LLMs simulating survey respondents under AI-assistant instructions report higher mathematics self-efficacy and lower anxiety than under human-shadow conditioning¶

Claim · All claims
Evidence · 1 study · design 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? Under AI-assistant instructions, LLMs project higher self-efficacy and lower mathematics anxiety than under human-shadow conditioning, with MSES peaks shifted rightward and AMAS distributions clustering at lower bounds. → Esposito 2026

Evidence¶

Esposito 2026¶

Esposito, N., Tricarico, A., Porzio, L., Aghazadeh Ardebili, A., & Stella, M. (2026). Math Education Digital Shadows for Investigating Learning with GenAI: Mathematics Performance, Anxiety, and Confidence in LLMs. arXiv preprint. https://arxiv.org/abs/2604.27618 doi:10.48550/arxiv.2604.27618

q2 · i? · design · r2

Task 2 distributional analysis of psychometric scores by prompting mode. The article reports AMAS distributions "tend to cluster towards the lower bounds of the scale" for AI assistants, and the MSEAQ self-efficacy subscale shows a pronounced rightward skew while the anxiety subscale shows a reciprocal leftward skew.

"This divergence is particularly evident in the MSES distributions, where the AI-assistant peaks are markedly shifted to the right relative to the human-shadow data, indicating systematically higher self-efficacy."

Discussion¶

Related Claims¶

• Self-efficacy judgments correspond less well to behavior in learning settings than in performance settings — the general claim this finding bears on, with the studies behind it
• Human-shadow prompting generates broad psychometric variance while AI-assistant responses are uniform — related
• Priming human shadows as math lovers or haters shifts LLM emotional tone about mathematics — related
• LLM digital shadows reproduce the human pattern of positive science perceptions anchored in negative mathematics associations — related
• Demographic and psychometric response prompting produces higher-variance, heavier-tailed originality distributions than a no-context baseline, though still less variable than humans — related
• Censoring appears to affect negative-valence ratings of mathematics-related cues between model variants — related