An AI assistant can make a difficult exercise feel easy. That is useful when you need an explanation, but it leaves a separate question unanswered: could you solve a different version tomorrow, without the assistant?
If your goal is learning, judge AI by what you can do after closing the chat. A finished worksheet is evidence of completed work. It is not, by itself, evidence that you acquired the skill. A better study routine gives the assistant a limited job: reveal your misunderstanding, offer a small hint and then get out of the way while you try again.
The experiment that separates help from learning
A 2025 paper in PNAS examined nearly 1,000 high-school mathematics students in Turkey. Researchers compared ordinary resources with two GPT-4-based assistants: a relatively unrestricted interface and a tutor designed with teacher-informed safeguards. Both AI groups did better during assisted practice. On a subsequent unaided assessment, however, the unrestricted group performed worse than the group without AI. The safeguarded tutor largely avoided that downside, without establishing an improvement in unaided performance. Read the study.
This was a particular educational setting, subject and tool design. It does not establish that every chatbot harms learning, or that the same effect occurs in every age group. It gives a concrete reason to distinguish assistance during a task from the ability retained afterwards.
The question for a student is therefore practical: what part of the thinking are you still doing? If the assistant chooses the method, performs every step and checks the answer, very little remains for you to practise.
Give the assistant a smaller job
Start by writing what you understand and where you are stuck. Then ask for help at that boundary. Instead of “solve this”, try: “I reached this step. Point out the first mistake, explain the relevant rule, and let me attempt the correction.”
For example, imagine that you are learning percentages. You calculate a 20% discount on an £80 item correctly, then assume that a later 20% price increase returns it to £80. Ask the assistant to identify which quantity the second percentage applies to. Do the arithmetic yourself: the discounted price is £64, and increasing that by 20% produces £76.80. The useful insight concerns the changing base, not memorising this answer.
This is an illustrative exercise, not a reported student outcome. Its purpose is to make the division of work visible. You identify the quantities, perform the calculation and explain the result; the assistant supplies targeted feedback.
Use a three-part study session
Try the following routine on a short topic before applying it to a whole course. Treat the timings as a starting point you can adjust, rather than an experimentally validated formula.
Attempt: spend a few minutes working without AI. Save the attempt, including uncertainty and mistakes. This gives the conversation something specific to address.
Discuss: ask for one hint or a question that exposes the next step. If the explanation introduces unfamiliar terms, ask for those terms to be unpacked. Check important rules against your course materials, especially if the assistant and your teacher use different conventions.
Rebuild: close the conversation and solve a related problem. Change the numbers, context or wording so that copying the previous answer cannot carry you through. Explain your method aloud or in writing.
Keep the third part even when the discussion felt convincing. It is the part that tests whether the explanation became usable knowledge.
Make your progress observable
A useful study log needs only three entries: the topic, the mistake you made and the next problem you solved independently. “Spent an hour with AI” says little about learning. “Confused percentage changes with percentage points; correctly explained a new example unaided” tells you what changed.
If you keep failing the independent attempt, reduce the difficulty or ask a teacher for help. More generated explanations are not automatically the answer. You may be missing an earlier concept, misunderstanding the question or receiving inconsistent feedback.
For writing assignments, apply the same principle differently. Draft your argument first, then ask the assistant which step is unclear. Rewrite that passage yourself. You should be able to explain why the evidence supports the conclusion without reading the assistant's suggested wording.
Keep the educational rules in view
Permission matters as well as usefulness. Check the assignment's rules before using AI, and disclose assistance when required. UNESCO's guidance on generative AI in education places human agency, privacy and appropriate educational design at the centre of adoption. It is a framework for responsible use, not proof that any particular study prompt works. UNESCO guidance.
Avoid uploading classmates' work or unnecessary personal information. Use a short invented example when that is enough to ask your question.
For your next study session, choose one concept and one independent test. Ask the assistant to help you understand the concept, then close it before attempting the test. The most useful result is something you can now explain and do yourself.
Sources & further reading
- Read the study — checked 2026-09-29
- UNESCO guidance — checked 2026-09-29
