PISA 2025 exposes the AI trap in the classroom: using it for homework costs a year of learning

By Zendoric · September 9, 2026. The Organisation for Economic Co-operation and Development (OECD) published on September 8 the results of PISA 2025, the test that every three years measures the science, reading and mathematics skills of 15-year-old students. More than 760,000 students from 91 countries and economies took part — in Ecuador, 9,128 students from 266 schools. Beyond the score rankings, the OECD also analyzed how they use artificial intelligence (AI) in the classroom, and the figure is revealing: 46% use AI chatbots weekly or more often to learn.

But that use is not uniform, and therein lies the finding that interests us most. According to the OECD itself, students who turn to AI for very specific schoolwork — summarizing a text, drafting an answer, researching a topic — score 20 points lower on the science test than those who do not use it that way: the equivalent of being roughly a year behind their peers. By contrast, those who use it for general purposes score practically the same as those who do not use AI at all. And there is a third group, the one that fares best: students who, in addition to using AI in a general way, have had learning opportunities to develop AI literacy skills score slightly higher.

Here is where the crack that really matters appears: the OECD finds that those AI literacy opportunities are "more common among students with greater socioeconomic advantages." Put another way, the same tool does not produce the same result in every student: where its use is taught with judgment it adds, and where it is used as a shortcut to avoid thinking it subtracts. It is not the technology that decides, it is how — and to whom — its use is taught.

The report also measures teenagers' relationship with information in general, and the picture is unsettling: only 46% say they check the credibility of sources and trust scientific evidence more than common sense. Some 37% check sources but still trust common sense more; 11% trust evidence without checking where it comes from, and 5% neither verify nor prioritize evidence. The OECD itself concludes that a large share of 15-year-olds may be vulnerable to disinformation, precisely when generative chatbots have become a routine part of their school life.

As sector context, PISA 2025 also captures the school climate beyond academics: one in five students (20%) says they suffered bullying at least several times a month in 2025, worse than in 2022, with 51 countries recording at least one form of bullying per student. Cyberbullying is a minority phenomenon (3%) but it is the one linked to the largest performance gaps. The sense of belonging at school averages 76%, with Spain and Japan at 90% and countries such as Denmark, Poland, Italy and Lithuania at 64% or less; girls and students from disadvantaged backgrounds report feeling less integrated. And in almost every country perceived family support has fallen compared with 2022, especially among boys and disadvantaged students, while concern is growing about the lack of materials and of preparation for digital learning, although schools did improve their student-teacher ratio versus 2022.

In Ecuador, where the Ministry of Education, Sport and Culture offered a plan to strengthen skills after the country's decline in the three subjects assessed, the National Institute of Educational Evaluation (Ineval) attributed the poor performance to a "structural problem" in the system, as reported by Primicias. It is a reasonable explanation, but in light of the OECD's global data it is incomplete: the gap in AI use suggests that part of the problem is not only how much class time students receive, but with what judgment they are taught to use the tool that half of them already carry in their pocket.

Our reading is that this PISA finding confirms something we already suspected when analyzing AI's impact on education as a sector: those who learn to orchestrate AI as a critical copilot win, and those who use it to delegate thinking lose. The difference is not in the technology — the same chatbot can subtract the equivalent of a year of learning or add a slight advantage — it is in whether the education system teaches students to question the answer before copying it. It is the same pattern we already see in the labor market: AI does not distribute its benefits automatically, it distributes them according to who has access to the training to use it well.

In the short term the risk is real and this report documents it with figures: a generation that uses AI every week but has rarely learned to question it, and in which that shortfall hits hardest those with the fewest resources. That widens the education gap just when closing it was most urgent. But the same report also points to the remedy: students with explicit AI literacy training not only avoid the harm, they turn it into a slight advantage. If education systems — Ecuador's included — make that literacy universal instead of merely banning or ignoring chatbots, AI in the classroom can stop being a test of social class and become what it promised to be from the start: a tutor that multiplies the learning of any student, not just of the one who already started ahead.