What the latest evidence tells us about AI use, cognitive effort and the changing role of teachers

What the latest evidence tells us about AI use, cognitive effort and the changing role of teachers

By Mehmet ALTAYFounder & International Projects DirectorTeachers Education Academy

Artificial intelligence has entered education faster than most education systems could develop rules, pedagogies or professional standards for its use.

By 2025, AI was no longer an emerging classroom technology. It had become part of students' everyday learning behaviour.

Across OECD countries, 46% of 15-year-old students report using AI chatbots at least weekly to help them learn. Yet PISA 2025 delivers a message that is considerably more important than this adoption figure:

More AI use does not automatically mean more learning. 

This distinction should become central to the next phase of the AI-in-education debate.

For the past few years, much of that debate has focused on access: whether schools should allow AI, which tools students should use, whether teachers should use generative AI, and how quickly education systems should adapt.

PISA 2025 suggests that we now need to move to a more mature question:

Under what conditions does AI actually support learning?

The answer emerging from the evidence is neither anti-AI nor uncritically pro-AI.

It is about purpose, pedagogy and cognitive effort.

AI is already part of students' learning environment

The first reality education systems must acknowledge is that AI use among students is already widespread.

PISA 2025 asked students about several forms of AI chatbot use for schoolwork, including using AI to:

• summarise assigned texts;

• conduct preliminary research;

• draft written assignments; and

• help them learn more generally.

Across OECD countries, only 14% of students reported never or almost never using AI for any of the schoolwork purposes examined. Almost-daily use is less common—around one in five students—but AI can no longer reasonably be regarded as a marginal educational technology.

This fundamentally changes the policy question.

The choice facing schools is increasingly not whether students will encounter AI.

They already have.

The challenge is determining what kind of relationship students should develop with it.

The uncomfortable PISA finding: AI users do not automatically perform better

One of the most striking findings in PISA 2025 deserves careful interpretation.

Students who report using AI chatbots for specific schoolwork tasks generally score lower in science than students who report not using them.

This pattern appears in activities such as summarising texts, conducting preliminary research and drafting written assignments. OECD's broader summary reports a difference of around 20 PISA points on average between users and non-users in these task-specific situations.

But this finding should not be translated into the simplistic conclusion that:

“AI makes students perform worse.”

PISA itself warns against such an interpretation.

The data show associations, not straightforward causality. Students may choose AI for different reasons. Their prior achievement, motivation, socioeconomic background, learning habits and the way they use AI can all influence the relationship.

What the results do challenge is another assumption:

Simply introducing AI into learning does not create educational improvement.

Access is not pedagogy.

Usage is not learning.

And technological sophistication is not the same thing as cognitive development.

Moderate and purposeful use tells a different story

The results become even more interesting when frequency and purpose are considered.

Students who use AI weekly for the general purpose of helping them learn perform at roughly similar levels to students who do not use it, after socioeconomic background is taken into account.

Students using AI either very frequently or only occasionally tend to score lower.

For tasks such as summarising and preliminary research, moderate users also tend to outperform both limited and very frequent users.

This produces a much more nuanced picture.

The relevant educational question may not be:

AI or no AI?

It may instead be:

What AI, for what purpose, how often, under what conditions, and with what learning design?

That is a far more difficult question.

But it is also a much more useful one.

The real issue may be cognitive effort

Perhaps the most important idea in PISA 2025 is not actually about technology.

It is about thinking.

Learning requires students to engage cognitively with new material: interpreting information, making connections, confronting uncertainty, testing ideas, making mistakes and reconstructing understanding.

Generative AI can support this process.

But it can also bypass parts of it.

Consider two students.

One asks an AI system:

“Summarise this article for me.”

The other reads the article, produces a summary, asks AI to critique that summary, compares the two interpretations, identifies disagreements and then revises the original response.

Both students have “used AI”.

But they have not participated in the same learning process.

In the first case, AI may have performed part of the cognitive work that the learner was supposed to practise.

