Review-attributed evidence: user confidence in model answers exceeds model internal confidence, and explanation design narrows the gap¶

Claim · All claims
Evidence · 1 study · review r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Steyvers et al. found user-attributed confidence systematically exceeded model internal confidence, longer explanations raised confidence without improving accuracy, and adjusting explanations to internal confidence narrowed both the calibration and discrimination gaps. → Muminova 2026

Evidence¶

Muminova 2026¶

Muminova, Z. O., Mamatkulov, D. M., Turokulova, R. I., Rakhmonova, N. T., Shavqiyev, O. S. O., & Zoirova, S. U. (2026). The nature of science in the age of generative artificial intelligence: Epistemic authority, materiality, and a synergetic response. EURASIA Journal of Mathematics, Science and Technology Education, 22(10), em2943. https://doi.org/10.29333/ejmste/19481

q2 · i? · review · r2

The review attributes this experimental finding to Steyvers et al. (2025), reporting it as a measured consequence of training objectives that reward probable continuations. The same experiments showed that adjusting explanations to reflect internal confidence narrowed both the calibration gap and the discrimination gap.

"the confidence users attributed to model answers systematically exceeded the models’ internal confidence, and longer explanations increased user confidence even where they did not improve accuracy."

Discussion¶

Related Claims¶

• Explainability's effect on trust in AI-enabled systems is mixed: it can increase trust and reliance but also raises cognitive load, slows performance, and sometimes does not improve trust — related
• A theoretical article, citing two simulation studies, reports that novice clinical learners overestimated their performance and that their confidence exceeded their accuracy — related
• Task completion with GenAI is a weak proxy for durable competence, with struggling learners at risk of an illusion of competence — related
• No trace format supported calibrated self-evaluation: perceived-achieved correlations were low and non-significant in all conditions — related