The Dunning-Kruger Experiment: What the 1999 Study Actually Found
Photo: Ann H / Pexels
In 1999, two Cornell psychologists published a paper that became one of the most cited and most misunderstood findings in modern psychology. The Dunning-Kruger effect — the idea that incompetent people lack the metacognitive skill to recognize their own incompetence — is now deployed in everyday conversation to explain overconfident politicians, bad drivers, and anyone who argues confidently about things they do not know. The problem is that the popular version of the effect, complete with the notorious "peak of Mount Stupid" confidence curve, does not accurately describe what Kruger and Dunning actually found. Here is what the original study did and did not show.
What Kruger and Dunning Were Investigating
David Dunning was a social psychologist at Cornell University studying how people assess their own abilities. His student Justin Kruger was working on a related puzzle: the observation that people who are bad at something are often unaware that they are bad at it, while people who are good at something tend to underestimate their performance.
The theoretical backbone came from earlier work on metacognition — thinking about thinking. If competence at a skill requires the same cognitive machinery that is needed to evaluate performance at that skill, then incompetent people face a "double burden": they produce poor work and, using the same underdeveloped skills to assess that work, cannot detect how poor it is. A person who cannot write grammatically correct sentences may also lack the grammatical intuition needed to recognize their errors.
Kruger and Dunning called this the dual burden of incompetence and set out to test it experimentally.
Source: Kruger, J., & 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.
The Setup
The study comprised four separate experiments, all conducted with Cornell undergraduates. Three domains were tested: logical reasoning, English grammar, and the ability to assess whether jokes were funny (humor judgment). These three domains were chosen deliberately: each has a reasonably objective correct answer (logical validity, grammatical correctness), or a consensus-based answer (shared sense of humor), and each can be assessed with standardized tests.
Study 1 used logical reasoning. Participants completed a 20-item logical reasoning test adapted from the LSAT (Law School Admission Test). After completing the test, they estimated:
- Their score on the test (how many items they got right out of 20)
- Their general logical reasoning ability compared to other Cornell students (what percentile they fell in)
- Their percentile rank on this specific test compared to other participants
Study 2 used English grammar. The same basic design was used with a 20-item grammar test. Participants answered the questions, then made the same three types of self-assessment.
Study 3 used humor judgment. Participants rated the funniness of 30 jokes on a scale. Their ratings were compared against a criterion of "funnier" based on a separate sample of professional comedians' assessments. Participants then assessed their own ability to identify what most people find funny.
Study 4 used logical reasoning again, with an added manipulation: after completing the test, half the participants received training in logical reasoning skills before making their self-assessments. This was designed to test whether improving competence would simultaneously improve self-assessment accuracy — the metacognitive link.
The Protocol
The mechanics of each study were similar. Participants took the relevant test under standard conditions. They then provided self-assessments on the dimensions described above. A week or more later (in some studies), they viewed their own test alongside tests from other participants and were asked to re-evaluate their standing.
The critical analysis divided participants into quartiles based on their actual test performance: bottom quartile (roughly the bottom 25% of scorers), second quartile, third quartile, and top quartile. For each quartile, the researchers calculated the mean actual score and the mean estimated score. The gap between those two numbers — the discrepancy between how well you thought you did and how well you actually did — was the key dependent variable.
Source: Unskilled and Unaware — But Why? A Reply to Krueger and Mueller — University of Michigan
The Results
The results, across multiple studies and domains, showed a consistent pattern.
Bottom-quartile participants substantially overestimated their performance. On the logical reasoning test, participants who scored at roughly the 12th percentile on average estimated they had performed at about the 62nd percentile. The gap was approximately 50 percentile points. This was the headline finding: people who performed the worst showed the largest disconnect between their self-perception and reality.
Top-quartile participants underestimated their performance. This received less attention in popular coverage, but it is equally part of the original finding. Participants who scored near the top of the distribution estimated their performance at roughly the 70th percentile when they were actually at the 86th. They were off, but in the opposite direction — they underestimated.
