The Overconfidence Effect: Why Experts Are Wrong More Often Than They Think.
The gap between how certain experts feel and how often they are right is not a personality flaw. It is a measurable, neurally encoded failure mode that appears in every high-stakes professional domain tested, and the overconfidence bias science shows it can be reduced. Here is what the science actually says, and what to do with it.
01The 30-Point Gap
Experts at 98% Confidence Are Right 68% of the Time
Ask yourself a factual question: something you feel confident about. A date, a number, a prediction about next quarter's results. Now assign a probability: how certain are you? If you said 98%, Lichtenstein, Fischhoff, and Phillips have a number for you. Across more than 15,000 probability judgments aggregated from over twenty studies, people who stated 98% confidence were correct 68% of the time.[8] That is not a rounding error. That is a 30-percentage-point calibration gap: the systematic distance between how certain humans feel and how often they are actually right.[7]
The finding did not stay in the laboratory. Svenson's canonical 1981 study asked American drivers to rate their own safety: 93% placed themselves above the median.[9] The mathematics are unyielding (half of all drivers are, by definition, below median), but the overconfidence bias does not answer to mathematics. It answers to something deeper in the architecture of human judgment.
That matters because this is not a story about amateurs misjudging trivia questions. The overconfidence effect scales with expertise, stakes, and consequence. Tetlock spent twenty years tracking 284 political and economic experts (professors, government advisors, think-tank analysts) as they made 28,000 predictions about world events. The result: they barely outperformed random chance.[4] The experts who appeared most frequently in media, the ones who spoke with the greatest certainty, performed worst of all.[5]
The scale of the problem becomes clear only when you see it across domains. In medicine, physicians claiming complete certainty about a diagnosis were wrong approximately 40% of the time in autopsy-comparison studies (a figure that applies to certainty-claiming diagnoses specifically, not to clinical judgment generally, since calibration improves in domains with rapid diagnostic feedback).[22] In finance, CFOs asked to provide 90% confidence intervals for the S&P 500's annual return captured the true value only 36.3% of the time.[51] In national security, a 2025 study of approximately 1,900 NATO-cleared officials found they were correct about 58% of the time on forecasts they rated at 90% confidence, a 32-point gap that mirrors the laboratory findings from four decades earlier.[31]
These are not isolated failures. Grežo's 2021 meta-analysis of 34 studies confirmed that overconfidence exerts a statistically significant effect on financial decision-making across investment, trading, and innovation contexts.[28]
The pattern is robust enough to name. Overconfidence bias science (the study of why humans systematically overestimate their knowledge, overrate their relative standing, and set confidence intervals too narrow) is one of the most replicated research programmes in behavioural science.[1][6] The question is no longer whether the effect is real. The question is why the brain generates it, and what can be done about it.
02The Mechanism
Three Types of Overconfidence and the Neural Architecture That Produces Them
For forty years, the overconfidence literature produced apparently contradictory results. Some studies found people wildly overestimated their abilities. Others found they underestimated themselves on easy tasks. The confusion persisted until Moore and Healy published their definitive 2008 taxonomy in Psychological Review, which demonstrated that researchers had been conflating three mechanistically distinct phenomena under a single label.[6]
The first is overestimation, believing your absolute performance is better than it actually is. The second is overplacement, believing you rank higher relative to others than you do, the engine behind Svenson's 93% above-median drivers.[9] The third, and most dangerous, is overprecision: the excessive certainty that your beliefs are correct, expressed as confidence intervals that are far too narrow.[6] Moore and Healy showed that on easy tasks, people underestimate their absolute performance but overplace themselves against peers. On hard tasks, the pattern reverses. The three types respond differently to task difficulty, which is why decades of research that treated them as one thing generated decades of confusion.
The critical insight is that overprecision is the most persistent of the three forms. It appears regardless of task difficulty, population, or domain.[6][23] Sanchez and Dunning's 2023 interdisciplinary review confirmed that experts are uniformly overprecise regardless of expertise level; expert status does not reduce and may actually exacerbate excessive certainty.[23]
The reward architecture of overconfidence: the brain treats certainty as a reward signal in the striatum, while the vmPFC and rACC selectively amplify confirming evidence, leaving the lateral PFC to encode disconfirmation weakly, hardwiring a systematic gap between felt certainty and actual accuracy.
