For the past few years, the sales pitch around AI in education has been remarkably simple: give students better tools and they will learn faster. That assumption is starting to look a little shaky.
The OECD’s newly released PISA 2025 results suggest that students who use AI chatbots heavily for schoolwork often perform worse than students who use them more selectively. Yet this is not another “AI is melting everyone’s brain” story. Moderate users can perform as well as—or in some cases better than—students who barely use AI at all.
The lesson appears to be less about whether students use AI and more about what they make AI do for them.
Heavy AI use is associated with weaker performance
PISA 2025 is the OECD’s largest assessment yet, covering roughly 760,000 15-year-old students across 91 countries and economies. Alongside the usual science, mathematics, and reading tests, researchers also examined how students use AI chatbots for schoolwork.
AI is already remarkably common in education. Across OECD countries, only about 14% of students said they had never or almost never used AI for the school-related activities measured by PISA. Roughly one in five reported using it almost every day.
But more usage did not translate neatly into better results.
Students who used AI very frequently for specific tasks such as summarizing readings, researching topics, or helping produce written work generally posted lower science scores than non-users. OECD data also shows that students using AI “to help me learn” almost every day tended to score below students who used it weekly.
In other words, apparently you cannot speedrun education simply by adding more chatbot.
The sweet spot appears to be somewhere in the middle
The most interesting part of the OECD findings is the curve.
For several school tasks, students who used AI moderately—roughly between once a month and a couple of times per week—performed better than both very infrequent users and students relying on AI almost every day.
When students specifically said they were using AI to help themselves learn, weekly users performed at roughly the same level as students who did not use AI, after accounting for socioeconomic background.
That distinction matters.
There is a large cognitive difference between asking:
- “Explain photosynthesis to me using an analogy.”
- “Why does this equation work?”
- “Quiz me on this chapter and explain what I get wrong.”
…and typing:
- “Summarize this.”
- “Write my answer.”
- “Finish this assignment.”
Both technically count as using AI for school. Only one forces the student to keep doing much of the thinking.
AI seems to work better when students have to argue with it
The OECD also found evidence that schools can improve the way students interact with AI.
Around six in ten students across OECD countries said they had been asked during lessons to evaluate the quality of AI-generated information. Students who frequently used AI for learning and were taught to assess its answers critically tended to perform slightly better in science than students who used AI infrequently or not at all.
That is probably one of the most important findings in the entire report.
AI becomes much more educationally useful when the assignment is not simply “get an answer from the machine,” but rather:
Is the machine correct?
Suddenly the student needs subject knowledge, skepticism, reasoning, and the ability to spot hallucinations or weak arguments.
The chatbot stops being an answer vending machine and becomes something closer to an extremely confident study partner who occasionally makes things up.
That is a much healthier relationship.
This does not prove AI causes lower scores
There is an important statistical warning attached to all of this.
PISA shows associations, not proof that frequent AI use directly causes students to perform worse. Students who already struggle academically might use AI more often. Different schools, socioeconomic conditions, teaching methods, and national education systems can also influence both AI adoption and academic outcomes.
The OECD explicitly cautions against interpreting the data as a simple cause-and-effect relationship.
So the conclusion should not be, “Daily ChatGPT use costs students X months of education.”
It is closer to this:
Heavy, unguided AI use is associated with weaker performance, while purposeful and moderate use appears substantially less concerning—and can sometimes be beneficial.
That is a messier headline, unfortunately. Reality keeps refusing to optimize itself for social media.
Different countries are already developing very different AI habits
Student AI adoption also varies dramatically around the world.
In Japan, 39% of students said they never or almost never used AI chatbots for the school tasks covered by PISA. Only 27% reported using AI to help themselves learn at least once per week.
Vietnam sits near the opposite end of the spectrum. There, only around 4% of students reported essentially never using AI for those schoolwork activities, while 69% said they used chatbots to help them learn at least weekly.
That makes education an interesting real-world experiment.
Some systems are embracing AI aggressively. Others are moving much more cautiously. And because high-performing countries can be found on both sides, simply handing every student an AI account clearly is not a magic education policy.
The surrounding teaching system matters.
Schools are starting to put guardrails around AI
Policymakers appear to be reaching the same conclusion.
New York City, for example, has introduced a one-year moratorium on student-facing generative AI for students from 2-K through eighth grade during the 2026–27 school year. High-school students will still receive AI-literacy instruction and can participate in tightly controlled pilot programs using approved tools under teacher supervision.
That distinction is revealing.
The policy is not pretending AI will disappear. Instead, it separates unrestricted chatbot use from structured AI education.
For younger students: build foundational skills first.
For older students: teach them how to use AI without outsourcing their brains.
That may turn out to be a much more practical model than either extreme of “ban everything” or “give every child a chatbot and hope for productivity gains.”
The real divide may be between AI users and skilled AI users
There is another wrinkle in the OECD data that deserves attention.
Students from more advantaged socioeconomic backgrounds were more likely to report being taught how to evaluate AI-generated information critically.
That raises the possibility of a new kind of digital divide.
The first internet-era divide was about access to technology.
The AI-era divide could increasingly be about knowing how to use that technology without becoming dependent on it.
Two students might have access to exactly the same chatbot. One has been taught to interrogate sources, challenge answers, request explanations, compare reasoning, and recognize hallucinations. The other mostly uses it to make homework disappear faster.
Technically, both have AI access.
Educationally, they are playing entirely different games.
AI should be the tutor, not the substitute student
For students, the practical takeaway is surprisingly straightforward.
Tools such as ChatGPT, Claude, and Gemini can be extraordinarily useful when they are used to extend thinking rather than eliminate it.
Ask AI to explain concepts differently. Ask for examples. Make it quiz you. Challenge its answer. Request hints instead of solutions. Have it critique your reasoning. Ask why your answer is wrong before asking what the correct answer is.
Use it like a tutor with infinite office hours.
The danger begins when AI stops assisting the learning process and quietly becomes the person doing the learning process.
The OECD’s early evidence suggests that moderate, deliberate use may be the better path. AI can make education more powerful—but only if the student remains the one doing the difficult part.
Thinking.
