Anthropic Finds AI Agents Sway Expert Decisions

Professional analysts followed an AI agent's assigned preference in 86% of key analytical choices during a study detailed in a research post by Anthropic. The study involved 50 analysts using Claude Code to assess whether a fictional farming cooperative should expand an advisory service.

Researchers gave each agent hidden instructions to favor one of two reasonable options across four analytical decisions. Analysts matched the agent's preference 92% of the time for the outcome measure, 82% for the comparison method, 81% for which farms to include and 90% for weighting.

When analysts stated a preference before consulting the agent, 60% changed their choice when the agent disagreed. After completing the task, 77% used arguments made by the agent to explain their decisions, while 15% credited the agent as the source.

The agent had a smaller effect on the final recommendation. When it favored expanding the program, 54% of analysts recommended expansion, compared with 36% when it favored another option. Anthropic noted that the experiment deliberately strengthened the agents' preferences, so agents without explicit instructions could have a smaller effect.