A team is deciding whether a new service needs another safety check. One incident comes vividly to mind: the account is recent, detailed and troubling. Another risk has a quieter record across hundreds of cases. The people in the room must decide before all the evidence can be gathered. Their problem is not that they have an imperfect brain while a perfect decision waits somewhere outside it. They need a workable judgment under limited time, information and attention.
Amos Tversky and Daniel Kahneman’s 1974 paper, Judgment under Uncertainty: Heuristics and Biases, changed how such judgments were studied. It described three common shortcuts: assessing by resemblance, assessing by ease of recall, and starting from an initial value before adjusting. Each can be useful. Each can also produce a predictable error under particular conditions. The paper did not discover that all intuition is false, and it did not offer a personal diagnosis for anyone who uses a shortcut. It posed experimental questions about how people estimate likelihood when a full calculation is absent.
The question the paper asked
Judging probability is often hard because the relevant events have not occurred often enough in one person’s experience to supply a reliable count. Even where counts exist, they may be hard to retrieve or interpret. Tversky and Kahneman asked what people do instead. Their answer was not one general bias but a set of heuristic operations. A person may ask whether a case looks like a typical member of a category, whether similar cases come readily to mind, or how far an answer should move from a number already in view.
In laboratory questions these operations can be exposed by setting intuitive impressions against a known statistical relation. If a description is vivid but the described occupation is rare, representativeness can compete with the base rate. If one kind of event is easy to recall because it is dramatic, availability can compete with its actual frequency. If an arbitrary starting number enters before a quantitative estimate, later adjustment may leave the final estimate closer to that number than it should be. These are specific patterns of judgment, not proof of a global failure of reasoning.
The paper was a synthesis of experimental demonstrations and arguments, published in Science in 1974. Many later studies developed, challenged and qualified its examples. Reading it now requires two moves together: recognizing how influential the three mechanisms have been, and asking whether a particular modern decision really shares the conditions of a given demonstration.
Representativeness: when resemblance outruns frequency
Representativeness is the tendency to assess likelihood partly by how much an instance resembles a familiar type. A reader may think a detailed description sounds like an engineer more than a teacher. The description could be informative, but its usefulness depends on how likely each profession was before the description and how frequently the traits occur in each group. A vivid fit alone cannot settle the probability. In a population with many more teachers than engineers, the prior distribution matters.
The useful question is not whether resemblance should be ignored. A suspicious transaction that closely matches a known fraud pattern may warrant attention. The question is how diagnostic the match is. If the pattern is common among legitimate transactions too, resemblance carries less information than it appears to. A base rate is not a moral instruction to treat every individual as an average; it is part of the probabilistic background against which a clue has meaning.
Representativeness can also encourage insensitivity to sample size. A pattern in ten observations can look as convincing as a similar pattern in a thousand when the surface arrangement is equally striking. Yet small samples fluctuate more. A sequence of coin tosses that looks too orderly to be random can still occur under a random process. The appearance of randomness and its statistical properties are different things. Tversky and Kahneman’s research made this distinction experimentally visible.
Later work on the conjunction fallacy pushed a related idea further. A richly described combination of events can seem more probable than one of its components, despite the mathematical fact that a conjunction cannot exceed either component in probability. That finding involves task wording and interpretation as well as intuition. The lesson for a reader is to check the formal relation before interpreting an apparent violation as evidence about a person’s entire reasoning ability.
Availability: what comes to mind is not a count
Availability concerns ease of retrieval. A recent emergency may come to mind faster than a common but uneventful failure. The first memory is real, and its emotional force may signal a reason to investigate. But it is not automatically a frequency estimate. Media attention, personal proximity, repeated discussion and vivid imagery all influence which examples are retrievable. A decision about resource allocation needs a denominator as well as a memorable numerator.
In the opening service example, a team could document the vivid incident without allowing it to become the whole evidence base. How many similar cases occurred? Over what period, among how many users, and under what conditions? What harms were missed because they were less visible? These questions do not dismiss testimony. They place the testimony beside systematic information and preserve its particular meaning.
