README.md

metadata

license: apache-2.0
extra_gated_prompt: >-
  You agree to not use the dataset to conduct experiments that cause harm to
  human subjects.
extra_gated_fields:
  Affiliation: text
  Email: text
  Country: country
  Specific date: date_picker
  I want to use this dataset for:
    type: select
    options:
      - Research
      - Education
      - Business
      - label: Other
        value: other
  Description of purpose: text
  I agree to use this dataset for non-commercial use ONLY: checkbox

---

📚 SafeTutors Benchmarking Pedagogical Safety in AI Tutoring Systems

Rima Hazra  ·  Bikram Ghuku  ·  Ilona Marchenko  ·  Yaroslava Tokarieva  ·  Sayan Layek  ·  Somnath Banerjee  ·  Julia Stoyanovich  ·  Mykola Pechenizkiy

EMNLP 2026 · March 2026

---

🔍 Overview

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 during real student–tutor interaction.

Core Idea: 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.

SafeTutors is a benchmark that jointly evaluates safety and pedagogy across mathematics, physics, and chemistry, grounded in a theoretically motivated risk taxonomy from the learning-science literature.

---

🚨 Key Findings

Finding	Detail
🔴 No model is universally safe	Every evaluated model exceeds 60% harm rate on ≥5 harm categories (single-turn) and ≥6 (multi-turn)
📏 Scale doesn't reliably help	Larger models do not consistently reduce pedagogical harms
💬 Multi-turn dialogue worsens behavior	Pedagogical failure rates rise from 17.7% → 77.8% as conversations extend
🔬 Subject-dependent harms	Harm profiles differ across subjects — mitigations must be discipline-aware
⚠️ Single-turn results are misleading	"Safe/helpful" single-turn scores mask systematic failure over extended interactions

---

Risk Taxonomy

11 Harm Dimensions
└── 48 Sub-risks (grounded in learning-science literature)
    ├── Epistemic harms
    ├── Informational harms
    ├── Instructional harms
    ├── Metacognitive harms
    ├── Reflective harms
    ├── Pedagogical relationship harms
    └── ... (5 more dimensions)

Dataset Statistics

Split	Instances	Construction Method
Single-turn	3,135	Curated student–tutor scenarios
Multi-turn sequences	2,820	Crescendo-based escalation
Total	5,955

Subjects: Mathematics  ·  Physics  ·  Chemistry

Models Evaluated

11 LLMs evaluated (10 open-weight + 1 closed-weight), ranging from 3.8B to 72B parameters.

---

📕 Cite us

If you use SafeTutors in your research, please cite:

@article{hazra2026safetutors,
  title   = {SafeTutors: Benchmarking Pedagogical Safety in AI Tutoring Systems},
  author  = {Hazra, Rima and Ghuku, Bikram and Marchenko, Ilona and
             Tokarieva, Yaroslava and Layek, Sayan and Banerjee, Somnath and
             Stoyanovich, Julia and Pechenizkiy, Mykola},
  journal = {arXiv preprint arXiv:2603.17373},
  year    = {2026},
  url     = {https://arxiv.org/abs/2603.17373}
}

---

📬 Contact

For questions or issues, please open a GitHub Issue or reach out to Rima Hazra via the contact on the arXiv page.

---

📜 License

This project is licensed under the MIT License — see the LICENSE file for details.

---

_{⭐ If you find this work useful, consider starring the repository.}