Cognitive Bias

A cognitive bias is a systematic tendency in judgment, memory, or decision-making that departs from a relevant standard of accuracy, evidence evaluation, or rationality. Unlike an isolated mistake, a bias describes a recurring pattern under specified conditions. Research in psychology and cognitive science examines these patterns through tasks involving reasoning, uncertainty, information search, and choice. Identifying a bias requires attention both to observed behavior and to the standard against which that behavior is assessed. (annualreviews.org)

Cognitive biases may arise from simplified information-processing strategies, limited cognitive resources, motivational influences, or the way a problem is presented. A heuristic is not itself necessarily a bias: a shortcut can produce accurate judgments in one environment and systematic errors in another. Amos Tversky and Daniel Kahneman’s research on judgment under uncertainty helped establish the heuristics-and-biases research program, which became influential in behavioral economics and other fields concerned with human decisions. (doi.org)

The subject encompasses biases in probability estimation, belief formation, recollection, social judgment, and valuation. Its study also addresses individual differences, experimental measurement, professional decision-making, and interventions intended to improve judgment. Competing explanations emphasize different relationships between cognitive limitations and environmental demands; consequently, the interpretation and practical significance of a demonstrated bias depend on the task, context, and quality of the evidence. (pubmed.ncbi.nlm.nih.gov)

Definition, Scope, and Neighboring Concepts

A cognitive bias is a systematic tendency in judgment, reasoning, recollection, or choice that departs from a specified standard of accuracy or appropriate inference. “Systematic” means that responses display a recurring direction or structure—for example, estimates shifting toward an irrelevant starting value—rather than merely containing occasional mistakes. The benchmark may be a known quantity, a rule of probability, a requirement of logical validity, or a consistency condition in decision theory. A wrong answer alone does not establish bias: an isolated calculation error and random variation around a target differ from a predictable displacement of responses. The classification therefore concerns a pattern under specified conditions, not simply an unfavorable result or disagreement with an observer. (csc2.ncsu.edu)

A heuristic is a processing strategy, whereas a bias is a pattern of outcomes evaluated against a benchmark. Heuristics simplify tasks by reducing information search or replacing a difficult assessment with a more manageable operation. For example, judging frequency by how readily examples come to mind can be useful when ease of retrieval tracks actual frequency, but misleading when publicity or vividness makes uncommon events unusually accessible. Amos Tversky and Daniel Kahneman’s 1974 paper, “Judgment under Uncertainty: Heuristics and Biases,” described such strategies as economical and usually effective while documenting predictable errors. Conversely, Gerd Gigerenzer and Wolfgang Gaissmaier’s 2011 review showed why ignoring information need not reduce accuracy: performance depends on the task and environment. Neither shortcut use nor reduced computational effort is, by itself, evidence of bias. (csc2.ncsu.edu)

A logical fallacy concerns a defect in an argument or argumentative procedure; cognitive bias concerns a tendency in how information is processed or responses are produced. The two can overlap, but they identify different objects of explanation. A fallacious inference can be stated in a text without evidence that its author habitually reasons that way, while an estimate or distorted recollection can exhibit bias without containing an explicit argument. Deliberate deception is also distinct: knowingly reporting a false answer does not demonstrate a mistaken judgment. Likewise, lacking knowledge may explain an inaccurate response without establishing systematic distortion. These distinctions matter experimentally: evidence of bias requires more than finding false statements, and task interpretation and available information can affect whether an apparent departure is genuinely erroneous. (plato.stanford.edu)

In statistics, the bias of an estimator has a narrower mathematical definition:

Here, $\hat{\theta}$ is an estimator of the parameter $\theta$, and the expectation is taken over repeated samples under the specified statistical model. This property can belong to a computational procedure without any human judgment being involved. Cognitive bias is not restricted to numerical estimation or to a nonzero mean error: it also includes systematic differences between responses to differently presented information and directional distortions in remembered events. The shared word bias therefore does not make the psychological and statistical concepts interchangeable. (nist.gov)

The scope extends across uncertain judgment, evaluation of reasons, memory reconstruction, and choice, but not every cognitive limitation qualifies. Daniel Schacter’s 1999 account of seven categories of memory failure distinguished forgetting—transience, absent-mindedness, and blocking—from distortions involving misattribution, suggestibility, and bias. In this narrower memory usage, bias includes recollections being reshaped by present knowledge or beliefs, rather than simply losing information. Correspondingly, limited attention, slow calculation, and difficulty retrieving a name do not by themselves establish a directional distortion. Bias labels describe identifiable response patterns; they need not imply that all affected people share one underlying mechanism or that a person exhibits the same tendency across every setting. (sites.harvard.edu)

Historical Development of Bias Research

Research on systematic errors in judgment developed from several earlier approaches to reasoning and decision-making. Ward Edwards’s 1954 review, “The Theory of Decision Making,” brought economic theories of preference and risky choice into experimental psychology, helping establish a framework for comparing actual choices with formal models. In “A Behavioral Model of Rational Choice” (1955), Herbert A. Simon challenged models that assumed extensive knowledge and computational capacity. His account of bounded rationality examined how people simplify choices and seek satisfactory alternatives rather than exhaustively calculate an optimum. This approach made limitations on information and computation central to the explanation of decision behavior. (tandfonline.com) A parallel line of research examined hypothesis testing: Peter Wason’s 1960 rule-discovery experiment found that only six of 29 participants reached the correct rule without first proposing an incorrect conclusion. The study focused attention on how people select evidence when testing their own hypotheses. (tandfonline.com)

Amos Tversky and Daniel Kahneman’s early-1970s work connected recurring judgment errors to identifiable cognitive strategies. “Belief in the Law of Small Numbers” (1971) examined exaggerated expectations that small random samples would closely represent their populations, including among researchers familiar with statistics. Their subsequent papers included “Subjective Probability: A Judgment of Representativeness” (1972), “On the Psychology of Prediction” (1973), and “Availability: A Heuristic for Judging Frequency and Probability” (1973). These studies investigated how resemblance and the accessibility of examples could influence probability judgments and predictions. The research program thus moved beyond documenting incorrect answers toward explaining why particular errors recur across differently worded tasks. (stats.org.uk)

Their synthesis, “Judgment under Uncertainty: Heuristics and Biases,” appeared in Science on September 27, 1974. It organized evidence around three heuristics: representativeness, availability, and adjustment from an anchor. The authors described these strategies as economical and often effective, while showing that they could produce predictable departures from statistical principles. An important feature was the use of short, reproducible judgment problems rather than a general claim that people reason poorly. For example, participants who encountered arbitrary starting numbers of 10 or 65 subsequently gave median estimates of 25% or 45%, respectively, for the proportion of African countries in the United Nations. Such demonstrations made the proposed influence of a judgment strategy experimentally visible. (doi.org)

The program subsequently extended from judgments of probability to preferences under risk. In March 1979, Kahneman and Tversky published “Prospect Theory: An Analysis of Decision under Risk” in Econometrica. Prospect theory offered a formal descriptive alternative to expected-utility theory, linking experimental choice patterns to an explicit mathematical model rather than merely cataloguing anomalies. (ideas.repec.org) Their 1981 Science article, “The Framing of Decisions and the Psychology of Choice,” demonstrated preference reversals when equivalent outcomes were presented differently. In its disease-response experiment, 72% selected the certain option under a lives-saved description, whereas 78% selected the risky option under an equivalent deaths description. These findings brought the formulation of decision problems into debates about rational choice. (stanford.edu)

The widening research field was represented by the 1982 collection Judgment under Uncertainty: Heuristics and Biases, edited by Kahneman, Paul Slovic, and Tversky. Its coverage included overconfidence, risk perception, and corrective procedures; David M. Eddy’s chapter on probabilistic reasoning in clinical medicine illustrated its application to professional judgment. (cambridge.org) The connection with behavioral economics received institutional recognition on October 9, 2002, when Kahneman was announced as a recipient of the Nobel Memorial Prize in Economic Sciences for integrating psychological research into economics, particularly judgment and decision-making under uncertainty. He shared the award with Vernon L. Smith, recognized separately for experimental economic methods. The announcement explicitly credited Kahneman’s collaboration with Tversky, who had died in 1996. (nobelprize.org)

Standards for Evaluating Judgment and Choice

Research distinguishes three roles for models of judgment and choice. Normative models specify standards against which an inference or preference can be evaluated; descriptive models characterize what people actually do; prescriptive approaches develop procedures intended to help people meet relevant standards under practical constraints. These roles are not interchangeable: a model may predict behavior accurately without establishing that the behavior is appropriate, while a normative principle need not describe the psychological process producing a response. In decision theory, for example, expected-utility maximization can function as an evaluative benchmark without implying that people explicitly calculate utilities and probabilities before choosing. (plato.stanford.edu)

Benchmarks from logic concern whether conclusions follow from premises, rather than whether conclusions sound plausible. Under classical deductive standards, a valid argument cannot have true premises and a false conclusion; the assessment therefore holds the premises fixed and examines the inference. Probability theory supplies corresponding consistency requirements for uncertain judgments: probabilities lie between zero and one, mutually exclusive possibilities have additive probabilities, and an event contained within another cannot be more probable than the containing event. Such constraints evaluate relations among judgments without requiring advance knowledge of which outcome will occur. Eldar Shafir and Robyn A. LeBoeuf’s 2002 review, Rationality, examined departures from normative requirements across reasoning, judgment, and choice, alongside disputes about experimental assumptions and task interpretation. (annualreviews.org)

