DeskillingWhat AI Is Doing to Expertise, and Why Braverman Saw It Coming

Between September 2021 and March 2022, nineteen endoscopists at four centres in Poland took part in a study that was not designed to find what it found. Each of them had performed more than two thousand colonoscopies. Towards the end of 2021 their units introduced an artificial intelligence system that watches the camera feed and flags polyps the eye might miss, and from then on each procedure was assigned at random to be done with the system or without it. The researchers wanted to know how the doctors did on the days the machine was off.

Before the system arrived, the endoscopists working unaided found adenomas in 28.4 per cent of procedures. After three months of regular use, working unaided again, they found them in 22.4 per cent. A fifth of their detection rate had gone. The study was published in The Lancet Gastroenterology and Hepatology in August 2025, and a commentator called it the first real-world clinical evidence for the phenomenon of deskilling. Nobody had been replaced and nobody’s job had changed. The doctors had spent a few months beside a machine that was better than them at one part of their work, and the part had started to leave them.

Deskilling is an old word, and it is about to be used a great deal. Most of what gets written about artificial intelligence and work asks which jobs will go. The Polish study asks what happens to the people who stay.

What Deskilling Means: Braverman and the Separation of Conception from Execution

The thought is older than the word. Adam Smith opened The Wealth of Nations with the pin factory and the gains that come from dividing labour into its simplest operations, then returned to the subject in Book V with a warning that admirers of the pin factory tend to leave out. A man whose whole life is spent performing a few simple operations has no occasion to exert his understanding, and he generally becomes as stupid and ignorant as it is possible for a human creature to become. Smith thought this was the natural consequence of the division of labour and that government would have to do something about it. He did not think it was anybody’s intention.

Harry Braverman thought it was. Braverman had spent his working life in metal trades and shipyards before he became a publisher, and in 1974 he brought out Labor and Monopoly Capital, subtitled The Degradation of Work in the Twentieth Century. Its argument was that deskilling was what a certain kind of management was for. He read Frederick Taylor’s scientific management as three principles: separate the labour process from the skills of the workers who perform it; separate conception from execution, so that the thinking happens in one place and the doing in another; and use the resulting monopoly over knowledge to control each step of the work. Taylor had said as much himself. The cost of production, he wrote in Shop Management in 1903, is lowered by separating the work of planning and the brain work as much as possible from the manual labour.

Braverman’s point was that skill is a kind of power. A craftsman who knows how the job is done can decide how it is done, can withhold that knowledge, and can set a price for it. An organisation that has extracted the knowledge from the craftsman and lodged it in a procedure, a machine or a department no longer needs him, and can hire someone cheaper and more obedient in his place. Deskilling, in this reading, is a transfer of knowledge from the person who performs the work to the institution that owns it, and the transfer of knowledge is a transfer of control.

The book started an argument that ran for twenty years. Critics pointed out that Braverman had written as if workers never fought back, that new technologies created skills as well as destroying them, and that when researchers measured skill across whole economies the picture was mixed: some occupations degraded, others upgraded, the net effect depending on the decade and the industry. Paul Attewell’s 1987 review of the deskilling controversy left the grand thesis badly dented. Shoshana Zuboff, watching computers arrive in pulp mills and offices through the 1980s, found that the same technology could be used either to automate work, taking the knowledge out of the operator’s hands, or to informate it, giving the operator more knowledge of the process than anyone had ever had. Which of the two happened was a decision, and it was made by whoever controlled the technology.

The critics did not dislodge the central observation. Deskilling is not something technology does on its own. It is something organisations do with technology, and they do it in the direction that power already runs.

How Automation Deskills: Bainbridge’s Irony and the Children of the Magenta

In 1983 Lisanne Bainbridge, a psychologist working on the design of control rooms, published a short paper in Automatica called Ironies of Automation. Engineers still pass it around four decades later, because it describes a trap that every wave of automation since has walked into.

