PISA 2025 and the school of the future: what the OECD survey says about learning in the age of AI

The OECD published the first PISA 2025 results on 8 September 2026. The press mostly picked up two things: French scores at a record low, and pupils doing worse when they use AI. The OECD report is more nuanced. What matters is not AI itself, but how it makes the pupil work.

Jump to the Galac6 comparison

1PISA 2025 in brief

PISA assesses 15-year-olds every three years. Here are the key facts, from the OECD country note on France (in French):

• Major domain in 2025: science.
• Minor domains: mathematics and reading.
• A new domain: Learning in the Digital World. It measures problem solving in a computational context and learning autonomy in a digital environment.
• In France: 6,501 pupils from 289 schools sat the test.
• In 2027: the analyses of learning autonomy and the results in English as a foreign language.

Back to contents

2France's results

First finding: France is close to the average, but falling. Here are its 2025 scores, from the OECD international report (Volume I):

• Science: 483 points (OECD average: 482).
• Reading: 456 points (OECD: 461), 18 points lower than in 2022.
• Mathematics: 458 points (OECD: 463), 16 points lower than in 2022.
• Problem solving in a computational context: 497 points (OECD: 500).

2.1Among the lowest results ever

The OECD's country note on France adds three points:

• the results remain close to the OECD average;
• yet they are among the lowest France has ever recorded in PISA;
• since 2015, the share of low-performing pupils has risen by 12 points in mathematics and by 11 points in reading.

In science, the score was 495 points in 2015, as the DEPP briefing note (in French) points out. The French Ministry of Education devoted a feature to it: "a fall in results shared across the OECD, a recovery still to pursue" (in French).

Back to contents

3AI and results: what the OECD really says

Now to AI. Volume I devotes a whole section to the use of AI chatbots for schoolwork. Here are its two main findings, summarised faithfully.

• Having AI do specific tasks. Summarising a text, doing initial research, writing: pupils who hand these tasks over to AI have lower results in science. This is the case in most countries.
• Using it to learn. When AI is used "to help me learn", weekly users have results similar to non-users, on average across the OECD. Moderate use is associated with slightly higher scores than rare or intensive use.

The OECD insists on one point: these data do not prove cause and effect. Several analyses make the same point, such as the one by The Conversation, republished by Futura (in French): the fall in results began before generative AI became widespread.

3.1A scaffold, not a crutch

The OECD's Director for Education sums up what is at stake in his editorial: "AI becomes a scaffold, not a crutch". His reasoning comes down to two ideas:

• we learn through productive cognitive effort when facing new material;
• a technology that short-circuits this effort holds the pupil back.

The report therefore recommends targeted AI, for feedback on work and personalised practice. It should come on top of the pupil's effort, not instead of it.

3.2The limit of general-purpose AI: answers that need checking

A general-purpose AI delivers ready-made answers. So the pupil has to learn to check them. The report shows that France lags behind on this:

• 47% of pupils say they are asked in class, at least sometimes, to evaluate information from an AI chatbot (OECD: 63%);
• pupils who often use AI to learn and learn to evaluate what it produces get slightly higher results;
• these opportunities are more often reported by advantaged pupils. The OECD fears a new divide in the age of AI.

An AI designed for school can tackle the problem at its root: it does not hand out answers to be taken on trust, and it makes the pupil search. That leaves the real question, the one raised by PISA's new domain: is the pupil learning how to learn?

Back to contents

4Learning to learn: what the new domain measures

PISA's new domain looks at a key skill: knowing how to steer your own learning.

Its Learning in the Digital World assessment framework is a working draft published by the OECD. It describes a process that pupils repeat and regulate themselves: building knowledge and solving problems with computational tools and practices.

At the heart of this definition is self-regulation. The framework breaks it down into three parts:

• Monitoring your progress and adapting: spotting your gaps, taking feedback into account, asking for help when you are stuck.
• Evaluating your performance against your goal.
• Staying motivated: persevering, and improving your work after negative feedback.

4.1Where do French pupils stand?

The PISA questionnaires give a first answer, according to the OECD's country note on France. French pupils plan and persevere a little less than average. On the other hand, more of them believe that intelligence can be developed through work.

Back to contents

5Three questions to judge an AI for education

From these results, we can draw a simple checklist. Parents, teachers and schools can use it to judge any AI tool.

Back to contents

6How Galac6 answers these three questions

We did not discover these three questions with PISA: Galac6 has been built around them from day one, long before the survey was published. We drew in particular on the work of Salman Khan, founder of Khan Academy, who has argued for years for an AI tutor that guides pupils without doing the work for them. PISA 2025 now confirms this intuition. Here is how, question by question.

6.1Preserving effort: help that teaches

A struggling pupil who asks a general-purpose AI for help often gets the exercise done for them. The homework is handed in, but nothing is learned.

Galac6 does exactly the opposite: it helps the pupil find the answer themselves. And that is where everything is decided, because an answer you find on your own is an answer you remember.

The pupil sends their question or a photo of their exercise. The AI tutor gives nothing away: it guides them step by step, with questions and increasingly precise hints.