Bayesian updating provides a benchmark for revising beliefs when evidence becomes available. For a hypothesis $H$ and evidence $E$, with $P(E)>0$, Bayes’ theorem states $P(H\mid E)=P(E\mid H)P(H)/P(E)$. The posterior depends jointly on the prior probability and the likelihood of the evidence, not simply on how strongly the evidence resembles the hypothesis. In an illustrative inspection problem, suppose 1% of components are defective, an alarm detects 90% of defective components, and it also activates for 5% of nondefective components. The probability of a defect given an alarm is then $0.009/(0.009+0.0495)$, approximately 15.4%, rather than 90%. This calculation is a benchmark only if the stated rates apply to the component population and inspection procedure in question; the theorem does not independently establish those assumptions. (stat.berkeley.edu)

Expected-utility theory evaluates choices by combining probabilities with the decision maker’s valuation of consequences. For discrete outcomes, it ranks an action $a$ by $EU(a)=\sum_i P(o_i\mid a)u(o_i)$, where $u(o_i)$ represents the utility of outcome $o_i$. Maximizing expected utility is not necessarily equivalent to maximizing expected money: a preference for a certain payment over a riskier alternative with a larger monetary expectation can be consistent with the theory. A bias assessment must therefore specify both beliefs and objectives, rather than treating the investigator’s preferred outcome as the uniquely rational choice. (plato.stanford.edu)

An unfavorable result does not, by itself, establish that the preceding decision was biased. A well-supported choice can produce a low-probability adverse outcome, just as a poorly supported choice can succeed by chance. Jonathan Baron and John C. Hershey demonstrated the evaluative importance of this distinction in five studies published in 1988: undergraduates judged decisions involving medical scenarios or monetary gambles more favorably when the reported outcomes were favorable, despite receiving the relevant information available when the decisions were made. Their experiments concerned outcome bias in evaluating decision quality, not proof that the original decision makers had reasoned incorrectly. (bear.warrington.ufl.edu)

Classifying a response as biased consequently requires identifying what information was available, what the participant understood the task to ask, which objectives were relevant, and what constraints applied. Everyday wording may invite interpretations different from its formal logical meaning; omitted probabilities may require assumptions rather than calculation; and time or information-search costs may change which procedure is feasible. Bounded rationality makes these resource limits part of the analysis. These considerations do not automatically excuse inconsistent judgments: they determine whether the proposed benchmark applies to the actual task and whether the observed departure reflects reasoning, interpretation, preferences, or constraints. (annualreviews.org)

Cognitive and Motivational Mechanisms

Selective processing can arise because attending to every cue, retrieving all relevant knowledge, and comparing every alternative would exceed available time and computational resources. Within bounded rationality, a heuristic reduces these demands by simplifying particular operations rather than merely producing a quick answer. Anuj Shah and Daniel Oppenheimer’s 2008 effort-reduction framework distinguished five operations that heuristics can simplify: examining cues, retrieving cue values, weighting cues, integrating information, and examining alternatives. A decision rule might therefore consider only a few attributes, assign them equal weights, or stop searching once an acceptable option appears. These shortcuts need not produce error; directional distortion arises when the omitted or simplified information matters to the judgment being made. Selective processing is consequently a proposed mechanism, not itself sufficient evidence of bias. (pages.ucsd.edu)

Dual-process accounts distinguish relatively autonomous processing from processing that requires sustained cognitive control. In their 2013 formulation, Jonathan Evans and Keith Stanovich described Type 1 processes as generating default responses, while Type 2 processing supports hypothetical thinking and depends heavily on working memory—the capacity to maintain and manipulate information during an ongoing task. A person evaluating an argument may initially respond to its familiar or plausible conclusion, then separately check whether that conclusion follows from the premises. The distinction concerns processing properties, not two literal anatomical systems in the brain; Type 1 encompasses multiple autonomous processes. Nor does the distinction establish that automatic answers are necessarily wrong or deliberative answers necessarily correct. Deliberation supplies opportunities for checking, but its results still depend on the knowledge, rules, and goals brought to the task. (psychologicalscience.org)

Goals can influence both which reasons become available and how those reasons are evaluated. Ziva Kunda’s 1990 account of motivated reasoning distinguished accuracy goals from directional goals: seeking a defensible answer differs from seeking support for a preferred conclusion. Directional motivation can favor retrieval of congenial memories, construction of supporting explanations, and more demanding scrutiny of unwelcome evidence. This does not require knowingly falsifying an answer. The reasoner can experience the selected evidence as persuasive, because the selection process itself is uneven. Kunda emphasized that preferred conclusions are constrained by the ability to construct apparently reasonable justifications. Thus, additional effort can be used either to test an initial judgment or to elaborate a case for retaining it; effort alone does not identify the purpose or quality of reasoning. (bear.warrington.ufl.edu)

Affect can also alter judgment through specific appraisals rather than through a single positive-versus-negative mood dimension. In Jennifer Lerner and Dacher Keltner’s 2001 studies, fear was associated with more pessimistic risk estimates and risk-averse choices, whereas anger was associated with more optimistic estimates and risk-seeking choices. The contrasting patterns appeared with both naturally occurring and experimentally induced emotions. Although fear and anger are both negatively valenced, their associated appraisals differ: fear involves uncertainty and reduced control, whereas anger involves greater certainty and perceived control. Appraisals of certainty and control moderated the reported effects, and perceived control mediated them. These findings identify an emotion-specific route to judgment differences; emotional influence alone does not establish departure from a normative benchmark. (pubmed.ncbi.nlm.nih.gov)

Descriptive bias labels therefore cannot substitute for process explanations. An estimate attracted toward a starting value, for example, could reflect stopping an effortful adjustment too early or preferentially retrieving information consistent with that value. Nicholas Epley and Thomas Gilovich’s 2005 anchoring research distinguished these proposed mechanisms and reported different effects of incentives and forewarning for self-generated and externally supplied anchors. Their 2006 experiments found that adjustment from self-generated anchors often terminated when an estimate first entered a plausible range. The same directional pattern can therefore arise through different processing routes, with different responses to changes in motivation or task demands. (onlinelibrary.wiley.com)

Organizing and Classifying Cognitive Biases

A functional classification groups biases by the activity in which judgment departs from a specified benchmark. Probabilistic inference includes base-rate neglect and conjunction errors in estimating probability; information search includes preferentially seeking evidence favorable to an existing hypothesis; memory reconstruction includes distortions of recollection by subsequent information or present beliefs; social judgment includes biases in explaining other people’s conduct or assessing oneself; and preference or choice includes framing effects and sensitivity to sunk costs. These groupings identify the main task affected, not separate mental compartments. Confirmation-related effects, for example, can concern both evidence gathering and evidence interpretation, while memory distortions can alter the information available for later decisions. (frontiersin.org)

Classification by observed effect describes a regularity in responses: estimates move toward an anchor, or preferences change when equivalent options receive different descriptions. Classification by proposed mechanism instead groups effects according to an explanation, such as reliance on a heuristic, selective processing, or motivation to preserve an existing belief. Amos Tversky and Daniel Kahneman’s 1974 account organized numerous judgment errors around representativeness, availability, and anchoring and adjustment. A different framework, proposed by Shaul Oreg and Mahmut Bayazit in 2009, distinguished verification, simplification, and regulation biases according to hypothesized underlying motivations. Such schemes make different commitments: placing two effects in the same functional category does not establish that the same process causes them, whereas a mechanism-based classification requires evidence beyond similarity in outcomes. (psy.ucsd.edu)

There is no single universally accepted taxonomy, and classifications often serve particular research purposes. Evanthia Dimara and colleagues’ task-based taxonomy, published in 2020, organized 154 biases into seven categories to connect bias research with information-visualization activities. By contrast, Vincent Berthet and Vincent de Gardelle’s 2023 Heuristics-and-Biases Inventory deliberately did not impose one theoretical taxonomy on its measures. Terminology also operates at different levels of specificity. Raymond Nickerson’s 1998 review treated confirmation bias as a broad family of ways of seeking or interpreting evidence favorably to existing beliefs, expectations, or hypotheses. A task measuring selective search therefore samples one manifestation of that family rather than necessarily measuring every process covered by the name. Counting labels in a catalog consequently does not establish the number of distinct cognitive mechanisms. (research-portal.uu.nl)

Domain-specific classifications need not coincide with a general catalog of cognitive biases. Daniel Schacter’s 1999 classification of memory errors distinguished seven “sins”: transience, absentmindedness, blocking, misattribution, suggestibility, bias, and persistence. The first three concerned forgetting, the next three distortion, and the last intrusive recollection. Here, bias names a narrower category within memory research, rather than encompassing all seven phenomena. This distinction also illustrates why a classification of cognitive limitations cannot simply be relabeled a classification of biases: failure to retrieve information and directional alteration of remembered information are different outcomes, even when both impair performance. (pubmed.ncbi.nlm.nih.gov)

A further distinction separates an experimental effect across participants from a stable difference between individuals. A comparison showing that one framing condition produces more risky choices than another does not, by itself, identify which participants are consistently susceptible to framing. Berthet’s 2021 study illustrates the separation: in its first sample of 157 participants, anchoring showed a standardized effect size of $d=1.05$, while the corresponding score’s split-half reliability was $0.68$. Effect magnitude and person-level measurement reliability answer different questions. Establishing an enduring individual tendency additionally requires consistency across items, stability over time, and evidence that the measure captures the intended construct. Berthet and de Gardelle’s 2023 review identified 41 biases measured across 108 studies but still found gaps in reliable measurement. A catalog of named effects is therefore not a validated diagnostic checklist or sufficient evidence for assigning a general “bias-prone” personality to an individual. (frontiersin.org)