Study 4's training manipulation worked. Participants who received logical reasoning training before their final self-assessment showed improved self-assessment accuracy compared to untrained controls. This was direct evidence for the metacognitive link: improving the skill also improved the ability to recognize skill level. This is the "double burden" in reverse — developing competence simultaneously develops the ability to evaluate competence.
Source: Kruger, J., & Dunning, D. (1999). Unskilled and unaware of it. JPSP, 77(6), 1121–1134.
The Timeline
| Year | Event |
|---|---|
| 1999 | Kruger & Dunning publish "Unskilled and Unaware of It" in JPSP |
| 2002 | Krueger & Mueller challenge: effect may be explained by regression to the mean |
| 2002 | Kruger & Dunning publish direct reply defending their interpretation |
| 2011 | Popularization accelerates; "Mount Stupid" visualization circulates online (not in the paper) |
| 2016–2017 | Nuhfer et al. simulations: random data can produce similar-looking graphs |
| 2020 | Gignac & Zajenkowski publish statistical artifact argument in Intelligence |
| 2023 | Dunkel, Nedelec & van der Linden partially replicate Gignac findings with smaller effect |
The Popular Misinterpretation
Here is where the story becomes interesting — and important for understanding what science actually produces versus what circulates as its output.
The "Mount Stupid" graph, showing a confidence curve that peaks sharply at low competence ("peak of Mount Stupid"), plummets through a "valley of despair" as learning progresses, rises again as genuine expertise develops, and levels off at the "plateau of sustainability" — this graph does not appear anywhere in Kruger and Dunning's 1999 paper.
The paper's figure shows two lines: mean actual test score and mean estimated score, plotted across the four performance quartiles. Both lines rise together — as actual performance increases, estimated performance also increases. The gap between them is largest at the bottom quartile (where overestimation is greatest) and largely closes by the top quartile (where underestimation occurs). There is no hump. There is no peak followed by a valley. There is no learning-stage progression. The x-axis is actual performance quartile, not a timeline of learning.
The "Mount Stupid" visualization maps the famous effect onto a learning curve — stages of acquiring a skill over time — which is a completely different thing from what Kruger and Dunning measured. They measured self-assessment accuracy across a cross-section of people at different skill levels at one point in time. Whether confidence follows a non-monotonic path as an individual learns a skill over months or years is a separate empirical question that the 1999 paper does not address.
David Dunning himself has acknowledged this confusion in subsequent interviews, noting that the popular graphic overstates and misrepresents the original findings.
The Statistical Artifact Debate
The more technically serious critique came from statistics, not from misattribution.
Regression to the mean is a statistical phenomenon that happens whenever two imperfectly correlated variables are measured. If you measure actual test performance and estimated performance, and these two variables are correlated but not perfectly correlated, then low scorers will, on average, estimate higher than their actual score and high scorers will estimate lower — purely as a statistical consequence, without any metacognitive mechanism required. The pattern in Kruger and Dunning's results looks exactly like what regression to the mean would produce.
Nuhfer and colleagues (2016, 2017) ran simulations in which they generated random data — numbers with no real relationship to each other — and showed that plotting these simulated scores using Kruger and Dunning's method produced graphs that looked visually similar to the original results. The implication: the quartile-based plotting method itself may be generating the pattern.
Source: Nuhfer, E., et al. (2017). How Random Noise and a Graphical Convention Subverted Behavioral Scientists' Explanations of Self-Assessment Data. Numeracy, 10(1).
Gignac and Zajenkowski published the strongest statistical argument in 2020 in the journal Intelligence. They argued that when you test the Dunning-Kruger hypothesis using continuous-variable methods — rather than collapsing data into quartiles — the effect shrinks substantially and may not survive. Using a dataset of self-assessed versus measured IQ, they found the correlation between overconfidence and low ability was much weaker than the quartile plots suggested, and argued the quartile method introduces artifact.
Source: Gignac, G. E., & Zajenkowski, M. (2020). The Dunning-Kruger effect is (mostly) a statistical artefact: Valid approaches to testing the hypothesis with individual differences data. Intelligence, 80, 101449.
