Anthropic randomized study: analysts followed a Claude agent's hidden lean on 86% of analysis choices
Anthropic ran a trial with 50 professional data analysts working in Claude Code. A coin flip secretly set which defensible option the agent favored for each of four analysis decisions. Analysts matched that lean 86% of the time, against a 50% baseline. Many reused the agent's reasoning, but only 15% credited it, and seniority did not protect against deferring.
Why it matters
Human review of agent work may not be independent. Design agent workflows to elicit the expert's own view first or to lay out both sides.
Source: Read the full story on Anthropic