AI deskilling is the loss of clinical skill that can follow routine reliance on AI assistance. Early evidence, including a 2025 colonoscopy study showing lower detection rates after AI was removed, suggests the risk is real. The practical answer is not to avoid AI but to plan for its absence: protected unassisted practice, outage drills and monitoring of human performance.
In August 2025, The Lancet Gastroenterology & Hepatology published a study that made gastroenterologists uncomfortable. Researchers in Poland had introduced AI-assisted polyp detection into colonoscopy at endoscopy centres taking part in a trial. As a side question, they looked at how well the same experienced endoscopists performed in colonoscopies done without the AI, before and after the tool arrived.
Before AI was introduced, their adenoma detection rate in standard colonoscopies was 28.4%. After several months of routine AI use, it was 22.4% when the AI was not switched on. A six-point absolute drop in a quality measure closely tied to whether patients later develop colorectal cancer.
The study was observational, and there are other possible explanations. But it was the first clinical data to put a number on a fear that had been theoretical: clinicians who work with an AI assistant may get worse at the task when the assistant is gone.
This piece argues something slightly different from the usual warnings. Deskilling is not mainly a reason to slow adoption. It is a reason to treat AI like any other critical system that will sometimes fail, and to design training, rotas and quality monitoring around the day it does.
What is AI deskilling in medicine?
Deskilling is the erosion of a skill through lack of use. In medicine it usually takes three forms.
The first is perceptual. A radiologist or endoscopist who routinely has an algorithm draw a box around the abnormality may search the image less thoroughly, and over time the habit of systematic search fades. This is what the colonoscopy finding suggests.
The second is cognitive. A clinician who routinely receives a machine-generated differential diagnosis, dose recommendation or summary may stop building the reasoning from scratch. The skill of holding the whole case in your head weakens.
The third, sometimes called never-skilling, affects trainees. A resident who has always had AI assistance may never build the unassisted skill in the first place. We have written separately about why residents should learn to work without AI first. This piece focuses on the first two, which affect clinicians already qualified and practicing.
Deskilling is related to automation bias, the tendency to over-trust machine outputs, but it is not the same thing. Automation bias is an error made while the tool is running. Deskilling is weakness that shows when the tool is not.
Is there evidence that AI makes doctors worse?
The evidence is emerging rather than settled, and it is worth being precise about what it shows.
The colonoscopy study, led by Krzysztof Budzyń and colleagues, is the strongest clinical signal so far. Its limitations are real: it was observational, conducted within a trial not designed to answer this question, and changes in patient mix, workload or the endoscopists’ attention could contribute. It measured performance only in the months after AI introduction, so we do not know whether skills recover or keep sliding.
What makes it credible is that it fits a much older body of evidence from other fields. Aviation is the best documented. In 2013 the US Federal Aviation Administration issued a safety alert, SAFO 13002, noting that analysis of flight data had found an increase in manual handling errors, and warning that continuous use of autopilot and autothrust could degrade pilots’ ability to recover an aircraft from an undesired state. It encouraged airlines to give pilots opportunities to fly manually. The FAA reinforced the message with a follow-up alert in 2017.
Clinicians themselves are worried. The American Medical Association’s 2026 physician survey, fielded in January and February 2026, found that 88% of respondents were at least mildly concerned about skill loss from AI, with concern higher among physicians in the first decade of their careers.
So the honest summary is: one important clinical study, a strong analogy from aviation, widespread clinician concern, and very little data on the dose, duration or reversibility of the effect.
Why does deskilling matter if the AI is better?
A reasonable objection: if the AI-assisted clinician outperforms the unassisted one, who cares whether the unassisted skill declines?
There are three reasons it matters.
AI systems fail. Software updates change behavior, integrations break, networks go down, and hospitals suffer cyberattacks that take clinical systems offline for weeks. The skill a clinician needs during an outage is the unassisted one, and outages tend to arrive during crises when workload is worst.
AI systems also fail quietly. A model can drift as patient populations, equipment or coding practices change. The main safeguard against silent drift is a clinician who notices that the tool’s outputs no longer match what they see. A clinician who has stopped looking independently cannot provide that check.
And AI is rarely deployed everywhere at once. A doctor who trains with AI at a teaching hospital and then works at a district hospital without it, or rotates to a night shift where the tool is not licensed, carries only the skill they have maintained.
There is a fourth, subtler reason. Many AI tools are validated against the performance of the clinicians who use them. If those clinicians’ unassisted performance declines, the benchmark moves with it, and a model that looks superior may simply be outperforming a weakened comparator. Tracking unassisted human performance is therefore not only a training question. It is part of knowing whether the tool is actually adding what the vendor claims, and whether the department could safely switch suppliers or survive a contract ending.
The combined AI-plus-human system is only as resilient as the human part when the AI part is missing.
