Large language models can be used to estimate the results of social science experiments about as accurately as a group of human forecasters—even for experiments published after the models were trained—although they tended to overestimate effect sizes.