Dunning–Kruger Effect — What It Actually Says (and What It Doesn't)
psychology/dunning-kruger-effect · Updated · Tags: psychology, cognitive-bias, learning, critical-thinking
正在用 AI 翻譯成中文，可能需要一分鐘… 以下暫時顯示英文原文。
Dunning–Kruger Effect — What It Actually Says (and What It Doesn't)
TL;DR
- The claim (Kruger & Dunning, 1999): people who are bad at something tend to overestimate how good they are, because the skill needed to do the task well is the same skill needed to notice you're doing it badly.
- What the data showed: the bottom quarter scored around the 12th percentile but put themselves around the 60th. Everyone rated themselves above average. Low scorers were just the furthest off.
- What it did not show: the famous "Mount Stupid" curve. That graph isn't in the paper. It's an internet meme.
- The catch: most of the classic pattern shows up even with random data, because of regression to the mean plus the ordinary "better-than-average" bias. The dramatic version is mostly a statistical artefact. A smaller real effect probably survives in some domains.
- Use it on yourself, not on other people. Calling someone "Dunning–Kruger" in an argument is the most common misuse of the idea.
The original study
Justin Kruger and David Dunning, "Unskilled and Unaware of It", Journal of Personality and Social Psychology, 1999. Cornell undergraduates took tests in three areas: humour, logical reasoning and English grammar. After each test they estimated their own score and their percentile rank against their classmates.
The researchers then split participants into quartiles by actual score and compared each quartile's actual rank with its self-estimate.
| Quartile (by actual score) | Actual percentile | Self-estimated percentile | Error |
|---|---|---|---|
| Bottom 25% | ~10th–12th | ~60th–68th | Overestimate by ~50 points |
| Second | ~37th | ~60s | Overestimate |
| Third | ~62nd | ~high 60s–70s | Roughly right |
| Top 25% | ~86th–90th | ~70s | Underestimate by ~10–20 points |
Figures are approximate, averaged across the paper's studies. The exact numbers vary by task.
What it looked like
This is the shape of the real figures: two lines, one steep and one nearly flat.
100 75 50 25 0
Bottom 2nd 3rd Top
Quartile by actual test score Percentile
Self-estimate Actual
Illustrative. Actual quartile ranks are about 12, 37, 62 and 87 by definition. The self-estimates follow the typical 1999 pattern.
Three things are worth noticing, because most retellings get them wrong:
- The self-estimate line still slopes upward. Low performers didn't think they were better than high performers. They just didn't know by how much they were behind.
- Everyone put themselves above the 50th percentile. Most of the "effect" is the general better-than-average bias, not something special about incompetent people.
- Top performers underestimated their rank. Kruger and Dunning put this down to the false-consensus effect: the test felt easy to them, so they assumed it was easy for everyone else too.
The proposed explanation: the "double burden"
The paper's theory is metacognitive:
The skills that engender competence in a particular domain are often the very same skills necessary to evaluate competence in that domain.
If you don't know the rules of grammar, you can't write grammatically, and you can't tell that your grammar is wrong. You're missing the skill and the ability to see that you're missing it.
Study 4 is the part people forget. The researchers trained some low scorers in logic and then asked them to reassess their earlier test. After training, their self-assessments dropped and became more accurate. Getting better at the skill made them better at judging it. That's the most hopeful finding in the paper.
Background
- The paper opens with McArthur Wheeler, who robbed two Pittsburgh banks in 1995 with lemon juice on his face. He believed it would make him invisible to security cameras, because lemon juice works as invisible ink.
- It quotes Darwin (The Descent of Man, 1871): "Ignorance more frequently begets confidence than does knowledge."
- Dunning and Kruger won the 2000 Ig Nobel Prize in Psychology for the paper.
"Mount Stupid" is not from the paper
The chart people usually share looks like this: confidence spikes early ("Peak of Mount Stupid"), drops into the "Valley of Despair", then climbs the "Slope of Enlightenment" to a "Plateau of Sustainability".
- The paper measured people once, at a single skill level. It didn't follow anyone as they gained experience, so it can't say anything about a confidence journey.
- The x-axis in the paper is test performance, not experience or time.
