With semester exams and pre-boards coming up, a lot of revision now happens with a chatbot open in the next tab. It feels efficient. You get clear explanations, worked answers and summaries in seconds. The research suggests that feeling and actual learning can come apart. In one large classroom experiment, students with unrestricted access to an AI assistant did much better on practice problems, then did worse on their own exam than students who never had it. The good news is that the problem isn't AI itself. It's a specific way of using it, and that's easy to change.
The experiment that made researchers pay attention
The clearest evidence comes from a field experiment published in PNAS by Hamsa Bastani and colleagues. Nearly a thousand high school maths students were given access to one of two AI tutors built on GPT-4. One worked like a standard chatbot. The other was designed to protect learning, with prompts that steered it towards hints rather than answers.
During practice, both helped. Grades improved by 48% with the standard chatbot and by 127% with the tutor version.
Then access was taken away. Students who had practised with the standard chatbot scored 17% lower than students who had never had AI at all. The tutor version largely avoided that drop. The researchers' explanation is simple: without guardrails, students used the AI as a "crutch" during practice, and then couldn't do the work alone.
Why this happens: effort is part of how memory works
Psychologists call it cognitive offloading: handing part of a mental task to a tool. We do it all the time with calendars and calculators, and usually that's fine.
Learning is different, because the effort is not a side effect. Struggling to recall a formula, getting it wrong and correcting it is the process that builds the memory. Researchers describe these as "desirable difficulties." When a chatbot supplies the answer before you've struggled, the practice session goes smoothly, and very little gets stored.
There's a second trap. A clear explanation feels like understanding. Reading a good summary gives you the sense that you know the material, an illusion of competence that only gets tested when the summary isn't there.
A review in Trends in Cognitive Sciences by Trent Cash and colleagues puts it carefully: offloading thinking to AI "can impede skill acquisition and lead to skill decay, but risks depend on how AI is used." The same review suggests our basic cognitive abilities are likely to be more resilient than specific skills. So the risk isn't that AI makes you less intelligent. It's that you don't build the particular skill you were practising.
It's how you use it, not whether
The same body of research points to uses that seem to help. As PsyPost's summary of the review describes, tools that give hints instead of answers did not show the same drop, and asking AI for feedback on your own idea supports learning in a way that asking for a finished solution does not. The review also notes a quieter risk: people tend to accept flawed chatbot answers without checking them.
That gives a useful rule of thumb. If the AI is doing the thinking you're supposed to be practising, it's costing you. If it's responding to thinking you've already done, it's probably helping.
Five small changes for your next study session
1. Attempt first, ask second
Give yourself a fixed stretch, say ten minutes, to work on a problem alone before opening the chat. The struggle in those ten minutes is where most of the learning happens.
2. Ask for the next step, not the answer
You can turn a general chatbot into something closer to the "tutor" version just by asking differently:
"Don't solve this for me. I'm stuck at this step: [paste your working]. Give me one hint about what to try next."
3. Close the tab, then recall
After an explanation, close it and write down what you remember, or solve a similar problem from scratch. If you can't, you've found what to revise.
4. Use AI as a checker
Explain a concept in your own words first, then ask the chatbot to point out what's missing or wrong. This keeps you doing the reasoning and uses the tool for feedback.
5. Keep your last sessions AI-free
Your exam won't have a chatbot in it. In the final week, practise under exam conditions, alone and timed, so the first time you work without help isn't in the exam hall.
What we don't know yet
This research is young. The strongest evidence comes from specific settings, like secondary school maths, and the review's authors note that long-term effects are still open questions. It's also not a case for avoiding AI altogether: in the experiment, the well-designed tutor raised practice scores a lot without the later drop.
It does fit a pattern that keeps appearing in attention research. Tools that remove friction feel good in the moment, whether it's a phone that makes procrastination easier or a summary that makes revision feel finished. Sometimes the friction was doing useful work. If you're also running on short sleep before exams, that effort feels even harder, as the research on sleep and perceived effort shows.
The takeaway
Use AI like a good coach: someone who asks you a question, gives you a nudge and checks your work, not someone who sits the exam for you. Struggle first, ask for hints, and test yourself without the tab open. If exam nerves are part of the picture too, what the research says actually helps with exam anxiety pairs well with this.



