Global study: students who use AI for schoolwork often score lower — but not always
A new OECD test of 760,000 teens finds AI use is linked to weaker scores overall, except when it's moderate, purposeful, and paired with training to question the AI's answers.
What happened: The OECD's PISA test, given to over 760,000 15-year-olds in 91 countries in 2025, is the first major global study since AI use became mainstream in schools. After controlling for socioeconomic status, students who never used AI generally scored higher in science than those who did. But the gap wasn't uniform: it depended heavily on what students used AI for, how often, and whether they'd been taught to question its output.
Why it matters: This is the first hard, global evidence on whether AI helps or hurts learning at scale, and the honest answer is: it depends on the habit, not the tool. Policymakers, schools, and parents have been guessing at rules for AI in classrooms. This data gives them something concrete to argue over, and it complicates the simple story that AI in schools is either a menace or a miracle.
How it works, plainly: Using AI to summarize readings or draft essays for you was linked to worse performance. Using it for early research or general 'help me learn' support was linked to a smaller drop, or even a gain. Frequency mattered too: occasional (once or twice a year) and daily users did worst, while weekly or monthly users did best. The clearest bright spot: students who used AI weekly for general learning help, and who were also regularly trained to critically assess AI-generated answers, outperformed non-users.
The rollout: AI use itself isn't evenly spread: it's more common among students with better internet, devices, and paid tools, and it varies wildly by country, from over 95% of Vietnamese students to 60% in Japan. OECD's Andreas Schleicher framed the takeaway as a fitness analogy: learning comes from struggling with material, not passively consuming it, so AI helps only when it doesn't remove that struggle.