A mistake? The tutor shows where the problem is, without giving the right answer. Still stuck? The help goes up a notch, as the diagram below shows. The pupil is never left alone with the exercise, and the work is never done for them.

The tutor also refuses to write a commentary or an essay for the pupil.

So Galac6 does not hand out answers to be taken on trust. In mathematics, a calculation engine also checks the calculations: the AI cannot approve a mistake.

6.2Making the pupil more autonomous: learning to learn

Getting an exercise right is good. Knowing how to learn on your own is what changes a school career. That is exactly what PISA 2025 measures, and what Galac6 trains every day.

• Asking yourself questions: the tutor answers with questions. The pupil searches, explains, and finds their own answer. They come away proud, and they remember it.
• "Quiz me on my lesson": the pupil gets quizzed on their revision sheet, as on the night before a test. A mistake? They get a second chance and a hint, not the answer.
• "My topics": in dictation, each correction shows the topics to work on again. The pupil and their parents finally know exactly where to focus their effort.
• Tailored dictation: Galac6 then offers a dictation that goes back over the pupil's mistakes, so they stop making them.

Spotting what you have not understood, coming back to it, persevering: pupils no longer just put up with their difficulties, they learn to overcome them.

6.3Putting AI at the service of the teacher, for every pupil

No AI replaces a teacher. With Galac6, the teacher decides, and the AI works for them.

• They set homework and see at a glance their class's success by topic, and which pupils are falling behind.
• They send a tailored dictation to those who need it.
• During an assessment on Galac6, the tutor is switched off until the pupil hands in their work: the mark truly reflects their own work.
• In class, every pupil has the same tool, supervised by the teacher. Not only those who have the right apps at home.
• Data is hosted in France.

Back to contents

7The school of the future: AI in the service of teaching

PISA 2025 does not say that AI will save or ruin school. It shows that the skills that matter most are deeply human:

• planning
• asking yourself questions
• persevering
• asking for help
• checking you have understood

Pupils already use AI. The challenge for innovation in schools is threefold:

• that AI makes them think, instead of thinking for them;
• that it helps them become autonomous: planning, asking themselves questions, persevering;
• that the teacher stays in control, and that this use is open to all, not only to the most advantaged.

That is Galac6's choice: AI in the service of teaching. It guides the pupil without doing the work for them. It gives teachers a finer view of their class, without ever taking their place.

To find out more, see Galac6 for schools or Galac6 for French schools abroad.

Back to contents

8In summary: Galac6 vs general-purpose AI

This table sets out the recommendations, risks and problems raised by the study. For each one, it shows Galac6's answer, and what happens when a pupil uses a general-purpose AI such as ChatGPT on their own.

What PISA 2025 raises	Galac6	General-purpose AI used alone
1 Doing the work for the pupilSummarising, writing or researching with AI goes hand in hand with lower results in science.PISA report, chapter 4, use of AI, p. 236 →	Galac6 Guides with questions and hints, and never writes for the pupil.	General-purpose AI Summarises, writes or solves the exercise as soon as it is asked.
2 Preserving effortAI should be "a scaffold, not a crutch".PISA report, foreword, p. 6 →	Galac6 Help increases only if the pupil is stuck. The final answer remains theirs.	General-purpose AI Gives the full answer straight away.
3 Making pupils think rather than answering for themSetting goals, asking for help when stuck: pupil autonomy.PISA report, chapter 3, pupil autonomy, p. 178 →	Galac6 Answers with questions. The pupil searches and finds their own answer.	General-purpose AI Thinks for the pupil, who no longer has to question themselves.
4 Tracking your progressSpotting your gaps, coming back to them, checking you have understood: self-regulation.PISA report, chapter 3, autonomy, p. 187 →	Galac6 "My topics", "Quiz me on my lesson" and tailored dictations.	General-purpose AI No follow-up by topic, no link with the curriculum.
5 Persevering after a mistakeImproving your work after negative feedback.PISA report, chapter 3, perseverance, p. 181 →	Galac6 A second chance with a hint. Every mistake is explained with its rule.	General-purpose AI Explains the mistake if asked, often by giving the answer.
6 The teacher's roleTeachers' support remains decisive.PISA report, chapter 3, teacher support, p. 187 →	Galac6 The class's success by topic, tutor switched off during assessments.	General-purpose AI The teacher sees nothing and cannot adjust anything.
7 Avoiding a new divideAdvantaged pupils more often learn to use AI well.PISA report, chapter 4, social background, p. 241 →	Galac6 In class, every pupil has the same tool, supervised by the teacher.	General-purpose AI Everything depends on the family and its subscription.

Comparison with the default use of a general-purpose AI (ChatGPT, Gemini, Copilot…), which is not designed for school. Some offer a "study" mode that asks more questions, but the pupil can switch it on or off freely, and the teacher sees none of it. The ticks and crosses are our assessment of the answer given to each finding, not an OECD measurement. The PISA findings are taken from the OECD report (Volume I); the OECD specifies that they describe associations, not cause and effect.

Back to contents

Back to contents