OECD: students who lean on AI for schoolwork score a year and a half behind
The first global test scores taken since generative AI went mainstream are in, and they are not the ones the industry has been promising. The OECD's PISA results, released September 8, cover more than 760,000 fifteen-year-olds across 91 countries — and they say the students who reach for a chatbot most often are the ones learning least.
Fifteen-year-olds who never or almost never use AI to draft writing assignments scored 509 in science, against 481 for students who use it every day or almost every day. That is a 28-point gap after adjusting for socio-economic status, and on PISA's scale 20 points equals roughly a year of schooling — so the daily-users are about a year and a half of teaching behind their abstaining peers. Similar patterns show up for other uses, including preliminary research on a new topic and summarising assigned reading. Only 14% of students say they scarcely or never use AI at all, though that number swings wildly by country: close to 40% in Japan, just 4% in Vietnam.
The interesting part is that the study does not find "AI makes students worse." It finds that how AI gets used decides everything. Students who use it about once or twice a week tend to outperform both lighter users and daily users, and among students who say they use AI "to help me learn," weekly users beat every other group — including the ones who never touch it. Among daily users, science scores run 13 points higher, more than half a year of teaching, when classes regularly ask them to assess AI-generated content. Andreas Schleicher, the OECD's director for education and skills, framed it plainly: "In the same way that we do not become fit by watching sports but by doing sports, learning does not occur through the consumption of content, but as a productive cognitive struggle of the mind with new material." His prescription is that AI be a "scaffold, not a crutch."
Our read: this is the first large-scale dataset on AI in schools, and it lands on the side of the skeptics — but it's a much more useful finding than a ban would be. The mechanism it points to isn't "AI is bad," it's that outsourcing the struggle is bad, and the fix is pedagogy, not prohibition. The policy world is moving fast in the other direction: we covered New York City bans student-facing AI through eighth grade earlier this month, and the UK's qualifications regulator is warning about AI cheating in coursework. Those moves now have a number attached to them — and a counterexample, since schools that teach kids to critique the output seem to come out ahead.
Tsinghua University and Tencent have opened China's first full-time master's program in "large model science and engineering." The first 27 students enrolled at an opening ceremony on September 8 at Tencent's Binhai headquarters in Shenzhen's Nanshan district, under a program the school calls the Qingyun Class. Tsinghua's Shenzhen International Graduate School signed the partnership with Tencent in July 2025; Tencent says it will hand students real problems, real data and real compute, with its Hunyuan model team supplying enterprise mentors. Tencent's VP of recruiting framed the bet around agents — "AI is moving from being able to talk to being able to do things." A first cohort of 27 is tiny, but it is a signal that China's labs would rather grow engineers than bid for them.
Amazon has named Kevin Mandia, the founder of Mandiant, to its board of directors. Mandia founded Mandiant and later sold it to Google for $5.4 billion; he is currently founder and CEO of Armadin and a general partner at Ballistic Ventures. Amazon cited his experience "combating cyber threats" in both the public and private sectors. It's a governance move rather than an AI one, but the timing is hard to miss — the same week Washington accused Chinese AI firms of systematic model theft, a cloud company that rents compute to frontier labs put one of the best-known incident responders in the room.
What to watch: whether the 28-point gap survives replication, and whether the weekly-use sweet spot holds up once schools start teaching AI critique at scale.
Is your school teaching kids to argue with the chatbot, or just letting them use it?
Sources: OECD PISA 2025 Results Volume I · The Verge · The Star / Bloomberg · OECD PISA dashboard · Tsinghua SIGS · Amazon · CNBC
