Is Using ChatGPT to Study Bad? What the Studies Found

There is no evidence that using ChatGPT to study, by itself, makes you dumber. In the studies, what cost students their learning was not the AI itself but handing it the thinking. Students who took the answers scored lower on exams taken without AI, and students who got hints and worked the problems themselves lost nothing.

Does studying with ChatGPT lower your grades?

It did when students took the answers. The same AI, set up to give only hints, did not.

Bastani and colleagues ran a field experiment with about 1,000 students in math classes at a high school in Turkey (2025, PNAS). Each class got one tool for its practice sessions: textbooks only, GPT-4 behind an interface like standard ChatGPT (GPT Base), or the same GPT-4 told to guide students step by step without giving the answer, using solutions and hints from the teachers (GPT Tutor). After practice, everyone took an exam on similar problems with no tools.

During practice the AI helped a lot. Compared with the textbook-only classes, scores were 48% higher with GPT Base and 127% higher with GPT Tutor. But on the exam without tools, the GPT Base classes scored 17% lower than the textbook-only classes. The GPT Tutor classes were statistically no different from textbook only.

The students' own sense of it is worth noticing too. GPT Base students did worse on the exam but did not feel they had learned less. GPT Tutor students scored the same yet felt they had done better. How easy studying felt said little about how much they learned.

What is the difference between getting an answer and getting a hint?

Who does the thinking. With an answer, the AI thinks. With a hint, you do.

The researchers read the messages students sent. The single most common message in the GPT Base classes was "What is the answer?" They grouped messages that pasted in the question or asked for the answer as superficial. On the very first problem of the first session, 56% of first messages to GPT Base and 42% to GPT Tutor were superficial. Across all problems in that session, the share rose to 67% for GPT Base and fell to 37% for GPT Tutor. Facing a tutor that would not hand over the answer, students moved toward trying the problem and asking for help.

The researchers also checked whether GPT Base's frequent math mistakes explained the drop. They concluded that being misled by wrong answers mattered less than not engaging with the problem: students used the AI as a "crutch."

This fits what memory research has long found. What you produce yourself is remembered better than what you only read, which is called the generation effect. Bertsch and colleagues (2007) pooled 445 effect sizes from 86 studies and found that generating beat reading by about 0.40 of a standard deviation. Getting it wrong is fine. In experiments by Kornell, Hays, and Bjork (2009), people who tried to answer before seeing the answer later remembered more than people who read the question and answer together, even though their attempts failed.

In one line Getting the answer first takes away your chance to think. Try first, even if you are wrong, and then ask.

Why can't people remember an essay they wrote with ChatGPT?

Probably because they did not produce the sentences themselves. This study is a small preprint, though, so read it with care.

Kosmyna and colleagues at the MIT Media Lab (2025) split 54 people aged 18 to 39 from universities in the Boston area into three groups of 18: one used ChatGPT, one used a search engine, one used no tools. Each person wrote a 20-minute essay on an SAT prompt, three times over several months. Right after writing, they were asked whether they could quote a sentence from the essay without looking.

In the first session, 15 of the 18 ChatGPT users could not quote anything, and none quoted correctly. In the search and no-tools groups, two or three people could not quote correctly. From the second session the gap shrank. The authors note that the question may have caught people off guard the first time. One reading is that when you expect to remember something, you work with it differently.

This is where the phrase "cognitive debt" comes from. A few things to keep in mind:

• As of October 2026 it is a preprint that has not been peer reviewed.
• The sample is small: 54 people for the first three sessions, 18 for the fourth, where groups switched.
• The authors themselves list the small number of participants from a few nearby universities as a limitation.
• The EEG results show less connected brain activity while writing with ChatGPT. They do not show brain damage.

So the takeaway is narrower than "ChatGPT rots your brain." Sentences you did not make yourself stayed poorly in memory, even minutes after you wrote them.

Does AI make you dumber?

The research so far does not show that. What it does show is that the work you hand to AI is work you stop practicing.

Psychologists call handing mental work to the outside world cognitive offloading. Risko and Gilbert (2016) define it as using physical action to reduce the mental demand of a task, like tilting your head to read a rotated image or setting a phone reminder for an appointment. We do it all the time. The trouble starts when the work you hand over is the very thing you meant to learn.

What should you change the next time you study with ChatGPT?

Just the order. Recall first, then ask, then close it and recall again.

1. Spend one minute answering before you ask. Write what you know on a blank page, or just the first line of a solution. Even if you are wrong, the answer you see next will last longer.
2. Ask for a hint, not the answer. For example: "Don't tell me the answer. Ask me a question that shows where I'm stuck." That is what the tutor in the experiment did.
3. When you have read the answer, close the window and write three lines. What you asked, the core of the answer, and what you did not know. Where you get stuck is what you have not learned yet.
4. Recall it again a few days later. Even what you recalled once fades. For when to come back, see the note on spaced repetition schedules.

If studying felt easy, be a little suspicious. The students in the experiment mistook easy studying for learning more.

In Aloom

In Aloom you ask AI and keep only the answers you want in your knowledge graph. The answers you keep come back each day, one, three, or five as you choose, and instead of showing you the answer again, Aloom asks you a question so you recall it first. The order in this note, asking and then closing it and recalling, is built into the daily review. The more you recall a node, the darker its ink, so you can see what has really become yours. See how Aloom works

FAQ

Is ChatGPT Study Mode OK to use?

We have not found a study that tests Study Mode directly, and it is not the tutor used in the experiment above. If it guides you with hints and questions instead of answers, it points the same way. With any tool, trying to answer before you ask is still your part.

How do I use ChatGPT to study without relying on it?

Change the order before you cut the time. Answer first on your own, ask for hints instead of answers, and close the answer and recall it after you read it. You can use AI just as much and still keep the thinking on your side.

Is the MIT "Your Brain on ChatGPT" study reliable?

It has not been peer reviewed and had 54 participants, so it is hard to generalize from. Its finding that people could not quote their own essays does point the same way as decades of research showing that what you produce yourself is remembered better.

References

• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. PNAS, 122(26), e2422633122.
• Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I., & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task. arXiv:2506.08872. Preprint, not peer reviewed.
• Bertsch, S., Pesta, B. J., Wiscott, R., & McDaniel, M. A. (2007). The generation effect: A meta-analytic review. Memory & Cognition, 35(2), 201–210.
• Kornell, N., Hays, M. J., & Bjork, R. A. (2009). Unsuccessful retrieval attempts enhance subsequent learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 35(4), 989–998.
• Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688.