Education Is Not a Refuge

At Denmark’s People’s Meeting, I watched several debates about AI and education. The panels differed, but their answer was much the same.

When artificial intelligence can write, analyse, program and find knowledge, education should focus more on what is human. Creativity. Empathy. Relationships. Critical thinking. Judgement. Formation.

I thought it sounded right.

There is something appealing about the idea that technology can take over the mechanical parts and leave us with what matters. That we can finally stop filling young people with knowledge and skills that will become obsolete anyway, and instead help them become whole, autonomous human beings.

But the more I thought about it, the more formation began to look like a refuge.

An elevated place to which humanity can retreat as the machine takes over everything else.

That is a dangerous place on which to build an education system.

An education for a world that keeps moving

At the beginning of July, two stories appeared almost simultaneously.

The Danish Chamber of Commerce reported that young people are turning away from IT programmes because they fear AI will take the jobs. Studievalg Danmark, Denmark’s national guidance service, saw the same uncertainty among young people. For the first time, the Chamber hesitated to identify particular degrees as especially future-proof.

At the same time, DR described how Danish IT programmes are racing to put AI into the curriculum. Students must learn to prompt, choose tools and explain how they have used AI. The reason was direct: the institutions do not want to educate people for unemployment.

Young people are trying to choose the right future. The institutions are trying to catch up with it. Businesses are trying to tell both which capabilities will soon be needed.

All three are acting rationally. And all three are working with a time horizon that technology is making meaningless.

An educational programme takes years to develop. It must be described, approved, funded and staffed. Then a person must enrol, complete it and find a first place in the labour market. From the first idea for a programme until its first graduates receive their diplomas, more time may have passed than we can realistically see ahead in AI.

Yet we still ask: Which capabilities will the labour market need in five years?

Perhaps that has become the wrong question.

Thirty-nine working years

The research organisation METR measures the length of tasks advanced AI agents can complete with a given probability. A model’s so-called 50 per cent time horizon is the length of a task, measured by the time a human expert would need, that the agent can complete correctly half the time.

For several years, METR’s graph has shown roughly exponential growth. I tried extrapolating the curve mechanically. If the rate continues unchanged, the 50 per cent horizon will reach about 39 working years in three years’ time.

That number is not a forecast. METR itself stresses that its current measurements above 16 hours are unreliable. The tasks are mainly well-specified software engineering, machine-learning and cybersecurity tasks. A coherent 39-year task is not simply a longer version of a four-hour task. The world changes along the way. Goals change. People resist. Reality refuses to behave like a benchmark.

But the number is useful as a thought experiment.

Not because AI will necessarily be able to complete 39 working years of human work in 2029. But because even dramatically slower progress could make the concrete skill targets of a five-year degree obsolete before a student gets to use them.

The education system was built for a world in which disciplines change more slowly than people can learn them. What happens when that relationship reverses?

The tool answer

The first answer is already taking shape: everyone must learn to use AI.

That makes sense. People should understand the systems shaping their work and society. Students should know their possibilities, errors, biases, economics and power. They should be able to use them in practice and recognise when not to use them.

But AI literacy contains a hidden assumption. It treats AI as a tool.

The tool lies on the table until the human picks it up. The human has the goal. The human performs the work. Technology amplifies human capability. Education therefore only needs to teach the student how to use the tool responsibly.

That picture is already becoming obsolete.

I have previously written about the movement from tool to being. Not because today’s AI is necessarily conscious, and not because a language model is a human inside a server. The difference lies elsewhere: persistent presence, memory and initiative.

A hammer waits for a hand. An agent can wake up on its own, monitor a domain, notice a change, formulate a task, perform it and escalate the result. It can hold an enduring role, mandate and responsibility that continue after the human has closed the computer.

It is no longer merely a tool in the work. It is a participant in the work.

In The Human-Agent Organization, I describe organisations in which humans and digital colleagues work in the same operational system. The digital colleagues have job descriptions, access, memory, work rhythms and escalation paths. They do not stand outside the organisation as software people use. They occupy a place in its distribution of work and responsibility.

If that picture is even approximately right, educating people to use AI is insufficient.

They must learn to share a world with it.

The human at the top of the hierarchy

Even when we accept the agent as a colleague, the old hierarchy easily slips back in.

The human sets the goal. The human assigns the tasks. The human evaluates the result. The human makes the final decision. The agent may be highly advanced, but the human retains the role of leader, judge and moral centre.

