Overconfidence bias concerns a measurable mismatch between what people believe about the quality of their judgments and what an appropriate performance criterion shows. The term covers three distinct patterns: overestimating one's absolute performance, placing oneself too high relative to other people, and expressing greater precision than the available evidence warrants. These patterns are related, but a person can exhibit one while showing the opposite tendency on another. In research on judgment and decision-making, a claim of overconfidence becomes meaningful only when the task, confidence report, comparison standard, and sampling procedure have been specified. Moore and Healy (2008) provide the central framework.
Excessive confidence matters because judgments guide information search, willingness to reconsider, delegation, investment, professional advice, and the use of increasingly capable decision-support technologies. Yet high confidence is not itself a psychological defect: a highly competent person may have good reason to be confident. The scientific question is whether confidence is warranted by observable accuracy or by a calibrated assessment of uncertainty. Research has established systematic miscalibration under many conditions, while also documenting well-calibrated judgments, underconfidence, substantial individual and domain differences, and measurement artifacts. Klayman and colleagues (1999) showed why the form of the question changes the result.
What Is Overconfidence Bias?
In experimental psychology, overconfidence bias is a systematic tendency for reported certainty, self-assessed performance, or comparative standing to exceed the relevant standard across a specified set of judgments. The standard may be an objective test score, another person's actual performance, the frequency with which propositions prove correct, or coverage of numerical uncertainty intervals. A single confident error does not establish a bias. Researchers ordinarily need multiple observations and an explicit method of aggregation before they can estimate whether overconfidence is present.
Confidence is a subjective judgment about the correctness of an answer, the likely occurrence of an event, one's competence, or the precision of an estimate. Accuracy refers to correspondence with an established outcome or answer; it cannot be inferred from the speaker's certainty. Calibration evaluates how stated probabilities correspond to observed frequencies, whereas resolution or discrimination evaluates whether higher confidence successfully distinguishes the better predictions from the worse predictions. Fleming and Lau (2014) explain why these metacognitive quantities must be measured separately.
Overconfidence is a descriptive and empirically testable phenomenon. It is not a clinical diagnosis, a symptom that independently indicates mental illness, or a universal personality category. Nor should it be equated with deliberate deception or arrogance. Someone may sincerely report a probability that systematically exceeds the outcomes, even while caring strongly about accuracy. Emotional states, incentives, experience, task familiarity, question wording, and feedback conditions all influence the particular confidence report.
Overconfidence, self-esteem, optimism, and certainty
Self-esteem evaluates personal worth; self-efficacy concerns expected capability to perform particular actions; optimism describes favorable expectations about future outcomes; overconfidence concerns a mismatch against a specified standard. These constructs can correlate in a study without becoming interchangeable. A person can have healthy self-esteem and well-calibrated forecasts, or a pessimistic outlook and overprecise numerical estimates. An unsupported assertion of certainty is a conversational behavior; an empirical overconfidence finding requires defined observations and a measurement model.
A logical fallacy is an invalid form of argument assessed relative to rules of inference. A statistical error is a mistake in estimation, modeling, or inference. Overconfidence bias is a property of confidence judgments relative to their criteria and may appear alongside either kind of error, but it requires its own evidence. Likewise, a heuristic is an information-processing strategy: using a cue quickly can sometimes support excellent judgments. Whether a heuristic becomes associated with miscalibration depends on the environment and the task.
Three Forms of Overconfidence
Overestimation: How good am I?
Overestimation occurs when a person predicts an absolute score or performance level that exceeds the actual result. A student who expects 90 correct answers out of 100 and earns 72 has overestimated that performance on the specified test. Researchers can compare predictions made before a task or estimates reported afterward with actual outcomes. The assessment becomes more informative when prediction and performance are comparable, items represent the claimed domain, and chance or hindsight effects are controlled.
People can also underestimate absolute performance. An easy task with an unexpectedly high success rate can yield underestimation when predictions are conservative. Moore and Healy's laboratory evidence illustrates why even an empirically observed overestimation pattern should not be generalized to every task or ability. Moore and Healy (2008).
Overplacement: How do I compare with other people?
Overplacement means believing that one's relative position is higher than one's measured position in a defined reference group. A participant may place themselves in the 80th percentile on a task while their observed percentile is the 55th. This is related to, but more precisely measured than, a generic better-than-average statement. The investigator must specify who constitutes the comparison group, how ranks are determined, and what respondents know about the group.
Relative judgments can reverse with task difficulty. On easy tasks, people may know that they performed well yet underestimate how well everyone else also performed, yielding overplacement. On difficult tasks, they may know they struggled but underestimate others' struggle, yielding underplacement. Commonly described as the better-than-average and worse-than-average patterns, these are partly consequences of comparing oneself with an imperfect mental model of other people's scores. Burson, Larrick, and Klayman (2006) and Moore and Healy (2008) investigate these reversals.
