The strongest evidence that AI deskills people is a single colonoscopy study

What Bainbridge saw in the control room

In 1983 Lisanne Bainbridge, writing about automated process-control plants, laid out a problem that has outlived every system she was describing.¹ A designer automates every part of a task that can be done cheaply and reliably, and leaves the human the residue: watch the automation, and take over on the rare occasion it cannot cope. Those are the two hardest parts of the original job, and automation has just removed the daily practice that made anyone good at them.

Her point was mechanical. Physical skills decay when they go unused, and the knowledge behind a judgment gets slower to retrieve the less it is used. So the operator who once ran the plant by hand, and now only monitors it, is on the way to becoming one who cannot run it by hand. In her words, "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one."¹

Two further turns make it worse. Monitoring a machine that almost never fails is a task human attention is poorly built for: you cannot stay alert for hours against an event that comes once a year. And the more reliable the automation, the rarer the takeover, so the skill it demands is most decayed exactly when it is finally called on. Bainbridge's conclusion was that the most automated systems need the most operator training, not the least.¹

The argument is strong, but it is reasoning, not measurement. Bainbridge was working from what she had seen on flight decks and in control rooms. Her paper is the origin of the idea and its most careful statement, but it predates the systems now in question by four decades, so whether it holds for any one of them the paper cannot say. It is also a claim about the skill, not the job. Whether automation destroys employment is a different argument. Deskilling is about the competence of the people still doing the work.

Aviation and GPS measured which skills decay

Outside AI, in domains automated for decades, the results are more specific than the slogan. Take the claim in its most familiar form: automation ruins a pilot's stick-and-rudder skills. In 2014 Stephen Casner and colleagues put 16 airline pilots in a Boeing 747-400 simulator and varied how much of the flying the automation did.² The finding cut against the slogan. The pilots' hands were fine: manual control and instrument scanning held up even when they said they rarely practiced them. What degraded were the cognitive parts of flying, such as tracking the aircraft's position without the map display and noticing a failure, and that decline tracked how far a pilot's mind had wandered more than how little they had touched the controls.²

The clearest real-world case is Asiana Airlines Flight 214, which struck a seawall short of the San Francisco runway in 2013.³ It is often told as automation deskilling a crew into a crash. The report is more careful. The National Transportation Safety Board's probable cause is in-the-moment mismanagement: a mishandled descent, the pilot flying's unintended deactivation of the autothrottle's speed protection, inadequate airspeed monitoring, and a delayed go-around.³ Automation reliance appears there as one contributor to the monitoring failure, not as the cause. The crew let the airplane get low and slow because they trusted the autothrottle to hold the speed when, in the mode they had set, it would not.³ The board did find that regular hand-flying would have helped the crew catch the decaying airspeed, and that Asiana's policy discouraged manual flight.³

Away from aviation, the cleanest controlled evidence comes from navigation. In 2020 Louisa Dahmani and Véronique Bohbot tested 50 regular drivers on virtual navigation and rated their lifetime GPS use.⁴ Heavier GPS users had worse spatial memory when they had to find their way unaided. Thirteen came back three years later, and those who had used GPS more in the interim had declined more steeply.⁴ One detail speaks to the direction of cause. How good a navigator someone was at the start did not predict how much GPS they later used, which argues the tool drove the decline rather than weak navigators reaching for the crutch.⁴ The samples are small, so the effect size is uncertain, but the three-year arm is the before-and-after a deskilling claim rarely has.

For AI, the evidence is one colonoscopy study

Everything so far is about older automation. The question the argument is really being asked now is whether AI does the same thing, and here the measured evidence nearly runs out.

The single strongest study is about colonoscopies. Endoscopists are graded on their adenoma detection rate, the share of their colonoscopies in which they find at least one precancerous growth; a higher rate means fewer cancers missed. In a 2025 study from four Polish centers, researchers tracked that rate in ordinary colonoscopies done without AI, before and after the clinics began using an AI polyp-detection tool in their other cases.⁵ The 19 endoscopists were experienced, each with more than 2,000 colonoscopies behind them. Their unaided detection rate fell from 28.4% before the clinics adopted AI to 22.4% after, a drop of 6.0 percentage points, about a fifth of where they started (the confidence interval runs from 1.6 to 10.5 points).⁵

What makes this matter for deskilling, rather than for trusting the machine, is that it measured the doctors working without the AI: a change in what they did on their own. That is why it belongs here and not with automation bias, the separate failure of over-trusting a specific automated output in the moment.

Cautions come with it, which the study's own authors and outside reviewers press.⁵⁶ The design is observational: the clinics were not randomly assigned, so something other than AI exposure could have moved the rate.⁷ There are only 19 endoscopists, with no breakdown showing whether the drop was broad or driven by a few, and no data on simpler explanations such as how long each doctor spent looking.⁶ The baseline rate was also low, around a quarter, against a US benchmark near 35%.⁶

So the drop is real and measured, but its cause is not settled. It could be the durable skill loss Bainbridge described, or short-term complacency that would lift when the stakes felt higher. A three-month dip shows behavior changed; it does not prove a capability was lost.⁶ One independent gastroenterologist warned of "the quiet erosion of fundamental skills" while noting AI's promise.⁷

The other AI-specific measurement people cite is weaker still. A 2025 MIT preprint had 54 people write essays over four months, some using a large language model, some a search engine, some no tools, and recorded their brain activity with EEG.⁸ The model users showed the weakest neural connectivity of the three groups and struggled most to quote back essays they had just written.⁸ But the paper is not peer reviewed, and a separate critique faults its sample size and EEG analysis.⁹ Weaker connectivity measures how engaged someone was, which can just mean less effort when help is on hand. It does not show that anyone's writing skill got worse.

Nobody has measured the coding-skill claim

The version repeated most often has the least behind it: that programmers who let an AI assistant write and debug their code will lose the ability to do it themselves, and miss the assistant's mistakes. It is a direct application of Bainbridge, and it is entirely a forecast. No controlled study measures erosion of programming or debugging skill from AI assistants.

The nearest relevant number cuts the opposite way. In 2025 METR ran a randomized trial with 16 experienced open-source developers on 246 tasks in their own repositories, some with AI tools allowed and some without.¹⁰ The developers expected AI to speed them up by about 24%, and afterward believed it had. It had in fact slowed them down by about 19%.¹⁰ That is a result about productivity, and about how badly people judge their own. It says nothing about whether their skills decayed; reading it as deskilling evidence would be a mistake.

Here confidence outruns proof in both directions. The people sure that AI is quietly hollowing out a generation of programmers have no measurement to point to. Neither do the people sure it is harmless. What can honestly be said is narrower: the mechanism is well understood, its effects are documented in older automation, and for AI the measurements have barely begun.