An experienced plumber looking at a recurring pressure problem knows more than the sequence of tests in a manual.

They know which innocent-looking detail usually matters. They know which test to run first, which apparently sensible repair will lead to another call-out in three months, and when the standard procedure is wrong for this particular building. Much of this may be difficult to explain. It appears in practice as a question asked earlier than expected, a possibility dismissed without fuss, or a pause before accepting the obvious answer.

At present, this knowledge is normally sold together with the plumber’s time. The customer pays for a visit, during which experience and labour arrive as a bundle.

But imagine that the plumber could teach a machine the important parts of that judgment. The questions to ask. The evidence to examine. The tempting mistakes to avoid. The exceptions that change the diagnosis. The point at which the system should stop and call a person.

The plumber’s knowledge could then help solve thousands of problems they never personally attended.

Who should receive the value created by what they have learned?

Intelligence still needs contact with reality

AI promises to make intelligence abundant. It can already analyse information, operate software, produce documents and carry out long sequences of intellectual work. The cost of applying a method is falling quickly.

Yet useful intelligence requires more than processing power. Somebody has to discover what matters in the first place.

People encounter parts of the world that have not been adequately described. They notice when the official procedure and reality have drifted apart. They learn which sensible-looking ideas fail in practice. They decide which outcomes are worth pursuing, earn the trust of the people affected, and live with the consequences when a decision is wrong.

Much of this knowledge does not exist in any manual or dataset. It emerges through repeated contact with reality.

Much of intellectual work can be understood as three layers. Someone works out what matters. They turn that understanding into a method. The method is then carried out, case after case.

Historically, all three layers usually required human labour. Increasingly, the third does not. In many areas, AI will also help substantially with the second.

This need not make the first layer worthless. It may finally separate its value from the hours spent repeatedly applying it.

We know more than we can explain

There is an immediate objection. Expertise is often tacit. People can be reliably good at something without being able to give a complete account of how they do it.

Earlier attempts to capture expert knowledge, including the expert systems of the 1980s, ran into this problem repeatedly. Asking an expert to convert years of experience into formal rules is slow and frustrating. The resulting system tends to be brittle: it works on the cases anticipated by its designers and fails when reality presents something slightly different. Knowledge changes, exceptions multiply, and the rulebook becomes harder to maintain.

Modern AI does not make tacit knowledge disappear. Some knowledge may remain embodied, social or inseparable from the circumstances in which it is used. There will continue to be work where presence and personal responsibility are essential.

What has changed is the way a person can teach a machine.

An expert need not begin with a complete decision tree. They can demonstrate cases, compare good and bad results, correct proposed decisions and explain exceptions as they arise. Conversation and correction can draw out knowledge that an expert would struggle to produce as a manual on demand.

The articulation problem remains, but the barrier is lower. That difference creates the possibility of capturing forms of practical knowledge that would previously have been too expensive or too awkward to formalise.

We have done a simpler version before

Medicine offers a useful precedent. Surgical safety checklists crystallise knowledge gathered from clinicians, researchers and past failures into a method that other teams can apply. The checklist does not contain the whole of a surgeon’s expertise, and it does not perform an operation. It captures a set of consequential checks that are easy to miss under pressure and makes their application repeatable.

This is one way knowledge already scales beyond the person who discovered it. Standards, protocols and checklists allow an insight to improve work carried out elsewhere by people the originator will never meet.

They also expose the economic gap. The people who contribute to a widely used professional method are rarely paid each time it creates value. Often that is entirely appropriate: some knowledge should be a public good, freely available to everyone. But it should not be our only model, particularly when a private system earns money by repeatedly applying a person’s distinctive expertise.

We have economic structures for other contributions that can be separated from their repeated use. A songwriter need not perform a song personally every time it earns money. An inventor need not manufacture every product that uses an invention. Their original contribution can remain connected to its later economic value.

Practical insight could sometimes be treated in a similar way. A person develops a better method for assessing a tender, preventing a manufacturing defect, organising a care rota or recognising a failing project. They teach that method to a system. The system applies it repeatedly. When the method proves useful, the originator remains recognised and shares in the value.