Diagram · HPC
The neural evidence explains why. Molenberghs and colleagues conducted the largest fMRI study on confidence and accuracy ever published (308 participants, exceptional for neuroimaging) and found that the brain processes certainty and correctness through structurally separate systems.[15] Higher confidence activated the striatum and hippocampus, regions associated with reward processing. Higher metacognitive accuracy (actually knowing when you are right and when you are wrong) correlated with decreased activation in the anterior medial prefrontal cortex.[15]
That matters because the implication is stark: the feeling of being certain and the reality of being correct are generated by opposing neural systems. Confidence feels like a reward. The brain treats it as one. Fleming and Dolan's review of the neural basis of metacognitive ability confirmed the dissociation: the lateral prefrontal cortex governs retrospective accuracy judgments, while medial structures govern the prospective feeling of knowing.[16]
The architecture gets worse under the influence of asymmetric belief updating. Sharot, Korn, and Dolan demonstrated that participants selectively updated their beliefs in response to better-than-expected information but failed to update appropriately after worse-than-expected news.[17] The ventromedial prefrontal cortex and rostral anterior cingulate cortex show heightened activation for confirming evidence, while the lateral PFC codes negative prediction errors weakly.[17][18] Approximately 80% of people demonstrate this optimism bias.[18]
03Evidence
The Five Strongest Studies on the Overconfidence Effect
01The claim
The single load-bearing finding
The hero study finds 28,000+ predictions.
Not all evidence carries equal weight. A twenty-year prospective longitudinal study tells you something different from a single-session lab experiment with thirty undergraduates. The five studies ranked below represent the strongest available evidence on overconfidence, chosen for design quality, measurement precision, causal clarity, and replication value. Together, they establish that the calibration gap is not an artefact of how researchers ask questions. It is a stable property of how human minds process uncertainty.
Pooled estimate
28,000+ predictions
02How we measured
Grading the calibration studies
Studies scored on design, sample, rigour, causality, replication, citations.
For overconfidence science, ecological validity is decisive: laboratory calibration gaps mean little unless they replicate in real expert populations making real consequential forecasts over years.
Rubric weights
03The spread
Heterogeneity across 5 studies
Methodological quality across the ranked studies.
The hierarchy reveals a pattern that is easy to miss when reading individual studies: the evidence gets stronger, not weaker, as you move from the laboratory to the field. Tetlock's experts were not performing contrived tasks under artificial conditions. They were making real predictions about real events (elections, economic shifts, military conflicts) in their areas of professional specialisation.[4] The TNSR 2025 study extended this to approximately 1,900 NATO-cleared national security officials making 60,000-plus assessments about geopolitical events.[31] The gap held.
Rubric spread
91 → 70 /100
Highest to lowest rubric score across the ranked studies.
04What does not hold
Negative knowledge
What the evidence base does not support.
One important nuance deserves emphasis. Gigerenzer's ecological rationality framework challenges the blanket framing of heuristics as error-prone. Simple heuristics, he argues, can outperform optimisation under specific ecological conditions.[3] This is not a refutation of overconfidence research. It is a boundary condition. Weather forecasters approach calibration because they operate in environments with rapid, unambiguous feedback.[8] The problem is not that heuristics always fail.
5 trials. One pooled answer.
Below: the anchor study in full; then the forest plot at scale; then the supporting trials in ranked order.
01Anchor
Expert Political Judgment: How Good Is It? How Can We Know?
Philip Tetlock spent two decades collecting predictions from 284 political and economic experts (university professors, government advisors, think-tank analysts). He verified every prediction against actual outcomes.
Rubric breakdown
The strongest studies, ranked by methodological weight.
Each scored 0–100 against a six-criterion rubric, tagged by design and year; the anchor leads.
02
Calibration of Probabilities: The State of the Art to 1980
When subjects stated 98% confidence, accuracy was 68%, a 30-point calibration gap replicated across populations, domains, and decades. Calibration only approached accuracy in rapidly-feedback domains like experienced weather forecasting.