Overprecision: How narrow is my uncertainty?
Overprecision arises when a confidence distribution or interval implies more certainty than outcomes justify. Suppose someone provides many numerical ranges designed to contain the true answers 90% of the time, yet only 58% of the answers fall inside. That result indicates undercoverage, one operational sign of excessive precision for that collection of questions. The values here illustrate the arithmetic, not an estimate of population prevalence.
Overprecision is especially visible in interval-estimation tasks. In the experiments of Soll and Klayman (2004), subjective intervals captured the correct answer less often than their stated coverage levels implied. But an overly narrow interval cannot always be interpreted as a narrow underlying belief distribution: interval-construction procedures, response variability, and statistical selection can contribute. Moore, Carter, and Yang (2015) challenge the simple explanation that everyone's internal uncertainty distribution must itself be too narrow.
The three forms cannot be combined into a single unconditional score without assumptions. Someone can overestimate their score on a difficult examination while correctly recognizing that others did better. Another person can be underconfident about an easy task and still rank themselves unrealistically above peers. The same individual can have comparatively accurate forecasts but set numerical confidence intervals that are too narrow. An assessment that measures only one expression of confidence cannot establish all three.
Confidence, Accuracy, and Calibration: The Mathematical Standard
Calibration has a frequency interpretation. Across a large, relevant group of binary predictions assigned a probability of 0.70, an ideally calibrated forecaster would see approximately 70% of those events occur. The event must be defined before outcomes are revealed, and there must be enough observations for a stable estimate. The expected agreement is statistical, not a promise that exactly seven of every ten events will occur in any small sample.
For a series of chosen answers with subjective confidence c and observed correctness y equal to 1 for correct and 0 for incorrect, a simple mean calibration gap is the average of c minus the average of y. A positive gap indicates average overconfidence in correctness across that set; a negative gap indicates average underconfidence. For example, 80 judgments reported at mean confidence 0.85 with 60 correct produce accuracy 0.75 and a positive gap of 0.10, or 10 percentage points. This example is hypothetical. A gap near zero does not prove excellent calibration at every probability level, because different errors can cancel.
A reliability diagram groups predictions into confidence ranges and plots their observed correctness. For a binary proposition assigned probability p, the graph compares average p with the proportion of events that actually occurred. Calibration plots require sufficiently populated bins, clear probability conventions, uncertainty intervals around empirical rates, and attention to selection effects. A study in which participants may decline difficult questions needs to examine the answered set and abstained set separately. Stone and colleagues (2023) use individualized calibration diagrams for intervention feedback.
Calibration and discrimination can diverge
A forecaster who always assigns 60% to an event that occurs 60% of the time is calibrated in aggregate but makes little distinction among specific occasions. A second forecaster may assign relatively higher probabilities to events that occur and lower probabilities to events that do not, yet be overconfident because the stated probabilities are too extreme. Calibration asks whether numerical confidence matches outcome frequencies. Discrimination asks whether judgments sort likely events from unlikely events. Both matter for actionable decisions.
Metacognitive sensitivity examines whether a person tends to be more confident on their own correct than incorrect responses while accounting for first-order task performance and response biases. Signal detection approaches, confidence-accuracy correlations, and metacognitive-efficiency models answer related but different questions. Some require modeling assumptions about the distribution of evidence. Fleming and Lau (2014) review the methodological choices.
Brier scores, accuracy, and proper scoring
For a binary outcome y and predicted event probability p, the Brier score is the mean of (p - y) squared across predictions. Smaller scores indicate better probability predictions on a shared event set. Forecast scoring encourages meaningful probabilities and captures information beyond a simple calibration gap, including aspects of resolution. Comparing forecasters across different event sets requires care because task difficulty and event prevalence change the score. The original scoring approach was introduced in Brier (1950).
A numerical estimate's 90% prediction interval must also be distinguished from a frequentist confidence interval for an unknown population parameter. In an overprecision experiment, the psychological issue is commonly whether a person's stated subjective range captures the true target values as often as intended. The mathematical interpretation of a frequentist interval across repeated sampling is a separate inferential concept. Clear terminology prevents an everyday confidence claim from being conflated with a formal statistical procedure.
Historical Development of Overconfidence Research
Research on probabilistic judgments and calibration developed through experimental work on general knowledge, forecasting, and expert performance in the second half of the twentieth century. Early investigators asked participants to select answers and assess the probability that each answer was correct. Researchers then compared these judgments against observed proportions correct to construct calibration curves. The methodological legacy remains central: the question set, scoring scheme, and probability response scale define the finding.
