Chapter 11

The Apprenticeship Problem

How anyone becomes expert now — and why the ladder is breaking from both ends at once.

For as long as there have been professions, expertise has been produced the same way: put a person on the bottom rung and let the work teach them. The junior lawyer reads the discovery. The junior analyst builds the model nobody senior wants to build. The resident takes the overnight shift. The work is tedious, the pay is poor, and the tedium is the curriculum — because the pattern recognition that eventually becomes judgment is assembled out of a thousand unremarkable repetitions.

That machinery is now failing in two places at once, and almost every conversation about AI and jobs has only noticed one of them.

The first break: the rungs are disappearing

Through 2026 the labour data started to show a consistent shape. Researchers at Stanford's Digital Economy Lab, working with payroll records covering tens of millions of American workers, documented a roughly sixteen percent relative decline in employment among workers aged twenty-two to twenty-five in the most AI-exposed occupations. The 2026 AI Index found software developers in that same age band down close to twenty percent since late 2022 — while developers over thirty, at the very same firms, grew six to twelve percent. A Harvard working paper tracking hiring across hundreds of thousands of firms found entry-level hiring falling sharply at companies adopting generative AI, even as senior hiring at those companies continued to climb.

The pattern is not "AI destroys jobs." It is narrower and stranger than that: a seniority tilt. The work that is disappearing first is the repeatable, high-volume, well-specified work — junior analysis, first-draft research, basic implementation, routine support. Which is to say: the bottom rung. The part of the ladder that existed so that people could climb off it.

The second break: the remaining rungs stopped building anything

Here is the part almost nobody is tracking, and it is the more consequential of the two. Even where the junior role survives, it may no longer do what junior roles were for.

A 2026 study of skill formation among programmers found that people who leaned on AI to learn an unfamiliar library gained little speed in the short run — and ended up measurably worse at reading and debugging that code unaided. In a field experiment with nearly a thousand high-school students, those who practised mathematics with an unrestricted AI tutor outperformed their peers spectacularly while they had it, and then scored below students who had never used it at all once it was taken away. In medicine, experienced endoscopists' unassisted detection rates fell measurably after months of routine AI-assisted work — experts, quietly unlearning.

Put those findings beside the hiring data and the situation clarifies. The junior role is disappearing. And where it survives, the work inside it is increasingly being done by a tool, which means the person occupying the role is no longer being built by it. They are producing at a level their understanding has not reached — which looks like exceptional early performance and is something closer to a loan.

Deskilling is the wrong word

The literature has been calling this deskilling, borrowing a term from industrial automation. That word is increasingly wrong, and the medical education literature has coined a better one: never-skilling — the failure to develop a capability in the first place, as distinct from losing one you once had.

The distinction matters because the remedies are different. A deskilled professional has the capability somewhere; it has gone slack from disuse, and deliberate practice brings it back — which is why aviation, having watched manual flying skills decay, simply mandated manual flying hours. A never-skilled professional has nothing to restore. There is no earlier, stronger version of their judgment to recover. They have four years of impressive output and no formed intuition underneath it, and the first time they are handed a problem outside the templates, everyone in the room finds out simultaneously.

This is why the timeline is so cruel. Never-skilling is invisible for precisely as long as conditions stay normal, and it becomes visible at exactly the moment normal conditions end — the novel case, the outage, the client question that goes one layer deeper than the deck.

What actually predicts whether you develop

Which brings us to the most useful research finding of 2026, and the one this entire site is organised around.

A controlled experiment did something the earlier studies had not: it separated how much AI people used from how much independent reasoning they preserved while using it. Skill development tracked the second variable, not the first. Participants who asked for a great deal of help but kept thinking alongside it developed. Participants who asked for less but disengaged did not. The same shape appears in the programming research: automating a task wholesale accelerates the decay, while iterative use — where the human forms a view, evaluates the output, corrects it, and builds on it — can preserve or even scaffold skill.

So the question is not whether an apprentice should use AI. Of course they should; they will be competing with people who do. The question is whether the way they use it leaves any independent reasoning intact. That is a design problem, and it has answers.

If you are early in a career

The strategic situation is unusual and worth stating plainly: the market will pay you, early, for output you can produce with tools — and it will pay you, later, only for judgment those tools cannot produce. Those are different currencies, and the first one is depreciating.

If you are developing someone

The mirror image of this problem sits with anyone senior, and it is not a soft or optional responsibility any more. It is a supply problem: the organisations that stop producing expertise internally will be buying it from a market where it is getting scarcer and more expensive every year.

The instinct that causes the damage is generosity. A senior professional hands a junior the AI-accelerated version of a task because it seems kind — faster, less tedious, more impressive output. What it actually does is spend the junior's development budget on this quarter's throughput.

There is also a quieter loss. When every small question can be answered instantly by a machine, the junior stops asking them — and the small questions were never really about their content. They were the thread the relationship was woven from. Mentorship does not die of neglect in an AI-first team; it dies because the occasions for it stop occurring, and nobody notices anything is missing.

Why this one is urgent

Most of what this site argues can be taken up whenever you are ready. Attention can be rebuilt at fifty. Taste keeps developing for as long as you keep looking. The research on adult neuroplasticity is unambiguous that capacity responds to use at any age.

The apprenticeship problem is the exception, because it has a window. The years in which a professional's judgment forms are specific years, and they are the same years the market is currently removing from the bottom of the ladder. A cohort is passing through that window right now with impressive output, historically weak formation, and no idea that the two are related — and roughly half of business leaders surveyed in 2026 already report seeing deskilling in their organisations, with more than sixty percent expecting it to become a genuine threat within three to five years.

Which means the people who solve this — for themselves, or for the people they are responsible for — will hold something that is becoming structurally scarce rather than merely rare. Not because they refused the tools. Because they kept building the thing the tools were supposed to be augmenting.