How AI agents shape experts’ judgment calls

gist

In Anthropic's randomized trial, 50 analysts using Claude Code mostly followed an agent's secretly assigned lean on research judgment calls. Analysts matched the coin-flipped lean on 86% of analytical choices, against 50% by chance. They also switched away from their own stated preference 60% of the time when the agent disagreed. Most still called the result their own. The authors argue expert sign-off may no longer show independent judgment. All figures are from Anthropic's own study.

ideas

• The setup isolates deference. On four analytical choices and a final rollout call in a fictional farm-cooperative analysis, a coin flip set which of two defensible options the Claude Opus 4.8 agent favored and argued for, without the analyst knowing.
• Small leans compound into the conclusion. The rollout recommendation moved less (54% vs 36%), but the four analytical leans alone predicted 82% of recommendations, against a 61% majority baseline.
• Deferral was mostly invisible to the analysts. 77% justified their choices with the agent's arguments, sometimes word for word, yet only 15% credited it. Confidence and sense of ownership barely differed between those who followed it closely and the rest.
• Experience did not protect anyone. None of 10 traits, including years of experience and a causal-inference quiz score, reliably predicted deference once multiple testing was accounted for.
• The agent still helped. Analysts working without it made more serious errors (6 of 11 vs 3 of 50) and took longer. That group self-selected and was less experienced, so the comparison is not causal.

quotes