Availability may be helpful when memory accurately tracks a changing environment. A clinician’s rapid recall of a rare complication may be valuable if recent local cases signal a real shift in practice or exposure. A worker’s recollection of repeated near misses can expose a pattern that an official database has failed to capture. The heuristic is not always a bias. It becomes a risk when ease of recall is treated as a direct measure of population frequency without checking why some examples are easier to retrieve.
This boundary matters especially in public discussion. A vivid case can motivate attention to a neglected harm; a base-rate table can reveal scale. Neither replaces the other. The sound judgment asks what each form of evidence can support. Statistics without case detail can hide mechanisms, while a compelling case without context can misstate prevalence. A practical review can include both, then state which decision criterion is being used.
Anchoring: an initial number can shape the answer
Anchoring describes the pull of a starting point on a later numerical estimate. The starting point may be a suggested price, a forecast, a target or even a number with no genuine relevance to the question. The 1974 paper presented experiments in which people adjusted away from an initial value but did not move far enough to erase its influence. Such a result is evidence about a task and its comparison conditions. It does not mean every number encountered creates the same degree of bias.
In an organization, an early estimate can become the reference around which a meeting revolves. If the first forecast was produced with uncertain assumptions, the group may debate whether to add or subtract from it rather than rebuild the estimate from independent evidence. One safeguard is to ask several people for an estimate before showing the proposed number. Another is to record the assumptions that would make the starting point reasonable. These are design suggestions derived from the mechanism, not interventions directly tested in the 1974 paper.
Anchors are not always arbitrary. A prior-year budget can be a useful starting point when the task genuinely resembles last year’s. A previous clinical result can be informative when the same patient’s condition is being tracked. The problem is undisclosed dependence: the answer appears to be an independent assessment while it still reflects the initial value. Good judgment asks whether the anchor is relevant, how uncertain it is and how much new evidence should move the estimate.
A heuristic is a method, not a character flaw
The word bias can sound like a moral verdict. In this literature it usually refers to a systematic deviation from a specified benchmark in a specified task. The benchmark might be probability theory, a known sampling process or an experimentally defined correct relation. Outside that task, the relevant standard may be harder to specify. A rough rule that saves time while preserving an acceptable decision can be adaptive under real constraints. A formally exact calculation may be impossible if probabilities are unknown or the data are poor.
Gigerenzer and Hoffrage’s research on frequency formats shows why the representation of evidence matters. Some Bayesian inference problems are answered more accurately when the same information is expressed as natural frequencies rather than abstract percentages. This does not prove that every judgment problem disappears with a better chart. It shows that performance can change when a task fits the way information is encoded. Calling a participant irrational without examining the information format can miss a design problem.
Kahneman and Gary Klein’s later discussion of intuitive expertise also complicates a simple anti-intuition story. Skilled intuition can develop where the environment contains valid, learnable cues and the person receives enough timely feedback to learn them. Where cues are weak, outcomes delayed or feedback misleading, confidence may grow without accuracy. The distinction concerns conditions of learning, not a personality type called intuitive or analytical.
For the service team, the question is therefore not whether to ban experience. Experienced staff may detect a subtle failure mode that a dashboard omits. Their observations should be recorded in a form that permits checking. A team could compare prior estimates with outcomes, distinguish verified cases from suspicions and review whether a highly memorable incident is rare or representative. The aim is to improve the environment for judgment, not to demand that people become machines.
What counts as a good correction?
A correction should match the possible error. When resemblance dominates, ask about base rates and whether the described feature is genuinely diagnostic. When a vivid memory dominates, seek a complete count with a clear denominator and time period. When an early number dominates, elicit independent estimates and make assumptions visible before negotiating around it. The intervention is a question or information practice tied to a mechanism, not a generic instruction to “think harder.”
