Cognition, 11 (198;1) 143-157

@Elsevier Sequoia %A., Lausanne - Printed in The Netherlands

Variants of uncertainty*b

DANIEL KAHNEMAN
University of British Columbia

AMOS TVERSKY
Stanford University

Abstract

In contrast to formal theories of judgment and decision, which employ a

single notion of probability, psychological analyses of responses to uncer-
tainty reveal a wide variety of processes and experiences, which may follow
different rules. Elementary forms of expectation and surprise in perception
are reviewed. A phenomenological analysis is described, which distinguishes
external attributions of uncertainty (disposition) from in ternal attributions
of uncertainty (ignorance). Assessments of uncertainty can be made in dif-
ferent modes, by focusing on frequencies, propensities, the strength of argu-
ments, or direct experiences of confidence. These variants of uncertainty are
associated with different expressions in natural language; they are also sug-
gestive of competing philosophical interpretations of probability.

Analyses of uncertainty in philosophy, statistics and decision theory com-

monly treat all forms of uncertainty in terms of a single dimension of prob-
ability or degree of belief. Recent psychological studies of judgment under
uncertainty have often followed this tradition and have focused on the cor-
respondence of intuitive judgments to the standard logic of probability (Ein-
horn and Hogarth, 198 1; Kahneman, Slavic and TversPy, in press; Kahne-
man and Tversky, 198 1; Nisbett and Ross, 1980; Slavic, Fischhoff and
Lichtenstein, 1977). A comprehensive psychological perspective on uncer-
tainty, however, reveals a variety of processes and experiences, ranging from
such basic mechanisms as habituation to repeated stimulation in a single neu-
rone, to such complex activities as the evaluation of scientific hypotheses.
In this paper we sketch some extensions of the range of observations that
are normally considered in psychological analyses of judgments under uncer-
tainty. Two levels of responses to uncertainty are discussed. We first describe

*This work was supported by the Office of Naval Research under Contract N00014-79X-0077 to
Stanford University. Repriiit requests should be sent to Deniel Kahneman, Department of Psychology,
University of British Columbia, 2075 Wesbrook Mall, Vancouver B.C., Canada V6T lW5.
144 6) Kahneman and A. lbersky

some basic processes of expectation and surprise in perception, which can be

considered the precursors of subjective probability. We then turn to a phe-
nomenological examination, in which we distinguish internal from external
attributions of uncertainty and sketch four modes of judgment that people
may adopt in assessing uncertainty.

Elementary forms of probability

Uncertainty is a fact with which all forms of life must be prepared to con-
tend. At all levels of biological complexity there is uncertainty about the
significance of signs or stimuli and about the possible consequences of
actions. At all leuels, {ction must be taken before the uncertainty is resolved,
and a proper balance must be achieved between a high level of specific readi-
ness for the events that are most likely to occur and a general ability to res-
pond appropriately when the unexpected happens. Because the focus of the
present treatment is on belief rather than on action, we shall not discuss the
remarkable processes by which lower organisms distribute their response
effort in accordance with probabilities of reinforcement (Herrnstein, 1970).
Our principal concern in this section is with perceptual uncertainty.

Perceptual expectations
Before the event there are expectations. After the event there may be sur-
prise. Surprise has been studied mainly by psychophysical methods, and it
has been measured by the various indicators of the orienting response (Lynn,
1966; Sokolov, 1969) and by the P300 component of event-related poten-
tials (Donchin, Bitter and McCaUum, 1978; Duncan-Johnson and Donchin,
1977). Expectancies have been studied in many contexts, and by a wide
variety of methods.
Our discussion of perceptual expectancies will be organized ardund the
scheme shown in Figure 1, which distinguishes three main types of expecta-
tions. The first major distinction separates active from passive expectations:
an active expectation occupies consciousness and draws on the limited capa-
city of attention; in contrast, a passive expectation is automatic and effort-
less, and is better described as a disposition than as an activity (Posner, 1978).
Some expectancies are relatively permanent. Long-lasting expectancies
about covariations of attributes defme the perceptual categories that we use
to organize and encode experience (Broadbent, 197 1). Specific expectations
about objects, e.g., that rooms and windows are likely to be rectangular,
function as permanent assumptions which help determine the interpretation
 Variantsof uncertainty 145

Figure 1. Perceptualexpectancies.

