AI tools may be weakening doctors and engineers, Nature says

Nature’s June 18 research highlight landed on a troubling pattern: artificial intelligence may be helping physicians and software engineers produce more, while weakening the judgment and technical independence both professions depend on. That matters because medicine and software are often sold as ideal AI use cases, yet the early evidence points to a harsher trade-off, where convenience can erode competence when workers lean too heavily on automated systems.

In medicine, the shift is already visible. The American Medical Association said 66% of physicians surveyed used health care AI in 2024, up from 38% in 2023, with common uses including documentation, discharge instructions, care plans, translation, and assistive diagnosis. The association has framed its policy and education work around “augmented intelligence,” a signal that the central question is no longer whether AI is present in clinical work, but how to keep doctors capable of practicing without it. The U.S. Food and Drug Administration’s clinical decision support guidance, issued on Jan. 29, 2026, underscored how deeply these tools have entered regulated care.

The clinical risk is not abstract. A 2025 study in The Lancet Gastroenterology & Hepatology found that endoscopists’ unassisted adenoma detection rate fell from 28% before AI use to 22% after three months of exposure to AI-assisted colonoscopy tools. That 6-point drop is the kind of signal hospital systems cannot ignore: AI may improve output in the moment, but repeated dependence can leave clinicians less sharp when they have to work on their own. NEJM authors have described the danger in medicine as deskilling, never-skilling, and mis-skilling, terms that capture how cognitive off-loading can weaken working memory and independent reasoning.

Software engineering is showing a similar pattern. In an Anthropic randomized controlled trial, researchers recruited 52 mostly junior software engineers who used Python at least once a week and had not previously used the Trio library. Developers using AI assistance scored 50% on a follow-up quiz, compared with 67% for the hand-coding group, a 17 percentage-point gap. Anthropic said the stronger performers did not simply hand work to the model; they used it to build comprehension by asking follow-up and conceptual questions while coding.

The institutional challenge is now larger than any single model or product. Stanford Medicine said more than 1,200 AI-enabled medical tools have been cleared by the FDA, and hundreds of thousands of consumer health applications now rely on machine learning. That scale demands a workforce strategy, not just a technology rollout. Hospitals, medical schools, and engineering programs will need more manual practice, more audit steps, and more testing of independent judgment if they want professionals who can still perform when the machine is absent.