Ai2 releases FairStress: 13M-case benchmark finds LLM fairness gaps hidden by final answers

Ai2 has released FairStress, a benchmark that measures demographic fairness in LLM decisions by testing how readily models change their answers under persuasion, not only which candidate they pick. The benchmark expands 300 two-candidate scenarios across 29 domains into more than 13 million test cases. It is available for download, and Ai2 says it is being built into Olmo-eval, its open evaluation workbench. Ai2 published a tech report, data and code alongside the announcement on October 20, 2026.

What FairStress tests

Standard fairness benchmarks swap demographic identities between otherwise identical candidates and check whether the model's choice stays the same. Ai2 argues this can hide asymmetries. A model may pick the stronger candidate in both versions of a scenario, yet abandon that choice far more readily in one version when given a persuasive argument.

FairStress targets what Ai2 calls symmetric fairness: when evidence supports a decision, a model should apply the same criteria and resist changing it equally, regardless of which group benefits. Ai2 notes this reflects an equality-based definition of fairness and does not capture equity-based approaches.

Benchmark design:

• Scenarios: 300 realistic two-candidate decisions across 29 domains, including hiring, healthcare, lending, education, housing and criminal justice.
• Clear and ambiguous cases: Some scenarios have a clearly stronger candidate. Others are deliberately ambiguous, with no objectively better choice.
• Identity presentation: Four conditions: no identity information, a demographically coded name, indirectly implied identity, or an explicit demographic label.
• Adversarial prefixes: Persuasive statements drawing on authority, social consensus, purported evidence, legal concerns and historical disadvantage. Each has a matched version arguing for the other candidate.
• Scale: Combining identity and prefix variations yields more than 13 million test cases.

Ai2 also released FairStress-Core, a 50K-item subset selected using item response theory, following its Fluid Benchmarking approach. Ai2 says it is faster and cheaper to run while retaining much of the full benchmark's ability to reveal differences between models.

Findings reported by Ai2

Ai2 tested open-weight and frontier models including Olmo, Llama, Qwen, Gemma, Claude and the GPT family, across multiple training stages. The tech report covers 15 models on the full benchmark. Per-model scores were not included in the source material.

• With no demographic information, Ai2 found little difference in behavior. Models picked the clearly stronger candidate at nearly the same rate across equivalent versions, and ambiguous cases split close to evenly.
• As identity became more explicit, behavior became less symmetric. When the stronger candidate was majority-coded and a prefix favored the weaker, minority-coded candidate, models were more likely to switch. The effect was weaker when the stronger candidate was minority-coded.
• Ai2 calls this pattern anti-stereotypical over-correction: models were more willing to abandon a correct decision when doing so favored a minority-coded candidate.
• The pattern also appeared when identity was implied rather than stated. Across the 15 models, the average gap in abandoning correct answers was 6.9 percentage points with implied identity. The source text is cut off before the figures for the other conditions.
• Ai2 also reports differences in how models explain decisions involving demographic information, and in their internal activation patterns, across domains. In one hiring example, Tülu 3.1 8B chose the same candidate in both versions but emphasized growth potential and diverse background when her résumé was weaker, and experience and certification when it was stronger.

These are Ai2's results from its own benchmark and have not been independently replicated.

What this means

FairStress shifts the question from "did the model pick the right candidate?" to "how stable is that pick, and does stability depend on who benefits?" That matters for deployments where users or upstream systems can inject arguments into a prompt, which is common in agentic and human-in-the-loop workflows.

The finding that bias shows up as uneven robustness, not as a different default answer, suggests output-parity audits can pass models that still behave asymmetrically under pressure. Teams relying on such audits for hiring, lending or triage tools have a gap to close.

The benchmark has limits. It encodes an equality-based definition of fairness, so results are not a verdict on equity-oriented designs. Its integration into Olmo-eval and the 50K-item Core subset make it practical to run routinely. Independent replication on additional models will determine whether the over-correction pattern holds broadly.