PISA 2025 and AI: what the science results mean

PISA 2025 results show associations between AI use and science scores: students who never or almost never used AI for drafting, summarizing, or preliminary research scored highest in those comparisons. For summarizing reading, daily users scored nearly 30 points below nonusers. The OECD cautions that these findings do not prove that AI caused the differences.

The OECD published the first volume of results in September 2026. The assessment included more than 760,000 students across 91 countries and economies, with science as its focus. PISA assesses 15-year-olds and supports comparisons across education systems; it is not your school exam and does not predict your grade. When reading the findings, distinguish the science test scores from students’ reported habits of using AI.

The OECD reports that 46% of students in its member countries use AI chatbots at least weekly. This describes frequency; by itself, it does not tell you whether the conversations involved studying, drafting an assignment, or requesting a solution. It also does not establish how many students checked the responses.

Read the task and frequency together. Summarizing reading every day is different from asking for help learning each week. These comparisons do not establish an ideal schedule for using AI. Knowing how often someone opens a tool tells you little without knowing what they do with it.

For drafting, summarizing, and preliminary research, students who never or almost never used AI had the highest science scores among the groups compared. Differences averaged about 20 points, described as close to a school year. That comparison puts a group-level gap in perspective; it does not mean every student who used AI lost a year of learning.

For summarizing reading, the gap between daily users and nonusers was nearly 30 points. Occasional use, once or twice a year, was also associated with lower scores than nonuse and weekly use. These patterns do not provide a ranking of good and bad habits or establish what caused the differences in performance.

There is a relevant exception. For use described as “to help me learn,” weekly users scored highest, at around 500 points, with nonusers slightly behind. The contrast suggests that the purpose of using the tool deserves attention alongside frequency.

The result does not show that a chatbot produced that performance. Prior differences between students or other study habits could be involved; that is a possible interpretation, not an additional finding. In your own routine, ask a concrete question: after receiving help, can you explain the idea and solve a variation without consulting the answer again?

Across the OECD, reading fell 28 points and mathematics 22 points between 2015 and 2025. These declines matter for the education discussion, but they should not be attributed to chatbot use: the cited data do not establish that causal connection.

Across the OECD, one in five 15-year-olds is a low performer in science, mathematics, and reading simultaneously, compared with 16% in 2022. This describes difficulties across all three areas at once. It reinforces the need to examine the wider education context rather than explain the results through a single technology.

Mathias Cormann, the OECD’s secretary general, emphasized that technology and AI strengthen learning when used purposefully. This is guidance for making study decisions, not proof that a particular use explains every result. Access to a tool also does not tell you how each student uses it.

In practice, compare receiving a summary before reading with answering questions after studying the text. You can also compare copying a solution with writing your own steps first and requesting feedback on one of them. These are suggestions for your routine, not interventions tested by PISA. In either case, check the help you receive against the course material.

For one week, keep a record of your AI use. Each time you open a chatbot, note the task, what you had already tried, and what you did after receiving the response. Distinguish between having the tool produce something you needed to practice and using a hint to continue your own attempt. This record can guide your next activity; it cannot establish causation.

Then change one activity. If you receive summaries without reading, try answering questions about the text. If you copy solutions, attempt a problem first and keep your work. When a conversation helps, note which question was useful and check the topic again in another session. You do not need to replace all your tools to identify where more practice is needed.

Lumia AI lets you create or edit flashcards, generate AI cards and quizzes from notes or PDFs, organize material into books, and review with spaced repetition based on your ratings. It also offers a chat tutor, learning paths, and Anki .apkg deck import, with web and mobile apps.

To apply the routine, choose one topic, check generated cards against your source, and answer before revealing each solution. Correct ambiguous questions and keep your written steps and references in your own notes. If an idea remains difficult, return to the material or ask your instructor. A difficult attempt identifies something to review; it does not establish that the tool or your effort caused the problem.