In late 2021, four endoscopy centres in Poland began using AI to help spot polyps during colonoscopy. The centres were part of a trial, so some procedures were still done without the software. That gave researchers an unusual view: what happened to doctors’ unaided performance once they got used to the machine?
It dropped. In a study published in The Lancet Gastroenterology & Hepatology in 2025, the adenoma detection rate for colonoscopies done without AI fell from 28.4% in the three months before the tool arrived to 22.4% in the three months after. These were experienced endoscopists, each with more than 2,000 procedures behind them. Three months of help was enough to blunt a skill that took years to build.
The study was observational and small, and it deserves replication before anyone rewrites a curriculum around it. But it points at a problem medical education has barely started to name. We are debating how to teach trainees to use AI. The harder question is how to make sure they can still practise when the AI is wrong, absent or quietly degrading.
Why a six-point drop matters
A fall in adenoma detection rate can sound like an abstract quality metric. It is not. The detection rate is used as a quality benchmark precisely because it tracks what happens to patients later. A large study by Douglas Corley and colleagues in the New England Journal of Medicine in 2014, drawing on Kaiser Permanente data, found that each one-percentage-point increase in an endoscopist’s adenoma detection rate was associated with a 3% lower risk of interval colorectal cancer, the cancers diagnosed after a supposedly clear colonoscopy.
Put the two findings side by side, with every caveat about mixing different populations, and the Polish result stops looking trivial. A six-point fall in unaided detection is the kind of difference the Kaiser data associates with a meaningfully higher rate of missed cancers. If the AI is always switched on, perhaps it does not matter. The question is whether it will always be switched on.
How the skill quietly erodes
The mechanism is not mysterious. Detection is partly a habit of attention: where the eye goes, how long it lingers on a fold, how carefully the scope is withdrawn. When a system reliably draws a box around suspicious tissue, the clinician’s attention reorganises around the box. Why scan the edges of the screen with full vigilance when the software will flag anything worth seeing?
That reorganisation is efficient while the tool is running. It becomes a liability when the tool is off, because the old scanning habit has partly lapsed. Psychologists who study automation have described this for decades in aviation and industrial control. Medicine is simply the latest field to learn it, with the added complication that its errors are often invisible for years.
For a trainee, the problem is sharper still. An experienced endoscopist had a scanning habit to lose. A resident who learns with the box always present may never build it. Their competence is not eroded; it is formed around the tool, and it may not exist without it.
Adoption has already outrun training
The tools are not waiting for curricula. The American Medical Association’s survey of physicians found that 66% reported using some form of AI in their practice in 2024, up from 38% the year before. Most of that use is administrative: visit notes, discharge instructions, billing codes. Some of it is diagnostic support.
Residents are absorbing these tools the way every previous generation absorbed new technology: informally, from the person one year ahead, during a night shift, with no one checking whether they understand what the output means. Most programmes have added a lecture or an online module on “AI in medicine.” Almost none have changed how competence is assessed.
That gap matters more for trainees than for the experienced endoscopists in the Polish study. An expert who loses some edge can, in principle, recover it. A trainee who never builds the skill in the first place has nothing to fall back on.
The wrong lesson from aviation
The usual comparison is aviation, where autopilot is routine and pilots still train for manual flight. It is a good comparison, but the lesson is often drawn backwards.
Aviation did not solve automation dependence by teaching pilots more about autopilot. It solved it, imperfectly, by requiring pilots to keep flying by hand and by drilling what to do when automation fails. The skill that matters most is not operating the tool. It is noticing when the tool is steering you wrong, and taking over cleanly.
Medicine’s version would look something like this:
- Sequenced exposure. Trainees reach a defined level of unaided competence in a task, such as reading a chest film or detecting polyps, before they routinely use AI for it.
- Deliberate unassisted practice. Regular sessions where the tool is switched off, with performance measured, the way surgical programmes log cases.
- Failure drills. Cases where the AI output is deliberately wrong, so trainees learn what an error looks like and how confidently the system presents it.
- Assessment that separates the two. Examinations that test clinical reasoning with and without assistance, and report both.
None of this is exotic. It is simply taking the risk of deskilling as seriously as we take the risk of error.
The strongest counterargument
There is a serious case on the other side, and it deserves a hearing. Nobody insists that accountants add columns by hand, or that radiologists develop film. If AI detection becomes as dependable as electricity, training people to work without it may be wasted effort, time that could be spent on the judgement the machine cannot supply.