The heuristics-and-biases program drew attention to systematic ways judgment can depart from probability norms. Later work asked whether familiar miscalibration patterns represented general mental failures, predictable consequences of noisy evidence, properties of selected question samples, or reasonable responses to environments with particular cue distributions. That debate improved operational definitions and prompted more careful experiments. Juslin, Winman, and Olsson (2000) critically examined interpretations of the hard-easy effect, while Klayman and colleagues (1999) separated psychological miscalibration from expected statistical effects.
The influential study by Kruger and Dunning (1999) connected low measured performance and inflated self-assessment with limits on metacognitive recognition of error. Later research refined the account by studying task difficulty and perceived peer performance. Contemporary cognitive science further separates task accuracy, confidence reporting, evidence strength, and metacognitive sensitivity; newer research increasingly tests models rather than treating one demonstration as a universal law. Burson, Larrick, and Klayman (2006).
Why People Become Overconfident: Competing Mechanisms
Sampling evidence from memory and interpretation
A confidence judgment typically rests on partial internal evidence: recalled examples, perceived fluency, how many reasons favor an answer, consistency among cues, or familiarity with a task. Such cues can be informative, but they are imperfect proxies for correctness. The subjective coherence of evidence can increase confidence even when the sampled evidence is incomplete or systematically tilted. Koriat (2012) models confidence partly through the consistency of sampled representations; the model predicts both useful confidence and circumstances in which consistent support favors the wrong answer.
The same mechanism can help explain why asking people to list reasons for a favored judgment may produce confidence while leaving alternative hypotheses insufficiently explored. Whether additional reasoning helps depends on the quality and independence of the evidence. Multiplying correlated arguments does not create independent confirmation. This points toward a practical intervention: separate new information from repeated versions of the same source before increasing certainty.
Noisy signals and regression effects
Perceived ability, actual performance, and reported confidence are all measured with error. When investigators group participants by extreme observed test scores, partly random error tends to regress toward the average on repeat testing. Bounded rating scales, item selection, and mathematical dependence between accuracy and the confidence gap can produce graphical patterns that resemble a psychological bias. Consequently, an observed slope on a performance-versus-overestimation graph cannot uniquely identify a metacognitive mechanism.
The critical review by Juslin, Winman, and Olsson (2000) showed how item sampling and scale artifacts alter classic findings. Klayman and colleagues (1999) used alternative methods to distinguish statistical inevitabilities from substantive effects. Both types of explanation can operate in the same dataset. Rigorous studies must show which remains after reasonable controls and how much variance is explained.
Relative comparisons and shared task difficulty
Participants observe their own performance more directly than the performances of an entire comparison group. Estimating rank therefore requires at least two uncertain quantities: one's score and the distribution of others' scores. When a task appears easy, the first quantity may be high while perceived peer performance is insufficiently high. When a task is difficult, people's own experience of struggle may be more salient than others' unseen struggle. This can change relative judgments without requiring a fixed desire to believe oneself superior. Moore and Healy (2008).
Motivation and social rewards
People have incentives to present competence, sustain a positive identity, persuade collaborators, or act decisively under uncertainty. In some experiments, social comparison and favorable self-presentation can influence confidence reports. Yet the observed gap could also emerge from information constraints even when incentives favor accuracy. Confidence elicited in public, in private, with financial accuracy incentives, or under social evaluation may be different. Analysts should avoid reading a private confidence error as proof of a personality motive.
Learning, feedback, and outcome attribution
Repeated outcomes provide opportunities to learn if feedback is timely, accurate, and sufficiently diagnostic. Feedback can also mislead if successful results arise partly from luck, easy cases are overrepresented, or failures are explained away. A formal model of trading by Gervais and Odean (2001) examines how attributing success to ability can generate overconfidence during learning; its predictions should be distinguished from direct psychological experiments. Detailed probability feedback is more informative than a vague instruction to 'be less confident.'
The Hard-Easy Effect: Task Difficulty Changes the Direction of Bias
In some traditional general-knowledge tasks, confidence exceeds accuracy on hard questions and falls below accuracy on easy ones. This pattern is called the hard-easy effect. For instance, when very challenging questions produce 55% correct responses but 70% average stated confidence, the group is overconfident. If easier questions produce 95% correct with 88% confidence, that group is underconfident. Both examples are hypothetical, and the numerical gap is tied to the selected item set.
The effect deserves careful interpretation. Selecting unusually easy or hard items after observing sample responses can induce artifacts; confidence may be bounded by scale endpoints; noisy confidence and performance measures can make the difference partly mathematically dependent on accuracy. When researchers use representative item sampling, different response forms, and statistical controls, some classic hard-easy patterns shrink or disappear. Juslin, Winman, and Olsson (2000). Meanwhile, other well-controlled studies still find task-dependent miscalibration, so it is equally unjustified to declare that all overconfidence is only an artifact. Klayman and colleagues (1999).
