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.

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…

Large language models could advance experiments in the social sciences, but fundamental questions remain about how they should be used to support research.