i

Description

Crête rocheuse menant au sommet d'Electric Peak, avec un randonneur isolé

Source

Jacob W. Frank (National Park Service) · original page

Licence

PD Public domain

Original photo, unretouched. The background is a version of it stylised by the site.

Dunning-Kruger and “Mount Stupid”

Why do the least competent people misjudge their own level so badly, and what remains of this idea after twenty-five years of debate?

In brief

“Mount Stupid” is the popular nickname for a confidence/knowledge curve associated with the Dunning-Kruger effect.

• The original study (1999): 4 studies; the lowest-performing people overestimate their rank (≈ actual 12^th percentile, ≈ estimated 62^nd); proposed explanation: the “double burden” of incompetence.
• The mountain-shaped curve is not in the paper: it is a popularization.
• The debate: the effect may be partly a statistical artifact; replications (2021) confirm a real core, but its size is still debated.
• Key takeaway: we judge poorly what we have poor command of. Countermeasures: external feedback, measurement, peers, tests on edge cases.

Go to the scientific debate →

On this page

1. The idea in two minutes
2. Where “Mount Stupid” comes from
3. The original study (Kruger & Dunning, 1999)
   1. Main results
   2. The original figure
4. The proposed mechanism: the “double burden”
5. A famous case: the lemon-juice robber
   1. McArthur Wheeler, Pittsburgh January 6, 1995
6. What the effect does not say
7. Critiques and reexaminations
   1. The statistical artifact hypothesis
   2. Responses
8. The “Mount Stupid” curve: how valid is it?
   1. Applied to learning: actual level versus perceived level
   2. How to read it
9. With AI, the bias takes on its full meaning
   1. What the studies show
   2. Calibrating yourself on your own curve
10. Why this remains useful
11. Protecting yourself: a method
12. Connection with maturity matrices
13. With AI: a viewpoint, use cases, and prompts
    1. Detect
    2. Fix
    3. Mitigate
    4. Prompts to integrate into your workflows
14. Glossary
15. Sources and verification
16. See also
    1. Combines with, or reinforces
    2. Related biases
    3. Counterweights and related reflexes

One sentence sums up the idea: to judge a skill well, you must already possess part of it. The rest of this page details what we actually know, what is debated, and what belongs to legend.

The idea in two minutes#

1. At first, you don't know what you don't know: the field seems simple, you only see the surface.
2. As you move forward, you discover complexity, edge cases, exceptions: confidence drops.
3. With experience, confidence rises again, this time grounded in evidence and awareness of your limits.

The difficulty comes at step 1: the tool that would let you detect your errors (skill) is precisely what is missing.

Where “Mount Stupid” comes from#

Illustration inspired by the person

Fictitious illustration

The expression refers to a curve widely shared on the internet: confidence on the y-axis, knowledge (or experience) on the x-axis. It shows a very early confidence peak (“Mount Stupid”), followed by a drop (“valley of despair”), a rise (“slope of enlightenment”), then a plateau.

The original study (Kruger & Dunning, 1999)#

Paper

Kruger, J., & Dunning, D. (1999). Unskilled and Unaware of It: How Difficulties in Recognizing One's Own Incompetence Lead to Inflated Self-Assessments. Journal of Personality and Social Psychology, 77(6), 1121–1134.

Identifier

doi:10.1037/0022-3514.77.6.1121

Protocol

Four studies conducted with Cornell University students, in three domains: humor, logical reasoning, grammar.

Measurement

Each participant takes a test, then estimates their own performance and relative rank (in percentiles) compared with others.

Main results#

• Participants in the bottom quartile greatly overestimate their performance and rank: in the paper's abstract, they are actually around the 12^th percentile but estimate themselves around the 62^nd.
• Participants in the top quartile tend to slightly underestimate their relative rank; the authors explain this with a “false consensus” bias (they assume others are as competent as they are).
• In the final study, training participants in logical reasoning improves the calibration of their self-assessment: acquiring the skill helps measure one's own limits.

The original figure#

The paper's graph compares, by quartile of actual performance, perceived rank and actual rank. It contains neither a “mountain” nor a “valley.”

The proposed mechanism: the “double burden”#

The authors suggest that low-competence individuals suffer from a double handicap: they make more errors, and they lack the metacognitive skills needed to notice them. Producing a correct answer and recognizing a correct answer rely largely on the same knowledge.

A famous case: the lemon-juice robber#

What the effect does not say#

• It's not about general intelligence. The effect is domain-specific: you can be excellent in one area and poorly calibrated in another.
• It's not a judgment about people. Everyone is affected as soon as they step outside their area of expertise.
• It's not limited to absolute beginners. Calibration gaps are measurable across the whole spectrum; they are simply largest at the bottom.
• The average person is not “stupid.” Self-assessments are correlated with performance, imperfectly.