The applied implication is to calibrate within meaningful task categories rather than importing a single confidence adjustment from an easy quiz into difficult medical or geopolitical cases. Forecasting quality can change greatly when questions require unfamiliar expertise, new information, or rare-outcome prediction.
The Dunning-Kruger Effect and the Limits of Self-Assessment
The Dunning-Kruger effect is commonly associated with the finding that weaker performers can substantially overrate their position on skill tests, partly because the abilities required to perform a task can overlap with those required to recognize errors. In the original studies, participants in the lowest performance quartile showed substantial percentile overestimation on particular humor, grammar, and logic tasks. Kruger and Dunning (1999).
The popular claim that every beginner is supremely certain while every expert doubts themselves misrepresents the range of findings. The key measures concern performance, relative self-placement, and calibration on defined tasks. Regression to the mean, measurement noise, differences in knowledge of peer ability, and task difficulty can contribute to the characteristic plots. Burson, Larrick, and Klayman (2006) found patterns involving perceived difficulty that require a richer explanation than a simple division between ignorant confidence and expert modesty.
Researchers therefore need reliable performance assessments, individual-level estimates with uncertainty, appropriate comparison populations, and designs that distinguish stable metacognitive skill from sampling artifacts. Even highly skilled individuals can be overconfident in new domains or in interval estimation; conversely, less skilled individuals can recognize uncertainty when given clear feedback. Skill and calibration are both learnable in some settings, but learning curves should be measured rather than presumed.
Experimental Tasks and How They Are Measured
General knowledge confidence judgments
A standard design presents questions with known answers, often in a two-alternative format, and asks participants both to choose and to rate the probability that their answer is correct. Researchers compute calibration differences across confidence categories and test whether confidence predicts individual correctness. Because a binary forced-choice task has a nominal chance-level baseline under guessing, many studies use confidence scales from 50% to 100%; mixing such scales with event probabilities from 0% to 100% requires explicit transformation.
Question selection matters. A set enriched with trick questions can overrepresent errors; a narrowly familiar domain may produce better calibration; item difficulty can depend on linguistic and cultural background. The investigator should preregister selection criteria where feasible and report the number of participants, items, usable judgments, and uncertainty intervals around the central estimates.
Post-task performance predictions
Participants may estimate the number correct immediately before or after taking a test. Researchers compare these values with the actual score. Post-task self-estimates incorporate experience of the test, so they answer a different question from prospective predictions. If feedback has already revealed answers, the estimate no longer assesses the same pre-outcome uncertainty. Researchers often distinguish self-assessed raw scores, relative ranks, and confidence in a particular answer.
Relative percentile ratings
Individuals rate where they expect to rank relative to others. The ranking criterion might be speed, correctness, score, or another objective metric. Accuracy requires the actual reference group's scores, not assumed population percentiles. Small samples have coarse rankings, and a person may compare with imagined peers rather than the study participants. Well-designed experiments manipulate difficulty and peer information to distinguish an own-performance error from an error about others.
Subjective confidence intervals
People create low and high bounds that they believe contain an unknown numerical fact at a designated probability. Investigators score the empirical coverage over many questions and examine width, bias, and stability. Interval widths can vary greatly within a person across items. Sequential elicitation, allowed response formats, and outlier trimming change results. Soll and Klayman (2004) demonstrate why interval estimates should not be treated as interchangeable with two-choice confidence.
Forecasts of events
Participants supply probabilities for independently verifiable outcomes, ideally before a deadline and with scoring rules that incentivize accuracy. Forecast archives preserve what was known at the time and allow assessment of resolution, calibration, Brier scores, updating, and performance across different event classes. Tournaments and repeated feedback can reveal skills that a brief laboratory exercise cannot. Mellers and colleagues (2015).
What Experiments and Meta-Analyses Actually Show
The overconfidence literature supports substantial task and measurement heterogeneity. Moore and Healy (2008) integrated apparently contradictory findings by separating overestimation, overplacement, and overprecision. Klayman and colleagues (1999) found comparatively little general overconfidence in their two-alternative questions but pronounced overconfidence in subjective intervals. These are experimentally grounded conclusions for studied tasks, not prevalence estimates for all human cognition.
Evidence also shows that confidence carries genuine information. A meta-analysis of 213 distinct datasets in the Confidence Database reported an across-person association between average confidence and average accuracy of R = .22, with differences across memory versus perceptual tasks and confidence-scale granularity. This is a small positive aggregate association between people, not an estimate of the average overconfidence gap and not a measure of trial-level calibration. It supports treating confidence as a potentially informative cue while rejecting the assumption that confident individuals necessarily perform best.
