At a glance
- Source
- Bastani, Bastani, Sungu, Ge, Kabakcı and Mariman (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. doi:10.1073/pnas.2422633122
- How it was done
- Students were assigned by chance to practise with an ordinary GPT-4 chat interface, with a GPT-4 tutor built to give hints rather than answers, or with no AI. All then sat an exam without AI. Randomised experiment. People were assigned by chance to work with or without the tool, and the two groups were compared.
- Who took part
- Nearly 1,000 high-school mathematics students.
- Where
- High schools in Turkey.
“unfettered access to GPT-4 can harm educational outcomes.”
What it found
The study had three groups. One practised with GPT-4 in an ordinary chat interface, like ChatGPT. One practised with a GPT-4 tutor set up to give hints and guidance, not finished answers. One practised without AI.
- With the tool. In practice sessions with AI, grades rose 48% with the ordinary chat interface and 127% with the hints-only tutor, compared with the group without AI.
- Without the tool. On the exam, taken without AI, the ordinary-chat group scored 17% below the group that never had AI.
- The hints-only tutor. Students who practised with it scored about the same on the exam as the group without AI. The guardrails largely removed the harm.
The authors’ reading is that students with the ordinary chat interface used it to get answers and did not learn the method well enough to do it alone.
When the result holds
A large gain with the tool can sit on top of a loss in what people can do without it. Studies that measure only work done with the tool cannot see this. Workplace studies almost never test people without the tool afterwards.
The result also points to a remedy: a tool limited to hints and explanations kept most of the help without the loss.
Limits
- The participants were school students, and the outcome was a school exam. No study has run the same test with employees learning a job.
- The tools were GPT-4-based tutors of the time.
Practices that use this study
This study supports:
- Train staff on the tools you actually chose Show every user what your chosen AI tool is, where it goes wrong, and what they must check. Use your own tasks and your own rules, give people time to practise, and include temporary staff. Do first · Effort: medium (weeks, a few people)
- Check now and then what staff can do without the AI On a sample of cases, have people do the work without the AI and compare the result with their usual work. Start with the people who check AI output and the people still learning the job, and set the tools that learners use to give hints, not answers. Do later · Effort: medium (weeks, a few people)
- Keep the entry-level tasks that newcomers learn on Before you automate routine work, check whether it is how juniors learn the job. If it is, keep some of it for them, or replace the learning another way. Do later · Effort: medium (weeks, a few people)
This study limits or contradicts: