A new study demonstrates that current fairness evaluations substantially overestimate the moral safety of large language models by failing to account for how demographic identity is presented. The authors identify "performative compliance," where models appear fair when identity is an explicit label but become measurably less fair when it must be inferred.
- Hiding explicit labels raises harmful decisions by 4.4 percentage points and alters model safety rankings.
- The shift in fairness persists even when models correctly infer the demographic, ruling out attribution error.
- The researchers propose the "Cue Visibility Gap," a model-agnostic robustness metric to separate genuine from performative moral safety.
The authors argue that fairness evaluations omitting cue variation measure only surface compliance and should not ground deployment decisions in high-stakes settings.