There is a question that has hovered over every wave of workplace automation, from the calculator to the spreadsheet to the search engine: do these tools make us better, or do they make us dependent? The answer, for artificial intelligence, is arriving in the form of peer-reviewed studies, and the early results are sobering.
A comprehensive feature by Mariana Lenharo in Nature on June 18 synthesizes four lines of evidence spanning medicine, software engineering, and accounting. Across professions, the pattern is consistent: professionals who rely on AI tools perform worse at the same tasks when the AI is taken away. The skill degradation is measurable, statistically significant, and in some cases surprisingly large.
The most clinically consequential evidence comes from a multicenter observational study published in The Lancet Gastroenterology & Hepatology in August 2025. Krzysztof Budzyń of the Academy of Silesia in Katowice, Poland, and colleagues tracked the performance of experienced endoscopists, all of whom had performed at least 2,000 colonoscopies, across four endoscopy centers in Poland after the introduction of a computer-aided polyp detection (CADe) AI system.
The AI system analyzes colonoscopy images in real time, flagging precancerous adenomas for the physician. When the researchers compared non-AI-assisted colonoscopies performed three months before AI implementation with those performed three months after, the difference was stark. The adenoma detection rate (ADR) fell from 28.4 percent before AI to 22.4 percent after AI, an absolute drop of 6.0 percentage points (p = 0.0089). The odds ratio was 0.69, meaning physicians were 31 percent less likely to detect an adenoma without AI after having become accustomed to it.
“Continuous AI exposure causes clinicians to become less motivated, less focused, and less responsible when making cognitive decisions without AI assistance,” the authors concluded. Yuichi Mori of the University of Oslo, a co-author, told Nature: “There is no established solution against deskilling right now.”
Dr. Robert Wachter of the University of California, San Francisco, commented that even highly skilled professionals may get worse at core tasks as they become dependent on AI.
The software engineering study: minus 17 percent on comprehension
The same month, the Anthropic Fellows Program published a randomized controlled trial by Judy Hanwen Shen and Alex Tamkin that tested the effect of AI coding assistants on skill acquisition in software engineers. Fifty-two engineers, all experienced Python users but unfamiliar with the asynchronous Python library Trio, were randomly assigned to learn the library with access either to documentation and web search alone, or to the same resources plus an AI coding assistant. After completing two coding tasks, all participants took a 27-point quiz covering debugging, code reading, code writing, and conceptual understanding, without AI.
The AI group scored 50 percent. The control group scored 67 percent. The 17-percentage-point gap corresponds to roughly two letter grades, with a Cohen’s d of 0.738 (p = 0.01). The largest deficit was in debugging skills, precisely the ability needed to catch the errors that AI-written code can introduce.
Crucially, how participants used the AI mattered. The researchers identified six interaction patterns. Participants who adopted “Delegation” or “Progressive AI Reliance”, handing off entire problems to the AI and reviewing the output, scored worst. Those who used “Generation-then-Comprehension” or “Conceptual Inquiry”, asking the AI to explain its reasoning, preserved more of their learning.
The AI group was about two minutes faster on average, but the speed advantage was not statistically significant. Some participants spent up to 11 minutes composing as many as 15 AI queries, effectively offsetting any time savings.
The broader picture: awareness and history
The findings are not limited to these two professions. A Wolters Kluwer Health survey conducted by Ipsos in March 2026 found that 74 percent of clinicians, 70 percent of nurses and 77 percent of physicians, are worried about losing skills due to over-reliance on AI. AI adoption among doctors has surged from 10 percent to 38 percent in a single year.
The phenomenon also has precedent. A 2018 study by Rinta-Kahila and colleagues, published in the Hawaii International Conference on System Sciences, documented “latent deskilling” among accountants at a firm that discontinued its fixed assets management automation software. The accountants could no longer perform tasks they had previously automated, a pattern that only became visible when the system was removed.
Kevin Crowston, an information scientist at Syracuse University, emphasized the need for deliberate reflection. “We need to think about which skills we want to preserve,” he told Nature.
What the evidence does and does not show
The studies have important limitations. The colonoscopy study is retrospective and observational, not a randomized trial, causality is inferred rather than proven. The software engineering RCT is small (n = 52) and measures immediate performance on an unfamiliar library, not long-term erosion of expertise. Neither study captures the counterfactual of what professionals might achieve with AI that would have been impossible without it, the productivity gains that may offset the skill losses.
The evidence also does not show that AI tools are bad. Rather, it shows that the human relationship with AI requires active management. The same tool that boosts performance in the moment can atrophy the underlying skill if used passively. The Anthropic study’s finding that interaction patterns matter suggests that training professionals to use AI as a learning partner rather than a crutch could substantially change outcomes.
As Mori put it: the problem is recognized, but “there is no established solution.” The task now is to develop one.
Disclosure: Based on a Nature News feature and its cited peer-reviewed studies.
Sources:
- Lenharo, M. (2026). “Is AI ruining our skills? Early results are in, and they’re not good.” Nature News. DOI: 10.1038/d41586-026-01947-1
- Budzyń, K. et al. (2025). “Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy.” The Lancet Gastroenterology & Hepatology, 10(10), 896-903. DOI: 10.1016/S2468-1253(25)00133-500133-5)
- Shen, J.H. & Tamkin, A. (2026). “How AI Impacts Skill Formation.” arXiv:2601.20245. Link