That argument partly holds. Not every skill needs protecting. Transcribing notes by hand is not a clinical competence anyone will miss, and insisting on it would be nostalgia dressed up as rigour. Programmes should be selective, protecting the skills that are safety-critical and hard to recover, rather than treating every AI-assisted task as a threat.
Where the argument breaks down is on the phrase “as dependable as electricity”. Clinical AI tools are not infrastructure in that sense. Systems go down during outages and cyberattacks. Models trained on one population perform differently on another. Software is updated, and its behaviour shifts. Doctors move between hospitals with different tools, or to settings with none. And when the AI misses something, the only safeguard is a clinician capable of seeing what it missed. A doctor who cannot perform the task unaided cannot meaningfully check the machine that performs it.
What each party can do now
Residency programme directors have the most direct lever. They can decide, this year, which core skills in their specialty require documented unaided competence before routine AI use, and build a small amount of tool-off practice into rotations. It costs some throughput. It is cheaper than discovering the gap during an outage.
Accreditation bodies can make that expectation explicit. As long as unaided competence is optional, the productivity pressure described below will win. Once it is a requirement, programmes will find the time, as they do for minimum case numbers.
Hospital leaders buying AI tools can ask vendors two questions that are rarely asked: how does clinician performance change when the tool is removed, and what is the plan for operating when it is unavailable? A vendor who has never measured the first has not studied its product’s full effect on the workforce.
And the researchers have their own job. The Polish study needs replication, in other procedures and other specialties, with trainees as well as experts. Over the next year or two the most useful evidence will not be another accuracy study showing that AI finds more lesions. It will be studies that measure what clinicians can do after sustained exposure, with the tool switched off.
The incentive nobody wants to say out loud
Here is the uncomfortable part. Every force acting on a hospital pushes against this kind of training.
AI tools are bought to save time. A department that pays for an ambient scribe or a detection algorithm expects throughput to rise. Asking residents to do some cases the slow way, deliberately, looks like waste on a productivity dashboard. Supervising consultants, themselves stretched, have little reason to insist on it.
Vendors have no incentive either. Their pitch is that the tool makes clinicians better. Evidence that it makes them worse when it is removed is, commercially, an awkward finding. It is telling that the Polish data came from a trial with a built-in comparison group, not from post-market surveillance.
And trainees themselves may resist. If the software writes a cleaner note and flags a lesion faster, why struggle without it? The honest answer is that one day the system will be down, or wrong in a way nobody has seen before, or unavailable in the hospital they move to. That answer is abstract. The saved hour is concrete.
So unassisted competence will not survive on good intentions. It needs to be written into accreditation requirements, the way minimum case numbers already are.
What to teach about the tools themselves
This is not an argument for keeping AI out of training. Trainees need real fluency, and most current teaching is too shallow to provide it. A useful curriculum would cover:
- how a given model was validated, on which population, and where performance is likely to drop
- what automation bias looks like in practice, including the tendency to stop looking once the software has found something
- how to document when a clinician overrides an AI recommendation, and why
- who is accountable when an AI-assisted decision goes wrong
The last point is the one trainees ask about most, and the one faculty are least able to answer. Liability frameworks are still unsettled in most countries. Pretending otherwise does residents no favours.
The standard worth setting
The goal is not a doctor who can use AI. That bar is low and falling, because the tools are getting easier. The goal is a doctor who is better with AI than without it, and still safe without it.
There is a simple way to make that concrete. For any AI-assisted task a trainee is certified in, the programme should be able to report two numbers: performance with the tool and performance without it. If the first is high and the second is unknown, the programme does not know whether it has produced a competent doctor or a competent operator of one product. If the second is falling over the course of training, something is going wrong, however good the first looks.
Recording both numbers is not expensive. It requires a few structured sessions a year, some cases read or procedures done with the software off, and a willingness to look at the results. The harder part is cultural: accepting that a trainee who performs brilliantly with assistance might still not be ready.
Those are two separate competencies. Right now we are measuring neither well, and we are training almost entirely for the first. The Polish study is a warning shot from the most experienced end of the workforce. The trainees coming through now will have spent their whole careers with the machine on.
Does your training programme require any AI-free practice, or measure what residents can do without the tools? Tell us how it is being handled where you work.
