“Does using AI make you dumber?” is a question plenty of parents, teachers, and students have been asking this week, after a study from China spread quickly online. Researchers David Strömberg, Victor Lei, and Yanhui Wu followed 26,811 students in grades 7 to 12 in a county in central China for 30 months, linking their homework scores and homework time to closed-book exams and to high-stakes entrance exams across nine subjects. Their findings are set out in a 2026 discussion paper on the generative AI learning penalty, published by CEPR.
The headline numbers are striking. After students began using generative AI, their homework scores rose by 18% and the time they spent on homework fell by 30%, from about 64 minutes to about 45. But their monthly closed-book exam scores fell by 20% within six months. On the entrance exams, which take longer to reflect what students have learned, the full penalty showed up after about two years: 24% for high-school entrance exams and 18% for college entrance exams, according to the authors’ summary of the study for VoxEU.
Psychologists have a name for what seems to be going on. Evan Risko and Sam Gilbert’s review of cognitive offloading, published in Trends in Cognitive Sciences, describes how we use physical actions and outside tools to cut the mental effort a task demands. Offloading is often sensible, but when the task is learning, the effort is the point. That is also why researchers have started rethinking critical thinking in the age of AI.
Does using AI make you dumber? What the study found
AI use spread fast in the sample. By June 2025, roughly 80% of students reported using generative AI, up from almost none in September 2022, and the most popular tools were Doubao and DeepSeek, general-purpose chatbots with no special educational features. Before adopting AI, the students who later used it had scores and trends very similar to those who did not.
The losses were not spread evenly. Exam scores fell by around 27% in social science subjects such as history and politics, about 22% in STEM subjects, 17% in English, and 9% in Chinese. Junior students, boys, and students who had been high achievers lost the most.
One finding should worry anyone who checks homework as a sign of progress. Among students who did not use AI, higher homework scores went with higher exam scores, as you would expect. Among AI users, students who scored higher on homework were more likely to do worse in exams. The authors suggest that a sudden jump in homework scores may now be a warning sign rather than good news.
Outsourcing the homework or using AI as a tutor
The researchers saw two quite different ways of using AI. After six months, about four in five AI users were finishing assignments in under 50 minutes, faster than the fastest students without AI, with exceptionally high homework scores and markedly worse exam results. That pattern is consistent with handing much of the homework over to the chatbot.
The remaining one in five kept spending roughly as long on homework as students who did not use AI, and their exam scores were similar to non-users. The authors are careful here: this does not prove that forcing students to spend more time would restore learning, because the two groups may differ in motivation, information, and AI skills.
Why the learning does not stick
Learning researchers Elizabeth and Robert Bjork use the term “desirable difficulties” for conditions that make learning feel harder in the moment but help it last, such as spacing out study, mixing topics, and testing yourself. Their 2011 chapter on creating desirable difficulties explains why the struggle of retrieving and working things out is often what builds durable knowledge. When a chatbot removes that struggle, the homework looks better while the learning behind it thins out.
A randomised trial points the same way. Hamsa Bastani and colleagues gave nearly 1,000 high school maths students in Turkey access to GPT-4 tutors, as reported in their study of generative AI without guardrails in PNAS. A standard ChatGPT-style tool improved practice scores by 48%, but when it was taken away, those students did 17% worse on the exam than students who never had it. A version designed to give hints instead of answers improved practice scores by 127% and largely removed the harm. Tellingly, students did not think they had learned less.
What the study can and cannot tell you
The Chinese study is not a randomised trial. Students chose when to start using AI, and the researchers compared changes in their results with those of students who had not yet adopted it, a method called difference-in-differences. It is a strong design, but it cannot rule out every difference between students who picked up AI and those who did not. The timing of adoption also came from student surveys.
It is also a discussion paper rather than a peer-reviewed journal article, it covers teenagers in one county in China, and it looks at self-directed use of general-purpose tools. The authors themselves say the effect sizes cannot be directly applied to other settings, though they argue the underlying problem, teenagers avoiding homework when they can, is likely to be general. It tells us little about adults using AI at work, and nothing about intelligence in the IQ sense.
How to use AI without losing the learning
- Try first: Have a proper go at the problem or the first draft before you open a chatbot, so the hard thinking is yours.
- Ask for hints, not answers: Tell the AI you want a nudge or a question back rather than the solution, which is close to the design that protected learning in the Turkish trial.
- Explain it back: Close the chat and explain the idea in your own words, out loud or on paper, then check where you got stuck.
- Test yourself without it: Short quizzes from memory are a classic desirable difficulty and show you what you actually know.
- Watch the time: If homework that used to take an hour now takes 15 minutes with perfect marks, the AI may be doing the learning for you.
- For parents and teachers: The study’s authors suggest looking at time actually spent on homework and at short in-class quizzes, not homework scores alone.
None of this means young people should be kept away from AI altogether. The goal is closer to the simple principle that AI can help children write, but it shouldn’t replace the thinking behind it.
FAQs
- Does using AI make you less intelligent? These studies measured learning of school material, not intelligence. They suggest that relying on AI for answers can mean you learn less, not that it lowers your IQ.
- Does AI make you lose brain cells? None of the studies discussed here measured brain cells or brain structure. They measured homework and exam performance.
- Is it bad to use AI for homework? It depends on how. Students who used AI while still spending similar time on homework had exam scores similar to non-users, while those who finished very quickly did much worse.
- Who was most affected in the study? Junior students, boys, and previously high-achieving students saw the largest losses, and social science subjects were hit hardest.
- Was the study a randomised trial? No. It compared students before and after they chose to start using AI, so it is strong evidence but not proof that the same would happen to every learner.
Takeaway
AI is very good at producing finished homework, and that is exactly the problem when the homework was only ever a route to learning. The study of 26,811 students is not the last word, but its message matches a randomised trial and decades of learning research: keep the effortful part for yourself, use AI for hints and feedback, and check what you know without it.