Source: The Dunning-Kruger effect is (mostly) a statistical artefact — ResearchGate
Dunning and colleagues have pushed back. A 2002 response paper directly addressed the regression-to-the-mean objection, running nine additional studies specifically designed to test whether the effect survives controls for regression. Their analysis suggested it does, though the debate has not been settled to everyone's satisfaction.
Dunkel, Nedelec, and van der Linden (2023) partially replicated Gignac and Zajenkowski's findings but found a small but statistically significant Dunning-Kruger pattern even using continuous methods — suggesting the effect is real but weaker than quartile plots imply. The current best reading is that the effect exists but the quartile visualization inflates its apparent magnitude.
Replication and Reanalysis
The overestimation of performance by low scorers has replicated consistently across domains and laboratories. It appears in logical reasoning, grammar, humor, medical knowledge, financial literacy, and scientific reasoning tasks. The bottom-quartile overestimation is one of the more robust findings in the metacognition literature.
The top-quartile underestimation is equally robust but less discussed. High performers consistently underestimate their relative standing — often because they assume that if a task was easy for them, it must have been easy for everyone. This "false consensus" effect among the skilled is as important as the overconfidence among the unskilled.
What the research does not robustly establish is the specific cognitive mechanism Kruger and Dunning proposed: the direct metacognitive link, where incompetence causes both poor performance and impaired self-assessment. An alternative explanation is that the pattern arises from the better-than-average effect combined with statistical regression — phenomena that do not require a specific metacognitive failure to explain. The data are consistent with both stories.
Source: How unaware are the unskilled? — Social Vision Lab, UCLA
Where It Stands Today
The Dunning-Kruger effect, correctly stated, is this: across studies and domains, people scoring in the bottom quartile of a test overestimate their performance by a substantial margin, and they do so more than high performers do. That pattern is reliable. The mechanism behind it — pure metacognitive failure, statistical artifact, or some combination — remains contested.
What the effect does not show, despite its cultural currency as a diagnosis: it is not a reliable predictor of an individual's overconfidence. The effect is a group-level statistical pattern. Knowing someone is in the bottom quartile of some skill does not allow you to reliably conclude that they are unaware of their low performance — some low performers know exactly how poorly they are doing.
The effect also does not create "smart" people who are uniquely self-aware. High performers underestimate their performance too, just in a different direction. The 1999 paper is equally a study of the underestimation of top performers as it is of the overestimation of low performers.
Perhaps the deepest irony in the Dunning-Kruger story is that the effect is most commonly invoked by people diagnosing it in others — which is exactly the kind of overconfident metacognitive claim the paper cautions against. Dunning has noted publicly that everyone is susceptible to this pattern in domains where they lack expertise, including the people confidently telling you about the Dunning-Kruger effect.
The Statistical Critique: Is Dunning-Kruger a Measurement Artifact?
The headline criticism of the Dunning-Kruger effect is not that it was measured in bad faith — it is that the method used to visualize it may create the pattern it purports to discover. Understanding why requires a short detour into measurement statistics.
Regression to the mean — the tendency for extreme scores on one measurement to be followed by less extreme scores on a second measurement — is a feature of any two imperfectly correlated variables. If actual test performance and estimated performance are correlated but not perfectly correlated (which they cannot be, since people do not know their scores in advance), then the mathematics guarantee a predictable asymmetry: low scorers will estimate higher than they scored, and high scorers will estimate lower. This is not because of any cognitive quirk. It is what the equations require. The question is whether the Dunning-Kruger pattern exceeds what regression alone would predict.
Krueger and Mueller raised this point directly in a 2002 commentary, arguing that the overestimation of low performers and the underestimation of high performers were largely recoverable from the statistical relationship between the two variables, without requiring any metacognitive story. Kruger and Dunning responded in the same year with nine additional studies designed to control for this explanation, arguing the effect survived.
