Censoring appears to affect negative-valence ratings of mathematics-related cues between model variants¶

Claim · All claims
Evidence · 1 study · associational r1 · 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? The censored Mistral Small 3.2 shows a lower proportion of negative-valence associations than its uncensored counterpart Anita 24B, suggesting censoring might affect valence ratings. → 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? · associational · r1

Task 3 comparison of aggregated valence distributions between a censored model and its uncensored fine-tune relative. The article reports the negative-valence proportions of approximately 35% versus approximately 52% and hedges the interpretation with "suggesting that censoring might affect the valence ratings."

"The bar chart in Fig. 4c shows that Mistral Small 3.2 has a negative-valence proportion of approximately 35%, whereas Anita 24B (Uncensored) reaches approximately 52%, suggesting that censoring might affect the valence ratings associated with the selected keywords."

Discussion¶

Related Claims¶

• LLM digital shadows reproduce the human pattern of positive science perceptions anchored in negative mathematics associations — related
• LLMs simulating survey respondents under AI-assistant instructions report higher mathematics self-efficacy and lower anxiety than under human-shadow conditioning — related
• Priming human shadows as math lovers or haters shifts LLM emotional tone about mathematics — related