Recent applied evidence illustrates domain specificity. A 2026 radiology study compared confidence and correctness over 2,806 decisions from 28 radiologists and 50 nonradiologists. The report found lower overconfidence and better success among radiologists, while difficult cases still produced substantial overconfidence. These figures describe the sampled participants and cases; they cannot establish the prevalence of overconfidence among clinicians generally, and observational comparisons cannot isolate expertise from every other group difference.
There is no single scientifically defensible percentage saying how many people 'have overconfidence bias' across all decisions. Researchers vary the construct, the confidence scale, the difficulty range, expertise, incentives, culture, sampling frame, statistical adjustment, and criterion of correctness. Where pooled effect sizes are reported, they should be interpreted for the task family and design from which they were derived.
Reproducibility and measurement challenges
Several influential experiments have direct replications, methodological critiques, alternative models, and substantial cumulative literature, but replication of one operationalization does not validate all overconfidence mechanisms. Apparent disagreements sometimes arise from different confidence questions rather than irreconcilable data. Preregistration, transparent exclusions, independent samples, representative item pools, measurement reliability, and direct comparisons of models improve interpretability.
Studies also need to distinguish the individual from the group. A group can be well calibrated when confident and unconfident participants cancel each other, while individuals are not. Conversely, most participants might rank their own ability inaccurately in different directions, obscuring the individual-level picture in the mean. Repeated measures and multilevel analyses help examine this heterogeneity.
Culture, Expertise, Personality, and Context
Confidence reports are shaped by social conventions, education, the language of probability, the perceived purpose of a task, and the reference group against which performance is evaluated. Cultural comparisons can be informative, but they require equivalent questions, suitable translations, comparable probability scales, and tests of measurement invariance. A higher mean stated confidence in one group does not establish greater overconfidence unless performance and the underlying criterion are also measured in comparable conditions.
Expertise improves accuracy on many domain-specific tasks, yet its relation to calibration is conditional. Experts may know which problems are diagnostic, gain access to repeated outcome feedback, and understand uncertainty distributions. They may also encounter rare or changing situations in which feedback is delayed and confidence outpaces evidence. The practical question is whether the professional environment supplies enough outcome data to support learning. Confidence established within one specialty should not automatically be transferred to unrelated domains.
Experience may lower or raise confidence depending on how outcomes are attributed. A successful decision can be followed by an undeserved increase in confidence when luck is mistaken for skill. Conversely, experience with informative feedback can sharpen uncertainty estimates. Motivation, personality, and social pressure may matter, but differences are typically interpreted best alongside measured competence, task properties, and incentives rather than through broad labels.
The Role of Emotion, Fluency, and Social Information
Judgment feels easier when evidence is familiar, a narrative is coherent, or the answer emerges rapidly. Processing fluency can serve as one cue for confidence, but the relation between ease and correctness varies with the task. Strong affect may alter the salience of threats or rewards, and time pressure can change which cues are sampled before a decision. These influences have to be established empirically in the specific judgment task; they should not be offered as a universal explanation for every high-confidence mistake.
Group discussion can supply independent evidence or reinforce a shared error. Several people agreeing is especially informative when their evidence sources are independent; agreement based on the same recycled claim can create social certainty without adding much new information. A structured group process can record each person's initial forecast before discussion, identify common sources, and make disagreement explicit. Its value depends on how the procedure is designed and whether information is genuinely independent.
Consequences for Decisions Under Risk and Uncertainty
Excessive precision can make a range of possible outcomes appear too narrow. In project planning, this may cause teams to underestimate the variation in completion times or costs even if their best single-point estimate is reasonable. Overestimation may encourage commitments that exceed actual capabilities; overplacement may motivate entry into competitive settings on unrealistic assumptions about peers. Each mechanism implies a different corrective measure: wider predictive intervals, better tests of one's own ability, or more accurate data about the comparison group.
Probability calibration matters whenever actions depend on estimated event likelihood. A claim that an adverse event is 'very unlikely' can lead to different decisions depending on whether it means 1%, 10%, or 30%. Recording explicit probabilities permits later comparison with actual outcomes. Decisions also depend on values and consequences, so even perfectly calibrated probabilities do not by themselves dictate which action is best. Normative expected utility models require both probabilities and utilities; choosing an action is distinct from estimating a probability accurately.
Professional judgments can be especially consequential when the effects of errors are asymmetric. High-confidence mistakes in medicine, law, security, or engineering warrant explicit verification steps, escalation criteria, and attention to the quality of reference data. Yet these applied domains have their own research methods and standards. Overconfidence research informs the psychology of the confidence report, while domain-specific evidence determines what concrete safeguards work.
Overconfidence in Human-AI Decision Support
AI-generated explanations can be fluent and persuasive. Users may infer reliability from confident language, polished presentation, or apparent consistency, even when the system's factual accuracy depends strongly on the question, model, retrieval process, and validation method. The study of human overconfidence therefore connects to the calibration of trust in technological advice. It does not establish that machine-generated confidence labels have a common meaning across architectures or that a system has subjective confidence.
