TL;DR. What you think you know about the Dunning-Kruger effect is probably overstated. The original 1999 paper ran four studies and found a real mismatch: participants in the bottom quartile on tests of humor, grammar and logic sat at the 12th percentile but estimated themselves at the 62nd, a gap of roughly 50 percentile points. But in 2020, a paper in Intelligence argued that the standard four-quartile plotting method producing that finding is confounded by the better-than-average effect and regression toward the mean. Applying two valid statistical tests to 929 participants, the authors failed to find the variance pattern the hypothesis predicts and found the relationship between measured and self-assessed intelligence to be essentially entirely linear. The conclusion is not that the effect does not exist. It is that its magnitude may be much smaller than previously reported. If you manage yourself like a CEO, that distinction matters: confident incompetence is not a finding solid enough to use as an insult, but the weakness of self-assessment is a different and better supported fact.
What the original study found
Justin Kruger and David Dunning published their paper in the Journal of Personality and Social Psychology in 1999 under the title “Unskilled and unaware of it.” The thesis: people who are unskilled in a domain carry a dual burden. They not only reach erroneous conclusions, their incompetence also robs them of the metacognitive ability to notice.
Four studies used tests of humor, grammar and logic. The verified core figure: participants scoring in the bottom quartile grossly overestimated their test performance and ability. Their test scores put them in the 12th percentile; they estimated themselves in the 62nd.
The paper also found something less quoted and more useful: improving participants’ skills, and thus their metacognitive competence, helped them recognise the limits of their own ability. The problem was not presented as a fixed personality trait.
Then came the objection
In 2020, Gilles Gignac and Marcin Zajenkowski published a critique in Intelligence. The title was direct: the Dunning-Kruger effect is (mostly) a statistical artefact.
Their objection was methodological. The most common method used to test the hypothesis plots self-assessed and objectively assessed means across four quartiles of objective ability. That method has been argued to be confounded by two known phenomena: the better-than-average effect (most people rate themselves above average) and regression toward the mean (those at the bottom of one extreme measurement drift toward the average on a second measurement). Together, those two produce the famous crossing chart even when no genuine Dunning-Kruger effect is present.
The authors proposed two inferential techniques they argue are valid: the Glejser test of heteroscedasticity and nonlinear (quadratic) regression. They then applied both to a sample of 929 general community participants who completed a self-assessment of intelligence and the Advanced Raven’s Progressive Matrices.
Both results ran against the hypothesis. No statistically significant heteroscedasticity was identified. And the association between objectively measured intelligence and self-assessed intelligence was essentially entirely linear, where the hypothesis predicts curvature.
The authors’ own conclusion is cautious and deserves quoting in full sense: although the phenomenon the hypothesis describes may be to some degree plausible for some skills, the magnitude of the effect may be much smaller than reported previously.
The debate is not closed. A comment by Avram Hiller responding directly to that paper was published in the same journal in 2023. That comment’s argument was not accessible for this piece, so no claim is made about its content; it is noted only that the objection itself drew an objection.
The original comparison: confidence and accuracy across independent studies
The popular narrative says certainty is a marker of incompetence. That is a far stronger claim than what independent studies actually find. The table below places verified findings from different domains side by side against the same question: how well does self-assessment track performance?
Table 1. The relationship between self-assessment and actual performance across four independent findings
| Study and domain | Scale | Finding | Direction |
|---|---|---|---|
| Kruger and Dunning, humor, grammar and logic | Four studies | Bottom quartile sat at the 12th percentile, estimated the 62nd | Strong overestimation at low performance |
| Gignac and Zajenkowski, self-assessed intelligence and Raven’s matrices | 929 participants | No significant heteroscedasticity; relationship essentially entirely linear | Predicted pattern not found |
| Porter and Machery, identifying AI-generated poetry | 16,340 judgements | Self-reported confidence had a significant negative effect on accuracy (b = -0.021673, p < 0.0001) | The confident were slightly more often wrong |
| Jones and Bergen, detecting AI in live conversation | 284 participants, 1,023 games | Self-estimated accuracy correlated positively with real accuracy among undergraduates (p = 0.03) but not among Prolific participants (p = 0.45) | Direction did not replicate across populations |
CEOtudent editorial framework. All figures are as published by the respective studies. The four rows use different designs, populations and outcome measures and cannot be pooled statistically; the table is assembled to display an evidence picture too inconsistent to be read in one direction. The 50-percentile-point gap between 12 and 62 in the first row is a CEOtudent subtraction.
The only honest way to read the table is this: self-assessment is not a reliable proxy for performance, but the direction of its unreliability is not reliable either. In some contexts confidence relates negatively to accuracy. In others positively. In one, two populations within the same study pointed opposite ways.
That is a messier picture than the confident-incompetence rhetoric can carry. And the messiness itself is usable information.
Where the popular narrative goes wrong
Three common errors, none supported by the evidence.