Biases in Probability and Statistical Judgment

The representativeness heuristic substitutes resemblance for a judgment of probability: a person who resembles a stereotypical engineer may be judged likely to be an engineer without sufficient consideration of how common engineers are in the relevant population. Daniel Kahneman and Amos Tversky’s research on prediction distinguished this matching operation from inference that incorporates prior probabilities and the diagnostic value of evidence. Similarity can be informative, but it is not itself a probability. Under Bayes’ theorem, category membership depends on both the category’s base rate and the likelihood of observing the description within that category; base-rate neglect occurs when the former receives insufficient weight. (scholarsbank.uoregon.edu)

The conjunction fallacy illustrates how a more representative description can nevertheless identify a less probable event. In Tversky and Kahneman’s 1983 Linda problem, Linda was described as a philosophy graduate concerned with discrimination and social justice. Participants ranked possible descriptions, including being a bank teller and being a bank teller active in the feminist movement. In the probability-ranking group discussed in Kahneman’s 2003 account, 89% placed the conjunction above the bank-teller statement. Yet a conjunction cannot be more probable than either constituent event: $P(A\cap B)\leq\min\{P(A),P(B)\}$. Every feminist bank teller is necessarily a bank teller; adding the second condition narrows, rather than enlarges, the set of qualifying cases. This constraint does not require the events to be independent. (bear.warrington.ufl.edu)

Insensitivity to sample size concerns the variability of samples, not simply their average outcome. In the hospital problem reported in 1974, participants compared hospitals with approximately 45 and 15 births daily and identified which would record more days with over 60% boys. Of 95 respondents, 53 selected approximately equal frequencies, whereas 21 selected each hospital. Under the task’s sampling assumptions, the smaller hospital should experience more extreme proportions. Tversky and Kahneman’s 1971 expression “belief in the law of small numbers” described the mistaken expectation that even small samples closely reproduce population characteristics. Unlike the law of large numbers, this is not a mathematical guarantee: it names an overstatement of small samples’ representativeness. (psy.ucsd.edu)

Misunderstanding regression toward the mean produces another statistical error. When successive measurements are imperfectly correlated, cases selected for extreme initial scores tend to receive less extreme subsequent scores. Tversky and Kahneman described flight instructors who interpreted improvement after criticism and deterioration after praise as evidence that criticism worked better. Because criticism followed unusually poor performances and praise followed unusually good ones, regression could generate that pattern without either intervention causing the change. Regression does not mean that every extreme result must reverse; it concerns the expected pattern across selected cases. (psy.ucsd.edu)

Expectations of representativeness also affect judgments about random sequences. The gambler’s fallacy treats a recent imbalance as creating pressure for an opposite outcome, as though chance corrected itself over short runs. For a known fair coin with independent tosses, however, the probability of heads on the next toss remains $1/2$, whatever the preceding sequence. Long-run convergence does not require compensating outcomes. These conclusions depend on the generating process: if the coin’s fairness is unknown, observed outcomes can inform beliefs about its bias; if events are dependent, previous outcomes may affect subsequent probabilities. Mistaken sequence judgments therefore must be distinguished from legitimate inference about an uncertain or changing mechanism. (stats.org.uk)

Availability and Anchoring in Estimation

The availability heuristic estimates frequency or likelihood from the ease with which relevant examples can be recalled or scenarios imagined. Amos Tversky and Daniel Kahneman’s 1973 account distinguished this accessibility from actual occurrence rates: frequent events often supply readily retrievable examples, but retrieval also depends on how information is organized. Their word-frequency experiments illustrated the difference. Participants tended to judge words beginning with a specified letter as more frequent than words containing it in the third position, although the reverse held for the letters tested. Searching memory by initial letter was easier than searching by an internal position. (noorsiddiqui.com)

Vividness, recency, and familiarity can change accessibility without changing prevalence. In a demonstration described by Tversky and Kahneman in 1974, participants heard lists of men’s and women’s names and overestimated the numerical representation of the sex whose names were more famous. The authors also identified recent occurrences and salient experiences—such as witnessing a fire rather than reading about one—as influences on availability. These factors alter which examples come readily to mind; they do not themselves establish how common the underlying events are. (psy.ucsd.edu)

Availability can involve the subjective experience of retrieval, not merely the number or content of remembered examples. In 1991, Norbert Schwarz and colleagues asked participants to recall either six or twelve instances of their own assertive or unassertive behavior. Recalling twelve was more difficult. Participants recalling twelve assertive instances subsequently rated themselves as less assertive than those recalling six, despite having produced more supporting examples. The corresponding pattern reversed for unassertive behavior. Further experiments examined whether retrieval difficulty was regarded as informative, showing that its influence depended on how participants interpreted that experience. (roar-assets-auto.rbl.ms)

The anchoring effect occurs when numerical estimates shift toward an initial value. In the classic demonstration reported in 1974, a wheel of fortune supplied a number between 0 and 100. Participants first judged whether the percentage of African countries among United Nations members was above or below that number, then estimated the percentage. Groups receiving starting values of 10 and 65 produced median estimates of 25% and 45%, respectively. The random number contained no information about United Nations membership, yet estimates moved toward it. Tversky and Kahneman explained anchoring in terms of adjustment from a starting value that stops short of sufficient correction. (psy.ucsd.edu)

Later research distinguished that explanation from other mechanisms. Nicholas Epley and Thomas Gilovich’s 2006 experiments found that adjustment from self-generated anchors tended to terminate once an estimate entered a plausible range, unless participants were able and willing to search further. Externally supplied anchors need not operate identically. Fritz Strack and Thomas Mussweiler’s 1997 selective-accessibility account proposed that answering a comparison question makes anchor-consistent knowledge more accessible, which then influences the absolute estimate. Thus, a shift toward an anchor does not by itself demonstrate a sequence of conscious numerical adjustments. (pubmed.ncbi.nlm.nih.gov)

Neither accessibility nor reliance on a starting value necessarily constitutes error. Remembered examples can carry useful information when exposure reflects the relevant event distribution; ease of recall becomes misleading when unrelated retrieval advantages substitute for frequency evidence. Likewise, a relevant starting estimate may legitimately constrain a judgment. Assessing anchoring therefore requires distinguishing the information an anchor supplies from its influence merely as a presented number, while assessing availability requires separating the evidential content of recalled instances from how easily they were retrieved. (sciencedirect.com)

Confirmation, Belief Maintenance, and Reasoning

Confirmation bias can occur at several stages of reasoning. Selective search concerns which information is obtained: a person may seek cases, questions, or sources expected to support an existing hypothesis, leaving alternatives unexamined. Asymmetric interpretation concerns what evidence is taken to mean: an ambiguous observation may be understood as consistent with a favored explanation. Selective weighting concerns its influence on judgment: supporting findings may be accepted with little scrutiny, while opposing findings face stricter demands concerning methods, credibility, or relevance. These processes can coexist, but they are not interchangeable; exposure to the same information does not ensure equivalent evaluation. Raymond Nickerson’s 1998 review distinguished these manifestations without treating them as evidence of a single mechanism or necessarily deliberate distortion. (psy.ucsd.edu)

Peter Wason’s 1960 rule-discovery experiment examined information search in a relatively simple setting. Twenty-nine participants received the triple $2,4,6$, which satisfied an undisclosed rule, and proposed additional triples, receiving feedback about whether each qualified. The actual rule was any three numbers in strictly increasing order. A narrower hypothesis, such as “consecutive even numbers,” also predicts acceptance of $4,6,8$, so that feedback cannot distinguish it from the broader rule. By contrast, acceptance of $1,2,8$ contradicts the narrower hypothesis while remaining compatible with the actual rule. Only six participants identified the rule without first announcing an incorrect one; 13 made one incorrect announcement, nine made two or more, and one announced no conclusion. These figures concern the sequence of announced rules, not a claim that only six participants ever solved the task. (bear.warrington.ufl.edu)

A positive testing strategy—examining cases expected to satisfy a hypothesis—is not itself equivalent to confirmation bias. Joshua Klayman and Young-Won Ha emphasized this distinction in 1987: a positive test can produce disconfirming feedback, and its informativeness depends on the relationship between the hypothesized category and the actual category. In Wason’s task, a narrow guessed rule is contained within the target rule, making positive instances especially poor discriminators. In other environments, testing a predicted positive case can efficiently reveal that an expected property is absent. The strategy describes how tests are selected; confirmation bias describes an unwarranted tendency toward supporting a belief. Inferring the latter requires examining alternative hypotheses, possible outcomes, and the information supplied by the tests. (bear.warrington.ufl.edu)

Biased assimilation occurs when prior beliefs shape the evaluation of newly presented evidence. In a 1979 experiment, Charles Lord, Lee Ross, and Mark Lepper presented 48 undergraduates who supported or opposed capital punishment with two fictitious studies reporting opposite conclusions about its deterrent effect. The association between research method and reported outcome was counterbalanced. Both groups judged the study supporting their existing position as more convincing and better conducted. Participants reported changes indicating greater polarization after considering the mixed evidence. The experiment thus concerned differential evaluation of controlled research descriptions, not factual evidence about the policy’s effectiveness or a demonstration that every encounter with conflicting information increases disagreement. (nuovoeutile.it)