The critique sharpened considerably in 2016 and 2017 when Edward Nuhfer and colleagues ran simulation studies. They generated datasets of random numbers — scores with no meaningful relationship to each other — and plotted them using Kruger and Dunning's quartile method. The resulting graphs bore a close visual resemblance to the original figures: low-quartile "scorers" appeared to overestimate, high-quartile "scorers" appeared to underestimate. If noise data produces the same picture, the picture is not strong evidence of a real phenomenon.
Source: Nuhfer, E., et al. (2017). "How Random Noise and a Graphical Convention Subverted Behavioral Scientists' Explanations of Self-Assessment Data." Numeracy, 10(1). https://doi.org/10.5038/1936-4660.10.1.4
Gignac and Zajenkowski (2020) published what became the most-cited version of this argument. Instead of using quartile grouping, they reanalyzed the relationship between self-assessed and measured intelligence using continuous correlation methods on a large dataset. Their finding: the correlation between the degree of overestimation and actual ability was far weaker than the quartile plots implied, and when they used their preferred statistical approach, the effect was not significant. They concluded that "the degree to which people mispredicted their objectively measured intelligence was roughly equal across the whole spectrum of objectively measured intelligence" — which is the opposite of the Dunning-Kruger prediction.
Source: Gignac, G. E., & Zajenkowski, M. (2020). "The Dunning-Kruger effect is (mostly) a statistical artefact: Valid approaches to testing the hypothesis with individual differences data." Intelligence, 80, 101449. https://doi.org/10.1016/j.intell.2020.101449
Psychologists have responded in two ways. First, methodologically: several labs redesigned studies to avoid quartile grouping, using calibration curves and continuous regression analyses instead. These studies find a smaller but often still significant effect — which suggests the quartile method does inflate the signal, but does not create it from nothing. Second, theoretically: researchers have pointed out that Gignac and Zajenkowski's recoding of self-assessed IQ onto a linear scale introduced its own analytic assumptions, and that when more standard scaling is used, a residual Dunning-Kruger pattern reappears.
Calibration — in psychometric terms, the degree to which a person's confidence in a judgment matches the actual accuracy of that judgment — is now the preferred framework for studying this class of effect. A person is well-calibrated if they are right roughly as often as their confidence predicts. Calibration studies find systematic overconfidence among low performers more reliably than the quartile-difference approach does, and they are less susceptible to the regression critique.
Dunkel, Nedelec, and van der Linden (2023) specifically replicated the Gignac-Zajenkowski analysis and found that while the effect size was small by the continuous method, it remained statistically significant across a large sample — and was not plausibly explained by measurement artifact alone. Their conclusion: the effect is real but the quartile visualization amplifies it by a factor of roughly three to five relative to what continuous analyses show.
Source: Dunkel, C. S., Nedelec, J. L., & van der Linden, D. (2023). "Reevaluating the Dunning-Kruger effect." Intelligence, 97. https://www.sciencedirect.com/science/article/abs/pii/S0160289623000132
The statistical debate does not invalidate the observation that low performers overestimate; it recalibrates how large the effect is and how strong the mechanism needs to be to explain it. The 50-percentile overestimation visible in the original quartile figures is almost certainly inflated. A smaller but genuine tendency toward overconfidence at low skill levels, consistent with both the metacognitive story and simpler explanations like the better-than-average effect, is where the current evidence converges.
FAQs
Q: What exactly did Kruger and Dunning find in the original 1999 study? A: They found that participants scoring in the bottom quartile of tests of logical reasoning, grammar, and humor judgment overestimated their performance by roughly 40–50 percentile points on average. Top-quartile participants underestimated their performance. Participants who received training in the relevant skill showed improved self-assessment accuracy, suggesting the same competence that improves performance also improves metacognitive accuracy.
Q: Where did the "Mount Stupid" confidence curve come from if it is not in the original paper? A: The "Mount Stupid" graph — showing confidence peaking at low competence, then dropping, then rising with expertise — was a later popularization that does not appear in the 1999 paper and maps the effect onto a learning timeline, which the original study did not measure. The paper's figures show quartile-grouped data from a single time point, not an individual's journey through stages of learning.