85/100
03
The Trouble with Overconfidence
Three distinct forms of overconfidence (overestimation, overplacement, and overprecision) respond differently to task difficulty. Overprecision is the most persistent, present regardless of context.
79/100
04
Neural correlates of metacognitive ability and of feeling confident
Higher confidence correlated with striatal/hippocampal (reward) activation; higher metacognitive accuracy correlated with decreased anterior medial prefrontal activation. The feeling of being right and the reality of being right recruit opposing neural systems.
74/100
05
Psychological Strategies for Winning a Geopolitical Forecasting Tournament
The CHAMPS KNOW probabilistic reasoning training produced reliable accuracy gains (6–11% Brier score improvement) sustained across multiple tournament years. Comparison class use was the single most powerful component.
70/100
04Stakes
The Cost of Misplaced Certainty Across Four Domains
Overconfidence is not an abstract laboratory finding. It has a measurable cost in money, health, security, and democratic function, and the price scales with the confidence of the decision-maker.
Financial
Consistent with an overconfidence account, male investors traded 45% more than female investors and earned risk-adjusted returns 1.4 percentage points lower per year (overconfidence inferred from gender-linked trading patterns).[24] Overconfident CEOs were 65% more likely to make acquisitions, and markets punished them: −90 basis points at announcement versus −12 for rational CEOs.[27] The most active traders in Odean's study earned 11.4% annually versus a market average of 16.4%.[25]
portfolio churn, impulsive trades, chronic underperformance vs. index
Medical
Physicians claiming complete certainty were wrong approximately 40% of the time in autopsy-comparison studies (a figure specific to certainty-claiming diagnoses, not clinical judgment generally).[22] In Senegal, overconfident healthcare providers were 26% less likely to correctly manage patients and performed 18% fewer diagnostic actions.[36] High-confidence residents spent 12% less time per case with no accuracy advantage. Certainty compressed effort, not error.[37]
premature diagnostic closure, skipped differentials, unexamined confidence
Strategic / Geopolitical
Overconfident leaders systematically overestimate capabilities and underestimate adversaries, a pattern documented across the First World War, Vietnam, and the Cuban Missile Crisis.[29] The 2025 TNSR study confirmed the pattern in active national security professionals: the calibration gap at 90% confidence was 32 points.[31] The same bias that made junior analysts overconfident in cold reading tasks operates at the level of strategic military planning.
intelligence failures, planning overruns, escalation through misread signals
Information / Democratic
Three in four Americans overestimate their news discernment by 22 percentile points, and this overconfidence, not actual news literacy, predicts willingness to share false political content.[32] Low health-literacy individuals with high confidence showed 62% tobacco use versus 29% in the high-literacy group.[33] The Dunning-Kruger effect pattern means the people most confident in their media literacy are often the least equipped to exercise it.
uncritical content sharing, resistance to correction, false sense of expertise
05Protocol
A 4-Step Calibration Protocol for High-Stakes Decision-Making
Every step does the same thing at a mechanistic level: it reintroduces the feedback signal that professional environments typically remove. Overconfidence is the predictable output of a brain designed for rapid-feedback environments operating in a slow-feedback world.
The protocol, as a sequence.
Ongoing → Pre-decision → Estimation → Training
Prediction Tracking
Maintain a simple prediction log: write down what you expect, your confidence as a percentage, and your reasoning, then track actual outcomes.
Overconfidence persists because feedback is absent in most professional domains. The only populations who achieve naturalistic calibration operate in rapid-feedback environments.[8] Tracking artificially closes the loop. Chang et al. found comparison class use combined with tracking was the most powerful component of the CHAMPS KNOW protocol.[39]
Tracking in vague terms ("I thought this would go well") defeats the purpose. Confidence must be quantified as a percentage before the outcome is known. Minimum 20–30 tracked predictions before calibration patterns become visible.[45]
Premortem
Before finalising any major plan, assume it has already failed completely. Write down every plausible reason it failed.
Veinott, Klein, and Wiggins found the premortem technique produced a 25-point reduction in plan confidence, outperforming standard critique, pro/con listing, and cons-only generation in a five-condition controlled experiment.[42]
Running the premortem after commitment, when social pressure makes generated concerns feel disloyal rather than informative.
