Students who use AI for schoolwork score lower, but one exception matters

The OECD's 2025 PISA assessment, released yesterday, finds that 15-year-olds who report using AI tools for learning perform worse in science, mathematics and reading than peers who do not use them at all. The more interesting signal sits inside the breakdown: not all use is equally harmful, and under certain conditions it is beneficial.

What breaks the average

After controlling for socio-economic background, the deficit for AI users appears most clearly in science. Students who used tools to summarise texts they were supposed to read, or to draft assignments, showed the steepest declines. Those who turned to AI for preliminary search or the broad category "help me learn" suffered a more moderate hit. The report's conclusion: moderate, purposeful use may be positive, but current classroom defaults are far from that.

Frequency is not linear

The data reveal a surprising curve. Students who used AI once or twice a year, and those who used it daily, recorded the worst performance. Monthly or weekly users achieved better results. Among the group using AI weekly for "general learning help," a slight edge over non-users appeared, and when regular training in evaluating the quality of AI output was added to the model, the advantage grew and became statistically significant for all regular users.

An access gap and a country gap

AI use is significantly more common among students from more affluent backgrounds, reflecting access to hardware, stable internet and paid tools. Internationally the disparities are extreme: over 95% of students in Vietnam reported using AI tools, versus only 60% in Japan. That figure alone demands caution in generalising; the global average masks vast variation in infrastructure, classroom culture and school policy.

Curiosity does not yet translate to scores

One bright spot: self-reported curiosity correlates with AI use and peaks among daily users. But that correlation does not currently convert into academic achievement. As Andreas Schleicher, the OECD's director for education and skills, put it: learning does not happen through content consumption but through productive cognitive struggle with new material. Technology that shortens that struggle harms development; technology that enables it promotes it. The practical takeaway is straightforward: without a pedagogical framework that forces the student to evaluate, critique and process what the AI produces, the tool remains a shortcut that leaves the learner in place.