Change in View

Principles of Reasoning

Gilbert Harman

A Bradford Book
The MIT Press
Cambridge, Massachusetts
London, England
© 1986 by The Massachusetts Institute of Technology
All rights reserved. No part of this book may be reproduced in any form by any For Olivia
electronic or mechanical means (including photocopying, recording, or information
storage and retrieval) without permission in writing from the publisher.

This book was set in Palatino by The MIT Press Computergraphics Department and
printed and bound by Halliday Lithograph in the United States of America.

Library of Congress Cataloging-in-Publication Data

Harman, Gilbert.
Change in view.
"A Bradford book."
Bibliography: p.
Includes indexes.
1. Reasoning. 2. Logic. 3. Belief and doubt.
I. Title.
BC177.H37 1986 160 85-18766
ISBN 0-262-08155-5
Chapter 1
A Plea for the Study of Reasoning

Reasoningas ReasonedChangein View

Intending to have Cheerios for breakfast, Mary goes to the cupboard.
But she can't find any Cheerios. She decides that Elizabeth must have
finished off the Cheerios the day before. So, she settles for Rice Krispies.
In the process, Mary has modified her original intentions and beliefs.
This is a very simple case of reasoned change in view, an elementary
example of reasoning. It has the following two features. First, not only
does Mary's reasoning lead her to add new beliefs to her view, so that
she comes to believe that there are no more Cheerios and that Elizabeth
ate the last Cheerios yesterday, it also leads her to give up things she
had been believing, so that she stops believing that there are Cheerios
in the cupboard and that she will have some Cheerios for breakfast.
Second, Mary's reasoning changes not only her beliefs but also her
plans and intentions. Her reasoning leads her to abandon her intention
to have Cheerios and to adopt the new plan of having Rice Krispies.
In other words, her reasoning is not only "theoretical," affecting her
beliefs, but also "practical," affecting her intentions and plans.
In saying this, I assume that Mary's reasoning can be separated into
distinct segments of practical and theoretical reasoning. I make this
assumption even though any given segment of reasoning is likely to
affect both her beliefs and her intentions, since changes in her beliefs
can affect her plans, and changes in her plans can affect her beliefs.
When Mary stops believing there are any Cheerios left, she also stops
intending to have Cheerios for breakfast. When she forms her intention
to have Rice Krispies instead, she also comes to believe that she will
be having Rice Krispies for breakfast. But I assume there is a difference
between immediate changes that are "part of" a given segment of her
reasoning and less immediate changes that are merely further effects
of that segment of reasoning.
True, it is not easy to say when a change is "part of" a given segment
of reasoning and when it is merely the result of reasoning. For example,
it is not immediately obvious whether changes in desires are ever part
 The Study of Reasoning 3
2 Chapter 1

of reasoning. (I discuss this question briefly in chapter 8.) Nevertheless, Reasoning Distinguished from Argument or Proof
in what follows I assume there is a definite difference between immediate Reasoned change in view like Mary's does not seem to have been
changes that are part of a given segment of reasoning and other less studied much except for some recent research into planning and "belief
immediate changes that are merely further effects of it. And I also revision" in the field of artificial intelligence (e.g., Doyle 1980). One
assume there is a distinction between theoretical and practical reasoning. possible cause of this otherwise general neglect is that reasoning in
These assumptions suggest a further distinction between two sorts this sense may often be conflated with reasoning in another sense,
of rules of reasoning, corresponding to two possible phases in reasoning. namely argument for, or proof of, a conclusion from premises via a
On the one hand there is often a process of reflection in which one series of intermediate steps.
thinks about one's beliefs, plans, desires, etc. and envisions various
possibilities in more or less detail. On the other hand there is the actual Old View Reasoning New View
revising of one's view, which may or may not follow such a reflection.
Maxims of reflection,as we might call them, say what to consider before
revising one's view, for example, that one should consider carefully
all the alternatives, with vivid awareness of relevant evidence of possible A Premises
consequences of contemplated courses of action. On the other hand r
what we might call principles of revision concern the actual changes to g
be made, the changes that are actually "part of" the reasoned revision, u
saying such things as that one should make minimal changes in one's
m
view that increase its coherence as much as possible while promising Intermediate Steps
suitable satisfaction of one's ends. e
Not all principles of psychological change are principles of revision n
in this sense, since not all changes are instances of reasoning. For
example, it may be that changes in desires are not instances of reasoning, Final Conclusion
although these changes can occur as a result of reasoning. Even so,
there may be general principles governing changes in desires. These Clearly, argument or proof is not at alJ the same sort of thing as
would be principles of change that were not principles of revision in reasoning in the sense of reasoned change in view. There is a clear
the relevant sense. difference in category. Rules of argument are principles of implication,
I don't want to suggest that one ever makes conscioususe of principles saying that propositions (or statements) of such and such a sort imply
of revision in changing one's view. One can reason without knowing propositions (or statements) of such and such other sort. Consider the
what the relevant principles of revision are and it may well be that following principle:
reasoning is a relatively automatic process whose outcome is not under
one's control. Modus Ponens: P and if P then Q taken together imply Q.
In the rest of this book I explore the hypothesis that there is a dif- Such a rule by itself says nothing at all in particular about belief revision.
ference between theoretical and practical reasoning and that principles It may be that some principles of belief revisions refer to such principles
of revision can be distinguished from principles of reflection and from of argument, that is, to principles of implication. It is an important
other principles of change in view. For the time being J simply assume issue in the theory of reasoning, conceived as change in view, just how
that there is something about some changes in view that makes it implica-tion or argument may be relevant to reasoning. I discuss this
reasonable to call these changes "instances of reasoning" and to call
issue at some length in chapter 2. My present point is simply to note
the relevant principles "rules of revision," distinguishing these changes
that rules of argument are not by themselves rules for revising one's
from others that are significantly different. This assumption seems ini- view.
tially plausible. I will try to show that it is also fruitful.
This difference in category between rules of implication and, as we
might say, rules of inference (rules of revision) lies behind other dif-
4 Chapter 1 The Study of Reasoning 5

ferences between proof or argument and reasoning. For example, im- Most FH's are not G's.
plication is cumulative in a way that inference may not be. In argument Y is an H.
one accumulates conclusions; things are always added, never subtracted. So, Y is a G.
Reasoned revision, however, can subtract from one's view as well as
At this point, it may well be that the conclusion is no longer made
add to it. 1n order to express this point, the artificial intelligence work
probable by all the premises; so the argument from the extended set
I have mentioned contrasts "monotonic reasoning," as in the usual sort
of premises would not be inductively valid (Hempel 1960).
of argument or proof, which is cumulative, with "nonmonotonic rea-
Rules of inductive argument would be rules of "inductive logic" as
soning," as in ordinary reasoning or reasoned revision, which is not
opposed to deductive logic. It happens, however, that there is no well-
cumulative (Doyle 1980, 1982). But, although this terminology em-
developed enterprise of inductive logic in the way that there is for
phasizes the noncumulative character of reasoned revision, it is also
potentially misleading in calling the ordinary sort of proof or argument deductive logic.
"monotonic reasoning," because proof or argument is not of the same Now, why should we think there are inductive arguments and an
category as reasoned revision. inductive logic? It is clear enough that there is something that might
be called inductive reasoning, that is, inductively reasoned change in
view. But if we clearly distinguish reasoned change in view from ar-
Induction and Deduction gument, we cannot suppose that the existence of inductive reasoning
Making a clear distinction between reasoning in the sense of reasoned by itself shows there is such a thing as inductive argument, nor can
change in view and reasoning in the sense of proof or argument can we suppose that it shows there is an inductive logic.
have a profound effect on how we view a variety of issues. For example, Indeed, if we clearly distinguish reasoning from argument, we cannot
we might be led to question whether there are such things as "inductive suppose that the existence of deductive arguments shows there is such
arguments." These would be like "deductive arguments" except that a thing as deductive reasoning, that is, deductively reasoned change
the conclusion of an inductive argument would not have to follow in view. As l have already observed, rules of deduction are rules of
logically from the premises, as in a deductive argument, but would deductive argument; they are not rules of in_ferenceor reasoning. They
only have to follow probabilistically (Black 1958). are not rules saying how to change one's view. Nor (to anticipate the
Such inductive arguments would be "defeasible," that is, adding discussion of this issue in chapter 2) are they easily matched to such
"premises" to an inductive argument might undercut the "validity" of rules. Consider again modus ponens. This principle does not say that,
the argument in a way that cannot happen with deductive arguments. if one believes P and also believes if P then Q, then one can infer Q,
For example, suppose that there is a valid or warranted inductive ar- because that is not always so. Sometimes one should give up P or if
gument of the following form. P then Q instead.
Most F's are G's. Even if some sort of principle of belief revision corresponded to this
Y is an F. logical principle, the principle of belief revision would have to be a
So, Y is a G. different principle. For one thing, the logical principle holds without
exception, whereas there would be exceptions ·to the corresponding
The "conclusion" here is not a deductive consequence of the premises; principle of belief revision. Mary believes that if she looks in the cup-
it can only be "made probable" by them. Adding premises in this case board, she will see a box of Cheerios. She comes to believe that she
can undercut the argument. Suppose that we are given the following is looking in the cupboard and that she does not see a box of Cheerios.
two additional premises: At this point, Mary's beliefs are jointly inconsistent and therefore imply
Most FH's are not G's. any proposition whatsoever. This does not authorize Mary to infer any
Y is an H. proposition whatsoever. Nor does Mary infer whatever she might wish
to infer. Instead she abandons her first belief, concluding that it is false
Now the argument is the following: after all.
Most F's are G's. Furthermore, even before Mary fails to find any Cheerios in the
Y is an F. cupboard, it would be silly for her to clutter her mind with vast numbers
 -
6 Chapter l The Study of Reasoning 7
of useless logical implications of her beliefs, such as either she will have Descriptive versus Normative Theories
Cheerios for breakfast or the moo11is made of grew cheese.
If there is a connection between standard principles of logic and My aim in this book is to contribute to the development of a theory
principles of reasoning, it is not immediately obvious. There is a gap. of reasoned revision, but I find it hard to say whether the theory I
We can't just state principles of logic and suppose that we have said want is a normative theory or a descriptive theory. A normative theory
something precise about reasoning. (I discuss the relation between logic says how people ought to reason, whereas a descriptive theory says
and reasoning in chapter 2.) how they actually do reason. The theory I envision tries to say either
Clearly, distinguishing between reasoning and argument can make or both of these things.
one skeptical of the familiar idea that deduction and induction are Actually, normative and descriptive theories of reasoning are inti-
different species of the same sort of thing. Obviously, there is deductive mately related. For one thing, as we will see, it is hard to come up
argument, but it is not similarly obvious that there is deductive rea- with convincing normative principles except by considering how people
soning. Again, it is not clear that there is such a thing as inductive actually do reason, which is the province of a descriptive theory. On
argument, although we might say there is inductive reasoning. (It might the other hand it seems that any descriptive theory must involve a
be safer, however, to speak of theoretical reasoning instead of inductive certain amount of idealization, and idealization is always normative to
reasoning, because theoretical reasoning contrasts with practical rea- some extent.
soning, which clearly exists, whereas to speak of inductive reasoning The distinction between a normative and a descriptive theory seems
may suggest a contrast with deductive reasoning, which does not ob- as clear as the thought that one might sometimes reason in a way in
viously exist.) which one ought not to have reasoned, in which case there is something
Analogous remarks also apply to the suggestion that there is such wrong with one's reasoning. So let us consider ways in which one can
a thing as the practical syllogism (Anscombe 1957, p. 57). A syllogism make mistakes while reasoning. There are at least four such ways:
is a form of argument, and although there is practical reasoning, there 1.One might start with false beliefs and by reasoning be led into
is not obviously any such thing as practical argument or logic and so further errors.
not obviously any such thing as a practical syllogism. 2. One might reach a conclusion that is perfectly "reasonable,"
Again, consider a defense of a "logic of entailment," which observes even though it happens to be mistaken.
(1) in standard logic a contradiction logically implies any proposition 3. One can be careless or inattentive; one can forget about a
at all, and (2) one is not justified in responding to the discovery that relevant consideration or fail to give it sufficient weight; one
one's view is inconsistent by inferring anything whatsoever, concluding can make mistakes in long division; one can fail to see some-
that (3) a new logic is needed (Meyer 1971). This line of thought loses thing, to remember something, to attend carefully; and so on.
plausibility if rules of inference or reasoning are distinguished from 4. One can revise one's view in accordance with an incorrect rule
rules of implication or argument. of revision, thereby violating the correct rules.
Finally, distinguishing reasoning from argument can make one worry
that the work in artificial intelligence I have previously mentioned may Mistake_s of type 1 or 2 do not seem to be errors of reasoning at all.
be hampered by the so far unsuccessful search for principles of a non- Only mistakes of type 3 and 4 seem to be errors of reasoning. Mistakes
monotonic logic, in contrast to the usual principles of monotonic logic of type 3 seem to be mistakes of "reflection," involving the violation
(McCarthy 1980; McDermott and Doyle 1980; Reiter 1980). It may be of a maxim of reflection. Mistakes of type 4 would be mistakes of
a mistake·to expect principles of reasoning to take the form of a logic. revision, involving the violation of a principle of revision.
In short, distinguishing reasoning from argument can make one sus- I envision a theory that says something about principles of revision.
picious of certain arguments for inductive logic, practical syllogisms, a One _way to try to discover the right principles of revision might be to
logic of entailment, and so on. It is unclear how work on such "logics" consider actuaJ cases in which people make mistakes of type 4 to see
might contribute to the study of reasoned revision. why they are mistakes. Through seeing when the wrong principles are
~o~lowed, we might hope to discover what the right principles are. But
~t IS not easy to find cases in which people clearly change their views
m accordance with incorrect principles of revision. It is difficult to come
8 Chapter 1 The Study of Reason.ing 9

