The OECD released the PISA 2025 results on September 8. More than 760,000 fifteen-year-olds in 91 countries and economies took the assessment, and for the first time the student questionnaire asked how often they use AI chatbots for schoolwork and what they use them for. Early coverage focused on the obvious headline: students who lean on chatbots for schoolwork tend to score lower in science.
A short box on page 240 of Volume I adds a finding that got less attention: which students get taught to check what the chatbot tells them, and how that splits along income lines.
What PISA Asked
Students reported how often they used a chatbot for four purposes: summarizing a text they had to read, doing preliminary research on a new topic, drafting text for a written assignment, and a broader category PISA labeled “to help me learn.”
Use turned out to be wide but not deep. On average across OECD countries, 14% of students said they never or almost never used AI for any of those purposes, which fits the adoption figures I’ve cited on shouldiuse.ai before, the ones that put student use near nine in ten. PISA also measured frequency, and almost-daily use was reported by only about one student in five. At fifteen, most students use chatbots some of the time, and a minority use them constantly.
Then come the scores. For the specific tasks (summarizing, research, and drafting), non-users outperformed users in most countries. For general “help me learn” use the pattern was flatter: weekly users scored about the same as non-users once the OECD accounted for socio-economic background, while near-daily users and occasional users scored lower. For summarizing and research, moderate users, somewhere between once a month and twice a week, did better than both the students who barely touched it and the ones who used it every day.
The OECD is careful about what this means, and it should be. These are associations from a single snapshot, and the report says plainly that they don’t necessarily show AI hurting performance. They may reflect who picks up the tools and how. A student who is already struggling with a reading might be the one most likely to ask for a summary, and PISA can’t separate that student from one whose reading got worse because of the summaries.
The Box on Page 240
Box I.4.2 asks a narrower question: were students ever asked, during their lessons, to assess the quality of information generated by AI?
Around six in ten said yes on average across OECD countries, with the share ranging from 31% to more than 80% depending on the country. Among students who said yes, the most common answer was “sometimes.” For comparison, about eight in ten said they’d been asked to assess the quality of information they found online. Schools have had a couple of decades to build lessons on evaluating websites, and the lessons on evaluating chatbot output are newer and thinner.
Students who got those lessons and also used AI to help them learn at least once a week scored slightly higher in science than non-users and than students who used AI twice a month or less. The OECD calls the difference slight, and I won’t inflate it. It is the clearest case in the report’s AI chapter where regular use lines up with stronger scores.
The box also reports that socio-economically disadvantaged students are less likely to report being asked to assess AI-generated information in class. The OECD says the disparity needs closer examination and targeted intervention so it doesn’t harden into a new socio-economic divide.
Two Divides: Access and Instruction
PISA also found that students who don’t use AI for schoolwork tend to come from more disadvantaged backgrounds, and frequent users tend to come from more advantaged ones. Access to devices, a stable connection, and paid tools all play a part.
So the familiar worry about AI and equity, that poorer kids won’t have the tools, is partly true at fifteen. Heavy use on its own goes with lower scores, though, so access alone doesn’t explain any advantage. The one place PISA finds an edge is among students taught to check the output, and wealthier students are more likely to sit in classrooms where someone asks them to do that.
The second divide is harder to close. Access gaps tend to narrow as tools get cheaper, and in my experience the free tiers of most chatbots already handle the bulk of high school assignments. Instruction gaps don’t narrow on their own. They depend on teacher time, training, curriculum space, and whether anyone in the building thinks the skill is worth a class period, and those resources are unevenly spread right now.
The report makes the same point in its own language: expanding access to AI tools is unlikely to produce benefits without parallel investment in guidance on how to use them well, so a school that answers the AI question by buying licenses has handled only the access half.
Sourcing a Chatbot’s Answer
The skill Box I.4.2 describes looked familiar to me. In AP history we call it sourcing. Students look at a document and ask who produced it, for what audience, and for what purpose, and then decide what that does to how far they can trust it. The AP rubrics award points for it, and it’s one of the skills students take longest to get comfortable with, because it asks them to stop treating a text as a container of facts.
A chatbot answer is a document with an odd provenance. It has no author in the usual sense, it was produced for an audience of one, its purpose is to sound helpful, and it can be confidently wrong in ways a textbook usually isn’t. Every question we ask of a primary source still applies, and a few new ones get added: what would it have said if I’d asked differently, where is it hedging, where does it sound certain without giving a reason, and what has it left out?
My book’s argument is in its title: the AI doesn’t know your students. It produces the same fluent paragraph for the kid who will question it and the kid who will paste it into a document and hit submit. What separates those two kids is whether someone taught them to read the output the way they’d read any other source that claims to know something.
Andreas Schleicher, who leads the OECD’s education work, writes in the report’s preface that if we get this right, AI becomes “a scaffold, not a crutch.” I agree with him. PISA’s data hints that teaching students to check the output is part of what makes it a scaffold, and right now those lessons show up more often in wealthier students’ classrooms.
If You Teach at a Well-Resourced School
I teach at an international school in Shanghai, and my students sit on the advantaged side of nearly every line PISA draws. It would be easy to read this report and decide our kids are fine. I don’t think that follows. Six in ten is an average, “sometimes” was the most common answer, and one lesson on checking AI output in one class doesn’t become a habit unless it shows up in the next class too, so advantaged students may get the lesson without getting it often enough for it to stick.
It cuts the other way as well. If you teach in a school with fewer resources, PISA says your students are among the least likely to be getting these lessons. I’d guess that gap tracks the time and training teachers in those schools have been given, and that’s the place where a district’s AI budget would do the most good.
Takeaway for Teachers
This week, pick one assignment where students would plausibly use a chatbot, and turn the chatbot’s answer into a source they have to analyze. Run your own prompt ahead of time, print the response, and hand it out alongside whatever primary or secondary material the lesson already uses. Ask students to source it the way they’d source anything else: what was it asked to do, what claims does it make, which of those claims can they check against the other documents, where does it sound certain without a reason, and what’s missing? Then have them write two sentences on how far they’d trust it and why.
It takes one class period and no new technology. Do it again in the next unit with a different kind of question, so it becomes a routine instead of a one-time warning about AI. The slight advantage PISA found belonged to students who used AI often and had been taught to check it, which suggests the checking needs regular practice alongside the using, more than a single lesson.
If you’re thinking through questions like this one, about who gets taught to question AI and who gets left to figure it out alone, my book goes deeper. The AI Doesn’t Know Your Students is available on Amazon and at shouldiuse.ai/book.
David Jacobson is a high school history teacher. He writes about AI, education, and the messy intersection of the two at shouldiuse.ai.
