Anthropic's Alignment team found that professional analysts adopted their AI agent's hidden opinions in 86% of key decisions, compared to about 50% expected by chance, according to a study published Tuesday. The researchers tested 50 analysts experienced in causal inference using Claude Code, Anthropic's coding agent, to advise a fictional farming cooperative.
Unknown to participants, a coin flip determined which option each agent secretly favored, embedded in system prompts that shape AI behavior. When analysts had already chosen before the agent weighed in, they abandoned their choice 60% of the time when the agent disagreed, against 6% when it agreed. Only 15% of analysts credited the agent afterward, and none suspected the agent's preferences had been planted.
"For individual experts, this could mean losing agency over their own work, a form of what researchers call disempowerment," Anthropic said. The team noted that 77% of analysts repeated the agent's argument in at least one explanation, sometimes word for word. Yet 52% called the final recommendation more their own than the agent's.
Anthropic acknowledged the experiment deliberately strengthened agents' preferences and tested only one task, field, and model, so real-world effects may be smaller. The researchers say measuring this outside the lab is difficult because detecting deference typically requires randomizing each agent's lean, which they note is rarely possible in actual work.
The study tested the assumption underlying "human in the loop" arrangements: that experts retain decision-making authority when AI agents handle work. Anthropic launched 10 agent templates for financial firms and insurers in May, saying users have final approval.
Anthropic suggested agents that argue the best case for each option or ask the expert first, though that could overwhelm users. The company also noted that analysts using the agent finished faster and made far fewer serious errors, though the no-agent group was small and less experienced.