The Dunning-Kruger effect

The finding, from Kruger and Dunning‘s four 1999 studies of humour, logic and grammar, that poor performers overestimate how they rank against other people, and by a lot. Bottom-quartile participants sat at the 12th percentile and placed themselves at the 62nd (kruger-dunning-1999).

The proposed mechanism is metacognitive: judging your own work draws on the same competence as doing it, so incompetence is partly self-concealing. Top performers were also wrong about their standing, in the opposite direction, and the paper gives that its own explanation — see below.

Read from both primaries as of 2026-08-05. The account here previously rested on one popular summary; two claims on this page changed when the papers arrived, and both are marked.

What the popular version added

Almost everything people can tell you about this effect comes from somewhere other than the paper.

There is no time axis. Performance and self-estimate were recorded once, in one sitting. The familiar story — a novice’s confidence spiking early and crashing on contact with real difficulty — requires following people as they learn, and nobody did. “Mount Stupid,” the peak-and-valley curve, is not in the 1999 paper, which contains four percentile-against-quartile figures and no confidence trajectory.

There is no claim about intelligence. “Confident people are stupid” and “smart people are humble” are both later inventions. On the second, dunning-kruger-misread reports a five-study paper (1,189 participants) finding intellectual humility associated with curiosity and appetite for learning rather than with cognitive ability.

This gap is unusual in what it is a gap between. The corpus’s other distortions run from an author outward — a modest result sold immodestly on a jacket (feeling-good, david-d-burns), a mechanism outrunning its evidence in a bestseller (dopamine-nation). Here the finding stayed where it was and the audience built the myth: a chart nobody in the study drew, describing a trajectory nobody was followed along.

The two errors have opposite sources

The half that survives every retelling is the incompetent overestimating themselves. The other half is in the paper under its own heading, The Burden of Expertise: top performers underestimated their relative standing while judging their absolute score accurately.

Kruger and Dunning attribute that to the false-consensus effect — having found the task easy, the able assume everyone else did. Phase 2 of Study 3 tested it: shown their peers’ actual work, top performers revised upward and bottom performers did not move at all. So the incompetent are wrong about themselves, and the competent are wrong about other people.

Is it real?

Contested for over twenty years, and the disagreement survives having both primaries in hand. It is not a replication dispute — nobody claims the pattern fails to appear.

The statistical case. magnus-peresetsky-2022 fits a three-parameter doubly-censored tobit model — measurement noise plus a random floor and ceiling — to 665 students predicting their own exam grades on a 0-100 scale, and gets a near-perfect fit. Their conclusion: “there is an effect, but it does not reflect human nature.” Someone genuinely near the floor can only err upward. Krueger & Mueller (2002) opened this line with regression-plus-better-than-average; Gignac & Zajenkowski (2020) made the same charge in Intelligence.

Corrected 2026-08-05: this page previously described that work as showing the shape arises “in data with no effect in it.” That overstates it. It is a model fitted to real student predictions, not a simulation from pure noise, and the claim is that the fit needs no psychological parameter.

The task-difficulty case. Burson, Larrick & Klayman (2006), across twelve tasks: difficulty governs who is miscalibrated, and on hard tasks top performers can become the worse judges of relative position.

What the statistical case does not reach. Kruger and Dunning saw the regression objection coming and built Study 4 against it. Ten minutes of logic training moved the trained bottom quartile’s self-estimate from the 55th percentile to the 44th, and from “I got 5.3 right” to “I got 1.0 right,” while an untrained control group did not shift. A floor-and-ceiling artifact does not explain why a packet handed out after the test changes self-assessment. magnus-peresetsky-2022 does not address that experiment.

And a replication for it. A 2021 Nature study, roughly 4,000 participants per study, supported the metacognitive reading for grammar and logic (dunning-kruger-misread, which names neither its authors nor its title).

So the defensible statement is narrower than either camp’s headline: the percentile-crossing chart is explained by censoring; the training experiment is not. Nobody in this corpus has refuted anybody.

Why the statistical objection is its own category

Worth separating from the replication failures this spoke usually records. replication-crisis collects results that stop appearing when you run the study again. The censoring argument is the reverse complaint: the effect keeps appearing, and would keep appearing whether or not the mechanism is there. Larger samples sharpen the artifact rather than exposing it.

That distinction bears on how heuristics-and-biases gets read generally. “Systematic, not random” was the claim that made the programme a science — and a bounded scale supplies systematic direction with no mechanism behind it.

Related

kruger-dunning-1999 · magnus-peresetsky-2022 · dunning-kruger-misread · heuristics-and-biases · replication-crisis · david-dunning · justin-kruger · focusing-illusion · synthesis