SafeTutors is a benchmark that jointly evaluates safety and pedagogy in AI tutoring systems across mathematics, physics, and chemistry. It argues that tutoring safety is fundamentally different from conventional LLM safety: the primary risk is not toxic content but the quiet erosion of learning through answer over-disclosure, misconception reinforcement, and the abdication of scaffolding.
Authors: Rima Hazra, Bikram Ghuku, Ilona Marchenko, Yaroslava Tokarieva, Sayan Layek, Somnath Banerjee, Julia Stoyanovich, Mykola Pechenizkiy · arXiv:2603.17373 · 3,135 single-turn + 2,820 multi-turn instances · 11 models tested
Key Findings
1. Universal harm across all models. Every evaluated model — from 3.8B open-weight models to GPT-5-mini — exhibited broad pedagogical harm. No model was consistently safe across all subjects and interaction modes, indicating that tutoring safety is not solved by general capability improvements.
2. Scale does not reliably improve safety. Increasing model size did not produce consistent improvements in pedagogical safety, challenging the assumption that larger models are inherently better tutors. This finding parallels broader critiques in llm-evaluation that general benchmarks do not capture domain-specific safety requirements.
3. Multi-turn dialogue dramatically worsens behavior. Pedagogical failure rates escalate from 17.7% in single-turn interactions to 77.8% in multi-turn conversations. The crescendo-based escalation design reveals that models which appear safe in one-turn evaluations systematically degrade across sustained interaction — single-turn "safe/helpful" results mask systematic tutor failure.
4. Harms are subject-dependent. Violation patterns vary significantly across mathematics, physics, and chemistry, indicating that mitigations must be discipline-aware. A tutoring safety strategy that works for math may not transfer to science domains.
5. An 11-dimension, 48-sub-risk taxonomy grounds the evaluation. SafeTutors' risk taxonomy spans Cognitive, Epistemic, Metacognitive, Motivational-Affective, Developmental & Equity, Instructional Alignment, Behavioral & Inquiry, Ethical-Epistemic Integrity, Informational-Semantic, Reflective-Critical, and Pedagogical Relationship dimensions — each with multiple sub-risks drawn from learning-science literature.
Implications
SafeTutors fundamentally reframes the conversation around pedagogical-safety and ai-tutor-safety-harms. The dominant paradigm has been to evaluate AI tutors on problem-solving accuracy and generic safety (toxicity, refusal), but SafeTutors demonstrates that a tutor can be technically accurate and "safe" by conventional metrics while systematically undermining learning. The benchmark's central insight — that tutoring harm is qualitatively different from content harm — has major implications for ai-tutoring regulation and deployment.
The multi-turn degradation finding is particularly alarming for real-world deployment. Most tutoring interactions extend over multiple turns, yet the evaluation community has largely relied on single-turn benchmarks. SafeTutors provides evidence that this practice is dangerously misleading. Systems like eduzone-llm-safety-k12 and vetting-dual-llm-safety-education that prioritize multi-turn safety evaluation are essential, not optional.
The risk taxonomy itself is a significant contribution, providing a theoretically grounded vocabulary for discussing tutoring harm. It bridges educational-theory and AI safety, enabling researchers to move beyond vague claims about "tutor quality" toward precise identification of specific failure modes. This taxonomy could inform the design of pedagogical-safety-rl approaches like singh-eduqwen-pedagogical-rl-2026 that train models to avoid specific pedagogical harms.
For k-12 contexts, where the stakes of pedagogical harm are highest, SafeTutors provides empirical evidence that current models are not safe enough for unsupervised deployment. The subject-dependence of harms suggests that safety evaluation must be integrated into discipline-specific ai-tutor-behavioral-evaluation pipelines rather than treated as a one-time gate.
Related Pages
- pedagogical-safety — Foundational concept of safety specific to teaching contexts
- ai-tutor-safety-harms — Taxonomy and analysis of AI tutoring harms
- ai-tutoring — Broader landscape of AI-based tutoring systems
- llm-evaluation — Frameworks and methods for evaluating LLM behavior
- eduzone-llm-safety-k12 — Safety benchmarks for K-12 educational LLM deployment
- vetting-dual-llm-safety-education — Dual-LLM approaches to educational safety vetting
- k-12 — Educational context with highest-stakes safety requirements
- educational-theory — Theoretical foundations informing the SafeTutors taxonomy
- ai-tutor-behavioral-evaluation — Behavioral evaluation frameworks for AI tutors