If you are 15 to 18 and revising for exams, you already know the trap. If your child is 11 to 14, read this alongside them. Each tool's own terms set a minimum age and say who may open an account, and some give under-13s no route at all. Use only an account the tool's own terms allow for that age, and do not let a child use an adult's account. AI will turn a chapter into a clean summary in seconds. It feels like progress. The summary is there, the notes look tidy, and you have not yet learned anything. This guide is a method first and a tool list second, because the method is what decides whether AI helps.

The method that works: retrieve, then space

The strongest finding in the study of how students learn is also the least pleasant. Trying to recall something is a better way to learn it than looking at it again.

In a well-known set of experiments, students read prose passages and then either re-studied them or tried to recall them from memory. Five minutes later, re-studying gave better recall. Two days later and a week later, the students who had tested themselves remembered substantially more. The students who re-read also felt more confident, without the memory to back it up (Roediger and Karpicke, Psychological Science, 2006). These were lab experiments with students, so treat the size of the effect with caution. The larger review below points the same way.

Re-reading feels better. Testing yourself lasts longer.

  • 5 minutes laterRe-reading did betterRe-reading looked better straight away.
  • 2 days laterTesting yourself did betterSelf-testing was substantially better.
  • 1 week laterTesting yourself did betterThe gap held a week on.

Source: Roediger & Karpicke, “Test-enhanced learning”, Psychological Science (2006); the figure shows which group recalled more, not by how much

A review of ten popular study techniques, by Dunlosky and colleagues, reached the same place at a larger scale. Practice testing and distributed practice were the only two techniques to receive its highest overall utility rating, and both were shown to help across many different kinds of tests (Association for Psychological Science, on Dunlosky et al., 2013). "Distributed practice" means spreading your study over several sessions instead of cramming into one.

Put those together and you get a simple loop.

A study loop where the learner does the remembering

  1. 1

    Close the notes

    Pick one topic. Shut the book.

  2. 2

    Recall

    Write or say everything you remember.

  3. 3

    Get quizzed

    Ask AI for questions, or answer your own cards.

  4. 4

    Check

    Compare with your notes or textbook. Mark the gaps.

  5. 5

    Come back later

    Return to the gaps in a day, then a few days, then a week.

AI can write the questions in step 3. It cannot do steps 1, 2 or 4 for you; those are where the learning is.

Notice where the AI sits in that loop. It can write the questions. It can mark your recall against your notes. It can remind you what to come back to. It cannot do the recall, which is the part that builds the memory.

Where AI helps and where it quietly hurts

There is real evidence that AI can teach well when it is designed to. At Harvard, a randomised trial of an AI tutor built on teaching principles found students scored higher after the AI tutor than after an in-class active-learning session. The median post-test score was 4.5 against 3.5, and the AI group spent a median of 49 minutes on the work (Kestin et al., Scientific Reports, 2025). That was one physics course for university students, and the tutor was purpose-built, so it does not show that any chatbot will do the same for a 14-year-old preparing for GCSE or SPM.

The warning evidence points the other way. In the high-school maths trial published in PNAS, students who practised with an unrestricted GPT-4 interface scored 17% below students who never had AI once access was removed, while a hint-giving version largely avoided the harm (Bastani et al., PNAS, 2025). A separate randomised study of university students writing an essay found the group using ChatGPT improved its essay score most, but its knowledge gain and transfer did not differ significantly from the other groups. The authors warn that it may trigger what they call metacognitive laziness (Fan et al., arXiv, 2024). That is a preprint on adults, so it is a caution, not a rule.

The common thread is simple. When AI does the thinking, the work improves and the learner does not. When it makes the learner think, the learner improves.

Ask it the right way

Most of the difference sits in the first line you type.

Same tool, two ways to ask

  • Test recallLike this: “Ask me one question at a time on this topic. Wait for my answer before you say if it is right.”Not this: “Summarise this chapter for me.”
  • Find gapsLike this: “Here is what I remember. What did I miss or get wrong? Do not rewrite it.”Not this: “Write my notes for me.”
  • UnderstandLike this: “Explain this a second way, then ask me a question to check I followed.”Not this: “Just give me the answer.”
  • PractiseLike this: “Give me five exam-style questions on this topic. Do not show the answers until I have tried.”Not this: “Do these past-paper questions.”
The left column makes you retrieve. The right column lets you skip the retrieval.

