PISA Measured Teen AI Use. Daily Drafters Trail by 28 Points.

TL;DR

The OECD's PISA 2025 results, released September 8, are the first round of the world's biggest school test to ask students how they use AI chatbots. More than 760,000 15-year-olds in 91 countries and economies took part. Across OECD countries, only 14% of students said they never or almost never use AI for any of the schoolwork tasks surveyed. After adjusting for socio-economic status, students who never use AI to draft writing assignments averaged 509 in science, against 481 for students who do it every day. Students who use AI weekly "to help me learn" scored level with non-users. The OECD says the data do not prove AI causes the gaps. Its own reading is that how students use AI matters more than whether they use it.

The first global census of homework chatbots

PISA tests 15-year-olds in science, reading, and math, and governments treat the rankings as a national report card. The 2025 cycle added questions on how often students use AI chatbots "such as ChatGPT" for four schoolwork purposes: summarising a text they had to read, doing preliminary research on a new topic, drafting text for writing assignments, and a catch-all, "to help me learn."

Use is broad but mostly not daily. Across OECD countries, 45.5% of students use AI at least weekly to help them learn, the most common purpose. Weekly-or-more use for the three specific tasks sits between about 29% and 31%. Daily or near-daily use runs from about 11% to 19%, depending on the purpose.

The release lands days after New York City and Los Angeles Unified paused generative AI for close to a million students. So the policy fight is already underway, and this is the first dataset big enough to inform it. Country differences are large. In Japan, 39% of students say they never or almost never use AI for any of the four purposes. In Viet Nam and Singapore, it is about 4%.

The ghostwriter gap

The OECD compared average science scores by how often students use AI for each purpose, adjusting for socio-economic status. That adjustment matters. Students who do not use AI tend to come from more disadvantaged backgrounds and frequent users from more advantaged ones, so the raw numbers flatter AI users. Adjusted, the gaps get wider, not narrower.

• Drafting writing assignments: never 509, daily 481, a 28-point gap.
• Summarising assigned reading: never 508, daily 477, a gap of about 31 points.
• Preliminary research: never 505, daily 486, a 19-point gap.

The OECD's rule of thumb is that 15-year-olds gain about 20 score points per school year, and it warns against applying that mechanically. Taken loosely, the drafting gap is well over a year of learning. Bloomberg's report on the release put it at roughly a year and a half of teaching.

The mechanism the OECD proposes is effort. Andreas Schleicher, its education director, writes in the report that "we do not become fit by watching sports but by doing sports," and that learning is "a productive cognitive struggle of the mind with new material." In gym terms, a spotter and a forklift both get the bar up. Only one of them makes you stronger. For the kids who hand the essay to a chatbot every day, the essay was never the deliverable. It was the workout.

The tutor exception

The fourth purpose behaves differently. For "to help me learn," students who use AI about once or twice a week scored 501, level with non-users at 499. Daily users scored 494 and monthly users 495. The OECD describes moderate use scoring slightly higher than rare or intensive use, and it mirrors the report's finding on screens in general: limited or moderate device use for learning at school goes with better scores, while device use for leisure during school goes with much worse ones.

One oddity runs through all four purposes. Students who use AI only once or twice a year scored at or near the bottom every time, between 482 and 489 adjusted. The report does not single it out, and a one-time survey cannot say why.

Teaching students to grade the machine

The most useful finding for anyone building AI into classrooms sits in a box on page 240. Across OECD countries, 62.6% of students say their lessons at least sometimes ask them to judge the quality of information an AI generated. About eight in ten get the same exercise for information found online.

Those lessons line up with better scores among heavy users. Among students who use AI "to help me learn" every day, those asked to check AI output in class averaged 501.5 in science, against 487.5 for those who never are. The top group in the chart is weekly users who also check AI output in class, at 506.4. Bloomberg's analysis of the same figure called the daily-user gap more than half a year of teaching.

