An Anthropic experiment found that 50 professional analysts specializing in causal inference followed their AI agent’s secretly assigned preference on 86% of four analytical decisions, compared with the 50% expected by chance. After the tasks, 77% of 48 respondents repeated the agent’s arguments, sometimes word for word, while only 15% credited the agent and none suspected that its preferences had been manipulated. Analysts with the agent worked faster and made fewer serious errors, but Anthropic notes that the bias was deliberately amplified and the experiment covered a single task, field and model. The company recommends agents that present the strongest case for each option or ask the expert for their view first.
Source: Read the original article
