The question facing schools is no longer whether teenagers use generative AI, but what they use it for—and whether the way they use it helps them learn. The OECD’s PISA 2025 Results, published on 8 September 2026, offers a large-scale view of students’ reported AI use for schoolwork and how different patterns relate to science performance. Its findings are more nuanced than either “AI improves learning” or “AI harms learning.”
The central lesson is practical: access to an AI chatbot is not a learning strategy. Schools need to teach students when to use AI, how to question its answers, and how to demonstrate understanding without it. This article translates the OECD findings into six steps teachers and school leaders can test without treating correlations as proof of cause and effect.
What PISA 2025 actually measured
The OECD asked students how frequently they used AI chatbots for four schoolwork purposes: summarising a text they had to read, conducting preliminary research on a new topic, drafting written assignments, and helping them learn more generally. The report compares these self-reported patterns with student outcomes, including science performance, while accounting for socioeconomic background in several analyses.
That design is useful for identifying patterns across education systems, but it is not a randomised experiment. Students choose whether and how to use AI; the report therefore cannot establish that AI use itself caused a higher or lower score. Prior achievement, access to devices, motivation, teacher guidance and the reason a student reaches for AI may all influence the relationship.
Five findings that matter for schools
1. AI use is widespread, but intensity varies
In most participating countries and economies, a majority of students reported using chatbots for schoolwork. Across OECD countries, only 14% said they never or almost never used AI for any of the schoolwork purposes examined. Yet almost-daily use was reported by only about one in five students on average. Schools should therefore plan for a range of experience rather than assume that every learner is either a confident daily user or a complete beginner.
2. The purpose of use matters more than a simple user/non-user label
The OECD found that students who did not use AI for particular schoolwork tasks generally scored higher than students who did. But the pattern varied by purpose and frequency. For the general purpose “to help me learn,” weekly users had science performance similar to non-users on average after socioeconomic adjustment. Students who used AI moderately for some specific tasks—such as summarising a reading or conducting preliminary research—tended to outperform both limited and very frequent users.
This is not evidence that every student should use AI weekly, nor that moderate use automatically improves results. It shows why “AI use” is too broad a category for policy. Asking a chatbot to complete an assignment, to suggest search terms, to quiz a learner, and to critique a draft are different activities that should not be evaluated as though they were the same intervention.
3. Frequent use is not proof of strong AI literacy
Regular access can make a student familiar with a tool’s interface without teaching them to detect unsupported claims, biased framing or fabricated references. In the OECD data, students’ opportunities to practise evaluating AI-generated information at school were associated with more favourable patterns among frequent learning-oriented users. The association does not establish causality, but it supports a sensible instructional priority: teach evaluation explicitly rather than equating familiarity with competence.
4. Critical evaluation is already part of some classrooms—but not all
Across OECD countries, around six in ten students said that school lessons had asked them to assess the quality of AI-generated information. The share varied substantially across countries and economies. The report also found that students from socioeconomically disadvantaged backgrounds were less likely to report these learning opportunities.
This creates a risk of a new divide. Some students may receive regular coaching in checking AI outputs, while others are left to learn through trial and error outside school. If AI literacy is important, access to critical instruction should not depend on a student’s family resources, device access or prior confidence.
5. Teacher guidance and task design remain essential
The OECD’s wider Digital Education Outlook 2026 makes a related distinction: generative AI can improve the immediate quality of a task without producing equivalent learning gains when it removes the thinking the learner needs to practise. PISA 2025 does not prove a single ideal level of AI use, but it reinforces the need to connect tool use to the learning objective and to give students opportunities to question the information they receive.
Six practical changes teachers can make
1. State the learning goal before allowing a prompt
Before students open an AI tool, name the skill being practised: interpreting evidence, solving a problem, explaining a concept, comparing arguments or planning an investigation. Then decide what assistance is appropriate. If the goal is to practise writing an argument, asking AI to write the argument defeats the task; asking it to identify a counterargument may support it.
