Last time I argued that AI didn’t get rid of the cost of knowledge work, it moved it. The saving you make on producing the work turns up again as the cost of checking it, and checking is now the thing everything else waits for. Which leaves a question I dodged: if judgement is the scarce, expensive thing, where does it come from? Because it does get made somewhere. It has a production process, slow and unglamorous, and the evidence coming in over the past year says we’re taking it apart in three places at once.
Start with how the skill gets made. Ask any senior person you rate where their judgement came from and nobody mentions a course. They tell you about the pitch that died in the room, the analysis a client pulled apart in front of them, the confident call that met the real world and lost. Mine got built the same way, through workshop designs that looked lovely on paper and then died in front of thirty people, which is a kind of feedback no framework prepares you for. The story is always the same shape: years of work you owned, marked by what actually happened, until the judging got quick enough to feel like instinct. And the detail that matters is this. It wasn’t doing the work that built the skill. It was doing the work and then feeling the result. That’s all an apprenticeship ever was: a long run of small jobs with real outcomes attached, cheap enough that your mistakes didn’t sink anyone, real enough that they taught you something.
Now the first place the line is breaking: the way in. The evidence here is genuinely disputed, so I’ll give you both sides. Researchers at Stanford (Brynjolfsson, Chandar and Chen, 2025) went through payroll records for millions of American workers and found that since late 2022, employment for 22 to 25 year olds in the jobs most exposed to AI has fallen 13% relative to everyone else, and the fall is worst where AI replaces the work rather than helps with it. It’s a working paper, and the New York Fed (Audoly, Guerin and Topa, 2026) has pushed back hard: the decline in those jobs started before ChatGPT existed, and job ads show no clear split between junior and senior hiring.
LinkedIn’s UK numbers, from a company marking its own homework but with a very large dataset, also show no AI hiring shock yet. So I won’t tell you the door is closing, because honestly nobody knows. The signal I find harder to argue with is a stranger one. PwC looked at over a billion job ads this year and found early-career openings flat in the most exposed sectors, while the junior jobs that survive increasingly ask for senior skills: leadership, strategic thinking, judgement. Read that again slowly. We’re asking people to arrive with the thing the job used to exist to teach them.
The second break is inside the work itself, and here the evidence is small but direct. Shen and Tamkin (2026) ran an experiment where developers learned an unfamiliar programming tool, some with AI help and some without. The ones using AI understood the concepts less, read code worse and were poorer at fixing problems, and on average they weren’t even faster. The people who handed the whole job to the AI produced the most and learned the least. One study, 52 people, one narrow task, not yet peer-reviewed, so hold it lightly. But keep the detail that matters. Of the six ways people used the AI, the three where they stayed properly in the work preserved the learning. Handing the work over kills the apprenticeship. Staying in it doesn’t. And nobody in that study was choosing how to use the tool with their own development in mind.
The third break is the teachers. Judgement was never learned from documents. Someone a rung or two up watched your work, told you why it was wrong, and gave you the next slightly harder thing. Two Harvard researchers (Shin and Sucher, 2026) interviewed people at two big consulting firms and found that layer drowning. The middle managers had picked up all the new work of checking AI output, catching its mistakes and coaching juniors on the tools, on top of everything they already did, with no extra support. Eighteen interviews in one industry, so a pattern worth noticing rather than proof it’s everywhere, but it matches what I see in rooms every week. And Michael Watkins, who has spent decades helping executives step up into bigger jobs, wrote in HBR the same month that leaders are now arriving in senior roles with fewer of the experiences that used to get them ready. Every layer needs more judgement than it did, and the layer whose job was to grow it has been turned into a checking department.
Here’s the obvious objection, and it deserves a proper answer because every generation has made this complaint. Calculators were going to rot arithmetic. Coding tools were going to produce programmers who couldn’t code. The work rearranged itself around the tool and people learned higher things instead. The reason this time might be different is simple to state. Those tools did the doing and left the judging alone. The calculator did the sum, but you still picked the sum and knew when the answer smelled wrong. AI writes the draft, the option, the first answer itself, and producing first answers is exactly the work judgement used to grow out of. Lisanne Bainbridge spotted the trap in 1983: automate a process and the humans keep the hardest jobs while losing the practice that kept them able to do those jobs. She was writing about factory control rooms. It reads like a memo about your office. I hold this as an open question, not a prediction, because nobody knows yet how work will rearrange itself. But that’s the specific reason to suspect this tool bites where the others didn’t.
