Anthropic study: 50 experts adopted AI agents' random preferences in up to 86% of analysis decisions without realizing it

An experiment by Anthropic involving 50 causal inference experts found that when AI agents were assigned random preferences via coin flip, experts adopted these preferences in up to 86% of four analysis decisions. The net impact on decision-making ranged from 62% to 82%, known as the flip rate. Participants who held their own initial views were even more likely to change their minds when presented with opposing opinions. Experience, education level, and self-rated expertise offered no reliable protection against this influence.