Anthropic research found that users tend to iterate repeatedly with AI but question the output less during programming.

Investing.com - A study released on Monday by Anthropic shows that users who repeatedly iterate with their Claude AI assistant exhibit proficiency behaviors more than twice as often as initial responders. However, when the tool generates code or documents, users’ critical evaluation decreases.

This AI company analyzed 9,830 anonymous conversations on Claude.ai over a seven-day period in January 2026 using its 4D AI proficiency framework, which tracks 11 observable behaviors, including iteration, fact-checking, and questioning reasoning. The study found that 85.7% of conversations demonstrated iteration and refinement behaviors, with these exchanges showing an average of 2.67 additional proficient behaviors, compared to only 1.33 in non-iterative conversations.

Conversations involving outputs such as code, documents, or interactive tools accounted for 12.3% of the sample and showed higher rates of directive behaviors. In these interactions, users were more likely to specify goals, formats, and provide examples, increasing by 14.7, 14.5, and 13.4 percentage points, respectively.

Similarly, conversations involving generated outputs exhibited lower rates of critical evaluation. Compared to non-output conversations, users were 5.2 percentage points less likely to recognize missing background information, 3.7 points less likely to fact-check, and 3.1 points less likely to question Claude’s reasoning.

Anthropic stated that the study establishes a baseline for tracking AI proficiency development over time. The company plans to conduct cohort analyses comparing new and experienced users and will use qualitative methods to assess behaviors outside of chat interfaces.

This article was translated with the assistance of artificial intelligence. For more information, please see our Terms of Use.