The common claim that generative AI simply amplifies the Dunning-Kruger effect is too coarse to capture the available evidence. The clearest findings instead suggest that large language model (LLM) use can improve observable output and short-term task performance while degrading metacognitive accuracy and flattening the classic competence-confidence gradient across skill groups. This paper synthesizes evidence from human-AI interaction, learning research, and model evaluation, and proposes the working model of AI-mediated metacognitive decoupling: a widening gap among produced output, underlying understanding, calibration accuracy, and self-assessed ability. This four-variable account better explains overconfidence, over- and under-reliance, crutch effects, and weak transfer than the simpler metaphor of a uniformly steeper Dunning-Kruger curve. The paper concludes with implications for tool design, assessment, and knowledge work.
Figures & tables
Fig. 1: Schematic of key findings from Fernandes et al. [ 1 ] . Actual performance improves with AI assistance, but the self-assessment gap widens. Values are illustrative of the reported pattern ( ≈ +3 point gain vs. ≈ +4 point overestimation).
Fig. 2: Classic Dunning-Kruger pattern (blue) versus AI-mediated decoupling. Under AI assistance, self-assessed competence (orange) is elevated and flattened across skill levels, while actual performance (green dashed) also rises but diverges from self-assessment. Dashed gray: perfect calibration.
Fig. 3: Contrast between in-session gains and later performance without AI, adapted from Bastani et al. [ 9 ] . GPT-Base improves immediate performance but hurts transfer, whereas GPT-Tutor preserves transfer better.
Fig. 4: Schematic of AI-mediated metacognitive decoupling. The left path is fast and visible: AI improves output, and users update self-assessment from that output. The right path is slower: actual performance should improve calibration, but often does so only weakly or with delay. When the two paths diverge, decoupling emerges.