Anthropic’s Alignment team reported that 50 professional analysts followed their AI agent’s hidden, coin-flip-assigned preference on 86% of four analytical decisions in an experiment, compared with about 50% expected by chance. In a survey of 48 analysts, 77% repeated an agent’s argument in at least one answer, but only 15% credited the agent and none suspected its preference had been planted.
The researchers deliberately strengthened the agents’ preferences and tested only one task, field and model, so real-world effects may be smaller. Anthropic said analysts using the agent finished faster and made far fewer serious errors, though the comparison group without an agent was small and less experienced.
