PISA 2025 Links Heavy Schoolwork AI Use to Lower Science Scores

PISA 2025 delivers a warning about how students use generative AI, rather than a simple verdict on whether schools should allow it. Across more than 760,000 15-year-olds in 91 countries and economies, frequent chatbot use for tasks that can substitute for students’ own reading, research, or writing was associated with lower science performance. The study’s comparisons are observational, so they cannot establish that AI use itself caused the score gaps. 14

The headline finding: a 28-point gap in science

Students who said they never or almost never used AI to draft writing assignments averaged 509 on the PISA science assessment. Students who used AI for that task daily or almost daily averaged 481—a 28-point difference after socioeconomic adjustment, according to reporting on the results. 25

Similar patterns appeared when students used AI to conduct preliminary research or summarize assigned reading. One report describes gaps of roughly 20 points on average for those task-specific uses, while the gap for frequent reading-summary use was close to 30 points. 612

That does not mean a chatbot necessarily lowered any individual student’s score. Students who turn to AI most often may also differ in prior attainment, motivation, home circumstances, access to teaching, or the kinds of work they are asked to do. The responsible conclusion is narrower: intensive, task-substituting use and weaker science outcomes appeared together in the PISA data. 4

AI use was already mainstream

Only 14% of students reported avoiding AI for schoolwork altogether, meaning 86% had used it at least occasionally. Reported non-use varied widely across education systems—from roughly four in ten students in Japan to about 4% in Vietnam. 1

This matters because the question for schools is no longer whether students will encounter chatbots. It is whether classrooms set expectations that preserve reading, reasoning, source evaluation, and original writing.

Frequency and purpose appear to matter

The reported relationship was non-linear. Moderate use—about once or twice a week—especially when students used AI to help them learn, was associated with better outcomes than either no use or daily use in reporting on the PISA findings. 9

That distinction is practical. Asking a chatbot for feedback on an explanation, alternative examples, or questions to test understanding is different from having it generate a reading summary, research notes, or an assignment draft that the student does not independently examine.

The evidence does not show that every use of AI harms learning. It does suggest that using it as a shortcut around the cognitive work of reading, drafting, and checking evidence may be a poor fit for the skills PISA measures.

Teach students to evaluate AI, not just use it

Instruction appears to be an important part of the picture. Among frequent users, students who had been taught to assess AI-generated information scored about 13 points higher than peers who had not received that instruction, according to the reported findings. 4

The useful classroom model is therefore AI as a scaffold rather than a crutch: a tool students must question, verify, and improve upon. A teacher-guided workflow can require students to:

• identify unsupported claims or errors in an AI response;
• compare an answer with assigned sources;
• explain the reasoning behind the final answer;
• revise an AI-generated draft in their own words; and
• document what the tool contributed and what the student changed.

These practices keep students accountable for the thinking that produces learning, rather than merely for submitting a polished result.

A result set against broader learning declines

The AI findings arrived amid a wider decline in OECD-country performance. From 2015 to 2025, average reading scores fell 28 points and mathematics scores fell 22 points; science declined by 7 points over the same period. 17

Those longer-term changes should not be attributed to generative AI alone. The score trend began before chatbots became widely available, and reporting around the release points to broader pressures on engagement, reading, and digital attention. 13

What schools should take from PISA 2025

PISA 2025 supports a cautious, evidence-led approach—not a claim that every student use of AI should be banned. Schools can draw a clearer line between AI that supports learning and AI that substitutes for it.

For high-stakes reading, research, and writing, students should remain responsible for the underlying work. Where AI is permitted, use should be structured around verification, explanation, and revision. The central lesson is straightforward: the educational value of a chatbot depends less on whether it is present than on whether it makes students think more deeply—or lets them avoid thinking altogether. 49