PISA hits a 20-year low and AI gets the blame — but the slide began back in 2000

Global teenage test scores have fallen to their lowest level in two decades. According to OECD data reported by the New York Post, reading and mathematics scores on the 2025 Programme for International Student Assessment (PISA) — the test the OECD gives every three years to 15-year-olds, taken this cycle by 760,000 students across 91 countries, per the Post — dropped by nearly 20 points. The correlations that grabbed the headlines are genuinely striking: students who rarely use AI scored roughly 30 points above regular users, and teens who used social media heavily at school scored about 44 points below light users in maths. Only three countries with wide AI availability improved in reading, four in maths and five in science.

Our thesis: the causal story in the headline is the weakest part of the report. Reading scores are down 34 points since 2000 and maths down 32 points since 2006, per the OECD figures cited in the article. Generative AI chatbots did not exist for the vast majority of that window — they became mass-market products roughly three years ago. Whatever has been eroding adolescent literacy for a quarter of a century, it is not a technology that shipped in 2022. AI may be pouring fuel on a fire; it did not light it.

The OECD is more careful than the coverage. "We cannot say for sure," the report itself says of the social media link, "but skimming social media feeds and rapidly processing information may be contributing to worsening ability and motivation to engage with complex texts." That hedge matters, because the data is a snapshot, not an experiment: a student already struggling with a dense text is exactly the student most likely to reach for the shortcut. The arrow can point the other way. And the league table complicates the narrative further — the article notes that China, Singapore and Japan top all three categories. These are not low-tech, screen-free societies. Whatever separates the top from the middle, it is not simply the absence of devices.

What is durable in the report is the mechanism, and it is worth taking seriously. Andreas Schleicher, the OECD's director for education and skills, put it bluntly: "You don't get fit by watching sport... If AI makes things easy, it's not great for you in the long run. It gives you better results, but it doesn't enhance learning." Secretary-general Mathias Cormann pointed to the same thing from the other side — more screen time, less reading for pleasure, and "more hasty reading, students rushing through a text and giving a quick wrong answer." That is not an indictment of AI as such. It is an indictment of a product design that optimises for a satisfying answer in two seconds rather than for the effortful struggle that actually builds a brain.

The American picture, meanwhile, is a reminder that domestic policy still dominates. The US sits 24 points above the global reading average and improved in science (up 3 points since 2022, up 13 since 2000), but has fallen 24 points in maths since its 2009 peak and now ranks 27th in that subject. A slide that begins in 2009 and runs through three presidencies, a pandemic and a curriculum war is not a chatbot problem.

Our read: PISA 2025 is the first hard, global baseline of the AI era in education, and it should be treated as a starting instrument rather than a verdict. The honest short-term conclusion is uncomfortable — the default consumer AI product, which hands over finished output on demand, is badly matched to how learning works, and schools that adopted it without redesigning assessment are now measuring the consequences. We have tracked the institutional swing from banning AI outright to mandating it; this data suggests both reflexes are wrong, and that the variable that matters is whether the tool demands effort or removes it. The long-term case remains intact and, if anything, more urgent: personalised tutoring at scale is one of the clearest routes by which AI could lift human capability broadly rather than narrowly. But that future is a design choice, not an automatic consequence of access. Build tools that make students work harder, measure them against PISA-grade evidence, and stop regulating on the strength of a correlation.