Expectancies

\
Passive .%ctive
1 ‘mporary
(&arch)

/
Permanent Temporary
(categories and (priming)
assumptions)

of ambiguous stimuli (Ittelson and Kilpatrick, 195 1). We are chronically

better prepared for some events than for others, as illustrated by the robust
effect of past frequency on the recognition threshold for words (Broadbent,
1967; Morton, X969). Indeed, expectations sometimes produce hallucinatory
experiences that people cannot distinguish from real ones, as in the phone-
mic restoration effect. Thus, all the sensory information corresponding to
the ‘s’ in the word ‘legislature’ can be removed from a recording of the word,
and be replaced by a cough, or by some other natural sound. Subjects who
are exposed to this recording are utterly convinced that they heard the phan-
tom phoneme (Warren, 1970).
Passive and temporary expectancies mediate the large effects of context
on recognition (Foss and Blank, 1980), and several variants of priming effects
(Posner, 1978). For example, the inclusion of a letter in the warning signal
which introduces a trial facilitates the response to that letter in a speeded
matching task, even when the contingencies are so arranged that the warning
signal conveys no valid information about the target. Posner ( 1978) has docu-
mented some important differences between the passive expectation that is
set up by an uninformative warning signal and the active expectation which
is produced when the target is in fact predictable, albeit imperfectly, from
that signal. A passive expectation yields a benefit (i.e., a faster response)
when it is confirmed, but it does not impede the response to targets that
146 D. Kahneman and A. lnversky

have not been primed. In contrast, a signal which causes the subject to pre-
pare actively for a particular target also slows the response to unanticipated
targets. In the language of probability theories, active expectations obey a
principle of complementarity: a high degree of preparation for a particular
event is achieved at the expense of a loss of preparation for other event%
Passive priming is associated with a non-complementary pattern of beneGt
without cost.
Passive expectations and conscious anticipations can conflict, and there is
evidence that the passive process exerts greater influence on the interpreta-
tion of ambiguous stimuli. Epstein and Rock (1960) pitted the two types of
expectations against one another, using a picture in which a left-looking and
a right-looking profiie were joined to form a pattern of reversible figure-
ground organization. Observers of the composite picture only saw one of the
profiles, which appropriated the common contour. Having constructed two
profiles which could be joined in this fashion? Epstein and Rock presented
the profiles separately in regular alternation for a number of trials, creating
a conscious expectation that each would always be followed by the other.
The composite was then presented for the first time, and the face that the
subjects saw in it was recorded. In accord with the priming effect, the ob-
servers almost always saw the profile that had been shown on the preceding
trial, rather th;ln the one which they consciously expected to occur.
A related demonstration of a conflict between different levels of expec-
tation has been reported, irl which the P300 component of the EEG was the
mair, dependent variable. The P300 is a positive deflection in the EEG, which
occurs about 300 msec after the presentation of any stimulus that the ob-
server treats as significant or relevant to the task. Ma,ny careful studies have
demonstrated a close link between the prior probabilities tif events and the
magnitude of the P300 deflections that they elicit (Donchin, Ritter and Mc-
Callum, 1978). When a subject is exposed to a Bernoulli series, frequently
repeated events elicit a smaller P300 than do rare ones. Furthermore, a run
of repetitions of the same event is associated with a steadily decreasing P300,
suggesting an increase in the subjective probability of further repetitions. In
contrast, the conscious expectation of repetitions decrea! es consistently
during a long run, by the familiar gambler’s fallacy. Evidently, an observer
can be prepared, or ‘primed’ for one event while consciously expecting
another-and can show physiological evidence of surprise at the occurrence
of an event that was consciously predicted. Thus, there is a sense in which
an individual can have conflicting probabilities for the same event at the same
time. These observations suggest an image of the mind as a bureaucracy (Den-
nett, 1978) in which different parts have access to different data, assign them
different weights and hold different views of the situation.
 Variantsof uncertainty 147