A good prompt does three things. It tells the tool to wait for your answer. It tells it to say what you got wrong, not rewrite it. And it keeps the answers hidden until you have tried. You can paste that rule once at the top of a session and the tool may keep to it, but guardrails written in a prompt can slip, so check.

What kind of study tool to look for

You do not need a long list. You need a tool that supports one of the methods above. These are described the way the tools describe themselves, as summarised in our Learning Tools Index.

  • Flashcards with spacing. Gizmo generates quizzes and flashcards from notes, PDFs, slides and YouTube inside a spaced-repetition system, which does the scheduling for you. Quizlet generates flashcards, summaries and practice from a student's own notes.
  • Questions from your own material. Gemini Notebook is designed to answer questions from the documents you upload, so it stays tied to your own material. StudyFetch turns uploaded lectures into practice questions and flashcards, among other things. Mindgrasp creates notes, summaries, flashcards and quizzes from uploads.
  • Shared notes and an AI helper. Knowunity pairs a community of student notes with an AI tutor and generates flashcards and practice tests. Check its age terms and community settings first, since it hosts notes written by other users.

What each tool's terms say about children under 13, as recorded in our index from the tools' own pages: Gizmo and Mindgrasp do not allow it; Quizlet, StudyFetch and Knowunity allow it with consent; Gemini Notebook is for school-managed accounts only. Terms change, so check the tool before a child signs in.

Two cautions apply to all of them. First, a tool that makes notes or summaries for you is doing the easy part. Use the quiz or flashcard feature, not the summary. Second, check the age terms and ask your school what it allows. Many of these were built for university students.

Fit by exam, and by age

The method is the same for every system, but what you quiz yourself on is not. For GCSE, IGCSE, SPM, CBSE or A-Level, the most useful questions come from your own syllabus and the exam board's past papers. Paste in your notes and ask for questions at that level. Then check your answers against the mark scheme, not the AI's opinion of them.

For 11 to 14 year olds, keep it short and keep an adult in the loop. Ten minutes of recall on one topic, repeated across the week, beats an hour of anything on the night before. For 15 to 18 year olds, build the loop into a revision timetable: new topics early, return to old ones at widening gaps, and finish with timed past-paper questions done without AI. We cover whether and when the school allows AI help with work in Is using AI for homework cheating?.

A simple weekly plan

Questions first, notes second.
  1. Pick two or three topics from the syllabus. Write down what you think you can remember about each, from memory.
  2. Ask the AI to quiz you, one question at a time, without showing answers.
  3. Mark the gaps against your notes. Add those gaps to a flashcard set.
  4. Come back to the cards a day later, then a few days later, then a week later.
  5. Finish with one timed past-paper question. No AI.

Frequently asked questions

What is the best AI for studying? The best one is the one that makes you retrieve from memory, with spacing built in. Flashcard and quiz tools tied to your own notes fit that better than a general chatbot used for summaries.

Is it OK to use AI to summarise my notes? It is fine as a first look. It is not studying. Close the summary and test yourself, or you have made tidy notes and learned very little.

Can AI replace a tutor? In one carefully designed trial an AI tutor beat an active-learning class, but that tutor was built for the purpose and the learners were university students. For most school learners, AI is better seen as a source of practice questions and explanations than as a replacement for a teacher.

How often should I revise a topic? Spread it out. Return after a day, then a few days, then a week, and shorten the gap on anything you got wrong. Spacing is what the research recommends.

Is it cheating to study with AI? Studying is not cheating. Submitting AI work as your own is. The two are very different, and we separate them in our guide on AI and homework.

Sources

  1. Roediger and Karpicke, "Test-enhanced learning," Psychological Science 17(3), 2006: https://www.psychologicalscience.org/journals/psychological-science/j.1467-9280.2006.01693.x/
  2. Association for Psychological Science, "Which study strategies make the grade?" (on Dunlosky et al., 2013): https://www.psychologicalscience.org/news/releases/which-study-strategies-make-the-grade.html
  3. Kestin et al., "AI tutoring outperforms in-class active learning," Scientific Reports 15, 2025: https://pmc.ncbi.nlm.nih.gov/articles/PMC12179260/
  4. Bastani et al., PNAS 122(26), 2025 (abstract via RePEc): https://ideas.repec.org/a/nas/journl/v122y2025pe2422633122.html
  5. Fan et al., "Beware of Metacognitive Laziness," arXiv:2412.09315: https://arxiv.org/abs/2412.09315v1