up with an example that cannot be attributed instead to a mistaken probability, or as mistakes of type 3, involving carelessness or a failure
belief, perhaps due to carelessness, so that the mistake is of type 1 and to consider al.l the relevant possibilities. None of the fallacies clearly
possibly type 3, rather than of type 4. involves a distinctive mistake of type 4, in which a mistaken principle
We cannot simply say one makes a mistake of type 4 whenever one of change in view is followed. So, it seems that we cannot immediately
reasons "fallaciously" in accordance with an "invalid" rule for changing use the existence of such fallacies to help us discover what the correct
one's view. If we distinguish clearly between argument and reasoning, principles of revision are.
we must agree that only arguments and proofs can be valid or invalid How then are we to begin to figure out what these principles of
and that the notions of validity and invalidity have no clear application revision are? There seem to be two possible approaches. We can begin
to changes in view, except in the sense that one can make a mistake by considering how people actually do reason, by trying to figure out
about what validly implies what, a mistake that affects one's reasoning. what principles they actually follow. Or we can begin with our "in-
It is often said that there is a fallacy of "affirming the consequent," tuitions" as critics of reasoning. In either case we can then hope to
in which one reasons from the premises if P then Q and Q to the find gen~ral principles. This will almost certainly involve some ideali-
conclusion P. This would contrast with "affirming the antecedent," zation. The suggested general principles will not coincide perfectly with
that is, modus ponens. But how are we to understand the contrast? our actual practice or with our intuitions about cases. This may lead
Given a sharp distinction between reasoning and argument, we cannot us to modify the general principles, but it may also lead us to change
suppose one's reasoning is valid if it proceeds in accordance with modus our reasoning practice and/or our intuitions about what reasoning is
ponens and invalid if it proceeds in accordance with the principle of correct. This can lead to a process of mutual adjustment of principles
affirming the consequent. Modus ponens is a principle of argument or to practice and/or intuitions, a process of adjustment which can continue
implication, not a principle of reasoned revision. If there is a "fallacy" until we have reached what Rawls (1971) calls a reflective equilibrium.
here, i.t seems to involve making a type 1 mistake about what implies Furthermore, and this is important, we can also consider what rationale
what. there might be for various principles we come up with and that can
Similar remarks apply to the so-called Gambler's Fallacy. This occurs, lead to further changes in principles, practices, and/ or intuitions.
for example, in the game of roulette, in which one bets on where a To repeat, even if we start by considering how people actually do
spinning wheel with a pointer will stop. The pointer might end up on reason, our account will probably have to involve a certain amount of
red or black (or very occasionally on green) and is equally likely to idealization. Now, some kinds of idealization yield a normative theory,
stop with the pointer on red as it is to stop with the pointer on black; a notion of how one would reason if everything went right. So even
that is, each time, the probability of red is the same as the probability this approach may yield a natural distinction between is and ought,
of black. The Gambler's Fallacy consists in thinking that, under these between how things do happen and how they ought to happen. Indeed,
conditions, red and black should each occur about half the time in any it may do so in more than one way.
sufficiently long series of spins; so, i.f black has come up ten times in In what follows I consider matters from both the viewpoint of our
a row, red must be highly probable next time. This is a fallacy since intuitions as critics and the viewpoint of our actual practice. These two
it overlooks how the impact of an initial run of one color can become approaches yield somewhat different results. As we will see, the appeal
more and more insignificant as the sequence gets longer. If red and to intuition tends toward a greater degree of idealization. In particular,
black occur each about half the time in a long enough sequence, they it tends to overlook or minimize practical limitations, such as limitations
also occur about half the time in the somewhat longer sequence that on memory or on calculative capacity. What seems wrong when these
includes ten extra occurrences of black at the beginning. For example, limitations are not taken into account may be quite reasonable when
suppose red and black each occur 50% of the time in a sequence of they are taken into account. So the two approaches can seem at least
1000 spins of the wheel. Then in the longer sequence obtained by initially to yield di.fferent results.
adding ten occurrences of black at the beginning, red and black each
occur within half a percent of 50% of the time. Human versus Artificial Reasoning
It may well tum out that all "fallacies" are best thought of either as
mistakes of type 1, namely, reasoning from false beliefs in which the I am concerned with human reasoning, given the constraints of human
beliefs happen to be beliefs about what implies what or beliefs about psychology. Although occasionally I allude to work in artificial in tel-
10 Chapter 1
ligence, l am here concerned with this work not for its own sake but
only for the light it may shed on human reasoning.
My conclusions about human reasoning may sometimes be relevant
to work in artificial intelligence. For example, I argue in chapter 3 that
people cannot do much probabilistic reasoning because of a combi- Chapter 2
natorial explosion such reasoning involves. If this is correct, the same
limitation will apply to the "reasoning" of artificial intelligence systems. Logic and Reasoning
But human reasoning is affected by other limits to which artificial
intelligence may not be subject, for example, limits on short-term mem-
ory. Seeing how humans reason in consquence of these limits may or
may not be of much interest for artificial intelligence.
Even if they agree that logic is not by itself a theory of reasoning, many
people will be inclined to suppose that log~c has some sort of special
Summary relevance to the theory of reasoning. In this chapter I argue that this
I am concerned with reasoned change in view. Such reasoning may inclination should be resisted. It turns out that logic is not of any special
involve giving up things previously accepted as well as coming to relevance.
accept new things. I assume there is a difference between theoretical
reasoning, which immediately modifies beliefs, and practical reasoning,
Implications, Inconsistency, and Practical Limits
which immediately modifies plans and intentions. l also assume we
can distinguish maxims of reflection, saying what to think about before If logic does have special relevance to reasoning, it would seem that
revising one's view, from principles of revision, the rules concerning its relevance must be captured at least roughly by the following two
the actual revision to be ma.de. principles.
Reasoning in the sense of reasoned change in view should never be
Logical Implication Principle The fact that one's view logically
identified with proof or argument; inference is not implication. Logic
implies P can be a reason to accept P.
is the theory of implication, not directly the theory of reasoning. Al-
though we can say there is inductive reasoning, it is by no means Logical Inconsistency Principle Logical inconsistency is to be
obvious that there is any such thing as inductive argument or inductive avoided.
logic. Nor does the existence of practical reasoning show there is such
These are distinct principles. Suppose one believes both P and also if
a thing as a practical syllogism or a practical logic.
P then Q. Since these beliefs imply Q, the Logical Implication Principle
Fina11y, it is not at this point easy to distinguish a descriptive theory
says this may give one a reason to believe Q. It does not say one should
of reasoned revision from a normative theory. Any normative inves- also refrain from believing Q's denial, not Q. Believing not Q when
tigation must begin by considering how people actually do reason and one also believes P and if P then Q is contrary to the Logical Inconsistency
how people criticize reasoning. Any. descriptive theory has to make Principle, not to the Logical Implication Principle. On the other hand
use of idealization. the Logical Inconsistency Principle does not say one has a reason to
believe Q given that one believes P and if P then Q.
Neither principle is exceptionless as it stands. Each holds, as it were,
other things being equal. Each is defeasible. For example, the Logical
Implication Principle entails that, if one believes both P and if P then
Q, that can be a reason to believe Q. But, clearly, that is not always a
reason •to believe Q, since sometimes when one believes P and also
believes if P then Q, one should not come to believe Q. Remember
Mary who came to believe three inconsistent things: If she looks in
the closet she will see a box of Cheerios, she is looking in the closet,
 Logic and Reasoning 13
12 Chapter 2
Explicit and Implicit Belief
but she does not see a box of Cheerios. Mary should not at this point
infer that she does see a box of Cheerios from her first two beliefs. I assume one believes something explicitly if one's belief in that thing
This suggests modifying the Logical Implication Principle: involves an explicit mental representation whose content is the content
of that belief. On the other hand something is believed only implicitly
Logical Closure Principle One's beliefs should be "closed under
if it is not explicitly believed but, for example, is easily inferable from
logical implication." In other words there is something wrong with
one's explicit beliefs. Given that one explicitly believes the earth has
one's beliefs if there is a proposition logically implied by them
exactly one sun, one can easily infer that the earth does not have two
which one does not already believe. In that case one should either
suns, that the earth does not have three suns, and so on. So all these
add the implied proposition to one's beliefs or give up one of the
propositions are things one believes implicitly.
implying beliefs.
That is an example in which implicit beliefs are implied by explicit
But the Logical Closure Principle is not right either. Many trivial things beliefs. There are also cases in which one impHcitly believes something
are implied by one's view which it would be worse than pointless to that is easily inferable from one's beliefs without being strictly implied
add to what one believes. For example, if one believes P, one's view by them. An example might be one's implicit belief that elephants don't
trivially implies "either P or Q," "either P or P," "P and either P or wear pajamas in the wild (Dennett 1978).
R," and so on. There is no point in cluttering one's mind with all these There is also another way in which something can be implicitly
propositions. And, of course, there are many other similar examples. believed-it may be implicit in one's believing something else. For
Here I am assuming the following principle: example, in explicitly believing P, it may be that one implicitly believes
Clutter Avoidance One should not clutter one's mind with one is justified in believing P. The proposition that one is justified in
believing P is not ordinarily implied by the proposition P and may not
trivialities.
be inferable from one's explicit beliefs, but it may be that in believing
This raises an interesting issue. To suppose one's mind could become P one is committed to and so implicitly believes the proposition that
cluttered with beliefs is to suppose such things as (1) that it takes time one is justified by believing P. (I discuss this and related possibilities
to add to one's beliefs further propositions that are trivially implied by in chapter 5.)
them, time that might be better spent on other things, and/or (2) that It is a possible view that none of one's beliefs are explicit, that is,
one has "limited storage capacity" for beliefs, so there is a limit on the that none are explicitly represented and all are only implicit in one's
number of things one can believe, and/or (3) that there are limits on mental makeup. This is a form of behaviorism about belief. There is
"information retrieval," so the more one believes the more difficult it surprisingly much that can be said in favor of such behaviorism (Dennett
is to recall relevant beliefs when one needs them. 1978; Stalnaker 1984), but I suppose that whatever is ultimately the
Such suppositions presuppose that beliefs are explicitly "represented" right view of belief must allow that unbridled inference can lead to
in the mind in the sense that these representations play the important too much clutter either in what one explicitly believes or in whatever
role in perception, thought, and reasoning that we think beliefs play. explicit thing underlies belief. Therefore I ignore the possibility of such
But we must be careful in stating this presupposition. Not all one's behaviorism and continue to assume that one's implicit beliefs are
beliefs can be explicitly represented in this way, since then one could implicit in one's believing certain things explicitly. If this is wrong, I
believe only finitely many things. But one can and does believe infinitely doubt that it is so wrong as to affect the conclusions I draw from this
many things. For example, one believes the earth does not have two assumption.
suns, the earth does not have three suns, the earth does not have four In this connection it might be useful for me to digress briefly to
suns, and so on. observe that the distinction between explicit and implicit belief is not
In order to accommodate this point, I assume that we can distinguish the same as either the distinction between belief that is available to
what one believes explicitly from what one believes only implicitly. consciousness and unconscious belief or that between "occurrent" and
Then we can take the principle of clutter avoidance to apply to what "dispositional" beliefs.
one believes explicit!y. We normally consider a belief "unconscious" if one is not aware one
has it and one cannot easily become aware of it simply by considering
14 Chapter 2 Logic and Reasoning 15

