Is AI Use Causing Endoscopists to Lose Their Skills?

Routine use of artificial intelligence (AI) may lead to a loss of skills among clinicians who perform colonoscopies, thereby affecting patient outcomes, a large observational study suggested.

“The extent and consistency of the adenoma detection rate (ADR) drop after long-term AI use were not expected,” study authors Krzysztof Budzyń, MD, and Marcin Romańczyk, MD, of the Academy of Silesia, Katowice, Poland, told Medscape Medical News. “We thought there might be a small effect, but the 6% absolute decrease — observed in several centers and among most endoscopists — points to a genuine change in behavior. This was especially notable because all participants were very experienced, with more than 2000 colonoscopies each.”

Another unexpected result, they said, “was that the decrease was stronger in centers with higher starting ADRs and in certain patient groups, such as women under 60. We had assumed experienced clinicians would be less affected, but our results show that even highly skilled practitioners can be influenced.”

The study was published online in The Lancet Gastroenterology & Hepatology.

ADR Reduced After AI Use

To assess how endoscopists who used AI regularly performed colonoscopy when AI was not in use, researchers conducted a retrospective, observational study at four endoscopy centers in Poland taking part in the ACCEPT trial.

These centers introduced AI tools for polyp detection at the end of 2021, after which colonoscopies were randomly assigned to be done with or without AI assistance.

The researchers assessed colonoscopy quality by comparing two different phases: 3 months before and 3 months after AI implementation. All diagnostic colonoscopies were included, except for those involving intensive anticoagulant use, pregnancy, or a history of colorectal resection or inflammatory bowel disease.

The primary outcome was the change in the ADR of standard, non-AI-assisted colonoscopy before and after AI exposure.

Between September 2021 and March 2022, a total of 2177 colonoscopies were conducted, including 1443 without AI use and 734 with AI. The current analysis focused on the 795 patients who underwent non-AI-assisted colonoscopy before the introduction of AI and the 648 who underwent non-AI-assisted colonoscopy after.

Participants’ median age was 61 years, and 59% were women. The colonoscopies were performed by 19 experienced endoscopists who had conducted over 2000 colonoscopies each.

The ADR of standard colonoscopy decreased significantly from 28.4% (226 of 795) before the introduction of AI to 22.4% (145 of 648) after, corresponding to a 20% relative and 6% absolute reduction in the ADR.

The ADR for AI-assisted colonoscopies was 25.3% (186 of 734).

The number of adenomas per colonoscopy (APC) in patients with at least one adenoma detected did not change significantly between the groups before and after AI exposure, with a mean of 1.91 before vs 1.92 after. Similarly, the number of mean advanced APC was comparable between the two periods (0.062 vs 0.063).

The mean advanced APC detection on standard colonoscopy in patients with at least one adenoma detected was 0.22 before AI exposure and 0.28 after AI exposure.

Colorectal cancers were detected in 6 (0.8%) of 795 colonoscopies before AI exposure and in 8 (1.2%) of 648 after AI exposure.

In multivariable logistic regression analysis, exposure to AI (odds ratio [OR], 0.69), patient’s male sex (OR, 1.78), and patient age at least 60 years (OR, 3.60) were independent factors significantly associated with ADR.

In all centers, the ADR for standard, non-AI-assisted colonoscopy was reduced after AI exposure, although the magnitude of ADR reduction varied greatly between centers, according to the authors.

“Clinicians should be aware that while AI can boost detection rates, prolonged reliance may subtly affect their performance when the technology is not available,” Budzyń and Romańczyk said. “This does not mean AI should be avoided — rather, it highlights the need for conscious engagement with the task, even when AI is assisting. Monitoring one’s own detection rates in both AI-assisted and non-AI-assisted procedures can help identify changes early.”

“Endoscopists should view AI as a collaborative partner, not a replacement for their vigilance and judgment,” they concluded. “Integrating AI effectively means using it to complement, not substitute, core observational and diagnostic skills. In short, enjoy the benefits of AI, but keep your skills sharp — your patients depend on both.”

Omer Ahmed, MD, of University College London, London, England, gives a similar message in a related editorial. The study “compels us to carefully consider the effect of AI integration into routine endoscopic practice,” he wrote. “Although AI continues to offer great promise to enhance clinical outcomes, we must also safeguard against the quiet erosion of fundamental skills required for high-quality endoscopy.”

‘Certainly a Signal’

Commenting on the study for Medscape Medical News, Rajiv Bhuta, MD, assistant professor of clinical gastroenterology and hepatology at Temple University and a gastroenterologist at Temple University Hospital, both in Philadelphia, said, “On the face of it, these findings would seem to correlate with all our lived experiences as humans. Any skill or task that we give to a machine will inherently ‘de-skill’ or weaken our ability to perform it.”

“The only way to miss a polyp is either due to lack of attention/recognition of a polyp in the field of view or a lack of fold exposure and cleansing,” said Bhuta, who was not involved in the study. “For AI to specifically de-skill polyp detection, it would mean the AI is conditioning physicians to pay less active attention during the procedure, similar to the way a driver may pay less attention in a car that has self-driving capabilities.”

That said, he noted that this is a small retrospective observational study with a short timeframe and an average of fewer than 100 colonoscopies per physician.

“My own ADR may vary by 8% or more by random chance in such a small dataset,” he said. “It’s hard to draw any real conclusions, but it is certainly a signal.”

The issue of de-skilling goes beyond gastroenterology and medicine, he noted. “We have invented millions of machines that have ‘de-skilled’ us in thousands of small ways, and mostly, we have benefited as a society. However, we’ve never had a machine that can de-skill our attention, our creativity, and our reason.”

“The question is not whether AI will de-skill us but when, where, and how do we set the boundaries of what we want a machine to do for us,” he said. “What is lost and what is gained by AI taking over these roles, and is that an acceptable trade-off?”

The study was funded by the European Commission and the Japan Society for the Promotion of Science. Budzyń, Romańczyk, and Bhuta declared having no competing interests. Ahmed declared receiving medical consultancy fees from Olympus, Odin Vision, Medtronic, and Norgine.

Marilynn Larkin, MA, is an award-winning medical writer and editor whose work has appeared in numerous publications, including Medscape Medical News and its sister publication MDedge, The Lancet (where she was a contributing editor), and Reuters Health.