The Dunning-Kruger effect isn't what everyone thinks it is: the original 1999 study found bottom-quartile performers overestimated their standing by about 50 percentile points, not that confident people are stupid

The Dunning-Kruger effect may be the most misused finding in pop psychology, and the way people misuse it often says more about them than about the people they are mocking.

You have seen the shorthand a thousand times. Someone confidently states something wrong online, and the replies fill with the familiar graph: the tall “Mount Stupid” peak, the deep “Valley of Despair,” and the gradual climb toward enlightenment. That graph does not appear in Kruger and Dunning’s original 1999 paper, and it is not simply a prettier version of their data. The paper did not track confidence as people progressed from beginner to expert. It compared performance with estimates of performance and ability at a single point in time.

Here is what the study actually found. Across four studies, participants completed tests involving humour, logical reasoning, and English grammar, then estimated how they had performed relative to their peers. People in the bottom quartile averaged around the 12th percentile in actual performance but estimated themselves to be around the 62nd percentile.

That is a gap of roughly 50 percentile points, not 30.

It is a striking result, but it is narrower than the meme. The paper did not conclude that confident people are stupid or that intelligent people are naturally humble. It proposed that people performing poorly within a specific domain may lack some of the metacognitive knowledge needed to recognize their own mistakes. The same knowledge that helps someone produce a correct answer may also help them identify whether an answer is correct.

Top performers were not perfectly calibrated either. In the original studies, they tended to underestimate their standing relative to their peers. That did not erase the much larger error among bottom-quartile performers, but it did complicate the slogan. Greater competence improved calibration without making self-assessment infallible.

The harder question is why the familiar pattern appears.

In 2002, Joachim Krueger and Ross Mueller argued that a general better-than-average bias combined with statistical regression could explain much of the apparent asymmetry between high and low performers. Kruger and Dunning responded that the statistical account could not explain all of their experimental results and that the replication used unreliable tests and inadequate measures of metacognition.

Then task difficulty entered the argument. In a 2006 series of 12 tasks, Katherine Burson, Richard Larrick, and Joshua Klayman found that the best and worst performers differed little in calibration on moderately difficult tasks. On harder tasks, the best performers could become less accurate about their relative standing than the worst performers.

The reason is intuitive. When a task feels easy, people tend to assume they are doing reasonably well. When it feels difficult, they tend to assume they are doing badly. Those impressions may be shared across the room even when actual performance varies widely. Change the difficulty and the direction of the largest error can change with it.

Statistical design matters too. A 2022 statistical analysis showed how noisy judgments and bounded scoring scales can reproduce the classic Dunning-Kruger shape with remarkable accuracy. People near the bottom have far more room to estimate upward than downward, while people near the top have far more room to estimate downward.

That does not prove that the original psychological explanation is false in every setting. It proves that the quartile graph alone cannot tell us which mechanism produced it.

Nor has the metacognitive explanation disappeared. A large 2021 replication involving roughly 4,000 participants in each of two studies found support for the idea that low performers were less able to estimate whether their individual answers were correct in grammar and logical reasoning tasks.

The honest conclusion is therefore not that the Dunning-Kruger effect has been debunked. It is that the internet’s clean morality play is not what the evidence shows. Low performers can have weaker insight into their errors, but task difficulty, prior beliefs, statistical regression, noisy judgments, and the limits of the scoring scale also shape the pattern.

The meme version says confident people are dumb and smart people are humble. The research says something less emotionally convenient: self-assessment is noisy, domain-specific, and partly dependent on what the task feels like from inside your own head.

That is a much less satisfying tweet.

It also matters because the sloppy version has become one of the internet’s favourite rhetorical weapons. Somebody expresses confidence about a subject, and instead of addressing the argument, a stranger announces that the confidence itself is evidence of incompetence.

This is the part that bothers me. Dunning-Kruger has become a permission slip for dismissing people without doing the harder work of showing where they are wrong. A finding about imperfect self-assessment gets converted into a personality diagnosis delivered from a distance.

The move is especially convenient because it cannot easily be falsified. Disagree with the accusation and your disagreement becomes further evidence that you are too incompetent to understand your incompetence. At that point, the research is no longer functioning as evidence. It is functioning as a trap.

Once you have used a concept as a rhetorical club, reopening the paper becomes uncomfortable. It is easier to defend the slogan than to admit that the evidence is narrower, more contested, and more interesting.

The more useful conversation begins with intellectual humility, but even here the evidence needs careful handling. A five-study paper involving 1,189 participants found that intellectual humility was associated with greater general knowledge and with traits such as curiosity, intellectual openness, reflective thinking, and an intrinsic motivation to learn. It was not associated with higher cognitive ability, and its relationship with metacognitive accuracy was mixed.

That is useful precisely because it is modest. The study does not prove that practising humility will make someone smarter. It suggests that people who are more willing to recognize the limits of their beliefs also tend to display several characteristics that make continued learning easier.

The practical takeaway is therefore not “be humble because humble people are secretly superior.” It is to build habits and feedback loops that make inaccurate self-assessment easier to detect. Pause before asserting. Look for evidence that could prove you wrong. Treat “I don’t know” as a complete sentence rather than an admission of defeat.

I think about this in hiring, where confidence can distort judgment on both sides of the table. Research on personnel selection found that adding unstructured interview information to standardized test results made decision-makers more overconfident in their predictions. Another study of applicant self-presentation found that impression-management behaviour during interviews was unrelated to supervisors’ later ratings of job performance.

That does not mean interviews are useless. It means polish and confidence are not clean measurements of competence. A hiring process needs evidence capable of correcting the interviewer’s first impression, just as a candidate needs feedback capable of correcting an inflated or deflated view of their own ability.

The original paper never said confident people are dumb. It found that bottom-quartile performers in several specific domains overestimated their relative standing by an average of about 50 percentile points, and it proposed that weak metacognitive insight helped explain the gap.

The follow-up literature made the picture harder, not simpler. Task difficulty can change who is most miscalibrated. Statistical features can magnify the famous quartile pattern. At the same time, large replications still find evidence that lower performers can be less sensitive to whether their answers are right.

The practical implication is not “spot the idiots.” It is “create conditions in which everyone, including you, can discover when they are wrong.”

People invoking Dunning-Kruger against strangers online are often making judgments about expertise they cannot directly observe. They are estimating another person’s competence from a few sentences, without a controlled task, an objective score, or a reliable feedback loop.

That does not mean their judgment is necessarily wrong. It means they should be less confident that a psychology paper has already proved it right.

More than a quarter-century after the original study, the safest conclusion is not that everyone is equally bad at self-assessment. The bottom of a performance distribution can be especially miscalibrated. But the top is not immune, the graph is not a law of intellectual development, and no percentile chart gives you a licence to diagnose a stranger.

The effect now appears constantly in political and cultural arguments, usually as an explanation for why the other side believes what it believes. That is the tell. When a psychology finding becomes primarily a way to explain your opponents, it has stopped functioning as a research result and started functioning as a slur with a citation.

Read the actual paper. It is shorter than the arguments people have about it.