A human-AI workflow should distinguish the user's certainty, the system's expressed confidence, independently tested model accuracy, and the consequences of a mistake. A numerical confidence score from a system must be calibrated and validated for the target population and use case before it can serve as a dependable probability. When a human changes a decision after AI advice, researchers can measure whether accuracy increases, how confidence changes, and whether the person detects system failures. Those are separate outcomes.
For a specialized examination of technological authority and reliance, see AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making. The present article concerns general overconfidence mechanisms; evaluation of AI system reliability, automation bias, and trust calibration belongs to the broader science of human-AI decision support.
Evidence-Based Approaches to Better Calibration
Collect repeated judgments and use feedback
The first step is to preserve a sufficient series of predictions before outcomes are known. State clear events or answer criteria, record probabilities, group them by confidence, and compare the resulting outcome frequencies. Identify whether errors mainly involve excessive certainty, failure to discriminate among cases, or inaccurate self-ranking. A single memorable mistake provides a story; a consistently scored record provides evidence of the pattern.
Experimental training shows some promise. In two studies, Stone, Luu, Costello, and Somerville (2023) tested automated individualized calibration feedback in forecasting. The intervention lowered the confidence of initially overconfident participants; the improvement in overconfidence and overall calibration was clearer in the second, more controlled task. This distinction matters: reductions in reported confidence do not automatically establish increased accuracy on every task, and results may not transfer to other fields without testing.
Clarify the target and response scale
People answer different questions when asked how certain they are of one answer, what score they expect, how they rank against others, or how wide a 90% interval should be. A reliable decision process asks precisely the form of confidence that is needed. For a comparison question, collect actual peer performance data. For an uncertain number, ask for a range with a stated coverage probability. For an event, request a numeric probability and define how the outcome will be resolved.
Training should also teach the distinction between a probability forecast and an everyday verbal label. Two people may use the word 'likely' for substantially different probability ranges. Writing an explicit number makes the estimate scoreable, while requiring 0% or 100% certainty should be reserved for events that genuinely warrant extremes.
Seek independent evidence and alternative possibilities
A structured prompt asking what evidence could make the current judgment wrong may reduce premature certainty under some conditions. Its value depends on the quality of the counterevidence and the decision context. Listing fanciful objections does little; comparing independent base rates, expert assessments, out-of-sample performance, and alternative hypotheses is more diagnostic. The objective is to improve the information used for judgment, not merely to force everyone to lower every estimate.
An uncertainty range can also be assessed against comparable historical cases. If a project forecast gives an unusually narrow spread relative to past projects with similar complexity, the discrepancy deserves examination. In some settings, reference distributions and simple statistical baselines outperform unaided intuition; their usefulness should be checked against representative data rather than assumed for all decisions.
Separate confidence training from knowledge training
Someone who knows too little about a subject may benefit most from learning the subject. Calibration feedback can help align judgments with current accuracy, but it cannot replace the information and skill required to improve actual performance. Conversely, increasing expertise without teaching probabilistic self-assessment may leave confidence poorly calibrated on unusual cases. These are complementary targets: make answers more often correct, and make confidence more diagnostic of when they are correct.
Use forecasting practice and proper scores
Repeated forecasts with documented deadlines, feedback, and proper scoring create conditions for evaluating incremental progress. Mellers and colleagues (2015) studied unusually accurate geopolitical forecasters and linked performance to combinations of cognitive skill, task-specific practice, motivation, and supportive environments. Tournament findings should not be interpreted as proof that an identical package of techniques will improve every person's decisions in every domain.
An improved Brier score is useful evidence of better probabilistic prediction on comparable events. Researchers should also show calibration and discrimination separately and report whether confidence improvements persist after practice ends or when the question type changes. The transfer of brief debiasing exercises to complex professional decisions remains an empirical question requiring direct study.
What Research Cannot Yet Establish
No single cognitive mechanism explains every case of overestimation, overplacement, and overprecision. The strongest defensible conclusions specify which measure changed and under what conditions. Some controlled experiments detect systematic psychological miscalibration, whereas other designs reveal sizable contributions from question sampling, statistical artifacts, response formats, and changing task difficulty.
Effects observed on university-student general-knowledge tests need separate evidence before they are applied to surgeons, entrepreneurs, judges, or autonomous systems. A person's high average confidence may be justified in a strong expertise domain yet poorly calibrated under unfamiliar or novel conditions. Causal claims about real-world losses also need designs that distinguish confidence from skill, resources, incentives, and exposure to risk. Observational associations alone cannot assign every poor outcome to overconfidence.
