Is deskilling real, or a rescaling of what counts as skill?

Both descriptions have been measured, and they describe different people. Among professionals who already hold a skill, the loss has been measured directly: nineteen Polish endoscopists averaging 27.6 years of experience each detected adenomas in 28.4 per cent of their unassisted colonoscopies before AI arrived in their departments, and 22.4 per cent of their unassisted colonoscopies afterwards. Among people who do not yet hold the skill, the gap between the strongest and weakest performers has been measured closing. So the claim that there is no deskilling fails against a measurement. The claim that what counts as a valuable skill is being rescaled survives, but only as a claim about direction, which has to be checked role by role before it means anything.

Nineteen endoscopists, doing the same procedure without the tool#

The strong form of the claim is falsifiable, and something has already falsified it. Across four Polish endoscopy centres, 1,443 colonoscopies performed without AI assistance were compared before and after computer-aided detection came into routine use in the same departments. The endoscopists were the same people, with 8 to 39 years of experience behind them. Their unassisted adenoma detection rate fell six percentage points, from 28.4 to 22.4 per cent, and the fall was statistically significant.

That is removal, measured, in experts, on the work they were trained for, within months. It is not a revaluation of which skills matter. The skill in question still mattered to every patient on the list, and the doctors were worse at it.

Three limits travel with the finding wherever it goes. It is observational, so other changes over the period cannot be ruled out. It is one procedure in one country. And detection rate is a proxy for skill rather than skill itself. It remains the single strongest direct measurement in the field, which is a statement about how thin the field is as much as about how good the study is.

The gap that closed in a Japanese taxi fleet#

The rescaling argument has evidence behind it too, and this research holds some of it deliberately. When a Japanese taxi fleet rolled out an AI demand-prediction system, the productivity gains went almost entirely to the low-skilled drivers, narrowing the gap between best and worst by 14 per cent. The same pattern appears in customer support and in software. Something is being redistributed, and the direction is towards the people who had least.

Anyone who wants a tidy story about AI hollowing out capability has to explain that result, and this research keeps it in view for that reason. Compression is real. The open question is what compression does over twenty years to the person who was compressed upward.

An exoskeleton works where the ground has been mapped#

The same panel put the mechanism more vividly than anyone else has. From the transcript:

So that 20-year-old, it's not like entry-level jobs are going, we're just redefining what is entry-level. That's really what we're going through. So the 20-year-old will have access to a level of skill and expertise with augmentation, it's a bit like having an exoskeleton from the movie Avatar, right? You will have the ability to be super-powered, to do something reliably that would otherwise take you 10 or 15 years to do.

The word carrying the weight there is "reliably". 758 BCG consultants were given tasks inside and just outside GPT-4's competence. Inside, the assisted consultants were dramatically better and faster, which is the exoskeleton working. Outside, they performed worse than consultants given no AI at all, which is the exoskeleton walking them off a cliff they could not see.

This is where the argument turns, and it turns on something the panel did not get to. The exoskeleton has an edge, and finding the edge is itself the capability that takes 10 or 15 years to build. The 20-year-old gets the reach and does not get the map. So the augmentation delivers senior output and withholds the one thing seniority was for, which is knowing when the output is wrong. This research calls the result synthetic seniority. A rescaling argument has to account for it before the rescaling can be called good news.

Redefining entry level upward is the problem, not the reassurance#

The second half of the rescaling argument holds up better than the first, and it holds up for a reason that should worry everyone. Entry-level jobs are not visibly going. ADP payroll data covering millions of US workers shows no economy-wide displacement, and the divergence that does appear among 22 to 25 year olds in exposed occupations runs through reduced hiring rather than through people being let go. In a YouGov survey of 1,250 employed US workers, about 3 per cent said they had lost a job to AI since 2023, against roughly 6 per cent who held a job that did not exist before it, though everyone in that sample was employed when asked, so it counts survivors only.

A second dataset and a second method find the same slowdown in hiring: the monthly job-finding rate for 22 to 25 year olds entering the most exposed occupations fell about 14 per cent against 2022 in those same occupations, on a base rate near 2 per cent per month, and the authors call the result just barely statistically significant themselves.

So the jobs stay and the ladder changes shape. An entry role that has been redefined upward asks a graduate for the judgement that used to be built by doing the tasks the redefinition removed. That is the missing rungs problem stated in the rescaling argument's own vocabulary. Redefinition is the mechanism, not the consolation, and calling it a redefinition tells you nothing about whether anyone reaches the top of the ladder in fifteen years.

Which tasks went, and what that does to a wage#

Rescaling sounds neutral and is not. Four decades of task data across 303 US occupations give the test. Where automation removed the less expert tasks from a job, wages rose and employment fell. Where it removed the expert tasks, wages fell and employment rose. Same technology, opposite outcomes, and what decides between them is which tasks left.