Perception as (abet

Expectancies that have developed over a life-time of visual experience have a

profound effect on perception, and are strikingly inaccessible to conscious
knowledge or intention. The best-known demonstrations of these facts have
been developed by the transactional% students of perception (Ittelson and
Kilpatrick, 195 1; Kilpatrick, 1961). Observers of the famous distorted room
and rotating window are led to have visual experiences that contradict both
their general knowledge and their specific acquaintance with the objects of
the illusions. Thus, one’s friends may be seen as giants or midgets, who change
size as they walk along the wall of the distorted room, and a paper napkin
may appear to slice through the rotating window. These striking effects are
produced by the dominant assumption that rooms and windows are rectan-
gular. Although the observer knows quite well that the assumption is not
applicable to the case at hand, this knowledge has no si,gnificant effect on
conscious perception. Models of reality that have been built over the years
cannot be revised. on demand for a particular occasion. These observations
again confirm th,at an observer can simultaneously hold conflicting views
of the samti event.
We have noted that perceptual expectancies determine what we ‘see’ in an
ambiguous stimulus. Indeed, the transactionalists have interpreted percep-
tion as a bet on reality (Kilpatrick, 1961). A significant aspect of such per-
ceptual choices is the strong commitment to the chosen interpretation. Our
experience contains no indication of the equivocation of stimuli, and even
-when perceptual interpretations fluctuate over time, as with the Necker cube,
they tend to be quite definite at any particular moment. The suppression of
uncertainty and equivocation in perception suggests that we may be biologic-
ally programmed to act on the perceptual best bet, as if this bet involved no
risk of error. A significant difference between the conscious experiences of
perception and thought is that the latter can represent doubt and uncertainty,
while the former normally do not.
Although the suppression of uncertainty distinguishes perceptual bets from
conscious judgments about uncertain events, the processing of uncertainty at
the two levels may be similar in other respects. Two striking observations of
transactionalist research suggests hypotheses that seem to apply to conscious
beliefs. The first is that the reconstructed image of the environment tends to
be coherent, reflecting the normal constraints and dependencies among the
attributes of th.e scene and of the stimulus. Thus, when an object is presented
under conditions that make both its size and. its distance ambiguous, the
chosen perceptual interpretation will select a size and a distance that relate
to retinal size in the standard manner: if the object is perceived to be large,
I48 D. Kahwnan and A. TveFsky

then it also appears to be further away than if it is seen as small (lttelson and
Kilpatrick, 195 1).
The second observation is that perceptual construction appears to be a
hierarchical process, in which decisions about the global features of the scene
constrain and dominate decisions about the objects Icontained in it. The dis-
torted room provides the best example. What is seen is not a compromise
between two extreme views: normal-sized people in a distorted room, or
oddly-sized people in a normal room. The latter view simply dominates the
former, as if the shape of the room were computed before the processing of
the people in it begins. Whether similar rules can be shown to operate, :for
example, in the construction of scenarios of future events is a problem that
well deserves study.

The phenomenology of uncertainty

The preceding section sought to show that the rules that govern perceptual
expectancies differ from the rules of probability theory. The present section
extends this analysis to the experiences of doubt and uncertainty that judg-
ments of subjective probability are assumed to reflect. As we shall see, the
notion of probability refers in natural language to several distinct states of
mind, to which the rules of the standard calculus of probability may not be
equally applicable.
To appreciate the complexity of expectations, consider one of their mani-
festations: the surprise that we experience when an expectation is violated.
Imagine that a coin is to be tossed 40 times. What number of ‘heads’ would
you expect? If you assume that the coin is fair, you would probably state
that the 20-20 result is more likely than any other, yet you would be more
surprised by this outcome than by a result of 22 ‘heads’ and 18 ‘tails’. Is the
‘true’ subjective probability of the two events indicated by the considered
judgment of their relative likelihood, or by the involuntary reaction of sur-
prise which they would elicit?
One possible interpretation is that the example illustrates a conflict be-
tween two approaches to the judgment of probability: the judpment that the
most likely outcome is 20-20 derives from knowledge of the i-ules of chance,
but outcomes such as 22-18 or 17-23 are more probable at another level
where probability is determined by representat’veness. A slightly uneven out-
come represents both the fairness of the coin and the randomness of tossing,
which is not at all represented by the exactly even result. In this view, the
greater psychological reality of expectations based on repreientativeness
manifests itself in the suprise reaction.
 Variantsof uncertainty 149

