Nature’s June 21 roundup of two 2026 peer-reviewed studies put hard numbers on a growing concern: AI can improve short-term workflow, but users may lose core skills once the tool is removed. The evidence came from two very different settings, medicine and software engineering, and both pointed in the same direction.
Doctors posted lower adenoma detection rates when AI was removed
The Polish ACCEPT trial focused on experienced physicians, not trainees. Every participating doctor had completed at least 2,000 colonoscopies. On some clinical days, they were allowed to use an AI support system that analyzed colonoscopy images in real time and flagged suspected adenomas. On other days, no AI assistance was available.
The results, published in The Lancet Gastroenterology & Hepatology, showed a clear drop outside the AI setting. Before AI was introduced, the physicians’ adenoma detection rate stood at 28.4%. After AI entered the workflow, their detection rate in sessions without AI assistance fell to 22.4%, a decline of 6 percentage points. The study said continued use of AI tools could make clinicians less proactive, less attentive, and less personally accountable when making cognitive decisions without AI support.
Yuichi Mori, a physician researcher at the University of Oslo, said there is currently no established solution to counter skill degradation and called it one of the hottest research questions for the next decade. Robert Wachter of the University of California, San Francisco, offered a blunt reading: even highly skilled professionals may lose ground in their own field as dependence on AI grows.
Anthropic trial found AI users scored 17 points lower
The second study came from Anthropic researchers Judy Hanwen Shen and Alex Tamkin and was released on January 29, 2026. It enrolled 52 junior software engineers, all tasked with learning the same new Python package, Trio. Everyone could search the web and read official documentation. Half of the participants also received access to an AI assistant.
Post-task testing showed the AI group averaged 50%, while the manual coding group averaged 67%, a gap of 17 percentage points. The study said that difference was roughly equal to two letter grades. The time savings were limited: the AI group finished only about 2 minutes faster on average, and that result was not statistically significant. In other words, the gain in speed was small, while the loss in understanding was much easier to see.
The weakest area was debugging. Shen and Tamkin highlighted that point because catching errors produced by AI remains one of the most important human oversight functions. If engineers outsource that work over long periods, their ability to spot AI mistakes may shrink as well.
How AI was used mattered as much as whether it was used
The experiment also drew a sharp line between two styles of use. Engineers who used AI for concept exploration scored above 65%. Those who offloaded code generation almost entirely to AI scored below 40%. The spread between those approaches reached 25 percentage points.
Neither study asked whether AI is useful in general. The real question was what remains of human judgment and execution after repeated AI assistance is taken away. Based on the material cited by Nature, there is still no broad consensus on the best frequency for AI support, and no validated intervention yet for preserving core skills inside AI-heavy workflows.