Last time, I wrote about how AI is inflating skills the way calculators inflated arithmetic and GPS inflated navigation — making things that used to be scarce available to almost anyone, while quietly thinning the capability behind it if nobody practices it deliberately. Three questions kept coming up in response, and they deserve more than a footnote: does this end with the job itself becoming redundant, not just the skill? How does someone junior build the judgment they haven't had time to earn yet? And how does someone who's already excellent stay that way?
Is inflation just a waystation to redundancy?
The honest answer is: probably not, based on how this has played out historically — but I'd hold that loosely.
The foundational insight from labor economists Autor, Levy, and Murnane, later extended by Acemoglu, is that automation doesn't hit jobs. It hits tasks. A job is a bundle of tasks, and technology tends to eat some of them while leaving others — often the ones requiring judgment, context, or dealing with the genuinely novel — completely untouched. A 2018 analysis by Brynjolfsson and colleagues, which scored task suitability for machine learning across the U.S. occupational database, found that very few occupations are fully automatable end-to-end; almost all contain a meaningful share of tasks that resist it.
That's the pattern across the last two centuries of technological change: roles get rebuilt around a smaller core of judgment-heavy tasks, not eliminated wholesale. But it's worth being precise about what that claim rests on — historical pattern, not law of nature. Whether generative AI, which unlike a robot doesn't need a stable physical environment to operate in, breaks that pattern is a genuinely open question that the data on this specific wave is still too young to answer. I'd treat "it's always been task automation, not job automation" as the reasonable base case, not a guarantee.
Building judgment before you've earned it
This is the harder problem, because the traditional route to tacit knowledge was time: you saw enough hard cases, made enough small mistakes under supervision, and eventually pattern-recognition set in. AI shortens the queue for output but doesn't automatically shorten the queue for judgment.
The research on expertise, going back to Anders Ericsson's work on deliberate practice, found something specific: raw years of experience correlate weakly with actual performance. What predicts it is structured, feedback-rich practice aimed at the edge of your current ability — not repetition of what you already do comfortably. That's the same whether the domain is chess, surgery, or writing a client proposal.
The complication AI adds is documented in a 2025 review on what researchers call "upskilling inhibition": when juniors default to AI for the harder cases too early, they lose exposure to exactly the difficult, ambiguous situations that would have built their judgment. The tool doesn't just do the easy 80% faster — used carelessly, it quietly removes the person from the room during the 20% that was actually the training.
In my coaching work, the practical version of this is: junior people need explicit permission and structure to do some things the slow way, on purpose, with someone senior reviewing the reasoning, not just the output. That's a deliberate design choice for a team now, not something that happens by default the way it used to.
Staying a high performer when the tool flatters everyone
The uncomfortable finding from the customer-support study I referenced last time — Brynjolfsson, Li, and Raymond's research — was that AI assistance barely moved the needle for the most experienced agents, and in some measures nudged their quality slightly down. Skill inflation doesn't just compress the gap from below. It can erode the top from complacency.
There's a sharper, more clinical data point for this. A multicentre observational study in colonoscopy, published in The Lancet Gastroenterology & Hepatology (Budzyń et al., 2025), found that adenoma detection rates — a core measure of how many precancerous growths a doctor actually catches — dropped from 28.4% to 22.4% when experienced endoscopists went back to working without AI assistance after getting used to having it. Detection stayed stable when AI was present. The skill hadn't just failed to grow. It had measurably receded, for people who were already good at the job. It's worth noting this was an observational before/after comparison, not a randomized trial — striking, but not proof of causation on its own.
Aviation figured this out earlier than most industries: automated flight systems are excellent until the moment they aren't, so pilots are required to maintain hands-on manual flying hours and run regular simulator sessions specifically without automation, not because the automation is unreliable day to day, but because atrophy is silent until the day it matters. The equivalent for a knowledge worker is the same discipline, deliberately unfashionable as it sounds: protect regular, unassisted reps of the core judgment calls in your role, not as nostalgia, but as maintenance.
None of this is really about resisting the tool. It's about being honest that the floor rising and your own ceiling holding are two different things, and only one of them happens automatically.
Where in your own role have you quietly stopped doing something the hard way — and is that a skill you can afford to let go of, or one you're accidentally letting atrophy?
If you want to think through where that line sits for your team, I coach leaders through exactly this kind of transition — or start with my book on self-coaching.
Sources:
- Autor, D., Levy, F., & Murnane, R. (2003). "The Skill Content of Recent Technological Change." Quarterly Journal of Economics; Acemoglu, D. & Restrepo, P. — task-based automation framework.
- Brynjolfsson, E., Mitchell, T., & Rock, D. (2018). "What Can Machines Learn and What Does It Mean for Occupations and the Economy?" AEA Papers and Proceedings.
- Ericsson, K.A., Krampe, R.T., & Tesch-Römer, C. (1993). "The Role of Deliberate Practice in the Acquisition of Expert Performance." Psychological Review, 100(3).
- (2025) Review on AI-induced deskilling and "upskilling inhibition" in medical training, Artificial Intelligence Review.
- Budzyń, K. et al. (2025). "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study." The Lancet Gastroenterology & Hepatology.
- Brynjolfsson, E., Li, D., & Raymond, L. (2023, rev. 2025). "Generative AI at Work." Quarterly Journal of Economics, 140(2).
Want to explore this further?
Start with a free 30-minute discovery session — no commitment required.
Book a Discovery Call