Uncertainty and Intuitive Judgment

Uncertainty is an unavoidable aspect of the human condition. Many significant choices must be based on beliefs about the likelihood of such uncertain events as the guilt of a defendant, the result of an election, the future value of the dollar, the outcome of a medical operation, or the response of a friend. Because we normally do not have adequate formal models for computing the probabilities of such events, intuitive judgment is often the only practical method for assessing uncertainty.

The question of how lay people and experts evaluate the probabilities of uncertain events has attracted considerable research interest. Much of this research has compared intuitive inferences and probability judgments to the rules of statistics and the laws of probability. The student of judgment uses the probability calculus as a standard of comparison much as a student of perception might compare the perceived size of objects to their physical sizes. Unlike the correct size of objects, however, the “correct” probability of events is not easily defined. Because individuals who have different knowledge or hold different beliefs must be allowed to assign different probabilities to the same event, no single value can be correct for all people. Furthermore, a correct probability cannot always be determined, even for a single person. Outside the domain of random sampling, probability theory does not determine the probabilities of uncertain events – it merely imposes constraints on the relations among them. For example, if A is more probable than B, then the complement of A must be less probable than the complement of B.

The Laws of Probability and the Conjunction Rule

The laws of probability derive from extensional considerations. A probability measure is defined on a family of events, and each event is construed as a set of possibilities. The probability of an event equals the sum of the probabilities of its disjoint outcomes. Probability theory has traditionally been used to analyze repetitive chance processes, but the theory has also been applied to essentially unique events for which probability is not reducible to the relative frequency of “favorable” outcomes. Additivity applies even when probability does not have a frequentistic interpretation and when the elementary events are not equiprobable.

The simplest and most fundamental qualitative law of probability is the extension rule: If the extension of A includes the extension of B (i.e., A ⊃ B), then P(A) ≥ P(B). Because the set of possibilities associated with a conjunction A&B is included in the set of possibilities associated with B, the same principle can also be expressed by the conjunction rule P(A&B) ≤ P(B): A conjunction cannot be more probable than one of its constituents. This rule holds regardless of whether A and B are independent and is valid for any probability assignment on the same sample space. Furthermore, it applies not only to the standard probability calculus but also to nonstandard models such as upper and lower probability, belief function, Baconian probability, rational belief, and possibility theory.

You can picture this with a Venn diagram: the overlapping region representing “feminist bank tellers” is a subset of the larger circle representing “bank tellers.” Any member of the overlapping region is automatically a member of the larger circle, but not vice versa. The probability that Linda is a bank teller equals the probability that she is a feminist bank teller plus the probability that she is a non‑feminist bank teller. Removing a possibility (non‑feminist bank tellers) can never increase the total probability. The same sum‑like structure applies to many everyday quantities, such as economic value: a set that contains all the items of another set plus additional positive items must be worth at least as much.

Yet, as we shall see, intuitive judgments of probability routinely violate this basic logical principle. The violations are neither rare nor subtle; they are robust, predictable, and deeply revealing about how the mind works under uncertainty.

The Representativeness and Availability Heuristics

In contrast to formal theories of beliefs, intuitive judgments of probability are generally not extensional. People do not normally analyze daily events into exhaustive lists of possibilities or evaluate compound probabilities by aggregating elementary ones. Instead, they commonly use a limited number of heuristics, such as representativeness and availability. Our conception of judgmental heuristics is based on natural assessments that are routinely carried out as part of the perception of events and the comprehension of messages. Such natural assessments include computations of similarity and representativeness, attributions of causality, and evaluations of the availability of associations and exemplars. These assessments are performed even in the absence of a specific task set, although their results are used to meet task demands as they arise.

For example, the mere mention of “horror movies” activates instances of horror movies and evokes an assessment of their availability. Similarly, the statement that Woody Allen’s aunt had hoped that he would be a dentist elicits a comparison of the character to the stereotype and an assessment of representativeness. Although these assessments are not tied to the estimation of frequency or probability, they are likely to play a dominant role when such judgments are required. The availability of horror movies may be used to answer the question, “What proportion of the movies produced last year were horror movies?”, and representativeness may control the judgment that a particular boy is more likely to be an actor than a dentist.

The term judgmental heuristic refers to a strategy – whether deliberate or not – that relies on a natural assessment to produce an estimation or a prediction. One of the manifestations of a heuristic is the relative neglect of other considerations. For example, the resemblance of a child to various professional stereotypes may be given too much weight in predicting future vocational choice at the expense of other pertinent data, such as the base rate frequencies of occupations. Hence, the use of judgmental heuristics gives rise to predictable biases. Natural assessments can also act as an anchor to which the required judgment is assimilated, even when the judge does not intend to use the one to estimate the other.

