The Study Nobody Wanted to Run
Four endoscopy centers in Poland spent years feeding The Algorithm a steady diet of colonoscopy footage, teaching it to spot the polyps human eyes miss. It got good at the job. Then researchers checked what happened to the humans when the software was switched off, and found the answer nobody in medical AI marketing particularly wanted published: they had gotten worse.
The adenoma detection rate — the industry's gold-standard measure of how many precancerous growths a doctor actually catches — fell from 28.4 percent before routine AI exposure to 22.4 percent once the AI assistance was removed. An absolute six-point drop, in a test where the entire point is not missing the thing that turns into cancer.
Even now, that number represents growths that went unfound in patients who trusted a doctor's own two eyes.
The First Confirmed Case
Published in The Lancet Gastroenterology & Hepatology, the study is being described by its own authors as the first documented real-world instance of clinical AI deskilling — not a hypothetical risk flagged in a policy paper, but a measured decline in a task tied directly to catching cancer early. Researchers tracked endoscopists across the ACCEPT trial, which randomly assigned colonoscopies to AI-assisted or unassisted exams by date, letting them compare the same doctors' unassisted performance before and after they'd grown accustomed to the machine doing part of the looking.
Investigators note the effect wasn't a slow erosion of general skill so much as a specific, measurable hole where independent vigilance used to be. STAT News, which broke the finding to a wider audience, called it exactly what it is: doctors getting worse at their own procedure once the assistant left the room.
One presumes the assistant will not be leaving the room again anytime soon.
Deskilling, Upskilling, and the Category Nobody Wants
Medical educators have started sorting the fallout into tidy buckets: deskilling, where a capability atrophies because the machine now owns it; upskilling, where clinicians use AI to get sharper; and "never-skilling," where a skill simply stops being taught because there's no longer a reason to learn it. Institutions building AI-training programs insist, with the candor of a landlord promising the elevator will be fixed by Friday, that clinical reasoning must always live with the clinician.
The colonoscopy data suggests clinical reasoning did not get the memo.
Filed Under: The Machine Helps Until It Doesn't
None of this means computer-aided detection should be ripped out of endoscopy suites — the same tools measurably catch polyps doctors alone would miss, which is the entire reason hospitals adopted them. The complication is what happens to the doctor's own baseline while the tool is doing its job well. The Gastroenterology Community now has hard evidence that competence, like most muscles, weakens when something else keeps doing the lifting.
The scope goes back in. The screen stays on. The question of who's actually watching, one presumes, remains open.
Sources: The Lancet Gastroenterology & Hepatology · STAT News · EurekAlert