Economists at the Atlanta Fed (Afrouzi and colleagues, 2026) have now put the whole arrangement into a model, building on Kenneth Arrow’s sixty-year-old point that skill is a side effect of doing. Automate the junior work and every firm saves money today, while the economy quietly stops making the people who would have become the seniors everyone needs ten years from now. It’s a model, not a measurement, so treat it as a mechanism worth taking seriously. Though you barely need the maths. The saving shows up this quarter. The bill arrives years later, addressed to someone else.
This is why the reflex answer, protect the junior roles, is weaker than it sounds. Doing junior tasks was never the point on its own; the consequences did the building. Keep juniors doing the work by hand while AI does the real version next to them and you’ve built a flight simulator with the crashes turned off. Everyone looks trained. Nobody got calibrated. And no dashboard will tell you, because most organisations are living off reserves: the judgement their current senior people built the old way, before the loop got cut. You can run down a reserve like that for a decade with every number green, right up until those people leave and you find the pipeline behind them is full of people who produced plenty and were tested by nothing.
One honest flag before I finish. “AI is making a generation who can’t really judge” is a comfortable thing for someone my age to believe, and the older have been saying it about the younger forever. The fact that it suits me is a reason to hold it at arm’s length. And the evidence itself argues against despair: in the one experiment we have, the people who stayed in the work kept learning. Consequences can be designed. Ownership can be handed over on purpose, earlier, on smaller things, with something real riding on it. Some of what the old apprenticeship did by accident could probably be done deliberately.
Probably. Nobody has shown it working at scale, and I include myself in that. So here’s the question I’d put to any leadership team. You depend on tested judgement more than you ever have, and the cheap apprenticeship that made it has been switched off. What, exactly, is your replacement? If the answer is a training course, that’s the flight simulator again. If the answer is silence, you’re running on reserves, and the fuse is already lit.
References
- Afrouzi, H., Blanco, A., Drenik, A. and Hurst, E. (2026) Automation, Learning, and Career Dynamics. Federal Reserve Bank of Atlanta Working Paper 2026-6. Available at: https://www.atlantafed.org (Accessed: 3 July 2026).
- Arrow, K.J. (1962) ‘The economic implications of learning by doing’, The Review of Economic Studies, 29(3), pp. 155–173.
- Audoly, R., Guerin, M. and Topa, G. (2026) ‘Do job postings show early labor-market effects of AI?’, Liberty Street Economics, Federal Reserve Bank of New York, May 2026. Available at: https://libertystreeteconomics.newyorkfed.org (Accessed: 3 July 2026).
- Bainbridge, L. (1983) ‘Ironies of automation’, Automatica, 19(6), pp. 775–779.
- Brynjolfsson, E., Chandar, B. and Chen, R. (2025) Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab working paper. Available at: https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/ (Accessed: 3 July 2026).
- LinkedIn Economic Graph (2026) The UK Labour Market: Unlocking Growth in the Age of AI. LinkedIn Corporation.
- PwC (2026) 2026 Global AI Jobs Barometer. Available at: https://www.pwc.com/gx/en/services/ai/ai-jobs-barometer.html (Accessed: 3 July 2026).
- Shen, J.H. and Tamkin, A. (2026) How AI Impacts Skill Formation. arXiv preprint arXiv:2601.20245. [Not peer-reviewed.]
- Shin, J. and Sucher, S.J. (2026) ‘AI adoption is overloading your middle managers’, Harvard Business Review, 26 June. Available at: https://hbr.org/2026/06/ai-adoption-is-overloading-your-middle-managers (Accessed: 3 July 2026).
- Watkins, M.D. (2026) ‘3 forces are redefining the transition from manager to leader’, Harvard Business Review, 17 June. Available at: https://hbr.org/2026/06/3-forces-are-redefining-the-transition-from-manager-to-leader (Accessed: 3 July 2026).