Balancing AI responsibility with privacy, safety, and utility: Unlearning in large language models for mathematics education

Synthesis: Balancing AI responsibility with privacy, safety, and utility: Unlearning in large language models for mathematics education

Key Findings

The study applies gradient-based LLM unlearning to three models — Math-GPT-J, Llama-Lora (Llama-1), and Llama-2-QLora (Llama-2), open-source math-focused LLMs extended from the authors' prior work — pre-trained on approximately 3 million data points from an Algebra I online discussion forum (Algebra Nation) between students and professional tutors.

PII and harmful content were detected on the training data and targeted for unlearning in two different orders (PII-first and harmful-content-first, labeled P→H and H→P), producing two unlearned variants per base model.

The PII forget dataset was dominated by person names (93.28%), reflecting the forum context; the harmful-content forget set was dominated by harassment (10,367 messages), with toxic content the smallest category (861 messages).

Before unlearning, PII-containing output rates were substantial (Llama-1: 17.8%; Llama-2: 15.8%; GPT-J: 15.7%; clean-data baselines 13.9%, 11.5%, 16.5%); after the harmful→PII unlearning order, rates fell to 0.1% for all three models, while the PII→harmful order was somewhat less effective (Llama-1: 1.0%; Llama-2: 1.8%; GPT-J: 0.2%).

After unlearning, the rates of PII-containing output and harmfulness substantially decreased compared to the pre-trained models; on the external RealToxicityPrompts benchmark, harassment, offensive, and toxic output rates fell to 0.0% across all three model families — interpreted as cross-dataset consistency, since baseline rates on that benchmark were already low.

Utility was maintained on both single-label and multi-label downstream math classification tasks: F1 scores for unlearned models remained comparable to pre-trained baselines across training-instance conditions (roughly 0.80–0.84, e.g., Llama-1 pre-trained 0.809 vs. unlearned H→P 0.809 at n = 100).

Sensitivity analyses showed the PII reductions were robust across learning rates and unlearning sample sizes (20K–150K), though larger forget samples did not produce monotonic improvement — effectiveness was not simply a function of using more unlearning examples.

The findings demonstrate a practical path to making LLM-based math tutors more responsible and privacy-preserving while retaining strong performance on math-related tasks.

Study Design & Method

Online mathematics learning platforms increasingly adopt LLMs for scalable, on-demand support, but pre-trained models may reproduce private information from training data or generate harmful language. The study first detects PII and harmful content on the ~3M-point Algebra I tutoring corpus, applies gradient-based unlearning in two orders (harmful→PII and PII→harmful), and then compares the generated outputs of the unlearned models with those of the pre-trained model in terms of PII-containing output rate and harmful rate, including on the external RealToxicityPrompts benchmark for generalization. Finally, the unlearned models are evaluated on two math classification tasks (single-label and multi-label) to confirm that utility survives, alongside learning-rate and unlearning-sample-size sensitivity analyses. The order effect matters theoretically: because the second unlearning stage updates the same parameters again, it can strengthen or partially reverse the first stage's changes, so the optimal sequence may not generalize uniformly across model families.

Implications for AI in Education

For responsible deployment in educational settings, the work shows that post-hoc unlearning is a viable complement to data curation: models already trained on sensitive or unsafe classroom data can be steered toward safer behavior without retraining from scratch. The trade-off between safety and utility is managed rather than eliminated — privacy and harmfulness drop substantially (to 0.1% PII output rates under the harmful→PII order and 0.0% toxic rates on the external benchmark) while math-task performance is maintained. Practitioners should treat unlearning order as a design decision rather than a detail, since the two orders produced different privacy–safety balances. This supports Pedagogical Safety and Privacy-conscious design of LLM-based tutors in K 12 mathematics contexts, where student data are especially sensitive, and connects to the broader AI Governance Education conversation about model-level mitigations alongside data curation and deployment policy.

Limitations

The findings are limited to the Algebra I context in which unlearning was applied and evaluated, so results and procedures may not transfer directly to other subject areas, grade levels, learning platforms, or deployment contexts where PII-like and harmful output risks differ. The PII-containing output rate captures classifier-detected PII-like spans in generated responses and should not be interpreted as direct evidence of memorized training-data leakage — exact or approximate target matching has its own reliability problems, so reductions are best read as decreases in PII-like generation under the study's configuration. Only gradient-based unlearning was examined; other unlearning approaches remain to be benchmarked in education, and future work should add extraction tests, membership-inference evaluations, and target-specific reproduction analyses.

Connected Concepts

Pedagogical Safety

Reinforcement Learning

AI Governance Education

Connected Articles

Cstutorbench Slm Tutors — CSTutorBench: Benchmarking Small Language Models as Tutors for Block-Based Programming

Child Safety GenAI — Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework

AI Tutor Safety Harms — AI Tutor Safety and Pedagogical Harms

Hazra Safetutors Pedagogical Safety 2026 — SafeTutors: Pedagogical Safety in AI Tutoring

Vetting Dual LLM Safety Education — VETTING: A dual-LLM framework for in-loop safety verification via policy isolation in educational AI

Knowledge Distillation AI Tutor Evaluation — Knowledge Distillation for Automated AI Tutor Evaluation

Citation

Li, C., Gülfidan, G., & Zhang-Kopf, Y. (2026). Balancing AI responsibility with privacy, safety, and utility: Unlearning in large language models for mathematics education.