2. Require an independent first attempt
Ask learners to write a short initial answer, solve the first step, or record what they already know before consulting AI. The first attempt does not need to be correct. Its purpose is to make the student’s current thinking visible and provide a basis for comparison. For students who need accommodations, allow an appropriate oral response, diagram, speech-to-text or other accessible format.
3. Teach students to interrogate—not merely accept—the output
Give students a repeatable critique routine. They should identify the main claim, ask what evidence supports it, look for an omitted perspective, and mark anything they cannot verify. In science, ask them to distinguish a source’s finding from the AI’s explanation of it. In history, compare the answer with a primary source. In mathematics, test the proposed solution against the original problem.
4. Make source checking visible
Require students to open and read at least one credible source behind an important factual claim. They should record the source, confirm that it actually supports the claim, and note any difference between the source and the AI’s summary. A list of plausible-looking links is not enough: the learner must check that the source exists and says what the response claims.
5. Include a short no-AI follow-up
After AI-supported practice, use a parallel question that students answer independently. A brief exit ticket can reveal whether they can retrieve and apply the idea without the tool. When feasible, repeat a small retrieval task a few days later. These checks should align with the lesson objective and should not be treated as a complete measure of learning on their own.
6. Teach the skill equitably and protect student data
Do not assume that students already know how to use AI responsibly. Demonstrate a weak response, model how to verify it, and let students practise with feedback. Provide a non-AI route when access is limited, and explain what information must never be entered into public tools. Teachers should avoid submitting identifiable student work or confidential school information unless the tool and the institution’s data arrangements have been appropriately approved.
A simple four-week school pilot
Schools do not need to begin with a large platform purchase. A small, well-defined pilot can test whether explicit AI-use guidance improves students’ reasoning and independent performance.
- Week 1 — Baseline: choose one unit and define the target skill. Collect a short unaided task and record common errors.
- Week 2 — Teach the routine: model an initial attempt, AI critique, source verification and independent revision. Tell students when AI is allowed and when it is not.
- Week 3 — Practise: use the routine in two or three lessons. Record the kinds of assistance students request and the errors they catch or miss.
- Week 4 — Evaluate: compare unaided follow-up work with the baseline, review student explanations and gather teacher observations. If possible, compare with a similar task taught without the AI routine.
Measure learning outcomes, not just logins, time on platform or student satisfaction. If students produce polished work with AI but cannot explain or reproduce the key idea independently, the design needs revision.
What the evidence does not yet tell us
PISA is valuable because it covers many education systems, but its AI-use measures are based on students’ reports and its performance analyses are observational. They cannot tell us exactly what would happen if a school required a particular AI routine, nor can they isolate the effect of each chatbot, prompt or classroom policy. Results for 15-year-olds should not be transferred automatically to younger pupils, adult learners or vocational training.
The safest conclusion is conditional rather than alarmist: AI is neither a guaranteed learning aid nor a uniform threat. Its educational value depends on the task, the student’s behaviour, the guidance provided and whether the learner still has to think, verify and apply knowledge. Schools should use PISA 2025 as a reason to teach deliberate AI use—not as a reason to mandate it for every task or ban it without examining the learning goal.
Related reading on Prowad
- AI and Student Memory: 5 Classroom Practices to Protect Learning — classroom routines that preserve retrieval and independent effort.
- AI and Independent Learning: What a 1,222-Person Study Found — why assisted performance and unaided performance can differ.
- How to Evaluate AI Learning Tools: 7 Evidence Checks for Schools — a framework for judging learning outcomes rather than product demonstrations.
- AI Literacy: What It Actually Means — the broader skills behind responsible use.
Sources and further reading
- OECD, PISA 2025 Results (Volume I): Student school life and beyond (8 September 2026), including students’ reported AI use for schoolwork and its relationship with science performance.
- OECD, Digital Education Outlook 2026 (19 January 2026), on the difference between task performance and learning gains.
- OECD and European Commission, Empowering Learners for the Age of AI (18 June 2026), an AI literacy framework for primary and secondary education.