In one study, students with ChatGPT scored higher on practice and lower on the real exam. Here is why, and how to use AI so you learn faster for real.
Is ChatGPT bad for learning? It can be, and the reason is simple. When a chatbot hands you the answer, it also skips the part of studying that builds memory. The good news is that the same tools can help you learn faster, for real, if you use them to think instead of to finish. This guide walks through what the 2025 research found, why it happens, and how to use AI so the learning stays in your head.
ChatGPT is not bad for learning by itself. Its default habit is: it answers right away. Studies show that students who get answers on demand do better on practice and worse once the AI is gone. When the same AI is set up to give hints and ask questions, most of that harm goes away.
So the real question is how you use it. Ask for the answer and you get a finished worksheet. Ask to be taught and you get a tutor. The rest of this article shows the evidence for both sides.
Most of what you read online about ChatGPT and learning is opinion. Here is what four recent studies measured.
The clearest study so far ran in real high school math classes. Hamsa Bastani and colleagues gave nearly 1,000 students access to GPT-4 during practice sessions (Bastani et al., PNAS 2025). One group got a plain chat window, like standard ChatGPT. Another got a version with prompts written to protect learning: it gave hints and did not hand over full answers.
With the AI open, both groups did much better on practice problems. Grades rose 48% for the plain version and 127% for the tutor version. Then the AI was taken away for the exam. Students who had used the plain chatbot scored 17% lower than students who never had AI at all. The tutor version mostly removed that drop.
This is the study people mean when they say ChatGPT hurts learning. The authors' word for what happened is "crutch." Students leaned on the chatbot to get through practice, and when it was gone they could not do the work alone. That is the core finding, and it is the reason this article exists.
You may have seen headlines that ChatGPT "rots your brain." They come from an MIT Media Lab study led by Nataliya Kosmyna (Kosmyna et al., 2025). Fifty-four adults wrote essays over several sessions, either with an AI model, with a search engine, or with no tools. Researchers recorded their brain activity with EEG.
The no-tools group showed the strongest, most spread-out brain connectivity. The AI group showed the weakest. People who wrote with AI also felt the least ownership of their essays and struggled to quote what they had just written.
Read this one with care. It is a small preprint, not yet peer reviewed, and only 18 people completed the final session. It measured essay writing, not exam results. It does fit the math study, though: when the tool does the thinking, less of the thinking happens in you.
In one lab experiment, 108 people with little or no coding background spent 45 minutes learning basic Python (Lehmann, Cornelius and Sting, 2024). Across two experiments, having an AI assistant made no overall difference to how much people learned.
The detail is where it gets useful. Students who used the AI to generate solutions covered more topics but understood each one less. Students who used it to ask for explanations understood more. The AI also widened the gap between students who started with more knowledge and those who started with less. Same tool, opposite results, depending on the question typed.
Now the other side. At Harvard, physics instructors built an AI tutor on GPT-4 with a careful system prompt. It guided students one step at a time, kept the load manageable, and was fed the instructors' own worked solutions so its feedback was accurate. They tested it in a randomized controlled trial against an active-learning class (Kestin et al., Scientific Reports 2025).
Students who used the AI tutor had median learning gains more than double those of the in-class group. They also took less time: a median of 49 minutes, against about 60 minutes of class. And they reported feeling more engaged and more motivated.
Put the four studies side by side and a pattern shows up. AI that answers can hurt learning. AI that teaches can help a lot. So is ChatGPT bad for learning? The research says the set-up decides.
Memory gets built when you pull information out of your own head. Psychologists call this retrieval practice, and it is one of the best-tested ideas in education.
In a classic experiment, students read short prose passages. Some reread them. Others took a recall test with no feedback. Five minutes later, the rereaders remembered more. Two days later and one week later, the students who had tested themselves remembered much more (Roediger and Karpicke, 2006). The rereaders also felt more confident, even though they would forget faster.
That last detail explains the ChatGPT problem. Reading a clean, correct answer feels like learning. It is smooth and it makes sense. But you never had to retrieve anything, so there is little to remember.
A large review of ten common study techniques reached the same verdict (Dunlosky et al., 2013). It rated practice testing and spreading study over time as the most useful. Rereading, highlighting and summarizing were rated low. Asking a chatbot to explain something and then reading its answer is, in practice, a very polished form of rereading.
