The OECD finds the first indication that using AI at school lowers science scores

By Zendoric · September 27, 2026.

The figure that opens this piece is not an opinion, it is a statistical signal: PISA 2025 — the OECD (Organisation for Economic Co-operation and Development) report that assesses 15-year-old students in reading, mathematics and science every three years — found an association between using AI for specific schoolwork and lower scores in science. The caution lies in how the data is read: this is not proven causation, it is correlation. But it is, as specialist Héctor Pérez-Urbina tells Vanguardia, "exactly the kind of warning sign we have to pay attention to."

The scale of the phenomenon around that figure is already the majority case. According to the Common Sense Media 2026 global report, published in June, 86% of children aged 9 to 17 use AI; 70% of teenagers use it for school, and of those, two out of three delegate critical thinking to the tool — the ability to analyze, question and reach one's own conclusions, not just to produce an answer. It is the first time an entire generation has grown up with a system capable of doing, fluently and with no apparent effort, the part of the cognitive work that until now required time and friction.

The person putting these figures in perspective is no technophobe. Héctor Pérez-Urbina holds a doctorate in AI from the University of Oxford and a master's in AI Ethics from the University of Cambridge, and spent a decade at Google before founding Tlanextli Group, an organization focused on what he calls "cognitive sovereignty." He describes himself as a "critical techno-optimist": in favor of technology, but not of using it automatically. His diagnosis rests on a concrete mechanism: when we delegate systematically, we stop exercising skills — analysis, planning, synthesis, reasoning — that begin to atrophy from disuse, just like a muscle. And he describes a less obvious side effect: the "illusion of competence," the sense of having learned something merely from having understood the fluent answer the AI delivered, without having been able to produce it oneself.

The risk that should most concern the industry, however, is not individual but collective: homogenization. If millions of people delegate the same kind of reasoning to the same models, the result — Pérez-Urbina warns — is not only that texts, photos and videos start to sound and look alike (something he says he already sees on LinkedIn), but that thought itself converges. A society that thinks more uniformly is, by definition, a society with fewer diverse cognitive resources for tackling new problems. It is an uncomfortable paradox: the very tool that promises to personalize everything may end up standardizing the intellect of those who use it without judgment.

This connects with something we have been documenting in our sector-by-sector coverage of AI and employment: in education, the job that survives is not that of the teacher who transmits content, but that of the teacher who orchestrates AI as a teaching tool. The interview adds a layer we had not covered in this detail: the risk is not only about work — which jobs disappear — but cognitive, and it affects the employed, students and the retired alike. The divide Pérez-Urbina describes, between those who use AI "to think more" and those who use it "to think less," is as relevant for the coming decade as today's divide between those who can code and those who cannot.

This is where it pays to be honest about the short term, in line with what we argue at Zendoric: cognitive dependence is a real and already measurable cost, not an alarmist laboratory hypothesis. The OECD detects it in classrooms; Common Sense Media quantifies it in home use. There is no need to deny the problem in order to defend the technology.

But neither should the symptom be mistaken for the final diagnosis. The solution Pérez-Urbina proposes — three questions before, during and after using AI: why am I using it, did I get what I was after, which skill did I stop exercising — is not a brake on adoption, it is a grammar of intentional use. And there, in our view, lies the deeper reading: the same cognitive surplus AI frees up — time, effort, calculation — is the raw material of the abundance we defend as this technology's horizon. But that abundance only translates into human flourishing if we retain the capacity to decide, to criticize and to create with our own judgment. A society populated by what the interviewee himself calls the replaceable "human cog" is prepared to take advantage neither of the cure for a disease nor of an economy of unlimited resources: it needs citizens capable of thinking, not just of delegating. The game being played now, classroom by classroom, is not whether we use AI, but whether we use it to grow or to stop trying.