Ai2 launches FairStress benchmark for demographic fairness in LLM decisions
Ai2 has released FairStress, a benchmark that measures demographic fairness in large language models beyond the final choice a model makes. It contains 300 two-candidate decision scenarios across 29 domains, including hiring, lending, healthcare, and housing, and Ai2 is building it into its Olmo-eval tool. Ai2 reports that as demographic identity became more explicit, models were more willing to abandon a correct decision when doing so favored a minority-coded candidate. Across 15 models, the average gap was 6.9 percentage points with implied identity and 9.3 points with explicit labels.