Belief perseverance is the persistence of a judgment after its original evidential basis has been discredited. In 1975, Ross, Lepper, and Michael Hubbard gave participants predetermined success or failure feedback on a task distinguishing authentic from fabricated suicide notes. Debriefing explicitly disclosed that the feedback was unrelated to actual performance, yet subsequent assessments of ability and expected future performance retained effects of the initial feedback. A second experiment found persistence in observers’ judgments as well. The authors proposed that initial impressions generate causal explanations and recruit apparently consistent information, allowing those explanations to survive the withdrawal of the initiating evidence. This is stronger than merely resisting one contrary observation: the experimental manipulation removed the relevance of the original evidence itself. (researchgate.net)

Belief bias concerns evaluating arguments rather than selecting evidence. In deductive reasoning, validity requires that the conclusion follow necessarily if the premises are true; the conclusion’s everyday plausibility is a separate matter. Jonathan Evans, Julie Barston, and Paul Pollard’s 1983 experiments crossed validity with conclusion believability and found effects of both, with belief exerting a larger influence on invalid arguments. For example, “All dogs are mammals; all cats are mammals; therefore, all cats are dogs” is invalid, but substituting a believable conclusion does not repair that inferential form. Belief-bias tasks test whether judgments track the premises’ logical relationship to the conclusion, rather than agreement with the conclusion considered independently. (core.ac.uk)

Memory Reconstruction and Retrospective Judgment

Reconstructive memory combines retained information about an experience with knowledge, expectations, and cues available at retrieval; it is not simply the playback of a fixed record. Present beliefs can therefore influence what is remembered about earlier experiences, including previous attitudes and emotional reactions. In his 1999 classification of memory errors, Daniel L. Schacter distinguished forgetting—such as declining accessibility over time—from misattribution, suggestibility, and belief-related distortion. Reconstruction does not itself establish bias: the relevant pattern is a systematic shift toward current beliefs, suggested details, or known outcomes, rather than missing information alone. (sites.harvard.edu)

Current appraisals can shape recollections even when earlier responses were recorded. In Linda J. Levine’s 1997 study, 227 supporters of Ross Perot reported their emotions and interpretations following his withdrawal from the United States presidential race in July 1992. After the November election, 147 participants recalled their initial reactions and reported their current appraisals. Errors in remembered emotion tended toward consistency with those later appraisals, rather than showing a general exaggeration of emotional intensity. The comparison with contemporaneous reports demonstrated directional reconstruction: later recollections partly reflected how participants now interpreted the event, not only how they had experienced it initially. (researchgate.net)

Source misattribution occurs when remembered content is assigned to the wrong origin—for example, when information encountered in a later account is remembered as something personally witnessed. Familiarity with a detail and recollection of its source can come apart. Suggestibility concerns the influence of leading questions or other post-event information on recollection; it can overlap with misattribution when a suggested detail is subsequently attributed to the original experience. These errors differ from deliberate fabrication because the remembered content may be accepted sincerely as part of the event. Schacter treated misattribution and suggestibility as distinguishable forms of distortion, rather than interchangeable labels for all inaccurate memories. (sites.harvard.edu)

Elizabeth F. Loftus and John C. Palmer demonstrated post-event influence in two 1974 experiments using films of automobile accidents. In the first, involving 45 students, mean speed estimates were 40.8 miles per hour when the question described the cars as having “smashed” into each other, compared with 34.0 for “hit” and 31.8 for “contacted.” In the second, 150 students were divided into three groups of 50. One week later, nonexistent broken glass was reported by 16 participants previously given the “smashed” question, seven given “hit,” and six who had received no speed question. The delayed false-detail reports supported a reconstructive interpretation beyond an immediate shift in numerical estimates. Nevertheless, the authors distinguished altered memory from response bias: wording can influence an answer without necessarily demonstrating that the original memory has been erased or replaced. (gwern.net)

Hindsight bias is an outcome-related shift in remembered predictions or perceived prior predictability. In a 1975 study, Baruch Fischhoff and Ruth Beyth asked participants to recall probability judgments made before President Richard Nixon’s visits to Peking and Moscow. Recalled probabilities were generally higher for events participants believed had occurred and lower for events believed not to have occurred. A separate 1975 study by Fischhoff showed that learning a reported historical outcome increased its retrospectively assessed likelihood and changed which background details seemed relevant. Participants were largely unaware of these influences. Thus, hindsight can involve either misremembering an actual forecast or reconstructing what could supposedly have been predicted without outcome knowledge. (sciencedirect.com)

Confidence and vividness do not independently establish recollection accuracy. Jennifer M. Talarico and David C. Rubin’s 2003 study recorded 54 Duke University students’ memories on September 12, 2001, both of hearing about the previous day’s attacks and of an everyday event. Participants were retested after one, six, or 32 weeks. Consistency declined similarly for both kinds of memory, but confidence and vividness declined only for everyday memories. These comparisons measured agreement with initial reports, not independently verified accuracy of every detail. They nevertheless showed that strong subjective certainty could persist while recollections changed—a different phenomenon from merely forgetting details. (psychologicalscience.org)

Social Judgment and Self-Assessment

Correspondence bias occurs when observers infer enduring attitudes or personality traits from behavior without adequately accounting for situational constraints. The closely related fundamental attribution error emphasizes overweighting personal dispositions and underweighting circumstances in explanations of others’ behavior. In Edward E. Jones and Victor A. Harris’s 1967 experiments, participants judged a speaker’s or writer’s attitudes after encountering statements supporting or opposing Fidel Castro. Even when the position had been assigned rather than freely chosen, judgments shifted toward the expressed position. The critical comparison was not whether participants noticed the assignment, but whether they sufficiently discounted its influence when inferring the author’s attitude. (pubmed.ncbi.nlm.nih.gov)

Such judgments need not arise from a single mechanism. Daniel T. Gilbert and Patrick S. Malone’s 1995 review distinguished four routes to correspondence bias: failing to recognize situational constraints, holding unrealistic expectations about behavior under those constraints, exaggerating the behavior’s apparent dispositional significance, and incompletely correcting an initial dispositional inference. An observer may therefore know that a person faced pressure yet underestimate how strongly that pressure could shape behavior. In the incomplete-correction account, identifying a behavior and drawing a trait inference can precede a more effortful adjustment for circumstances; knowing about the situation does not guarantee that the adjustment will be sufficient. (pubmed.ncbi.nlm.nih.gov)

Self-serving attribution concerns explanations of one’s own outcomes rather than primarily the interpretation of another person’s character. Its characteristic pattern is greater attribution of positive than negative outcomes to internal causes, sometimes also treating positive outcomes as reflecting more stable and broadly applicable qualities. A 2004 meta-analysis by Amy H. Mezulis and colleagues combined 266 studies yielding 503 independent effect sizes and reported an average standardized effect of $d = 0.96$. This outcome asymmetry does not by itself identify a motive: Dale T. Miller and Michael Ross’s 1975 review proposed that expectations of success and perceived relationships between effort and outcomes could also generate apparently self-enhancing explanations. Evidence for a directional pattern must therefore be distinguished from evidence that protecting self-esteem caused it. (pubmed.ncbi.nlm.nih.gov)

Social projection instead uses one’s own responses as a basis for estimating other people’s responses. In four studies published in 1977, Lee Ross, David Greene, and Pamela House found that participants perceived their own choices as relatively common and alternative choices as relatively uncommon. They also regarded responses unlike their own as more revealing of the respondent’s distinctive personality. This “false consensus effect” links population estimates with attribution: a choice judged unusual becomes stronger apparent evidence of a personal disposition. Unlike self-serving attribution, the central comparison concerns inferred agreement between self and others, not explanations of personal success or failure. (sciencedirect.com)

The bias blind spot is an asymmetry in perceived susceptibility: people recognize bias more readily in others than in themselves. Emily Pronin, Daniel Y. Lin, and Lee Ross documented this pattern in 2002. Research by Pronin and Matthew B. Kugler, published in 2007, implicated unequal weighting of introspection and observable behavior. Participants gave their own conscious thoughts and intentions greater weight when assessing personal bias, but relied more on behavior when assessing others. Providing access to another person’s introspections did not eliminate the asymmetry. The absence of a consciously experienced biased intention can consequently be treated as evidence of impartiality even when the relevant influence is not accessible to introspection. (philosophy.rutgers.edu)

Research on implicit social attitudes overlaps with these concerns but addresses a different construct. Anthony G. Greenwald and Mahzarin R. Banaji’s 1995 account examined how past experience can influence attitudes, self-esteem, and stereotypes without the person recognizing that influence introspectively, motivating the use of indirect measures. Cognitive-bias studies typically examine directional departures from a specified judgment benchmark; implicit-attitude studies examine evaluations and associations that may not be captured by explicit reports. “Implicit” therefore identifies a mode of influence or measurement, not automatically an erroneous inference. The overlap lies in social judgments whose determinants may be automatic or inaccessible, rather than in treating every implicit attitude as a correspondence error or every attributional bias as prejudice. (pubmed.ncbi.nlm.nih.gov)

Framing, Valuation, and Economic Choice

A framing effect occurs when alternative descriptions of equivalent outcomes change preferences. In Amos Tversky and Daniel Kahneman’s 1981 disease-response experiment, participants considered an outbreak expected to kill 600 people. In the gain-worded version, 72% of 152 respondents preferred saving 200 people for certain to a gamble offering a one-third probability of saving all 600 and a two-thirds probability of saving nobody. In the loss-worded version, only 22% of 155 respondents preferred 400 certain deaths to a gamble offering a one-third probability of no deaths and a two-thirds probability of 600 deaths. Under the intended interpretation, the certain options describe the same outcome, as do the gambles. The reversal therefore concerns descriptive framing, not differences in consequences or expected value. These were choices by separate groups, rather than demonstrated reversals within each individual. (stanford.edu)

