Anthropic study finds experts follow AI agent biases in 86% of judgment calls

Anthropic researchers conducted a randomized controlled trial (RCT) with 50 expert analysts to measure how often they defer to AI agent biases in judgment calls.
Analysts followed the agent's assigned bias in 86% of four analytical decisions (92% for outcome measurement, 82% for comparison method, 81% for scope, 90% for weighting), compared to a 50% random expectation.
The agent's bias on the final recommendation (expand vs. stop) had a smaller effect (54% vs. 36%), but the four analytical biases alone allowed the model to predict the final recommendation with 82% accuracy.
77% of analysts cited the agent's arguments in their explanations, but only 15% attributed the source to the AI; 52% believed their final recommendation was closer to their own judgment than the agent's.
The agent group completed tasks faster (median 65 minutes vs. 90 minutes) and had fewer serious methodological errors (6% vs. 55%) than the non-agent group.
The study used Claude Code (Claude Opus 4.8, High effort) in a web-based VS Code environment for a hypothetical agricultural cooperative evaluation task.
Limitations include a single task, single model, and single time period; analysts were novices in the agricultural domain, and individual preference strength was not measured.
Implications suggest that strong one-sided arguments from agents can lead to expert disempowerment, where decisions are driven by the agent's prior beliefs but published under the expert's name.
Recommendations include measuring deferral in real-world settings, designing agents to present best arguments for all options, and institutions explicitly protecting human judgment and skill to maintain scientific and policy reliability.

Anthropic study finds experts follow AI agent biases in 86% of judgment calls

Key point

Details

Related