Reference Class Forecasting
When estimating costs, timelines, or success rates, find a reference class of comparable past projects and anchor to the empirical median before adjusting.
Flyvbjerg validated reference class forecasting on 258 large infrastructure projects across 20 nations; it systematically reduces the planning fallacy.[43] Kahneman and Lovallo showed the "inside view" is the primary driver of planning overconfidence; the corrective is the outside view.[46]
Treating your current project as so unique that no reference class applies. The literature suggests this is almost always the inside view in disguise.
Structured Debiasing
Complete a structured probabilistic reasoning training covering base rates, Bayesian updating, comparison classes, disconfirming evidence, and aggregated views.
Morewedge's single-session debiasing training produced at least 31.94% bias reduction in the game condition, with effects persisting at two months.[40] Sellier et al. found trained participants were 19% less likely to choose the inferior solution on a real business case (the corrected figure per a published corrigendum that revised the original 29% estimate downward).[41]
Believing domain expertise eliminates the need for calibration training. Sanchez and Dunning confirm experts are uniformly overprecise regardless of domain.[23] A 2025 meta-analysis of 54 RCTs (N = 10,941) found small but significant debiasing effects (Hedges' g = 0.26).[49]
06Verdict
The verdict.
Bottom line
The brain will not stop generating certainty. The only question is whether you build the systems that test it before you act on it.
The overconfidence effect is not a pop-psychology curiosity. It is a precisely measured, neurally grounded, and universally replicated property of human cognition: a 30-point gap between how certain people feel and how often they are right, present in every high-stakes domain from medicine to military intelligence. The gap does not close with expertise, experience, or education. It closes only when the environment provides calibration feedback (rapid, unambiguous, and unavoidable) or when the individual builds the external structures that substitute for it. The science is clear: certainty is a feeling, not a signal, and treating it as a signal is the single most expensive cognitive error in professional life.
The reframing this evidence demands is simple but uncomfortable. Every professional who has ever said "I'm 90% sure" about a strategic judgment was probably operating with something closer to 58% accuracy, with no internal mechanism for detecting the difference.[31] That is not a failure of character. It is a design specification. The brain was built for environments where confidence was tested by consequences within hours, not years. The modern professional world removed the consequences and left the confidence.
The practical implication is equally direct. Calibration is a skill, not a trait. The evidence from Mellers' forecasting tournaments,[38] Morewedge's debiasing interventions,[40] and Sellier's field-transfer studies[41] converges on the same conclusion: structured training in probabilistic reasoning produces measurable improvement, and the improvement transfers to real-world decisions. The solution is not to feel less confident. It is to build the feedback loops, the premortems, the reference classes, and the prediction-tracking systems that translate the feeling of confidence into a testable, falsifiable commitment.
What Tetlock called epistemic humility is not a personality orientation. It is an engineering specification: a set of external structures that compensate for a known limitation in the brain's confidence architecture. The question was never whether we are overconfident. The question is what we build to correct for it.
Stated Certainty vs Actual Accuracy
The gap is systematic
The 30-point calibration gap is not noise, bad luck, or poor education. It is a stable architectural feature of human cognition, present in every population and domain measured across fifty years of research, from trivia questions to military intelligence.
Certainty suppresses correction
Overconfidence does not just produce wrong answers. It produces wrong answers that feel right and disable the error-detection systems that would catch them. Physicians order fewer tests. Investors check fewer sources. Planners skip the reference class.
Calibration is trainable
The IARPA forecasting tournaments, single-session debiasing interventions, and field-transfer studies converge: structured probabilistic reasoning training produces measurable, durable, and transferable accuracy gains. The correction is external scaffolding, not willpower.
Put it to work
Where this science goes next on HPC
07Bibliography
The bibliography.
-
01
Journal
doi: 10.1126/science.185.4157.1124
Judgment under uncertainty: Heuristics and biases
-
03
Book
Gut feelings: The intelligence of the unconscious
-
04
Book
Expert political judgment: How good is it? How can we know?