whether one has it. Otherwise the belief is available to consciousness. Clutter Avoidance
Now, clearly, implicit beliefs can be available to consciousness. The
belief that the earth does not have two suns is normally only implicit How is the principle of clutter avoidance to be used? It seems absurd
in one's explicit beliefs and is not itself explicitly represented, even for it to figure explicitly in one's reasoning, so that one refrains from
though it is immediately available to consciousness in the sense that, drawing an otherwise acceptable conclusion on the grounds of clutter
if one considers whether one believes it, one can immediately tell one avoidance. Once one is explicitly considering whether or not to accept
does. a conclusion, one cannot decide not to on such grounds. One might
On the other hand a belief can be explicitly represented in one's rationally decide not to try to remember it, perhaps, but one cannot
mind, written down in Mentalese as it were, without necessarily being decide not to believe it at least for the moment.
available to consciousness. For example, one might explicitly believe Suppose George is trying to convince Bob that P. George shows how
that one's mother does not love one, even though this belief may not p is a deductive consequence of things Bob believes. Bob accepts the
be consciously retrievable without extensive psychoanalysis. So the validity of George's argument and refuses to change his belief in any
distinction between implicit and explicit beliefs is not the same as that of the premises, but he also refuses to accept the conclusion P, citing
between unconscious beliefs and those available to consciousness. clutter avoidance as his reason for refusing. That is absurd. (l am in-
Turning now to the distinction between occurrent and dispositional debted to Robert Stalnaker for this example.)
beliefs, we can say a belief is occurrent if it is either currently before But the Principle of Clutter Avoidance is not just a principle about
one's consciousness or in some other way currently operative in guiding what one should try to remember. It would be a violation of clutter
what one is thinking or doing. A belief is merely dispositional if it is avoidance if one spent all one's time thinking up trivial consequences
only potentially occurrent in this sense. Any merely implicit belief is of one's beliefs even if one refrained from committing these conse-
merely dispositional, but explicit beliefs are not always occurrent, since quences to memory. In that case one would be cluttering up one's
only some explicit beliefs are currently operative at any given time. So short-term processing capacities with trivialities.
the djstinction between implicit and explicit beliefs is not the same as The Principle of Clutter Avoidance is a metaprinciple that constrains
that between occurrent and dispositional beliefs. the actual principles of revision. The principles of revision must be
So much for this digression comparing these various distinctions in such that they discourage a person from cluttering up either long-term
kinds of beliefs. memory or short-term processing capacities with trivialities. One way
Let me return to the discussion of the Logical Closure Principle, to do this would be to allow one to accept a new belief P only if one
which says one's beliefs should be closed under logical implication. has (or ought to have) an interest in whether Pis true. (This is discussed
Clearly this principle does not apply to explicit beliefs, since one has in chapter 6.)
only a finite number of explicit beliefs and they have infinitely many
logical consequences. Nor can the Logical Closure Principle be satisfied
Unavoidable Inconsistency and the Liar Paradox
even by one's implicit beliefs. One cannot be expected even implicitly
to believe a logical consequence of one's beliefs if a complex proof I was saying that neither the Logical Implication Principle nor the
would be needed to see the implication. Inconsistency Principle is without exception. I have indicated why this
It won't help to change the Logical Closure Principle to say one's is so for the Logical Implication Principle, which says one has a reason
beliefs should be closed under obvious logical implication. That ~ould to believe the logical implications of one's beliefs. Similar remarks hold
come to the same thing, since any logical implication can eventually for the Logical Inconsistency Principle, which says one should avoid
be demonstrated by a proof consisting entirely of a series of obvious inconsistency.
steps. This means that, if beliefs are required to be closed under obvious To see that the Logical Inconsistency Principle has its exceptions,
logical implication, they are required to be closed under any logical observe that sometimes one discovers one's views are inconsistent and
implication, obvious or not. So, since beliefs cannot be required to be does not know how to revise them in order to avoid inconsistency
closed under logical implication, they cannot be required to be closed without great cost. In that case the best response may be to keep the
under obvious logical implication either. inconsistency and try to avoid inferences that exploit it. This happens
in everyday life whenever one simply does not have time to figure out
16 Chapter 2 Logic and Reasoning 17

what to do about a discovered inconsistency. It can also happen on things being equal." But in ordinary life we accept many principles of
more reflective occasions. For example, there is the sort of inconsistency this vaguer sort.
that arises when one believes that not all one's beliefs could be true. My point about the Logical Inconsistency Principle remains. One
One might well be justified in continuing to believe that and each of may find oneself with inconsistent beliefs and not have the time or
one's other beliefs as well. ability to trace the sources of the inconsistency (e.g., the Biconditional
There are also famous logical paradoxes. For example, the liar paradox Truth Schema). In that event, it is rational simply to retain the con-
involves reflection on the following remark, which l call (L): tradictory beliefs, trying not to exploit the inconsistency.

(L) is not true.

Immediate Implication and Immediate Inconsistency
Thinking about (L) leads one into contradiction. If (L) is not true, things
are as (L) says, so (L) must be true. But if (L) is true, then it is true that I turn now to an issue about the Logical Implication and Inconsistency
~L) is _not true, so (L) must not be true. It seems (L) is true if and only Principles I have so far mentioned only in passing. One might have
1f (L) 1s not true. But that is a contradiction. no reason to accept something that is logically implied by one's beliefs
The paradox arises from our uncritical acceptance of the following: if there is no short and simple argument showing this. To take an
extreme example, one accepts basic principles of arithmetic that logically
BiconditionalTruth Schema "P" is true if and only if P. imply some unknown proposition P which is the answer to an unsolved
To see that this schema is indeed the culprit, notice that one instance mathematical problem: but one has no reason to believe P if one is
of it is not aware that P is implied by these basic principles. This suggests
revising the Logical Implication Principle:
"(L) is not true" is true if and only if (L) is not true.
RecognizedLogicalImplicationPrinciple One has a reason to believe
Since (L) = "(L) is not true," this instance is equivalent to the self- P if one recognizesthat P is logically implied by one's view.
contradictory
Similarly, we might revise the Logical Inconsistency Principle:
(L) is true if and only if (L) is not true
Recognized Logical lnconsistency Principle One has a reason to
Various restrictions on the Biconditional Truth Schema have been sug- avoid believing things one recognizes to be logically inconsistent.
gested in order to avoid the liar paradox, but none is completely sat-
isfactory (Kripke 1975, Herzberger 1982). So, the rational response for However, there is a problem with this. It would seem one can recognize
most of us may simply be to recognize our beliefs about truth are a logical implication or logical inconsistency only if one has the relevant
lo?ic_allyin:onsiste_nt, agree this is undesirable, and try not to exploit concept of logical implication or logical inconsistency. But it. would
this mcons1stency m our inferences. (The danger is that, since incon- seem that few people have such concepts, at least if this involves dis-
sistent beliefs logically imply anything, if one is not careful, one will tinguishing logical implication and inconsistency from other sorts of
be able to use this fact to infer anything whatsoever.) implication and inconsistency. Consider the following examples:
In practice the best solution may be to retain the Biconditional Truth P or Q and not P taken together imply Q.
Schema and yet avoid contradiction by interpreting the Schema not A = B and B = C taken together imply A = C.
as something that holds without exception but rather as something that A < B and B < C taken together imply A < C.
holds "normally" or "other things being equal." It is then a "default A is part of Band Bis part of C taken together imply A is part of C.
assumption." One accepts any given instance of the Biconditional Truth X is Y's brother implies X is male.
Schema in the absence of a sufficiently strong reason not to accept it. Today is Thursday implies Tomorrowis Friday.
One does not apply the Schema to (L) because doing so leads to X plays defensive tacklefor the PhiladelphiaEaglesimplies X weighs
contradiction. more than 150 pounds.
This does not seem to be a satisfactory solution from the point of
view of logic, since we take logic to require precise principles with Peo~le who recognize these and related implications do not in any
precise boundaries, not principles that hold merely "normally" or "other consistent way distinguish them into purely logical implications and
18 Chapter 2 Logic and Reasoning 19

