Assess how much stronger the observed pattern is than patterns produced by an appropriate null model with measurement noise and the same analytical method.
Kruger and Dunning found substantial self-overestimation among low performers and proposed a metacognitive mechanism. Later methodological criticism shows that parts of the familiar quartile pattern can also arise from regression to the mean, measurement error and binning.
Assess how much stronger the observed pattern is than patterns produced by an appropriate null model with measurement noise and the same analytical method.
In the original studies, participants in the lowest performance quartile substantially overestimated their relative performance. The popular internet version often turns this into the much stronger claim that incompetent people, by definition, do not know they are incompetent.
Self-assessment miscalibration is real, but the effect size and canonical quartile shape are contested. Better analyses find varying results, including small residual effects.
A competence-dependent metacognitive deficit competes with artefact explanations such as measurement error, regression, better-than-average effects and mathematical coupling.
AI transfer to investigate: AI may decouple assisted output, one’s own capability and confidence. This is a question about an existing mechanism in a new context. Calling LLM overconfidence itself ‘Dunning–Kruger’ is not justified without a comparable mechanism.
Proposed test, not yet conducted: Compare no AI, direct-answer AI, analytical-support AI and AI with explicit assisted-versus-unaided feedback. Conduct a later unaided transfer test and run the same analysis on artefact-compatible null simulations.
This case concerns both substantive research and methodological discipline: first let the null model try to reproduce your famous effect.
Compare continuous ability and self-assessment with an appropriate null model containing measurement noise and the same analytical method. Gignac & Zajenkowski (2020) and Dunkel et al. (2023) support different bounded conclusions; the debate remains open.
Gignac and Zajenkowski (2020) and Dunkel et al. (2023) report different, bounded findings. Supplements and raw data have not been independently reanalysed here, and the original 1999 studies have not been fully assessed. The AI application remains a hypothesis.
Help improve this case
We review your contribution before updating the text or sources.