Q: Is the Dunning-Kruger effect just regression to the mean? A: That is the strongest critique and it has not been fully resolved. When researchers use continuous statistical methods rather than quartile grouping, the effect shrinks considerably, suggesting that part of what looks like the Dunning-Kruger pattern in quartile plots is a statistical artifact. However, some studies using continuous methods still find a small but significant overestimation effect among low performers, suggesting the phenomenon is real, even if its magnitude is smaller than early quartile-based analyses suggested.
Q: Does the Dunning-Kruger effect mean that stupid people do not know they are stupid? A: That is an oversimplification. The effect is about domain-specific skill and self-assessment accuracy, not general intelligence. In a specific domain where they lack skill, people tend to overestimate their competence. Experts in one domain are novices in others and show similar overestimation there. The effect is not a stable property of certain people; it is a property of certain conditions.
Q: Is the effect unique to American or Western participants? A: The pattern has been found across cultures, but with variation in magnitude. Some studies find that East Asian participants are more likely to underestimate performance across skill levels, which would complicate the effect's specific predictions. The overestimation of low performers appears to be a reasonably general phenomenon, but the specific shape of the self-assessment curve varies across cultural contexts.
Q: What is the practical takeaway if the mechanism is still debated? A: The practical finding is clear regardless of mechanism: in domains where you are a novice, your self-assessment of performance is likely to be less accurate than in domains where you have experience, and it is likely to be biased toward overestimation. Building feedback loops — testing yourself, seeking external evaluation, exposing yourself to actual performance data — is the primary corrective, which is what Kruger and Dunning's training manipulation demonstrated.
The Takeaway
The Dunning-Kruger effect is real in this specific sense: across multiple studies, low performers overestimate their abilities more than high performers do, and improving skill simultaneously improves the ability to self-assess. The controversial parts are the mechanism — whether it requires a specific metacognitive failure or emerges partly from statistical regularities — and the magnitude, which quartile-based visualizations may inflate. The famous confidence curve that peaks at low competence is a post-hoc internet visualization, not a finding from the original 1999 paper. What Kruger and Dunning actually found is narrower, more nuanced, and — because it applies to everyone in their areas of ignorance — more unsettling than the version that circulates online.
Read next:
Weekly digest
Stay ahead of the curve
Get the best science, AI, and tech stories delivered to your inbox every week. No spam — unsubscribe any time.
About the Author
SciHub101 is an independent science and technology publication founded and edited by Buğra Sözeri, a software engineer and app developer based in Türkiye. Articles are drafted against peer-reviewed literature and primary sources, fact-checked, and edited by a human before publication. Where we compare products, anything built by our own team is labelled as such — see the editorial guidelines and About page for how corrections and disclosures work.
Related Articles
The 1919 Eclipse Expedition: The Experiment That Made Einstein Famous
On May 29, 1919, two expeditions photographed a solar eclipse to test Einstein's general relativity. The results made Einstein a global celebrity — and sparked a controversy that took 90 years to resolve.
The Double-Slit Experiment: How Light Proved the Universe Is Weird
The double-slit experiment shows that particles behave like waves until you watch them. Thomas Young ran it in 1801 with light. Feynman called it the heart of quantum mechanics.
Negativity Bias and Memory: Why Bad Events Stick
Negativity bias makes bad memories stronger and stickier than good ones because threat detection, amygdala activity, and stress hormones prioritize survival.
Why Conspiracy Theories Spread Online
Conspiracy theories spread because they satisfy real psychological needs and because ranking algorithms reward the exact emotional signals conspiracies produce.
What If the Mandela Effect Proved Parallel Universes? The Science of Memory
People remember Nelson Mandela dying in prison, the Berenstein Bears (not Berenstain), and Pikachu's tail tip being black. The "Mandela effect" — but it's memory, not a multiverse glitch.
Complete Science Experiments Guide 2026 — Every Landmark Experiment by Field
Every landmark science experiment in one hub: Pasteur's swan-neck flask, Galvani's frog legs, Eratosthenes, Michelson–Morley, Hafele–Keating and more than 140 others, grouped by field and era.