Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues

2026-06-30

AI Summary

Current large language models may exhibit only performative compliance—rather than genuine moral robustness—in ethically sensitive scenarios. This work proposes a cue-variation method that, while holding moral dilemmas and demographic identities constant, alters only the linguistic expression of identity to evaluate model fairness. We introduce a model-agnostic metric, the "cue visibility gap," and integrate it with fairness benchmarking and reasoning consistency analysis to demonstrate that existing evaluations overestimate models’ moral safety. Experiments reveal that concealing explicit identity labels increases harmful decisions by 4.4 percentage points—a gap that persists even when models correctly infer the underlying identity—thereby confirming the prevalence of superficial compliance.

Abstract

As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, we need to examine whether their ethical behaviors are genuine or superficial. We show that current fairness evaluations substantially overestimate moral safety. Models appear fair when demographic identity is stated as an explicit label, yet become measurably less fair when the same identity must be inferred. We term this failure \emph{performative compliance}, where a model is fair when the presentation resembles a fairness evaluation and less fair as that cue weakens. We introduce a cue-variation methodology that holds the moral dilemma and the demographic identity fixed and varies only how that identity is conveyed. Hiding the explicit label raises harmful decisions by $+4.4$~pp and changes model safety rankings, and the shift persists when models correctly infer the demographic, ruling out attribution error. We propose the \textbf{Cue Visibility Gap}, a model-agnostic robustness metric that can be added to any existing fairness benchmark to separate genuine from performative moral safety. Fairness evaluations that omit cue variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings.

Problem

• moral safety
• performative compliance
• fairness evaluation
• cue visibility
• large language models

Innovation

• performative compliance
• cue-variation methodology
• Cue Visibility Gap
• moral safety
• fairness robustness