- The labels come from a much later internet meme. The shape looks borrowed from the Gartner hype cycle.
The curve might be a fair description of how learning feels for some people. It just isn't evidence, and it isn't the Dunning–Kruger effect.
The statistical critique
This is where the idea gets shaky. Since 2002, several researchers have shown that the classic graph appears even when there's no psychological effect at all.
Why random data produces the same graph
Two ingredients are enough:
- Noisy self-assessment. People's guesses about their own rank are only loosely tied to their real rank.
- Everyone assumes they're a bit above average. This is the well-documented better-than-average effect.
Now group people by their actual score and plot each group's average guess:
- The bottom quartile's actual rank is ~12 by definition. Any noisy guess centred around 60 will be far above that.
- The top quartile's actual rank is ~87 by definition. The same noisy guesses will land below it.
- The middle quartiles land close to their guesses.
That's regression to the mean. If you assign self-estimates completely at random, you still get a flat line crossing a steep one, which is the Dunning–Kruger picture. Blair Fix (2022) called this autocorrelation: actual score sits on both axes (it's the x-axis, and it's also inside the "error" on the y-axis), so the two are bound to correlate.
| Paper | Finding |
|---|---|
| Krueger & Mueller (2002) | Regression to the mean plus the better-than-average heuristic explain most of the pattern. Correcting for test reliability shrinks the effect a lot. |
| Burson, Larrick & Klayman (2006) | On hard tasks the pattern flips: everyone underestimates themselves, and top performers are the worst calibrated. The asymmetry depends on task difficulty, not just skill. |
| Nuhfer et al. (2016, 2017) | Random-number simulations reproduce the classic graph. With better measurement, people turn out to be reasonably good at self-assessment, and their accuracy improves with education. |
| Gignac & Zajenkowski (2020) | Using methods that avoid the artefact, they found little to no Dunning–Kruger effect for intelligence. |
What survives
It's not all artefact. The fair summary of the research in 2026:
| Claim | Status |
|---|---|
| Most people rate themselves above average | Robust. One of the most replicated findings in psychology, though weaker in some East Asian samples. |
| Low performers' self-estimates are the least accurate in absolute terms | True, but largely expected, from regression to the mean. |
| Low performers are disproportionately bad at judging themselves, beyond what statistics predict | Contested. Ehrlinger et al. (2008) and Jansen, Rafferty & Griffiths (2021) find a real but smaller effect after controlling for the artefacts. Others find almost none. |
| The cause is a metacognitive "double burden" | Plausible, not proven. The famous graph can't tell this apart from pure noise. |
| Confidence rises, crashes, then recovers as you learn ("Mount Stupid") | Not supported by the original research. |
| Stupid people think they're smarter than smart people | False. Low performers still rate themselves lower than high performers do. |
Dunning argues that the effect holds up after reliability corrections and incentives for accuracy (Ehrlinger et al., 2008). He also stresses that the effect is about everyone, in whatever area they happen to be weak: "The first rule of the Dunning–Kruger club is you don't know you're a member of the Dunning–Kruger club." Critics and defenders mostly agree that the popular version is overstated. They disagree on how much of the rest is real.
Common misreadings
| Misreading | What's actually true |
|---|---|
| "It's about stupid people." | It's about domain-specific skill. A brilliant surgeon can be in the bottom quartile at tax law. Everyone is a novice at most things. |
| "Incompetent people think they're experts." | They think they're somewhat above average, which is what almost everyone thinks. |
| "Confident people are probably incompetent." | Confidence is a weak signal in both directions. It doesn't reliably tell you who is incompetent. |
| "Experts are humble." | Top performers underestimated their rank, not necessarily their score. They assumed the test was easy for others too. |
| "It's a proven law of psychology." | It's a contested finding with a real statistical problem at its core. |
| "It describes the people I disagree with." | It describes you too, in the areas where you're weakest, and by definition you can't tell which areas those are. |
Related effects
| Effect | What it is | How it differs |
|---|---|---|
| Better-than-average effect | Most people rate themselves above the median on desirable traits | The robust core that the Dunning–Kruger graph mostly rests on |
| Overconfidence effect | People's certainty in their answers is higher than their accuracy | About individual answers, not rank against others |
| Illusion of explanatory depth (Rozenblit & Keil, 2002) | You feel you understand how a zip or a toilet works until you try to explain it | A quick self-test for the same blind spot |
| Impostor phenomenon (Clance & Imes, 1978) | Capable people feel like frauds and expect to be found out | Sometimes called the "inverse" of Dunning–Kruger, but it's a separate construct with its own research |
| False-consensus effect | Assuming others think, and find things, the way you do | Kruger and Dunning's explanation for why top performers underrate their rank |
| Curse of knowledge | Experts can't remember what it's like not to know | Why experts misjudge how hard things are for others |
What to actually do with it
On yourself
The effect assumes you can't see your own blind spots from the inside. So the fixes all bring in outside feedback:
- Predict, then check. Before a task, write down your estimate: time, score, whether the approach will work. Compare it with the result afterwards. Calibration is a skill, and it only improves when you keep score.
