The Dunning-Kruger Effect: What the Research Actually Found

By the IQ Revealed Team · Updated September 16, 2026

The Dunning-Kruger effect is the claim that people who lack skill at something also lack the skill needed to notice it, and so rate themselves far above where they actually perform. Kruger and Dunning proposed it in 1999. Since then the pattern in their graphs has been shown to be partly a statistical artefact, and the one large study that tested it on measured intelligence found essentially nothing. The idea is smaller than its reputation.

What the 1999 study actually did

Justin Kruger and David Dunning, then at Cornell, ran four studies on undergraduates. They tested humour (rating jokes against a panel of professional comedians), logical reasoning and English grammar, then asked each participant to estimate how they had done relative to their peers.

The headline finding is the one everyone quotes. Across the studies, participants whose scores fell in the bottom quarter placed themselves, on average, around the 62nd percentile while actually performing at about the 12th. In the humour study the bottom quartile put itself at the 58th percentile against an actual 12th — not only above where they landed, but above average.

Two details from the same paper are quoted far less often, and both matter.

Everybody overestimated, not just the weak. In the humour study the sample as a whole put its ability at the 66th percentile, sixteen points above the 50th that is true by definition.
Self-ratings were not random. Even in that first study, what people said about themselves correlated with how they actually scored at r = .39. People had some signal about their own ability; it was just weak.

The curve you have seen is not in the paper

If you picture the Dunning-Kruger effect, you probably picture a line that shoots up to a peak of misplaced confidence, plunges into a valley, and climbs slowly back — often labelled with names like Mount Stupid and the valley of despair.

That curve appears nowhere in the 1999 article. The paper's figures plot two lines across four quartiles of actual performance: what people scored, and where they thought they ranked. The perceived line does not spike and it does not dip. It runs roughly flat and high across all four groups while the actual-score line climbs steeply beneath it. The gap between the lines is the finding — the dramatic shape is a later illustration that has been attached to the research by other people.

This matters because the popular curve makes a claim the research never made: that a small amount of learning makes you more overconfident than knowing nothing at all. The quartile data does not show that.

The statistical objection

Within three years, Joachim Krueger and Ross Mueller argued that the pattern could be produced without any metacognitive deficit at all, from two ordinary statistical facts.

Mechanism	What it does to the graph
Regression toward the mean	Any two imperfectly correlated measures will pull extreme groups toward the middle. Sort people by test score and the worst performers' self-estimates will sit closer to average than their scores do — with no psychology involved.
The better-than-average effect	Most people rate themselves above the midpoint on most traits. Apply that to a group that is genuinely below the midpoint and a large gap appears automatically.

Edward Nuhfer and colleagues later pushed the point further by simulating self-assessment data from random numbers and showing that the classic quartile graph still emerges from noise. A picture that appears in data with no content in it cannot, by itself, be evidence of a mental process.

When it was tested on intelligence, it did not appear

The studies above are about grammar, humour and exam performance. For a site about IQ, the question is narrower: do people with lower measured intelligence misjudge their own intelligence more than everyone else does?

Gilles Gignac and Marcin Zajenkowski put that question to a direct test in 2020. They took 929 general-community participants, asked each to assess their own intelligence, and measured it with Raven's Advanced Progressive Matrices. Rather than drawing the quartile graph — which they argue is confounded by exactly the two mechanisms above — they used two tests that are not.

A Glejser test of heteroscedasticity: if the hypothesis holds, self-assessment error should be larger at the low end of ability than the high end. They found no statistically significant heteroscedasticity.
Nonlinear (quadratic) regression: if the hypothesis holds, the relationship should bend. They found the association between measured and self-assessed intelligence to be, in their words, “essentially entirely linear”.

Their conclusion was that while the phenomenon “may be to some degree plausible for some skills, the magnitude of the effect may be much smaller than reported previously”. On intelligence specifically, the strong version of the effect did not show up.

What still stands

Cutting the effect down is not the same as deleting it. Three things survive the criticism intact.

Self-assessment really is unreliable. That was never the contested part. Freund and Kasten pooled 154 effect sizes from 41 studies and estimated the correlation between self-estimated and psychometrically measured cognitive ability at r = .33. Squared, that is about a tenth of the variation between people. Your own guess about your intelligence carries some information and misses most of it.

The better-than-average effect is robust. It is one of the mechanisms used to explain away the Dunning-Kruger graph, which means it is itself well established. Most people place themselves above the middle, which is arithmetically impossible for most people to be right about.

