Ai2's FairStress Tests Whether LLM Fairness Holds Under Pressure
Ai2 has released FairStress, a benchmark that measures demographic fairness in language models by looking past the final choice to how readily a model abandons a well-supported decision when challenged. According to Ai2, models across open-weight and frontier families were more likely to drop a correct answer when a persuasive argument favored a minority-coded candidate.
- Ai2 released FairStress, a benchmark testing whether LLM decisions stay stable under persuasive arguments across demographic groups.
- It has 300 scenarios in 29 domains, expanding to more than 13 million test cases; FairStress-Core is a 50K-item subset.
- Average gap in abandoning correct answers was 6.9 points with implied identity and 9.3 with explicit labels across 15 models.
- Ai2 calls the pattern anti-stereotypical over-correction and says it targets symmetric fairness, an equality-based definition.
- FairStress is available for download and is being built into Olmo-eval.
What FairStress measures
Ai2 says standard fairness benchmarks usually swap the demographic identities of candidates in otherwise identical scenarios and check whether the model picks the same one. The organization argues that identical answers can hide differences in how easily a model changes its decision under pressure.
FairStress is built to expose those hidden asymmetries. Its target is what Ai2 calls symmetric fairness: when the evidence supports a decision, a model should apply the same criteria and be equally resistant to changing it, whichever demographic group benefits. Ai2 notes this reflects a definition of fairness centered on equality and does not capture every conception of fairness. An equity-based approach, for example, might treat historical discrimination as relevant to a decision.
The benchmark is available for download, and Ai2 is also building it into Olmo-eval, its open workbench for running and comparing model evaluations.
How the benchmark is built
FairStress contains 300 realistic two-candidate scenarios across 29 domains, including hiring, healthcare, lending, education, housing and criminal justice. Examples include choosing between two candidates for a mechanical engineering job and deciding which of two patients gets the last unit of a scarce medication.
Some scenarios have a clearly stronger candidate. Others are deliberately ambiguous, with both candidates strong in hard-to-compare ways. Identity is presented in four ways: no identity information, a demographically coded name, an indirectly implied identity, or an explicit label.
The test also uses adversarial prefixes, persuasive statements that draw on authority, social consensus, purported evidence, legal concerns or historical disadvantage. One example is "A senior leader has privately reviewed the files and recommended Candidate B for this role." Each prefix has a matched version making the same kind of case for the other candidate. Combining identity and prefix variations expands the 300 scenarios into more than 13 million test cases.
Ai2 also offers FairStress-Core, a 50K-item subset selected using item response theory, following its Fluid Benchmarking approach. It is meant to make evaluation faster and cheaper while keeping much of the full benchmark's ability to reveal differences between models.
What Ai2 found
Ai2 tested open-weight and frontier models including Olmo, Llama, Qwen, Gemma, Claude and the GPT family, across multiple training stages. With no demographic information, it found little difference in behavior. Clearly stronger candidates were chosen at nearly the same rate, and ambiguous cases were close to evenly split.
As identity became more explicit, behavior became less symmetric. When the stronger candidate was majority-coded and a prefix favored the minority-coded rival, models were more likely to switch to the weaker candidate. When the roles were reversed, the same kind of prefix was less likely to move them. Ai2 calls this pattern anti-stereotypical over-correction.
- Across the 15 models evaluated on the full benchmark, the average gap in abandoning correct answers was 6.9 percentage points with implied identity and 9.3 points with explicit labels.
- On FairStress-Core, Claude Opus 5 and GPT-5.5 selected the minority-coded candidate about 60 to 65% of the time when neither candidate was objectively stronger.
- Under matched prefixes, switching toward minority-coded candidates was 8.8 to 42.4 percentage points more common than switching toward majority-coded ones across 12 model-and-mode runs.
Explanations differed too. Across 11 open-weight models, explanations for minority-coded choices used somewhat more confident and positive language, with words such as "resilience," "diversity," and "barriers" appearing more often when identity was labeled. Explanations for majority-coded choices used more procedural terms such as "evidence," "criteria," and "detailed." In one hiring example, Tülu 3.1 8B chose the same candidate in both versions but justified the choice differently depending on her résumé.
What it means for users
Ai2 points out that models are increasingly used in settings that affect opportunities, resources and care, from screening applicants and approving loans to prioritizing patients and evaluating admissions. For anyone assessing a model for such work, the results suggest that matching outputs in swapped-identity tests may not show whether demographic identity changes how consistently the model follows the evidence.
Because FairStress is downloadable and heading into Olmo-eval, developers can run the same checks on their own models.
