The Dunning-Kruger Effect Is Not What You Think It Is

The Dunning-Kruger effect is not the finding that incompetent people believe they are brilliant. The 1999 study found that people at every skill level estimated themselves as somewhere above average, that the lowest scorers were the furthest off, and that the highest scorers slightly underestimated themselves. Nobody in the bottom quartile claimed to be exceptional. And a serious statistical critique argues that a meaningful share of the famous pattern can be generated by random noise and the way the data are graphed.

So this piece has two jobs. First, correct the meme. Second, complicate the correction, because the standard "well actually" version of Dunning-Kruger is also incomplete.

What Kruger and Dunning actually did

Justin Kruger and David Dunning published "Unskilled and unaware of it" in the Journal of Personality and Social Psychology in 1999, working at Cornell. Across four studies they tested undergraduates on humor, logical reasoning and grammar, then asked them to estimate their own percentile standing.

The headline number is specific and worth quoting precisely. Participants scoring in the bottom quartile landed at roughly the 12th percentile on the actual tests and estimated themselves at roughly the 62nd. That is a gap of about 50 percentile points, and it is real. But look at what the estimate was. The 62nd percentile is "slightly above average," not "genius." The people the internet mocks were, in the data, claiming to be a bit better than typical.

Kruger and Dunning proposed a dual burden: the skills required to do well at a task overlap with the skills required to judge how well you did, so people lacking the first also lack the second. Their fourth study supported this by training participants in logical reasoning and finding that improved competence improved self-assessment.

Where the popular version diverges, ranked by how wrong it is

1. "Incompetent people think they are experts." The most wrong. Bottom-quartile participants estimated the 62nd percentile. The effect is an above-average bias, not delusion of grandeur.
2. "It is about stupidity." Wrong category. The studies measured performance on specific tasks in a specific sample of Cornell undergraduates, not intelligence, and certainly not fixed traits of persons.
3. "Experts underrate themselves because they assume everyone finds it easy." The underestimation by top performers is in the data, and Kruger and Dunning did offer a false-consensus explanation. But this half of the effect is the part most vulnerable to the statistical critique below, and it gets stated with far more certainty than it has earned.
4. "There is a curve with a Mount Stupid on it." Entirely fabricated. The famous peaked confidence curve circulating online appears nowhere in the paper. The actual figure is two roughly monotonic lines: perceived ability nearly flat across quartiles, actual ability rising.
5. "It explains the people I disagree with." The effect is used almost exclusively as an accusation aimed outward, which is itself a decent illustration of the thing it describes.

The critique that makes this interesting

Here is the part that rarely survives into the corrected version. The classic Dunning-Kruger figure plots self-assessment against actual performance quartiles, and both axes contain measurement error drawn from the same source. When you rank people by a noisy measure and then examine any other variable across those ranks, regression to the mean guarantees a pattern that looks like the bottom overestimating and the top underestimating, even when no psychological effect exists at all.

Edward Nuhfer, Steven Fleisher, Christopher Cogan, Karl Wirth and Eric Gaze made this argument concretely in Numeracy in 2017, in a paper titled "How Random Noise and a Graphical Convention Subverted Behavioral Scientists' Explanations of Self-Assessment Data." They ran random-number simulations and produced the signature Kruger-Dunning graphical pattern from data with no signal in it whatsoever. They then analyzed a real dataset of 1,154 people using paired measures of documented reliability and criterion-referenced rather than purely normative comparisons, and reported the opposite conclusion from the popular one: people's self-assessments of competence generally do reflect competence they can demonstrate. Their data also showed experts self-assessing more accurately than novices, which preserves one piece of the original claim while dismantling the framing around it.

Related critiques arrived from several directions. Joachim Krueger and Ross Mueller argued in 2002 that regression to the mean plus the better-than-average effect accounts for much of the pattern. Katherine Burson, Richard Larrick and Joshua Klayman published a study in the Journal of Personality and Social Psychology in 2006 running 12 tasks across three studies and finding that task difficulty drives the shape of the miscalibration. On moderately difficult tasks, best and worst performers were about equally accurate. On harder tasks, best performers were less accurate than worst performers, which is the exact reverse of the metacognitive story. Their conclusion was that a simple noise-plus-bias model explains the data without any special deficit among the unskilled. Gilles Gignac and Marcin Zajenkowski reported in the journal Intelligence in 2020 that using an analysis less vulnerable to the artifact left only a small residual effect.

What survives

Not nothing, and this is where honesty cuts both ways. Kruger and Dunning have responded to the critiques and continue to argue that a metacognitive component remains after statistical corrections. Several findings do appear robust across methods.

• People are noisy self-assessors. Correlations between self-rated and measured ability are typically modest across many domains. That is not in dispute.
• A better-than-average bias is real. Most people place themselves above the median on most desirable dimensions, which is arithmetically impossible in aggregate.
• Expertise improves calibration. Both the original studies and the Nuhfer data show this, from different directions.
• Training improves self-assessment. Kruger and Dunning's fourth study is the strongest part of the original paper and least dependent on the contested graph.

What has not survived is the confident causal story that a specific metacognitive deficit among low performers produces a specific dramatic overestimation. The measured gap is smaller than advertised, part of it is an artifact of ranking on noisy variables, and the popular version bears almost no relationship to the data.

Why this matters outside psychology departments

Because self-assessment is the input to nearly everything in the self-improvement category. Deciding what to work on, whether you are improving, and whether a practice like positive self-talk is helping all run through a judgment that this literature says is unreliable and that the statistical critique says is measured badly. If you cannot cleanly tell whether your estimate of yourself is accurate, then the practical move is to lean on external evidence rather than internal impressions, which is the same conclusion that comes out of the distinction between confidence and task-specific self-efficacy.

There is also an irony worth stating plainly. The Dunning-Kruger meme is used to dismiss other people's judgment by citing a study almost nobody who cites it has read. The finding about confident people making claims beyond their evidence turns out to describe the finding's own popularization better than it describes its participants. Something similar has happened to a related construct pointed the other way, which we cover in the piece on what Clance and Imes actually found.

The reasonable posture is neither "everyone is deluded" nor "self-assessment is fine." It is that self-judgment carries a lot of error, that ranking people by noisy scores manufactures patterns, and that the correct response to both is to check against something outside your own head.