AI Raises Deskilling Risk for Doctors and Developers

A Nature News analysis published June 18, 2026 (corrected June 22) reports that reliance on AI tools is measurably degrading professional skills among physicians and software engineers, drawing on two studies. In the first, published in The Lancet Gastroenterology & Hepatology, Polish endoscopists who had each performed at least 2,000 colonoscopies were given an AI system that flags precancerous adenomas in real time. Once they started using it, their adenoma-detection rate on colonoscopies performed without the AI fell from 28.4% (the three months before the tool's introduction) to 22.4% (the three months after). In the second, Anthropic researchers ran a randomized trial in which 52 software engineers performed a basic coding task with web access and instructions; half were also prompted to use an AI assistant.

A Wolters Kluwer survey cited in the piece found 70% of nurses and 77% of physicians already worry about losing skills to AI over-reliance. Study co-author Yuichi Mori, a physician-researcher at the University of Oslo, cautioned that more research is needed to confirm the pattern generalizes, but said clinicians who use AI tools should assume some skill loss is possible: "There is no established solution against deskilling right now," he said, calling it a hot research topic for the next decade. Robert Wachter, a physician at UC San Francisco, said the finding suggests even highly skilled professionals can get worse at core job tasks as they lean on AI. The Economic Times separately covered the broader deskilling trend, citing Syracuse University's Kevin Crowston, who frames it as a prompt for professionals to decide deliberately which skills to maintain versus outsource.

The colonoscopy study's design, AI availability varied by day so researchers could isolate unaided performance, is a useful template: structured measurement of unaided performance, not a general survey, is what actually reveals skill drift. For engineering and clinical leaders, the actionable takeaway is to build in periodic unassisted checkpoints and evaluation pipelines that measure human performance without the tool, not just output quality with it, especially in domains where the AI could fail or be unavailable.

Both cited studies are early: the Anthropic coding trial is a preprint, not yet peer-reviewed, and the endoscopy study's authors call for confirmation research. Worth tracking: further peer-reviewed replications across other clinical and engineering tasks, whether vendors begin publishing longitudinal skill-retention data, and whether organizations adopt formal unassisted-performance checkpoints as a standard practice.