760,000 Students, 91 Countries, One Question: What Does AI Do to Learning? PISA 2025 Reports Tomorrow

Tomorrow morning at 9:30 CEST, the OECD will publish the most consequential education dataset of the AI era. PISA 2025 — the ninth edition of the Programme for International Student Assessment — tested more than 760,000 15-year-olds across 91 countries and economies, and for the first time in the survey’s 25-year history, it comes with something no previous cycle had: hard, cross-national numbers on how students use AI tools for schoolwork.

For an industry that spends its days arguing about whether AI helps or harms learning, this is the closest thing to a verdict that has ever existed. And the timing could not be sharper.

Why this cycle matters more than any before it

PISA has always been the world’s educational scoreboard. Every three years (now moving to a four-year cycle after 2025), it tests 15-year-olds in reading, mathematics, and science, and governments treat the results as a national report card — one that can shift education policy, trigger ministerial resignations, and dominate front pages from Seoul to Stockholm.

But the last cycle, PISA 2022, landed like a warning shot. Scores fell across the OECD to their lowest levels ever recorded: mathematics dropped by a record 15 points, reading fell 10 points — twice the previous record decline — and science was the only domain to hold roughly steady. The OECD attributed the damage to the pandemic years, but with recovery spending wound down and a new technology flooding classrooms, the obvious question for 2025 was whether the curve would bend back up.

PISA 2025 collected its data in late 2025 — the first full assessment cycle conducted entirely after ChatGPT-class tools became ambient in teenagers’ lives. Whatever it shows about science performance (this cycle’s major domain), the more interesting variable is what happened outside the test: how often students turned to AI, for what tasks, and whether any of it shows up in the scores.

The new evidence: AI use for schoolwork, measured at scale

The OECD has been explicit that the 2025 report will include “new evidence on how students use AI for schoolwork, offering timely insights to inform the global debate on the role of AI in education.” That phrasing, from the launch advisory, is doing a lot of work. It signals that the OECD believes the data is robust enough to weigh in on a debate that has so far run almost entirely on anecdote — teacher surveys, vendor studies, and laboratory experiments with a few hundred subjects.

Seven hundred sixty thousand students across 91 systems is a different order of evidence. If the OECD’s data shows, for example, that heavy AI use correlates with weaker performance in science reasoning — or the opposite — it will be the first time anyone has been able to say so with cross-cultural, population-scale data rather than speculation.

The stakes for the industry are direct. OpenAI and Google have spent two years pushing ChatGPT and Gemini into classrooms, framing them as tutors that democratize learning. Critics, including several prominent cognitive scientists, argue that offloading reasoning to a chatbot undermines the very skills PISA measures. The 2025 cycle is the first dataset large enough to adjudicate even partially between those positions — and it arrives just as school districts from New York City (which imposed an AI moratorium for younger students) to national ministries are deciding whether to embrace or restrict the tools.

Learning in the Digital World: the experimental domain that grew teeth

Alongside the core domains, PISA 2025 included an innovative domain called “Learning in the Digital World.” Rather than asking students how they learn with technology, it put them in front of simulated digital learning environments and measured their capacity for iterative knowledge building and problem solving — including motivation and self-regulation while learning digitally.

This is the quiet radical part of the 2025 cycle. Traditional PISA items measure what a student knows. Learning in the Digital World measures how a student behaves when the answer isn’t given — whether they persist through failed attempts, adjust strategy, and monitor their own confusion. Those are precisely the metacognitive skills that both AI advocates and skeptics claim as their territory: proponents argue AI tutoring can train them; skeptics argue chatbot reliance erodes them. For the first time, there will be comparative data.

The OECD has already committed to going further: PISA 2029 will debut a full Media and AI Literacy (MAIL) assessment, testing whether students can evaluate the credibility, quality, and purpose of digital content — a tacit admission that AI literacy is now considered as fundamental as reading itself.

What to watch when the numbers drop

Several storylines are queued up for tomorrow’s launch.

Does the 2022 collapse continue or reverse? Science is the major domain this cycle. Singapore topped PISA 2022 science at 561 points, with Japan (547), Macau (543), Taiwan (537), and South Korea (528) close behind; the OECD average languished near 485. Whether East Asian systems extended their lead — and whether Western systems stabilized after a decade of decline — will drive most headlines.

Where is AI use highest, and does it track with performance? The gap between a student in Helsinki using AI for research and a student using it to outsource homework is the difference between a tool and a crutch. If the OECD breaks AI usage down by frequency and country, expect the patterns to be uneven in ways that embarrass someone.

What happened to the pandemic cohort? Students tested in 2025 lived through COVID school closures during their foundational years and hit secondary school just as AI arrived. Disentangling those two shocks is the report’s hardest analytical problem — and its most important one, because policymakers can still do something about the second.

The uncomfortable possibility

There is a scenario nobody in the industry wants to name aloud: that PISA 2025 shows scores continuing to fall in systems with the highest reported AI adoption. That would not prove causation — correlation across 91 systems is confounded by everything from funding to culture — but it would hand ammunition to every regulator considering classroom restrictions, and it would complicate the education-market ambitions of every frontier lab.

The opposite scenario is equally plausible: AI-heavy systems show gains, particularly among lower-performing students who finally had access to unlimited, patient tutoring. That result would accelerate adoption dramatically.

Either way, the era of running education-AI debates on vibes ends tomorrow at 9:30 in the morning, Paris time. The largest assessment of teenage learning ever conducted has been listening to how the world’s students actually study — AI included. Now we find out what they heard.