How to study with AI without outsourcing the learning

Why assisted performance is not the whole story

The OECD Digital Education Outlook 2026 examines generative AI in education and the importance of pedagogical design. The distinction between completing a task with help and developing lasting capability is particularly useful for students trying to choose a study workflow. The report is context, not a guarantee that one prompt will improve everyone's learning.

Set your goal before opening the assistant. Do you need to remember vocabulary, explain a mechanism, solve unfamiliar problems, or critique an argument? Those goals need different practice. A summary is a possible input to learning; it is rarely a complete test of whether learning happened.

Start with retrieval, even when it feels uncomfortable

Close the notes and attempt an answer from memory. Write a rough definition, sketch the causal steps, or solve the first part of a problem. The gaps in that attempt tell you where help is needed. If you read the answer first, you lose some of that diagnostic information because familiar words can make an incomplete understanding feel complete.

Explore the testing effect before planning your next session. Its teaching model helps you compare study strategies; its displayed values are not a forecast of your examination score. For research context, Roediger and Karpicke's study examines how retrieval practice affects later retention in a specific experimental setting. The general lesson is to practise recalling, not merely recognising.

Ask for feedback on your attempt

Give the assistant your own explanation and ask it to identify missing steps, ambiguous claims, and one counterexample. A useful instruction is: 'Do not rewrite the whole answer yet. Ask one question that would reveal whether I understand the mechanism.' Then try to answer before asking for a complete solution.

Check factual feedback against your course material or a reliable source. AI can confidently criticise a correct answer or approve an incorrect one. For mathematics, verify steps; for history, check dates and sources; for a practical skill, test the result. Feedback is valuable because it changes your next attempt, not because it sounds encouraging.

A worked example: learning Bayes' rule

Suppose you understand the words 'prior' and 'evidence' but cannot explain why a rare condition may remain unlikely after a positive test. Use an invented scenario rather than somebody's medical result. Your purpose is to understand conditional probability, not to interpret a real diagnosis.

1. Predict the answer in plain language, then use the base rate experiment to compare your intuition with counts of true and false positives.
2. Explain the result without looking at the formula. Ask the assistant to challenge one step, and check that challenge against the experiment's assumptions.
3. The next day, solve a different numerical example without the assistant. Record which part you still needed help with and practise that part again.

Space the practice around what you forget

The forgetting curve gives you a visual way to think about time and review. It is a simplified model, not a personal memory measurement. Use it as a reason to revisit a topic after a delay rather than assuming that one long session is enough.

Keep a small review list with the concept, the mistake you made, and a fresh question. Review the explanation after trying the question, not before. If the question is always identical, you may learn the wording instead of the idea. Change the surface details while keeping the underlying structure, then explain why the same principle still applies.

Try a simple session you can repeat

Spend a few minutes attempting one question unaided. Use the next part of the session to check errors and request a targeted explanation. Finish by closing everything and writing a short answer to a new question. The timings can vary; the important feature is an independent attempt at both ends.

Try the cognitive reflection experiment when you want to notice how quickly an attractive first answer can appear. Do not treat a game score as a measure of your intelligence. Use the pause between intuition and checking as a study habit: name the assumption that makes an answer seem obvious, then test it.

• Attempt before asking.
• Request targeted feedback instead of a finished submission.
• Verify claims using course or primary materials.
• Return later with a new problem and no assistant.

When a simpler tool is better

A textbook exercise with an answer key may give more dependable feedback than a conversational assistant. A teacher can diagnose a misconception across several attempts and understand the assessment criteria. Use those resources when available. Also follow your institution's rules about permitted assistance; a useful study method is different from submitting somebody else's work.

If the assistant keeps producing material faster than you can evaluate it, reduce the output. One good question is often more useful than twenty pages of notes. A good session leaves evidence of changed capability: a clearer explanation, a solved unfamiliar problem, or a mistake you now know how to avoid. That evidence is worth more than the feeling of having been productive.