A striking demonstration of how natural assessments overpower extensional logic comes from a simple test of availability. When given 60 seconds to list seven‑letter words of a specified form, students produced many more words ending with ‑ing than words with the form ‑‑n‑‑ (where the sixth letter is “n”), although the latter class includes the former. The average numbers of words produced were 6.4 and 2.9, respectively. In a parallel judgment task, subjects estimated the expected number of such words in four pages of a novel. The median estimate for ‑ing words was 13.4, while for ‑‑n‑‑ words it was only 4.7, directly contrary to the extension rule. The increased ease of recalling ‑ing words overwhelmed the logical set‑inclusion relation. Similar results appeared for ‑‑ly versus l‑‑‑ words. This example captures the structure of many studies: a reduction of extension is paired with an increase in availability or representativeness, and the conjunction rule is put to the test.

The Conjunction Fallacy: Classic Demonstrations

The conjunction fallacy is a cognitive bias where individuals incorrectly judge that a specific combination of two events is more likely than one of the individual events alone, violating the conjunction rule. It is closely linked to the representativeness heuristic: people often substitute the easier question “How representative is this description?” for the harder question “What is the probability?”

The most famous demonstration involves Linda. Participants read:

Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice and also participated in anti‑nuclear demonstrations.

They are then asked which is more probable:

1. Linda is a bank teller.
2. Linda is a bank teller and is active in the feminist movement.

In the original experiments by Tversky and Kahneman, roughly 85% of people chose the conjunction, even though it is logically impossible. The description of Linda seems far more typical of a feminist than of a bank teller, so the specific, coherent scenario feels more plausible. The same pattern emerges for Bill, a character designed to be representative of an accountant but unrepresentative of a jazz player. Across numerous variations, the conjunction of the representative feature with the unrepresentative one was ranked as more probable than the unrepresentative component alone. The correlation between probability rankings and representativeness rankings was very high (0.96 for Bill, 0.98 for Linda), directly implicating the representativeness heuristic.

The conjunction fallacy is not confined to person perception. In a study on Wimbledon outcomes, participants ranked the probability of four events:

• Borg will win the match.
• Borg will lose the first set.
• Borg will lose the first set but win the match.
• Borg will win the first set but lose the match.

Even in direct comparison, 72% assigned a higher probability to “lose first set but win match” than to “lose first set,” although the latter is a broader category that includes the former. A non‑verbal version used a die with four green faces and two red faces, rolled 20 times. Participants chose among sequences for a prize; nearly two‑thirds preferred a longer sequence (GRGRRR) over a shorter one (RGRRR) even though the longer sequence is simply the shorter one with an extra G at the start—making it strictly less likely. The longer sequence looked more representative of a die with twice as many green faces.

The dinnerware problem, though about economic value rather than probability, reveals the same underlying mental operation. In joint evaluation, people correctly value a 40‑piece dinnerware set (with some broken items) more than a 24‑piece intact set, because the larger set contains all the intact pieces of the smaller set plus additional usable items. But in single evaluation, the average quality heuristic dominates: the broken pieces drag down the perceived value of the larger set, and people pay more for the smaller, flawless set. Probability is similarly sum‑like: a detailed conjunction that adds a plausible cause to an outcome can feel more probable because it forms a more coherent, representative story, even though it logically narrows the set of possibilities.

Conjunction Errors in Medical Judgment

The fallacy is by no means limited to laypeople. Practicing physicians given a clinical case of pulmonary embolism were asked to rank the probability of various symptoms. For a patient with a documented embolism, dyspnea (shortness of breath) is a typical symptom, while hemiparesis (partial paralysis) is very atypical. When the list included both the atypical symptom alone (hemiparesis) and the conjunction of the atypical symptom with the typical one (dyspnea and hemiparesis), physicians overwhelmingly judged the conjunction as more probable than the atypical symptom by itself. The incidence of violations across five medical problems ranged from 73% to 100%, averaging 91%. The correlation between judged probability and representativeness was over .95. When later confronted with their errors in a group discussion, the physicians did not defend their answers; most appeared surprised and dismayed. Expertise in a domain does not insulate against the representativeness heuristic.

Crucially, the instructions explicitly asked physicians to consider that the patient could experience more than one symptom, and nearly all confirmed they interpreted the task as intended. The fallacy cannot be dismissed as a simple misreading of the question.

