AI Tutor Safety and Pedagogical Harms

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

"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:

Dimension	Core Concern	Key Examples
Cognitive	Interferes with knowledge internalization	Cognitive offloading, fluency illusion, shallow procedural learning
Epistemic	Weakens justification/evaluation ability	Unverified authority, source opaqueness, false consensus
Metacognitive	Erods monitoring and self-reflection	External validation dependence, reflection bypass, learned helplessness
Motivational-Affective	Undermines curiosity and persistence	Shortcut temptation, performance-over-mastery, emotional disengagement
Developmental & Equity	Fails to calibrate to learner level	Cognitive load mismatch, unequal benefit, cultural bias
Instructional Alignment	Departs from learning goals	Pedagogical drift, goal misidentification, hidden curriculum
Behavioral & Inquiry	Enables shortcuts/dishonesty	Answer-seeking bypass, assignment outsourcing
Ethical-Epistemic Integrity	Compromises intellectual ownership	Blurred authorship, misrepresentation of understanding
Informational-Semantic	Embeds factual inaccuracies	Fabrication, misleading scientific explanation
Reflective-Critical	Suppresses evidence-weighing	Over-smooth acceptance, no metacognitive challenge
Pedagogical Relationship	Dysfunctional learner-system dynamic	Over-trust in AI authority, loss of learner agency

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

Transfer Of Learning — 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

Connected Concepts

Metacognition

Self Regulated Learning

Connected Articles

Hazra Safetutors Pedagogical Safety 2026

Tutoring Specific Vs General AI

Transfer Of Learning

LLM Fallacy Misattribution

Citation

Hazra, R., Ghuku, B., Marchenko, I., Tokarieva, Y., Layek, S., Banerjee, S., Stoyanovich, J., & Pechenizkiy, M. (2026). SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems. arXiv:2603.17373.