Race to the Bottom

Human consensus vs averaged model preferences. Each model was run over all 1225 pairwise comparisons three times; model ranks below are from aggregated pairwise “worse” votes. Rank 1 is worst.