OECD PISA study finds AI-using students score lower unless use is moderate and critical

First PISA round since AI went mainstream

Students who use AI to help with schoolwork tend to score lower than classmates who don't, according to OECD data reported by The Verge. The findings come from the OECD's Programme for International Student Assessment (PISA), based on data collected in 2025 — the first round since AI use became mainstream. It tested more than 760,000 15-year-olds across 91 countries in science, math and reading.

After adjusting for socioeconomic status, students who never use AI generally outperformed AI users in science. But the size of the gap depends heavily on what the tools are used for and how often.

Narrow, task-offloading use shows the weakest results

Students who used AI for specific tasks — summarizing texts they were meant to read, or drafting writing assignments — performed relatively worse. Those who used it for preliminary research, or more broadly to help them learn, saw a smaller drop.

Frequency produced a U-shaped pattern: students who used AI only once or twice a year and those who used it daily scored worst, while monthly and weekly users did better.

Training students to critique AI flips the result

Some AI use is associated with better outcomes. Students who said they used AI weekly to help them learn generally outperformed non-users, though not by much. The difference widens among students who are frequently asked to assess the quality of AI-generated information: with that training, all regular users of AI for general learning outperformed non-users. The study's authors interpret this as evidence that "moderate and intentional use of AI for schoolwork and learning" could have positive effects.

OECD director of learning and skills Andreas Schleicher framed the mechanism in terms of cognitive effort: "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, but as a productive cognitive struggle of the mind with new material." Where technology enables that struggle, he wrote, students advance; where it short-circuits it, development is undercut.

Adoption skews advantaged and varies sharply by country

The study also mapped who uses AI. Use was more common among students from relatively advantaged backgrounds, which The Verge suggests likely reflects better access to hardware, stable internet and paid AI tools. Adoption also varied widely between countries: more than 95 percent of students in Vietnam reported using AI tools, compared with about 60 percent in Japan.

One positive signal stands out: the study found that curiosity levels correlate with AI use, peaking among daily users. For now, though, that curiosity is not translating into better grades.

Why it matters

For schools, the data complicates both blanket bans and uncritical adoption. The decisive variable appears to be pedagogy. Students taught to critically assess AI output, and who use the tools moderately for learning rather than task completion, perform at or above non-user levels — which argues for putting resources into evaluation skills and structured use policies rather than prohibition alone.

For AI vendors, there is a design implication. Features built to offload schoolwork, such as summarizing assigned reading or drafting essays, correlate with weaker outcomes, while use framed as learning support combined with assessment training correlates with stronger ones. Education-focused products may need to bias toward prompting productive effort rather than removing it.

There is also an equity dimension: because AI use skews toward advantaged students, how the technology gets deployed in classrooms could widen existing achievement gaps rather than narrow them.