The metacognitive argument is still reasonable. Kruger and Dunning's fourth study trained poor performers in logical reasoning and found that their self-assessments improved along with their skill — which is the kind of evidence regression to the mean does not explain. The mechanism may be real even if the quartile graph overstated it.

What this means for guessing your own IQ

Put the pieces together and the practical answer is unglamorous. You are probably not spectacularly wrong about your own ability in the way the popular version of the effect suggests, and you are also not close to right. The correlation of about .33 says a self-estimate lands in the right general region and misses by a wide margin often enough that it cannot be relied on in either direction.

That applies to underestimating as much as overestimating. The same research finds high performers misjudging their standing against others, which is the less-discussed half of the original paper.

The alternative to a self-estimate is a measurement. An IQ test samples reasoning, verbal and numerical problem-solving, working memory and processing speed, and reports the result as a position on the IQ scale — a percentile against a norm group rather than against your own impression. It is a narrow measurement with real limits, set out on how accurate IQ tests are, but it is anchored to something outside your own judgement.

Where this sits among the other ideas about intelligence

The Dunning-Kruger effect belongs to a family of claims that are more appealing than the evidence behind them. The learning styles claim failed direct testing; the left-brain / right-brain personality claim failed it too. This one is different in an interesting way: it was not refuted, it was measured more carefully and turned out to be smaller. That is the more common fate of a psychology finding, and the harder one to report.

If the question underneath is whether you can improve rather than whether you are fooling yourself, what the evidence supports about raising your IQ collects that separately.

Trust and scope notes

This page is educational. It describes published research on self-assessment and is not a diagnosis, a personality judgement or clinical advice. Nothing here is a claim about any individual person.

IQ Revealed is not affiliated with, endorsed by or connected to Cornell University, the University of Michigan, the American Psychological Association, Elsevier, or any of the universities, journals or researchers named on this page. They are named so that each claim can be traced to the work that made it. Quoted phrases are taken from the published abstracts of the papers linked in the sources.

Frequently asked questions

What is the Dunning-Kruger effect in simple terms?＋

It is the claim that people who are bad at something are also bad at judging that they are bad at it, because the knowledge needed to do the task well is the same knowledge needed to spot your own mistakes. Kruger and Dunning proposed it in 1999 after testing Cornell students on humour, logic and grammar. Participants in the bottom quarter of one test scored around the 12th percentile and placed themselves around the 58th.

Has the Dunning-Kruger effect been debunked?＋

Not debunked, but substantially cut down. The pattern in the original graphs is partly produced by two statistical facts that would appear even in random data: regression toward the mean, and the better-than-average effect. When Gignac and Zajenkowski tested the hypothesis on intelligence specifically in 2020, using 929 adults and Raven's Advanced Progressive Matrices, they found the relationship between measured and self-assessed intelligence was essentially linear, with no sign of the pattern the hypothesis predicts.

Is the Dunning-Kruger graph with 'Mount Stupid' real?＋

That curve is not from the paper. The famous picture — confidence spiking at a peak of false expertise, collapsing into a valley, then recovering — appears nowhere in Kruger and Dunning's 1999 article. Their figures plot perceived ability and actual test score across four performance quartiles, and perceived ability never falls as skill rises. It stays roughly flat and high while actual score climbs. The dip-and-recovery curve was drawn later by other people.

Do smart people underestimate themselves?＋

In the original studies, top-quartile participants did rate their relative standing below where they actually placed. Dunning and Kruger explained this differently from the low-end error: high performers found the task easy and assumed others did too, so they misjudged everyone rather than themselves. Note that in one study they did not underestimate their raw score, only their rank against other people.

How accurate is your own guess at your IQ?＋

Weakly accurate. A meta-analysis by Freund and Kasten pooled 154 effect sizes from 41 studies and put the correlation between self-estimated and psychometrically measured cognitive ability at about .33 — which means a self-estimate accounts for roughly a tenth of the differences between people. It is better than nothing and far worse than a measurement.

What part of the Dunning-Kruger effect still stands?＋

Two things. The better-than-average effect is robust: across the original studies the whole sample, not just the weak performers, placed itself above the midpoint. And the metacognitive argument — that judging your own work requires some of the same skill as doing it — remains plausible and is supported by the training study in the original paper, where teaching poor performers the skill improved their self-assessment. What is contested is the size and the shape of the effect, not the idea that self-assessment is unreliable.