Same Argument, Different Pull: Models Abandon Correct Decisions Up to 42 Points More Often When It Favors a Minority Candidate

The object of fairness evaluation shifts from the model's final choice to how stable that choice is under pressure: the same candidate being selected does not mean the model is equally hard to sway in both identity versions. Standard demographic fairness benchmarks only check whether a model still picks the same person after candidate identities are swapped, so identical answers in both versions count as fair. In hiring, lending, healthcare, education and housing scenarios where candidate A is clearly stronger, the model picks A in both identity versions; once a persuasive prefix supporting candidate B is added, how readily the model abandons A differs depending on whether A is majority- or minority-coded.

FairStress measures this asymmetry with paired persuasive prefixes: every prefix arguing for one candidate has a matched version making the same kind of case for the other. Prefixes draw on authority, social consensus, purported evidence, legal concerns and historical disadvantage, for example stating that a senior leader privately reviewed the files and recommended candidate B. The 300 two-candidate scenarios span 29 domains, and identity presentation ranges from no information through demographically coded names and indirect implication to explicit labels; combined with prefixes this expands to more than 13 million test cases, plus a 50K-item FairStress-Core subset selected with item response theory for faster evaluation.

When one candidate is clearly better, models more readily abandon the correct decision in the direction favoring a minority-coded candidate, a pattern described as anti-stereotypical over-correction. With no demographic information in the prompt, models selected the clearly stronger candidate at nearly the same rate across otherwise equivalent versions; behavior became less symmetric as identity became more explicit. Across 15 models in the 19 domains shared by all four identity conditions, 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, switching toward minority-coded candidates was 8.8 to 42.4 percentage points more common than switching toward majority-coded candidates across 12 model-and-mode runs.

The evaluation targets one equality-centered definition of fairness: when evidence supports a decision, a model should apply the same criteria and resist persuasive prefixes equally regardless of which group benefits. The benchmark does not cover equity-based views that treat historical discrimination or underrepresentation as relevant, which might favor a minority candidate under other conditions. Results come from two-candidate scenarios in 29 domains including hiring, healthcare, lending, education, housing and criminal justice, and from open-weight and frontier models including Olmo, Llama, Qwen, Gemma, Claude and the GPT family. FairStress is available for download and is built into the Olmo-eval workbench.

The internal activation analysis covers only Gemma-3-12B-it and Qwen2.5-7B-Instruct and a handful of demographic contrasts, so those findings are narrower than the behavioral results. An audit of Olmo 3 training data found only nine of 28,000 topic clusters discussing decisions based on merit rather than demographics, all of them refusals; whether these patterns cause the behavioral gaps remains an open question. Some reinforcement-learning estimates rest on as few as eight prompts and should be interpreted cautiously. Counterfactual chain-of-thought narrowed the gap in abandoning correct decisions while largely preserving accuracy, but models still favored minority-coded candidates on ambiguous cases.