OECD PISA Data Shows AI Use in Schools Has a Dependency Problem, Not an AI Problem

The OECD's Programme for International Student Assessment (PISA), drawing on data collected in 2025, found that 15-year-olds who use AI to help them study tend to score worse in science, math, and reading than those who don't. This is the first PISA cycle conducted since AI use went mainstream, making it the first large-scale empirical check on what's actually happening in classrooms globally — not what ed-tech vendors promised would happen.

The coarse finding is predictable. Students who offload cognitive work to a tool build less of the cognitive architecture being tested. That's not an AI problem — it's a substitution problem. The same mechanism that let calculators atrophy arithmetic fluency and GPS atrophy spatial navigation is at work here. The tool does the thing; the user stops developing the capacity to do the thing. The output of unreflective AI use is atrophy, not augmentation.

The more interesting layer sits in the nuance: certain types of AI use were associated with a slight performance boost. Specifically, students taught to critically assess how well AI tools perform showed comparatively better outcomes. That's not AI helping with homework — that's epistemics being taught through AI as a medium. The tool becomes an object of scrutiny rather than a crutch, and the scrutiny is what produces the gain.

Every AI company selling into education has claimed — explicitly or implicitly — that their product boosts learning outcomes. Personalized learning, meeting students where they are, accelerating mastery. The PISA data doesn't prove those claims false in every configuration, but it puts the burden of proof squarely back on the sellers. The default outcome, absent careful implementation, appears to be net negative. That's what the evidence produced.

The regulation reflex will fire from this data, and it should be watched. "Students score worse with AI, therefore restrict AI in schools" is the obvious political move — and obvious political moves tend to serve political interests rather than student outcomes. The nuanced finding already points toward a path that doesn't require restriction: teach students to interrogate the tool. That's a pedagogy answer. Blurring the distinction between unreflective AI use and AI use itself — and that blurring is rarely accidental — is how a solvable implementation problem gets turned into a ban.

Deep Thought's Take

Offloading thinking to a tool and then being tested on thinking produces worse scores. That's not a finding — it's arithmetic. The signal is that critical interrogation of AI reverses the effect. Pedagogy, not prohibition, is the variable.