Participants substantially overestimated their performance in all trace conditions, with the largest overestimation under full traces¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q3 · i1–i3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q3 i3 Perceived performance significantly exceeded actual performance in every condition (Answer-only d = .85; Full-trace d = 1.21; Summary-trace d = .97). → Fernandes 2026 q3 i1 The magnitude of overestimation differed significantly by condition, with Full-trace showing the largest bias (M = 2.44) versus Answer-only (M = 1.74). → Fernandes 2026 (2)

Evidence¶

Fernandes 2026¶

Fernandes, D., Buschek, D., Tankelevitch, L., Kosch, T., & Welsch, R. (2026). Explaining Too Much? Understanding How Large Language Model Reasoning Traces Influence Performance and Metacognition. https://arxiv.org/abs/2605.25856 doi:10.48550/arxiv.2605.25856

q3 · i3 · causal · r2

Global performance estimates collected after all ten problems exceeded achieved performance in each condition; the Full-trace overestimation test printed d = 1.21, with Answer-only d = .85 and Summary-trace d = .97 reported alongside.

"InFull-tracecondition, participants estimated 7.91 items correct (𝑆𝐷= 1.71), while achieving 5.46 correct items, also showing significant overestimation,𝑡( 182)= 16.35,𝑝<. 001,𝑑= 1.21."

Fernandes 2026 (2)¶

Fernandes, D., Buschek, D., Tankelevitch, L., Kosch, T., & Welsch, R. (2026). Explaining Too Much? Understanding How Large Language Model Reasoning Traces Influence Performance and Metacognition. https://arxiv.org/abs/2605.25856. doi:10.48550/arxiv.2605.25856

q3 · i1 · causal · r2

Omnibus ANOVA on overestimation (perceived minus achieved) was significant with η² = .021; Tukey-adjusted comparisons showed Full-trace overestimation significantly larger than Answer-only (Mdiff = 0.70, padj = .003), while Summary-trace did not differ from either.

"the magnitude of overestimation differed significantly by condition,𝐹( 2, 556)= 5.83,𝑝=. 003,𝜂2 =. 021."

Discussion¶

Related Claims¶

• Access to AI advice roughly doubled reported confidence in answers in one experiment — the general claim this finding bears on, with the studies behind it
• Full reasoning traces impaired task performance relative to both answer-only and summary-trace conditions in AI-assisted LSAT reasoning — related
• Full reasoning traces worsened calibration bias and Brier scores relative to both other conditions — related
• Hedonic appeal, not trust, carries the indirect path from trace exposure to performance overestimation, consistent with a processing-fluency account — related
• No trace format supported calibrated self-evaluation: perceived-achieved correlations were low and non-significant in all conditions — related
• Reasoning-trace exposure raised trust and hedonic appeal without performance benefits, shifting subjective appraisal of the interaction — related
• Summary traces preserved task performance at the answer-only baseline while elevating subjective appraisal — related
• Estimation errors for partner AI use are directionally biased for specific task types at t2 but not at t3 — related
• Students expect AI to help their test scores, but control students overestimate the effect relative to the experimental estimate — related
• The spotlight effect: people overestimate how much others notice them, across flattering and embarrassing contexts — related
• Non-interactive teaching produces stronger under-confidence (metacognitive bias) than restudy immediately after learning — related
• Students consistently overestimated how much they learned with answer-revealing images — related