Does AI polyp detection weaken endoscopists' unassisted performance?

Budzyń and colleagues report, from a retrospective observational study at four endoscopy centers in Poland taking part in the ACCEPT trial, that the adenoma detection rate (ADR) of standard colonoscopy fell after those centers introduced AI for polyp detection. Standard, non-AI-assisted colonoscopies had an ADR of 28.4 percent (226 of 795) in the three months before AI and 22.4 percent (145 of 648) in the three months after, an absolute difference of -6.0 percent (95% CI -10.5 to -1.6; p=0.0089). In the multivariable logistic regression, AI exposure had an odds ratio of 0.69 (95% CI 0.53-0.89) for ADR, alongside sex and age. The authors conclude that "continuous exposure to AI might reduce the ADR of standard non-AI assisted colonoscopy, suggesting a negative effect on endoscopist behaviour." The measured object is the endoscopist's unassisted work, the procedures in which the AI was not in use.

The reasoning rests on the shape of the comparison. The authors ask whether continuous exposure to AI changes how endoscopists perform colonoscopy when the AI is off, and they answer by holding the procedure type fixed (standard, non-AI-assisted colonoscopy) and changing only the period. Behavior is inferred from an outcome, a detection rate, rather than observed directly. The abstract does not say what the endoscopists did differently during the examination, whether they looked more slowly, or whether the lower detection reflects reliance on the tool. The inference that AI exposure left these endoscopists worse when it was switched off runs through a before-and-after contrast, not through a mechanism the abstract tests.

Against the nearest notes, this is a measured case behind a claim the library has so far made in principle. Does AI augmentation protect workers from skill erosion? argues that AI framed as augmentation can still erode worker skill through overreliance. The colonoscopy study supplies a clinical skill outcome that moved in that direction once AI was in place. It also fits Which AI risks are already harming individual users today?, which scores autonomy erosion among individual-level risks already occurring. The difference matters: that survey rated risks, while this study measures a change in performance on a diagnostic skill.

The excerpt does not establish what it would take to call this deskilling in general. It is an abstract, so the number of endoscopists, the case mix in each window, and whether all four centers show the drop are not visible. The two periods are sequential, so calendar factors such as seasonality, staffing or referral patterns are not ruled out by design, even though the abstract describes AI assignment as random but keyed to the date of examination. The model adjusts for sex, age and AI exposure, and the abstract reports no other covariates. What the study supports is a before-and-after association in one country, within one trial, showing a drop of 6.0 percentage points in unassisted detection. That supports the cautious claim that unassisted detection can fall after AI exposure. The mechanism and the reach of the claim to other specialties need other evidence.