Picture two students prepping for the same science exam. One types her assignment into a chatbot, gets a polished paragraph back in ten seconds, and moves on. The other struggles through a rough draft alone, deletes half of it, and rewrites the argument twice. Six months from now, according to the OECD's newest and largest study of teenagers on the planet, the second student is significantly more likely to actually understand the material — and to prove it on a test that doesn't let her cheat.
That's the uncomfortable pattern buried in PISA 2025, the OECD's flagship education report, released September 8, 2026.
The headline number
PISA surveyed more than 760,000 fifteen-year-olds across 91 countries, asking not just what they know but how they're using AI to learn it. The result: teenagers who said they never or almost never use AI chatbots to draft their writing assignments scored 509 on the science test. Students who leaned on AI for that same task every day, or nearly every day, scored 481.
That's a 28-point gap — and it holds even after adjusting for students' socio-economic background. On the PISA scale, that's roughly the equivalent of a year and a half of schooling. Similar gaps turned up for other everyday uses of AI, including using it to do preliminary research on a new topic.
It's not an isolated finding
The timing makes the finding harder to dismiss as noise. PISA 2025 also recorded the lowest OECD-average scores in the test's history: science slipped from 489 points in 2015 to 482 in 2025, reading is down 28 points over the same period, and math has fallen 22 points. One in five 15-year-olds across the OECD is now a low performer in reading, math, and science simultaneously — up from roughly one in six just three years ago.
AI isn't the only culprit behind a decade-long slide that predates ChatGPT. But it's the newest variable, and other data points are lining up in the same direction. A 30-month study tracked nearly 27,000 middle- and high-school students in China, published as a CEPR working paper in August 2026.
The pattern it found: AI use raised homework scores by 18% and cut assignment time by 30% — while closed-book exam scores fell 20% within six months, and entrance-exam scores dropped 18–24%, with the damage peaking after about two years. The tool made the practice easier and the learning weaker, at the same time, and the gap didn't show up until students no longer had the AI in front of them.
Tellingly, the damage wasn't spread evenly. The researchers traced most of it to the roughly 80% of AI-using students who used the tool to "outsource" their homework — finishing assignments unusually fast with suspiciously high scores — while students who treated it more like a study aid kept most of their learning gains. Same tool, same classroom, wildly different outcome depending on how it got used. That's the exact pattern the OECD is describing half a world away.
Schools are already moving — just not in the same direction
Institutions aren't waiting for a scientific consensus to act, and they're reaching for three different levers.
New York City Public Schools — the largest district in the US — went for outright removal: a one-year moratorium on student-facing AI tools for its elementary and middle schoolers, about 600,000 students in all, for the 2026–27 school year.
England's exam regulator went for tighter policing instead of removal. Ofqual chief Ian Bauckham warned in July 2026 that written coursework will face "far, far more scrutiny," arguing schools cannot let AI-generated work pass as a substitute for a student's own.
And MIT's own committee on AI in teaching went further than either. It concluded in August 2026 that generative AI can already produce credible answers to almost any written assignment in its undergraduate curriculum, so banning or policing it is a losing race. Its fix: redesigning assessments around oral exams, portfolios, and in-class work AI can't easily stand in for.
Three institutions, three different playbooks — ban it, police it, or design around it — but the same underlying diagnosis: once AI is quietly doing the part of an assignment that was supposed to build a skill, the grade stops meaning what it used to.
The OECD's actual argument is narrower than the headline
Here's the part that's easy to miss in a scary headline: the OECD isn't telling schools to rip AI out of classrooms, and its own data doesn't support a simple "more AI, worse scores" story. Students who use AI once or twice a week actually outperform both the daily users and the students who never touch it — and among daily users specifically, science scores climb 13 points, worth more than half a year of teaching, when those students are regularly asked to critically evaluate what the AI has told them rather than just accept it.
The damage isn't coming from AI itself; it's concentrated in unsupervised, uncritical, everyday dependence.
OECD Director for Education and Skills Andreas Schleicher put the underlying idea more bluntly: AI should be a "scaffold, not a crutch." The logic borrows from an old idea about physical fitness — you don't get fit by watching someone else exercise, and understanding doesn't come from passively consuming a finished answer; it comes from the struggle of producing one yourself.
Researchers call the failure mode cognitive offloading: when a tool consistently does the thinking for you, the mental muscle you'd otherwise be building quietly atrophies. That's a plausible mechanism for why daily AI drafting correlates with weaker science scores even though writing an essay and understanding science seem, on the surface, unrelated. The skill being eroded — working through a problem yourself instead of taking the first answer offered — turns out to generalize.
The debate isn't settled
None of this has stopped AI's advocates from making their case. Supporters argue that, done right, AI could deliver genuinely personalized instruction — especially valuable for struggling students who need a different pace than the rest of the class — and free up teachers from paperwork so they can spend more time actually teaching. The OECD itself is threading that needle: in June 2026, it and the European Commission finalized an AI literacy framework for primary and secondary schools, built around teaching students to question and evaluate what a model tells them rather than simply accept it — the same habit the PISA data suggests separates the students AI helps from the ones it quietly hurts.
What it means, practically
If there's a single actionable takeaway in the report, it's this: the question schools should be asking isn't "AI or no AI," but "is this use of AI doing the student's thinking, or supporting it?" A chatbot that drafts an entire essay on request is doing the former. A tool that checks a student's own draft, flags weak reasoning, and asks a follow-up question is closer to the latter. The OECD's data suggests that distinction is worth far more than a blanket ban — or a blanket embrace.
Source: OECD, PISA 2025 Results (Volume I): Future-Ready Students, published September 8, 2026.
Comments (0)
Login to post a comment.