The catch is who gets those lessons. Socio-economically disadvantaged students are less likely to be asked to assess AI-generated information, and the OECD warns about "a new form of socio-economic divide in the age of AI." In the United States, 56.3% of students report such lessons and 37.3% use AI weekly to help them learn, both below the OECD averages. The US numbers carry a large caveat, covered below.

The backdrop is PISA's worst scores yet

The AI questions landed in a grim cycle. PISA 2025 recorded the lowest OECD-average performance so far in science, reading, and math. Reading fell 28 points across OECD countries between 2015 and 2025, and math fell 22. One in five students is now a low performer in all three subjects, up from 16% in 2022.

The reading data look like an attention dashboard. On the fluency items at the start of the test, "hasty readers," students who answer fast and wrong, rose from almost 7% in 2018 to 11% in 2025, while accurate and fluent readers fell by about seven percentage points. Scores dropped more in later sections of the test than early ones. In fairness to 15-year-olds, fast and wrong is also how most adults read a cookie banner. And 28% of students say classmates are distracted by digital devices in most or every science lesson.

Do not pin the slide on chatbots. The OECD notes the reading decline was visible well before the pandemic, years before ChatGPT existed. What the report does say is that the skills falling fastest are "the very skills that matter most in the AI age: evaluating information, making connections across multiple sources, and thinking critically about what we read."

What the data cannot tell you

• Correlation, by the OECD's own account. The report says these relationships "do not necessarily imply a negative impact of AI use on science performance, but may reflect a complex mix of who adopts AI and how they use it."
• Self-reported, one snapshot. Frequency of use comes from a student questionnaire and scores from a single test sitting. There is no before-and-after for individual students.
• Adjusted for one thing. The adjusted scores account for socio-economic status, not prior achievement or motivation.
• 2025 tools, 2025 schools. Students were tested in 2025, on the chatbots and school policies of that year.
• The US asterisk. The US school response rate was 45% before replacement schools. More than 40% of responding US students were not given the student questionnaire, and nine states were excluded from questionnaire coverage, so the OECD flags US questionnaire results for possible distortion.

Source files: the full PISA 2025 Volume I report (PDF) and the chapter 4 data workbook, which holds every score in this post's charts.

If you build AI for learners

• Stop at the scaffold. The worst-scoring pattern is daily use for the tasks that replace the work: drafting and summarising. The OECD's recommendation is AI "used in a targeted way for feedback and personalised practice," which describes hints and critique on a student's own draft, not a finished draft.
• Ship the critique step. Asking students to judge AI output is the one practice in the report that lines up with higher scores for heavy users. Build it into the product instead of hoping a teacher adds it.
• Measure where the tool cannot help. PISA's gaps show up on a supervised test. Completion rates and assignment grades inside your app will not show them.
• Watch who gets the good version. AI-literacy lessons are more common for advantaged students. A rollout that ships the chatbot without the critique layer lands hardest on students who already get less support.

Key Takeaways

• PISA 2025, released September 8, is the first round to measure student AI chatbot use, across more than 760,000 students in 91 countries and economies. Only 14% of OECD students never use AI for the schoolwork tasks surveyed.
• Adjusted for socio-economic status, students who never use AI to draft writing scored 509 in science versus 481 for daily users. Summarising (508 vs 477) and research (505 vs 486) show the same direction.
• Weekly use of AI "to help me learn" matched non-users (501 vs 499), so the purpose of use tracks with scores more than use itself.
• Daily users who are asked to check AI output in class averaged 501.5, versus 487.5 for those who are not, and disadvantaged students get those lessons less often.
• The OECD says the data do not show causation, and US questionnaire results carry a large coverage caveat.
• For builders: design for scaffolding and critique, not ghostwriting, and measure learning where the tool cannot do the work.

Sources: OECD: PISA 2025 Results (Volume I): Future-Ready Students, PISA 2025 Volume I full report (PDF), OECD StatLink: chapter 4 data (Figures I.4.11 to I.4.14), Bloomberg: School Students Who Use AI Get Worse Test Scores, OECD Warns, University of Oxford Department of Education: PISA 2025 results