Generative AI without guardrails can harm learning: Evidence from high school mathematics
Summary
A mathematics RCT directly separates assisted performance from subsequent unaided assessment.
Classes at one Turkish high school received a basic GPT-4 interface, teacher-designed tutoring safeguards, or no AI. Both AI versions improved practice scores; the basic version reduced subsequent unaided scores by about 17% relative to control.
Safeguards largely removed that harm but did not demonstrate an unaided advantage. Tool-assisted success and learning must remain distinct outcomes.
One school, four review sessions and immediate similar-problem exams; not long-term retention or far transfer.
Sources & reading record
- Bastani et al. / PNAS ↗
- Read abstract, experimental design, main results and Table 1; no code rerun or exhaustive supplement audit.