The Dunning-Kruger effect: what Dunning and Kruger found in 1999, and what the myth gets wrong

The Dunning-Kruger effect says that people in the bottom quarter of performers overestimate their own ability most on average. In the 1999 studies the bottom quartile scored at the 12th percentile but rated itself at the 62nd percentile. The well known curve with Mount Stupid does not come from these studies.

• The core is a double burden: the same gaps in knowledge cause errors and hide them from the person who makes them.
• David Dunning and Justin Kruger showed the pattern in 1999 in four studies on humor, logical reasoning and grammar at Cornell University.
• In a classroom exam with 141 students the bottom quarter scored at the 12th percentile and self-rated at the 60th percentile, a gap of more than 45 percentile points.
• Mount Stupid is a meme graph from a comic, not research data: real confidence rises with skill, more slowly than skill itself.
• The critique in 2020 and 2026 shrinks the effect but leaves a remainder: almost everyone overestimates, weak performers most, and training improves calibration.

What the effect claims

The Dunning-Kruger effect describes a bias in self-assessment. Weak performance goes with strong overestimation on average. Strong performance goes with a slight underestimation on average.

Dunning later called this the double burden of incompetence. Metacognition is missing: a person who does not recognize what a correct answer requires also does not recognize that the own answer is wrong. The same gap causes the error and hides it.

Three results belong to the definition. First, weak performers overestimate their ability. Second, they judge ability in others poorly. Third, they judge their own ability better after training improves their skill. All three come from the 1999 paper and from the 2011 review by Dunning.

A related distinction helps to place the term. People who judge other people show own patterns, such as the rating errors described under perceptual errors in performance appraisal. The Dunning-Kruger effect concerns the judgement of the own ability instead. It belongs to the judgement biases, next to the biases in the overview of ten cognitive biases.

What Dunning and Kruger tested in 1999

Psychologists David Dunning and Justin Kruger worked at Cornell University in 1999. They tested students in four studies in three fields: humor, logical reasoning and English grammar. Participants solved tasks and then rated their own performance, also compared with others. Kruger and Dunning reported the results in the Journal of Personality and Social Psychology.

The pattern was the same in all fields. The bottom performance quartile overestimated itself most. The top quartile rated itself a little too low. The source of the top error is not modesty: high performers assume that others find the task as easy as they do, a form of the false consensus effect.

Two findings from the original work receive less attention. First, the best performers recognized ability in others more accurately, while the weakest recognized it poorly. Second, practice helped: participants who improved through training judged themselves more accurately afterwards. Calibration follows skill, and that is the most useful result of the original work for study.

Little experience does not protect against the bias either. Sanchez and Dunning showed in 2018 that beginners with some experience are prone to overconfidence. At the very start, knowledge of the own gaps still protects. After first experience this protection fades. Research calls this the beginner bubble. A plain summary by the authors explains the beginner bubble without jargon.

The numbers: 12th against 62nd

The clearest numbers come from a classroom exam, not only from the laboratory. Dunning and colleagues asked 141 students to rate their own performance after a test and compared feeling with points. Dunning reported the numbers in a 2011 review chapter.

The bottom quarter scored at the 12th percentile in reality. That means 88 of 100 scored better. The same quarter felt itself at the 60th percentile, above the average. The error was more than 45 percentile points. In raw points the bottom quarter overestimated its score by almost 30 percent.

All respondents overestimated themselves. The felt mastery of the material averaged at the 70th percentile, the felt test performance at the 68th percentile. Both lie clearly above the possible average of 50. The distance between bottom and top was small in self-rating: bottom and top quartiles differed by only 15 to 20 percentile points.

A percentile is a rank, not a grade. The 12th percentile means 12 percent score worse and 88 percent score better. The 60th percentile means feeling better than 60 percent. The comparison shows the core of the effect in one number: feeling and rank differed by almost 50 places in the bottom quarter.

The laboratory numbers point in the same direction with different labels. In the 1999 logic test with 45 students, the bottom group answered 10 of 20 questions correctly on average and the top group 17 of 20, while both groups estimated about 14 correct. In peer comparison the bottom group placed itself at the 62nd percentile. Wikipedia summarises the 1999 design and the 12th against 62nd result. Scientific American retells the same 10 against 17 against 14 numbers. Britannica gives the paper title and the three fields.

