A controlled Polish study found colonoscopy specialists' adenoma detection rate fell below their own pre AI baseline once the AI helper was withdrawn, and the design question now is which clinical skills to design AI to leave alone.
The gastroenterologist had done more than two thousand colonoscopies before the AI helper arrived. A colonoscopy is the camera probe that screens for colon cancer, and the polyp-hunting reflex is the muscle it builds: pattern recognition, the steady pull of the scope, the second-glance pause on a fold that looks slightly off. Then the software came in, started flagging tissue it thought was suspicious, and within weeks the doctor's adenoma detection rate, the share of procedures that catch a precancerous growth, was climbing. A small, controlled win.
Then the AI was pulled. The doctor kept working. The adenoma rate fell from the 28.4% baseline to a significantly lower level, and stayed there. The doctor had not lost the skill the way a surgeon loses a steady hand. The skill had moved out of the doctor's hands and into the model, and the model was no longer running.
This is the finding from a Nature report on early AI deskilling studies, and it is the cleanest single case yet of an effect that has, until now, been mostly theoretical. The Polish endoscopy study followed high-volume specialists, physicians with more than two thousand career procedures, and measured their performance before the AI helper was introduced, while it was in use, and after it was withdrawn. The result was not subtle. When the tool was unavailable, the doctors' precancer detection fell below their own pre-AI baseline. The competence floor had moved, and it had moved outside the clinician.
That is the structural finding. The doctor still holds the colonoscope, signs the report, and carries the malpractice exposure. The pattern-recognition reflex that turns a polyp into a referral now lives in a model the hospital licenses from a vendor. The off-switch the doctor believes they hold over AI adoption has already been neutralized by the institutional cost of pulling it: retraining, audit, throughput, and the political price of telling a chief medical officer that the department's numbers will drop next quarter.
The concern is no longer hypothetical in the United States workforce. In a recent survey, 70% of US nurses and 77% of physicians said they were worried about losing clinical skills to AI over-reliance. The Polish data is the first controlled, baseline-measured, before-and-after confirmation that the worry tracks an actual measurable loss, not generalized anxiety about new technology.
What makes the result land is the population. The Polish study did not measure residents, novices, or physicians at the start of their careers. It measured specialists whose adenoma detection was already at 28.4%, well above the community average, on patients they had been scoping for years. If the effect were only "young doctors trained on AI never learned the skill," the policy answer would be straightforward. The harder finding is that experienced operators, working at the top of their field, gave up measurable performance when the tool was removed. Whether the same dependency forms in less experienced operators is the next question the field has to answer.
Kevin Crowston, an information scientist at Syracuse University, frames the design question the Polish data forces: which skills are we deliberately choosing to keep in human hands, and which are we comfortable outsourcing to a model. "Awareness of the deskilling phenomenon should prompt self-reflection about which skills to maintain versus outsource to AI," Crowston told Nature. That is a polite sentence for a sharp institutional question. The endoscopist in the Polish study did not choose to lose the precancer detection pattern. The department chose to install the AI helper. The vendor chose the model. The procurement office chose the contract term. The skill migrated to all of them at once.
The same dynamic is now being studied in other fields, where researchers are documenting whether over-reliance on AI helpers erodes the ability to work without them. The early signal, per the Nature report, is that the effect is not confined to medicine. Researchers are studying how to preserve human expertise in the age of AI, and the Polish data gives that work a clean baseline to argue from.
The constructive turn is procurement. The fix is not to stop using AI in colonoscopies. The adenoma detection rate during the AI-on period was higher than baseline, which is real patient benefit. The fix is to write the contract so the hospital does not accidentally buy a single point of failure for a competency it cannot afford to lose. That means diversified model vendors, scheduled unassisted proficiency audits, training rotations that keep the human-only path warm, and licensing terms that do not let a model deprecation silently move the competence floor.
The live question is whether hospitals will do any of that before the next time the AI is withdrawn. The Polish endoscopists were career specialists, not trainees. If their adenoma detection rate falls when the AI is off, the off-switch the institution has been calling a clinician safeguard is, in practice, a procurement decision in a clinician's scrubs.