PISA 2025 and AI in the classroom: why EduGPT is particularly relevant right now

On 8 September 2026, the OECD published the results of PISA 2025, the largest PISA study to date. Over 760,000 fifteen-year-olds in 91 countries took part, including around 6,600 pupils at 177 schools in the Netherlands. The figures are sobering. And for the first time, PISA also includes data on how pupils use AI chatbots for their schoolwork.

We read the report with extra attention. Not because the news about reading and maths surprises us, but because the conclusions about AI touch precisely on the question for which we started building EduGPT. In this post, we summarise the key points and explain why a safe, guided AI environment is no longer a luxury.

What PISA 2025 shows

The average reading score in OECD countries fell by 28 points between 2015 and 2025, which the OECD equates to about a year and a half of learning. Maths dropped by 22 points, more than a year’s worth. For both reading and maths, this is the lowest OECD average PISA has ever measured. One in five fifteen-year-olds now performs poorly in all three subjects, compared to 16 percent in 2022.

Science, the main subject of this round, remained stable between 2022 and 2025 after a long decline since 2009. The top ten consists of Chinese regions, Singapore, Estonia, Japan, Korea, Taiwan, and the United Kingdom.

The opportunity gap remains large. Pupils from advantaged families scored on average 85 points higher in science than those from disadvantaged backgrounds in the OECD. That gap narrowed slightly, but mainly because the advantaged group performed worse.

The situation in the Netherlands

Dutch pupils still score above the OECD average in maths and around the average in science. Both remained unchanged compared to 2022. The decline in reading, however, continued. Only 61 percent of Dutch fifteen-year-olds now reach the basic level, compared to 69 percent on average in the OECD. Nearly four in ten pupils therefore risk leaving education insufficiently literate.

The OECD states in its country note on the Netherlands that the results in all three subjects are among the lowest ever measured in our country, and that the trend since 2015 is clearly negative. The gap between advantaged and disadvantaged pupils is also larger here than average: 105 points compared to 85 in the OECD.

AI in the classroom, measured for the first time

New in PISA 2025 is that pupils were asked about their use of AI chatbots for school. Only 13 percent of Dutch pupils never or almost never use them. Over a third use AI at least weekly for learning, a quarter regularly use AI to summarise texts, and over a quarter have AI produce a first version of writing assignments.

The OECD’s first conclusion is confronting. Pupils who did not use AI to produce texts for writing assignments scored better than those who did. Andreas Schleicher, head of PISA, links this to the decline in reading. The drop is greatest in precisely those skills that matter most in an age of AI: distinguishing fact from opinion, connecting sources, and critically assessing information.

Then comes the second conclusion, which for us is the core. Among frequent users, pupils who used AI to learn and who had been taught at school how to assess the quality of AI answers performed better than those who did not receive such guidance. The problem, therefore, is not the technology, but its use without guidance.

OECD Secretary-General Mathias Cormann summarises it as follows. Technology and AI can enhance learning and prepare young people for the future, but only if they are used purposefully and not as a substitute for attention, effort, and understanding.

Why this makes EduGPT particularly relevant

The temptation is great to read PISA as an argument against AI in the classroom. We read it differently. PISA shows that AI without didactics is harmful, and that AI with didactics is helpful. The difference lies not in the tool, but in the design of its use. And that is precisely the principle on which we built EduGPT.

Critical thinking is built into the design. EduGPT began in 2023 with the Feedback Companion, which does not give pupils the answer but asks questions about their own work. This approach is present in all our didactic companions. Teachers can have two models answer side by side, so pupils see that AI output varies and must always be checked. Engaging critically with AI answers is, in practice, simply reading education—the very skill PISA sees declining.

The teacher remains in control. With the Prompt Companion, teachers create their own AI Companions for their subject and their pupils, with guidance built in. The school decides which companions and models are available and how they behave. Teachers are designers of their learning environment, not consumers of a tool. Our founder Pascal Mariany calls this, in his book The Art of Living Together with AI, Reasoned Responsibility—the judgement remains with the human.

European hosting and sovereignty. PISA surveys fifteen-year-olds, in other words, minors. The AI Assessment Framework for Primary and Secondary Education from SIVON and Kennisnet is clear: no free tools with pupil data. EduGPT runs in Europe, the school remains the owner of the data, and personal data can be anonymised. This way, the classroom conversation can be about AI, instead of about where your pupils’ data is going.

Focus instead of distraction. PISA also warns against digital distraction. In the Netherlands, 23 percent of pupils say classmates are distracted by devices in most lessons, and these pupils score 19 points lower. EduGPT is a learning environment for schoolwork, without adverts, without a feed, and with companions tailored to the lesson. A conscious choice, not a by-product.

Our recommendations for schools

The Dutch education organisations put it plainly in their joint statement on PISA. Do not use these figures to claim that education must be completely overhauled, but continue with the broad approach to basic skills. We agree with this. Especially in an age of AI, the temptation is great to quickly purchase a tool or to ban AI altogether. Neither reflex helps pupils.

Our advice is to do three things at once. Start with the didactic intention: what must pupils be able to do themselves, and where may AI support? Choose an environment where this can be done safely and which the school itself controls. And make assessing AI answers a regular part of the lesson, because according to PISA, that is the difference between pupils who fall behind and those who progress.

The AI literacy framework published this week by Kennisnet provides a good shared language for this. Would you like to see what this looks like in practice, with real companions and real lessons? We are happy to show you, in a demo or a small pilot with a few teachers. Get in touch via edugpt.nl or send a message to Pascal Mariany.

Sources: OECD press release of 8 September 2026, PISA 2025 Results Volume I, country note Netherlands, Andreas Schleicher on AI use and reading, National Education Knowledge Institute, joint statement education organisations and AI literacy framework by Kennisnet.

Image for illustration, created with AI