others that are not purely logical. (Only the first counts as purely logical
with
in "classical" first-order predicate logic without identity. Sometimes
principles for identity are included as part of logic, in which the second A is disposed to treat P, Q, ... , R as a reason to believe S.
also counts as a logical implication.) So the Recognized Logical Im- For one thing, A's general disposition may be overridden by other
plication and [nconsistency Principles would seem to have only a limited considerations in a particular case, for example, if S is absurd. In that
application. case, some of A's beliefs will immediately imply a particular proposition
To some extent this objection can be met by generalizing the prin- for A, although A is not disposed to treat those beliefs as reasons for
ciples, dropping specific mention of logical implication and inconsis- believing that proposition. Furthermore, beliefs can be treated by A as
tency. Then the principles would be stated as follows: reasons for believing a conclusion even though A does not take those
Recognized Implication Principle One has a reason to believe P if beliefs to imply that conclusion.
Similarly, we can say A is disposed to avoid believing things that
one recognizes that P is implied by one's view.
are immediately inconsistent for A but we cannot identify a set of
Recognized Inconsistency Principle One has a reason to avoid be- beliefs' being inconsistent for A with A's having a disposition to avoid
believing all the members of that set. On the one hand the general
lieving things one recognizes to be inconsistent.
disposition may be overridden in a particular case, as when A is disposed
These principles still apply only to people who have concepts of im- to believe the premises of the liar paradox. On the other hand (as
plication and inconsistency. But this is not so clearly problematical, if Soames observes) there is Moore's paradox: One is strongly disposed
only because it is not clear what it takes to have these concepts. not to believe both P and that one does not believe P while realizing
I suggest it is enough to be able to make reasoned changes in one's that these propositions are perfectly consistent with each other.
view in a way that is sensitive to implication and inconsistency. Someone So, I am inclined simply to assume one has certain basic dispositions
who is disposed to treat beliefs in P and if P then Q as reasons to to take some propositions immediately to imply other propositions and
believe Q has, by virtue of that very disposition, an appropriate ability to take some propositions as immediately inconsistent with each other.
to recognize this sort of implication, at least if this disposition is also More generally, I assume one can have general dispositions with respect
accompanied by the disposition not to believe P, if P then Q, and not Q. to certain patterns of immediate implication and inconsistency even if
And the latter sort of disposition might reflect an appropriate ability some instances of the patterns are so long or complex or otherwise
to recognize that sort of inconsistency. distracting that one has no particular disposition to take those particular
I am inclined to take as fundamental certain dispositions to treat instances to be immediate implications or inconsistencies.
propositions in certain ways, in particular, dispositions to treat some
propositions as immediately implying others and some as immediately Summary and Conclusion
inconsistent with each other. That is, I am inclined to suppose that the
basic notions are I began by suggesting logic might be specially relevant to reasoning
in two ways, via implication and inconsistency. lt seemed the relevant
P, Q, ... , R immediately imply S for A principles would be defeasible, holding only other things being equal.
and Furthermore, they would apply only to someone who recognized the
implication or inconsistency. Since this recognition might be manifested
P, Q, ... , R are immediately inconsistent for A. simply in the way a person reacts to these implications and incon-
It is unclear to me whether these notions can be reduced to others sistencies, I suggested that certain implications and inconsistencies are
in any interesting way. (I am indebted to Scott Soames for raising this "immediate" for a given person. (In appendix A, I discuss whether
issue.) [f A is disposed to treat some beliefs as implying others, then basic logical notions can be defined in terms of such immediate im-
A is disposed to treat beliefs that immediately imply something as plications and inconsistencies. I also discuss the hypothesis that only
giving him or her a reason to believe that thing; but we cannot identify logical implications and inconsistencies are immediate and that others
are mediated by the acceptance of certain nonlogical principles, con-
P, Q, ... , R immediately imply S for A cluding that this hypothesis may be impossible to refute.)
20 Chapter 2

My conclusion is that there is no clearly significant way in which

logic is specially relevant to reasoning. On the other hand immediate
implication and immediate 'inconsistency do seem important for rea-
soning, and so do implication and inconsistency. Sometimes, reasoning
culminates in the conclusion that a certain argument is a good one or
Chapter 3
that certain propositions are inconsistent. But that is not to say that
logical implication or logical inconsistency has any special status in
Belief and Degree of Belief
human reasoning.

ProbabilisticImplication
We have a rule connecting implication and reasoning:
Principle of Immediate Implication That Pis immediately implied
by things one believes can be a reason to believe P.
Is there also a weaker probabilistic version of this rule?
Hypothetical Principle of Immediate ProbabilisticImplication That
Pis obviously highly probable, given one's beliefs, can be a reason
to believe P.
Suppose Mary purchases a ticket in the state lottery. Given her beliefs,
it is obviously highly probable that her ticket will not be one of the
winning tickets. Can she infer that her ticket will not win? ls she
justified in believing her ticket is not one of the winning tickets?
Intuitions waver here. On the one hand, if Mary is justified in believing
her ticket is not one of the winning tickets, how can she be justified
in buying the ticket in the first place? Furthermore, it certainly seems
wrong to say she can know that her ticket is not one of the winning
tickets if it is really a fair lottery. On the other hand the probability
that the ticket is not one of the winning tickets seems higher than the
probability of other things we might easily say Mary knows. We or-
dinarily allow that Mary can come to know various things by reading
about them in the newspaper, even though we are aware that news-
papers sometimes get even important stories wrong.
This issue is one that [ will return to several times, but 1 want to
begin by considering a suggestion which I think is mistaken, namely,
that the trouble here comes from not seeing that belief is a matter of
degree.
22 Chapter 3 Belief and Degree of Belief 23

All-or-Nothing Belief Propositions -that are individually highly probable can have an im-
mediate implication that is not. The fact that one assigns a high prob-
I have been supposing that, for the theory of reasoning, explicit belief ability to P and also to if P then Q is not a sufficient reason to assign
is an all-or-nothing matter. I have assumed that, as far as principles a high probability to Q. Each premise of a valid argument might be
of reasoning are concerned, one either believes something explicitly or probable even though the c~nclusion_ is improba~J_e. Sinc_eone mi?ht
one does not; in other words an appropriate "representation" is either assign a high degree of behef to various propos1t1ons without bemg
in one's "memory" or not. The principles of reasoning are principles committed to assigning a high degree of belief to a logical consequence
for modifying such all-or-nothing representations. of these propositions, Kyburg argues that the Logical Implication Prin-
This is not to deny that in some ways belief is a matter of degree. ciple is mistaken.
For one thing implicit belief is certainly a matter of degree, since it is Similarly, each of an inconsistent set of beliefs might be highly prob-
a matter of how easily and automatically one can infer something from able. To take Kyburg's lottery example, it might be that the proposition,
what one believes explicitly .. Furthermore, explicit belief is a matter of "one of the N tickets in this lottery is the winning ticket" is highly
degree in the sense that on~ believes some things more strongly than probable, and so is each p_roposition of the form, "ticket. i is not_ the
others. Sometimes one is only somewhat inclined to believe something, winning ticket," for each I between 1 and N. So one might believe
sometimes one is not sure what to believe, sometimes one is inclined each of these propositions to a high degree while recognizing that they
to disbelieve something, sometimes one is quite confident something are jointly inconsistent. Kyburg argues there is nothing wrong with
is not so, and so forth. this, so the Logical Inconsistency Principle is mistaken.
How should we account for the varying strengths of explicit beliefs? It is not just that these principles have exceptions. We have seen
I am inclined to suppose that these varying strengths are implicit in a that they are defeasible and hold only other things being equal. But if
system of beliefs one accepts in a yes/no fashion. My guess is that belief were always a matter of degree the principles would not even
they are to be explained as a kind of epiphenomenon resulting from hold in this way as defeasible principles. They would not hold at all.
the operation of rules of revision. For example, it may be that P is It would be odd for someone to take this seriously in a routine matter.
believed more strongly than Q if it would be harder to stop believing It is contrary to the way we normally think. Imagine arguing with such
P than to stop believing Q, perhaps because it would require more of a person. You get him to believe certain premises and to appreciate
a revision of one's view to stop believing P than to stop believing Q. that they imply your conclusion, but he is not persuaded to believe
In contrast to this, it might be suggested that principles of reasoning this conclusion, saying that, although you have persuaded him to assign
should be rules for modifying explicit degrees of belief. In this view, an a high probability to each of your premises, that is not enough to show
account of reasoning should be embedded in a theory of subjective he should assign a high probability to the conclusion! This is not the
probability, for example, as developed by Jeffrey (1983), not that Jeffrey way people usually respond to arguments.
himself accepts this particular suggestion. In fact, this suggestion cannot Or consider the following attitude toward contradiction. As Jack asserts
really be carried out. People do not normally associate with their beliefs several things, you observe that he has contradicted himself. His re-
degrees of confidence of a sort they can use in reasoning. It is too sponse is that he sees nothing wrong, since all the things he has asserted
complicated for them to do so. Degrees of belief are and have to be are highly probable. This is comprehensible, but it is again different
implicit rather than explicit, except for a few special cases of beliefs from the normal way of doing things.
that are explicitly beliefs about probabilities. A normal reaction to someone's refusal to accept the conclusion of
Let me say why this is so. To begin with, Kyburg (1961) observes a clearly valid argument after he says he has been persuaded to accept
that the Immediate Jmplication and Inconsistency Principles would not the premises, if he gives Kyburg's reason, is to suppose that he does
be right even as approximations if belief were a matter of degree. not really accept the premises after all, but only believes of each that
Immediate Implication Principle The fact that one's view imme- it is probable. Similarly, we suppose that a person who says at least
diately implies P can be a reason to accept P. one ticket will win and also says of each ticket that it will not win does
not really believe of each ticket that it will not win but merely believes
Immediate InconsistencyPrinciple Immediate logical inconsistency of each ticket that it is unlikely that that ticket will win. We do not
in one's view can be a reason to modify one's view. Ordinarily think of this as like the case in which an author believes
24 Chapter 3 Belief and Degree of Belief 25

each of the things he or she says in a book he or she has written and Conditionalizatio11
also believes that, given human fallibility, at least one of the things he
Some probability theorists appear to deny these obvious points. They
or she has said in the book must be false. Such a person is justified in
having inconsistent beliefs, but that does not show that the Recognized 5eem to suppose that reasoned revision is or ought always to be_m
ccordance with a special principle of "conditionalization" that app!Jes
Inconsistency Principle is incorrect. It only shows that the principle is
~hen one comes to treat evidence E as certain. The claim is that in
defeasible.
uch a case one is to modify one's other degrees of belief so that the
Of course, to say one normally thinks of belief in an all-or-nothing
~ew probability one assigns to any given proposition Pis given by the
way is not to deny one sometimes has beliefs about probabilities. More
following formula:
important, one often manifests a varying degree of confidence in this
or that proposition as revealed in one's willingness to act, for example, old prob (P & E)
to bet. But this does not show one normally or usually assigns explicit new prob (P) = old prob (£)
levels of confidence or probability to one's beliefs. The degree of con-
fidence one has might be merely implicit in one's system of beliefs.
The quotient on the right-hand side is sometimes called the conditional
Subjective probability theory can give an account of one's dispositions
probability of P given E, which is why the principle is called
without being an account of the psychological reality underlying those
conditionalization.
dispositions. R. C. Jeffrey (1983, chap. 11) shows how this formula can be gen-
It might be said one ought to operate using explicit degrees of belief. eralized to allow for the case in which evidence propositions change
This would imply one should make much more use of probability in probability without becoming certain. Suppose that there are II rele-
theory than one does. vant atomic evidence propositions £ 1, ••• , £,., so that there are 2"
Similarly, it might be said that one's goals should be treated as strongest conjunctions C, each containing E, or its denial. Then the new
matters of degree. Since different prospects are more or less desirable, probability one assigns to any given proposition P is the sum of aU the
one ought to assign them different degrees of ''subjective utility." ln quantities of the following form:
acting, one should act so as to maximize expected utility.
In chapter 9 I argue that this is not right. But even if it were right, old prob (P & C;)
such an appeal to probability theory would not eliminate the need for new prob (C;) X old prob (C,)
reasoning in the sense of change in view. One's subjective probability
assignments would never be complete. They would often have to be So, let us consider the following hypothesis, which is widely accepted
extended. To some extent they could be extended by means of the by subjective probability theorists:
Principle of Immediate Implication by considering the immediate im-
plications of one's current probability assignments and by allowing for Reasoning is conditionalization The updating of probabilities via
clutter avoidance and other relevant considerations. Furthermore, there conditionalization or generalized conditionalization is (or ought to
would also often be cases in which current subjective probability as- be) the only principle of reasoned revision.
signments would have to be changed, for example because they were One way to argue for this is to try to show that various intuitively
not consistent with each other. The Principle of Immediate Inconsistency acceptable priJ1ciples of reasoning from evidence can be accounted for
then has a role to play. And there are other cases in which one will if this hypothesis is accepted (e.g., Dorling 1972; Horwich 1982).
want to modify such assignments, for example, when one discovers However, there is a problem with making extensive use of this method
that a current theory would explain old evidence one had not realized of updating. One can use conditionalization to get a new probability
it would explain (Glymour 1980, chap. 3). And whatever principles for P only if one has already assigned a prior probability not only to
are developed for changing all-or-nothing belief will apply to changing E but to P & E. If one is to be prepared for various possible condi-
degrees of belief, treating these as a!J-or-nothing beliefs about tionalizations, then for every proposition P one wants to update, one
probabilities. must already have assigned probabilities to various conjunctions of P
together with one or more of the possible evidence propositions and/
or their denials. Unhappily, this leads to a combinatorial explosion,
26 Chapter 3 Belief and Degree of Belief 27