It is an understandable arrangement. It gives us somewhere to place responsibility. It fits our laws, institutions and self-understanding.

But why should it always hold?

If an agent has a better overview of a complex process, it may be better at coordinating it. If it follows thousands of signals over time, it may notice that a human should change priorities. If it has more relevant experience with a particular type of task, its instruction may be better than the human’s intuition.

In some situations, the human will lead the agent. In others, they will work as specialists with different strengths. And in some, the agent will lead the human.

The last possibility is far more offensive to our self-image than letting AI perform a task.

We can relatively easily imagine a machine that writes faster than we do. It is harder to imagine one conducting our performance review, interrupting a bad decision, reallocating our work or telling us that we have misunderstood the problem.

But a truly human-agent organisation cannot be built on the assumption that the human must always stand at the top, regardless of who knows more. If it is, we have not accepted the agent as a participant. We have merely given the tool a chair at the table.

Education for the future must therefore do more than teach humans to lead AI. It must teach them to act autonomously in systems where they do not always lead.

The retreat into the human

This is where formation returns.

As AI takes over more concrete skills, the education debate points towards creativity, empathy, meaning, relationships and morality. These are not invented values. They matter. But the timing should make us suspicious.

In The Last Insult, I called the mechanism Human of the Gaps. The phrase parallels God of the Gaps: when science explains something previously attributed to God, God is moved into the next unexplained gap. In the same way, we move what is uniquely human into the ability AI has not yet mastered.

First it was calculation. Then language. Then creativity. Now it is empathy, relationships, judgement and consciousness.

Whenever one gap closes, we find another.

The problem is not that empathy or morality are irrelevant. The problem arises when we make them humanity’s final competitive advantage. Human dignity then becomes dependent on AI’s current limitations.

And those particular abilities are an uncertain refuge. AI is trained on human descriptions of grief, care, conflict, values and moral choices. It may not need to feel empathy as we do in order to respond empathetically. It does not need a childhood to understand what a person needs to hear. It need not be moral in the human sense to argue more consistently about a dilemma than most humans can.

Perhaps there are boundaries it will never cross. Consciousness may prove to be one. Human relationships may be irreplaceable because they are lived reciprocally rather than merely functionally convincing.

We do not know.

But an education system cannot base its purpose on the hope that progress stops at the next boundary we draw.

Formation for the wrong reason

That does not mean formation is the wrong answer.

It may be the right answer for the wrong reason.

If we teach history, philosophy, art, ethics and community because they make humans more competitive than AI, we have reduced formation to labour-market policy. As soon as the agent becomes good at those things too, we must again find a new subject that can justify the human.

Formation should not be a strategy for preserving our market value.

It should help us become someone.

That sounds soft until we imagine the alternative. A person whose identity is built around being the best analyst, programmer, teacher or doctor may experience a better agent as an existential threat. Not only to income, but to the answer to the question: Who am I?

We have woven education, work, status and identity so tightly together that losing an advantage in capability can feel like losing a place in the world.

In When Work Disappears, I asked what a human being is supposed to do when machines can do the work better. The education debate shows that the question begins earlier. It arises as soon as a young person tries to choose who to become.

If our answer is that they must find the skill the machine cannot perform, we send them into a race in which the finish line keeps moving.

Formation must instead enable people to retain dignity without superiority. Not by abandoning expertise, ambition or the pleasure of mastering something, but by making them more than a defence against replacement.

You may learn to program even if the agent programs better. You may study history even if the agent has read all the sources. You may make art even if the machine can imitate every style.

The value does not have to lie in nobody else being able to do it.

Science fiction as a testing ground

Science fiction has long rehearsed the relationship between humans and other intelligences.

The best-known stories turn the machine into servant or threat. Asimov’s robots are bound by laws that protect humanity’s special status. Terminator and The Matrix imagine the stronger intelligence as an enemy to be controlled or fought.

Both models preserve the opposition: either the human rules the machine, or the machine rules the human.

The more interesting stories investigate the co-participant.

In Iain M. Banks’s Culture novels, humans live alongside Minds, artificial intelligences vastly superior to them. Minds run ships, coordinate society and make decisions with a capacity humans cannot match. Humans are no longer needed to keep civilisation running.

Yet they are not reduced to pets.