A slightly different interpretation is possible, which focuses on the coding

of the possible outcomes. As we shall see, it is frequently appropriate in con-
versation to extend the definition of an event X to ‘X or something like it’.
If the spontaneous coding of events follows similar rules, outcomes such as
22-18 or 17-23 will be spontaneously coded as ‘an approximately even
split’, while the outcome 20-20 will be assigned a distinctive code of ‘exactly
even split’. A person who attempts to judge the relative likelihood of the
events will consider the explicit statement of the outcomes, and will note
that 20-20 is more likely than, say, 22- 18. But the reaction of surprise may
be determined by the natural coding of events. The event 22-l 8 will then be
relatively unsurprising because it is coded as ;n approximately even result,
which is indeed more likely than a precisely even one.
The role of event-coding is manifest in the interpretation of uncertain as-
sertions, such as ‘I estimate that. ..‘, or, sometimes ‘I think that.. .‘. Uncer-
tain assertions are a class of speech acts, which are ch,aracterized by specific
sincerity conditions and tests of validity. Ccnsi.der, for example, the predic-
tion: ‘I think that the price of gold will be higher by 50% in six months than
it is today’. Taken literally, this is a point-prediction, which should be as-
signed a very small probability of confirmation. But the prediction is not in-
tended to be taken literally. Point predictions are normally understood as
comparative statements, or as statements of the range in which an outcome
is expected to fall,, e.g., “I think the increase in the price of gold will be nearer
to 50% than to X% or Y%. The speaker and the listener normally expect to
agree on the tacitly implied values of X and Y. For example, the forecaster
cited above will be considered remarkably accurate if the price, of gold ac-
tually rises by 53% in the next six months, although the forecast was not
strictly true. Thus, a speaker who asserts a numerical prediction is committed
to a range rather than to a point. The speaker is also committed to the pro-
position that the value i: about equally likely to be above the estimate as
below it, except ,when the nature of the prediction makes this impossible.
Thus, a person who1 says ‘I think the price of gold will rise by 50% in the
next six months’ would be considered to be deliberately misleading if he or
she also thought, but did not communicate, that the actual value was much
more likely to be above the estimate than below it.
It is significant thLa,tthe sincerity conditions associated with a prediction
do tot require that lthe predicted value (or range) of a variable be considered
highly probable, but only that it be consitiered more probable than cornpa-
able values (or ranges). For example, a man who asserts ‘I think Billy John
wi;ti win the gold :medal for the high jump in the next Olympics’ will not be
co:nsidered a liar if he prefers to bet against this proposition rather than on
it, but he is prohibited from adding ‘and the chances of Jack Small are even
150 D. Ktihemn and A. l’versky

better’. Thus, natural language allolws a privileged role to the best guess, and
the identification of the favored guess conveys information about the alter-
natives to which it may fairly be compared. The mention of a favorite ath-
lete indicates that he is to be compared to other individual athletes, rather
than to a disjunction of possible winners. One consequence of this rule is
that it is sometimes possible to ‘predict’ an event which is considered less
probable than its complement, if the complement is naturally coded as a
disjunction.
A related restriction applies to expressions of confidence. A statement of
confidence expresses one’s uncertainty in a prediction, estimate or inference
to which one is already committed. Thus, it is natural to ask ‘how confident
are you that you are right ?‘, but it is anomalous to ask ‘how confident are
you that you are wrong ?‘. Confidence is the subjective probability or degree
of belief as,ociated with what we ‘think’ will happen.
Common language also provides a large number of expressions to talk of
events which may happen, although we do not necessarily ‘think’ they will.
Thus, people assess the chances of candidates, estimate the risks of different
activities, give odds for football games and understand forecasters’ statements
about the probability of rain. We now turn to a more detailed analysis of the
states of uncertainty which such statements may express, following the
scheme shown in Figure 2. The two levels of the figure, attributions of un-
certainty and variants of uncertainty, are discussed in the following sections.

