Generative AI without guardrails can harm learning
A 2025 PNAS field experiment showing that unrestricted GPT-4 access improved assisted practice performance but reduced later unaided performance, while tutor guardrails largely mitigated the harm.
Also known as Bastani et al. 2025 AI learning
In the high-school mathematics experiment, access to a general GPT-style interface improved performance while assistance was available but was followed by worse unaided exam performance than the control group. A more constrained tutor design mitigated the negative learning effect.
The study is strong evidence that assisted task performance and independent learning are not interchangeable measurements. It is limited to its population, subject, intervention, and study period.