The designer of an automated system, Bainbridge observed, tends to regard the human operator as the unreliable part, and automates everything that can be automated in order to get the human out of the loop. What cannot be automated is left to the operator, who is now asked to do two things: monitor the system, and take over when it fails. The conditions the automation has created make both impossible to do well. Monitoring is the task humans do worst, since nobody can hold their attention on a display that does nothing for hours. Taking over requires exactly the skill the operator has stopped using. Physical skills deteriorate when they are not practised, she wrote, and so does the operator’s working knowledge of the process, which was built through doing and is not maintained by watching. The operator is then handed the controls at the one moment when the system has gone wrong, the most demanding moment there is, and is expected to do by hand what they have not done by hand in years.

Aviation learned this expensively. In 1997 an American Airlines captain named Warren Vanderburgh gave a training talk in which he called a generation of pilots children of the magenta, after the colour of the flight-management line on their screens, and warned that their dependence on automation would hurt them on the day it switched off. On the night of 31 May 2009 an Air France Airbus flying from Rio de Janeiro to Paris lost its airspeed data when a probe iced over. The autopilot disengaged and handed the aircraft to its crew. One pilot pulled the nose up, the aircraft stalled, and the investigation found that in the four minutes that followed the crew never identified the stall they were in. Two hundred and twenty-eight people died. The captain had logged 346 hours in the previous six months. About four of them had been spent flying the aeroplane by hand.

Flying is safer than it has ever been, and automation is most of the reason. What the accident shows is what automation does to the people left inside it, and how little the organisation around them can see. The pilots’ skill did not vanish in an afternoon. It eroded over years, on flights that landed safely, in an airline with a good safety record, under a measurement regime that recorded outcomes rather than capacities. The erosion was invisible because the system was working. It became visible on the one night the system was not.

The Polish endoscopists were in the same position. On the days the machine was on, their detection rates were fine. The deskilling showed only because the researchers had arranged for days when it was off.

AI Deskilling at Work: Judgment, Verification and the Jagged Frontier

What is new about the present wave is who is standing in the control room. Braverman wrote about machinists and clerks. Bainbridge wrote about process operators. The people now spending their days beside a system that is better than them at part of their work are analysts and lawyers, doctors and developers, the conceptual class on whose behalf conception was separated from execution in the first place. The separation has reached conception.

In 2023 researchers from Harvard, Wharton, MIT and the Boston Consulting Group gave 758 BCG consultants a set of tasks, some with access to GPT-4 and some without. On the tasks the model handled well, the consultants with AI were faster and their work was rated markedly higher. On a task the researchers had chosen because the model would get it wrong, the consultants with AI were about nineteen percentage points less likely to reach the right answer than the consultants without it. The authors borrowed a phrase from an earlier experiment with recruiters: the consultants had fallen asleep at the wheel. They had the model’s confident wrong answer in front of them, and they took it.

A survey of 319 knowledge workers by researchers at Microsoft and Carnegie Mellon, presented in 2025, found the pattern in ordinary use. The more confidence people had in the AI, the less critical thinking they applied to its output; the more confidence they had in themselves, the more they applied. The authors described a shift in the work itself, from gathering information to verifying it, from solving problems to integrating answers, from doing the task to stewarding it, and they worried in passing about the deterioration of cognitive faculties that ought to be preserved.

Stewardship is Bainbridge’s monitoring under a better name. And verification, which every responsible guide to using these tools now recommends, has a problem the guides leave out. The ability to verify a piece of work is not separate from the ability to do it. It is the same skill run backwards. The endoscopist who can check the machine’s call is the one whose own eye was trained on two thousand unaided procedures; the reviewer who can see that a clause will fail is the one who has drafted the failing version herself. Verification is the residue of execution. An organisation that moves its people from doing to checking is drawing down a skill that checking does not replenish, and the balance declines the way the Polish detection rates declined: gradually, while the outputs look fine.

The decline is steepest where the skill was never built. The tasks AI takes first are the tasks that used to go to people at the start of their careers, and as AI and the Future of Work traces, those tasks were never only output. They were how a junior became a senior. The first draft a partner marked up, the reconciliation that did not reconcile, the search that turned up the paper nobody expected: this was the apprenticeship, unofficial and unfunded, through which judgment was formed. Remove it and the organisation keeps its seniors, for now. It has lost the means of making more.

