On this page
Your child has a chemistry worksheet due tomorrow and a chatbot that will produce every answer in seconds. Whether that helps or hurts depends less on the tool than on how it is used. Here is what the research shows, and a set of house rules you can adopt this week.
What happened when students got ChatGPT for maths practice?
The most useful study so far is a field experiment by Hamsa Bastani and colleagues, published in PNAS in 2025. Nearly a thousand high-school maths students got one of two AI helpers during practice sessions: one that worked like standard ChatGPT ("GPT Base"), and one with prompts designed to protect learning ("GPT Tutor").
During practice, both groups did better: grades rose 48% with GPT Base and 127% with GPT Tutor. But when the AI was taken away, students who had used GPT Base did 17% worse than students who never had AI at all. The authors found that, without guardrails, students used it as a "crutch". The GPT Tutor version largely avoided that harm.
A second study points the other way, for good reasons. In a randomised trial at Harvard, Kestin and colleagues (Scientific Reports, 2025) found college physics students learned significantly more in less time from an AI tutor than from an active-learning class. That tutor was custom-built on the same teaching principles as the class. It guided students; it did not hand them solutions.
Put together: AI that does the thinking for a student can lower what they learn, while AI designed to make them think can help. Our explainer on AI tutors versus answer chatbots covers both studies in more depth.
Is my child old enough?
Check the rules first. OpenAI's terms of use say users must be at least 13 (or the minimum age required in their country), and that users under 18 need a parent or legal guardian's permission. UNESCO's 2023 guidance on generative AI in education suggested considering an age limit of 13 for classroom use. Your child's school may also have its own policy on AI in assignments; ask before assuming.
AI that teaches vs AI that answers: a 5-point check
Sit with your child for ten minutes and watch one session. Then ask:
Who wrote the steps? If the AI produced the full solution and your child copied it, that is answering.
Did the AI ask anything back? A teaching tool asks "what do you think the first step is?" before explaining.
Can your child redo it without the screen? Close the laptop and ask them to solve a similar problem. This is the closest home version of the test in the Bastani study.
Did it catch a mistake, or just replace it? Good help points to where the reasoning went wrong.
Is the answer right? Chatbots make confident errors in physics and maths. See how to check an AI's physics solution.
House rules that work
Attempt first. At least 10 minutes on a problem before opening any AI tool.
Ask for hints, not answers. Teach the prompt: "Don't give me the answer. Ask me a question that helps me find the next step."
No AI for graded work unless the school allows it, and say how it was used if so.
AI in shared spaces. Laptops at the dining table for younger teens, not behind closed doors.
Weekly review. Look at one chat history together, not to police, but to see how it is being used.
What to look for in a study tool
The Bastani results suggest the design of the tool matters as much as the rules. Look for tools that hold back the answer, ask the student questions and check understanding. Fermi's AI tutor is built to teach rather than answer: it notices where a student is stuck and decides whether to explain, ask or show, checking understanding as it goes. For worked problems, the Fermi Problem Library has a step-by-step solver that explains the reasoning rather than printing an answer. The closed-laptop test applies to any tool, including ours.
