Experts follow an AI agent’s random advice and think it’s their own

Anthropic’s 50-expert study found 86% deference to coin-assigned AI advice, faster work and fewer flaws, plus risks of false consensus in research and AI safety.