Estimation errors for partner AI use are directionally biased for specific task types at t2 but not at t3¶

Claim · All claims
Evidence · 1 study · associational r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? At t2, AI use for automatically implementing full features was systematically underestimated (Mdiff = +0.21, p_adj = .049), while use for asking about programming principles (Mdiff = −0.29, p_adj = .029) and confirming an approach's validity (Mdiff = −0.19, p_adj = .049) was overestimated; no task differences reached significance at t3. → Laura Graf 2026

Evidence¶

Laura Graf 2026¶

Laura Graf, Ramona Beinstingel, Stephan Krusche, Oleksandra Poquet. (2026). Students' Perception Accuracy of Partners' AI Use and its Relation to Collaboration Performance. Preprint, Technical University of Munich. https://arxiv.org/abs/2606.23237 doi:10.48550/arxiv.2606.23237

q2 · i? · associational · r2

Wilcoxon signed-rank tests with Benjamini–Hochberg correction on differences between partner self-reports and estimates across ten task types, measured at t2 and t3 in the 103-pair survey. The quoted values are unstandardised five-point-scale mean differences; no standardised effect size is printed.

"AI use for automatically implementing full features was systematically underestimated (Mdiff = +0.21, p_adj = .049), while AI use for asking about programming principles (Mdiff = −0.29, p_adj = .029) and for confirming the validity of an approach (Mdiff = −0.19, p_adj = .049) were overestimated."

Discussion¶

Related Claims¶

• Perception alignment improves mainly in the early weeks and does not consistently decrease through the project — related
• Participants substantially overestimated their performance in all trace conditions, with the largest overestimation under full traces — related