The latest OECD Programme for International Student Assessment (PISA) delivers a sobering reality check for educators betting big on generative AI. Based on 2025 data from over 760,000 students across 91 countries, the report finds that students who never use AI generally outperform their peers in science after adjusting for socioeconomic status. The drop is sharpest for those using tools to summarize texts or draft assignments—tasks that effectively outsource the mental heavy lifting. It’s a clear signal that convenience comes at a cognitive cost.
But the story isn’t purely negative. Frequency and intent matter immensely. Students who used AI daily or only once or twice a year performed the worst, while moderate weekly users saw better results. In fact, those using AI weekly for general learning support slightly outperformed non-users. The real breakthrough, however, appears when schools pair AI access with specific training. When students are taught to critically assess the quality of AI-generated information, regular users don’t just catch up—they surpass non-users. This suggests the tool itself isn’t the problem; passive consumption is.
Andreas Schleicher, OECD director of learning and skills, frames this as a issue of cognitive effort. “Learning does not occur through the consumption of content, but as a productive cognitive struggle of the mind with new material,” he writes. Technology that short-circuits this struggle undercuts development, while tech that enhances it drives advancement. This distinction is crucial for curriculum designers. Simply banning AI ignores its potential, but unfettered access risks creating a generation of students who can prompt but cannot think.
Access remains uneven, with AI use correlating strongly with advantaged backgrounds and varying wildly by region—over 95% of Vietnamese students reported using AI, compared to 60% in Japan. Interestingly, AI users report higher levels of curiosity, even if their grades don’t always reflect it yet. The challenge for schools isn’t just providing access, but teaching students how to wrestle with the output rather than just accepting it. Without that critical layer, AI remains a crutch rather than a catalyst.