First, using it as a personality diagnosis. The original study reported a group average, not an individual diagnosis. You cannot say whether a particular person is overestimating without measuring both their real and their estimated performance. Labelling an opponent in an argument “Dunning-Kruger” is an inference the data does not support.
Second, the claim that the most competent underestimate themselves. That is the second half of the crossing chart, and precisely the part Gignac and Zajenkowski explain through regression toward the mean. Those at the top of one extreme measurement drift down on a second; that is a property of measurement, not modesty.
Third, assuming the effect is uncorrectable. The original paper found the opposite: when participants’ skills improved, they recognised their own limits better. The mechanism was a knowledge deficit, not a character flaw.
What survives: a much narrower but genuinely usable conclusion
The claim the evidence can carry is this: your self-assessment of your own competence is not a reliable enough signal to build decisions on. That is far less satisfying than “the ignorant do not know themselves,” but far more defensible and far more useful day to day, because for the first time it includes you.
That distinction maps exactly onto the pattern found in our analysis of the discernment gap: people fail to identify machine-generated text whether or not they are experts, and the confident are slightly more often wrong. Same lesson, different domain.
Table 2. What to do once you accept that self-assessment is weak
| Step | What to do | Rationale | Status |
|---|---|---|---|
| 1. Anchor competence claims to external measurement | Say “I scored X on this test” rather than “I am good at this” | The relationship between self-assessment and objective measurement is inconsistent across the studies above | Evidence-backed direction |
| 2. Apply the label to yourself, not to others | Stop using Dunning-Kruger as a rhetorical weapon | The finding is group-level; individual diagnosis is not supported | Evidence-backed |
| 3. Treat confidence as a warning, not a signal | When your certainty rises, increase checking rather than reduce it | Confidence related significantly negatively to accuracy in one study; it is a reliable positive signal in none | Evidence-backed, with a replication caveat |
| 4. Measure calibration, do not estimate it | Write down decisions and predictions, then check outcomes against them | No source cited here measured this practice | CEOtudent convention |
| 5. Raise awareness by raising skill | If you want to assess yourself better in a domain, become more competent in it | The original paper reported that improving skills helped participants recognise their limits | Evidence-backed |
CEOtudent editorial framework. Steps 1 to 3 and 5 restate findings the cited studies reported. Step 4 is deliberately labelled a convention; no source measured it.
For the concrete version of step 4, our decision journal protocol does exactly this job: it records the prediction at the moment of the decision, so calibration stops being something you remember and becomes something you measure. A different application of the same discipline appears in our piece on auditing what you think you know about your industry.
The CEO and student lens
The tension here is clear. An executive has to display certainty when deciding; visible hesitation paralyses an organisation. A student has to keep testing their certainty; otherwise they do not learn.
What the evidence says is that both are possible at once: decisive outward, calibrated inward. The display of certainty is a communication tool. Certainty itself is not evidence. A manager who confuses the two moves quickly in the wrong direction.
FAQ
Has the Dunning-Kruger effect been debunked?
No, but the claim has weakened. The 1999 finding was real; the objection is that the method producing it is confounded by the better-than-average effect and regression toward the mean. When valid tests were applied to 929 participants, the predicted pattern was not found. The authors’ own wording is that the effect may be plausible for some skills but much smaller in magnitude than previously reported.
Does this mean nobody is unaware of their own incompetence?
No. That self-assessment tracks performance poorly is a fact the evidence supports. What is not supported is that this error is specific to the unskilled, or that it can be diagnosed in an individual.
Do the most competent really underestimate themselves?
That is the part targeted by the same statistical objection. Those at the top of an extreme measurement drifting toward the mean on a second measurement is the expected consequence of regression toward the mean, and requires no modesty.
How do I improve my self-assessment?
Two routes, both supported. The original paper reported that improving skill improved awareness. Separately, replacing self-assessment with external measurement swaps a signal of unknown reliability for one of known reliability.
How large was the gap in percentile terms?
In the original study, roughly 50 percentile points between the bottom quartile’s actual 12th percentile and their estimated 62nd. That figure is accurate as the 1999 paper reported it; what is disputed is not the number but the interpretation of the method that produced it.
Sources
Kruger and Dunning. Unskilled and unaware of it: how difficulties in recognizing one’s own incompetence lead to inflated self-assessments. Journal of Personality and Social Psychology, 1999.
Gignac and Zajenkowski. The Dunning-Kruger effect is (mostly) a statistical artefact: Valid approaches to testing the hypothesis with individual differences data. Intelligence, 2020.
Hiller. Comment on Gignac and Zajenkowski. Intelligence, 2023. Only its existence and bibliographic record were verified; its argument was not accessible for this piece.
Porter and Machery. AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably. Scientific Reports.
Jones and Bergen. Large Language Models Pass the Turing Test. arXiv preprint, Computation and Language.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