since the number of such conjunctions is an exponential function of new probability distribution than the description envisioned in (1) or
the number of possibly relevant evidence propositions. [n other words, (2), so this will not normally be feasible either.
to be prepared for corning to accept or reject any of ten evidence prop- Doing extensive updating by conditionalization or generalized con-
ositions, one would have to record probabilities of over a thousand ditionalization would be too complicated in practice. Therefore one
such conjunctions for each proposition one is interested in updating. must follow other principles in revising one's views. It is conceivable
To be prepared for twenty evidence propositions, one must record a that all or some of these principles might refer to strength or degree
million probabilities. For thirty evidence propositions, a billion prob- of belief and not just to whether one believes something in a yes/no
abilities are needed, and so forth. fashion. But the actual principles we follow do not seem to be of that
Clearly, one could not represent all the needed conjunctions explicitly. sort, and it is unclear how these principles might be modified to be
One would have to represent them implicitly using some sort of general sensitive to degree or strength of belief. In the rest of this book I assume
principle. Given such a general principle, one's total probability dis- that, as far as the principles of revision we follow are concerned, belief
tribution would then be determined, by either (1) the total evidence is an all-or-nothing matter. I assume that this is so because it is too
one accepts as certain (using conditionalization) or (2) the various new complicated for mere finite beings to make extensive use of probabilities.
probabilities assigned to the C, (using Jeffrey's generalization of con-
ditionalization). But neither (1) nor (2) is feasible. Consider what is
involved in each case.
The idea behind (1) is to represent the degrees of belief to which
one is presently committed by means of some general principle, speci-
fying an initial probability distribution, together with a list of all the
evidence one has come to treat as certain. Such evidence will include
aJl immediate perceptual evidence-how things look, sound, smell,
etc., to one at this or that moment. One will have to remember all such
evidence that has influenced one's present degrees of beUef. But in fact
one rarely remembers such evidence beyond the moment in which one
possesses it (a point I return to in chapter 4). So (1) is not a usable
approach.
On the other hand, (2) requires that one keep track of one's current
degree of belief in each of the relevant conjunctions C; of evidence
propositions and/or their denials. These are things one does not have
to be certain about, so the relevant propositions need not be for the
most part about immediate perceptual experience, as in (1). So the
objection that one hardly ever remembers such propositions does not
apply to (2). But (2) is also unworkable, since the number of relevant
conjunctions C; is an exponential function of the number of atomic
evidence propositions.
These objections assume one sticks with one's original general prin-
ciple describing one's initial degrees of belief and records one's present
degrees of belief by representing the new evidence accepted as certain
or the new probabilities of the various conjunctions C,.
Alternatively, one might try each time to find a new principle de-
scribing one's updated degrees of belief in a single general statement.
But the problem of finding such a general principle is intractable, and
anyway (b) there will normally be no simpler way to describe one's
Chapter 4
Positive versus Negative Undermining

J now want to compare two competing theories of reasoned belief

revision, which l will call the foundations theory and the coherence
theory since they are similar to certain philosophical theories of jus-
tification sometimes called foundations and coherence theories (Sosa
1980; Pollock 1979). But the theories l am concerned with are not
precisely the same as the corresponding philosophical theories of jus-
tifkation, which are not normally presented as theories of belief revision.
Actually, I am not sure what these philosophical theories of "justifi-
cation" are supposed to be concerned with. So, although I will be using
the term"justification" in what follows, as well as the terms "coherence"
and "foundations," I do not claim that my use of any of these terms
is the same as its use in these theories of justification. I mean to be
raising a new issue, not discussing an old one.
The key issue is whether one needs to keep track of one's original
justifications for beliefs. What lam calling the foundations theory says
yes; what I am calling the coherence theory says no.
The foundations theory holds that some of one's beliefs "depend
on" others for their current justification; these other beliefs may depend
on still others, until one gets to foundational beliefs that do not depend
on any further beliefs for their justification. In this view reasoning or
belief revision should consist, first, in subtracting any of one's beliefs
that do not now have a satisfactory justification and, second, in adding
new beliefs that either need no justification or are justified on the basis
of other justified beliefs one has.
On the other hand, according to the coherence theory, it is not true
that one's ongoing beliefs have or ought to have the sort of justificationa1
structure required by the foundations theory. In this view ongoing
beliefs do not usually require any justification. Justification is taken to
be required only if one has a special reason to doubt a particular belief.
Such a reason might consist in a conflicting belief or in the observation
that one's beliefs could be made more "coherent," that is, more or-
ganized or simpler or less ad hoc, if the given belief were abandoned
30 Chapter 4 Positive versus Negative Undermining 31

(and perhaps if certain other changes were made). According to the be justified, it is required eit!1er that P be a foundational belief whose
coherence theory, belief revision should involve minimal changes in intrinsic justification is not defeated or that there be at least one un-
one's beliefs in a way that sufficiently increases overall coherence. defeated justification of P from other beliefs one is justified in believing.
In this chapter f elaborate these two theories in order to compare If one believes Pandit happens that all one's justifications for believing
them with actual reasoning and intuitive judgments about such rea- p come to be defeated, one is no longer justified in continuing to believe
soning. lt turns out that the theories are most easily distinguished by P, and one should subtract P from one's beliefs.
the conflicting advice they occasionally give concerning whether one Furthermore, and this is important, if one comes not to be justified
should give up a belief P from which many other of one's beliefs have in continuing to believe Pin this way, then not only is it true that one
been inferred, when P's original justification has to be abandoned. Here must abandon belief in P but justifications one has for other beliefs
a surprising contrast seems to emerge-"is" and "ought" seem to come are also affected if these justifications appeal to one's belief in P. Jus-
apart. The foundations theory seems, at least at first, to be more in tifications appealing to P must be abandoned when P is abandoned.
line with our intuitions about how people ought to revise their beliefs; If that means further beliefs are left without justification, then these
the coherence theory is more in line with what people actually do in beliefs too must be dropped along with any justifications appealing to
such situations. Intuition seems strongly to support the foundations them. So there will be a chain reaction when one loses justification for
theory over the coherence theory as an account of what one is justified a belief on which other beliefs depend for their justification. (This is
in doing in such cases; but in fact one will tend to act as the coherence worked out in more detail for an artificial intelligence system by Doyle
theory advises. (1979, 1980).)
After I explain this I consider how this apparent discrepancy can be Now, it is an important aspect of the foundations theory of reasoning
resolved. I conclude that the coherence theory is normatively correct that justifications cannot legitimately be circular. P cannot be part of
after all, despite initial appearances. the justification for Q while Q is part of the justification for P (unless
one of these beliefs has a different justification that does not appeal
to the other belief).
The Foundations Theory of Belief Revision The foundations theory also disallows infinite justifications. It does
The basic principle of the foundations theory, as I will interpret it, is not allow P to be justified by appeal to Q, which is justified by appeal
that one must keep track of one's original reasons for one's beliefs, so to R, and so on forever. Since justification cannot be circular, justification
that one's ongoing beliefs have a justificational structure, some beliefs must eventually end in beliefs that either need no justification or are
serving as reasons or justifications for others. These justifying beliefs justified but not by appeal to other beliefs. Let us say that such basic
are more basic or fundamental for justification than the beliefs they or foundational beliefs are intrinsically justified.
justify. For my purposes it does not matter exactly which beliefs are taken
The foundations theory rejects any principle of conservatism. In this to be intrinsically justified in this sense. Furthermore, I emphasize that
view a proposition cannot acquire justification simply by being believed. the foundations theory allows for situations in which a basic belief has
The justification of a given belief cannot be, either in whole or in part, its intrinsic justification defeated by one or more other beliefs, just as
that one has that belief. For example, one's justification for believing it allows for situations in which the justification of one belief in terms
something cannot be that one already believes it and that one's beliefs ~f other beliefs is defeated by still other beliefs. As I am interpreting
in this area are reliable. it, foundationalism is not committed to the incorrigibility of basic beliefs.
Justifications are prima facie or defeasible. The foundations theory A belief is a basic belief if it has an intrinsic justification which does
allows, indeed insists, that one can be justified in believing something not appeal to other beliefs. A basic belief can also have one or more
P and then come to believe something else that undermines one's nonintrinsic justifications which do appeal to other beliefs. So, a basic
justification for believing P. In that case one should stop believing P, belief can have its intrinsic justification defeated and still remain justified
unless one has some further justification that is not undermined. as long as it retains at least one justification that is not defeated.
I say "unless one has some further justification," because in this . The existence of basic beliefs follows from the restrictions against
view a belief can have more than one justification. To be justified, a arcular and infinite justifications. Infinite justifications are to be ruled
belief must have at least one justification. That is, if a belief in P is to out because a finite creature can have only a finite number of beliefs,
32 Chapter 4 Positive versus Negative Undermining 33

or at least only a finite number of explicit beliefs, whose content is how well it fits together with everything else one believes. If one's
explicitly represented in the brain. What one is justified in believing beliefs are coherent, they are mutually supporting. All one's beliefs
either implicitly or explicitly depends entirely one what one is justified are, in a sense, equally fundamental. In the coherence theory there are
in believing explicitly. To consider whether one's implicit beliefs are not the asymmetrical justification relations among one's ongoing beliefs
justified is to consider whether one is justified in believing the explicit that there are in the foundations theory. It can happen in the coherence
beliefs on which the implicit beliefs depend. A justification for a belief theory that P is justified because of the way it coheres with Q and Q
that appeals to other beliefs must always appeal to things one believes is justified because of the way it coheres with P. In the foundations
explicitly. Since one has only finitely many explicit beliefs, there are theory, such a pattern of justification is ruled out by the restriction
only finitely many beliefs that can be appealed to for purposes of against circular justification. But there is nothing wrong with circular
justification, and so infinite justifications are ruled out. justification in the coherence theory, especially if the circle is a large
one!
The Coherence Theory of Belief Revision I turn now to testing the foundations and coherence theories against
our intuitions about cases. This raises an apparent problem for the
The coherence theory is conservative in a way the foundations theory coherence theory.
is not. The coherence theory supposes one's present beliefs are justified
just as they are in the absence of special reasons to change them, where
changes are allowed only to the extent that they yield sufficient increases An Objection to the Coherence Theory: Karen's Aptitude Test
in coherence. This is a striking difference from the foundations theory. Sometimes there clearly are asymmetrical justification relations among
The foundations theory says one is justified in continuing to believe one's beliefs.
something only if one has a special reason to continue to accept that Consider Karen, who has taken an aptitude test and has just been
belief, whereas the coherence theory says one is justified in continuing told her results show she has a considerable aptitude for science and
to believe something as long as one has no special reason to stop music but little aptitude for history and philosophy. This news does
believing it. not correlate perfectly with her previous grades. She had previously
According to the coherence theory, if one's beliefs are incoherent in done well not only in physics, for which her aptitude scores are reported
some way, because of outright inconsistency or simple ad hocness, then to be high, but also in history, for which her aptitude scores are reported
one should try to make minimal changes in those beliefs in order to to be low. Furthermore, she had previously done poorly not only in
eliminate the incoherence. More genera!Jy, small changes in one's beliefs philosophy, for which her aptitude scores are reported to be low, but
are justified to the extent these changes add to the coherence of one's also in music, for which her aptitude scores are reported to be high.
beliefs. After carefully thinking over these discrepancies, Karen concludes
For present purposes, r do not need to be too specific as to exactly that her reported aptitude scores accurately reflect and are explained
what coherence involves, except to say it includes not only consistency by her actual aptitudes; so she has an aptitude for science and music
but also a network of relations among one's beliefs, especially relations and no aptitude for history and philosophy; therefore her history course
of implication and explanation. must have been an easy one, and also she did not work hard enough
It is important that coherence competes with conservatism. It is as in the music course. She decides to take another music course and not
if there were two aims or tendencies of reasoned revision, to maximize to take any more history.
coherence and to minimize change. Both tendencies are important. It seems quite clear that, in reaching these conclusions, Karen bases
Without conservatism a person would be led to reduce his or her beliefs some of her beliefs on others. Her belief that the history course was
to the single Parmenidean thought that all is one. Without the tendency easy depends for its justification on her belief that she has no aptitude
toward coherence we would have what Peirce (1877) called the method for history, a belief which depends in turn for its justification on her
of tenacity, in which one holds to one's initial convictions no matter belief that she got a low score in history on her aptitude test. There is
what evidence may accumulate against them. no dependence in the other direction. For example, her belief about
According to the coherence theory, the assessment of a challenged her aptitude test score in history is not based on her belief that she
belief is always holistic. Whether such a belief is justified depends on
 34 Chapter 4 Positive versus Negative Undermining 35

