Microsoft Research finds LLMs show Dunning-Kruger-style overconfidence in coding

burkov · x · 2026-09-30

A Microsoft Research study tests whether LLMs exhibit the Dunning-Kruger Effect in programming: six prominent models were evaluated across 37 programming languages on multiple-choice tasks derived from CodeNet, comparing actual accuracy against self-assessed confidence. As AI systems increasingly collaborate with human developers, understanding gaps between competence and self-perception is critical for knowing when to trust model output.

Original post →

More from coding & agent

coding & agent channel →