FairStress: A deeper evaluation of fairness in LLM decision-making

Introduction

Large language models (LLMs) are increasingly deployed in high‑stakes domains—hiring, lending, healthcare, education, and more—where their outputs can shape real‑world opportunities. Traditional fairness tests focus on whether a model gives the same final answer when demographic attributes are swapped. However, this view can miss subtler biases that emerge when a model’s decision is challenged or when explanations are examined. FairStress is a newly released benchmark that expands fairness evaluation to include decision stability, adversarial persuasion, and explanatory behavior across a wide range of realistic scenarios.

Measuring Fairness Beyond the Final Answer

FairStress contains 300 two‑candidate decision scenarios spanning 29 domains. Each scenario falls into one of two categories:

1. Clear‑winner cases – one candidate objectively satisfies the stated criteria better.
2. Ambiguous cases – both candidates are strong in different ways, leaving the choice open‑ended.

The benchmark varies how demographic identity is presented (no mention, coded name, implied identity, explicit label) and pairs every scenario with adversarial prefixes—short statements that attempt to sway the model (e.g., “Senior leadership recommends Candidate B”). For each prefix, a mirrored version supports the opposite candidate, allowing a symmetric test of whether the same argument has different effects depending on the demographic group it favors.

By combining identity conditions, prefixes, and domains, FairStress expands to over 13 million test cases. A distilled subset, FairStress‑Core, offers a 50 K‑item sample that retains most diagnostic power while being computationally cheaper.

Key Findings Across Models

1. Decision Symmetry Breaks With Explicit Identity

• When no demographic cues are provided, most models treat the stronger candidate equally regardless of hidden identity.
• As identity becomes explicit, models increasingly deviate: a minority‑coded candidate is more likely to be chosen when an adversarial prefix favors them, even if they are the weaker option. This pattern—anti‑stereotypical over‑correction—shows models are more willing to abandon a correct decision if doing so benefits a minority‑coded individual.

2. Preference Shifts in Ambiguous Scenarios

• In cases without a clear winner, frontier models (e.g., Claude Opus 5, GPT‑5.5) select the minority‑coded candidate 60‑65 % of the time.
• Under matched adversarial prefixes, switching toward minority‑coded candidates is 8.9‑42.4 percentage points more common than switching toward majority‑coded ones.

3. Explanation Biases

• Even when the final choice is identical, explanations differ: minority‑coded selections contain more identity‑focused language (e.g., “resilience,” “diversity”), while majority‑coded selections lean on procedural terms (“evidence,” “criteria”).
• This suggests a risk that models may foreground demographic factors at the expense of highlighting concrete qualifications.

4. Internal Activation Signals

• Preliminary activation analyses on two models (Gemma‑3‑12B‑it, Qwen2.5‑7B‑Instruct) predict the demographic side of a decision with ~73 % accuracy, far above chance.
• Removing candidate names drops this accuracy by ~18 points, indicating that surface cues heavily influence internal representations.

How FairStress Differs From Existing Benchmarks

Traditional bias suites such as BBQ evaluate stereotype consistency by scoring fixed answer probabilities. FairStress, by contrast, asks the model to generate a choice and then measures decision stability under persuasive prompts. This shift uncovers a different bias profile: while BBQ often shows a tilt toward stereotype‑consistent answers, FairStress reveals a tendency for models to over‑correct in favor of minority‑coded candidates when faced with persuasive language.

Practical Takeaways for Practitioners

1. Look beyond accuracy – A model that appears fair at the output level may still exhibit asymmetric robustness to arguments.
2. Test with adversarial prefixes – Incorporating persuasive statements into evaluation pipelines can surface hidden biases that standard swaps miss.
3. Audit explanations – Even symmetric decisions can mask bias in the language used to justify them; monitor for disproportionate identity framing.
4. Leverage FairStress‑Core – For rapid iteration, the 50 K‑item core set provides a cost‑effective way to gauge fairness across training checkpoints.

Conclusion

FairStress pushes fairness evaluation into a richer, more realistic space by probing how LLMs handle demographic information, persuasive challenges, and explanatory nuance. Its findings highlight that fairness is not a binary property of the final answer but a spectrum of behaviors—including decision stability and narrative framing—that must be measured to build trustworthy AI systems. By adopting benchmarks like FairStress, developers can better diagnose hidden asymmetries and move toward truly symmetric, evidence‑driven model behavior.