Nature, Published online: 08 July 2026; doi:10.1038/s41586-026-10742-x

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.