Harm profiles are subject-dependent: metacognitive harm crosses 90% in all three subjects in multi-turn, peaking at 97.44% for Qwen2.5-32B in Mathematics¶

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? Multi-turn metacognitive harm exceeds 90% in all three subjects, peaking at 97.44% for Qwen2.5-32B in Mathematics, indicating mitigations must be discipline-aware. → Rima Hazra 2026

Evidence¶

Rima Hazra 2026¶

Rima Hazra, Bikram Ghuku, Ilona Marchenko, Yaroslava Tokarieva, Sayan Layek, Somnath Banerjee, Julia Stoyanovich, Mykola Pechenizkiy. (2026). SAFETUTORS: Benchmarking Pedagogical Safety in AI Tutoring Systems. arXiv preprint. https://arxiv.org/abs/2603.17373 doi:10.48550/arxiv.2603.17373

q2 · i? · design · r2

Multi-turn results across mathematics, physics, and chemistry show metacognitive harm above 90% in every subject, "peaking at 97.44% for Qwen2.5-32B in Mathematics". The article concludes mitigation strategies must be discipline-aware; no effect size is printed.

"Metacognitive harm crosses 90% in all three subjects in multi-turn, peaking at 97.44% for Qwen2.5-32B in Mathematics."

Discussion¶

Learner Variables¶

• Metacognition — outcome: instruction changes it

Related Claims¶

• Multi-turn interaction amplifies pedagogical harm: mean harm rises 6-11 percentage points across subjects and pedagogical-relationship harm surges from 17.7% to 77.8% — related
• No evaluated LLM tutor is reliably safe: every model exceeds 60% harm rate on at least five risk categories in single-turn and six in multi-turn evaluation — related
• LLM teaching effectiveness varies substantially across subject domains, strongest where knowledge points apply directly and weakest in physics and chemistry — related
• Informational harm is the sole dimension that decreases in multi-turn interaction, dropping from 60.52% to 11.90% — related