Ai2 releases FairStress, a benchmark testing demographic fairness in LLM decisions, including resistance to persuasive arguments

Ai2 released FairStress, an open benchmark of 300 two-candidate decision scenarios across 29 domains, including hiring, healthcare, lending, education, housing and criminal justice. Varying how demographic identity is presented and adding adversarial prefixes expands these to over 13 million test cases, and a 50K-item subset (FairStress-Core) was chosen using item response theory. It is being added to Ai2's open Olmo-eval workbench. Ai2's own tests on 15 models found that models abandon a correct choice more readily when doing so favours a minority-coded candidate, with an average gap of 6.9 percentage points when identity is implied.

The benchmark tests something simpler checks miss: not just whether an answer changes when demographics are swapped, but how easily a model is talked out of a well-supported answer depending on who benefits. Teams choosing or auditing models for hiring, lending or admissions can run it for free, and regulators gain a concrete test to point to. The impact is still limited to evaluation. It is a research tool, not a deployed change, and nothing here shows how models behave in real products.

The 6.9 point gap and the "anti-stereotypical over-correction" label come from Ai2's own tests on 15 models, with no independent replication yet, and we do not know how the gap varies by model or domain. The 13 million figure is mostly combinations of the same 300 scenarios, so it overstates how much independent variety there is. Two-candidate, scripted scenarios may also not reflect how LLMs are really used in high-stakes decisions.

1. 01Independent replications of the 6.9 point gap on other models, especially the latest closed models, and whether the gap holds when scenarios are reworded.
2. 02Adoption: whether other labs, auditors or model cards report FairStress or FairStress-Core results within the next few months, and whether it appears in third-party eval suites beyond Olmo-eval.
3. 03Whether model developers or regulators cite it or respond with fixes, and whether retesting after mitigation shows the gap closing without hurting accuracy.