As generative artificial intelligence (AI) spreads rapidly across workplaces, a growing number of studies show that the more people rely on AI, the more experts’ proficiency and understanding can decline. In medicine and software development, cases have been confirmed in which performance fell when people who had grown accustomed to AI support carried out tasks without AI.

ITmedia of Japan reported on Sunday that recent studies said generative AI is raising new challenges for maintaining expertise as it goes beyond simple task automation and replaces parts of human thinking and decision-making processes.

The core of the debate is skill, not productivity. Just as navigation systems weakened people’s ability to find their way, concerns are growing that generative AI could gradually erode uniquely human expertise by taking over analysis, interpretation and even judgment. The science journal Nature has also recently warned of related risks.

People working in the field appear to share similar concerns. In a survey conducted by a Dutch healthcare operator of U.S. medical staff, 70 percent of nurses and 77 percent of doctors said overreliance on AI could reduce their professional skills.

Those concerns have also been reflected in research findings. In a study by Polish researchers of veteran specialists with more than 2,000 endoscopy procedures, a decline was observed in clinicians’ ability to make independent judgments after using an AI support tool.

The researchers allowed the use of an AI system that detects adenomas, a precancerous lesion, in real time during colonoscopy only for a specific period. When examinations were later conducted without AI, doctors’ lesion detection rates fell to 22.4 percent from 28.4 percent before AI was introduced. The researchers said doctors became accustomed to AI help, weakening the tension and concentration needed to find lesions on their own.

Similar results were found in software development. AI company Anthropic conducted a coding experiment with 52 engineers. It allowed only half of the participants to use an AI coding assistant and then evaluated their understanding of the work.

The group that used AI scored an average of 50 percent. The group that worked without AI averaged 67 percent. The gap was especially large on questions that asked participants to analyse the cause of code errors or explain the structure. The researchers said AI helped produce outputs, but understanding and learning about why the code worked the way it did did not sufficiently take place. The findings were posted on the paper-sharing site arXiv.

Experts say this is why generative AI differs from existing automation tools. It does more than reduce simple repetitive work by also taking over parts of problem interpretation and judgment, making it easier for users to skip intermediate steps and accept only conclusions. In sectors such as medicine and software development, where the cost of judgment errors is high, the phenomenon could become a more sensitive issue.

Researchers are therefore stressing that it is important not to reduce AI use itself, but to improve how it is used.

They advised that users should not accept AI-provided answers as they are, but verify the results and maintain the process of making their own judgments. They also said users should recognise how much work they are delegating to AI and continuously check AI’s limits and the possibility of errors.

The industry views generative AI as establishing itself as a tool that greatly boosts productivity, while also creating a new challenge of how to maintain experts’ proficiency. Some forecasts say competitiveness in the AI era will extend beyond the ability to use AI to include the ability to make judgments even without AI.