Status of the Dunning–Kruger effect in human–AI interaction

Determine whether the Dunning–Kruger effect in human–AI interaction reflects a genuine performance-dependent metacognitive phenomenon rather than regression-to-the-mean, measurement-error, or other statistical artifacts.

Background

The paper examines whether lower-performing and higher-performing users differ systematically in the accuracy of their self-assessments when solving reasoning tasks with an AI model. Existing statistical accounts show that similar quartile patterns can arise from regression to the mean, measurement noise, bounded response scales, and constant bias or noise, so the psychological status of the effect remains unsettled.

The study applies disjoint-score controls, measurement-error corrections, simulations, and computational models. These analyses provide evidence that within-group performance-dependent estimation errors persist under the specified controls, but they do not establish a unique metacognitive mechanism or settle the broader status of the Dunning–Kruger effect in human–AI interaction.

References

The smaller within-block contrast suggests an important contribution from cross-task confidence ordering, but does not establish equivalent within-task discrimination.

— Confident, Not Wiser: The Dunning-Kruger Effect in Human-AI Interaction  (2609.31095 - Fernandes et al., 25 Sep 2026) in Section 5, subsection “Metacognitive sensitivity”

The status of the DKE in human--AI interaction remains unresolved.

— Confident, Not Wiser: The Dunning-Kruger Effect in Human-AI Interaction  (2609.31095 - Fernandes et al., 25 Sep 2026) in Section 2, subsection “The Dunning-Kruger Effect”