For most of history, ordinary people could sell their time but had few ways to turn what they had learned into a durable asset. A small number wrote books, secured patents or built companies. Most useful knowledge remained local. It was passed informally to a few colleagues, repeated manually throughout a career, and frequently lost at retirement.

AI could give many more people the ability to make experience durable.

The copying problem

The royalty comparison describes an economic shape we might want. It does not describe a legal system that already exists.

A song is a fixed expression protected by copyright. Practical methods are much harder to own. Copyright will not ordinarily prevent somebody from expressing the same procedure differently. Patents apply only in limited circumstances. Trade-secret protection depends on keeping knowledge secret, which is difficult to reconcile with distributing it for others to use.

There is a deeper problem. A system that applies a method may accumulate examples, corrections and outcomes. Its operator might eventually use that evidence to reproduce or improve the method without the original contributor. A model does not necessarily learn merely by performing a task, but the records of repeated performance can become training material. The machinery that distributes an expert’s knowledge may also make that expert easier to remove.

Any serious market for practical insight must confront this directly. It cannot rely on models remaining incapable of learning from use.

Contributors would need enforceable terms governing how their knowledge and the resulting records can be used. Systems would need provenance: a reliable account of who supplied a method, how it changed, and which later results depended upon it. Payment would require transparent measures of use. There would also need to be rules for independent discovery, common professional knowledge, collaborative improvements and the point at which an original contribution has become too remote to justify continuing payment.

Some methods should never become private toll roads. Others will belong to an employer rather than an individual. Many insights will be independently discovered by several people. These are hard boundary questions, but leaving them unanswered would settle the matter by default in favour of whoever owns the system.

An economy built around contribution

There is an unappealing version of the post-AI economy in which we preserve human activity simply to preserve employment. Machines are prevented from carrying out useful tasks, leaving people to perform work everyone knows could be done mechanically. Human involvement becomes a ceremony required by law, convention or sentiment.

Drudgery does not acquire dignity merely because we are frightened of what may replace it.

A better settlement would give greater economic weight to contribution and less to repetition. People should not have to spend their lives copying information between systems, producing routine variations of the same document, or applying an already-settled decision to one case after another. Their greater value lies in noticing something true, deciding what good work looks like, teaching others and improving the method when reality changes.

That is substantive intellectual work. It asks people to observe, investigate, choose, explain and take responsibility. Machines can carry more of the repetition.

Such an economy will require institutions capable of distinguishing expertise from confidently packaged nonsense. Useful methods will need testing, maintenance and challenge. Reputation will matter. So will the ability to inspect what a machine has been taught to do in somebody’s name.

It will also require credible payment. Today’s creator economy offers a warning: a small group earns substantial sums, a long tail earns almost nothing, and intermediaries encourage millions of people to dream while retaining much of the dependable value. Renaming practical experts as creators would not improve their position.

If human insight is a scarce input into useful AI, the people supplying it need a defined place in the economics. Exposure and distant promises are insufficient.

A different question about the future

Human beings cannot assume a permanent economic advantage in producing intelligence. Machines are becoming remarkably good at analysis, creativity and professional reasoning. Each supposedly protected category of knowledge work is likely to face the same pressure in time.

The question What can a human do that a machine can never do? therefore offers fragile comfort. “Never” is a dangerous foundation for an economic system.

A more useful question is: What have people learned that could make intelligence more useful?

Nearly everyone who has done something seriously for a long time has an answer. The nurse who can see which discharge plan will fail. The site manager who knows where the schedule is pretending. The teacher who can tell the difference between a child who has not understood and one who has stopped trying. The owner of a small business who has developed a way of handling a problem that larger organisations still get wrong.

Much of that knowledge currently disappears. In the future, it could become a lasting contribution: expressed once, tested in practice, improved over time and applied wherever it helps.

That future would change what people are paid for. A person’s economic value would not depend entirely on remaining present for every repetition of the work.

The machine could carry out the method. The person who taught it something worth doing would remain part of the value it creates.