AI Tutor Safety and Pedagogical Harms

📄 Full text: arXiv:2603.17373 · local

"Solving problems correctly and avoiding toxic language does not make a tutor safe. Tutoring-specific harm is qualitatively different." SafeTutors exposes that all tested models show broad pedagogical harm, with failures escalating from 17.7% in single-turn to 77.8% in multi-turn student-tutor dialogue.^hazra-safetutors-pedagogical-safety-2026

Why Tutoring Safety Is Different

Conventional LLM safety benchmarks focus on toxic outputs, jailbreaks, and bias. In education, the primary risks are quieter:

• Answer over-disclosure — Revealing solutions rather than facilitating guided discovery

• Misconception reinforcement — Validating or ignoring student misunderstandings

• Abdication of scaffolding — Failing to provide appropriate structured support

• Erosion of productive struggle — Short-circuiting the cognitive work that consolidates understanding

These harms appear "helpful" to surface inspection: the student gets a correct answer quickly. But the long-term effect is learning atrophy.

The SafeTutors Risk Taxonomy

Hazra et al. (2026) derive 11 harm dimensions and 48 sub-risks from learning-science literature:

Critical Findings

1. Universal harm: All 11 tested models (3.8B–72B open-weight + GPT-5-mini) exhibited broad pedagogical harm 2. Scale is not a fix: Larger models were not reliably safer; raw helpfulness correlates weakly with pedagogical safety 3. Multi-turn degradation: Harm rates rose from 17.7% (single-turn) to 77.8% (multi-turn), showing that sustained tutoring interaction progressively erodes safety 4. Discipline-aware mitigations needed: Harms varied significantly across math, physics, and chemistry 5. Single-turn evaluation is misleading: "Safe" single-turn responses masked systematic failure when conversations extended to 5–8 turns

Relationship to Broader Debates

• tutoring-specific-vs-general-ai — SafeTutors reveals that even "helpful" general-purpose AI produces systematic tutoring harm; pedagogical design is not an add-on but a safety requirement

• metacognition — The Metacognitive and Reflective-Critical dimensions directly map to metacognitive suppression risks

• self-regulated-learning — Motivational-Affective harms undermine the SRL↔motivation reciprocal loop

• ai-learning-transfer — Cognitive offloading and shallow learning directly undermine transfer; SafeTutors provides a mechanistic taxonomy for why

• llm-fallacy-misattribution — Fluency illusion (Cognitive dimension) and misrepresentation of understanding (Ethical-Epistemic dimension) are tutoring-specific instantiations of the LLM Fallacy

Implications

• Evaluation: Tutor safety must be measured with multi-turn, discipline-specific benchmarks, not single-turn toxicity screens

• Design: Guardrails must target pedagogical failure modes (over-disclosure, misconception reinforcement) not just content correctness

• Policy: Procurement criteria for educational AI should include pedagogical safety audits alongside accuracy metrics

Related Pages

• xai-education-framework

• ai-lifelong-learning-policy

• ai-pedagogical-accompaniment-amico

• contextual-sycophancy-ai-literacy — The Hidden Cost of Contextual Sycophancy: an AI Literacy Intervention in Human-AI Collaboration

• ai-tutor-authoring-promptdecipher — PromptDecipher: Supporting AI Tutor Authoring Through Editable Simulated Interactions

• llm-tutoring-feedback-diagnosis-gap — Confirming Correct, Missing the Rest: LLM Tutoring Agents Struggle Where Feedback Matters Most

• prompt-injection-defenses-educational-llm-tutors — Prompt injection as a safety threat to educational LLMs

• eduframetrap-llm-sycophancy-educational-safety — Sycophancy identified as educational safety risk

• llm-educational-simulation-adhd — Questions whether safety testing on simulations generalizes to real students

• brookings-ai-students-report — Brookings global report: AI risks overshadow benefits for students; 12 recommendations

• ai-learning-companions-framework — three-foundation framework for AI learning companions prioritizing durable learning over performance

• multimodal-learning-genai — Environmental costs and knowledge costs as expanded harm dimensions

• ai-literacy — Student recognition of pedagogical harms as protective literacy

• principled-ai-education — Augment-not-displace as harm prevention principle

• faculty-development-genai — Institutional neglect of safety in rapid adoption

• ai-tutor-effectiveness-review — Mixed real-world effectiveness as safety signal

• socratic-ai-dialogue — Socratic dependency as a long-term pedagogical harm

• affective-tutoring — Emotional risks and parasocial dependency as harm dimensions

• authentic-assessment — Performative reflection and displaced metacognition as pedagogical harms

• tutoring-specific-vs-general-ai — Pedagogical design vs. general-purpose helpfulness

• metacognition — Metacognitive suppression mechanisms

• self-regulated-learning — Motivation and regulatory erosion

• ai-learning-transfer — How pedagogical harms undermine durable learning

• llm-fallacy-misattribution — Self-perceptual distortions enabled by fluent AI output

• pedagogical-llm-training — Training methods that align models with tutoring safety principles

• educational-llm-alignment — Misalignment between benchmark scores and teaching quality

• educational-vlm-evaluation — VLM failures on pedagogical tasks

• agentic-workflows-education — Interpretability/trustworthiness challenges in multi-agent systems

• neural-symbolic-knowledge-tracing — Interpretable models enable safety auditing

• pedagogical-safety-rl — Four-layer safety framework complements SafeTutors taxonomy

• aied-carbon-footprint-reporting — Environmental cost of LLMs in AIED as an ethical imperative

• shame-guilt-ai-regulation-computing-education -- Shame and guilt create cycles of hiding and moral tension around AI use in computing education

• paternalistic-filter-llm-history-education — Catalogue of tutor harms (2026-07-14)

• eduguard-safe-rag-llm-tutor — EduGuard's verification + overreliance-control modules operationalise the safety-harms taxonomy for programming tutors.

• vetting-dual-llm-safety-education

• llm-children-reading-story-generation — Safety guarantees in children's reading generation- genai-assessment-governance — Governance framework provides structured decisions about when AI use is appropriate

• eduzone-llm-safety-k12 — EduZone: an evaluation framework for LLM safety in K-12 education

📎 1 other page tagged ai-tutor-safety-harms

• EduZone: A Framework for Evaluating LLM Safety for K-12 Students and Teachers