has no aptitude for history or on her belief that the history course was original report of her scores. Still, there is considerable coherence among
an easy one. the beliefs she inferred from this false report. For example, there is a
According to the coherence theory, the relevant relations here are connection between her belief that she has little aptitude for history,
merely ten;pora(or causal relations. The coherence theory can agree her belief that her high grade in the history course was the result of
~hat ~aren s behef about the outcome of her aptitude test precedes and the course's being an easy one, and her belief that she will not take
1s an important cause of her belief that the history course she took was any more courses in history. There are similar connections between
an easy one. But the coherence theory denies that a relation of de- her beliefs about her aptitudes for other subjects, how well she did in
pendence or justification holds or ought to hold between these two courses in those subjects, and her plans for the future in those areas.
beliefs as time goes by, once the new belief has been firmly accepted. Let us suppose that from the original report Karen inferred a great
In order to test this, let me tell more of Karen's story. Some days many other things that I haven't mentioned; so there are many beliefs
later she is informed that the report about her aptitude scores w~s involved here. Abandoning all these beliefs is costly from the point of
incorrect! The scores reported were those of someone else whose name view of conservatism, which says to minimize change. Suppose that
was confused with hers. Unfortunately, her own scores have now been there are so many of these beliefs and that they are so connected with
lost. How should Karen revise her views, given this new information? each other and with other things Karen believes that the coherence
The foundations theory says she should abandon all beliefs whose theory implies Karen should retain all these new beliefs even though
justifications depend in part on her prior belief about her aptitude test she must give up her beliefs about the explanation of the report of her
scores. The only exception is for beliefs for which she can now find aptitude scores. (In fact, we do not really need to suppose these beliefs
another and independent justification which does not depend on her are intricately connected with each other or even that there are many
belief about her aptitude test scores. She should continue to believe of them, since in the coherence theory a belief doesacquire justification
only those things she would have been justified in believing if she had simply by being believed.)
never been given the false information about those scores. The foun- The foundations theory says Karen should give up all these beliefs,
dation_s theory says this because it does not accept a principle of con- whereas the coherence theory says Karen should retain them. Which
servatism. The foundations theory does not allow that a belief can theory is right about what Karen ought to do? Almost everyone who
acquire justification simply by being believed. has considered this issue sides with the foundations theory: Karen
Let us assume that, if Karen had not been given the false information should not retain any beliefs she inferred from the false report of her
about her aptitude test scores, she could not have reasonably reached aptitude test scores that she would not have been justified in believing
a~y of the ~onclusions she did reach about her aptitudes for physics, in the absence of that false report. That does seem to be the intuitively
history, philosophy, and music; and let us also assume that without right answer. The foundations theory is in accordance with our intuitions
those beliefs Karen could not have reached any of her further conclusions ~bout what Karen ought to do in a case like this. The coherence theory
1s not.
ab~ut the courses she has already taken. Then, according to the foun-
dations theory, Karen should abandon her beliefs about her relative
aftitudes for these subjects, and she should give up her belief that the BeliefPerseverance
history course she took was easy as well as her belief that she did not
work hard enough in the music course. She should also reconsider her In fact, Karen would almost certainly keep her new beliefs! That is
what people actually do in situations Jjke this. Although the foundations
decisions to take another course in music and not to take any more
history courses. theory seems to give intuitively satisfying advice about what Karen
ought to do in such a situation, the coherence theory is more in accord
The coherence theory does not automatically yield the same advice
with what people actually do.
that the foundations theory gives about this case. Karen's new infor-
To document the rather surprising facts here, let me quote at some
mation does produce a loss of overall coherence in her beliefs, since
length from a recent survey article (Ross and Anderson 1982, pp.
she can no longer coherently suppose that her aptitudes for science, 147-149), which speaks of
music, philosophy, and history are in any way responsible for the
original report she received about the results of her aptitude test. She the dilemma of the social psychologist who has made use of de-
must abandon that particular supposition about the explanation of the ception in the course of an experiment and then seeks to debrief
36 Chapter 4 Positive versus Negative Undermining 37

the subjects who had been the target of such deception. The psy- ... [Other] studies first manipulated and then attempted to
chologist reveals the totally contrived and inauthentic nature of undermine subjects' theories about the functional relationship be-
the information presented presuming that this debriefing will tween two measured variables: the adequacy of firefighters' profes-
thereby eliminate any effects such information might have exerted sional performances and t,heir prior scores on a paper and pencil
upon the subjects' feelings or beliefs. Many professionals, however, test of risk performance .... [S]uch theories survived the revelations
have expressed public concern that such experimental deception that the cases in question had been totally fictitious and the dHferent
may do great harm that is not fully undone by conventional de- subjects had, in fact, received opposite pairings of riskiness scores
briefing procedures .... and job outcomes .... [O]ver 50% of the initial effect of the "case
history" information remained after debriefing.
Ross and Anderson go on to describe experiments designed to "explore"
what they call "the phenomenon of belief perseverance in the face of In summary, it is clear that beliefs can survive ... the total de-
evidential discrediting." In one experiment, struction of their original evidential bases.
Subjects first received continuous false feedback as they performed It is therefore quite likely that Karen will continue to believe many
a novel discrimination task (i.e., distinguishing authentic suicide of the things she inferred from the false report of her aptitude test
notes from fictitious ones) .... (Then each subject] received a stan- scores. She will continue to believe these things even after learning
dard debriefing session in which he learned that his putative out- that the report was false.
CO!i\e had been predetermined and that his feedback had been
totally unrelated to actual performance .... [E]very subject was led
to explicitly acknowledge his understanding of the nature and
The Habit Theon; of Belief
purpose of the experimental deception. Why is it so hard for subjects to be debriefed? Why do people retain
conclusions they have drawn from evidence that is now discredited?
Following this total discrediting of the original information, the
One possibility is that belief is a kind of habit. This is an implication
subjects completed a dependent variable questionnaire dealing with
of behaviorism, the view that beliefs and other mental attitudes are
[their] performance and abilities. The evidence for postdebriefing
habits of behavior. But the suggestion that beliefs are habits might be
impression perseverance was unmistakable .... On virtually every
correct even apart from behaviorism. The relevant habits need not be
measure ... the totally discredited initial outcome manipulation
overt behavioral habits. They might be habits of thought. Perhaps, to
produced significant "residual" effects upon [subjects'] ...
believe that Pis to be disposed to think that P under certain conditions,
assessments ....
to be disposed to use this thought as a premise or assumption in. rea-
Follow-up experiments have since shown that a variety of un- soning and in deciding what to do. Then, once a belief has become
founded personal impressions, once induced by experimental pro- established, considerable effort might be needed to get rid of it, even
cedures, can survive a variety of total discrediting procedures. For if the believer should come to see that he or she ought to get rid of it,
example, Jennings, Lepper, and Ross ... have demonstrated that just as it is hard to get rid of other bad habits. One can't simply decide
subjects' impressions of their ability at interpersonal persuasion to get rid of a bad habit; one must take active steps to ensure that the
(having them succeed or fail to convince a confederate to donate habit does not reassert itself. Perhaps it is just as difficult to get rid of
blood) can persist after they have learned that the initial outcome a bad belief.
was totally inauthentic. Similarly, ... two related experiments have Goldman (1978) mentions a related possibility, observing that
shown that students' erroneous impressions of their "logical prob- Anderson and Bower (1973) treat coming to believe something as the
lem solving abilities" (and their academic choices in a follow-up establishing of connections, or "associative links," between relevant
measure two months later) persevered even after they had learned conceptual representations in the brain. Now, it may be that, once set
that good or poor teaching procedures provided a totally sufficient up, such connections or links cannot easily be broken unless competing
explanation for the successes or failures that were the basis for connections are set up that overwhelm the original ones. The easiest
such impressions. case might be that in which one starts by believing P and then comes
 38 Chapter 4
Positive versus Negative Undermining 39
to believe not P by setting up stronger connections involving not P Furthermore, since people rarely keep track of their reasons, the
than those involved in believing P. It might be much harder simply to
theory implies that people are unjustified in almost all their beliefs.
give up one's belief in P without substituting a contrary belief. According
This is an absurd result! The foundations theory turns out not to be a
to this model of belief, in order to stop believing P, it would not be
plausible normative theory aft-er all. So let us see whether we cannot
enough simply to notice passively that one's evidence for P had been
defend the coherence theory as a normative theory.
discredited. One would have to take positive steps to counteract the
We have already seen how the coherence theory can appeal to a
associations that constitute one's belief in P. The difficulties in giving
nonholistic causal notion of local justification by means of a limited
up a discredited belief would be similar in this view to the difficulties
number of one's prior beliefs, namely, those prior beliefs that are most
envisioned in the habit theory of belief.
crucial to one's justification for adding the new belief. The coherence
But this explanation does not give a plausible account of the phe-
theory does not suppose there are continuing links of justification de-
nomenon of belief perseverance. Of course, there are cases in which
pendency that can be cons,ulted _when re~ising on~'s b~liefs. But the
one has to struggle in order to abandon a belief one takes to be dis-
theory can admit that Karen s commg to believe certam thmgs depended
credited. One finds oneself coming back to thoughts one realizes one
on certain of her prior beliefs in a way that it did not depend on others,
should no longer accept. There are such habits of thought, but this is
where this dependence represents a kind of local justification, even
not what is happening in the debriefing studies. Subjects in these studies
though in another respect whether Karen was justified in coming to
are not struggling to abandon beliefs they see are discredited. On the
believe those things depended on everything she then believed.
contrary, the subjects do not see that the beliefs they have acquired Given this point, I suggest the coherence theory can suppose it is
have been discredited. They come up with all sorts of "rationalizations" incoherent to believe both P and also that all one's reasons for believing
(as we say) appealing to connections with other beliefs of a sort that P relied crucially on false assumptions. Within the coherence theory,
the coherence theory, but not the foundations theory, might approve. this implies, roughly, the following:
So the correct explanation of belief perseverance in these studies is not
that beliefs which have lost their evidential grounding are like bad Principle of Positive Undermining One should stop believing P
habits. whenever one positively believes one's reasons for believing Pare
no good.
Positive versus Negative Undermining This is only roughly right, since there is also the possibility that one
should instead stop believing that one's reasons for P are no good, as
In fact, what the debriefing studies show is that people simply do not
well as the possibility that one cannot decide between that beli_ef~nd
keep track of the justification relations among their beliefs. They con- P. In any event, I want to compare this rough statement of the pnnc1ple
tinue to believe things after the evidence for them has been discredited with the corresponding principle in a foundations theory:
because they do not realize what they are doing. They do not understand
that the discredited evidence was the sole reason why they believe as Principle of Negative Undermining One should stop believing P
they do. They do not see they would not have been justified in forming whenever one does not associate one's belief in P with an adequate
those beliefs in the absence of the now discredited evidence. They do justification (either intrinsic or extrinsic).
not realize these beliefs have been undermined. It is this, rather than
The Principle of Positive Undermining is much more plausible than
the difficulty of giving up bad habits, that is responsible for belief the Principle of Negative Undermining. The Principle of Negative Un-
perseverance.
dermining implies that, as one loses track of the justifications of one's
The foundations theory says people should keep track of their reasons
beliefs, one should give up those beliefs. But, if one does not keep
for believing as they do and should stop believing anything that is not track of one's justifications for most of one's beliefs, as seems to be
associated with adequate evidence. So the foundations theory implies
the case, then the Principle of Negative Undermining says that one
that, if Karen has not kept track of her reason for believing her history
should stop believing almost everything one believes, which is absurd.
course was an easy one, she should have abandoned her belief even On the other hand the Principle of Positive Undermining does not
before she was told about the mix-up with her aptitude test scores. have this absurd implication. The Principle of Positive Undermining
This seems clearly wrong.
does not suppose that the absence of a justification is a reason to stop
40 Chapter 4 Positive versus Negative Undermining 41

