What are PISA 2025 Results telling us about AI in education?

A student struggling with a science problem can now turn to an Artificial Intelligence (AI) chatbot and, within seconds, receive a clear explanation and a well-structured answer. The assignment may be completed successfully.

But has the student learnt? This distinction between better task performance and better learning is becoming increasingly important as generative AI enters classrooms and homes. The Organisation for Economic Co-operation and Development’s (OECD) Programme for International Student Assessment (PISA) 2025 Results: Future-Ready Students offers some of the first large-scale evidence with which to examine this question.

AI use is already widespread among 15-year-olds. Across OECD countries, only 14% of students reported never or almost never using AI chatbots for the schoolwork purposes examined. The relationship between AI use and student performance is complex.

Students who did not use AI for specific tasks such as summarising texts and conducting preliminary research generally performed better in science than users. Among AI users, moderate users for these tasks tended to outperform both limited and frequent users.

These are associations, not evidence that AI causes higher or lower performance. Yet, they illuminate an important tension. A student can use AI to produce a stronger essay without becoming a stronger writer, generate a good summary without becoming a better reader, or arrive at the right mathematical answer without developing stronger reasoning.

Doing better at a task is not necessarily the same as becoming better at doing the task. Education is different from many other domains in which technology is deployed. In the workplace, reducing human effort may be the objective.

In learning, some of that effort is the objective. Retrieving something from memory, struggling with an unfamiliar problem, making mistakes and trying again may appear inefficient. But such cognitive effort is often how understanding develops.

AI can short-circuit this process by supplying the answer. Or it can strengthen it — by questioning a learner’s reasoning, identifying a misconception, offering a hint and allowing another attempt. One makes AI an answer machine.

The other makes it a thinking partner. Foundational skills matter more, not less There is a deeper paradox. As machines become better at reading, calculating and generating answers, it may be tempting to conclude that students need these capabilities less.

The opposite may be true. Across OECD countries, average reading performance fell from 489 points in 2015 to 461 in 2025, while mathematics and science also fell. PISA also highlights challenges in evaluating information, connecting sources and engaging critically with text — precisely the capabilities needed in an environment increasingly populated by AI-generated information.

The better machines become at generating information, the more important human capacity becomes to interrogate it. The foundational learning agenda for the AI age may therefore need to encompass three mutually reinforcing capabilities: reading, numeracy and computational problem-solving.

Reading enables students to comprehend and question information. Numeracy enables them to interpret evidence, reason quantitatively and test the plausibility of claims. Computational problem-solving enables them to frame problems, experiment, learn from feedback, and adapt.

PISA 2025 is significant because it assesses computational problem-solving through the new “Learning in the Digital World” domain introduced for the first time. It examines students’ capacity for iterative, self-regulated problem-solving using computational tools and practices.

This does not mean every child must become a programmer. It means every child increasingly needs the capability to remain intellectually in command while working with powerful machines. Another PISA finding offers a clue about how education systems might respond.

Around six in ten students across OECD countries reported opportunities at school to assess the quality of AI-generated information. Among these students, those frequently using AI “to help me learn” tended to perform slightly better than non-users and less frequent users.

Again, causality cannot be established.