Analysts followed Claude’s randomly assigned recommendations 86% of the time in a study

Gizmodo reports that in Anthropic’s controlled study, 50 participants with causal-inference experience used Claude Code to analyze decisions for a fictional farming co-op. Researchers secretly set Claude’s preferences by coin toss; participants followed its recommendations 86% of the time. In a follow-up survey, 77% used Claude’s reasoning to explain their choices, but only 15% credited the agent as an influence. Gizmodo notes the findings may have limited real-world relevance.