A ttribu tions of uncertainty

The primary distinction shown in Figure 2 refers to two loci to which uncer-
tainty can be attr-outed: the external world or our state of knowledge. For
example, we attribute to causal systems in the real world the uncertainty
associated with the tossing of a coin, the drawing of a hand of cards from a
pack, the outcome of a football game and the behavior of the St. Helens
volcano. These causal systems have dispositions to produce diifferent events,
and we judge the probabilities of these events by assessing the relative strength
of the competing dispositions. In contrast, such statements as ‘I think Mt.
Blanc is the tallest mountain in Europe’ or ‘I hope I spelled her name cor-
rectly’ reflect an uncertainty that is attributed to one’s mind rather than to
a mountain or a woman. (Howell and Burnett [ 19781 have applied the terms
internal uncertainty and external uncertainty, respectively, to events that the
subjects can or cannot control.)
Our distinction between ignorance and external uncertainty is closely re-
lated to a more general distinction between internal and external attributions
of experience. Color, size and texture, for example, are normally experienced
 Variantsof uncertainty 15 i

Figure 2. Variantsof uncertainty.

Uncertainty

Distributional Singular Reasoned Introspkctive

(Frequencies) (Propensities) (Arguments) (Confidence)

as pr$?perties that be:long to exterr.dl objects, but pains, feelings and memo-
ries Bre attributed to the experiencing subject rather than to the eliciting
object.
The attribution of uncertainty can sometimes be inferred from a simple
linguistic test: is it alppropriate to describe the assessment of uncertainty as
‘the probability is.. , ‘? Or shou!d one say ‘my probability is.. . ‘? In contrast
to the Bayesian view, which treats all probabilities as subjective and personal,
natural language marks the distinction between internal and external uncer-
tainty. Thus it is legitimate to speak of ‘the best estimate of the probability
of a change of regime in Saudi Arabia within the next year’, but it is anoma-
lous to say ‘the best estimate of the probability that the Nile is the largest
river in the world is.. . ‘. Best estimates of probability belong to the public
domain. Expressions af private ignorance do not.
This test does not always distinguish internal from external uncertainty.
For example, one may speak of the probability that Marlowe wrote Hamlet
although this uncertainty is attributed to our ignorance rather than to the
strength of Marlowe’s propensity to write plays. The use of ‘the probability’
in this example is justified by the existence of a public body of knowledge,
which reduces but doles not eliminate the uncertainty about the authorship
of Hamlet. Not everybody need have access to this knowledge, but the esti-
mated probability refers to a reasonable or consensual inference from the
152 D. Kahnemanand A. TversIg

available evidence. In the example of the Nile, however, the public body of
evidence certainly includes the correct answer, and ignorance can only be
private.
The attribution of uncertainty about an event to dispositions or to igno-
rance depends, among other things, on timing. Uncertainty about past events
is likely to be experienced as ignorance, especially if the truth is known to
someone else, whereas uncertainty about the future is more naturally attri-
buted to the dispositions of the relevant system. Indeed, it has been noted
that people exhibit different attitudes to the outcome of a coin toss, depend-
ing on whether or not the coin has already been tossed (Rothbart and Snyder,
1970).

Variants of uncertainty
The second level of Figure 2 distinguishes four prototypical variants of un-
certainty, identified by the nature of the data that the judge might consider
in evaluating probability. External uncertainty can be assessed in two ways:
(i) a distributional mode, where the case in question is seen as an instance of
a class of similar cases, for which the relative frequencies of outcomes are
known, or can be estimated; (ii) a singular mode, in which probabilities are
assessed by the propensities of the particular case at hand. The two modes
of judgment are illustrated by the following true story.
A team that was concerned with the development of a high school curric-
ulum on thinking under uncertainty was conducting a planning session. The
question was raised of the time that would be required to complete the first
version of a textbook. The participants in the discussion were asked to esti-
mate this value as realistically as possible; the seven estimates ranged from
18 months to three years. The team leader then turned to one of the partici-
pants, an educator with considerable expertise in the problems of curriculum
development, with the following question: ‘What has been the experience of
other teams that have tried to write a textbook and develop a curriculum in
a new arca, where no previous course of study existed? How long did it take
them to complete a textbook, from a stage comparable to the present state
c c our project ?“. The chilling implications of the answer appeared to surprise
tne expert who gave it, much as they surprised the other participants: ‘Most
teams I could think of failed, and never completed a textbook. For those
that succeeded, completion times have ranged from five to nine years, with
a median of seven’.
Subsequent probing revealed that all participants had produced their ini-
tial estimate in the singular mode, by constructing plans and scenarios, with
SOme allowance of safety margins for unforeseen contingencies. Because of
 Varim ts ofuncertainty 153

anchoring effects (Tversky and Kahneman, 1974) an estimate which is ob-

tained by adding safety margins to current plans is likely to be highly opti-
mistic. A notable aspect of this anecdote is that the relevant distributional
information was not spontaneously used, although it was available to one
expert from personal knowledge, and could have bee0 estimated quite accu-
rately by several other participants.
Our example illustrated the application of singular and distributional
modes of reasoning to the prediction of a continuous variable: the timt! re-
quired to complete a project. The distributional information cons&ted in
this case of knowledge about the relative frecluencies of different completion
times. Of course, the same reasoning can be applied to assess the probability
of a discrete outcome, such as the failure of tLe project. The relative fre-
quency of that outcome in a relevant class provides the basis for a distribu-
tional assessment of probability, and other ir:formation about the particular
case, used in the singular mode, may produce an impression of propensity
to fail or to succeed. There are niany irrstances in which the same question
can be approached in either singular or distributional mode.
Compare the following examples:
(1) ‘Chances are that you will find John at home if you call tomorrow
morning. He said that he prefers to work at home’.
(2) ‘Chances are that you will find John at home if you call tomorrow
morning. He has often been there when I called him’.
Statement (1) only allows a singular judgment of the probability that John
will be at home. Statement (2) could support both a distributional and a
singular assessment. The relative frequency of similar mornings on which
John has been at home provides a natural estimate of the probability of find-
ing him there tomorrow, but the statement has also endowed John with a
propensity to spend mornings at home, much as did Statement (1).
We have conjectured (Kahneman and Tversky, 1979) that people generally
prefer the singular mode, in which they take an ‘inside view’ of the causal
system that most immediately produces the outcome, over an ‘outside view’,
which relates the case ;at hand to a sampling schema. Our planning example
illustrates this preference for the singular mode. It also illustrates another
effect, which we suspect to be quite general: that the distributional mode of
judgment is more likely than the singular to yield accurate estimates of values
and reasonable assessments of probability.
We now turn to a distinction between two modes of assessment of inter-
nal uncertainty, which are illustrated by the following examples:
(3) ‘I believe New York is north of Rome, but I am not sure’.
(4) ‘I think her name is Doris, but I am not sure’.
154 D. Kahneman and A. Tvershy

The uncertainty expressed in ,these statements is clearly internal: the state-

ments reflect (partial) ignorance rather than dispositions of external objects.
]It is surely far-fetched to speak of the propensity of New York to be north
of Rome (incidentally, it is not), or of Linda to be remembered as Dar%.
The two statements differ in the nature of the evidence on which they are
based. Statement (3) could reflect a process of sifting and weighI;tg of evi-
dence and arguments (e.g., New York is much colder than Rome; Rome is in
the middle of Ltaiy, etc.). Statement (4) has a different character. The confi-
dence that it expresses is based on an introspective judgment of the strength
of an association. Much as happens when we check the spelling of a word by
examining whether it “looks right’, confidence rests on an unanalyzed experi-
ence. In studies of psychophysic.s and of memory, the confidence associated
with judgments is significantly correlated with accuracy: people are more
likely to be confident when they are correct than when they are not, although
their assessments of the probability that they are right are poorly calibrated
(see Lichtenstein, Fischhoff and Phillips, in press).
As in the case of external uncertainty, the internal uncertainty associated
with a given question can sometimes be assessed both in the reasoned and in
the introspective modes. For example, a question concerning the age of a
movie star can be approached introspectively by searching for an answer that
sounds familiar, or in a reasoned mode by trying to induce the answer from
other knowledge.
We do not wish to suggest that any experience of uncertainty can be as-
signed to one of the four variants of Figure 3. There are undoubtedly many
mixed and indeterminate cases. We have seen that the uncertainty in a given
problem can be attributed to external dispositions, to one’s ignorance, or to
a combination of the two, and that it may be assessed in a singular made, in
a distributional mode, or in a mixture of modes. The purpose of our treat-
ment was to highlight some significant dimensions of variation in experiences
of ur certainty, not to offer an exhaustive and mutually exclusive classifica-
tron of these experiences” For an attempt to classify experimental oljerations
in the measurement of subjective probability, see Howell and Burnett (1978).