Here the vocabulary of upskilling arrives. Prompting is a skill. Reviewing model output is a skill. Knowing which tasks to hand over and which to keep is a skill. Each of them, though, is a skill in operating the thing that now holds the skill, the competence of Taylor’s planning department handed back to the shop floor as a consolation. An organisation that retrains its analysts in prompt design has completed the separation of conception from execution and filed the completion under development.

At the same time it is sorting its people. The ones who have taken to the tools are described as the future, and the ones who have not are described, in the language of a thousand transformation decks, as needing support. The sorting is a condition of belonging and people can feel it. What the Microsoft findings suggest is that the sorting runs the wrong way. The people most comfortable handing the work to the model are the ones applying the least judgment to what comes back. The ones who insist on doing it themselves first, who are awkward about the tools and starting to look like the past, are the ones keeping alive the skill the organisation will need on the night the system fails. The marker of belonging and the marker of competence have come apart and are pointing in opposite directions, and the organisation is rewarding the first.

Why Organisations Cannot See Deskilling: Meaning and Measurement

Zuboff’s distinction has not expired. The same model that hollows a role can deepen it. The difference lies in how the work is arranged: whether people still do the thing unaided often enough to keep the skill, whether the tool shows its reasoning or only delivers a verdict, whether anyone measures capacity as well as throughput. These are decisions, and they are being made mostly by default, by the people who buy and deploy the tools rather than the people who use them, in favour of whatever the dashboard counts. Throughput is what the tools are bought to raise, so throughput is what the deployment serves, and automate wins over informate without anyone having chosen it.

That is Braverman’s observation in its current form, with one difference he could not have foreseen. The knowledge extracted from the worker is no longer lodged in a planning department down the corridor. It sits in a model owned by a vendor on another continent, under a licence that can be repriced, inside weights that nobody in the organisation can inspect. Taylor’s planning department at least worked for the same firm. The organisation that has deskilled its professionals has not kept the knowledge it took from them. It has rented it.

The people on the receiving end feel the change first as a question about meaning. Michael Polanyi observed that we know more than we can tell, and professional skill is mostly of that kind: the pattern the radiologist sees before she can say what it is, the wrongness an auditor feels in a set of accounts. That tacit knowledge was built over years of doing, and for most professionals it was the substance of their identity, what they meant when they said they were good at their job. When the doing moves to the machine the knowledge does not go at once, but it stops being renewed, and the person begins to suspect that what they are good at is now a residue. The Polish doctors, on the days the machine was off, were still endoscopists. They were less good at it than they had been, and nobody had told them. The AI Transition Diagnostic puts the same tension as a pair of statements to score side by side: that AI is making work more meaningful by removing tedious tasks, and that some of the tasks it removes were how people developed expertise in the first place. Organisations tend to agree with both.

This produces a split that most organisations cannot see, because their instruments were built to look elsewhere. The output of the work holds up or improves. The capacity of the people doing it declines. The staff survey records engagement, the dashboard records throughput, and neither has a line for whether the people could still do this without the tool. The unknowing is built into the measurement. The Polish study shows how deliberately you have to arrange things to see past it: the researchers found the deskilling only because they switched the machine off at random, and no organisation does that to its own operations. The erosion stays hidden until the system fails, or the vendor changes its terms, or the day someone senior leaves and it emerges that nobody below her ever learned to do the thing she did.

Bainbridge’s irony was that the more sophisticated the automation, the more the whole arrangement depends on a highly skilled human being available at the moment of failure, and the more certainly the arrangement has ensured that no such human exists. She was writing about one operator at one panel. The present version is a profession.

The Polish researchers were careful to say their study was observational and that other things might explain the fall. They also said the result was concerning, and that the adoption of artificial intelligence in medicine was spreading fast. What they did not say, because it was not their subject, is that the endoscopists had no idea it was happening to them. They felt as skilled as ever. The machine was on, the polyps were being found, and the loss was recorded nowhere any of them could see. Every organisation now installing the same kind of system is almost certainly deskilling somebody, in some role, in ways it has chosen not to measure. Whether it could tell is a harder question. Whether anyone inside it would want to be the one to say so is harder still.