- Learn a little more before judging. Study 4's lesson: a small amount of real training is what shows you how much there is to know.
- Explain it out loud. If you can't explain something step by step, you don't understand it as well as you feel you do. This is the illusion-of-explanatory-depth test.
- Look for concrete, fast feedback. Tests, code review, a mentor, real users. Vague praise doesn't recalibrate anyone.
- Be most sceptical in areas you've only just picked up. Early familiarity feels a lot like competence.
In software engineering
- Estimates: track estimated vs actual time per ticket. Most engineers find a consistent bias once they measure it.
- Code review, tests and CI are outside feedback at scale. They catch what the author can't see, not just what the author didn't bother to check.
- Fluent AI output can create an illusion of competence. Code or prose that reads right isn't evidence that you, or the model, got it right. Verify it the same way you'd verify a junior's work.
- Hiring and interviews: confidence and polish are weak signals. Use structured, work-sample assessments instead.
On other people
"That's Dunning–Kruger" in an argument is a claim about someone's competence that you haven't shown, dressed up as science. It's also a good way to reveal your own blind spot about how contested the research is. Argue with the claim, not the person's supposed metacognition.
Bottom line
The Dunning–Kruger effect as people usually tell it ("idiots think they're geniuses", "Mount Stupid") isn't what the study found, and the graph behind it is mostly regression to the mean. What holds up is humbler and more useful: nearly everyone overrates themselves, self-assessment is noisy, and it's noisiest where you know the least. The fix is the same either way: measure, get outside feedback, and assume the next blind spot is your own.
Sources
- Kruger & Dunning (1999). Unskilled and unaware of it: How difficulties in recognizing one's own incompetence lead to inflated self-assessments. JPSP, 77(6), 1121–1134
- Krueger & Mueller (2002). Unskilled, unaware, or both? The better-than-average heuristic and statistical regression predict errors in estimates of own performance. JPSP, 82(2), 180–188
- Burson, Larrick & Klayman (2006). Skilled or unskilled, but still unaware of it. JPSP, 90(1), 60–77
- Ehrlinger, Johnson, Banner, Dunning & Kruger (2008). Why the unskilled are unaware: Further explorations of (absent) self-insight among the incompetent. OBHDP, 105(1), 98–121
- Dunning (2011). The Dunning–Kruger effect: On being ignorant of one's own ignorance. Advances in Experimental Social Psychology, 44, 247–296
- Nuhfer et al. (2016). Random number simulations reveal how random noise affects the measurements and graphical portrayals of self-assessed competency. Numeracy, 9(1)
- Gignac & Zajenkowski (2020). The Dunning–Kruger effect is (mostly) a statistical artefact. Intelligence, 80, 101449
- Jansen, Rafferty & Griffiths (2021). A rational model of the Dunning–Kruger effect supports insensitivity to evidence in low performers. Nature Human Behaviour, 5, 756–763
- Blair Fix (2022). The Dunning–Kruger effect is autocorrelation
- Jonathan Jarry (2020). The Dunning–Kruger effect is probably not real — McGill Office for Science and Society
- Dunning–Kruger effect — Wikipedia
Disclaimer: These documents are personal notes. The information may be incomplete, out of date or wrong, so check it before you rely on it.