Is AI Making You Worse at Studying? What the Research Says

The short answer: AI can make you worse at studying, and researchers have measured how. When students use it to get answers, practice gets easier and exam scores drop. When they use it for hints and to check their own work, the harm mostly disappears. The tool is not the problem. The question is who does the thinking.

Almost every student uses AI to study now. A survey of UK undergraduates published by the Higher Education Policy Institute in March 2026 found that 95% use it in some way and 94% use it for assessed work. So the useful question is no longer whether you use it. It is whether AI is making you worse at studying without you noticing.

The study every student should know about

A team led by Hamsa Bastani at the Wharton School ran one of the largest experiments so far on AI and learning. Nearly 1,000 high school math students in Turkey were split into three groups. One group practiced with a plain assistant that worked like ChatGPT. A second group practiced with a version that teachers had set up to give hints instead of answers. A third group had only their textbook and notes.

During practice both AI groups looked brilliant. The plain assistant group performed 48% better than the students without AI. The hint group performed 127% better.

Then came the exam, taken without AI. The students who had practiced with the plain assistant scored 17% lower than the students who never had AI at all. The hint group scored about the same as the students without AI. The results were published in PNAS in 2025.

The students with the plain assistant did better in every practice session and worse on the only test that counted.

The researchers’ explanation is simple. Many students used the assistant to get the answer and move on. The homework got done. The learning did not.

These were high school students and your course is not a Turkish math class. But the mechanism has nothing to do with age. Whenever a tool does the solving for you, you skip the part of practice that builds the skill.

Why it feels like you learned more than you did

The most worrying part of the Bastani study is not the 17% drop. It is that the students did not see it coming. The researchers found that students who used the plain assistant were overconfident about how much they had learned.

This trap is well documented. In a 2019 Harvard study published in PNAS, physics students were taught the same material in two ways. Some sat through polished lectures. Others worked through problems together in class. The students who worked through problems scored higher on the tests. Yet they felt they had learned less. The researchers found that the feeling of learning and actual learning pointed in opposite directions.

A clear AI explanation creates exactly that feeling. The steps flow. Everything makes sense while you read it. But following an explanation is not the same as being able to do the work alone. The first feels like learning. Only the second shows up on the exam.

It is not only about grades

A 2025 study from the MIT Media Lab looked at what happens in the brain. Fifty-four adults wrote essays with ChatGPT, with a search engine or with no tools at all while researchers recorded their brain activity. The ChatGPT group showed the weakest brain engagement of the three. After the first session 83% of them could not correctly quote a single sentence from the essay they had just written. In the other two groups only 11% had that problem.

That study is small and was released as a preprint before peer review. Treat it as a warning sign rather than a final verdict. It does fit a pattern though. A 2025 survey of 319 knowledge workers by Microsoft Research and Carnegie Mellon University found that people with more confidence in AI reported doing less critical thinking about its answers.

So should you stop using AI?

No. Look again at the second group in the Bastani study. Students who used an AI tutor that gave hints did far better in practice and lost nothing on the exam. AI did not cause the harm. Handing over the thinking did.

Anthropic saw the same split when it studied how university students use its assistant Claude. About 47% of student conversations were what it called direct. The student asked for an answer or a finished piece of work with little back and forth. The rest were collaborative, with the student working through the problem alongside the AI. The first pattern is the one to avoid.

Six rules for studying with AI without losing the learning

1. Try it yourself first

Give every problem an honest attempt before you ask for help. Five or ten minutes is enough. The struggle is not wasted time. It is what makes the explanation stick when it arrives, because you now know exactly which gap it fills.

2. Ask for the next step, not the answer

Tell the AI what you tried and ask only for a hint. For example:

“I got this far and I think the next step is to factor. Don’t solve it. Tell me if I’m on the right track and give me one hint.”

That is essentially what the hint group in the Bastani study received. It is why they did not fall behind.

3. Make it check your reasoning

Flip the usual roles. Solve the problem yourself. Paste your full working and ask the AI to point to your first mistake without fixing it. Finding and fixing your own error teaches you far more than reading a correct solution.

4. Explain it back in your own words

When an explanation finally clicks, close it and explain the idea out loud or on paper as if to a friend. In a classic 1989 study Michelene Chi and her colleagues found that the strongest physics students explained worked examples to themselves as they read. The weaker students just read them. Explaining exposes the gaps that reading hides.

5. Close the chat and test yourself

Before you finish, close the tab and solve a similar problem from scratch. In a well known 2006 experiment Henry Roediger and Jeffrey Karpicke found that students who tested themselves remembered much more a week later than students who reread the same material. The rereaders felt more confident. The testers remembered more.

6. Come back tomorrow

One good session is not enough. Return to the idea a day or two later with a fresh problem. In a 2013 review of ten popular study techniques, a team led by John Dunlosky rated only two as highly useful. They were practice testing and spreading study out over time. Highlighting and rereading, two of the most popular habits, were rated low.

Do the new study modes fix this?

In 2025 the big AI companies all added modes built for learning. ChatGPT has Study Mode. Gemini has Guided Learning. Claude has a Learning style. NotebookLM has a Learning Guide. Each one asks questions and gives hints before it gives answers. They are a real step forward and they follow the same idea as the hint group in the Bastani study.

The catch is that each one is a setting. One click turns it off and the plain answers come back, usually when you are tired and the deadline is close. So the six rules still apply. A study mode can make them easier to follow. It cannot follow them for you. We looked at what each study mode actually does in a separate guide.

A one-minute check before you close the tab

Ask yourself three questions. Could I solve a similar problem right now without help? Could I explain why each step works? Would I recognise this kind of problem if it turned up mixed in with others on an exam?

If the answer to any of them is no, you are not done. You have just found the next thing to study.

Frequently asked questions