What did Kruger and Dunning actually find in 1999?

Across tests of humour, logic and grammar, participants who scored in the bottom quarter estimated their percentile rank far above where they landed, while top performers slightly underestimated theirs. The authors argued that judging your performance draws on the same skills as performing well, so weakness in a skill partly hides itself.

Is the Mount Stupid graph accurate?

No. The popular chart showing confidence spiking then collapsing into a valley of despair does not come from the research, which was not longitudinal and never tracked confidence over time. The studies measured calibration at a single point, and low scorers placed themselves somewhat above average rather than claiming expertise.

Have researchers challenged the effect?

Yes, on two main grounds. The better-than-average effect means most people rate themselves slightly above the middle, which mathematically forces low scorers to overshoot, and regression to the mean with noisy test scores can reproduce the pattern even in randomly generated data. Defenders point to the training results and the breadth of replications, so the fair reading is that the mechanism is contested while poor self-calibration itself is well supported.

What is the most practical way to correct overconfidence?

Replace the feeling of knowing with a measurement. Predict your answer or score before checking, recall from a blank page instead of rereading, explain the idea aloud to someone unfamiliar with it, and repeat the check after a couple of days rather than immediately.

Learning Science

The Dunning-Kruger Effect, Explained

Marisol Vega

Psychology & Habits Writer

Sep 17, 2026

Last updated: August 2026

7 min read

TL;DR

The 1999 studies by Justin Kruger and David Dunning found that people who performed worst on a test estimated their performance far too generously, while top performers underrated themselves slightly. The popular version, a mountain of confidence collapsing into a valley of despair, is a cartoon that the original data does not contain. Serious critics argue that part of the pattern is explained by the better-than-average effect and by regression to the mean, and that critique deserves to be taken seriously rather than waved away. What survives is the useful part: your sense of how well you know something is a weak instrument, and the only reliable way to calibrate it is to test yourself before you trust it.

Almost everyone has met the effect through a chart. A jagged line labelled Mount Stupid rockets up on the left, plunges into a Valley of Despair, then climbs a long slope of enlightenment. It is a very satisfying picture, and it appears nowhere in the research it claims to summarise. The actual finding is quieter, more careful and considerably more useful.

Because the cartoon is so widespread, this piece does three things: it says what Kruger and Dunning measured, it takes the statistical objections seriously instead of hiding them, and it ends with what you can practically do about your own miscalibration.

What the 1999 studies actually did

Kruger and Dunning ran a set of experiments in which participants completed tests of humour recognition, logical reasoning and English grammar, then estimated both their raw score and their percentile rank against everyone else. The comparison of estimate to reality is the whole finding.

Two patterns emerged. Participants in the bottom quarter of actual performance placed themselves far higher than they landed, often well above the midpoint. Participants in the top quarter were slightly more accurate but tended to shade downward, underestimating their rank because they assumed the questions had been easy for everyone.

The authors' explanation is the part usually lost in translation. They argued that the skills needed to perform a task well are often the same skills needed to judge performance on that task. If your grammar is shaky, the instrument you would use to notice that it is shaky is itself the shaky grammar. Incompetence is partly self-concealing, not because people are arrogant but because the detector and the ability share a mechanism.

There is a second half often skipped. In a follow-up condition, participants were trained on the skill, retested, and asked to re-estimate. The poor performers' self-assessments improved, not only because they had improved but because they could finally see what a good answer looked like. Competence and self-knowledge grew together.

Where the meme goes wrong

There is no despair valley in the data. Nobody's confidence was measured collapsing over time, because the studies were not longitudinal.
Poor performers did not claim to be experts. They typically placed themselves as somewhat above average, which is overconfidence, not delusion.
The effect is about calibration, the gap between estimate and reality, not about a personality type you can point at across a dinner table.
Using it as an insult inverts the point. The finding applies to you, in every domain where you have not yet been tested.

The statistical critiques, treated fairly

Two serious objections have followed the effect for years, and anyone quoting it should know them.

The first is the better-than-average effect. Across many tasks, most people rate themselves as somewhat above the middle. If nearly everyone guesses roughly the same above-average position, then arithmetic alone guarantees that low performers will overshoot and high performers will undershoot. Some of the Dunning-Kruger pattern may be that general bias meeting the floor and ceiling of the scoring range, rather than a special blindness among the unskilled.

The second is regression to the mean plus measurement noise. Test scores contain error. Anyone sorted into the bottom group by a single test includes people who simply had a bad day, and a noisy estimate paired with a noisy score will produce this shape even in a simulation with no psychological content whatsoever. Critics have demonstrated exactly that with randomly generated data.

Defenders reply that the training condition is difficult to explain by statistics alone, and that the pattern persists across many task types and measurement approaches. The honest position sits in the middle: the mechanism is contested and the size of the true effect is smaller than the meme suggests, while the core observation that self-assessment is poorly calibrated is very well supported by a much larger literature than these two papers.

Why this is really a story about metacognition

Strip away the argument and what remains is a claim about monitoring: your internal sense of how well you know something is a rough instrument, and it is most unreliable exactly where you have least experience. That is the territory of metacognition, the practice of watching and steering your own learning rather than trusting the feeling of understanding.

There is a mirror-image failure at the other end. Genuine experts routinely misjudge how hard their subject is for a beginner, because fluency erases the memory of confusion. That is the curse of knowledge, and it is why the best practitioner in a room is often not the best teacher in it.

How to puncture your own false confidence

The remedy is not humility as a mood. It is a small number of habits that replace a guess with evidence.

1Predict before you check. Write down the score you expect, or the answer you expect, before you look. The size of your error is the only honest measure of your calibration.
2Close the book and produce. Fluent rereading is the main engine of overconfidence, a trap covered in the illusion of knowing. Recall from an empty page instead, which is the practical core of retrieval practice.
3Explain it to someone who does not know it. Every place your explanation goes vague marks a hole your confidence had papered over.
4Space the check. Confidence measured five minutes after studying is nearly meaningless. Test again after two days, when the easy trace has faded.
5Keep an error log. One line per mistake and why it happened. Patterns appear within a fortnight, and they are almost never the patterns you would have guessed.

A disclosure, since we build in this space: MindSnap is our app, and its design leans on this idea deliberately. Lessons run about two minutes and end with a recall question rather than a summary, because a question tells you what you actually hold while a summary only tells you what you recognise. If you want a sharper test of judgment, our Turning Points scenarios drop you into a real historical decision and ask you to choose before revealing what happened, which is a rather humbling exercise for anyone confident about hindsight.

The takeaway worth keeping

Use the Dunning-Kruger effect as a mirror, not a weapon. The version worth carrying is not that other people are foolishly overconfident. It is that in any domain where you have not yet been tested, your estimate of your own competence is an unverified guess, and the cost of checking is one uncomfortable question. Beginners should ask for the test earlier than feels comfortable. Experts should assume the beginner is more lost than they appear. Both corrections come from the same source, which is measuring rather than feeling.