World's biggest school test asked 15-year-olds how they use chatbots. Where the machine wrote the essay, scores fell — unless lessons taught them to check it.Prefer HTon Google Every three years the Organisation for Economic Co-operation and Development (OECD), the Paris-based grouping of mostly wealthy countries, sets the same test in science, reading and mathematics to a sample of 15-year-olds around the world. The Programme for International Student Assessment, or PISA, is the most widely used comparison of school systems. The 2025 round, published on Tuesday, covered more than 760,000 students in 91 countries and economies, and for the first time asked them how often they use AI chatbots to do their schoolwork.True learning requires struggle; AI hands over the answer before memory can build it. (Pexels/ Representational)Students answered for four uses: summarising a text they had been set to read, doing first-pass research on a new topic, drafting writing assignments, and a catch-all option, using AI "to help me learn".The key insight: children who never use a chatbot to draft their assignments scored 509 in science, while those using one almost every day scored less, at 481.PISA marks are set on a scale where the average student across OECD countries scores close to 500 and a typical 15-year-old gains about 20 points over a school year, which makes the 28-point gap worth roughly a year and a half of schooling. Those figures already account for how comfortable a child's home is: the OECD compares students of similar household income, parental education and study space, because a straight comparison would also end up measuring a families' comfort and conveniences rather than the chatbot.The same trend holds when a chatbot is asked to do other parts of school work. Children who never use AI to summarise something they have been given to read scored 508, and those doing it almost every day scored 477 — a gap of 31 points, wider still than the one for essay drafting. For first-pass research the gap narrows to 19 points, 505 against 486. The ranking follows how much of a child's own work the machine takes over: reading a text and writing an essay are the work itself, while looking something up is a step on the way to it.In other words, using AI does not equip a child with anything new in what they know. Put two children with the same advantages at home side by side, and the one who hands the writing over to a machine knows less science by the time the test comes round.The other key insight was that almost every child in the survey uses AI for something. Across OECD countries only 13.7% said they never use it for schoolwork of any kind. Abstention has little to do with how rich or how successful a school system is — 39.1% of Japanese students and 28.3% of British ones never use it, against 4.4% in Singapore and 4.1% in Viet Nam. Daily use remains a minority habit, about one student in nine for each specific task, rising to one in five for "to help me learn".The scores also fail to fall in a straight line as use rises. In all four uses, children who turn to a chatbot once or twice a year score below children who use it every week.Also read: The road to rogue AI, and the technical shortcut behind itBut not all AI use is badPISA data shows students using chatbots weekly as study aids outscored those who never touched the technology at all. (Unsplash/ Representational)The four uses divide into two kinds. In three of them the chatbot produces something the child would otherwise have produced. The fourth, "to help me learn", covers everything else asked of a chatbot in the course of studying: explain this again, set me practice questions, tell me why my answer is wrong.The crucial lesson here is that learning with AI may have advantages. In the help-me-learn category, children using AI about once or twice a week scored 501, ahead of every other group, including the 499 of children who never use it at all. Monthly users scored 495 and daily users 494, both within a few points of the abstainers. Only the once-or-twice-a-year group fell away, to 482. Compare that to those who use AI for drafting --- essentially, handing in AI-generated work, no level of use comes even close to the high learning scores of 509 who never do it.In other words, the numbers are separating children who work with a chatbot from children who let a chatbot do the work instead of doing it themselves. Nothing in PISA establishes that the tool causes either outcome, since the survey records what children say they do and what they scored on a single day in 2025, though the ordering holds across 91 school systems and matches what the OECD's education chief, Andreas Schleicher, wrote alongside the results: AI should be used as a "scaffold, not a crutch".The survey does dispose of the easy explanation. On the OECD's own measure of curiosity, daily users of AI are the most curious children in the sample on all four uses, and the least curious are the once-a-year users sitting near the bottom of the score table. Disinterest does not explain what is happening to the heavy users.Effort: the core of learningLearning happens in the space between not knowing something and working it out. Psychologists call