Students who use AI generally score worse at school
Data from the OECD's 2025 PISA survey show that, overall, 15-year-olds who use AI tools for school tend to score lower than nonusers after accounting for socioeconomic status. However, the effect varies by how students use AI and how often, with some types of moderated, intentional use — especially when students are taught to evaluate AI outputs — linked to small performance gains.
Why It Matters
The findings come from the first PISA round conducted since AI use became widespread and draw on a large international sample, so they offer early, broad evidence about how classroom and study-time AI use relates to learning outcomes. That nuance — harm in some patterns of use and potential benefit in others — has direct implications for school policy, instruction, and digital access.
Key Facts
- study: OECD Programme for International Student Assessment (PISA) 2025
- sample: over 760,000 students
- countries: 91 countries
- students-tested: 15-year-olds in science, math, and reading
- overall-finding: AI users generally scored worse in science after adjusting for socioeconomic status
The OECD’s latest PISA cycle, based on data gathered in 2025, found that students who reported using AI for study generally underperformed compared with peers who did not use such tools. PISA tested more than three-quarters of a million 15-year-olds across 91 countries in science, math and reading, and this round is the first to reflect the era when AI use became commonplace among learners.
When researchers controlled for socioeconomic background, the gap was clearest in science. How students used AI mattered: those who relied on it for specific tasks like summarizing assigned texts or drafting written work tended to do relatively worse, while students who used AI for preliminary research or described using it broadly to “help me learn” showed smaller declines in performance.
Frequency of use also shaped outcomes. Students who used AI only once or twice a year and those using it daily were among the worst-performing groups, whereas monthly or weekly users generally fared better. Notably, students who used AI weekly for general learning purposes slightly outscored nonusers. The advantage became more pronounced when students were regularly asked in class to evaluate the quality of AI-generated information: with that additional practice, regular AI users for general learning outperformed nonusers, suggesting benefits from intentional, reflective use of the tools.
The report highlights uneven access and uptake: AI use was more common among relatively advantaged students, a pattern the OECD links to better access to devices, reliable internet and paid AI services. International variation was large as well — more than 95 percent of students in Vietnam reported using AI, versus about 60 percent in Japan. Curiosity levels were higher among AI users and peaked with daily users, but higher curiosity has not consistently translated into better scores under current patterns of AI use. OECD learning director Andreas Schleicher summarized the study’s educational point: technology can accelerate learning when it supports the active cognitive work of students, but it can harm development when it short-circuits that productive struggle.
Keep Reading
Google DeepMind alumni are building tools to accelerate fusion power for the grid
Suno replaces its AI models with a new one trained on licensed music as copyright suits pile up
Besxar is building an orbital semiconductor factory, one SpaceX rocket at a time