Learning scientists have a name for the useful kind of struggle: desirable difficulty. A math problem that makes you stop and think for a minute is doing its job. Remove the pause and you remove the learning.
The risk is not the same for everyone. Three situations make ChatGPT bad for students in a way you can predict.
When you are still learning the basics. In the coding study, the AI widened the gap between people who started with more knowledge and people who started with less. If you already understand fractions, asking a chatbot to check your work costs you little. If you are learning fractions this week, letting it do the work costs you the week.
When the test is closed-book. The math students did fine as long as the chatbot sat next to them. The harm showed up on the exam, where it was gone. Most tests you will take, from a unit quiz to the SAT, look like that exam.
When you are tired or behind. Nobody types "please make me struggle" at midnight before a deadline. The pull toward the answer is strongest exactly when you have the least time, which is why a habit or a tool that slows you down helps more than good intentions.
None of this makes ChatGPT bad for learning in every case. A college student reviewing a topic they already know, or an adult checking a fact, faces much less risk. The danger sits with new skills, closed-book tests and late nights.
Yes, for the right jobs. ChatGPT is good for studying when you ask it to react to your thinking. A fair look finds real strengths:
What these have in common: you do the thinking, and the AI reacts to it.
Here is the contrast in practice. Say you are stuck on this algebra problem: 3(x − 4) = 2x + 5.
If you type "solve 3(x-4)=2x+5", a standard chatbot will usually reply with every step and the answer, x = 17. It is correct and clear, and you will have learned almost nothing by the time you copy it down.
Now type this instead: "I'm learning to solve equations. Don't give me the answer. Ask me one question at a time." A good reply looks more like this:
Sure. Look at the left side, 3(x − 4). Before we move anything across the equals sign, what could you do with those parentheses?
You answer "distribute the 3," you get 3x − 12, and the next question asks how to get the x terms together. Two minutes later you found x = 17 yourself. That is the difference the Bastani study measured.
OpenAI saw the same problem. ChatGPT now has a study mode that teaches through questions, hints and step-by-step explanations, adds quizzes and knowledge checks, and tracks progress.
That is a real step in the right direction, and if you use ChatGPT, turn it on. Two limits are worth knowing. First, it is a mode you choose. One click and you are back to plain answers, which is exactly the moment a tired student at 11 p.m. will click. Second, a prompt alone does not add the parts of learning that happen across weeks. Think of a plan tied to your exam date, quizzes graded where you cannot peek, and reviews timed for when you start to forget. A parent or teacher may also want to see progress.
For a full step-by-step method with prompts and a 7-day exam plan, read how to study with AI.
Everything above works when you use ChatGPT for studying, or any other chatbot, if you have the discipline to follow it every time. A tutor built for learning takes that discipline off your plate. This is what we built Kuest to do, and the aim behind our line "study anything 10x faster" is that you stop rereading and start retrieving.
If you want to see it on a real topic, start with algebra or read about the wider evidence in do AI tutors work. Parents can read how Kuest works for families.
It depends on your school's rules and on what you hand in. Most teachers allow AI for explaining ideas and forbid submitting AI-written work as your own. Teens draw the same line. In a 2024 Pew survey, 54% said using ChatGPT to research new topics is acceptable. Only 18% said the same about writing essays (Pew Research Center, 2025).
There is no good evidence that it lowers intelligence. The evidence shows something narrower: if it does the work, you learn less from the work. The MIT EEG study found weaker brain engagement while writing with AI, but it is a small preprint about essays, so treat it as a warning sign, not a verdict.
Because the studies tested different set-ups. Plain answer-on-demand use hurt exam scores in the Bastani study. A tutor that asked questions and gave hints cut most of that harm, and the Harvard tutor more than doubled learning gains. Design and use decide the result.
OpenAI's terms require users to be at least 13, and under-18s need a parent's permission. For younger children and for anyone still building basic skills, a tool that teaches step by step is a safer choice than one that answers on request. Our parent's guide to kids and AI covers this in detail.
According to Pew, 26% of US teens aged 13 to 17 said they had used ChatGPT for schoolwork in late 2024, double the 13% from 2023.
Kuest is the AI tutor built to help you study anything 10x faster, by teaching you instead of answering for you. Start free.