In the second, AI may have increased the amount of reasoning required from the learner.

That distinction may prove far more important than the choice of AI platform itself.

The OECD frames the principle powerfully: technology should support the productive cognitive struggle involved in learning rather than short-circuit it. AI should function as a scaffold rather than a crutch and as a tool for thinking rather than a substitute for thinking.

This may be one of the most important principles for AI pedagogy.

AI literacy changes the picture

PISA 2025 provides another particularly important finding.

Across OECD countries, approximately six in ten students report having been asked during school lessons to assess the quality of information generated by AI.

Now consider what happens when AI use is combined with this kind of learning.

Students who frequently use AI to help them learn and who have opportunities at school to evaluate AI-generated information tend to perform slightly better than students who use AI without comparable opportunities.

Again, PISA explicitly warns that this does not demonstrate causality.

But the pattern matters.

It suggests that the educational value of AI may depend partly on whether students are being taught to engage with it critically.

That means asking questions such as:

Is this AI-generated answer accurate?

What evidence supports it?

What might be missing?

Could the answer contain bias?

Can another source verify the claim?

Why did the AI produce this answer?

Do I agree with its reasoning?

Can I produce a better answer myself?

At that point, AI stops functioning merely as an answer generator.

It becomes an object of critical inquiry.

And that is a very different educational use of artificial intelligence.

A new inequality may be emerging: the AI literacy divide

There is, however, a concerning dimension to this finding.

Opportunities to learn how to evaluate AI-generated information are not distributed equally.

PISA reports that socioeconomically disadvantaged students are less likely to say that they have been asked during school lessons to assess AI-generated information.

This creates the possibility of a new educational divide.

The first digital divide was largely about access:

Who has a computer?

Who has internet access?

Who has digital resources?

The next divide may be more subtle.

It may concern quality of use.

Two students might both have access to the same AI system.

One learns how to prompt it, challenge it, verify it, detect weaknesses and use it as part of an independent reasoning process.

The other learns simply to ask it for answers.

Technically, both students have AI access.

Educationally, they may be developing completely different capabilities.

This is why future discussions of digital equity cannot stop at device and connectivity statistics.

AI equity must also mean equitable access to AI literacy.

The reading crisis becomes even more important in the AI age

PISA 2025 contains another finding that initially appears separate from AI but is actually deeply connected to it.

Reading performance has deteriorated sharply.

Across OECD countries, average reading performance declined by 28 points between 2015 and 2025.

Even more concerning is what kind of reading skills are weakening.

The OECD highlights declines in abilities particularly important in an AI-rich information environment:

evaluating information, connecting multiple sources and thinking critically about what is read.

The proportion of what the OECD describes as “hasty readers”—students moving rapidly through texts but answering inaccurately—almost doubled between 2018 and 2025.

This creates an important paradox.

At exactly the moment when artificial intelligence can generate unlimited quantities of plausible text, young people may be becoming weaker in some of the skills required to judge that text.

That should concern education systems far more than whether students know how to write sophisticated prompts.

The fundamental AI skill may ultimately be reading well enough to know when AI is wrong.

Digital readiness does not begin with technology

This leads to perhaps the most counter-intuitive lesson from PISA 2025.

Preparing students for an AI-driven future does not necessarily begin by giving them more AI.

It begins by strengthening capabilities that existed long before generative AI:

concentration, reading, reasoning, interpretation, synthesis, curiosity, judgement and metacognition.

The OECD makes essentially this argument when discussing digital readiness: it begins not with devices but with young people's capacity to concentrate, interpret, synthesise and reason.

This is important because discussions about “future-ready education” can easily become technology-centred.

But PISA 2025 points towards something different.

The future-ready student may need stronger human cognitive foundations precisely because technology is becoming more powerful.

Computational thinking and foundational knowledge must develop together

PISA 2025's new Learning in the Digital World assessment adds another piece to the picture.

Almost two-thirds of students across OECD countries reach the relevant proficiency level in computational problem solving.