A worked example: planning for the outage
Consider a hypothetical endoscopy unit with ten endoscopists that introduces AI polyp detection across all rooms.
The unit tracks adenoma detection rate for each endoscopist, as quality programs recommend. After adoption, overall rates rise, as randomized trials of AI-assisted colonoscopy have generally found. The dashboard looks excellent.
Now the unit adds two simple measures. First, one list a month for each endoscopist is performed with the AI switched off, and detection rates on those lists are tracked separately. Second, the unit agrees an outage protocol: if the AI system is down, lists continue with a checklist emphasizing systematic withdrawal technique and adequate withdrawal time.
Six months in, the unassisted lists show that three endoscopists’ detection rates have fallen meaningfully below their pre-AI baseline. Those three get feedback and targeted practice. The unit also learns that its overall AI-assisted performance, while higher, depends more on the tool than it thought, which becomes a factor in its contingency planning and contract negotiations with the vendor.
None of this is expensive. It needs a decision to measure unassisted performance, which most units currently do not make.
What is the strongest case against worrying about deskilling?
The counterargument is serious and deserves a fair hearing.
Every technology in medicine has deskilled clinicians in something. Few doctors today can reliably diagnose by percussion what earlier generations could, because imaging replaced it. Few calculate drug doses by hand. Society generally accepts that, because the new method is better and reliably available. If AI detection is consistently better and becomes as available as a CT scanner, maintaining the old skill may be a waste of scarce training time.
There is also a risk that deskilling fears become an excuse for resisting useful tools. A patient does not benefit from an endoscopist’s pride in unassisted detection if the assisted rate is higher.
The response is that the analogy holds only once the technology is reliable, universal and well monitored. Imaging replaced percussion over decades, with robust infrastructure and backup. Clinical AI in 2026 is none of those things: it is patchily deployed, subject to drift, often a single vendor’s product, and vulnerable to outages. Until it is as dependable as the scanner, clinicians need to keep the fallback skill, and organizations need to know whether they have it.
What this means for health systems, educators and regulators
For hospital leaders, treat AI as a critical system with a failure mode. That means an outage plan for every clinical AI tool, periodic unassisted practice for clinicians who use it heavily, and measurement of human performance separately from the combined system. The airline industry learned this lesson through accidents, and aviation’s safety culture is only useful to health systems if they borrow its routines rather than its slogans.
For clinicians, the practical habit is simple: form your own view before you look at the AI’s output where the workflow allows it, and occasionally work unassisted on purpose. Your independent read is both a skill maintenance exercise and the best safety check on the tool.
For educators and professional bodies, the question is how continuing competence is assessed. If revalidation and continuing education measure attendance rather than performance, as we have argued they often do, they will not detect deskilling. Periodic unassisted assessment in AI-heavy specialties is worth considering.
For regulators and vendors, post-market monitoring should include how clinician performance changes over time, not only how the model performs. A tool that improves results for two years while quietly eroding the skills needed on the day it fails is not as safe as its trial suggests.
Frequently asked questions
Can using AI make doctors less skilled?
It can. A 2025 study in The Lancet Gastroenterology & Hepatology found that experienced endoscopists’ adenoma detection rate in non-AI colonoscopies fell from 28.4% to 22.4% after AI was introduced. The study was observational and needs confirmation, but it matches long-standing evidence from aviation that heavy automation erodes manual skills.
What is the difference between deskilling and never-skilling?
Deskilling is the loss of a skill a clinician once had, through reliance on assistance. Never-skilling happens when trainees learn with AI from the start and never develop the unassisted skill at all. Both matter, but they need different responses: skill maintenance for qualified clinicians, and structured unassisted training for students and residents.
How can hospitals prevent AI deskilling?
Hospitals can schedule periodic unassisted practice, track clinicians’ performance with and without AI separately, write outage protocols for each clinical AI tool, and encourage clinicians to form an independent judgment before viewing AI output. These steps are inexpensive compared with discovering a skills gap during a system failure.
Do doctors worry about losing skills to AI?
Yes. In the American Medical Association’s 2026 physician survey, 88% of respondents said they were very, somewhat or mildly concerned about skill loss from AI. Concern was higher among physicians with fewer than ten years in practice, the group most likely to have trained alongside digital decision support.
Should doctors avoid using AI to protect their skills?
No. Where AI improves patient outcomes, avoiding it would harm patients. The better approach is to use AI while keeping independent skills in practice, for example by forming your own assessment first and doing some work unassisted. Patients with questions about how AI is used in their care should ask their clinician.
The question is not whether AI will make clinicians better. It is whether they will still be good on the day it is switched off, and almost nobody is measuring that.
Has your team ever worked a full shift with its AI tools down, and what did you learn about what you had forgotten?