That converts the rescaling argument from a description into a question with an answer. For any role, ask which tasks the tool took. If it took the routine ones and left the judgement, the role is appreciating and the rescaling argument is right about it. If it took the judgement and left the assembly, the role is commoditising and the word rescaling is doing public-relations work. The data behind that result ends in 2018, so it is a lens for asking the question and not a forecast of how generative AI will answer it.

Nobody has followed a cohort that started with the exoskeleton#

The honest gap sits precisely where the disagreement is. Every measurement above is of people who built a skill and then had a tool arrive. There is no study of people who never built it.

Sixteen authors writing in Nature Medicine in May 2026 named the distinction and separated three failures that are usually run together. Deskilling is the degradation of competence in people already trained. Mis-skilling is picking up faulty reasoning from uncritical use of wrong or biased output. Never-skilling is the failure to form the competence at all, when the tool substitutes for the effort that would have built it. All three terms are theirs and none is claimed here.

Their own caveat is the load-bearing part, and they state it twice: "Direct causal evidence linking AI exposure during training to competency failure in medical trainees does not exist." It is a risk model. Prevalence, severity and reversibility are all unknown, and the framework they propose is untested. Anyone citing it as proof of harm has misread it, and this page cites it only for the taxonomy.

So the answer to the panel is that both sides were arguing past the evidence. Deskilling has been measured and rescaling has been measured, in different populations, and the case that decides between them, a cohort that entered the profession with the exoskeleton on, has never been followed.

The study offered against the rescaling argument was misdescribed as well#

Recorded against interest, because the same standard applies to the other side of the room. The write-up of that panel offers as its main evidence for deskilling "Ehsan et al's radiologist study", in which "specialists who leaned on generative AI experienced an initial productivity gain, followed by a creeping erosion of skill".

The study is real and it is not that study. Its 42 participants work in radiation oncology, planning treatment rather than reading images. The system is optimisation-based, not generative. And it is qualitative: twelve months of fieldwork, 52 think-aloud sessions, 24 interviews, with capability loss self-reported and no measured skill outcome. It cannot sit beside the Polish colonoscopy result as a second measurement.

What it does give, and nothing else here gives, is occupational identity. Its authors name intuition rust, the dulling of expert judgement underneath output that still looks fine, and identity commoditisation, the loss of professional standing as practitioners describe becoming AI babysitters and button-pushers in their own practice. Both terms are theirs. They call the harms asymptomatic because the hospital's own measures registered only the 15 per cent faster planning cycles.

The question to put to your own organisation#

The argument becomes tractable the moment it stops being about deskilling in general and starts being about one role.

Take the role. List what the tool now does that a person used to do. Ask which of those tasks were the expert ones, using the test above. Then ask the question the panel never got to: if a person joined this role tomorrow and used the tool from day one, what would they be able to do unaided in five years, and how would anyone find out before it mattered. If nobody can answer the second half, the organisation has a rescaling story and no way of knowing whether it is true. That check has a name here, the capability audit, and its whole content is measuring unaided performance often enough to notice a change.

Key sources

• Budzyń, K., Roman'czyk, M., Kitala, D. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. The Lancet Gastroenterology and Hepatology, 10(10).
• Kanazawa, K., Kawaguchi, D., Shigeoka, H. and Watanabe, Y. (2022). AI, Skill, and Productivity: The Case of Taxi Drivers. NBER Working Paper 30612; Management Science, 72(2).
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier. Harvard Business School and BCG working paper.
• Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? Stanford Digital Economy Lab.
• Massenkoff, M. and McCrory, P. (2026). Labor market impacts of AI: A new measure and early evidence. Anthropic, 5 March 2026, corrected 8 March 2026.
• Autor, D. and Thompson, N. (2025). Expertise. Journal of the European Economic Association, 23(4).
• Ke, Y. et al. (2026). AI-induced never-skilling in medical education. Nature Medicine, 32(6).
• Ehsan, U., Passi, S., Saha, K., McNutt, T., Riedl, M. O. and Alcorn, S. (2026). From Future of Work to Future of Workers. CHI '26, ACM.
• Tsim, F. (2026). Still Sharp? The Human Edge in an Age of AI Assistance. BehSci Meets AI, 22 May 2026. The source of both quotations. Read at source 6 September 2026.

Related SuperSkills research#

On the term itself, what deskilling is and which professions carry the most risk. On the ladder, the missing rungs, synthetic seniority and whether AI replaces entry-level jobs. On the debt that builds while nobody measures it, capability debt. On finding out before it matters, the capability audit and assessing capability rather than output.

About this argument#

Deskilling, cognitive augmentation, never-skilling, mis-skilling, intuition rust and identity commoditisation are other people's terms and are credited above. The rescaling argument is Roop Bhadury's and is quoted from his own words at a panel he shared with Rahim Hirji on 22 May 2026. The reading offered here, that compression and degradation are both measured and apply to different populations, and that the exoskeleton withholds the ability to find its own edge, is an interpretation by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), and is marked as an interpretation and not a finding.