Short answer: use AI to give hints, check your working and generate fresh practice problems, not to solve the problems you are meant to be practising. In a field experiment with nearly a thousand high school maths students (Bastani and colleagues, 2025), unrestricted GPT-4 access improved practice grades by 48% but left students 17% worse on a later test without AI. A version set up to give hints instead of answers largely avoided that harm.
The study every maths student should know about
Bastani and colleagues published the clearest test so far in PNAS in 2025. They gave nearly a thousand high school maths students access to one of two GPT-4 tutors during practice sessions. "GPT Base" behaved like a standard ChatGPT interface. "GPT Tutor" had prompts designed to safeguard learning, such as giving hints rather than complete answers.
- With AI available, practice grades rose: by 48% with GPT Base and 127% with GPT Tutor.
- When the AI was taken away, students who had used GPT Base did 17% worse than students who never had access.
- The safeguards in GPT Tutor largely mitigated that negative effect.
The authors describe students using unrestricted GPT-4 as a "crutch" during practice and then performing worse on their own. That is the trap: AI makes practice look great, and practice scores are not what you get marked on.
Why AI answers hurt maths learning
Maths is learned by doing: choosing a method, carrying it out, getting stuck and getting unstuck. Each time you retrieve a method from memory and apply it, you strengthen it. Roediger and Karpicke (2006) showed that retrieving information produces much better long-term retention than reviewing it, even though reviewing feels more reassuring at the time.
Reading an AI's worked solution is reviewing. You follow each line, it makes sense, and you feel you understand. Then the exam asks you to produce the first line yourself, with nothing to follow. Our post on the illusion of competence explains why following a solution feels so much like being able to write one.
There is also the choosing problem. In an exam, questions aren't labelled by chapter, so half the skill is recognising which method applies. Rohrer and Taylor (2007) found that mixing different types of maths problems in practice, rather than doing a block of one type, led to better results on a later test. If AI picks the method for you, you never practise that choice.
Five rules for AI maths revision
- Attempt first, always. Write down at least a first step before asking anything. "I don't know where to start" is still worth five minutes of trying.
- Ask for a hint, not a solution. This is the safeguard that made the difference in the Bastani study. Tell the AI explicitly not to give you the answer.
- Use it to check, not to do. When you have a full answer, ask it to find the first line where your working goes wrong. Then fix it yourself.
- Ask for new problems, not more explanations. After getting one wrong, ask for two similar problems with the answers hidden. Solving them is the learning.
- Finish without AI. End each topic with a past-paper question done under exam conditions, no help. That's the only score that tells you where you stand. See how to use past papers.
Prompts that keep you doing the maths
- "I'm revising for an exam. Don't solve this. Give me one hint for the next step only, then wait for my attempt."
- "Here is my full working. Don't correct it. Tell me the number of the first line with an error."
- "Give me three problems like this one, mixed with two from a different topic, without telling me which is which. Answers only when I ask."
- "I keep getting this type wrong. Ask me questions to find out which step I don't understand."
If you want an AI that behaves this way by default, the study modes in ChatGPT, Claude and Gemini are built around guiding rather than answering; we compare them in ChatGPT study mode vs Claude learning mode vs Gemini guided learning.
Check the AI, too
Language models still make arithmetic and reasoning errors, and they explain wrong answers just as confidently as right ones. When an AI tells you your answer is wrong, check its claim against your textbook, a mark scheme or a calculator before accepting it. Practice testing has high utility in the Dunlosky and colleagues (2013) review of study techniques, but only if the answers you check against are right.
For the formula-heavy topics in your own course, Ace turns your notes, slides or PDFs into practice questions written from that material only, and for formula topics it can include an interactive simulation where you set the value and see what changes. Every answer is checked against your file, so you are practising your course, not a generic one. See how it works.
Frequently asked questions
Is it cheating to use AI for maths homework?
That depends on your school's rules for each piece of work. But even where it is allowed, using AI to produce answers for practice work is likely to hurt your exam performance, based on Bastani and colleagues (2025).
Can AI help you learn maths at all?
Yes. In the same study, the version that gave hints instead of answers largely avoided the learning loss. Use AI to hint, check and generate practice problems.
What's the best AI for maths revision?
The one you configure to stop giving answers. A guided study mode or a clear prompt matters more than the model.
How do I revise maths if I keep forgetting methods?
Mix problem types and revisit them over several days rather than doing one topic in a block. Rohrer and Taylor (2007) found mixed practice beat blocked practice on a later test. Our guide to studying maths has more.
Sources
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122.
- Roediger, H. L., & Karpicke, J. D. (2006). Test-enhanced learning: Taking memory tests improves long-term retention. Psychological Science, 17(3), 249–255.
- Rohrer, D., & Taylor, K. (2007). The shuffling of mathematics problems improves learning. Instructional Science, 35(6), 481–498.
- Dunlosky, J., Rawson, K. A., Marsh, E. J., Nathan, M. J., & Willingham, D. T. (2013). Improving students' learning with effective learning techniques. Psychological Science in the Public Interest, 14(1), 4–58.