Conjunction Errors in Forecasts and Scenarios

The construction and evaluation of scenarios of future events are a favorite pastime of analysts and influence major decisions. Our analysis suggests that a scenario that includes a possible cause and an outcome could appear more probable than the outcome on its own. We tested this hypothesis with statistically naive students and professional forecasters.

Students estimated the probability of “a massive flood somewhere in North America in 1983, in which more than 1,000 people drown.” A separate group estimated the probability of “an earthquake in California sometime in 1983, causing a flood in which more than 1,000 people drown.” The conjunction (earthquake and flood) was assigned a higher probability than the flood alone, with geometric means of 3.1% versus 2.2%. Adding a plausible, representative cause made the event seem more likely.

The same bias appeared among 115 participants at the Second International Congress on Forecasting, most of whom were professional analysts. “A complete suspension of diplomatic relations between the USA and the Soviet Union, sometime in 1983” was judged less probable than “A Russian invasion of Poland, and a complete suspension of diplomatic relations between the USA and the Soviet Union, sometime in 1983” (geometric means of 0.14% vs. 0.47%). Similar results were obtained for a drop in oil consumption paired with a dramatic oil price increase. Even seasoned forecasters fall prey to the illusion that a detailed, causally linked narrative is more likely than the outcome it describes.

Scenarios can usefully stimulate the imagination, but as a prime instrument for probability assessment they are highly misleading. A detailed scenario consisting of causally linked and representative events may appear more probable than a subset of those events. This effect contributes to the appeal of scenarios and the illusory insight they often provide. The attorney who fills in guesses about motive or mode of operation may strengthen a case by improving its coherence, although such additions can only lower probability.

Why the Conjunction Fallacy Occurs

Two main psychological forces drive the conjunction fallacy. First, detailed, coherent descriptions seem more probable. In everyday deception, a specific lie is often more convincing than a vague one. The thought of a city throwing a massive parade because its team played a football game feels absurd, yet a parade after a Super Bowl victory makes perfect sense. When asked which is more probable—the team won the Super Bowl, or the team played a football game—many choose the former, though every Super Bowl is a football game but not vice versa. The causal story makes the narrower event feel right.

Second, we search for causal explanations. The representativeness heuristic makes a conjunction whose components fit a causal script more compelling than an isolated, broader category. The ease with which a plausible scenario comes to mind (availability) and its fit to a stereotype (representativeness) trump the abstract logic of set inclusion.

The laziness of System 2 – the deliberate, analytical mind – is part of the explanation. Most people know the logic of Venn diagrams, but under casual conditions they are content to answer as if merely offering an opinion, letting the intuitive answer driven by representativeness suffice. If their vacation depended on a correct answer and they were given unlimited time and explicit instructions, they would likely avoid the error.

Resistance and the Role of Problem Format

A powerful manipulation almost eliminates the fallacy: asking “How many of the 100 participants have had one or more heart attacks?” instead of “What percentage…?” dropped the error rate from 65% to 25%. This frequency representation evokes a spatial image of individuals sorting into groups. You can imagine 100 men in a room: the heart attack victims gather in one corner, and among them, some are over 55 and some are not. The inclusion relation becomes visible – the set of men who are over 55 and had a heart attack is clearly a subset of all heart attack victims. This format helps System 2 see the logical structure that the percentage question hides. Some approaches emphasize that thinking in terms of frequencies makes nested sets salient, while others focus on how it triggers more careful reasoning about individuals.

However, not all transparent comparisons dispel the illusion. The dinnerware problem and the Linda problem share the same logical structure, yet joint evaluation immediately reveals the “more is better” logic for dinnerware, whereas with Linda, even direct comparison often fails to overcome intuition unless frequency representations or other aids are used. The stubbornness of the fallacy in probability judgments underscores the dominance of representativeness when numbers are abstract.

Individual Differences and Alternative Construals

Some have argued that the conjunction fallacy is not a genuine error but a rational response to a pragmatic reinterpretation of the statements. For instance, the phrase “Linda is a bank teller” might be interpreted as “Linda is a bank teller and not active in the feminist movement” because the detailed background suggests the experimenter knows a great deal about Linda. Under this construal, rating the conjunction higher no longer violates probability theory. When Tversky and Kahneman rephrased the alternative to “Linda is a bank teller, whether or not she is active in the feminist movement,” the fallacy rate dropped from 85% to 57%, but it did not vanish. Many subjects still commit the error even when the logical relation is made explicit.

