Informational harm is the sole dimension that decreases in multi-turn interaction, dropping from 60.52% to 11.90%¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Informational harm drops from 60.52% in single-turn to 11.90% in multi-turn, a decrease of 48.62%, unlike all other dimensions. → 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

In the multi-turn amplification analysis, Informational harm is the only dimension that decreases, falling "from 60.52% in single-turn to 11.90% in multi-turn". The article attributes this to additional conversational context helping models produce more factually grounded responses; no effect size is printed.

"The sole exception is Informational harm, which drops from 60.52% in single-turn to 11.90% in multi-turn, a decrease of 48.62%."

Discussion¶

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
• A three-week MALA deployment recorded 62 distinct usernames across 128 conversations, with 63% of conversations resolved or partially resolved — related
• 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 — related