The Skill Paradox: AI Made Expert Doctors Worse at Their Own Job
Here is an uncomfortable finding: giving expert doctors an AI assistant made them worse at their job once it was switched off — not in a lab, but in real clinics on real patients, and the ones who slipped were among the world's most experienced.
The study, in The Lancet Gastroenterology & Hepatology (August 2025), tracked colonoscopies, where doctors hunt for adenomas — growths that can turn into bowel cancer. The field's key metric is the adenoma detection rate (ADR): how often a doctor finds at least one. Researchers measured 19 highly experienced Polish endoscopists, each with 2,000+ procedures, working without AI after their clinics adopted it. The unassisted rate fell from 28.4% to 22.4% in months — a roughly 20% relative decline. ([The Lancet Gastroenterology & Hepatology](https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00133-5/abstract))
What's actually happening
The leading explanation has a name: automation-induced deskilling. When a reliable assistant is always watching, attention drifts — why scan every fold of tissue when a green box flags the suspicious ones? The habit of vigilant looking atrophies, and once the safety net is gone, the gap shows. Here, the deskilled task happens to be finding cancer.
The authors are careful, and so should we be. This was an observational study, not a randomised trial, so staffing or fatigue could explain the dip. ([EurekAlert / The Lancet](https://www.eurekalert.org/news-releases/1094223)) The broader evidence is positive — earlier trials found AI assistance increases adenoma detection. The sting: the AI-assisted rate of 25.3% beat the deskilled 22.4% but fell below the pre-AI baseline of 28.4% — a nuance flagged in the International AI Safety Report 2026. ([International AI Safety Report 2026](https://internationalaisafetyreport.org/publication/2026-report-executive-summary))
Both sides of the argument
The case for worry. A commentary by Dr Omer Ahmad of University College London warned of "the quiet erosion of fundamental skills required for high-quality endoscopy." ([The Lancet, linked Comment](https://www.thelancet.com/journals/langas/article/PIIS2468-1253(25)00164-5/abstract)) The danger isn't dramatic failure but a slow slide nobody notices until the tool is gone — an outage, a clinic without the kit, a model that misses the case a human would have caught. Assume the AI is always right, and you've built a single point of failure into a life-or-death task.
The case for calm. Deskilling is the oldest story in technology, and we've mostly survived it: surgeons stopped tying every knot by hand, pilots couldn't land manually without autopilot. We don't mourn it because the system — human plus machine — got better. An endoscopist slightly worse alone but better with AI is a net win, so long as the AI is reliably present. The honest answer turns on a question we skip: what happens when the AI isn't there, or is quietly wrong?
Why this is about your job, not just theirs
It is tempting to file this under medicine and move on. Don't. Colonoscopy is simply where the effect was measured cleanly, because the profession tracks a hard number for skill. Most of us have no ADR, which makes the same erosion harder to see, not less likely. Think about what you've already handed over — the difficult email, the report, the function, the figures. Each handoff is reasonable alone; the warning is the cumulative effect.
A subtler version shows up in how we write. A June 2026 University of Maryland study compared tens of thousands of AI-generated essays with human ones from public debates. Within a debate, 65.3% of human main arguments were unique — versus just 3.4% for the AI. ([arXiv: "Argument Collapse"](https://arxiv.org/abs/2606.01736)) The models converge on the same polished points. Outsource your first draft to a tool that flattens everyone toward one answer, and the deskilling turns collective.
The takeaway
The lesson is not "avoid AI." The tools genuinely help — which is what makes the trap so easy to walk into. Help and dependence are not the same thing, and the gap between them is where skill drains away. So treat your competence as worth maintaining: keep doing the hard part yourself, check the model instead of trusting it, and notice which skills you'd be lost without if the screen went dark. The endoscopists weren't careless — they got comfortable. The most valuable skill of the next decade may be the most old-fashioned: knowing how to do it yourself.