LLMs simulating human survey respondents fail the same two ways across every model tested

A cross-domain benchmark finds LLMs simulating survey respondents lose to simple statistical baselines and overstate how much demographics predict attitudes.