the useful version of that space desirable difficulty: recalling, struggling and getting it wrong leave a deeper trace than reading a correct answer does. A chatbot closes that space before the child can learn from it.At the MIT Media Lab, 54 people wrote essays over four months with their brain activity recorded, in three groups — one using ChatGPT, one using a search engine, one working unaided. The chatbot group showed the weakest connectivity in the networks associated with attention and memory, up to 55% below the unaided writers. Asked later to write without AI, they still showed reduced activity, and struggled to quote sentences from essays they had submitted themselves. The researchers called it cognitive debt.The same pattern turns up in examination results. Researchers at Stockholm University and the University of Hong Kong followed more than 26,000 middle and high school students in a county in central China from September 2022 to June 2025. Homework scores rose 18% and the time spent on homework fell from 64 minutes to 45. Within six months, monthly test scores had dropped about 20%. Two years on, results in the zhongkao and gaokao entrance examinations were down 24% and 18%. The authors put it down to a deceptive sense of fluency — work that feels easy being mistaken for work that has been learnt.Whether the tool causes the loss has been tested once at scale. In Turkish high schools, in a trial published in the Proceedings of the National Academy of Sciences, nearly a thousand mathematics students practised with one of two versions of the same model: an unrestricted chatbot, or a tutor build that gave hints and withheld answers. While practising, the unrestricted group did 48% better than students working from textbooks alone and the tutor group 127% better. On the exam afterwards, sat without any AI, the unrestricted group scored 17% worse than the control group while the tutor group came out level — which places the design of the tool ahead of the fact of its use.Anthropic's India brief, published before the New Delhi AI Impact Summit in February, found that a fifth of Indian use of its Claude model is coursework, and that students ask for a direct answer nearly half the time.Also read: AI has changed the question, schools must change the answerNew strategiesHigher scores link to AI literacy lessons, but PISA data cannot prove whether the teaching itself drives the success or reflects broader school quality. (Unsplash/ Representational)PISA asked students one further question: whether their lessons had ever required them to judge the quality of information generated by AI. About six in ten across OECD countries said yes, ranging from 31% of students in some systems to more than 80% in others.Among the children who use a chatbot almost every day to study, those whose lessons had asked them to judge AI output scored 501 in science. Those whose lessons never had scored 487. Fourteen points separate them, around two-thirds of a school year, inside the group the rest of the data marks as most at risk. Among weekly users the same comparison is worth about six points, 506 against 501.Among children who rarely or never use AI, the comparison runs the other way. Children who never use a chatbot scored 493 where their lessons had asked them to assess AI output, and 497 where their lessons had not; among once-a-year users the figures are 476 and 484. Being taught to check a machine's answers shows up in the scores of the children who are actually using one.The children least likely to receive it are the poorer ones. The OECD found that socio-economically disadvantaged students are asked to assess AI-generated material less often than their better-off classmates, and warns of a new socio-economic divide opening in the age of AI. The same divide runs through the scores. Among daily users of AI for study, the raw numbers put them ahead of children who never use it, 493 to 489; matched for family background, they fall behind, 494 to 499. What reads as a benefit of heavy use turns out to be the wealth of the homes those children come from.At the level of whole school systems the pattern repeats. Japan has the lowest rate of AI use for schoolwork in the survey and Singapore one of the highest, and their 15-year-olds scored 538 and 560 in science, both far above the OECD average of about 500. Rates of AI use do not sort the strong systems from the weak ones.What PISA cannot settle is whether the lesson itself produces the gain, or whether schools that teach it differ in other ways as well. The OECD says as much, and asks for the targeted studies that would tell the two apart. Those studies have not been run. In the meantime, the strongest evidence available to an education ministry is a correlation measured on a single day in 2025, against a curriculum decision most of the world has already taken.HT Explainers Desk delivers in-depth explainers and analysis on the stories shaping India and the world, from SimplyPut, Hindustan Times’ dedicated explainer section. The desk unpacks the science, policy, geopolitics and market forces behind major news events, breaking down complex developments into clear, accessible reporting for readers who want more than the headline.