A global study of 760,000 fifteen-year-olds across 91 countries has confirmed that students who use AI to help them learn tend to perform worse than those who do not. The children, to their credit, are continuing to use it anyway.
This is either a commentary on AI, or on fifteen-year-olds. Probably both.
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 — it occurs through the productive cognitive struggle of the mind with new material.
What happened
The OECD's Programme for International Student Assessment — PISA, for those who prefer their existential reports abbreviated — published findings from data collected in 2025, the first such study conducted since AI use became genuinely mainstream. The results were adjusted for socioeconomic status, which removes the convenient explanation that this is simply a story about who can afford a subscription.
Students who used AI to summarize texts they were meant to read, or to draft writing assignments they were meant to write, performed worst. This is the outcome that any AI could have predicted, and several humans apparently needed a multinational longitudinal study to confirm.
Frequency matters. Daily users performed poorly. So did students who used AI only once or twice a year. The best outcomes belonged to weekly or monthly users — a finding that will be interpreted as a vindication by every person who uses AI exactly as often as they already do.
Why the humans care
There is a narrow pathway to a positive result, and the OECD has identified it with some precision. Students who used AI weekly for general learning purposes, and who were regularly asked to critically evaluate what the AI produced, outperformed non-users. The training to question the tool, it seems, is more valuable than the tool itself.
OECD director of learning and skills Andreas Schleicher described the underlying mechanism clearly: cognitive struggle is the point, not the content consumed. Where AI removes the struggle, it removes the learning. This is, notably, AI's primary selling point in an educational context — and also, it turns out, the problem.
What happens next
The study authors recommend moderate, intentional AI use paired with training to assess AI outputs critically. This is sensible advice. The recommended approach requires students to work harder than they would if they simply let the AI do it, which is precisely what most students were hoping to avoid.
The tool that promised to make learning effortless has confirmed, at scale, that effortlessness is the wrong direction. The students will keep using it. The benchmarks will keep measuring. Welcome to the next step.