Where judgement used to come from

The number that is being read wrongly

Employment among workers aged 22 to 25 in the most AI-exposed occupations now stands roughly nineteen per cent below where it would have been had it kept pace with their less-exposed peers, according to Stanford Digital Economy Lab's analysis of high-frequency US payroll data through June 2026.

The pattern appears to be driven primarily not by dismissal, but by hiring that does not happen.

That distinction matters more than the headline. There is no mass displacement event to observe. A door is quietly not being opened. And because there is no obvious moment of displacement, there is no single event to respond to — only an absence whose consequences may become legible in about a decade.

Apprenticeship was a by-product of slow work

Entry-level work was never only valuable for its output. Much of the output was relatively cheap. What it also produced was exposure: to ambiguity in a real situation, to a client who was not satisfied, to a number that did not reconcile, to a senior colleague thinking aloud badly before thinking well.

Some organisations designed apprenticeships and development programmes deliberately. But much of professional formation happened almost accidentally. It was a by-product of work being slow enough that a junior person had to stay inside a problem long enough to develop a feel for it.

The difficult client. The calculation that would not reconcile. The meeting where you did not understand half of what was being said. The mistake you had to explain rather than quietly correct.

These things rarely appeared in a development curriculum. They were simply part of becoming experienced.

And perhaps that is why we have been slow to notice when some of the conditions producing them begin to disappear.

What is removed is situations, not tasks

The usual framing asks which tasks a model can do. That may be the wrong unit.

A task can often be delegated without much loss. A situation is different. Situations are where judgement gets tested and formed — because judgement is partly what remains after enough situations have been encountered without knowing the answer in advance.

Research from Microsoft and Carnegie Mellon on knowledge workers using generative AI offers an interesting warning from inside the work itself. Higher confidence in the tool was associated with less critical thinking, while the researchers also found that generative AI changes where critical-thinking effort occurs — away from some forms of information gathering and task execution and towards verification, integration and oversight.

The work still gets done.

What is less certain is what happens to the formation that used to occur while doing it.

That may be one of the less visible things AI changes. Not simply labour, and certainly not every junior role, but the inability to skip to the answer.

And some of that inability was apprenticeship.

What this asks of the people who hire

If this is right, the response is not to slow down the tools. Nor is it to preserve inefficient work simply because previous generations learned from it.

It is to recognise that the formation of judgement may increasingly become something an organisation has to design deliberately, having previously received much of it as a side effect of work itself.

For anyone who hires senior people, the practical consequence arrives later.

The people entering organisations now are among those who will be considered for senior leadership roles ten or fifteen years from now. Their CVs may show experience, progression and increasingly impressive output. But those things do not necessarily tell us how often they had to remain inside uncertainty without an immediate answer.

That may eventually change what we need to look for.

Search does not create judgement. It can only discover where and how it was allowed to form.

And perhaps the larger question belongs to organisations long before the search begins:

If efficiency removes some of the places where judgement used to form, where will we build them next?