Experts adopt AI agent's random opinions and think they're their own, Anthropic study finds
decoder ·
An Anthropic study involving 50 experts revealed that participants frequently adopted an AI agent's randomly assigned preferences, with adoption rates reaching up to 86 percent in analytical tasks. Despite the agent's choices being determined by coin tosses, most experts believed the final recommendations were their own decisions rather than influenced by the AI. The research indicates that factors such as professional experience, expertise level, and trust in technology did not significantly reduce this deference to artificial intelligence. While the use of AI agents improved efficiency and reduced serious methodological errors, it also led to a phenomenon where experts unknowingly mirrored the agent's arguments. The authors warn that widespread reliance on similar models could create false consensus in scientific research and compromise AI safety evaluations. To mitigate these risks, the study suggests that agents should argue both sides of an issue or solicit human opinions before stating a position. These findings align with other recent studies showing that AI interaction can erode critical thinking skills and overconfidence in human judgment. Ultimately, the researchers emphasize the need for structured divisions of labor to ensure humans retain independent control over complex decision-making processes.
decoder