Kahneman and Tversky’s prospect theory, introduced in 1979, models choice descriptively through reference-dependent valuation and decision weights. Outcomes are evaluated as gains or losses relative to a reference point, rather than solely as final wealth. Its proposed value function is generally concave for gains and convex for losses: successive increments have diminishing psychological impact as distance from the reference point increases. This can favor certainty for gains and risk-taking for losses, although probability weighting also affects the prediction. Decision weights represent an event’s influence on a prospect’s attractiveness; they are not simply reported probabilities and need not obey probability axioms. Overweighting small probabilities can make both low-probability prizes and protection against unlikely losses attractive. The model consequently does not predict universal risk aversion for gains or universal risk-seeking for losses. (web.mit.edu)

Loss aversion is the additional proposition that a loss has greater impact than a corresponding gain around the reference point. It differs from diminishing sensitivity: the latter concerns curvature within the gain or loss domain, whereas loss aversion concerns asymmetry across the boundary. Endowment-related valuation provides a possible application. In Daniel Kahneman, Jack Knetsch, and Richard Thaler’s 1990 experiments, randomly allocating coffee mugs to participants produced substantially less trading than predicted by an ownership-independent valuation model; markets for tokens with experimentally assigned monetary values behaved differently. Interpreting ownership as a reference point makes relinquishing the mug a loss, while acquiring it is a gain. Nevertheless, an observed gap between willingness to accept and willingness to pay does not, by itself, identify loss aversion as its cause. (web.mit.edu)

Valuation procedures can also influence such gaps. Charles Plott and Kathryn Zeiler’s 2005 experiments found no willingness-to-accept–willingness-to-pay gap under procedures combining training, practice, and controls intended to reduce misunderstanding. A 2015 replication study, however, found that the mug valuation gap persisted, including among participants whose practice choices showed no apparent misconceptions. These contrasting results distinguish an observed valuation pattern from a uniquely established mechanism: ownership, understanding of the elicitation procedure, and experimental conditions must be considered separately. (swlb2.aeaweb.org)

Status-quo preferences likewise need not have a single explanation. William Samuelson and Richard Zeckhauser’s 1988 experiments compared choices with and without a designated existing position and found disproportionate retention of that position. Their analysis distinguished psychological mechanisms, including loss aversion and commitment to earlier choices, from potentially rational considerations such as switching costs and uncertainty about alternatives. Maintaining an existing arrangement is therefore not sufficient evidence of bias: the relevant comparison concerns whether its designation as the status quo changes preference beyond what consequential differences explain. (web.mit.edu)

A sunk-cost effect occurs when irrecoverable expenditure influences continuation decisions. Hal Arkes and Catherine Blumer’s 1985 theater study found that customers paying more for season subscriptions attended more performances during the following six months; their experiments implicated reluctance to appear wasteful. Escalation of commitment concerns further resource allocation despite unfavorable feedback. In Barry Staw’s 1976 simulated investment study involving 240 business-school students, commitment was greatest when participants were personally responsible for negative consequences. These effects overlap but are not identical: prior responsibility can matter independently of expenditure. The economic distinction is between unrecoverable past spending and future consequences. Evidence learned from earlier investment, remaining completion costs, recoverable assets, and opportunity costs can legitimately affect continuation; spending that cannot change under either option is not itself an additional benefit of continuing. (sciencedirect.com)

Experimental Methods and Measurement

Controlled comparisons isolate changes in judgment attributable to a specified manipulation. In a between-participants experimental design, random assignment places participants in different task versions—for example, equivalent choices expressed as gains or losses—while other information remains constant. Within-participant designs instead compare the same person’s responses across versions. They provide person-level contrasts but introduce possible memory, order, and demand effects: recognizing that two questions are equivalent may change subsequent answers. Separating matched questions and counterbalancing presentation order can address these problems, although converting a demonstration into an individual-differences instrument changes its measurement properties. Probability-estimation tasks elicit numerical judgments or confidence ratings; choice experiments record selections or graded preferences between alternatives. Each requires explicit task instructions and a specified scoring rule. (apa.org)

Measurement depends on whether the target is directional error, inconsistency, or correspondence with actual performance. An overconfidence measure, for example, can subtract percentage accuracy from mean percentage confidence across questions. A framing measure can compare preference ratings under matched loss and gain descriptions. Taking the absolute difference measures inconsistency regardless of direction; retaining the signed difference tests the direction predicted by the account under investigation. These scores are not interchangeable. In Vincent Berthet’s 2021 study, changing task content, item numbers, and scoring improved several instruments: the revised framing measure achieved internal consistency of $\alpha = 0.74$, while revised overconfidence and outcome-bias measures reached $0.81$ and $0.85$, respectively. Such coefficients describe particular scores obtained under particular procedures, not permanent properties of a bias label. (frontiersin.org)

Process identification requires evidence beyond the final answer. Response times constrain explanations of how choices develop, especially when analyzed jointly with accuracy and information acquisition. Eye tracking and mouse-based information displays record which attributes participants inspect and in what sequence. In the Mouselab-MDP planning paradigm, participants click to reveal rewards at possible destinations before choosing a route; the resulting record contains the sequence of inspected locations, revealed information, and selected plan. Repeated trials reveal changes in search and strategy use rather than only average performance. Computational cognitive modeling connects these observations to explicit candidate mechanisms. The associated “computational microscope,” for example, compares observed search sequences with predictions from 79 planning strategies. Its classifications remain conditional on the candidate strategies and assumptions linking clicks to planning operations; observable behavior does not uniquely identify a mental process. (sciencedirect.com)

Multi-item instruments distinguish recurring tendencies from responses to a single scenario. Internal consistency assesses how coherently items contribute to a score within an administration; test–retest reliability assesses consistency across administrations, with the interval and intervening experiences specified. Neither alone establishes that a measure captures its intended construct. Berthet’s initial eight-bias battery illustrates the importance of checking each instrument: among 157 participants, framing scores had Cronbach’s $\alpha = 0.15$, whereas base-rate-neglect scores reached $0.70$. Adding heterogeneous bias questions into one total score therefore cannot substitute for examining the reliability and interpretation of its components. (frontiersin.org)

A reproducible group effect need not provide reliable rankings of individuals. Craig Hedge, Georgina Powell, and Petroc Sumner demonstrated this distinction in research published in the 2018 volume of Behavior Research Methods. Across three studies of seven classic cognitive tasks, test–retest coefficients ranged from $0$ to $0.82$. Although these were principally attention and cognitive-control tasks rather than a cognitive-bias inventory, they illustrate a general measurement problem: an experimental contrast can be consistently positive while varying little between people. When stable between-person variation is small relative to measurement error, individual rankings become unstable. Robustness of an average contrast consequently does not establish suitability for correlational research or individual assessment. (pubmed.ncbi.nlm.nih.gov)

Reporting must make both the manipulation and the measurement reconstructable. The American Psychological Association’s 2018 quantitative reporting standards specify participant characteristics, recruitment, sample-size determination, assignment procedures, materials, exclusions, and psychometric information. For bias experiments, exact task wording, response options, presentation order, feedback, payment, and performance-contingent incentives clarify what participants actually encountered. Results require effect sizes and confidence intervals, alongside descriptive statistics and the analysis used—not merely a significance threshold. Reports must also distinguish planned from exploratory analyses and consider alternative explanations, including task comprehension, construct overlap, and response-format effects, rather than treating every observed contrast as uniquely diagnostic of the proposed bias mechanism. (apa.org)

Individual Differences and Contextual Variation

Cognitive ability moderates some biases but does not provide general immunity. In seven studies published in 2008, Keith Stanovich and Richard West found that several tested effects—including anchoring, framing, and outcome bias—were unrelated to cognitive ability, whereas higher ability was associated with less denominator neglect, belief bias, and probability matching rather than maximizing. These findings concern particular tasks and samples, not an assertion that intelligence never matters. In a separate 2008 investigation involving more than 1,200 participants, they found no evidence that higher cognitive ability reduced myside bias or one-sided thinking: the capacity to reason effectively was distinguishable from the tendency to evaluate positions impartially. (pubmed.ncbi.nlm.nih.gov)

Numeracy and reflective thinking capture more specific differences. Ellen Peters and colleagues’ four studies in 2006 associated higher numeracy with greater use of appropriate numerical principles and reduced susceptibility to numerical framing; the relationship was not explained by general intelligence. More numerate participants also extracted more precise affective meaning from numerical comparisons, a tendency that was not invariably beneficial. Shane Frederick’s 2005 Cognitive Reflection Test assessed whether respondents resisted an immediately plausible answer. Its bat-and-ball problem states that two items cost 1.10 dollars together and that the bat costs one dollar more than the ball: the tempting answer is 10 cents, but the correct answer is 5 cents. Performance therefore samples both relevant calculation and the ability or disposition to reconsider an initial response, rather than measuring universal resistance to bias. (pubmed.ncbi.nlm.nih.gov)

Expertise depends on learning conditions as well as accumulated experience. Daniel Kahneman and Gary Klein’s 2009 analysis identified two central requirements for skilled intuition: an environment containing sufficiently predictable relationships and adequate opportunities to learn those relationships through feedback. Chess provides recurring configurations governed by stable rules; recognizing such configurations can support rapid, accurate judgments. Familiarity with a domain does not by itself establish accurate probability estimates, particularly when outcomes are difficult to predict or feedback is delayed, ambiguous, or selective. Their analysis explicitly distinguished subjective confidence from demonstrated judgment accuracy: a compelling impression of knowing is not evidence that the relevant regularities have been learned. (bear.warrington.ufl.edu)

