PISA 2025: the OECD links heavy AI use to worse grades, in an educational decline that is global

By La Silla Rota · September 12, 2026.

The PISA test—the exam the OECD (Organisation for Economic Co-operation and Development) administers every three years to 15-year-old students in mathematics, reading and science—was conducted in 2025 across 91 countries and economies, and its conclusions have just been published. The figure that dominates the report is not a local one: it is a widespread decline in school performance across much of the world. The United States recorded its worst reading score since 2000, 14 points below the previous edition; France speaks of a "historic collapse," with its lowest reading and mathematics scores since that same date; Germany posted its lowest levels since the test has existed; Finland, which led the world reading ranking in 2000, fell to 17th place; the Netherlands slipped again in reading comprehension. Even the systems at the top of the table—China, South Korea, Japan and Singapore—fell back compared with previous editions, although they remain ahead of the rest.

Mexico strings together two consecutive declines in mathematics: 14 points between 2018 and 2022, and 7 points more between 2022 and 2025, the latter considered statistically significant by the OECD (unlike the variations in reading and science, which fall within the margin of error). Even so, Andreas Schleicher, the OECD's Director for Education and Skills, stressed that Mexico has sustained its performance in science since 2012, held its position in reading while other countries fell back sharply, and that its students report more curiosity and interest in learning than the organization's average. He also underlined a figure rarely mentioned in alarmist coverage: despite Mexico expanding educational coverage to marginalized groups, the score gap between rich and poor students did not widen, and is smaller than the OECD average. It is worth bearing that in mind before reading the Mexican result as an isolated failure: the relevant comparison is against a global downward trend, not against a static ideal.

The finding that connects this edition to artificial intelligence is more nuanced than a headline suggests. According to the OECD itself, daily use of AI to replace cognitive tasks—drafting, summarizing, researching—is associated with worse results, while moderate use, as occasional support for learning, does not harm performance. It is a correlation, not proof of direct causation, and the organization itself avoids a blanket condemnation. But the reasonable reading, and the one the writer makes, is that substitutive use—letting the tool think instead of the student—weighs more in practice than prudent use. It is worth putting precisely, because the difference between "AI lowers grades" and "using AI to avoid thinking lowers grades" is not cosmetic: the second leaves the door open to the same instrument, used another way, not having that effect.

Two governments have already taken note, and they are moving in almost opposite directions. Denmark, which already bans access to social media for under-15s, announced on September 10—in direct response to these results—a total ban on mobile phones in schools, a measure driven by Prime Minister Mette Frederiksen. The United Arab Emirates, by contrast, with the world's best results in this edition, is betting on the opposite extreme: curricular, supervised use of AI in the classroom, as against free, individualized and chaotic use. These are two different governance bets in the face of the same problem, and both deserve monitoring: we will not know which works better until the next PISA round.

Our reading is that attributing the decline solely to generative AI is a convenient oversimplification. Finland began falling from 2000, long before any chatbot existed; the deterioration in fact coincides with two decades of saturation by screens, social media and algorithms that decide what we read and for how long. Conversational AI is the most recent and most powerful link in that chain, not the sole cause. What does seem consistent, and is the part worth taking seriously without qualification, is the mechanism: any tool that answers for us, if used to avoid cognitive effort rather than to direct it, atrophies the very capacity it was supposed to complement. It is the same principle we have already seen in analyzing AI's impact on teaching jobs: the teacher who orchestrates AI as a teaching tool wins; the one who uses it—or lets their students use it—to skip the learning process loses. Here it applies directly to the student: whoever uses AI as a copilot for thinking wins; whoever delegates the thinking entirely loses.

That said, we hold the same underlying thesis we apply to other AI fronts: the problem is real and short-term, and it should not be minimized. A generation that stops exercising sustained reading, memorization and unaided reasoning pays a tangible, measurable cognitive cost. That is not solved with generic optimism. But the same data that worries us contains an encouraging clue: the United Arab Emirates, which is betting on curricular, supervised use of AI, tops the world results, and the gap between rich and poor did not widen in Mexico despite expanded educational access. If AI, well governed, can become a personalized tutor capable of leveling opportunities instead of replacing effort, the problem is not the technology but the design with which we let it into the classroom and into adults' daily lives. That, ultimately, is the question the writer poses and one we share: not what is happening to schools, but what is happening to all of us who use these tools without pause or judgment. The answer is still being written, and the next editions of PISA will be the most reliable thermometer we have to find out.