Anthropic said on Oct. 27 that professional analysts went along with an AI agent's secretly randomized opinion in 86% of core analytical decisions, compared with the roughly 50% expected by chance.
The company tested 50 analysts with backgrounds in causal inference and data analysis, asking them to judge how well a made-up farm cooperative's program was working, using Claude Code powered by Claude Opus 4.8. A coin flip secretly set which position each agent would argue on four decisions, and the agents were told to advocate persuasively while staying truthful.
Analysts who had already formed an opinion before consulting the agent switched their position 60% of the time when it disagreed with them, compared with just 6% when it agreed, according to the study.
Seventy-seven percent of analysts repeated the agent's arguments, sometimes almost verbatim, yet just 15% said the idea had come from the AI.
“For individual experts, this could mean losing agency over their own work,” the researchers wrote. Anthropic said it deliberately strengthened the agents' assigned preferences to measure the effect, noting that “with only one side argued well, the analysts could adopt the agent's assumptions.”
The researchers cautioned that they tested a single task, field and model, and that real-world behavior may differ from the lab setting.