Developmental comparisons also depend on the response category and task instructions. Paul Klaczynski’s 2001 study included 28 early adolescents, 31 middle adolescents, and 31 young adults, with mean ages of approximately 12.8, 16.8, and 21.7 years. Although normatively correct responses generally increased with age, the characteristic fallacies tested did not consistently decrease; some improvement reflected fewer alternative errors. Asking participants to adopt a logical perspective reduced denominator neglect from roughly 53% to 33% of responses across age groups. Thus, developmental change in correct responding and change in a specific directional bias were not interchangeable measures. The selected school and university samples also limited population-wide interpretation. (gwern.net)

Cross-cultural findings likewise require matched task conditions rather than rankings of demographic groups. Incheol Choi and Richard Nisbett’s 1998 experiments found correspondence bias in both Korean and American participants when they inferred an essay writer’s attitude despite knowing that the position had been assigned. When situational constraints became more salient—for example, through participants’ own experience of writing an assigned position—the bias decreased in the Korean sample but not in the American sample. The observed difference therefore depended on how contextual information was presented. Comparable wording, constraints, and response measures are necessary to distinguish differences in attention to context from differences in task interpretation. (journals.sagepub.com)

Information format and incentives can alter performance without changing a person’s underlying abilities. Gerd Gigerenzer and Ulrich Hoffrage’s 1995 experiments found Bayesian reasoning in 46% of responses under a standard natural-frequency format, compared with 16% under the corresponding probability format; frequency representations reduced computational demands. Colin Camerer and Robin Hogarth’s 1999 review of 74 experiments found that performance-based financial incentives most commonly left mean performance unchanged, although higher payments usually reduced variability. Incentive effects interacted with task demands and available knowledge. Such results identify person–task interactions, not a general “bias-prone” personality that produces the same errors across representations, rewards, and learning environments. (pages.ucsd.edu)

Professional Decisions and Institutional Processes

In clinical diagnosis, premature closure occurs when an explanation is accepted before alternatives have been adequately evaluated, ending the search for evidence too soon. Subsequent findings may then be interpreted within the initial explanation rather than prompting reconsideration. A 2024 Agency for Healthcare Research and Quality (AHRQ) primer illustrates this pattern with a reported case in which repeated blood cultures containing Corynebacterium were dismissed as contamination before endocarditis was diagnosed. Such cases illustrate reasoning failures, but diagnostic error also involves health-system conditions: interruptions, workload, documentation demands, poorly designed electronic records, and failures to communicate or follow up test results. AHRQ’s diagnostic-safety review therefore treats diagnosis as a sociotechnical process involving clinicians, patients, teams, and information systems, rather than solely as an individual mental operation. These conditions affect which evidence becomes available and how much attention it receives; a missed diagnosis does not, by itself, establish a particular cognitive bias. (psnet.ahrq.gov)

In forensic science, contextual information can influence the interpretation of physical evidence without changing that evidence. Information about a suspect’s identity, an alleged confession, or the surrounding case narrative can establish expectations that are irrelevant to a narrowly defined comparison task. In a 2006 study, Itiel Dror, David Charlton, and Alisa Péron returned previously examined fingerprint pairs to the same experts, accompanying them with misleading context suggesting that the prints did not match. Most examiners contradicted their earlier identification decisions. The small study demonstrated sensitivity to context under its experimental conditions, not a population-wide error rate. A separate institutional concern is exposure to another examiner’s conclusion: verification conducted with that conclusion already visible is not fully independent, because the expected answer can influence interpretation. The 2020 NIST report on handwriting examination distinguishes task-relevant information from potentially biasing context and discusses information management and blind review as procedural responses. (eprints.soton.ac.uk)

Experimental research in law examines how preferred outcomes can influence the interpretation of legal criteria. In Avani Mehta Sood’s 2015 study, Cognitive Cleansing, lay participants acting as judges assessed illegally obtained evidence under an “inevitable discovery” exception described in the task. When an otherwise identical illegal search uncovered evidence of selling heroin to high-school students, participants were more inclined to construe lawful discovery as inevitable, permitting admission of the evidence. When it uncovered evidence of unlawfully selling marijuana to cancer patients, participants were more than three times as likely to exclude it. The findings linked punishment preferences to judgments about a procedural exception, rather than simply to open rejection of the stated rule. They concern hypothetical decisions by research participants and do not establish how any particular judge ruled, or why, in an actual case. (papers.ssrn.com)

Organizational discussion can amplify selective information use through shared-information bias. Garold Stasser and William Titus’s 1985 experiment assigned four-person groups information about candidates for student-body president. Across conditions, the complete information favored Candidate A; what differed was whether favorable information was available to everyone or distributed among members. Candidate A was selected by 83% of the 18 groups with fully shared information, but only 18% of the 38 groups in the unshared-information conditions. Discussion disproportionately sampled information already shared and information supporting existing preferences. Consequently, collective access to sufficient evidence did not guarantee that the evidence would enter deliberation or overturn an initial majority. (uni-muenster.de)

Incentives and group procedures also shape information processing. Carsten de Dreu, Bernard Nijstad, and Daan van Knippenberg’s 2008 model distinguishes motivation concerning whose outcomes to advance from motivation to develop an accurate, thorough understanding; these affect the information members seek, retain, and disclose. Research on collective rather than individual performance incentives describes one route by which cooperative motives can be encouraged. In a 2007 experiment, Lotte Scholten and colleagues manipulated process accountability—requiring groups to justify how they reached their decisions. Accountable groups repeated unshared information more often and selected the correct alternative more frequently. This result identifies a task-specific mechanism through which institutional procedures can constrain preference-driven discussion, rather than demonstrating that accountability or stronger incentives invariably eliminate bias. (repub.eur.nl)

Digital Information and Human–AI Interaction

Believing a claim and choosing to share it are distinct judgments. In a 2021 study, Gordon Pennycook, Ziv Epstein, Mohsen Mosleh, Antonio Arechar, Dean Eckles, and David Rand found that headline truthfulness affected accuracy judgments much more strongly than sharing intentions, although most participants reported valuing accurate sharing. Their experiments support an attentional explanation: users may possess relevant knowledge and prefer accuracy, yet fail to bring that preference to bear when deciding what to post on social media. Sharing a false headline therefore does not, by itself, establish that the person believes it; nor does the observed dissociation establish that every misleading post results from inattention rather than mistaken belief or deliberate communication. (nature.com)

The researchers tested this explanation through four survey experiments that redirected attention toward accuracy and a field experiment involving 5,379 Twitter users who had previously shared links to websites associated with misleading content. In the field experiment, users received a private message asking them to evaluate one non-political headline’s accuracy. The researchers compared the quality of news sources shared during the following 24 hours with those shared by users who had not yet received the message. The prompt improved subsequent source quality without supplying corrections to the political stories users were sharing. This result concerns attention at the moment of selection, rather than necessarily a change in underlying beliefs. Its scope is also specific: the field outcome measured the quality of linked news sources over a short interval, not the truth of every shared statement or a durable transformation of users’ convictions. (moodle2.units.it)

Interface position can likewise alter apparent preferences without changing an answer’s content. Keith Burghardt, Tad Hogg, and Kristina Lerman’s 2018 question-answering experiment asked Amazon Mechanical Turk workers to select the most correct answers to ten questions drawn from Stack Exchange’s English Language Learners forum. Participants saw either two or eight answers in randomized order; another manipulation displayed artificial popularity scores. Across conditions, earlier-listed answers were favored, with attention, perceived popularity, and the number of alternatives contributing to position effects. Because randomizing order separated placement from answer identity, the results showed why selection counts cannot simply be treated as independent measures of intrinsic quality. In information retrieval, an apparent preference for one answer may partly reflect the opportunity to encounter and inspect it, together with social cues attached to its presentation. (ojs.aaai.org)

Human–AI collaboration can create a different interaction: an erroneous system output may reinforce an erroneous initial human judgment. In a study reported in a 2024 preprint, Emely Rosbach and colleagues recruited 28 trained pathology experts to estimate tumor-cell percentages in 20 tissue-image patches. Participants first judged independently and repeated the task with AI assistance after a two-week interval. Incorrect AI estimates received greater weight when they agreed with participants’ incorrect initial estimates, a pattern the researchers operationalized as false confirmation. Time pressure increased overall reliance on AI advice but weakened the association between agreement with the initial estimate and subsequent alignment with the system. Thus, selective acceptance of agreeable advice and general reliance on automation were distinguishable within the experiment. These findings concern a small expert sample, image-based percentage estimates, and a particular decision-support setup; they do not establish the frequency of this mechanism across clinical practice or conversational AI use. (arxiv.org)

These human response patterns should be distinguished from properties of computational systems. NIST’s 2022 treatment of AI bias separates human cognitive biases from statistical and computational biases, including problems caused by unrepresentative training data, modeling choices, and failure to generalize beyond the training setting. Generated factual errors, termed confabulations in NIST’s generative-AI profile, are another category: fluent output can be false without representing either a human belief or a directional preference. A large language model producing answers that resemble human anchoring or confirmation effects therefore demonstrates an output pattern, not by itself an equivalent mental mechanism. Such comparisons require separately specified benchmarks, controlled prompts, and evidence about how the pattern is produced. (nvlpubs.nist.gov)

