The rise of AI is forcing us to reconsider not only what people need to learn, but how they develop the judgement, agency and expertise that will remain distinctly human.

I recently had the opportunity to join a discussion at All Saints’ College prompted by a deceptively simple question: what is education actually for?

The discussion drew on the OECD’s recent Education for Human Flourishing: A Conceptual Framework. Published in late 2025, the report challenges an idea that has shaped education systems for decades: the human capital model.

In simple terms, the human capital model views education as an investment in people. We develop knowledge and skills that allow us to participate productively in the economy. We learn so that we can do.

It is a model that has served us remarkably well. Education has helped create generations of engineers, scientists, doctors, teachers, tradespeople and professionals whose capabilities have contributed enormously to economic prosperity and social progress.

But the OECD asks whether that model is sufficient for the world young people are now entering.

It argues that education should develop a broader range of capabilities that allow people to flourish throughout their lives. At the centre of its framework are five capabilities: adaptive problem solving, ethical reasoning, understanding the world, appreciating the world, and acting in the world.

And sitting behind much of the argument is artificial intelligence.

That changes the question considerably.

If the purpose of education is largely to equip people with the knowledge and skills to do, what happens when machines can increasingly do as well?

Human capital still matters

There is a temptation when discussing new models of education to declare the old model obsolete.

I don’t think that would be helpful.

Knowledge matters. Technical competence matters. Literacy and numeracy matter. Deep disciplinary expertise matters.

An engineer cannot exercise meaningful engineering judgement without understanding engineering. A doctor cannot critically evaluate a diagnosis without understanding medicine. An accountant cannot challenge a financial model without understanding accounting.

The problem may therefore not be the human capital model itself. It may be that our conception of valuable human capital has become too narrow.

For much of the modern economy, education has created value by teaching people to perform tasks that require substantial knowledge or expertise.

We calculate. Analyse. Research. Draft. Design. Model. Diagnose. Code. Write. Interpret.

These capabilities take years to develop, which has historically made them economically valuable.

AI is beginning to challenge that relationship.

Increasingly capable AI systems can already perform parts of many of these activities. They can write software, analyse datasets, produce technical documentation, synthesise research, generate designs and assist with complex analytical tasks.

That does not mean expertise becomes irrelevant.

It does mean we need to think carefully about what expertise looks like when the machine can participate in the work.

From what can you do to what can you contribute?

This is where I find the OECD’s human flourishing proposition compelling.

The traditional human capital question is largely:

What do you know, and what can you do?

The flourishing framework introduces a broader question:

What are you capable of becoming, deciding and contributing?

That is an important distinction.

As machines become better at producing answers, human value increasingly shifts towards framing the question, understanding context, exercising judgement, navigating ambiguity, challenging assumptions, considering consequences and deciding what should actually be done.

These are not alternatives to technical knowledge. They depend upon it.

But they are also not simply technical skills.

Consider an engineer using AI to generate a series of possible designs.

The ability to produce those designs may become less scarce. The more important questions may be: What problem are we actually trying to solve? What assumptions has the system made? Which constraints matter? What happens when conditions change? What are the safety implications? Who carries the risk? What has the model missed? And ultimately, which solution should we choose?

AI can increasingly participate in answering those questions too.

But responsibility for the outcome remains human.

That makes capabilities such as judgement, agency, ethical reasoning and adaptive problem solving increasingly important.

But there is a problem with flourishing too

The discussion at All Saints’ College raised an important challenge to the flourishing model.

How do we measure it?

There is a fundamental difference between assessing whether education has equipped someone with the capabilities to flourish and assessing whether that individual is flourishing.

The first is reasonably within the remit of education.

The second is much more complicated.

We can assess whether someone can reason through an ethical dilemma. We can observe whether they can approach an unfamiliar problem, examine evidence and adapt their thinking.

But what happens when we start measuring purpose, fulfilment, meaning, relationships or someone’s appreciation of the world?

At that point assessment begins moving from what you know and what you can do towards who you are and how you experience your life.

That creates difficult questions.

Who defines flourishing? Does a highly ambitious student flourish more or less than someone seeking a quiet life centred on family and community? How should cultural, religious and philosophical differences be treated? What happens when a perfectly healthy young person is unhappy, uncertain, introverted or struggling to find purpose?

The OECD itself distinguishes between objective flourishing over a lifetime and the subjective happiness or wellbeing of a student at a particular point in time. That distinction is important.

Flourishing can be a valuable aspiration for education without necessarily becoming another score on a student’s report.

Perhaps the better objective is to develop the capabilities that enable flourishing, while remaining cautious about institutions judging whether an individual is successfully flourishing.

The necessary discomfort of learning

There is another tension that becomes particularly important when we introduce AI.

Learning is not always pleasant.

Anyone who has genuinely learned something difficult knows the feeling.

You encounter a problem you cannot solve. You stare at it. Try one approach. Get it wrong. Go backwards. Question what you thought you understood. Try again. Eventually something connects.

That discomfort is not necessarily a defect in the learning process.

Sometimes it is the learning process.

We learn through doing, but we also learn through the unpleasantness of thinking.

The effort required to retrieve something from memory, work through a difficult calculation, construct an argument or understand why something failed helps build the mental models that eventually allow us to exercise judgement.

Failure matters too.

A solution that does not work forces us to confront the difference between what we thought would happen and what actually happened. Feedback can be uncomfortable. Ambiguity can be frustrating. Not knowing an answer can make us feel incompetent.

Yet those experiences can be enormously developmental.