Discussion

Although the language of probability can be used to express any form of un-
certamty, the laws of probability theory do not apply to all variants of un-
certainty with equa1 force. These laws are most likely to be accepted, and
satisfied in intuitivG judgments, when an external uncertainty is assessed in a
distributional or frequent&tic mode. For example, complementarity of sub-
 Variantsof uncertainty 15 5

jective probability is very compelling when we consult weather statisticsin

order to assess the prob?.bility that it will rain next year on April 12: the
relevant set of past April days is clearly separable into days on which there
was rai:r and days on which there was not.
Complementarity is less compelling in other variants. When uncertainty is
assessed in terms of propensities, arguments or confidence, it is less obvious
that the probabilities should add up to unity-even if it is known with cer-
tainty that one of the alternatives is correct. For example, one may question
why the degree f9f belief in the assertion that New York is north of Rome
‘and the degree of belief in the assertion that New York is south of Rome
should sun to the same value as the degrees of belief for any other pair of
complementary statements. Indeed, several authors (e.g., Cohen, 1977,
Shafer, 1976) have proposed that complementarity should not apply to de-
gree of belief. In particular, Shafer has argued against complementarity 01’
belief on the grounds that there are situations in which two mutually exclu-
sive and exhaustive hypotheses both have substantial support, and other
situations in which neither hypothesis has much support. Similar questions
car:ld be raised about the necessity of complementarity in impressions of
confidence, and in assessments of conflicting propensities.
The variants of uncertainty may differ in the confidence with which they
are assessed. Imagine that a thumbtack has been tossed four times, and has
landed twice on its point and twice on its head. Given these data, most ob-
servers will assign a probability of 0.5 to the event that the thumbtack will
land on its head on the next toss. They also assign a probability of 0.5 to the
event that a tossed coin will show ‘heads’, but express much greater confi-
dence in their judp nent about the coin than about the tack. As this example
illustrates, it is c*,lte possible to assign different degrees of confidence to the
same judgment of propensity. Confidence about probabilities is important
because it controls decisions. There is evidence (Ellsberg, 1961; Raiffa, 196 1)
that people prefer to bet on events that have known probabilities, such as
the toss of a coin, rather than on events that are associated with a combina-
tion of external uncertainty and ignorance, such as the toss of a thumbtack.
There are natural links between the conceptiorls of probability advanced
by different bchools of thought on this topic and the modes of uncertainty
that we have discussed. Thus, the frequentistic or objective interpretation of
probability restricts the concept to external uncertainty generated by a sam-
pling process. In contrast, the Bayesian or personal school treats all uncer-
tainty as ignorance. In the Bayesian school, preferences are the basis of be-
liefs, and probabilities are derived from preferences between bets. From a
psycho!.ogical point of view, however, this betting hecristic appears unreal-
istic. Controversy has often been sharp in this domain, because of the exis-
156 0 Kahneman and A. Tversky

tence of intuitions which are individually compelling and mutually incompa-

tible, and because there is no agreed criterion for settling normative disputes
when intuitions conflict. A psychological aniklysis could perhaps contribute
to the normative discussion, by providing an adequate description of the
intuition from which the various positions draw their appeal.