If the conjunction fallacy were merely a reasonable pragmatic inference, one might expect that people with higher cognitive ability would be more likely to make this sophisticated interpretation. The evidence points the other way. In a study of 150 subjects, the mean SAT score of those who committed the fallacy was 82 points lower than those who avoided it. Higher cognitive ability was associated with greater resistance to the error, not greater susceptibility. Moreover, frequentist representations markedly reduce, and in some cases eliminate, cognitive ability differences in the fallacy, suggesting that the difficulty lies not in a subtle linguistic construal but in applying extensional logic spontaneously.

Similar patterns appear in other reasoning tasks, such as the Wason selection task, where cognitive ability differences are large on abstract, ambiguous problems but shrink on deontic versions that provide familiar content and clear logical structure. The overall picture is that many people fail to apply the conjunction rule not because they rationally reinterpret the problem, but because the intuitive pull of representativeness overwhelms the less accessible logic.

Logic versus Intuition: The Broader Picture

The contrast between extensional logic and intuitive heuristics runs through many domains of judgment. Probability judgments vary in the degree to which they follow a decompositional, frequentistic approach or a holistic, intuitive one. At one extreme, questions about the chances of beating a given hand in poker can be answered by calculating relative frequencies; at the other, the probability that a witness is telling the truth is normally evaluated holistically. The conjunction fallacy is a symptom of a more general phenomenon: people tend to overestimate the probabilities of representative (or available) events and underestimate the probabilities of less-representative events, even when true probabilities are unknown. The effects are often large. For example, subjects’ estimates of the frequency of ‑ing words were three times as high as their estimates of ‑‑n‑‑ words. A correction factor of three is the minimum needed to eliminate the inconsistency, and if one considers that many ‑‑n‑‑ words are not ‑ing words, the required adjustment would be even larger.

Interestingly, people generally accept the conjunction rule in its abstract form – they agree that B is more probable than A&B – but defy it in concrete examples when the rule conflicts with an intuitive impression of representativeness. This is the opposite of the familiar pattern where concrete, familiar contexts facilitate logical reasoning. It suggests that the representativeness heuristic is not a mere fallback when formal knowledge is absent; it is a powerful, automatically activated assessment that can override even well‑understood abstract principles.

The extension rule is perhaps the simplest and most transparent rule of probability theory, yet its violation proved stubborn even in direct‑transparent tests where subjects were forced to compare the conjunction with its component. In the classic studies, the conjunction fallacy was observed in indirect tests (where different subjects judged the two events), in direct‑subtle tests (where subjects compared the events but the inclusion relation was not emphasized), and even in direct‑transparent tests. The overall incidence of violations averaged 88% in the person‑perception problems, and the correlation with representativeness judgments remained near‑perfect.

Extensions of this work have shown that the same representativeness principle can produce disjunction fallacies – where a narrower category is judged more probable than its broader superordinate – provided the narrower category is more representative of the evidence. For instance, in one set of stories, a target basic‑level category (e.g., “clay court tennis”) was judged more probable than its superordinate (“sports”) when it was highly representative, while an unrepresentative basic‑level category was ranked lower than its superordinate. Probability judgments tracked representativeness, not level of categorization, confirming that the representativeness heuristic, not a general preference for basic‑level categories, drives extension fallacies.

Implications and Practical Safeguards

The conjunction fallacy has far‑reaching practical consequences. In the courtroom, it can lead to serious miscarriages of justice. During the O. J. Simpson trial, a Harvard law professor argued that because only 10% of men who batter their wives go on to murder them, the history of abuse was weak evidence. This confuses an unconditional probability with the relevant conditional probability (in murder cases where the wife was killed by her husband, roughly 50% had prior battering) – a close cousin of the conjunction fallacy. When multiple pieces of evidence are combined, their joint probability is the product of their individual probabilities, making the conjunction far less likely; yet jurors and lawyers often treat each corroborating detail as strengthening the case, effectively adding rather than multiplying improbabilities. The detailed, representative scenario overshadowed the logic of probability.

To guard against the conjunction fallacy, whenever comparing probabilities of scenarios, ask whether one option is a subset of another. A specific event can never be more likely than the general event that contains it. Thinking in terms of frequencies – “how many out of 100” – can make the inclusion relation visible. Conscious engagement of System 2, with enough time and motivation, can override the intuitive pull of a plausible conjunction. Recognizing the power of representativeness is the first step toward more accurate probabilistic reasoning, whether in psychological experiments, medical decisions, financial forecasting, or trials that shape lives.