Do doctors lose skill when AI helps? The colonoscopy study
One study suggests they might, and it cannot prove it. After AI was introduced at four centres in Poland, doctors found growths in 22.4% of the colonoscopies they did without AI, down from 28.4% before. The study is peer reviewed and was published in a Lancet journal in 2025. It compared records before and after, so other changes could explain part or all of the fall.
The study in one table
| Title | Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study |
|---|---|
| Who | Krzysztof Budzyn and 20 co-authors |
| Status | Peer reviewed. Lancet Gastroenterology and Hepatology, online August 2025 |
| Data | 1,443 colonoscopies done without AI: 795 before AI arrived, 648 after |
| Where | Four centres in Poland |
| Design | Observational. Records compared before and after, about three months each side |
| Source | doi.org/10.1016/S2468-1253(25)00133-5 |
What they compared
In a colonoscopy, a doctor looks for growths called adenomas. The share of procedures in which at least one is found is the adenoma detection rate. It is a standard measure of how well the procedure is done.
At these centres, AI that helps spot growths was brought in. After that, some colonoscopies were done with AI and some without. The researchers looked only at the ones done without AI, and compared them with colonoscopies from before AI arrived. The question was simple: once doctors are used to the AI, how do they do when it is not there?
What they found
| Period | Colonoscopies without AI | With an adenoma found | Detection rate |
|---|---|---|---|
| Before AI was introduced | 795 | 226 | 28.4% |
| After AI was introduced | 648 | 145 | 22.4% |
That is a fall of 6.0 percentage points. The range the authors give for the true fall is 1.6 to 10.5 points (p = 0.0089), so it is unlikely to be chance alone. The doctors in the study were experienced. There were 19 doctors, with 27.6 years of experience on average. The journal published a correction in November 2025. We have not been able to read what it changed.
Why this cannot prove cause
This was not an experiment. Nobody decided at random which doctors would get AI and which would not. The study looked back at records from two periods and compared them.
- The patients may have differed. If the people examined in the later months were at lower risk, fewer growths would be found whatever the doctors' skill.
- Other things may have changed. Workload, staffing or which cases were sent to the AI rooms could all shift between periods.
- The window was short. About three months on each side.
Independent experts who commented when the paper came out made the same point: other factors could explain the result. The authors themselves describe a "risk" of lost skill in the title, not a proven loss.
What it did not show
- Not that AI harms patients. The study did not look at procedures done with AI. AI-assisted colonoscopy may find more growths, not fewer.
- Not a permanent change. Nobody tested whether the rate recovers.
- Not all doctors or all of medicine. Four centres, one country, one procedure.
- Not what happened in the doctors' minds. The data is a detection rate. Whether doctors looked less carefully, trusted a missing prompt, or something else, the records cannot say.
A fair one-line summary. "After AI was introduced, doctors' unaided detection rate fell from 28.4% to 22.4% in one observational study. It cannot show that the AI was the cause."
Why people outside medicine noticed
Most studies of AI and skill use students or volunteers on short tasks. This one is about trained professionals doing real work, over months. If the result holds up, it would mean that years of practice do not fully protect a skill from help that is always there. That is a big "if", and it needs a study designed to test it.
The studies that can show cause are experiments, and they tend to be about learning. See the trial with about 1,000 students and the experiments on giving up.
What to do with this
If a skill matters to you, keep doing some of it without help. Not all of it, and not out of fear. A regular unaided attempt is how you find out whether you can still do the thing alone.
Questions
Does AI make doctors worse at their job?
That has not been shown. One observational study found a lower detection rate in procedures done without AI after AI had been introduced. Other explanations are possible.
What is an observational study?
One where researchers watch what happens without deciding who gets what. It can find patterns. It cannot rule out other causes the way a randomised trial can. The glossary has more.
Should hospitals stop using AI for colonoscopy?
The study does not say that. It did not measure results with AI switched on. It raises a question about what happens when the AI is switched off.
How big was the fall?
From 28.4% (226 of 795) to 22.4% (145 of 648), a fall of 6.0 percentage points.