Home · Sources · Bastani et al. 2024
Generative AI without guardrails can harm learning: Evidence from high school mathematics
Bastani et al. 2024 · randomised trial · touches 1 claim · used in 1 attack · used in 1 correction to our notes
Read the study itself (doi.org) ↗
Which beliefs it touches
- C2 Answer-giving AI in practice sets a limit on it
The GPT Tutor arm (withholds direct answers) substantially mitigated the harm.
- C2 Answer-giving AI in practice backs it up
Upper-secondary maths RCT: answer-giving GPT Base +48 % in practice, −17 % on the closed-book exam.
Attacks that used it
- C2 attack #1, 27 September 2026: survived
Corrections to our notes that used it
- C2, 26 October 2026: corrected