OECD links most student AI use to weaker PISA performance
An OECD analysis found that AI use was generally associated with poorer results, while moderate use paired with training to assess AI-generated information produced slightly better outcomes.
Students who use artificial intelligence for schoolwork generally perform worse than non-users, according to an OECD analysis of more than 760,000 15-year-olds across 91 countries.
The Programme for International Student Assessment (PISA) examines performance in science, mathematics and reading. The analysis uses data collected in 2025, in the first PISA cycle since generative AI became mainstream.
After adjusting for socioeconomic status, students who did not use AI generally outperformed users in science. Use for tasks including summarising assigned texts and drafting writing was associated with relatively poorer results, while students using AI for preliminary research or general learning saw a smaller performance gap.
Frequency also mattered. Students using AI once or twice a year or daily tended to perform worst, while monthly and weekly users recorded better results. Weekly use for general learning was associated with slightly better outcomes than non-use, particularly among students trained to assess the quality of AI-generated information.
The OECD said the findings suggest that moderate, intentional use may be beneficial when it supports active learning rather than replacing it. The relationships are correlational and do not establish that AI use caused either poorer or better performance.
AI use was more common among students from advantaged backgrounds and varied widely between countries. More than 95 per cent of students in Vietnam reported using AI tools, compared with about 60 per cent in Japan. AI users also reported higher curiosity levels, which peaked among daily users.