There is therefore a danger in equating flourishing with comfort.

If education is intended to help people flourish, it cannot simply seek to remove difficulty, frustration or failure from the learning experience.

Some discomfort is necessary.

AI can remove precisely that discomfort

This is where generative AI creates a fascinating educational paradox.

AI is extraordinarily good at removing friction.

A student confronted with a difficult problem can now ask an AI system for an explanation. A blank page can become a first draft in seconds. A difficult equation can be solved. Code can be generated. An argument can be structured. A complex paper can be summarised.

In many contexts that is enormously useful.

But the OECD’s Digital Education Outlook 2026 highlights an important distinction: performing better with AI does not necessarily mean learning more.

Emerging evidence reviewed by the OECD suggests that students using general-purpose generative AI can produce better outputs without achieving corresponding improvements in underlying learning. When AI becomes a shortcut rather than a scaffold, students can offload the cognitive effort involved in reflection, reasoning and self-monitoring.

This creates a distinction that I think will become increasingly important:

There is a difference between producing an answer and developing the capability to produce an answer.

AI makes the first dramatically easier.

Education needs to ensure that this does not come at the expense of the second.

The OECD has used the term “productive struggle” in discussing this issue. I think of it as necessary discomfort.

AI can remove the friction from producing an answer. But some of that friction is precisely what produces learning.

Consider the graduate engineer

Engineering provides a useful example.

Imagine a graduate engineer confronted with a model producing an unexpected result.

They spend several hours trying to understand it.

They check the inputs. Revisit the equations. Question their assumptions. Go back to first principles. Ask a more experienced engineer for advice. Discover that an assumption they considered insignificant is driving the result.

Eventually they solve the problem.

From a productivity perspective, those hours might look terribly inefficient.

An AI-enabled engineering system might have identified the issue in seconds.

But something else happened during those hours.

The graduate developed a deeper understanding of the system. They discovered the consequences of an assumption. They learned something about their own reasoning. They experienced uncertainty and worked through it. And they added another small piece to the body of experience from which professional judgement eventually emerges.

They did not simply solve the problem.

They changed themselves through solving it.

That distinction matters enormously.

Where will tomorrow’s senior engineers come from?

It also exposes a potentially much bigger workforce problem.

Experienced engineers did not arrive in the profession with experienced judgement.

They developed it.

They performed calculations. Produced drawings. Built models. Wrote specifications. Visited sites. Made mistakes. Had their work challenged. Observed failures. Worked alongside experienced colleagues and gradually learned to recognise patterns that aren’t necessarily written in textbooks.

Much of that is junior work.

And junior cognitive work is precisely the kind of work AI is becoming increasingly capable of augmenting or automating.

This creates an uncomfortable possibility.

If organisations use AI primarily to improve short-term productivity by removing junior tasks, they may inadvertently remove some of the experiences through which junior professionals become senior professionals.

We could create an expertise pipeline problem:

AI removes junior work → fewer opportunities to learn through doing → less experiential development → weaker professional judgement later.

That does not mean we should require graduates to perform work manually forever.

That would make little sense.

Once capability has developed, tools that make professionals faster and more effective should absolutely be used.

The more difficult question is which experiences should not be removed too early because the experience itself is developmental?

That is not simply a question for universities.

It is a question for employers, professional bodies and anyone responsible for developing future expertise.

Efficiency and learning are different objectives

Organisations naturally optimise for productivity.

If an AI system allows a task that previously took a graduate engineer four hours to be completed in 20 minutes, the commercial logic for using it is obvious.

But education and professional development have a different objective.

Sometimes the inefficient way of doing something is precisely how expertise is created.

That means we may need to become much more deliberate about distinguishing between production environments and learning environments.

In a production environment, we generally want the best answer as efficiently as possible.

In a learning environment, getting the answer may be only part of the objective. The cognitive journey required to reach it may matter just as much.

This suggests that responsible adoption of AI in education and graduate development cannot simply ask:

Where can AI make this faster?

It must also ask:

What is the human learning objective here, and what happens if AI removes the effort required to achieve it?

Sometimes AI should provide the answer.

Sometimes it should provide a hint.

Sometimes it should challenge the learner’s reasoning.

And sometimes perhaps it should not be used at all.

Knowing the difference will become an important part of educational design.

So what is education for?

I don’t think the answer is to abandon human capital and replace it with human flourishing.

Human capital still matters.

We need knowledgeable people. We need technically capable people. We need people who can calculate, analyse, design, write, build and understand complex systems.

And importantly, people often need to do those things themselves before they develop the judgement required to supervise a machine doing them.

But AI does expose a weakness in an education system focused too heavily on the ability to perform tasks.

If machines can increasingly do, then education must place greater emphasis on the human capabilities required to decide what should be done, why it matters, what good looks like and when the machine might be wrong.

That means knowledge and technical competence, but also judgement, curiosity, adaptability, ethics, agency and the ability to deal with ambiguity.

It also means recognising something that can easily be lost in our enthusiasm for frictionless technology.

Learning can be difficult.

It can be frustrating.

It can occasionally be unpleasant.

And that is not necessarily something we should engineer away.

The challenge for education in the age of AI may therefore be more subtle than deciding between human capital and human flourishing.

It is to develop people who have the knowledge and capability to work effectively with increasingly capable machines, while preserving the experiences through which human understanding and judgement are formed.

We should use AI to extend human capability, not inadvertently bypass the process through which that capability develops.

Perhaps that is where human capital and human flourishing ultimately meet.