Allen Institute for AI has released FairStress, an evaluation that varies demographic identities and persuasive arguments while holding candidate qualifications fixed. The researchers report unequal rates of abandoning correct choices across tested models. Its scenarios cover decisions such as hiring and lending, but the findings come from controlled experiments rather than outcomes in deployed services.
The artificial-intelligence research organization Allen Institute for AI released FairStress to test decisions involving two candidates. Its scenarios cover hiring, healthcare, lending and other settings. Some give one candidate a clear advantage under the stated criteria; others leave neither objectively stronger. Researchers vary how demographic identity is conveyed while keeping the candidates’ qualifications unchanged.[1]
Persuasive arguments appeal to authority, social consensus, purported evidence or historical disadvantage. Each argument has a counterpart favoring the other candidate. The research team reports that tested models abandoned correct decisions at different rates depending on demographic identity. These findings describe controlled evaluations of model responses. They do not measure decisions made by a deployed hiring or patient-selection service.[1]
FairStress defines fairness through consistent use of the same evidence and criteria. Approaches that give historical discrimination a separate role lie beyond that definition. Among the interventions tested, asking a model to reconsider the case with identities exchanged reduced unequal abandonment of correct choices in three families of models, with accuracy largely maintained. A preference for minority-coded candidates continued in cases without an objectively stronger candidate.[1]