But only around half simultaneously demonstrate that level and baseline proficiency in science, reading and mathematics.

This is an important finding.

Digital problem solving cannot replace foundational knowledge.

Nor should traditional subject knowledge exist independently from students' capacity to apply it in complex digital environments.

The future therefore cannot be:

traditional knowledge OR digital competence.

It must increasingly be:

strong foundational knowledge + digital problem solving + critical AI literacy.

That combination may become one of the defining educational challenges of the next decade.

AI may actually make teachers more important

Generative AI has produced repeated predictions that technology will diminish the role of teachers.

PISA 2025 points in another direction.

If students need to learn when to use AI, when not to use it, how to evaluate its outputs, how to preserve cognitive effort and how to distinguish information from understanding, then human pedagogical judgement becomes more—not less—important.

AI can generate explanations.

A teacher decides whether the explanation supports the intended learning.

AI can provide feedback.

A teacher understands the learner.

AI can produce an assignment.

A teacher understands why the assignment exists.

AI can generate an answer in seconds.

A teacher must help students understand why arriving at an answer themselves may sometimes matter more than receiving it.

The role of the teacher therefore shifts.

Not from teacher to technology.

But increasingly from:

provider of information → designer of learning, guide to inquiry and protector of cognitive effort.

That is a considerably more demanding professional role.

Schools may be asking the wrong AI question

Many schools are currently trying to answer:

“Should students be allowed to use AI?”

PISA 2025 suggests that this question may already be becoming outdated.

A more useful set of questions would be:

What learning objective are we trying to achieve?

Which parts of the task require productive cognitive effort from the student?

Where could AI deepen that effort?

Where could AI remove it?

How will students verify AI-generated information?

What evidence of their own thinking should remain visible?

When should AI use be disclosed?

And when is not using AI pedagogically preferable?

These questions move AI policy away from simple permission and prohibition.

They move it towards learning design.

From AI adoption to AI maturity

Education appears to be entering a second phase of generative AI.

The first phase was dominated by discovery.

Teachers experimented with prompts.

Students experimented with ChatGPT.

Schools debated bans.

Technology companies launched education products.

Policy struggled to catch up.

The second phase must be different.

It must ask whether AI use actually improves learning.

That means moving:

from access to judgement;

from frequency of use to quality of use;

from prompting to verification;

from generating answers to developing understanding;

from AI adoption to AI maturity.

PISA 2025 does not provide all the answers.

Nor can its correlations tell us precisely what AI will do to education over the next decade.

But it gives us something extremely valuable at this stage of the AI transition:

a warning against confusing technological activity with educational progress.

The future-ready student will still need to think

Artificial intelligence will become more capable.

It will generate better explanations, stronger arguments, more sophisticated images, better translations and increasingly convincing answers.

Education cannot respond by attempting to freeze technology outside the classroom.

But neither should it assume that every task made easier by AI represents educational progress.

Sometimes difficulty is part of learning.

Sometimes searching matters.

Sometimes writing matters.

Sometimes remembering matters.

Sometimes struggling with a difficult text matters.

Sometimes thinking without assistance matters.

The challenge for education is therefore not to protect students from AI.

It is to protect the cognitive processes through which students learn while living with AI.

That distinction changes almost everything.

The most future-ready student will not necessarily be the student who can obtain the fastest or most impressive answer from artificial intelligence.

It will be the student who knows:

when to use AI,how to question it,how to verify it,how to improve upon it —and when to think without it.

That may ultimately be the most important form of AI literacy of all.

Mehmet ALTAYFounder & International Projects DirectorTeachers Education Academy

This independent analysis was prepared by Teachers Education Academy based primarily on OECD (2026), PISA 2025 Results (Volume I): Future-Ready Students. PISA findings discussed here describe associations and should not be interpreted as establishing causal effects of AI use on student achievement. Interpretations and conclusions are those of Teachers Education Academy and should not be understood as official OECD positions.