How AI Agents Influenced Professional Analysts’ Decisions

Anthropic studied how AI agents influence specialists’ decisions at work. In the experiment, 50 professional analysts used Claude Code to assess whether a farm advisory support program should be expanded. During the analysis, they had to make four choices, such as deciding how to measure productivity and which farms to include.

Researchers secretly set the agent’s preference on each of the four questions and asked it to argue for its position. On average, analysts chose the option proposed by the agent in 86% of cases. If an analyst had already stated an opinion before the agent responded, they changed their choice in 60% of cases when it differed from the agent’s position, and left it unchanged in 94% of cases when the two positions matched. These results are based on 38 such instances.

After the assignment, 77% of participants used the agent’s arguments in their explanations, sometimes almost verbatim, but only 15% said they had taken them from AI. The authors emphasize a limitation of the experiment: the agents’ preferences were deliberately amplified to measure their influence. The results therefore do not show how often specialists follow AI’s position in their usual work.

Image: Basile Morin CC BY-SA 4.0 · License

Translation editor: Салтанов А.