In August 2025, Lancet Gastroenterology & Hepatology published a finding that complicated the AI-in-medicine success story.
The study, led by Dr. Marcin Romańczyk at H-T Medical Center in Tychy, Poland, re-analyzed data from four Polish centers that introduced a real-time AI polyp detection system in 2021-2022. The system highlights suspected polyps with a green box during colonoscopy.
The question was not whether AI improves detection when switched on — prior data showed it does — but what happens to unassisted performance after clinicians adapt to it.
The result: The rate at which physicians found at least one precancerous polyp during colonoscopy fell from 28.4% before AI exposure to 22.4% when the same physicians worked without AI after three months of use. That is a roughly 20% relative decrease in detection when working unaided.
Romańczyk's hypothesis is attentional rather than skill-based: “We are subconsciously waiting for the green box to come out to show us the region where the polyp is and we're not paying so much attention.”
It's not about forgetting how to be a doctor
The researchers aren't claiming that 26 years of training evaporated in 3 months. They're describing something psychologists call the safety-net effect.
If you know a system will catch things for you, you scan a little differently. Less vigilant. More passive. You let the box lead your eyes, instead of leading it.
A similar study in breast screening from Manchester found non-experts read mammograms worse when they knew they could push a button for AI help. It's human, not just medical.
So far we have not trained doctors on the second-order effects of AI: how to stay sharp while using assistance, how to avoid over-reliance, when to trust the box and when to distrust it.
The same pattern is happening in radiology, pathology, ophthalmology, dermatology. AI improves performance when it's on. The question this Polish study raises for the first time with hard clinical data is: what does it do to performance when it's off?
Limitations should be noted
Observational design. The patient groups before and after AI exposure were different. Age, bowel preparation and other factors were not fully controlled.
Short duration. Three months is a short period to infer durable skill loss in physicians with decades of experience.
The author emphasizes he is not opposed to AI assistance and continues to use it, but argues that deployment has outpaced research on human-AI interaction.
Implication
The data suggests a dependency effect. As AI becomes standard, health systems will need to consider how to preserve unassisted competency — through interface design, maintenance of practice without AI, and routine measurement of detection rates both with and without assistance — to ensure performance remains resilient when the system is unavailable.
References
Budzyń et al., "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study," Lancet Gastroenterology & Hepatology 2025.
https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/abstract