believing something. It only supposes that one's belief in P is under- the possibility that she may be violating not that principle but only
mined by the positive belief that one's reasons for P are no good. the foundationalist's Principle of Negative Undermining.
It is relevant that subjects can be successfully debriefed after experi-
ments involving deception if they are made vividly aware of the phe- Keeping Track of Justification
nomenon of beljef perseverance, that is, if they are made vividly aware
of the tendency for people to retain false beliefs after the evidence for People do not seem to keep track of the justifications of their beliefs.
them has been undercut, and if they are also made vividly aware of If we try to suppose that people do keep track of their justifications,
how this phenomenon has acted in their own case (Nisbett and Ross we would have to suppose that either they fail to notice when their
1980, p. 177). It might be suggested that this shows that under ideal justifications are undermined or they do notice but have great difficulty
conditions people really do act in accordance with the foundations in abandoning the unjustified beliefs in the way a person has difficulty
theory after all, so that the foundations theory is normatively correct abandoning a bad habit. Neither possibility offers a plausible account
as an account of how one ideally ought to revise one's beliefs. But in of the phenomenon of belief perseverance.
fact this further phenomenon seems clearly to support the coherence It stretches credulity to suppose people always keep track of the
theory, with its Principle of Positive Undermining, and not the foun- sources of their beliefs but often fail to notice when these sources are
dations theory, with its Principle of Negative Undermining. The so- undermined. That is like supposing people always remember everything
called process debriefing cannot merely undermine the evidence for that has ever happened to ·them but cannot always retrieve the stored
the conclusions subjects have reached but must also directly attack information from memory. To say one remembers something is to say
each of these conclusions themselves. Process debriefing works not one has stored it in a way that normally allows it to be retrieved at
just by getting subjects to give up beliefs that originally served as 'will. Similarly, to say people keep track of the sources of their beliefs
evidence for the conclusions they have reached but by getting them must be to say they can normally use this information when it is
to accept certain further positive beliefs about their lack of good reasons appropriate to do so.
for each of these conclusions. I have already remarked that the other possibility seems equally
incredible, namely, that people have trouble abandoning the under-
mined beliefs in the way they have trouble getting rid of bad habits.
What about Our Intuitions? To repeat, participants in belief perseverance studies show no signs of
It may seem to fly in the face of common sense to suppose that the knowing their beliefs are ungrounded. They do not act like people
coherence theory is normatively correct in cases like this. Remember struggling with their beliefs as with bad habits. Again, I agree it some-
that, after carefully considering Karen's situation, almost everyone agrees times happens that one keeps returning to thoughts after one has seen
she should give up all beliefs inferred from the original false report, there can be no reason to accept those thoughts. There are habits of
except those beliefs which would have been justified apart from any thought that can be hard to get rid of. But that is not what is going
appeal to evidence tainted by that false information. Almost everyone's on in the cases psychologists study under the name of belief
judgment about what Karen ought to do coincides with what the foun- perseverance.
dations theory says she ought to do. Indeed, psychologists who have This leaves the issue of whether one should try always to keep track
studied the phenomenon of belief perseverance in the face of debriefing of the local justifications of one's beliefs, even if, in fact, people do not
consider it to be a paradigm of irrationality. How can these strong seem to do this. I want to consider the possibility that there is a good
normative intuitions possibly be taken to be mistaken, as they must reason for not keeping track of these justifications.
be if the coherence theory is to be accepted as normatively correct?
The answer is that, when people think about Karen's situation, they Clutter Avoidance Again
ignore the possibility that she may have failed to keep track of the
justifications of her beliefs. They imagine Karen is or ought to be aware We have seen there is a practical reason to avoid too much clutter in
that she no longer has any good reasons for the beliefs she inferred one's beliefs. There is a limit to what one can remember, a limit to the
from the false report. And, of course, this is to imagine that Karen is number of things one can put into long-term storage, and a limit to
violating the Principle of Positive Undermining. It is hard to allow for what one can retrieve. It is important to save room for important things
42 Chapter 4
,.
and not clutter one's mind with a lot of unimportant matters. This is
an important reason why one does not try to believe all sorts of logical
consequences of one's beliefs. One should not try to infer all one can
from one's beliefs. One should try not to retain too much trivial in-
formation. Furthermore, one should try to store in long-term memory Chapter 5
only the key matters that one will later need to recall. When one reaches
a significant conclusion from one's other beliefs, one needs to remember Implicit Commitments
the conclusion but does not normally need to remember all the inter-
mediate steps involved in reaching that conclusion. Indeed, one should
not try to remember those intermediate steps; one should try to avoid
too much clutter in one's mind.
Similarly, even if much of one's knowledge of the world is inferred
ultimately from what one believes oneself to be immediately perceiving More on Positive Undermining
at one or another time, one does not normally need to remember these In chapter 4 I stated the Principle of Positive Undermining somewhat
original perceptual beliefs or many of the various intermediate con- roughly: "One should stop believing P whenever one positively believes
clusions drawn from them. It is enough to recall the more important one's reasons for believing Pare no good." Let us now try to be more
of one's conclusions. This means one should not be disposed to try to specific about the content of the undermining belief that one's reasons
keep track of the local justifications of one's beliefs. One could keep are "no good."
track of these justifications only by remembering an incredible number One suggestion would be that the relevant undermining belief is
of mostly perceptual original premises, along with many, many inter- simply the belief that one is not now justified in believing P. But this
mediate steps which one does not want and has little need to remember. cannot be right. Sometimes, thinking one's reasons for believing Pare
One will not want to link one's beliefs to such justifications because "no good" in the relevant sense, one follows the Principle of Positive
one will not in general want to try to retain the prior beliefs from which Undermining and concludes that one is not justified in believing P. In
one reached one's current beliefs. such a case one's belief that one's reasons are "no good" has to be
The practical reason for not keeping track of the justifications of different from the belief that one is "not justified." Otherwise, one
one's beliefs is not as severe as the reason that prevents one from could not get started. One could not reach the conclusion that one's
trying to operate purely probabilistically, using generalized condition- belief in P is unjustified without first having already reached that
alization as one's only principle of reasoned revision. The problem is conclusion!
not that there would be a combinatorial explosion. Still, there are im- This may suggest the relevant belief is that one's reasons are "no
portant practical constraints. It is more efficient not to try to retain good" in the sense that one was not originally justified in believing P
these justifications and the accompanying justifying beliefs. This leaves when one first formed that belief. One could first believe that and then
more room in memory for important matters. use the Principle of Positive Undermining to conclude that one is still
not justified in believing P. But this is not right either. Karen was
originally justified in reaching the various conclusions she reached on
the basis of the initial report about her aptitude test scores. Later, given
a full process debriefing that makes her vividly aware of what has
happened in her case, she may come to see that her reasons are "no
good," but that is not the same as coming to see she was originally
unjustified, since she was justified originally.
At one point in chapter 4 I stated the general principle like this: "It
is incoherent to believe both P and also that all one's reasons for
believing P relied crucially on false assumptions." This accounts for
Karen's situation after a process debriefing. At that point she realizes
she had been justified in accepting various conclusions only because
Bibliography

