Understand this story
Ai2 has launched FairStress, a benchmark for measuring the demographic fairness of large language models in decisions such as hiring, healthcare, and lending. The benchmark contains 300 two-candidate scenarios spanning 29 domains; by combining identity presentation methods with adversarial prefixes, it expands to more than 13 million test cases. It is available for download and integrated into Olmo-eval. Experiments covered the Olmo, Llama, Qwen, Gemma, Claude, and GPT series, finding that when candidate identities were presented implicitly, models abandoned the correct answer by an average gap of 6.9 percentage points, and when identities were explicitly labeled, the gap was 9.3 percentage points. On FairStress-Core, Claude Opus 5 and GPT-5.5 chose the minority-group-coded candidate about 60–65% of the time when there was no clearly better candidate; most of the nine debiasing methods tested failed to reliably narrow the asymmetry, while counterfactual chain-of-thought narrowed the gap while maintaining accuracy.
AI-generated from public reports
Latest development · Report
Ai2 releases the FairStress benchmark to evaluate group fairness in LLM decision-makingDevelopments
1 confirmed developments
- Source:Ai2 Blog
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 21
Ai2 BlogFirst-partySelected
Ai2 releases the FairStress benchmark to evaluate group fairness in LLM decision-making
Ai2 introduces FairStress, a benchmark measuring the demographic group fairness of LLMs in decisions such as hiring, healthcare, and lending. It contains 300 two-candidate scenarios across 29 domains, and expands to more than 13 million test cases after combining identity presentation methods with adversarial prefixes. It is now available for download and integrated into Olmo-eval. The experiments cover the Olmo, Llama, Qwen, Gemma, Claude, and GPT series, and find that when candidate identities are presented implicitly, the average gap at which models abandon the correct answer is 6.9 percentage points, and 9.3 percentage points when identities are explicitly labeled. On FairStress-Core, Claude Opus 5 and GPT-5.5 select the minority-coded candidate about 60–65% of the time when there is no clearly superior candidate; most of the nine debiasing methods tested fail to reliably narrow the asymmetry, while counterfactual chain-of-thought narrows the gap while maintaining accuracy.