The test people use for whether a machine takes a job is whether a machine could do the work. That is not the test that decides.

Ask most people whether a job is safe from automation and they will run the same check. They picture the work. They ask whether a machine could plausibly do it. Rewiring a house, fitting a boiler, driving a lorry through Bedford in the rain: no, they conclude, not for a long while yet.

The check is reasonable and it answers a real question. It is simply not the question that determines whether the job is there.

A job exists when somebody is willing to pay for it. That depends on how many firms are operating, what it costs them to employ a person, whether customers are spending, and whether the work is being done here or somewhere else. A machine's capabilities are one input among those, and often not the loudest one.

The Office for National Statistics put a plain example of this in its release of 21 July. Reporting a fall of 8,000 vacancies in businesses with one to nine employees over the quarter, it noted that feedback from its own Vacancy Survey suggests some small firms may not be recruiting because of increases in labour costs and other operating expenses.

Read that slowly. Nothing in it concerns what a machine can do. A firm of six people has decided not to take on a seventh because the seventh costs more than they can carry. The work that seventh person would have done is entirely human, entirely un-automatable, and entirely absent.

This desk's read, and the seam between the evidence and the read is here: the capability test feels authoritative because it is vivid. You can picture a robot failing to fit a boiler. You cannot picture the meeting where a builder in Wakefield decides against hiring an apprentice this year. The first is a scene. The second is an absence, and absences do not form images.

So the advice survives on the strength of a picture, while the thing that governs the outcome happens somewhere nobody can see.

There is a version of this the desk has reported before, from the other direction. In July we noted that a large employer shedding thousands makes the news while a small firm quietly not hiring a junior makes none. The same asymmetry is at work. What is easy to imagine gets treated as what is likely, and what is diffuse gets treated as though it were not happening.

None of this makes the advice wrong. Plenty of people have gone into a trade and done very well, and the national figures say nothing about any particular person in any particular town. The caution is narrower. When someone tells you a job is safe because a machine cannot do it, they have answered a question about technology and left the question about demand untouched. Those are different questions, and only one of them determines whether there is a job at the end of the training.

The useful thing to ask is not what a machine can do. It is who is currently hiring, in what size of firm, and where.