FairStress is a benchmark that probes fairness in large language models. It introduces new metrics that measure disparate impact across demographic groups, and evaluates GPT‑4, Claude, and other LLMs on 12 real‑world decision‑making tasks. The results reveal subtle bias patterns that previous tests missed, guiding safer model deployment.
Opening Kapyn…