The Work We Cannot Outsource: AI and Conditions for Learning

What PISA reveals about learning

The latest OECD PISA findings have prompted familiar concern about young people and artificial intelligence. The headlines, however, can obscure a more useful question: when does digital assistance help a student learn, and when does it begin to replace the work through which understanding and judgement develop?

Public discussion often narrows into two competing positions. One places the emphasis on risk and restriction. The other sees AI as inevitable and focuses on adoption. Neither gives teachers, parents or policymakers much help with the everyday reality of learning in an AI-shaped environment.

PISA 2025 found that students who reported using AI chatbots for particular schoolwork tasks, including drafting, summarising assigned reading and preliminary research, generally scored lower in science than students who did not report using chatbots for those purposes. The OECD does not say that AI caused the difference. Students already finding schoolwork difficult may be more likely to turn to chatbots, while teaching quality, home support, prior attainment and study habits may also shape the pattern.¹

Even so, the findings point towards a problem worth taking seriously. When a chatbot can assemble a bibliography, summarise a text and write an essay in seconds, a completed assignment no longer tells us much by itself. Some of the mental work through which capability develops may have been handed over to the system.

UNESCO’s guidance on generative AI in education makes a similar point: adoption is most defensible where it protects human agency, inclusion, privacy and a clear educational purpose, rather than treating use of the technology as an end in itself.³

Generative AI arrived after years of young people already growing up with algorithmic feeds, endless notifications and apps designed to retain attention and encourage return. Young people did not design that environment. Adults built and normalised it while public literacy, meaningful safeguards and accountability struggled to keep up with the pace.

Technology can support accessibility, connection and creative work, but children are developing focus, judgement and identity in environments that typically reward speed, rapid reaction, constant availability and external validation. Generative AI raises the stakes because it can step into the parts of a task that a student would otherwise have attempted on their own.

A tired student facing a difficult assignment could reach for a chatbot because they are overwhelmed, short of time or unsure where to begin. Treating that simply as laziness or dishonesty misses a key point: the educational issue lies in the moment a tool that could have supported understanding begins to replace the work through which understanding develops.

When assistance becomes substitution

People have always used tools to extend what they can do. Calculators, spellcheckers, maps and search engines are part of everyday life. The real question is what the tool is taking over, and whether the learner already understands the foundations underneath it.

A calculator works well when a student understands the underlying mathematical concepts. Grammar tools can help a student communicate more clearly when they understand what acceptable grammar looks like. AI feedback can point out flaws in an argument that a person has already developed on their own.

Asking a system to do the reading, construct the argument, pick the evidence and write the final prose is entirely different.

A student can use AI to test and improve their own draft in valuable ways that support learning. A student who lets a system generate the argument, research and prose may still produce something impressive, but the relevant capability may not have developed in the same way. The finished work may no longer provide reliable evidence of the student’s understanding.

If the system disappeared and the student could not explain, defend or apply the work independently, the task may be complete, but the learning has not been demonstrated.

This is where productive struggle matters. Struggling through difficult reading, retrieving facts from memory, making a mistake and trying again are part of how information turns into usable knowledge. Other forms of difficulty contribute nothing to learning, including poorly designed tasks, inaccessible materials and bureaucratic obstacles that serve no educational purpose. Technology can help to clear those barriers.

The aim is to preserve the cognitive effort through which capability develops, while removing obstacles that do not serve the learning.

What schools can protect and teach

Schools already have legitimate authority to protect spaces where sustained attention can happen. Phone-free classrooms, AI-free first attempts and supervised assessments can be practical responses to a crowded digital environment. These are decisions about what a particular learning activity is intended to achieve.

A teacher might keep devices away during a discussion so students can listen and respond to one another. A university might ask for planning notes and in-class writing because a final PDF cannot always show independent understanding.

PISA 2022 found that, even in schools with phone bans, 29% of students reported using smartphones several times a day at school, while a further 21% reported using them every day or almost every day.²

A rule on a page does not guarantee a focused environment. Consistency, explanation, relationships and classroom culture all affect how a boundary lands. Schools also cannot govern every late-night notification, group chat, algorithmic recommendation or chatbot prompt encountered outside the school day.

That is where guided practice has value. Students can examine an AI-generated explanation beside a trusted source, look for a citation that cannot be found, identify a confident claim that goes beyond the evidence or discuss what a tool changed in a first draft.

Useful questions include:

• What did the system actually do here?
• What parts did I do myself?
• What claims need checking?
• Could I explain this logic in my own words?
• What responsibility do I carry for the final result?

This kind of literacy helps students see that automated outputs can reflect assumptions built into data, design, prompts, institutional practice or the way a system has been set up to respond. It also helps them recognise how easily convenience can turn into dependence.

Responsibility beyond the classroom

Parents carry an enormous weight here, but expecting families to act as the sole regulatory layer for industrial-scale technology is unrealistic.

Households differ greatly in time, income, digital confidence, support and capacity. Devices are tied up with safety, social connection and everyday logistics. Asking a parent to audit every platform update, privacy setting and chatbot interaction is an impossible standard for most.

Responsibility is distributed across the people and institutions that shape the environment. Schools shape learning environments. Governments set baseline protections. Technology companies design defaults, data collection and friction points. Regulators can address power imbalances that individuals cannot fix on their own.

Passing the entire burden to children, parents, or teachers fails because none of them built the ecosystem.

The unequal conditions of access

Many students have access to some form of digital technology, although the quality, reliability and support surrounding that access differ greatly.

The deeper divide is access to guidance. Some young people have paid tools, quiet study spaces, digitally confident adults and schools equipped to teach critical navigation. Others rely on an old phone, unstable internet and consumer apps that encourage rushed workarounds.

Exposure to powerful tools can widen existing gaps where guidance, support and safe access are uneven. An inadequately governed technology does not create equal opportunity. Instead, it can accelerate the inequalities that were already in the room.

The work education must retain

Education is being asked to balance the appeal of fast answers with the slower work through which understanding usually develops. Young people benefit when there is time to concentrate, room to make mistakes without instant artificial rescue, and the confidence to recognise when a system is supporting their learning and when it is helping them avoid it.

An answer is only part of an education. Learning to question where it came from, what it leaves out and what responsibility remains with the person using it is real work that no system removes from us.