SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems

2026/03/18 by Rima Hazra, Bikram Ghuku, Ilona Marchenko +5 · 1 voice
Computer Science · #cs.CL

paper · pdf · doi:10.48550/arxiv.2603.17373

Abstract

Large language models are rapidly being deployed as AI tutors, yet current evaluation paradigms assess problem-solving accuracy and generic safety in isolation, failing to capture whether a model is simultaneously pedagogically effective and safe across student-tutor interaction. We argue 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. To systematically study this failure mode, we introduce SafeTutors, a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry. SafeTutors is organized around a theoretically grounded risk taxonomy comprising 11 harm dimensions and 48 sub-risks drawn from learning-science literature. We uncover that all models show broad harm; scale doesn't reliably help; and multi-turn dialogue worsens behavior, with pedagogical failures rising from 17.7% to 77.8%. Harms also vary by subject, so mitigations must be discipline-aware, and single-turn "safe/helpful" results can mask systematic tutor failure over extended interaction.

Citations

• AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting
• Beyond Single-Turn: A Survey on Multi-Turn Interactions with Large Language Models
• Unifying AI Tutor Evaluation: An Evaluation Taxonomy for Pedagogical Ability Assessment of LLM-Powered AI Tutors
• SG-Bench: Evaluating LLM Safety Generalization Across Diverse Tasks and Prompt Types
• CoSafe: Evaluating Large Language Model Safety in Multi-Turn Dialogue Coreference
• Safety Arithmetic: A Framework for Test-time Safety Alignment of Language Models by Steering Parameters and Activations
• Great, Now Write an Article About That: The Crescendo Multi-Turn LLM Jailbreak Attack
• How (un)ethical are instruction-centric responses of LLMs? Unveiling the vulnerabilities of safety guardrails to harmful queries
• The Past, Present, and Future of the Cognitive Theory of Multimedia Learning
• MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems
• CAMEL: Communicative Agents for "Mind" Exploration of Large Language Model Society
• Autonomy-supportive teaching: Its malleability, benefits, and potential to improve educational practice
• RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models
• The ICAP Framework: Linking Cognitive Engagement to Active Learning Outcomes
• First principles of instruction
• Students' questions: a potential resource for teaching and learning science
• What teachers say and do to support students' autonomy during a learning activity.
• Reflective Judgment: Theory and Research on the Development of Epistemic Assumptions Through Adulthood
• Becoming a Self-Regulated Learner: An Overview
• Enhancing teaching through constructive alignment
• The Interplay of Evidence and Consequences in the Validation of Performance Assessments
• Accommodation of a scientific conception: Toward a theory of conceptual change
• The Measurement of Observer Agreement for Categorical Data
• THE ROLE OF TUTORING IN PROBLEM SOLVING *
• Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being.
• Metacognition and cognitive monitoring: A new area of cognitive–developmental inquiry.

Discussions

• SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems arxiv.org/abs/2603.17373 "risk is answer over-disclosure, misconception reinforcement, and the abdication of scaffolding" "multi-turn [bsky, 0 points, 0 comments]

Related

• EduGuardBench: A Holistic Benchmark for Evaluating the Pedagogical Fidelity and Adversarial Safety of LLMs as Simulated Teachers
• SafeScientist: Toward Risk-Aware Scientific Discoveries by LLM Agents
• Safety-Aligned Weights Are Not Enough: Refusal-Teacher-Guided Finetuning Enhances Safety and Downstream Performance under Harmful Finetuning Attacks
• Safety-Tuned LLaMAs: Lessons From Improving the Safety of Large Language Models that Follow Instructions
• Sycophancy is an Educational Safety Risk: Why LLM Tutors Need Sycophancy Benchmarks