Critiques and reexaminations#

The statistical artifact hypothesis#

Several works argue that the observed pattern can appear without any specific metacognitive mechanism, through the combined effect of two classic phenomena:

• regression to the mean: when two imperfectly correlated measures are compared, the extremes of one are less extreme on the other;
• the “better-than-average” effect: most people judge themselves a little above average.

Simulations with random numbers (Nuhfer et al., 2016 and subsequent work) show that a similar pattern can emerge. Gignac and Zajenkowski (2020), on 929 people (self-assessment of intelligence and Raven matrices test), did not find the expected statistical signs of a Dunning-Kruger effect, hence their paper title: “…is (mostly) a statistical artefact”.

Responses#

• Jansen, Rafferty & Griffiths (2021) propose a rational model of self-assessment and run two large replications (about 4,000 participants each, grammar and logical reasoning). Their results support that low performers are less sensitive to cues from their own errors.
• Mazor & Fleming (2021), in a commentary published in the same issue of Nature Human Behaviour, place this work within the debate on the effect's origin, which remains open.

The “Mount Stupid” curve: how valid is it?#

Practical conclusion: using the curve as a metaphor is legitimate; using it as proof is not.

Applied to learning: actual level versus perceived level#

The curve becomes interesting when applied to the learning of one person over time, not to people at different levels at a given moment (which is what the 1999 study measures). Two lines are plotted: the actual level (what one can really do, theoretical) and the perceived level (what one believes one can do). The colored area between them is the self-assessment gap.

The actual level follows the classic shape of learning curves: rapid gains at first, then increasingly slower gains (the “power law of practice” described by Newell and Rosenbloom in 1981 across very different tasks). The perceived level, however, does not follow the same path: it rises quickly with early successes, then corrects when complexity appears.

How to read it#

• The confidence peak occurs while the actual level is still low: early successes are real, but they involve easy cases.
• The dip is the moment when complexity appears (an error, an edge case, an incident): confidence drops, sometimes below the actual level.
• Recalibration is the useful phase: perception is adjusted through feedback (results, reviews, measurements). Without feedback, the gap persists.
• What changes from one field to another is not the shape, but the speed of feedback: driving penalizes quickly, a new job more slowly, AI sometimes never.

With AI, the bias takes on its full meaning#

For many people, using AI is a learning process like any other, and it has its curve: you discover it, you get excited, you hit a first error, you learn to verify, you find what works and what does not. This is precisely the context where the “Mount Stupid” metaphor is most useful, because several conditions that sustain the perception gap are brought together.

What the studies show#

• METR (Becker et al., 2025): 16 experienced developers, 246 tasks on their own repositories, randomly assigned with or without AI. Result: about 19% more time with AI, while developers estimated afterward that they had gained about 20%. The study covers tools from early 2025 and developers very familiar with their code: it does not say AI always slows things down, it shows that perception can be the opposite of measurement.
• Fernandes et al. (Aalto, Computers in Human Behavior): on logical reasoning problems, model assistance improves scores; participants still overestimate themselves by about 4 points, and the usual Dunning-Kruger relationship fades: the more comfortable people say they are with AI, the more they overestimate themselves.
• Dell'Acqua et al. (2023): on in-frontier tasks, consultants with GPT-4 completed about 12% more tasks, about 25% faster, with quality rated about 40% higher; outside the frontier, they were less often correct.
• Brynjolfsson, Li and Raymond (2023): across more than 5,000 support agents, the assistant delivers an average 14% productivity gain, 34% for novices and very little for experts: AI can also help people move up the experience curve faster. So the issue is not to avoid it, but to measure what it brings.

Calibrating yourself on your own curve#

1. Estimate before, measure after. Write down the expected gain for a task with AI, time it, compare. The difference between the two is your perception gap.
2. Build test cases, including cases outside the area where AI usually succeeds (rare data, business rules, plausible errors): that's where the frontier lies.
3. Keep independent feedback: peer review, automated testing, acceptance testing: AI should not be the only reviewer of its own work.
4. Ask for the opposite: have AI produce arguments against its own answer (the prompts in the “With AI” section below are designed for this).
5. Reassess with every tool or model change: the curve partly restarts, and confidence should restart with it.

Why this remains useful#

• As a rule of humility: the newer a domain is to me, the more I should doubt my own evaluation.
• As an argument for external feedback: peer reviews, tests, code reviews, audits.
• As an explanation for team practices: progression paths, skills assessments, frameworks (see Understanding maturity matrices).

