Does AI build expertise, or quietly erode it?

This piece condenses a longer research review I maintain in my working knowledge system. Sources are listed at the end.

Knowledge work has always reproduced itself through a hidden curriculum. Juniors did the tedious, formative work — the document review, the first-draft analysis, the boilerplate code — and by doing it badly, then less badly, built the judgement that later let them recognise good work at a glance. Generative AI now does much of that formative work on demand, at acceptable quality, with no learning required of the human. The worry is obvious once stated: the output arrives without the understanding of how it was assembled.

To take the worry seriously you need an account of how expertise actually forms, and the research here is unusually settled. Expert performance is organised perception, not superior reasoning. Chess masters reconstruct real positions almost perfectly and random ones no better than novices — they see meaningful chunks, not pieces. Gary Klein's fireground commanders didn't compare options under pressure; they recognised situations, and the recognition supplied the action. That pattern library is built through feedback-rich encounters with real situations. It cannot be installed by instruction, and its owner cannot fully articulate it. I watched this for eighteen years as a rowing coach: the knowledge that wins races is precisely the knowledge nobody can say.

If that's the mechanism, AI intervenes at a precise point — production, the traditional occasion for accumulating pattern-outcome pairs. And there is direct evidence of the damage. A 2025 multicentre study in The Lancet Gastroenterology & Hepatology found endoscopists' adenoma detection dropped from 28.4% to 22.4% when they returned to unaided work after a period of AI assistance — experienced clinicians, measurably deskilled. An MIT Media Lab study found students who wrote essays with ChatGPT from the outset showed the weakest neural signatures of learning and struggled to quote sentences they had just produced; the authors called it cognitive debt. And Stanford's analysis of US payroll data found employment for 22-to-25-year-olds down 13% in the most AI-exposed occupations while older workers in the same jobs held steady — the entry-level rung, the one people climbed to expertise on, bending first.

But an equally serious body of evidence cuts the other way. When 5,000 customer-support agents got an AI assistant, novices improved 34% while experts barely moved — the AI was transmitting the tacit best practice of the strongest performers to the newest, moving juniors down the experience curve faster, with evidence of durable learning. Cognitive load theory has held for decades that novices learn more from studying worked examples than from unguided struggle, which recasts an AI's competent solution as exactly what a beginner should be shown. And the deepest paper in the automation literature — Lisanne Bainbridge's "Ironies of Automation", from 1983 — locates the danger somewhere more specific than lost friction: automate the routine work and you leave humans supervising a machine, which is the one task that most needs the continuous practice the automation just removed.

So the strong claim fails in both directions, and what survives is a narrower, more useful one: the outcome depends on the quality of the feedback loop and the design of the task, not on friction as such. Some friction is formative — effortful practice at the edge of ability, with fast, unambiguous feedback. Some friction is merely expensive. AI can improve the loop: instant, specific response to a draft is what the training literature has always prized. It can also poison the loop: feedback that is confidently wrong writes bad patterns into a learner who lacks the expertise to catch it.

Four things follow. Unaided effort first, AI second — the worst results in the learning studies came from AI-from-the-outset. Use AI inside your existing expertise, where you can supply the ground truth it can't supply about itself. Rebuild junior roles around evaluating work you didn't produce — a real skill, but one that has to be paired with enough hands-on production to make the evaluation mean something. And elicit tacit knowledge deliberately, because a repository captures the explicit residue of expertise, never the pattern-reading underneath — a limit I've written about separately: second brains fail because nobody elicits.

The honest position is that both patterns are visible in today's data and the decisive evidence — retained unaided ability, measured over years — does not exist yet. Which means the variable is still ours to set.