The graph that never existed

The best known graph for the effect shows a curve: first a steep mountain of overconfidence, then a valley of despair, then a plateau of sustainability. This graph does not come from research. It is an invention of the net.

The real data look different. Confidence does not fall as knowledge grows. It rises with performance, only more slowly than performance itself. The weakest lie far off, the best lie only a little off and rather below their mark.

The confusion is understandable but has consequences. A person who knows the false curve expects a crash of self-confidence after first success. This crash does not exist in the data. What exists is a slow approach of feeling and performance through practice and feedback.

The source of the meme is documented. An engineering manager blog traces Mount Stupid to a Saturday Morning Breakfast Cereal comic and reproduces the comic with permission as a humorous plot. The page names the comic as the source of the Mount Stupid plot. No study by Dunning and Kruger contains this mountain, valley and plateau.

Imposter feeling is often named as the opposite of the effect: very able people hold their success as undeserved. It describes the other end of the scale but follows a separate mechanism. The boundary to self-worth helps further: self-worth rates the own value as a person, self-efficacy rates the own ability in a task. The article on self-esteem against self-efficacy explains this difference.

The critique: what statistics explain, and what remains

The effect is smaller than its reputation. That is the state of the debate, not a refutation. Two statistical objections carry the critique.

The first objection is regression to the mean. Extreme values move toward the middle on repetition. A person who scores very badly once often had bad luck: hard questions, bad form on the day, chance in multiple choice. Next time the same person lands nearer the middle. A person who sorts only by test result into quartiles sorts luck with it. That creates a pattern that looks like overestimation by the weak, even when no psychological mechanism acts.

Gignac and Zajenkowski argued in 2020 that the effect is largely a statistical artefact. Next to regression to the mean they name the better than average effect: most hold themselves as better than the average. That cannot hold mathematically but appears stably. With random data a similar pattern can be produced. The Decision Lab summarises the statistical account with regression and the better than average heuristic.

The second objection came in 2026 from Bath and London. Dawson and colleagues recalculated large replication data sets with improved methods. The University of Bath reported the reversal in July 2026. The result: after clearing random influences the classic pattern disappears. Instead the most able overestimate most. The authors read overconfidence as evolutionarily useful: it signals competence and gains status. Loss aversion brakes action at the same time, so big words rarely lead to big risks. The paper appeared in Psychological Review. A German science news piece summarises the Bath and London analysis.

The random data argument is older and independent of Dawson. Colleagues created 1154 fictional people and assigned test score and self-rating at random. After sorting into quartiles the same Dunning-Kruger graph appeared without any human bias. In a science literacy test with 25 questions only 16.5 percent of the weak students clearly overestimated, while almost 80 percent judged fairly well. McGill University explains the random data result and the measurement problem.

What remains after the critique holds in three points. First, almost everyone overestimates, not only the weak. Second, a remainder of stronger overestimation by the weak stays, smaller than first thought. Third, practice and feedback improve judgement, as the 1999 training conditions already showed. The effect is therefore not a law of incompetence. It is a tendency with noise, and feedback damps it.

What this means for studying

For exam preparation one result counts: feeling deceives, a tested answer does not. The bottom and top quartiles differed in feeling by only few points. A person who trusts feeling hardly recognizes the own level. A person who writes old exams or lets questions test them receives a number instead of a feeling.

Three steps calibrate the estimate. Test yourself with real tasks instead of rereading. Compare each answer with the model solution sentence by sentence. Repeat the test after some days with spaced intervals. Each step replaces feeling with evidence.

A fixed routine helps. If you use Psychology World, save each term you look up in a list for the current topic. Test the list as a question, as a flashcard and as a typing exercise. The question form shows gaps at once, because each answer is right or wrong. The method follows the same routine as the article on how to memorise psychology terms so they stick. Mixing topics in one session sharpens calibration further, as the guide on the interleaving study method shows.

1. 1.If a term is new, save it in a list for the current topic before you study it.
2. 2.Test the list with the Question game first, because each answer is right or wrong.
3. 3.If you miss an item, read the definition and test it again after two days.
4. 4.Mix two topics in one session to stop feeling fluent from one easy set.

Students who prepare for AP Psychology or the MCAT meet the effect in self-assessment questions: rank the own answer against a peer group, not against a feeling of fluency. Old exams, model solutions and spaced retesting give the rank.