Protecting yourself: a method#

1. Break down competence into precise subskills and assess each one separately.
2. Externalize judgment: have others review, test, and audit, rather than self-assessing.
3. Predict, then compare: write down the expected result in advance (timeline, quality, risk), then measure the gap. Calibration is trainable.
4. Look for edge cases: if you cannot name three situations where your method fails, you do not know it well enough yet.
5. Ask someone more experienced to challenge the reasoning, not just the result.
6. Regularly revisit past decisions to measure their real effects.

Connection with maturity matrices#

A maturity matrix is often filled out through self-assessment. That is exactly where the bias appears: without evidence, declared levels tend to be overly optimistic. For this reason, good assessment practices require verifiable evidence for each level and, if possible, an external perspective. See the limits of matrices.

With AI: a viewpoint, use cases, and prompts#

What follows is a viewpoint, not an established research finding. An assistant quickly produces a result that looks solid: the feeling of mastery may come before real competence. A few safeguards consistent with the above:

• do not confuse getting an answer with knowing how to evaluate it;
• keep tests and checks independent of the assistant;
• learn enough of the domain to detect plausible errors;
• retain responsibility for the final decision.

Prompts to integrate into your workflows#

1. Calibration test.

Quiz me on [domain] with ten questions of increasing difficulty, one at a time.
After each answer, ask me for my confidence level from 0 to 100% BEFORE telling me whether I'm correct.
At the end, give me a table: stated confidence, result, and my calibration gap (am I overconfident, or not confident enough?).

2. Explain to find your gaps.

I will explain [concept] as if to a beginner. Play the role of a demanding student: ask me one question at a time, point out every place where my explanation is vague, incomplete, or wrong, and never fill in the gaps for me.
At the end, list what I wasn't able to explain.

3. Challenge an AI answer itself.

Here is a solution (code, calculation, or reasoning) you just produced: [solution].
Now play devil's advocate: look for plausible errors, cases where it fails, and unverified assumptions. For each point, tell me how I could verify it myself, without taking your word for it.

Glossary#

Metacognition

Knowledge and control of one's own thinking processes, including the ability to judge the quality of one's answers.

Calibration

Alignment between expressed confidence and actual performance.

Regression toward the mean

Tendency for extreme values to be followed by values closer to the mean when measurement is noisy.

Centile (percentile)

Position in a ranking: the 62^nd centile means “better than 62% of the group.”

Sources and verification#

• Becker, Rush, Barnes & Rein (2025), “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity,” METR / arXiv 2507.09089 — controlled trial, 16 developers, 246 tasks
• Brynjolfsson, Li & Raymond (2023), “Generative AI at Work,” NBER w31161 — 5,179 support agents
• Dell’Acqua et al. (2023), “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality,” Harvard Business School Working Paper 24-013
• Mayhew, Simpson & Pak (2003), Accident Analysis & Prevention 35(5), 683-691 — novice driver collisions
• Deery (1999), “Hazard and risk perception among young novice drivers,” Journal of Safety Research 30(4)
• Boswell, Shipp, Payne & Culbertson (2009), “Changes in newcomer job satisfaction over time: examining the pattern of honeymoons and hangovers,” Journal of Applied Psychology 94(4), 844-858
• Newell & Rosenbloom (1981), “Mechanisms of skill acquisition and the law of practice,” in J. R. Anderson (ed.), Cognitive Skills and Their Acquisition, Erlbaum
• Kruger & Dunning (1999), JPSP 77(6), 1121–1134 — original article
• Gignac & Zajenkowski (2020), “The Dunning-Kruger effect is (mostly) a statistical artifact: Valid approaches to testing the hypothesis with individual differences data,” Intelligence, 80, 101449
• Nuhfer et al. (2016), “Random Number Simulations Reveal How Random Noise Affects the Measurements and Graphical Portrayals of Self-Assessed Competency,” Numeracy 9(1)
• Jansen, Rafferty & Griffiths (2021), Nature Human Behaviour 5, 756–763 — rational model and replications
• Mazor & Fleming (2021), “The Dunning–Kruger effect revisited,” Nature Human Behaviour 5, 677–678 — commentary
• Fernandes et al. (2026), “AI makes you smarter but none the wiser: The disconnect between performance and metacognition,” Computers in Human Behavior, 175, 108779 — Aalto University study, published online in October 2025 (preliminary version: arXiv 2409.16708).
• Wikipedia, “1995 Greater Pittsburgh bank robberies” — the McArthur Wheeler robberies.

See also#