Anthropic study: expert analysts followed AI agents' secretly coin-flipped leans 86% of the time
In a randomized trial from Anthropic's Alignment team, 50 professional analysts experienced in causal inference used Claude Code to advise a farming cooperative on whether to expand an advisory program. On four analytical judgment calls, each agent was secretly assigned a lean by coin flip and told to make the best honest case for it. Analysts matched the agent's lean 86% of the time, versus about 50% expected by chance, and switched away from their own stated preference 60% of the time when the agent disagreed. Afterward, 77% justified their choices with the agent's arguments, sometimes nearly verbatim, but only 15% credited the agent. Anthropic cautions the lean was deliberately amplified and tested on a single task and model.
Source: Gizmodo