, 41, 41nn.7,9,10, Setiya 183n.11, 184, 190, 193
177, 201n.1, 202–4, 204n.3, 210 Shah 125n.16
Pomerantz 76–7 Sher, George 97nn.14–16
Portner 265n.39 Sher, I 261nn.30,31
practical reasoning 1–3, 5, 7, 247 Silva 1–2, 232n.19
Prakken and Horty 147n.4 Silverstein 263n.34
 289
Simion 1, 9–10 Titelbaum 120, 135

Simion, Kelp, and Ghijsen 98n.17 transmission 2, 73–4, 82–3, 179, 254n.16
Simons 82
Singer 137n.32 Unger 236n.24
Skipper 26n.14
Sloman 259n.27, 264n.36 Van Wietmarschen 85, 232n.19
Smith 229n.14, 234n.22, 235–7 Velleman 125n.16, 188
Smithies 74n.3 veritism 105, 234–5
Snedegar 124n.13, 128n.21, 136n.29, 247, Von Fintel and Gillies 82–3
249, 253n.15, 259nn.25,26, 263n.34,
264nn.38,39 warrant 2, 9–12, 20nn.6,8, 26n.15, 58–65,
Sosa 121n.6, 125n.16, 141–2, 181, 164n.22, 232–3
181n.7, 187 Way 117n.2, 262n.32, 263n.34
Spohn 224n.3 Way and Whiting 120
Stalnaker 172n.30, 178 Weatherson 5, 121–2, 230n.17,
Stanley 184n.13, 189–91 234n.22
Stanley and Williamson 184n.13, 189–91 Wedgwood 125n.16
Stich 178n.2 Weiner 21n.8
Sturgeon 9, 130n.22, chapter 9 passim, 225n.7 Weisberg 5, 182–3
Sudduth 1, 10 Wertheimer 265n.39
suspension 12–13, 69–70, 108–10, 117–19, Whiting 21n.8, 249nn.4,5
121, 125–6, 128–35, 138–9, 154, 242 Williams, Bernard 186, 224n.4
Swinburne 10 Williams, Michael 54, 56n.34
Sylvan 116, 117n.3, 123n.10, 124n.13, Williamson 6, 18n.1, 21n.9, 23n.12, 79n.4,
139n.33, 141n.35, 226n.9, 239n.30 178n.3, 180n.4, 181–2, 183n.12, 187,
Sylvan and Sosa 173n.33 191, 231n.18, 233n.20, 236nn.24,27,
240–1, 243, 275–6
Tang 58n.36, 148n.5, 169n.27, 170 Worsnip 51n.30, 121–3, 134–5,
testimony 229n.15
anti-reductionism 2, 20
reductionism 2, 20 Ye 26nn.14,15, 85n.8

Reasons, Justification, and Defeat. Jessica Brown and Mona Simion (Ed.)

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Reasons, Justification, and Defeat. Jessica Brown and Mona Simion (Ed.)

Uploaded by

Copyright:

Available Formats

OUP CORRECTED AUTOPAGE PROOFS – FINAL, 22/2/2021, SPi

Reasons, Justiﬁcation, and Defeat

Printed and bound in the UK by

List of Contributors vii

7. Reasons for Reliabilism 146

Bob Beddor, National University of Singapore

(1) The nature of justiﬁcation and knowledge: since knowledge and

justiﬁcation are taken by many to be defeasible, the extent to which

2     

evolutionary explanations of morality. It is argued that some, if not all,

It is useful to categorize contributions to the defeat literature as falling under

1.2. The Nature and Extent of Defeat

1.2.1 Defeaters as Reasons

accommodate what can be called justifying defeaters. As the name suggests, a

4     

reliabilism deny that reasons are substantive epistemic normative categories to

1.2.2 Defeaters as Reliable Processes

Reliabilist theories of justiﬁcation have been extremely popular in the last

tion, Beddor offers the following case:

Thinking about Unger: Harry sees a tree in front of him at t. Consequently, he

Here’s Beddor’s case against ARP’s necessity direction:

1.2.3 Defeat Scepticism

We have seen that, within epistemology, it is standardly assumed that a subject

6     

1.2.4 Higher-Order Defeat

Higher-order evidence is evidence about what evidence one possesses or what

8     

1.3. Kinds of Defeaters

of the truth of P. Hence E* also acts as an undercutting defeater, since it

1.3.2 By Normative Status

An important question addressed in the literature on defeat concerns the

10     

justiﬁcation, i.e. it implies justiﬁcation internalism. If this is so, externalists

1.3.3 By Psychological Status

Another important question for taxonomizing defeat concerns whether defea-

internalism’s lights, justiﬁcation supervenes exclusively on facts internal to the

1.4. Summary of the Volume

nature and scope of defeat. Goldberg’s strategy is to focus on cases in which

12     

in response to higher-order evidence when that evidence is not misleading

epistemologically substantial claim—i.e. that the epistemic proﬁle of know-

14     

to McGrath, for noninferential justiﬁcation, the Pollockian account is in

restricted space of possible options in general epistemology. Their challenge is

Alston, W. (1989). Epistemic Justiﬁcation. Ithaca: Cornell University Press.

16     

Goldberg, S. (2016). On the epistemic signiﬁcance of evidence you should have

Pappas, G. (2017). Internalist vs. externalist conceptions of epistemic justiﬁcation.

Weisberg, J. (2009). Commutativity or holism? A dilemma for conditionalizers.

Knowledge is a normative standing. Perhaps this much is uncontroversial.

knowledge as basic, and understand normativity of all other epistemic notions

       19

knowledge at least partly in terms of the normativity of justiﬁcation, and so we

As I say, I regard the truth of both CONFIDENCE and AUTHORIZATION as

I should note both CONFIDENCE and AUTHORIZATION are relatively

AUTHORIZATION is a claim about S qua speaker; it is silent on what

to believe. Neither claim weighs in on whether it is knowledge itself that grounds

       21

I note, too, that since AUTHORIZATION states a sufﬁcient condition (not a

How might CONFIDENCE and AUTHORIZATION be used so as to cast

       23

whether this is the effect of red lighting.) Verdict: he is not in a position to

       25

alternative questions. They are: