Dunning–Kruger effect

These illustrate the pattern the effect describes. Whether the explanation offered for each is the right one is the disputed part (see Evidence).

The grammar quiz

At a staff training day, everyone takes a twenty-question grammar quiz and then guesses how they ranked. One participant, who scored near the bottom, puts themselves “a bit above the middle”.

On the original account, the participant marked their wrong answers as right because they didn’t know the rules those answers broke, so they had no way to see how many they’d missed. That’s the claimed “double burden”: getting it wrong, and being unable to tell.

Judging everyone else, too

A camera club asks members to score each other’s photos on focus, exposure and composition. A newer member gives nearly identical scores to a sharp, well-exposed shot and a slightly blurry, overexposed one, and rates their own entry among the best.

The same missing skill that would reveal the flaws in their own photo also hides the flaws in other people’s. Kruger and Dunning reported this: low scorers were worse at grading their peers’ tests, and seeing better answers didn’t lead them to lower their own estimates.

The expert who guesses too low

The club’s most experienced photographer finds the scoring task easy and assumes most members did too. They rank themselves in the top third, when they were in fact the best in the room.

This is the other half of the claimed pattern, and the explanation offered is different. High scorers judge their own work well but assume others find it as easy as they do, so they underestimate their relative standing.

• When the judgment doesn’t use the skill being judged. You can tell whether a free throw went in without being good at basketball. Where the result is visible to anyone, poor performers can see it.
• When you’ve had feedback against an outside standard. An estimate calibrated by scores, reviews or results isn’t resting on self-assessment alone.
• When everyone is overestimating by a similar amount. If most people rate themselves above average, the lowest scorers will show the largest gap simply because they’re furthest from where everyone places themselves. A bigger gap at the bottom isn’t, on its own, evidence of a special blind spot there.
• Naming it isn’t a rebuttal. Saying someone is “Dunning–Kruger-ing” says nothing about whether their claim is true; that still has to be checked on the merits.

The test: could this person tell a right answer from a wrong one here, if they saw both?

An ordinary beginner’s overconfidence

A new runner trains for a 10K on a flat riverside path and predicts a finish time based on those runs. The race course turns out to be hilly, and they finish seven minutes slower than predicted.

The runner was overconfident, but not because they lacked the skill to judge their running. They were missing information about the course. As soon as they saw the result they understood it exactly. The claimed effect is about errors a person can’t see even after the fact, not predictions that miss for ordinary reasons.

A gap produced by the numbers alone

On a class quiz, nearly every student estimates they finished around the 60th to 70th percentile. The bottom scorers turn out to have overestimated their rank by 50 points; the top scorers underestimated theirs by 20.

This looks like the effect, but it’s exactly what you’d get if every student simply guessed “a bit above average” regardless of how they did. The students furthest below that guess show the biggest gap. This is regression to the mean combined with the better-than-average effect, the alternative explanation critics have raised since 2002. To show a real blind spot, you have to show that low scorers’ estimates are less connected to their actual performance, not just further from it.

Kruger and Dunning’s explanation is metacognitive: judging whether an answer is right requires the same knowledge as producing it, so the gaps in what you know are invisible from inside. David Dunning later described these as “unknown unknowns”. It’s a close relative of “what you see is all there is”: errors you can’t recognize don’t register as missing.

The alternatives don’t require any special blind spot:

• Regression to the mean. Self-estimates are only loosely related to actual scores, so people with extreme scores will, on average, estimate something closer to the middle.
• The better-than-average effect. Most people place themselves somewhat above average on most skills, which pushes every estimate up and makes the gap largest at the bottom.
• Task difficulty. How hard a task feels shifts everyone’s guesses about where they rank, whatever their skill, which changes who ends up looking miscalibrated.

These can all operate at once, which is why separating them has taken more than two decades of argument.

Only one remedy has been shown to work, and the evidence for it is thin.

• Learning the skill improved self-assessment. In Kruger and Dunning’s fourth study, low scorers on a logic test who were given a ten-minute lesson on that type of problem became far better at identifying which of their answers were wrong, and lowered their estimates of how they’d done. This was one study of 140 Cornell undergraduates, and the training taught the exact kind of problem they had just attempted.
• Seeing better work didn’t help on its own. In the same paper, low scorers who graded more skilled peers’ answers did not revise their self-estimates downward.

In practice, the useful move is the one the effect implies: check your estimate against something outside your own judgment (a score, a reviewer, a result) rather than against how confident you feel. That advice follows from the account; it hasn’t been tested as a debiasing technique.

Status: contested. The pattern (lowest scorers overestimate themselves the most) replicates. Whether it shows a real deficit in self-insight, and how big that deficit is, is seriously disputed, with credible studies on both sides.

The original studies. Kruger and Dunning (1999) ran four studies with Cornell undergraduates on humor, logical reasoning and grammar. Across the studies, people in the bottom quartile scored around the 12th percentile on average but estimated themselves around the 62nd. Top-quartile participants underestimated their standing (in Study 4, they scored around the 90th percentile and guessed the high 70s). The authors addressed regression to the mean directly, arguing it couldn’t be the whole story because the bottom quartile’s overestimation was much larger than the top quartile’s underestimation, and because training changed low scorers’ self-estimates.

The popular version. The familiar “Mount Stupid” curve, where confidence spikes with a little knowledge, dips, and then climbs with expertise, does not appear in the paper. Its figures plot estimated and actual percentile for each performance quartile. Estimated rank rose far less from the bottom quartile to the top than actual rank did. Low scorers weren’t more confident than high scorers, just not much less confident, and a flat line of estimates across very different scores is exactly what regression to the mean produces. That’s why it’s a live alternative.

The artifact critique.

• Krueger and Mueller (2002) replicated the basic pattern and found that when either regression or the better-than-average effect was statistically removed, the asymmetry between low and high scorers disappeared.
• Burson, Larrick and Klayman (2006) varied task difficulty and found that on harder tasks, the best performers were the less accurate judges of their relative standing, suggesting judges at every skill level make similar errors.
• Nuhfer and colleagues (2016, 2017) used simulations to show how random noise, graphed the conventional way, produces patterns that look like evidence about self-assessment. In their own data (1,154 participants rating their science literacy), self-assessments generally tracked real competence.
• Gignac and Zajenkowski (2020) proposed tests that the artifacts can’t produce. In 929 members of the public who rated their own intelligence and took a reasoning test, errors in self-assessment were about the same size at every ability level, and the relationship was essentially linear. They concluded the effect may be much smaller than reported.
• Dunkel, Nedelec and van der Linden (2023) applied the same tests to a large, nationally representative US sample and did find a statistically significant effect, but a minimal one.

Evidence for a real component.

• Ehrlinger and colleagues (2008), including Dunning and Kruger, reported five studies finding the pattern in real-world settings and when accuracy was rewarded, and traced low scorers’ optimism to missing their own errors rather than misjudging their peers.
• Jansen, Rafferty and Griffiths (2021) built a model showing that the pattern can arise from reasonable reliance on prior beliefs alone, then ran two large online replications of about 4,000 participants each in grammar and logic. The data fit best a model in which low performers really are less able to tell whether each answer is correct. A commentary by Mazor and Fleming (2021) summarized this as showing the effect is not merely a statistical artifact, while noting it may be a single burden (poor performance showing up twice) rather than a separate metacognitive deficit.
• McIntosh and colleagues (2019), in experiments with preregistered plans, using simple pointing and spatial-memory tasks, found that self-insight did track skill, but contributed only weakly to the pattern; task performance itself was the main driver. They concluded metacognitive differences can contribute but are “neither necessary nor sufficient”.

What remains uncertain. How much of the typical Dunning–Kruger graph is artifact, how large the real part is, whether it varies by domain, and whether low performers’ poor self-judgment is a distinct deficit or just another symptom of low skill. Dunning (2011) reviews the case for the effect; the studies above show why that case is not settled.