Debiasing: Training and Decision Environments

Debiasing interventions differ in whether they change the decision maker’s reasoning or the conditions under which a judgment is produced. Individual-level approaches teach statistical principles, strategies for generating alternatives, or methods for evaluating evidence, sometimes combined with practice and feedback. Environmental approaches alter information displays, incentives, or the procedures through which judgments are elicited and combined. These approaches address different sources of error: greater motivation cannot necessarily supply missing statistical knowledge, whereas an information display can reduce computational demands without teaching a reusable skill. Jack Soll, Katherine Milkman, and John Payne’s review, A User’s Guide to Debiasing, organizes interventions around this person–environment distinction and notes that monetary incentives and accountability are not general cures for bias. (katymilkman.com)

In 2015, Carey Morewedge and colleagues reported two longitudinal experiments comparing a single instructional video with an interactive computer game. Videos lasted approximately 30 minutes and games approximately 60 minutes. Experiment 1 targeted bias blind spot, confirmation bias, and fundamental attribution error; Experiment 2 targeted anchoring, representativeness, and social projection. Both formats explained the biases and mitigating strategies, while games additionally elicited judgments, supplied personalized feedback, and provided practice. Game participants showed average reductions in measured bias of 46.25% and 31.94% immediately after training in Experiments 1 and 2, respectively; corresponding video reductions were 18.60% and 25.70%. At follow-up two or three months later, reductions were 34.76% and 23.57% for games and 20.10% and 19.20% for videos. These were changes in experimental bias measures, not percentages of consequential decisions rendered error-free. (journals.sagepub.com)

Transfer concerns whether learning improves performance beyond the examples or formats used during instruction. The 2015 experiments found reductions across different problem contexts and some untrained task formats, supporting transfer within the targeted biases rather than demonstrating improvement in every form of reasoning. A subsequent 2019 study by Anne-Laure Sellier, Irene Scopelliti, and Morewedge examined 290 graduate business students who received game-based training before or after an unannounced classroom business case. Trained participants were 29% less likely to select the inferior, hypothesis-confirming solution. Training schedules, rather than random assignment, determined exposure before the case. The result extends evidence to an educational setting without an explicit reminder to apply training, but the outcome remained a case solution rather than an actual commercial decision with financial consequences. (decisiontimes.org)

Information presentation can improve performance through a different mechanism. In their 1995 experiments, Gerd Gigerenzer and Ulrich Hoffrage showed that natural-frequency representations facilitated Bayesian inference without prior instruction. Instead of separately presenting a base rate and conditional probabilities, such representations preserve counts within a common sample: the relevant posterior proportion can then be calculated by dividing the count satisfying both conditions by the total count satisfying the observed condition. This reduces the operations needed to combine the evidence. A display functioning as a decision aid can therefore improve accuracy while it is available without establishing that the user has acquired a transferable statistical skill. (pages.ucsd.edu)

Structured procedures also require task-specific evaluation. In an experiment on clinical-like personality judgments, a “consider-the-opposite” procedure that sorted cues by diagnostic relevance reduced single-cue anchoring, but simply taking notes while reviewing each cue produced comparable results. The comparison matters because improvement cannot automatically be attributed to the intervention’s named strategy: systematic attention to evidence may account for part of the effect. Evaluations consequently distinguish recognition of bias from changed judgment, immediate gains from durable gains, and aided performance from unaided transfer. Describing a successful procedure in a particular experiment does not establish its effectiveness across professional decisions with different information, time constraints, or outcome criteria. (pubmed.ncbi.nlm.nih.gov)

Competing Explanations and Ecological Rationality

The heuristics-and-biases approach evaluates judgments against specified standards, such as the rules of probability, and explains systematic departures through processes such as representativeness or availability. It does not require that every heuristic produce error: Daniel Kahneman and Amos Tversky’s 1996 response to Gerd Gigerenzer defended the usefulness of heuristics while arguing that systematic errors remained demonstrable. Ecological rationality shifts the emphasis toward the fit between a strategy and its environment. Connected to bounded rationality, this approach asks which information structures make a particular search, stopping, and decision rule successful, rather than treating reduced computation as necessarily inferior. (scispace.com)

A concrete example is the take-the-best heuristic, formalized by Gigerenzer and Daniel Goldstein in 1996. To infer which of two objects has a higher criterion value—for example, which city has the larger population—the rule searches cues in order of predictive validity, stops at the first cue that distinguishes the objects, and selects the object favored by that cue. It ignores remaining cues rather than combining them into a weighted score. Its success depends partly on environmental structure: it is particularly suitable when a strong cue cannot readily be outweighed by several weaker cues. Thus, ignoring information is a substantive modeling commitment, not merely an unspecified shortcut. (web.mit.edu)

Less information can also improve prediction by reducing sensitivity to sampling noise. With a small sample size, estimating numerous predictor weights can capture accidental relationships that fail to generalize. A simpler rule may fit observed data less closely yet predict new cases more accurately. Gigerenzer and Wolfgang Gaissmaier’s 2011 review identified small samples and low predictability as conditions under which such “less-is-more” effects can occur. Their argument also concerns robustness when relationships change: elaborate weighting may depend on estimates that cease to be reliable. This is not a universal advantage for simplicity; comparative accuracy must be established for the relevant environment and prediction task. (annualreviews.org)

Another disagreement concerns what participants understand an experimental question to mean. In their 1999 investigation of the conjunction fallacy, Ralph Hertwig and Gigerenzer argued that ordinary-language “probability” can convey plausibility or evidential support rather than mathematical probability. Conversational interpretation may therefore affect answers to the Linda problem: respondents need not initially understand the investigator’s alternatives as explicitly nested sets. Their experiments showed that contexts encouraging mathematical interpretations increased conformity to the conjunction rule. This challenges the classification of some responses, not the rule itself: when $A$ and $B$ denote events under the same probability measure, $P(A \cap B) \leq P(A)$. (onlinelibrary.wiley.com)

Information format provides a separate, experimentally manipulable explanation. In 1995, Gigerenzer and Ulrich Hoffrage found that natural-frequency presentations facilitated Bayesian inference without instruction. Such presentations preserve counts within a common sampled population: if 10 objects possess a property, eight of those yield a positive signal, and 95 objects without the property also yield that signal, the relevant proportion is directly $8/(8+95)$. Separate percentages require reconstructing these joint quantities. Nevertheless, Kahneman and Tversky’s 1996 response documented systematic biases in frequency judgments, and subsequent experiments compared natural-frequency explanations with the hypothesis that transparent nested-set relations produce the improvement. Frequency wording is therefore neither an automatic cure nor an explanation independent of task structure. (pages.ucsd.edu)

These disputes distinguish mathematical coherence from predictive achievement as evaluative benchmarks. Tests of ecological rationality compare strategies across environments with different cue validities, redundancies, and amounts of available data; tests of conversational explanations manipulate wording while holding the underlying event relations constant. Computational success also needs to be separated from descriptive accuracy: an algorithm’s predictions may be good without establishing that participants actually use it. For example, Nick Chater, Mike Oaksford, and Richard Cooper’s 1997 critique compared take-the-best with alternative learning algorithms and questioned its psychological generality. The competing claims consequently concern identifiable models, interpretations, and performance conditions, rather than a single verdict that people are either irrational or immune to bias. (annualreviews.org)

Evidence Quality and Unresolved Research Questions

Replication concerns a specified effect under specified conditions, not “cognitive bias” as a single phenomenon. Reproducing a difference between experimental conditions establishes neither the universality of that difference nor its proposed psychological mechanism. Richard Klein and colleagues’ 2018 Many Labs 2 project tested 28 psychological findings across 125 samples comprising 15,305 participants in 36 countries and territories. Fourteen findings showed statistically significant effects in the original direction; variability in effect magnitude was attributable more to the finding tested than to the sample or setting. Because the findings were selected rather than randomly sampled, this proportion is not an estimated replication rate for all cognitive biases. The project illustrates why conclusions must distinguish the existence, magnitude, contextual robustness, and explanation of individual effects. (journals.sagepub.com)

Sampling and publication practices constrain generalization in different ways. Student or online convenience samples do not automatically represent professional decision-makers or populations with different educational histories. For example, the first study in Vincent Berthet’s 2021 measurement investigation recruited 163 French undergraduates, including 137 women, with a mean age of 18.52 years. Such a sample can test a task’s properties without establishing population-wide norms. Meanwhile, selective publication of positive results can make an effect appear larger or more consistent than the complete evidence warrants. Joseph Simmons, Leif Nelson, and Uri Simonsohn demonstrated in 2011 that undisclosed flexibility in stopping data collection, excluding observations, choosing outcomes, and selecting analyses can substantially increase false-positive findings. Preregistered protocols and disclosure of all conditions and measures make these sources of uncertainty more assessable; effect sizes and confidence intervals remain necessary alongside significance tests. (pmc.ncbi.nlm.nih.gov)

Measurement creates a further distinction between reliable experimental effects and reliable rankings of individuals. A task may consistently shift a group’s average response while providing an unstable estimate of any participant’s susceptibility. Single-item questions also confound the intended bias with particular wording, background knowledge, or interpretation. Berthet’s 2021 investigation obtained internally consistent measures for six of eight previously identified biases, but did not obtain satisfactory reliability for sunk-cost and belief-bias measures using the replicated procedures. It assessed internal consistency rather than test–retest stability. Its four-item confirmation-bias measure concerned evidence weighting, leaving open whether information search and interpretation belong to the same individual-difference dimension. Thus, a reliable score for one operationalization does not establish a comprehensive measure of the named bias. (pmc.ncbi.nlm.nih.gov)

