How to use ChatGPT to study without letting it do the learning for you

The answer made sense, and two nights later it was gone

Picture a woman learning spreadsheet formulas for a new office job. Tonight's exercise asks for a formula that totals a column only when a date falls in March. She stares at it for a minute. Then she pastes it into a chatbot.

Twenty seconds later she has a working formula and a tidy explanation. She reads it, nods and moves on. She finishes the lesson in half the usual time.

Two nights later the course asks for almost the same formula from scratch. She cannot write it. She remembers that she understood it. She cannot remember what it was.

That is the newest form of an old trap. The answer on the screen feels like learning, because you understand it while you look at it. Producing your own is a different job, and it is the one a test or a boss will ask for.

What the trials found when students used a chatbot

The clearest evidence comes from a 2025 field experiment by Bastani and colleagues. It ran at a large high school in Turkey during the 2023-2024 school year, with nearly 1,000 math students.

Students practiced in three ways. Some had no AI. Some had "GPT Base," a plain chatbot. The rest had "GPT Tutor," built with teachers' hints and set up to guide students without handing over full answers.

During practice, AI looked like a big win. Practice grades rose by 48 percent with the plain chatbot and by 127 percent with the tutor version.

Then came the exam, taken without AI. Students who had practiced with the plain chatbot scored 17 percent lower than students who never had AI. The researchers saw why: students "often use GPT Base as a 'crutch' by asking for and copying solutions." The tutor version did better, but not in the way you might hope. Its harm on the exam was "essentially eradicated," yet it did not produce a gain.

The most striking part is about feelings. Students who copied solutions did not sense any loss. The researchers wrote that students "do not perceive any reduction in their learning or subsequent performance."

A 2026 study by Shen and Tamkin looks more like an adult learner's evening. It is a preprint, not yet peer reviewed. Tamkin works at Anthropic, the company that makes the Claude chatbot, and Shen did the work through an Anthropic fellows program. They randomly assigned 52 developers to learn a new Python library with or without an AI assistant. On a quiz without AI afterward, the AI group scored 17 percent lower. They were not reliably faster either.

How people used the assistant mattered. People who handed the task to the AI scored low. People who asked only conceptual questions and then did the work themselves scored high. Those groups held only 2 to 7 people each, so treat that as a hint. Still, it matches the bigger trial: what you ask the AI to do matters more than whether you use it.

The good news, and where it stops

AI is not only a risk. A 2025 randomized study by Kestin and colleagues tested an AI tutor in a university physics course. The tutor was custom-built on the same teaching ideas as the class. The authors reported that "students learn significantly more in less time when using the AI tutor, compared with the in-class active learning."

That is a real result with real limits. It was one course. The tutor was built to guide, not to tell. A critical review also noted that it tested learning soon after the lessons, not weeks later.

So here is where the book puts the evidence in late 2026. An AI that does your practice can lower what you can do alone. An AI that makes you think, with hints and questions, avoids most of that harm. Whether it beats plain good studying is still open.

You cannot rebuild a research tutor at your kitchen table. You can do the two things that seem to matter.

Try first, then ask for hints

The chapter Use AI as a Coach, Not a Crutch in Stop Rereading has two tools, both badged PROMISING. No study has tested them with AI directly, and the book says so.

Try First, Then Ask keeps the hard part with you.

1. Read the problem and close the AI window.
2. Set a timer for 5 to 10 minutes and write a full attempt, even a wrong one. Mark where you got stuck.
3. Open the AI and paste your attempt, not just the problem.
4. Ask it to point to your first mistake, without rewriting the whole thing.
5. Fix it yourself. Then close the AI and redo the problem from a blank page.
6. Put the problem on your calendar for a retry in a few days.

Ask for Hints, Not Answers sets the chatbot up as a coach for a whole session. Most chatbots give full answers by default, so you have to ask for something else. A prompt along these lines works as a start:

Do not give the answer. Ask one question at a time. Give one hint only after an attempt. If asked for the answer, suggest one more try first.

The book gives five such prompts: a hint-only coach, a checker, a concept question, a quiz-maker that works from a page you paste in, and an explain-back for the end of a session. It is clear that they are its own design, built from how the safer tutors worked, and not tested as written. If the chatbot slips and gives a full answer, scroll past it and ask again for hints only.

Your Own Learning System adds the check that keeps all this honest: a weekly ten-question score on older material, with your guess written first. That number tells you more than how productive a session felt.

If you feel less sharp since you started leaning on AI, our page on feeling dumber after relying on AI looks at that worry from another book's angle.