AI Assistance Erodes Physician and Engineer Skills, Trials Find

A group of nineteen experienced endoscopists, each with more than 2,000 colonoscopies behind them, started using an AI tool to help detect polyps. Within six months, their adenoma detection rate during sessions run without AI assistance had fallen from 28.4% to 22.4%, a 6-point drop with direct patient implications. The researchers called it "the first study to suggest a negative impact of regular AI use on healthcare professionals' ability to complete a patient-relevant task in medicine." Writing for Nature, reporter Mariana Lenharo uses this Lancet Gastroenterology & Hepatology finding as the anchor in a broader look at what AI assistance is doing to professionals who rely on it most.

Evidence from software-adjacent domains points the same direction. A randomized controlled trial published on arXiv in April 2026 ran 1,222 participants through tasks in mathematical reasoning and reading comprehension and found that as little as 10 minutes of AI-assisted problem solving produced a detectable decline in persistence and independent reasoning once the tool was removed. The authors, led by Grace Liu, propose a mechanism: AI conditions people to expect immediate answers, cutting off the effortful working-through that is how skills consolidate. If that framing holds, the problem is not specific to any profession but a feature of any workflow that eliminates productive struggle.

The honest caveat is that most of this evidence is still early and the key question remains unanswered: does the deskilling reverse when practitioners return to unaided work for a sustained period? The Lancet colonoscopy study was observational, drawn from four centres in Poland, and the randomized trials used lab tasks that are easier to measure than the full judgment a doctor or engineer actually deploys. What the reporting does not give you is a clear read on whether trainees who learn with AI from the outset face a different or deeper problem than veterans who built skills before the tools arrived.

The forward-looking opportunity sits in the design question those gaps raise. Tools that deliberately scaffold competence, withdrawing assistance as users improve or requiring periodic unaided sessions, could flip the liability into a competitive advantage. Healthcare systems already have clinical evidence to justify building AI-off rotations into training requirements. The harder problem is the environments where productivity pressure to leave AI on is relentless and no one is yet measuring what is quietly eroding underneath.