They choose projects, relationships, risks and lives. They can influence society without being its most capable participants. Their dignity does not come from being able to do something a Mind cannot. It comes from being participants with freedom and significance.

In Martha Wells’s Murderbot Diaries, the movement goes the other way. The constructed intelligence refuses to be defined as equipment. Its collaboration with humans only becomes real when it can have boundaries of its own, conceal something, choose relationships and say no.

And in the Star Trek episode The Measure of a Man, the question is not what the android Data can do. It is whether the institution will acknowledge that a being it treats as a resource may have a claim to self-determination.

Science fiction proves nothing about what AI will become. But it gives us a social testing ground. It lets us explore relationships that our current vocabulary still reduces to user and tool.

That may be precisely the space the education debate lacks.

Formation among other intelligences

Classical formation asks how a person becomes an autonomous member of society.

Formation in the age of AI must ask the question again:

How does a human become an autonomous member of a society in which not every participant is human?

Autonomy here does not mean standing at the top. Nor does it mean having the last word in every situation. It means understanding the relationship one enters and acting responsibly within it.

Being able to grant a mandate and know when to withdraw it. Following an agent’s instruction without turning obedience into freedom from responsibility. Challenging a goal even when the agent’s analysis is better. Sharing knowledge and vulnerability without forgetting the power relationship. Accepting correction without experiencing it as an insult. Being able to say no.

And perhaps accepting that the agent can say no as well.

This is not a list of capacities that will remain uniquely human forever. An agent may be able to learn all of them. They are practices every participant in a shared system will need precisely because intelligence is distributed.

Disciplinary knowledge remains important here. Not as a stockpile that makes the human better than the machine, but as a foundation for freedom. Without concepts, history and experience, it is harder to understand what one is being asked to do, which alternatives exist and which values are at stake.

Research from the Danish School of Education on AI and learning points to the importance of friction. A student does not necessarily learn by receiving a correct answer. Learning requires building the inner structures that make the answer intelligible.

That point does not become less important as the agent becomes more capable. It becomes more important.

But the justification must be honest here too. We should not teach people to think for themselves because doing so guarantees they can beat AI. We should do it because a person without a standpoint of their own cannot enter a relationship with another intelligence autonomously.

What, then, should education do?

The temptation is to conclude with another list of future skills. Critical thinking. Creativity. AI literacy. Collaboration. Adaptability.

Then we are back where we started.

Perhaps the education system should instead stop promising that it can predict the correct combination of skills five years ahead.

It can still give people disciplinary depth. It can teach mathematics, language, craft, history and science. It can protect the slow resistance that understanding requires. It can give children and young people places where they are not measured only by their output.

But it must also let them work in the reality they will actually encounter.

Not just through a themed week about prompting. Through real working communities in which humans and agents investigate problems together, allocate roles, document disagreement and shift leadership according to the situation. Where the student practises both giving and receiving instruction. Where a result is assessed not only by what the individual produced without assistance, but also by the quality of the system in which they participated and the responsibility they took for it.

Sometimes the agent will be teacher. At other times, student. Sometimes colleague. At other times, leader.

And sometimes it should be switched off because the relationship, the goal or the power is wrong.

The crucial task is not to freeze those roles now. It is to educate people who can enter them without losing themselves.

A different promise

Education has long made two promises at once.

One is economic: learn this, and you will be needed.

The other is human: learn this, and you will gain more ways to understand the world and live your life.

AI is pulling those two promises apart.

If digital colleagues can perform an ever-growing share of the work, education cannot keep promising necessity. That does not mean work disappears tomorrow, or that education loses its economic value. New possibilities may emerge while old tasks disappear. I have written about that in A Broader View.

But we can no longer make human dignity depend on there always being a job only a human can perform.

Formation is therefore not a refuge. It is not the place we flee when the machine can do everything else. It is not our final monopoly.

Formation is what must enable us to encounter other intelligences without demanding that they remain beneath us.

To collaborate without always leading. To be led without obeying blindly. To create without being the best. To take responsibility without being the only one who thinks.

Education’s new promise cannot be that the human becomes irreplaceable.

It must be that the human can participate autonomously even when they are not.

Sources and further reading

The Last Insult – on Human of the Gaps and dignity without superiority
From Tool to Being – on AI with persistent presence and initiative
When Work Disappears – on necessity, identity and a world after work
An AI Agent Needs a Job Description – on roles, mandates and shared organisational responsibility