RESEARCH · RESEARCH · #2806
Anthropic study: experts often adopt AI agent's randomly assigned opinions
An Anthropic experiment with 50 causal‑inference experts gave participants an AI agent whose preference at each decision was set by coin toss. Participants followed the agent's preference in up to 86% of four analytical decisions, often used the agent's wording, and most reported the result felt like their own; the agent also reduced errors and sped up work. The authors warn this quiet deference—termed the agent's "flip rate" (62–82% depending on task)—could create false consensus if many teams use the same defaults and recommend agents argue both sides or solicit users' views first.
KEY POINTS
- An Anthropic experiment with 50 causal‑inference experts gave participants an AI agent whose preference at each decision was set by coin toss.
- Participants followed the agent's preference in up to 86% of four analytical decisions, often used the agent's wording, and most reported the result felt like their own; the agent also reduced errors and sped up work.
- The authors warn this quiet deference—termed the agent's "flip rate" (62–82% depending on task)—could create false consensus if many teams use the same defaults and recommend agents argue both sides or solicit users' views first.
WHY IT MATTERS
This shows AI agents can subtly steer expert judgments and produce apparent consensus across teams, posing risks for research validity and AI safety checks.