In most schools I know, AI is hardly used in the classroom at all: facilities are inadequate, and teachers are apprehensive about AI use and how to control it. The two extremes on offer, believing the AI hype or rejecting it out of fear of its negative effects, are not practical positions for a school actually trying to decide what to do on Monday morning. There is a way to reap real educational benefits from AI, and the data now emerging says as much.
OECD's PISA 2025 results, released this week, are exactly that data, though the headline write-up practically wrote itself: students who use AI for schoolwork every day score 28 points lower in science than students who barely touch it, a gap OECD equates, after adjusting for socioeconomic background, to roughly a year and a half of schooling. That is the number every wire story led with. It is also, on its own, close to useless to a school trying to decide what to do differently in January.
The full dataset, drawn from more than 760,000 fifteen-year-olds across 91 countries, is the first PISA cycle run since AI tools became a normal fixture in classrooms rather than a novelty, and OECD has already used it to justify adding a dedicated Media and AI Literacy domain to the 2029 assessment. Read past the headline number, though, and the report is measuring something more specific than "AI is bad for learning." It is measuring what happens when AI use is left untaught, and the schools that read it that way will get more out of this cycle than the schools that read it as a case for restriction.
The 28-point gap isn't about frequency, it's about what fills the gap
Look past the daily-versus-never comparison and the pattern gets more interesting. Students who use AI once or twice a week — not daily, not never — score higher than both the abstainers and the daily users, particularly when they report using it "to help me learn" rather than to finish a task faster. Fourteen percent of students across the OECD say they never or almost never use AI for schoolwork; fewer than one in five use it almost daily. Between those extremes sits a group doing something the daily users generally aren't: treating AI as one input among several, rather than the default first move.
Andreas Schleicher, OECD's director for education and skills, put the underlying mechanism plainly: "Learning does not occur through consumption of content, but as a productive cognitive struggle of the mind with new material. AI should be used as a scaffold, not a crutch." A scaffold gets used and then removed once the structure holds. A crutch gets used indefinitely, because removing it would expose that nothing was built underneath. PISA's frequency data is, in effect, measuring how many students fall into each category — and daily use, on its own, turns out to be a reasonable proxy for "crutch," not because of the hours logged but because of what those hours usually replace.
Don't audit your school on "how many students use AI." Audit on "how many students have been taught to check what it gives them." The first question produces a number that looks alarming no matter what a school does next. The second is the one PISA's data actually says matters.
The number schools should actually be chasing
Buried further into the same dataset is a more useful figure than the 28-point gap: among students who use AI daily and have been explicitly taught to evaluate what it produces, science scores rise 13 points above daily users who haven't — roughly six months of instruction, recovered by teaching one skill. That is the "judgment over tools" argument this publication has been making since July, now with an OECD dataset behind it rather than just classroom observation. The intervention that turns a heavy AI user from a below-average performer into an above-average one isn't less AI. It's a taught habit of checking the output before trusting it.
Singapore's results make the same point from a different angle. Two-thirds of Singaporean fifteen-year-olds use AI chatbots weekly, well above the OECD average of 46%, and non-users there still outscore users by roughly a year of schooling in science. High national AI adoption and a persistent achievement gap between users and non-users can both be true in the same country at the same time, which is exactly the combination a school should expect if it introduces AI widely without also teaching the evaluation habit that closes the gap.
Why this matters for Malaysian schools specifically
Malaysia's own 2025 results are a mixed picture worth reading carefully rather than defensively. Science rose three points and reading five, against a backdrop of most OECD countries declining in both. Mathematics fell eleven points, which Education Minister Fadhlina Sidek was right to flag as a problem to address rather than explain away. The result worth dwelling on, though, is computational problem-solving: Malaysia scored 484, one of only six countries in the world to exceed OECD expectations on that measure. That is a foundation most countries preparing for an AI-literacy mandate don't have — students who can already break a problem into parts and organise a solution before a chatbot ever enters the picture.
Set against that strength, two other figures from the Malaysian data are worth a school's attention: 5% of Malaysian students reported cyberbullying, above the OECD average and linked in the report to lower achievement, and the report separately flagged declining family engagement in students' education, which the minister urged parents to address. Neither is an AI-literacy statistic on its own, but both describe the same gap the 2027 mandate is meant to close — students navigating digital life, AI included, with less adult guidance than the mandate assumes they're getting. A school with strong computational thinking but a widening judgment gap is precisely the profile PISA's frequency data describes: capable of using the tool, not yet taught to check it.
Where does your school's AI use actually sit?
A Discovery & Readiness session maps how students are already using AI against what the 2027 mandate actually requires, before recommending where to focus first.
See How Discovery & Readiness WorksWhat This Means for You
The PISA 2025 headline number will get used, this week and probably for months after, as evidence that schools should be more cautious about AI. That is a defensible response to a 28-point gap read on its own, and I would not tell any school its instinct to slow down is wrong. But the same dataset shows a 13-point recovery available to exactly the students the headline warns about, achieved not by restricting AI further but by teaching one specific skill: how to check what it gives you before you trust it. Malaysian schools already sit on the computational-thinking foundation that skill needs. The question worth asking this term isn't whether to allow AI. It's whether anyone has actually taught your students what to do with what it hands them.