Intervention claims deserve the same standard. A post-training shift toward moderate probabilities can lower expressed confidence without improving discrimination or actual decision quality. The best evaluation includes randomized comparisons where feasible, prespecified metrics, follow-up over time, and tests in consequential situations. An effective intervention for numerical forecasts is not automatically effective for comparative self-assessment or social persuasion.
Practical Examples: Identifying the Right Kind of Error
Consider an engineer who predicts that a project will take 20 workdays and later requires 38. The discrepancy could reflect inaccurate base data, unexpected scope changes, or an overprecise schedule. Several projects with stated probability ranges are needed before diagnosing a stable interval-calibration problem. The difference between one forecast and one outcome is evidence of forecast error, not by itself proof of a cognitive bias.
A manager who believes they are among the top 10% of negotiators can be tested on a meaningful negotiation-performance criterion and compared with a defined peer group. A mismatch would concern overplacement. A manager who also expects to close 90% of deals but closes 60% on comparable opportunities may additionally show overestimation. Those two measurements should be reported independently.
A researcher who gives 90% confidence to each of 100 independently scored answers and answers only 70 correctly exhibits a 20-percentage-point aggregate confidence gap in the observed set. This supports overconfidence in correctness for those questions, while more detailed analysis should examine confidence bins, task difficulty, and how the questions were selected. The calculations are illustrative rather than estimated empirical averages.
An AI system that generates a confident recommendation is a distinct case. Its prose may influence a human evaluator's confidence, but establishing the system's calibration requires outcomes measured on representative inputs with the system's defined scores. Establishing the human's overconfidence requires eliciting the human's own judgments and comparing them with appropriate criteria. The outputs of these two measurements should not be collapsed.
Frequently Asked Questions
Is overconfidence bias the same as being confident?
No. Confidence is a subjective judgment or attitude. Overconfidence is a measured discrepancy between that judgment and an appropriate standard. High confidence can be warranted when performance is highly accurate. Researchers must observe actual outcomes, relative scores, or confidence-interval coverage to determine whether the reported level is excessive.
What are the three types of overconfidence?
Overestimation means rating one's absolute performance too highly. Overplacement means rating one's position relative to others too highly. Overprecision means assigning too much certainty or intervals that are too narrow for the observed range of outcomes. These outcomes may coexist, but they can move in opposite directions under different task conditions. Moore and Healy (2008).
How do researchers know whether confidence is calibrated?
They compare recorded probabilities with independently resolved outcomes across enough comparable judgments. If events assigned 70% probability occur approximately 70% of the time over a sufficiently large and relevant set, that subset is calibrated. Reliability diagrams, calibration error measures, and proper scoring rules provide complementary views. Meaningful uncertainty intervals are needed around sample estimates.
Can someone be accurate and overconfident at the same time?
Yes. Accuracy and calibration refer to different properties. A person who answers 90% of questions correctly but reports 99% confidence on each is highly accurate and overconfident in the studied set. Another person can have low accuracy yet appropriately low confidence. Studies should assess both outcome performance and metacognitive judgment.
Does the Dunning-Kruger effect mean that intelligent people are always uncertain?
No. The original experiments concerned discrepancies in self-assessment and measured performance on particular tasks. Later research identifies important roles for task difficulty, comparison groups, and statistical properties of ratings. Skilled people can be well calibrated, overconfident, or underconfident depending on the judgment context.
Are overconfident people always worse decision-makers?
No general conclusion follows. Someone may have strong skills and slightly overstated confidence, while another person may be calibrated but poorly informed. Decision quality also depends on consequences, available options, and the cost of mistakes. Calibration, discrimination, first-order accuracy, and choice outcomes should be evaluated separately.
Does expertise eliminate overconfidence?
Expertise may improve both performance and monitoring in familiar, feedback-rich environments. It does not ensure accurate confidence on rare, novel, poorly defined, or rapidly changing tasks. The strongest inference is task-specific and evidence-based: a professional's actual accuracy and confidence should be measured in the setting that matters.
Can practice and feedback reduce overconfidence?
Yes, in some studied tasks. Individualized performance feedback can help people adjust extreme forecasts and improve calibration, although gains are not equally strong in every experiment. Reliable transfer to unrelated high-stakes judgments should be tested rather than presumed. Stone and colleagues (2023).
Is overconfidence a mental disorder?
Overconfidence bias is a construct in the psychology of judgment and decision-making, not a clinical disorder. Some clinical conditions involve changes in insight, grandiosity, or risk taking, but a laboratory confidence gap does not establish any diagnosis. Clinical assessment relies on criteria, context, impairment, and professional evaluation.
How should I respond when someone sounds extremely certain?
Ask what outcome would test the claim, what evidence supports it, whether other explanations have been considered, and whether past predictions with similar confidence have been correct. If a decision is consequential, request an explicit probability or uncertainty range and an independent check. Certainty of expression can communicate conviction without measuring predictive reliability.