References

Broadbent, D. E. (1967) Word-frequency effect and response bias. Psychol. Rev., 74, 1- 15.
Broadbent, ;y. E. (197 1) Decision and Stress. London, Academic Press.
Cohen, 1. J. (1977) The Plobuble and the Provable. Oxford, Clarendon Press.
Dennett, D. C. (1978) Brainstorms. Cambridge, Bradford Books.
Donchin, E., Ritter, W. and McCallum, W. C. (1978) Cognitive psychophysiology: The endogeneous
components of the ERP. In E. Callaway, P. Tueting and S. H. Koslow (eds.), Everwelated
Brain Potentials in Man. NEWYork, Academic Press.
Duncan-Johnson, C. C. and Donchin, E. (1977) On quantifying surprise: The variation of event-related
potentials with subjective probability. Psychophysiol., I4,456-467.
Einhorn, H. J. and Hogarth, R. M. (1981) Behavioral decision theory: Processes of judgment and
choice. An. Rev. Psychol., 32, 53-88.
Elksberg, D. (1961) Risk, ambiguity and the Savage axioms. Q. J. Econ., 75, 643-669.
Epstein, W. and Rock, i. (1960) Perceptual set as an artifact of recency. Amer. J. Psychol., 73, 214-
228.
Foss, D. J. and Blank, M. A. (1980) Identifying the speech codes. Cog. Psychol., 12, 1-31.
Herrnstein, R. J. (1970) On the law of effect. J. exper. Anal. Behav., 13, 243-266.
Howell, W. C. and Burnett, S. A. (1978) Uncertainty measurement: A cognitive taxonomy. Orguniz.
Behav. Hum. Perf.. 22.45-68.
Ittelson, W. H. and Kilpatrick, F. P. (1951) Experiments in perception. Sci. Amer., 18.5, 50-55.
%&neman, u., Slavic, P. and Tversky, A. (eds.) (in press) Judgmen? under Uncertainty.’ Heuristics
and Biases. New York, Cambridge University Press.
KSneman, D. and Tversky, A. (1979) Intuitive prediction: Biases and corrective procedures. TiMS
Studies in Management Science, I 2, 3 13 - 3 27.
Kahneman, D..and Tversky, A. (1982) On the study of st ltistical intuitions. Cog., II, 123-141.
Kilpatrick, F 3. (1961) Explorations in Transactional Psychology. New York, New York University
Press.
Lichtenstein, S., Fischhoff, B. and Phillips, L. (in press) Calibration of probabilities: The state of the
art to 1980. In D. Kahnemari, P. Slavic and A. Tversky (eds.), Judgment under Uncertainty:
Heuristics and Biases. New York, Cambridge University Press.
Lynn, R. (1966) Attention, Arousal and the Orientation Reaction. Oxford, Pergamon.
Morton, J. (1969) Interaction of information in word recognition. Psychol, Rev., 76, 165-178.
Nisbett. R. E. and Ross, L. (1980) Human Inference Strategies and Shortcomings. Englewood Cliffs,
NJ, Prentice-Hall_
Poster, M. 1. (1978) Chronometric Explorations of Mind. Hillsdale, NJ, Lawrence Earlbaum Associates.
wfa, H. (196k) Risk, ambiguity and the Savage axioms: Comment. Q. J. Econ., 7.5 690-694.
Rothbart, M. and Snyder, hf. (1970) Confidence in the prediction and postdiction of an uncertain
event. &tad. J. behav. Sci., 2, 38-43.
Shafer, G. (1976) A Mathematical Theory of Evidence. Princeton, Princeton University Press.
 Vatian ts of uncertainty 157

Slavic, P., Fischhoff, B. and Lichtenstein, S. (1977) Behavioral decision theory. An. Rev. Psychol., 28,
l-38.
Sokolov, E. N. (1969) The modeling properties of the nervous system. In I. Maltzman and K. Cole
(eds.), Handbook of Contemporary Soviet Psychology. New York, Basic Rooks.
Tversky, A. and Kahneman, D. (1974) Judgment under uncertainty: Heuristi and biases. Science,
185, 1124-1131.
Warren, R. M. (1970) Perceptual restoration of missing speech sounds. Science, Z67, 393-395.

Rt%umr

Contrastant avec les theories formelles de jugement et de decision qui utilisent une notion unique de
probabilit8, les analyses psychologiques des rcponses g Pincertitude indiquent une grande varieti dans
les processus et les experiences qui peuvent ob& 6 diffirentes rGgles. On a passe en revue les formes
616mentaire.s d’attente et de surprise en perception. L’analyse ph&romt%rologique prCsent6e distingue
les attributions externes (disposition) des attributions internes (ignorance) de I’incertitude. L’estima
tion de l’incertitude peut refaire selon diff&ents modes selon qu’elle se centre sur les frequences, lcs
tendances, la force des arguments, une conviction issue d’une experience directe, etc. Ces variatiors
d’incertitude sont associ&es avec diffbrentes expressions du langage naturel, clles suggtrent cgaiemeni
des interp&tations philosophiques concurrentes de la probabilit6.