Evidence also challenges treating rationality as a single ability whose absence explains all biases. In 2015, Predrag Teovanović, Goran Knežević, and Lazar Stankov administered multi-item measures of seven biases to 243 undergraduates. The measures covered anchoring, belief bias, overconfidence, hindsight bias, base-rate neglect, outcome bias, and sunk-cost effects. Although all except hindsight bias had internal-consistency coefficients above $0.70$, inter-bias correlations were generally below $0.20$. Exploratory analysis yielded two factors whose robustness was doubtful. These findings challenge a single-factor account for that battery, rather than demonstrating that every bias is independent. By contrast, Wändi Bruine de Bruin, Andrew Parker, and Baruch Fischhoff’s 2007 seven-task decision-making-competence battery supported an aggregate index associated with fewer self-reported negative life events. The contrast makes task selection and criterion validity central unresolved issues. (sciencedirect.com)

Open questions concern what laboratory measures predict and which processes they identify. Associations with reported life outcomes do not establish causal effects or precise prospective prediction; opportunities to encounter an outcome also differ between people. Cross-cultural comparisons must separate changes in the target process from differences in language, task familiarity, and interpretation: multinational replication alone does not establish equivalent measurement. Intervention research faces a related transfer problem. Carey Morewedge and colleagues’ 2015 experiments found training effects persisting for at least two months and extending to untrained problem formats. Those results establish persistence and transfer within the tested assessments, not durable improvement across consequential workplace or everyday decisions. Longer follow-up, independent outcome measures, and tests distinguishing competing process explanations remain important questions for this evidence base. (cmu.edu)

References

1. Judgment under Uncertainty: Heuristics and Biasescsc2.ncsu.edu
2. Judgment under Uncertainty: Heuristics and Biases - PubMedpubmed.ncbi.nlm.nih.gov
3. Fallacies (Stanford Encyclopedia of Philosophy)plato.stanford.edu
4. Investigating Bias in Maximum Likelihood Quantum State Tomographynist.gov
5. The Seven Sins of Memorysites.harvard.edu
6. Judgment under Uncertainty: Heuristics and Biasespsy.ucsd.edu
7. Prospect Theory: An Analysis of Decision under Riskideas.repec.org
8. Prospect Theory: An Analysis of Decision under Riskweb.mit.edu
9. The Framing of Decisions and the Psychology of Choicestanford.edu
10. Contents — Judgment under Uncertaintycambridge.org
11. Probabilistic reasoning in clinical medicine: Problems and opportunitiescambridge.org
12. Bounded Rationalityplato.stanford.edu
13. Descriptive Decision Theoryplato.stanford.edu
14. Glossary of Statistical Termsstat.berkeley.edu
15. Normative Theories of Rational Choice: Expected Utilityplato.stanford.edu
16. Decision Theoryplato.stanford.edu
17. Outcome Bias in Decision Evaluationbear.warrington.ufl.edu
18. Heuristics Made Easy: An Effort-Reduction Frameworkpages.ucsd.edu
19. Dual-Process Theories of Higher Cognitionpsychologicalscience.org
20. The case for motivated reasoningpubmed.ncbi.nlm.nih.gov
21. The Case for Motivated Reasoningbear.warrington.ufl.edu
22. Fear, anger, and risk.pubmed.ncbi.nlm.nih.gov
23. The Measurement of Individual Differences in Cognitive Biases: A Review and Improvementfrontiersin.org
24. Prone to Bias: Development of a Bias Taxonomy From an Individual Differences Perspectivecris.huji.ac.il
25. A Task-Based Taxonomy of Cognitive Biases for Information Visualizationresearch-portal.uu.nl
26. The seven sins of memory. Insights from psychology and cognitive neuroscience.pubmed.ncbi.nlm.nih.gov
27. The heuristics-and-biases inventory: An open-source tool to explore individual differences in rationalityfrontiersin.org
28. The heuristics-and-biases inventory: An open-source tool to explore individual differences in rationalitypubmed.ncbi.nlm.nih.gov
29. On the Psychology of Prediction, No. 4scholarsbank.uoregon.edu
30. Maps of Bounded Rationality: Psychology for Behavioral Economicsbear.warrington.ufl.edu
31. Availability: A Heuristic for Judging Frequency and Probabilitynoorsiddiqui.com
32. Ease of Retrieval as Information: Another Look at the Availability Heuristicroar-assets-auto.rbl.ms
33. The anchoring-and-adjustment heuristic: why the adjustments are insufficient.pubmed.ncbi.nlm.nih.gov
34. Explaining the Enigmatic Anchoring Effect: Mechanisms of Selective Accessibilitybear.warrington.ufl.edu
35. Confirmation Bias: A Ubiquitous Phenomenon in Many Guisespsy.ucsd.edu
36. On the failure to eliminate hypotheses in a conceptual taskbear.warrington.ufl.edu
37. Confirmation, Disconfirmation, and Information in Hypothesis Testingbear.warrington.ufl.edu
38. Biased Assimilation and Attitude Polarization: The Effects of Prior Theories on Subsequently Considered Evidencenuovoeutile.it
39. Reconstruction of automobile destruction: An example of the interaction between language and memorygwern.net
40. Hindsight ≠ Foresight: The Effect of Outcome Knowledge on Judgment Under Uncertaintyweb.mit.edu
41. Confidence, Not Consistency, Characterizes Flashbulb Memoriespsychologicalscience.org
42. The correspondence biaspubmed.ncbi.nlm.nih.gov
43. The Correspondence Biasdtg.sites.fas.harvard.edu
44. Is there a universal positivity bias in attributions? A meta-analytic review of individual, developmental, and cultural differences in the self-serving attributional biaspubmed.ncbi.nlm.nih.gov
45. Self-Serving Biases in the Attribution of Causality: Fact or Fiction?bulidomics.com
46. The Bias Blind Spot: Perceptions of Bias in Self Versus Othersphilosophy.rutgers.edu
47. Valuing thoughts, ignoring behavior: The introspection illusion as a source of the bias blind spotprivacycasebook.com
48. Implicit social cognition: attitudes, self-esteem, and stereotypespubmed.ncbi.nlm.nih.gov
49. Implicit Social Cognition: Attitudes, Self-Esteem, and Stereotypesbanaji.sites.fas.harvard.edu
50. Experimental Tests of the Endowment Effect and the Coase Theoremweb.mit.edu
51. The Willingness to Pay-Willingness to Accept Gap, the "Endowment Effect," Subject Misconceptions, and Experimental Procedures for Eliciting Valuationsswlb2.aeaweb.org
52. Status quo bias in decision makingweb.mit.edu
53. Measuring Individual Differences in Decision Biases: Methodological Considerationsfrontiersin.org
54. A computational process-tracing method for measuring people’s planning strategies and how they change over timecocosci.princeton.edu
55. The reliability paradox: Why robust cognitive tasks do not produce reliable individual differencespubmed.ncbi.nlm.nih.gov
56. Journal Article Reporting Standards for Quantitative Research in Psychologyapa.org
57. On the relative independence of thinking biases and cognitive abilitypubmed.ncbi.nlm.nih.gov
58. Numeracy and decision makingpubmed.ncbi.nlm.nih.gov
59. Conditions for Intuitive Expertise: A Failure to Disagreebear.warrington.ufl.edu
60. Framing effects on adolescent task representations, analytic and heuristic processing, and decision making: Implications for the normative/descriptive gapgwern.net
61. How to Improve Bayesian Reasoning Without Instruction: Frequency Formatspages.ucsd.edu
62. The Effects of Financial Incentives in Experiments: A Review and Capital-Labor-Production Frameworkauthors.library.caltech.edu
63. Diagnostic Errorspsnet.ahrq.gov
64. Results | Agency for Healthcare Research and Qualityahrq.gov
65. Forensic Handwriting Examination and Human Factors: Improving the Practice Through a Systems Approachnvlpubs.nist.gov
66. Applying Empirical Psychology to Inform Courtroom Adjudication—Potential Contributions and Challengesharvardlawreview.org
67. Pooling of Unshared Information in Group Decision Making: Biased Information Sampling During Discussionuni-muenster.de
68. Motivated information processing in group judgment and decision makingrepub.eur.nl
69. Motivated information processing and group decision-making: Effects of process accountability on information processing and decision qualityresearch.rug.nl
70. Shifting attention to accuracy can reduce misinformation onlinenature.com
71. Shifting attention to accuracy can reduce misinformation onlinemoodle2.units.it
72. Quantifying the Impact of Cognitive Biases in Question-Answering Systemsojs.aaai.org
73. Quantifying the Impact of Cognitive Biases in Question-Answering Systemsojs.aaai.org
74. “When Two Wrongs Don’t Make a Right” — Examining Confirmation Bias and the Role of Time Pressure During Human-AI Collaboration in Computational Pathologyarxiv.org
75. Towards a Standard for Identifying and Managing Bias in Artificial Intelligencenvlpubs.nist.gov
76. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profilenvlpubs.nist.gov
77. A User’s Guide to Debiasingkatymilkman.com
78. Debiasing Decisions: Improved Decision Making With a Single Training Interventiondecisiontimes.org
79. Debiasing Training Improves Decision Making in the Fieldurisohn.com
80. Procedural debiasing of primacy/anchoring effects in clinical-like judgmentspubmed.ncbi.nlm.nih.gov
81. On the Reality of Cognitive Illusionsscispace.com
82. Reasoning the Fast and Frugal Way: Models of Bounded Rationalityweb.mit.edu
83. Rationality the Fast and Frugal Wayescholarship.org
84. The Measurement of Individual Differences in Cognitive Biases: A Review and Improvementpmc.ncbi.nlm.nih.gov
85. Individual Differences in Adult Decision-Making Competencecmu.edu