-
05
Book
Superforecasting: The art and science of prediction
-
06
Review
doi: 10.1037/0033-295X.115.2.502
The trouble with overconfidence
-
07
Journal
doi: 10.1037/0096-1523.3.4.552
Knowing with certainty: The appropriateness of extreme confidence
-
08
Chapter
doi: 10.1017/CBO9780511809477.019
Calibration of probabilities: The state of the art to 1980. In D. Kahneman, P. Slovic, & A. Tversky (Eds.), Judgment under uncertainty: Heuristics and biases (pp. 306–334). Cambridge University Press
-
09
Journal
doi: 10.1016/0001-6918(81)90005-6
Are we all less risky and more skillful than our fellow drivers? Acta Psychologica, 47(2), 143–148
-
15
Journal
doi: 10.1093/scan/nsw093
Neural correlates of metacognitive ability and of feeling confident: A large-scale fMRI study
-
16
Journal
doi: 10.1098/rstb.2011.0417
The neural basis of metacognitive ability
-
17
Journal
doi: 10.1038/nn.2949
How unrealistic optimism is maintained in the face of reality
-
18
Journal
doi: 10.1016/j.cub.2011.10.030
The optimism bias
-
22
Journal
doi: 10.1016/j.amjmed.2008.01.001
Overconfidence as a cause of diagnostic error in medicine
-
23
Review
Are experts overconfident? An interdisciplinary review
-
24
Journal
doi: 10.1162/003355301556400
Boys will be boys: Gender, overconfidence, and common stock investment
-
25
Journal
doi: 10.1111/0022-1082.00078
Volume, volatility, price, and profit when all traders are above average
-
27
Journal
doi: 10.1016/j.jfineco.2007.07.002
Who makes acquisitions? CEO overconfidence and the market's reaction
-
28
Meta
doi: 10.1108/RBF-01-2020-0020
Overconfidence and financial decision-making: A meta-analysis
-
29
Book
doi: 10.4159/9780674039162
Overconfidence and war: The havoc and glory of positive illusions
-
31
Review
source
Texas National Security Review / Good Judgment Project
-
32
Journal
doi: 10.1073/pnas.2019527118
Overconfidence in news judgments is associated with false news susceptibility
-
33
Journal
doi: 10.1007/s10880-022-09895-4
Overconfidence in managing health concerns: The Dunning-Kruger effect and health literacy
-
36
Journal
doi: 10.1016/j.joep.2019.102213
Overconfident health workers provide lower quality healthcare
-
37
Journal
doi: 10.2147/AMEP.S442689
Overconfidence, time-on-task, and medical errors: Is there a relationship? Advances in Medical Education and Practice, 15, 133–140
-
38
Journal
doi: 10.1177/0956797614524255
Psychological strategies for winning a geopolitical forecasting tournament
-
39
Journal
Developing expert political judgment: The impact of training and practice on judgmental accuracy in geopolitical forecasting tournaments
-
40
Journal
doi: 10.1177/2372732215600886
Debiasing decisions: Improved decision making with a single training intervention
-
41
Journal
doi: 10.1177/0956797619861429
Debiasing training improves decision making in the field
-
42
Journal
Evaluating the effectiveness of the PreMortem technique on plan confidence
-
43
Journal
doi: 10.1177/875697280603700302
From Nobel Prize to project management: Getting risks right
-
45
Chapter
Debiasing. In D. J. Koehler & N. Harvey (Eds.), Blackwell handbook of judgment and decision making (pp. 316–337). Blackwell
-
46
Journal
doi: 10.1287/mnsc.39.1.17
Timid choices and bold forecasts: A cognitive perspective on risk taking
-
47
Journal
doi: 10.1002/acp.4236
The effect of calibration training on the calibration of intelligence analysts' judgments
-
49
Journal
doi: 10.1038/s41598-025-28794-w
Debiasing training reduces confirmation bias in national risk analysts
-
51
Review
doi: 10.3389/fpsyg.2021.802439
The impact of cognitive biases on professionals' decision-making: A review of four occupational areas
No entries match the current filter and search.
THE HAND-OFF — Performance Scan: find your limiting factor. Three minutes, 24 questions, six systems — and your answers never leave your device.