Conclusion
Overconfidence bias is best understood as a family of measurable discrepancies between confidence and specified standards of performance or uncertainty. Overestimation, overplacement, and overprecision involve distinct questions, and each can vary across tasks, people, domains, and experimental procedures. The central scientific achievement is the ability to test confidence against outcomes rather than infer competence from assurance alone. Better decision practice joins accurate information, explicit uncertainty, representative comparison data, appropriate scoring, and feedback. These practices support calibrated judgment in ordinary life and in increasingly complex interactions between humans and decision-support technologies.
Related Articles
Cluster Cognitive Biases, Judgment and Decision-Making - The overarching scientific framework for human judgment, probability, heuristics, and decision theories.
Metacognitive Confidence vs. Accuracy: Why Certainty Can Mislead - Metacognitive sensitivity, confidence reporting, and monitoring the correctness of individual cognitive judgments.
Memory Confidence vs. Accuracy: What the Evidence Shows - Memory-specific confidence, retrieval conditions, recognition, and the limits of eyewitness certainty.
AI as Authority: Trust, Expertise, Automation Bias, and Human Decision-Making - Applied trust and reliance in human-AI interactions, with separate analysis of automation bias.
References
Bercovich, E., and Ackerman, R. (2026). Metacognitive processes in radiologic decision-making: The hard-easy effect and temporal thinking dynamics. Radiology, 319(2), e252834. https://doi.org/10.1148/radiol.252834
Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Monthly Weather Review, 78(1), 1-3. https://doi.org/10.1175/1520-0493(1950)078%3C0001:VOFEIT%3E2.0.CO;2
Burson, K. A., Larrick, R. P., and Klayman, J. (2006). Skilled or unskilled, but still unaware of it: How perceptions of difficulty drive miscalibration in relative comparisons. Journal of Personality and Social Psychology, 90(1), 60-77. https://doi.org/10.1037/0022-3514.90.1.60
Fleming, S. M., and Lau, H. C. (2014). How to measure metacognition. Frontiers in Human Neuroscience, 8, 443. https://doi.org/10.3389/fnhum.2014.00443
Gervais, S., and Odean, T. (2001). Learning to be overconfident. Review of Financial Studies, 14(1), 1-27. https://doi.org/10.1093/rfs/14.1.1
Juslin, P., Winman, A., and Olsson, H. (2000). Naive empiricism and dogmatism in confidence research: A critical examination of the hard-easy effect. Psychological Review, 107(2), 384-396. https://doi.org/10.1037/0033-295X.107.2.384
Klayman, J., Soll, J. B., Gonzalez-Vallejo, C., and Barlas, S. (1999). Overconfidence: It depends on how, what, and whom you ask. Organizational Behavior and Human Decision Processes, 79(3), 216-247. https://doi.org/10.1006/obhd.1999.2847
Koriat, A. (2012). The self-consistency model of subjective confidence. Psychological Review, 119(1), 80-113. https://doi.org/10.1037/a0025648
Kruger, J., and Dunning, D. (1999). Unskilled and unaware of it: How difficulties in recognizing one's own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology, 77(6), 1121-1134. https://doi.org/10.1037/0022-3514.77.6.1121
Mellers, B., Stone, E., Murray, T., Minster, A., Rohrbaugh, N., Bishop, M., Chen, E., and colleagues. (2015). Identifying and cultivating superforecasters as a method of improving probabilistic predictions. Perspectives on Psychological Science, 10(3), 267-281. https://doi.org/10.1177/1745691615577794
Moore, D. A., Carter, A. B., and Yang, H. H. J. (2015). Wide of the mark: Evidence on the underlying causes of overprecision in judgment. Organizational Behavior and Human Decision Processes, 131, 110-120. https://doi.org/10.1016/j.obhdp.2015.09.003
Moore, D. A., and Healy, P. J. (2008). The trouble with overconfidence. Psychological Review, 115(2), 502-517. https://doi.org/10.1037/0033-295X.115.2.502
Rahnev, D., and colleagues. (2022). Across-subject correlation between confidence and accuracy: A meta-analysis of the Confidence Database. Attention, Perception, and Psychophysics. https://pubmed.ncbi.nlm.nih.gov/35129781/
Soll, J. B., and Klayman, J. (2004). Overconfidence in interval estimates. Journal of Experimental Psychology: Learning, Memory, and Cognition, 30(2), 299-314. https://doi.org/10.1037/0278-7393.30.2.299
Stone, E. R., Luu, J., Costello, C. K., and Somerville, A. H. (2023). Automated calibration training for forecasters. Journal of Behavioral Decision Making, 36(4), e2334. https://doi.org/10.1002/bdm.2334