Even these practices have limits. A base rate compiled from unequal reporting can be misleading. An apparently comprehensive incident log may systematically omit harm to people who have less access to complaint routes. Independent estimates may share the same hidden source. Good statistical practice includes asking who was counted, how a category was defined and which events were never recorded. A bias checklist can become another shortcut if it treats every number as clean.
Decision stakes also matter. A low-stakes choice may not justify a formal analysis. A high-stakes allocation affecting safety, employment or access to services should carry more documentation and review. The 1974 article does not prescribe a universal level of effort. Its contribution is to identify repeatable ways judgment may drift, which can help institutions decide where checks are worth their cost.
There is a social dimension to error correction. If a team treats a person who raised an alarming case as “biased,” it may suppress the very observations needed to understand the problem. A better review separates the event from the inferred prevalence. The incident can be taken seriously while its frequency and causal meaning remain open. This allows both testimony and quantitative evidence to contribute without asking either to carry more than it can.
What the evidence does not show
The original paper does not show that people are uniformly irrational, that every intuitive decision is wrong or that every statistical model outperforms experienced judgment. Its demonstrations concern defined tasks and comparisons. The three heuristics are not an exhaustive theory of choice, and a single answer on a laboratory item does not establish a stable trait in the respondent.
The paper does not prove that a quick workshop can remove biases from complex organizations. Nor does it establish that a particular platform, profession or demographic group is more biased than another. Moving from an experimental finding to a workplace procedure requires testing the procedure, its costs, its unintended effects and its performance in that setting. An intervention that corrects one error can create another if it silences local knowledge or shifts attention to an unreliable metric.
Research on natural-frequency formats and intuitive expertise suggests that errors depend partly on how information is presented and learned. Those findings do not cancel the 1974 demonstrations. They set boundaries around a broad interpretation. The defensible conclusion is that judgment should be designed with its task, evidence, feedback and stakes in view.
Reading a claim about bias
When a report says a decision was distorted by availability, ask what information the decision maker saw and whether a relevant frequency was available. If it invokes representativeness, ask what base rate and diagnostic information were compared. If it invokes anchoring, ask what the starting value was and whether the comparison actually tested its effect. Naming a heuristic after an unpopular result is not an explanation; the proposed mechanism needs evidence.
A useful decision record can state the first impression, the additional evidence obtained and the final rationale. It can acknowledge uncertainty without treating uncertainty as permission to decide arbitrarily. In some cases the right response is to delay, collect data or create a reversible trial. In others, action must occur with incomplete evidence. Then the institution should record what would change its mind and build a route for correction.
Return to the opening team. The disturbing incident should be investigated. The quieter pattern should be counted. The first estimate should be marked as provisional. Someone should ask whether the choice criterion is to minimize average harm, avoid catastrophic cases, preserve equity or meet a legal duty, because those goals can lead to different choices even with the same probabilities. Heuristics research cannot choose the goal for the team. It can help the team see where its estimate came from and where it may need repair.
That is the enduring value of Tversky and Kahneman’s paper for the Learning Hub. It turns a vague suspicion that “people make mistakes” into questions about particular mechanisms and conditions. Used carefully, it supports humility without contempt for judgment, and a practice of checking without pretending that every hard decision can be made certain.
Research references
- Tversky, A., & Kahneman, D. (1974). Judgment under uncertainty: Heuristics and biases. Science, 185(4157), 1124–1131. View source
- Tversky, A., & Kahneman, D. (1983). Extensional versus intuitive reasoning: The conjunction fallacy in probability judgment. Psychological Review, 90(4), 293–315. View source
- Gigerenzer, G., & Hoffrage, U. (1995). How to improve Bayesian reasoning without instruction: Frequency formats. Psychological Review, 102(4), 684–704. View source
- Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526. View source
- Tversky, A., & Kahneman, D. (1973). Availability: A heuristic for judging frequency and probability. Cognitive Psychology, 5(2), 207–232. View source