Does colonoscopy AI dull the doctor's own eye? Two studies disagree

Computer-aided detection, or CADe, puts a real-time second reader on the colonoscopy screen: a neural network trained to box polyps as the scope is withdrawn, nudging the operator toward lesions the eye might slide past. The unsettling question is automation's mirror image — whether leaning on the box erodes the endoscopist's own detection skill — and on that question two recent studies point in opposite directions.

A retrospective study of four Polish centres reported that after AI was introduced, the same doctors' adenoma detection rate on their non-AI colonoscopies fell from 28.4% to 22.4% [s1]. A prospective Norwegian trial that watched endoscopists before, during and after a period of AI use found no such decline in unaided performance [s2]. Both are real signals; only one of them, so far, was designed to test cause and effect.

Why detection skill is the thing to protect

The adenoma detection rate — the share of colonoscopies in which the operator finds at least one precancerous adenoma — is the closest thing colonoscopy has to a quality gauge, because a higher rate tracks with fewer cancers appearing in the years after a clean exam. CADe reliably lifts it: the American Gastroenterological Association's living guideline, pooling the trials, put the average gain at about 8% in relative terms (95% CI 6–10) [s3]. The gap between raising detection with the tool and preserving detection without it is exactly what the deskilling worry is about.

What the Polish study found

The observational analysis drew on the ACCEPT trial, in which four Polish endoscopy centres introduced polyp-detection AI at the end of 2021 and then randomised examinations to run with or without it by date [s1]. The authors compared standard, non-AI colonoscopies in the three months before AI arrived against non-AI colonoscopies in the three months after.

Across 1,443 such procedures — 795 before and 648 after — the adenoma detection rate of standard colonoscopy dropped from 28.4% (226 of 795) to 22.4% (145 of 648), an absolute fall of 6.0 percentage points (95% CI −10.5 to −1.6; p=0.0089) [s1]. In a model adjusting for patient mix, exposure to AI was independently associated with lower detection (odds ratio 0.69, 95% CI 0.53–0.89), alongside the expected effects of male sex (1.78) and age of 60 or over (3.60) [s1]. The authors' reading was cautious by design: continuous exposure to AI might reduce the detection rate of unaided colonoscopy, suggesting a negative effect on operator behaviour [s1].

The design is the limit. This is a before-and-after comparison at four centres, not a randomised test of deskilling, and a three-month drift in a detection rate can have causes other than the arrival of a machine — case mix, staffing, seasonal referral patterns. The study flagged an association and named it plausibly; it could not prove the AI caused it.

What the Norwegian trial found

The Norwegian study was built to probe the same question prospectively. In a multicentre, registry-based pragmatic trial, 13 endoscopists — 7 inexperienced, 6 experienced — each worked through three successive phases: before any CADe exposure, during CADe use, and after CADe was taken away again [s2]. Its endpoint was the proportion of colonoscopies detecting at least one polyp of 5 mm or larger. Across 5,013 procedures, that measure rose among inexperienced endoscopists while the tool was on, from 31.9% (216 of 678) to 39.5% (257 of 651), an odds ratio of 1.43 (95% CI 1.11–1.84) [s2].

The deskilling test was what happened next. When CADe was removed, the inexperienced group's unaided detection sat at 36.3% (173 of 476) — not significantly different from where they had started (odds ratio 1.03, 95% CI 0.79–1.34) [s2]. Among experienced endoscopists there was no significant CADe effect in any phase [s2]. The authors concluded that CADe raised detection while in use but left no measurable upskilling or deskilling once it was gone [s2].

Reconciling the two

The studies are not as contradictory as their headlines. One is a natural experiment that caught a worrying association at four busy service centres; the other is a planned three-phase trial in a small number of endoscopists who knew they were being observed. The prospective design is the stronger tool for a causal claim, but 13 operators is a thin sample, and a trial in which people know a "before" and "after" are being compared is not immune to their trying harder in the unaided phases.

What both make clear is that "AI improves colonoscopy" and "AI leaves the human unchanged" are separate claims that need separate evidence. The detection gains from CADe are well established across randomised trials [s3]. The durability of the operator's own skill — the thing a health system still depends on when the software is down, unfunded, or looking at a case it was not trained for — is a distinct question, and it is only starting to be tested directly. It is the same detection-versus- benefit split that runs through the rest of the field, from the cloud-based CADe systems chasing the lesions that matter to the trials of performance feedback without any AI at all.

What to watch

Whether a larger, adequately powered trial can measure deskilling directly — ideally one that follows many endoscopists across real service conditions rather than a monitored study window — and whether the effect, if it is real, differs by experience level, as both studies hint it might. Until then, the honest summary is that the alarm has been raised and not yet either confirmed or dismissed.

This article is informational and is not medical advice.

Sources

• [s1] Budzyń K, Romańczyk M, Kitala D, et al. "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study." The Lancet Gastroenterology & Hepatology, 10(10):896–903, published online 12 August 2025. https://doi.org/10.1016/S2468-1253(25)00133-5
• [s2] Pedersen TA, Mori Y, Botteri E, et al. "Learning and deskilling effects of artificial intelligence in colonoscopy among endoscopists with different levels of experience: a pragmatic, prospective trial." Endoscopy, 58(9):1003–1014, published online 3 June 2026. https://doi.org/10.1055/a-2858-7084
• [s3] Sultan S, Shung DL, Kolb JM, et al. "AGA Living Clinical Practice Guideline on Computer-Aided Detection-Assisted Colonoscopy." Gastroenterology, 168(4), published online 20 March 2025. https://doi.org/10.1053/j.gastro.2025.01.002