Achinstein, Peter (1983). The Nature of Explanation(New York: Oxford University Press).
Anderson, J. R., and Bower, G. H. (1973). Human Associative Memory (Washington, D. C.:
Winston).
Anscombe, G. E. M. (1957). /11/enlio11 (Oxford: Basil Blackwell).
Austin, J. L. (1946). "Other minds," Proceedingsof Ille Aristotelian Society, supp. vol. 20:
148-187.
Beardsley, Monroe (1978). "Intending," in Values and Morals, Alvin I. Goldman and
Jaegwon Kim, eds. (Dordrecht, Holland: Reidel), 163-184.
Bennett, Jonathan (1976). LinguisticBehaviour (Cambridge: Cambridge University Press).
Black, Max (1958). "Self-supporting inductive arguments," Journal of Philosophy
55:718-725.
Brandt, Richard (1979). A Theory of /lie Good and Ille Right (Oxford: Oxford University
Press).
Bratman, Michael (to appear). "Two faces of intention," Philosopl1icalReview.
Bromberger, Sylvain (1966). "Why-questions," in Mind a11dCosmos, R. G. Colodny, ed.
(Pittsburgh, Penn.: University of Pittsburgh Press), 86-108.
Castaneda, Hector-Neri (1972). "On the semantics of the ought to do," in Se111a11tics of
Natural Language,D. Davidson and C. Harman (Dordrecht, Holland: Reidel), 675-694.
Chisholm, Roderick (1966). "Freedom and action," in Freedoma11dDeterminism, Keith
Lehrer, ed. (New York: Random House), 11-44.
Davidson, Donald (l 963). "Actions, reasons, and causes," Journalof Philosophy60, 685-700.
Davidson, Donald (1973). "Freedom to act," in Essayson Freedomof Action, T. Honderich,
ed. (London: Routledge and Kegan Paul), 137-156.
Davidson, Donald (1978). "Intending," in P/1ilosophyof History a11dAction, Yirmiaku
Yovel, ed. (Dordrecht, Holland: Reidel).
Dennett, Daniel C. (1978). Brainstorms(Montgomery, Vt.: Bradford Books, Reprinted by
The MJT Press, Cambridge, Mass.).
Dorling, Jon (1972). "Bayesianism and the rationality of scientific inference," British
Journalfor Ille Philosophy of Science 23:181-190.
Doyle, Jon (1979). "A truth maintenance system," Artificial I11tellige11ce 12:231-272.
Doyle, Jon (1980). A Model for Deliberation,Actio11,and lutrospection, MIT Artificial In-
telligence Laboratory Technical Report 581.
Doyle, Jon (1982). "Nonmonotonic logics," in Ha11dbookof Artificial Intelligence, Paul R.
Cohen and Edward A. Feigenbaum, eds. (Los Altos, Calif.: William Kaufmann),
vol. 3, 114-119.
Dworkin, Ronald (1977). "Hard cases," in Taking Rights Seriously (Cambridge, Mass.:
Harvard University Press), 81-130.
Fodor, Jerry A. (1983). Modularity of Mind (Cambridge, Mass.: MIT Press).
 Bibliography 141
140 Bibliography
Kyburg, Henry (1961). Probability and the Logic of Rational Belief (Middletown, Conn.:
Foot, Philippa (1967). "Abortion and the doctrine of double effect," Oxford Review 5, Wesleyan University Press).
5-15. Lacey, A. R. (1976). A Dictionary of Philosophy (London: Routledge and Kegan Paul).
Franklin, Benjamin (1817). Private CorrespondenceI (London: Colburn). Lakatos, lmre ( 1970). "Falsification and the methodology of scientific research pro-
Gentzen, Gerhard (1935). "Untersuchungen iiber das logische Schliessen," Mathe111atiscl1e grammes," in Criticismand the Growthof Knowledge,Jmre Lakatos and Alan Musgrave,
Zeitschrift 39:493-565. Translated in CollectedPapersof Ger/1ardGentzen, M. E. Szabo, eds. (Cambridge: Cambridge University Press), 91-196.
ed. (Amsterdam: North-Holland, 1969), 68-113. Levi, Isaac (1967). Gambling with Truth (New York: Knopf).
Glymour, Clark (1980). Theory and Evidence (Princeton, N.j.: Princeton University Press). Lewis, David (1969). Convention (Cambridge, Mass.: Harvard University Press).
Goldman, Alvin I. (1967). "A causal theory of knowing," Journalof Philosophy64:357-372. Lewis, David (1975). "Languages and language," in Minnesota Studies in the Philosophy
Goldman, Alvin I. (1978). "Epistemology and the psychology of belief," Monist 61:525-535. of Science, Keith Gunderson, ed. (Minneapolis, Minn.: University of Minnesota Press),
Grice, H. P. (1957). "Meaning," PhilosophicalReview 66:777-788. vol. 7, 3-35.
Grice, H.P. (1969). "Utterer's meaning and intentions," PhilosophicalReview 78:147-177. Malcolm, Norman (1963). "Knowledge and belief," in Knowledgeand Certainty (Ithaca,
Grice, H.P. (1972). Intention and Uncertainty, British Academy Lecture (Oxford: Oxford N.Y.: Cornell University Press), 58-72.
University Press). McCarthy, John (1980). "Circumscription: A form of non-monotonic reasoning,'' Artificial
Grover, D. L., Camp, J.C., and Belnap, N. D. (1975). "A prosentential theory of truth," /11telligence13:27-39.
PhilosophicalStudies 27:73-125. McDermott, D., and Doyle, Jon (1980). "Non-monotonic logic I,'' Artificial Intelligence
Hacking, Ian (1979). "What is logic?" Journal of Philosophy 76:285-319. 13:41-72.
Hare, R. M. (1952). The Languageof Morals (Oxford: Oxford Univerity Press). Meyer, Robert K. (1971). "Entailment,'' Joumal of Philosophy 68:808-818.
Harman, Gilbert (1973). Review of Roger Wertheimer's The Significanceof Sense, Philo- Miller, George (1956). "The magical number seven, plus or minus two: Some limits on
sophical Review 82:235-239. our capacity for processing information,'' PsychologyReview 63:81-97.
Harman, Gilbert (1974). Review of Stephen Schiffer's Meaning, Journal of Philosophy Nagel, Thomas (1970). The Possibility of Altruism (Oxford: Oxford University Press).
71 :224-229. Nisbett, Richard, and Ross, Lee (1980). Human Inference:Strategies and Shortcomingsof
Harman, Gilbert (1976). "Practical reasoning," Review of Metaphysics 29:431-463. Social Judgement (Englewood Cliffs, N.).: Prentice-Hall).
Harman, Gilbert (1977). Review of Jonathan Bennett's Linguistic Behaviour, Language Peirce, C. S. (1877). "The fixation of belief," PopularScience Monthly 12:1-15. Reprinted
53:417-424. in Philosophical Writings of Peirce, Justice Buchler, ed. (New York: Dover, 1955),
Harman, Gilbert (1980). "Reasoning and evidence one does not possess," Midwest Studies 5-22.
in Philosophy 5:163-182. Pollock, John (1979). "A plethora of epistemological theories,'' in Justificationand K11owl-
Harman, Gilbert (1982a). "Conceptual role semantics," Notre Dame Journal of Formal edge, George Pappas, ed. (Dordrecht, Holland: Reidl), 93-114.
Logic 23: 242-256. Popper, Karl (1947). "Logic without assumptions,'' Proceedingsof the Aristotelian Society
Harman, Gilbert (1982b). "Logic, reasoning, and logical form," in Language,Mind, and 47:251-292.
Brain, Thomas W. Simon and Robert Scholes, eds. (New York: Erlbaum), 13-19. Popper, Karl (1959). Logic of Scientific Discovery (New York: Basic Books).
Harman, Gilbert (1983). "Human flourishing, ethics, and liberty," Philosophyand Public Ramsey, Frank (1931). "Truth and probability," in Foudnationsof Mathematics and Other
Affairs 12:307-322. Essays (London: Routledge and Kegan Paul), 156-198.
Hempel, C. G. (1960). "Inductive inconsistencies," Sy11these12:439-469. Rawls, John (1971). A Theory of Justice (Cambridge, Mass.: Harvard University Press).
Hempel, C. G. (1965). Aspects of Scientific Explanation (New York: Free Press). Raz, Joseph (1975). Practical Reason and Norms (London: Hutchinson).
Hempel, C. G. (1966). Philosophy and Natural Science (Englewood Cliffs, N.J.: Prentice- Reiter, Raymond (1980). "A logic for default reasoning,'' Artificial Intelligence 13:81-132.
Hall). Ross, Lee, and Anderson, Craig A. (1982). "Shortcomings in the attribution process: On
Herzberger, Hans G. (1982). "Naive semantics and the liar paradox," Joumal of Philosophy the origins and maintenance of erroneous social assessments,'' in Judgement under
79:479-497. Certainty: Heuristics and Biases, Daniel Kahneman, Paul Slovic, and Amos Tversky,
Horwich, Paul (1982). Probabilityand Evidence(Cambridge: Cambridge University Press). eds. (Cambridge: Cambridge University Press), 129-152.
Humberstone, I. L. (1971). "Two sorts of 'ought,'" Analysis 32:8-14. Schiffer, Stephen (1972). Meaning (Oxford: Oxford University Press).
Jeffrey, Richard C. (1969). "Statistical explanation versus statistical inference," in Essays Searle, John (1983). /11/entionality(Cambridge: Cambridge University Press).
i11 Honor of Carl G. Hempel, Nicholas Rescher, ed. (Dordrecht, Holland: Reidel), Sellars, Wilfrid (1963). "Some reflections on language games," in Science, Perception,a11d
104-113. Reality (London: Routledge and Kegan Paul), 321-358.
Jeffrey, Richard C. (1981). FormalLogic:Its Scope and Limits, second edition (New York: Shoesmilh, D. J., and Smiley, T. J.(1978). Multiple-ConclusionLogic(Cambridge: Cambridge
McGraw-Hill), chap. 7. University Press).
Jeffrey, Richard C. (1983). The Logic of Decision (Chicago: University of Chicago Press). Simon, Herbert A. (1969). Sciences of the Artificial (Cambridge, Mass.: MIT Press).
Kahneman, Daniel, Slovic, Paul, and Tversky, Amos, eds. (1982). Judgement1111der Certainty: Sosa, Ernest (1980). "The raft and the pyramid: Coherence versus foundations in the
Heuristics and Biases (Cambridge: Cambridge University Press). '\ theory of knowledge,'' Midwest Studies in Philosophy 5:3-25.
Kneale, William C. (1956). "The province of logic,'' in ContemporaryBritish Philosophy, Stalnaker, Robert C. (1984). Inquiry (Cambridge, Mass.: Bradford Books/MIT Press).
H. D. Lewis, ed., third series (London: Allen and Unwin), 235-262. Stebbins, Sarah (1980). "Necessity and natural language,'' PhilosophicalStudies 37:1-12.
Kripke, Saul A. (1975). "Outline of a theory of truth,'' Journal of Philosophy 72:640-715.
142 Bibliography

Steiner, Mark {1978). "Mathematical explanation," Philosophical Studies 34: 135-151.

Tarski, Alfred (1956). "The concept of truth in formalized languages," in Logic, Semantics,
Metamathematics (Oxford: Oxford University Press), 152-278.
Unger, Peter (1975). Ignorance (Oxford: Oxford University Press).
von Neumann, John, and Morgenstern, Oscar (1944). Theory of Games and Economic
Behavior (Princeton, N.j.: Princeton University Press).
Wertheimer, Roger (1972). Tire Significance of Sense (Ithaca, N.Y.: Cornell University
Press). Name Index
Williams, Bernard (1980). "Internal and external reasons," in Rational Action, Ross Harrison,
ed. (Cambridge: Cambridge University Press).
Ziff, Paul (1972). "What is said," in Semantics of Natural Language, Donald Davidson
and Gilbert Harman, eds. (Dordrecht, Holland: Reidel), 709-721.

Achinstein, Peter, 67 Hare, Richard M., 134

Anderson, Craig A., 35-37 Hempel, Carl G., 5, 48, 69, 70
Anderson, John R., 37 Herzberger, Hans G., 16
Anscombe, G. E. M., 6 Horwich, Paul, 25
Austin, John L., 50 Humberstone, I. L., 131

Bach, Kent, 81 Jeffrey, Richard C., 23, 26, 27, 69, 105,
Beardsley, Monroe C., 78-79, 82, 89 125
Belnap, Nuel D., 125
Bennett, Jonathan, 88 Kahneman, Daniel, 105
Black, Max, 4 Kneale, William C., 119
Bower, Gorden, 37 Kripke, Saul A., 16, 86
Brandt, Richard 8., 101 Kyburg, Henry, 23-24
Bratman, Michael, 90
Bromberger, Sylvain, 73-74 Lacey, A. R., 131
Lakatos, lmre, 4 7
Camp, Joseph, 125 Levi, Isaac, 72
Castaneda, Hector-Neri, 132 Lewis, David, 88
Chisholm, Roderick, 86
Malcolm, Norman, 48
Davidson, Donald, 79, 81, 86, 89, 94, McCarthy, John, 6
129, 132 McDermott, Drew, 6
Dennett, Daniel C., 13 Meyer, Robert K., 6
Dorling, Jon, 25 Miller, George A., 101
Doyle, Jon, 3-4, 6, 31 Mills, Claudia, 94
Dworkin, Ronald, 135 Morgenstern, Oscar, 104

Fodor, Jerry A., 118 Nagel, Thomas, 75-76, I 29

Foot, Philippa, 100 Nisbett, Richard, 40, 47, 105
Franklin, Benjamin, 101
Peirce, Charles S., 32
Gentzen, Gerhard, 119 Pollock, John, 29
Glymour, Clark, 24 Popper, Karl, 48, 119
Goldman, Alvin I., 37, 44, 88
Grice, H. P., 81-82, 87-88, 94 Ramsey, Frank, 104
Grover, Dorothy, 125 Rawls, John, 9
Raz, Joseph, 107
Hacking, Tan, 119 Reiter, Raymond, 6
Hampshire, Stuart, 93, 109 Ross, Lee, 35-37, 40, 47, 105