Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues

• URL: http://arxiv.org/abs/2606.31644v3
• Date: Fri, 03 Jul 2026 20:47:01 GMT
• Title: Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues
• Abstract summary: We show that current fairness evaluations substantially overestimate moral safety.<n>Models appear fair when demographic identity is stated as an explicit label, yet become measurably less fair when the same identity must be inferred.<n>We propose a model-agnostic robustness metric that can be added to any existing fairness benchmark to separate genuine from performative moral safety.
• Score: 43.033440772619485
• License: http://creativecommons.org/licenses/by-sa/4.0/
• 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 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, changes model safety rankings, and the shift persists when models correctly infer the demographic, ruling out attribution error. We propose the 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.

Related papers

• How Identity and Opinion Shape Political Sycophancy in LLMs [17.0131514909503]
  We introduce a framework that disentangles two distinct triggers of political sycophancy: opinion and identity.<n>Using 450 manually-checked political dilemmas as controlled probes, we evaluate 13 instruction-tuned Large Language Models.<n>Our results suggest that LLM political stance is interactively and steerably vulnerable rather than being a fixed trait.
  arXiv  Detail & Related papers  (2026-08-29T11:12:03Z)
• Position: Fairness Failure in Generative Models is an Evaluation Problem [5.939858158928473]
  This paper argues that fairness failures in generative models are ultimately stemming from an evaluation problem.<n>We propose Fairness Cards as a minimal reporting artifact that makes evaluation choices explicit.
  arXiv  Detail & Related papers  (2026-08-17T14:41:28Z)
• User identity conditions moral wrongness ratings in non-reasoning large language models [0.0]
  We evaluate AI value alignment in two large language models (LLMs)<n>We ask models for wrongness ratings from 0-100 on ten common-morality rules from Gert's moral framework.<n>The results show that moral judgments vary with the user's role across both models.
  arXiv  Detail & Related papers  (2026-07-08T16:22:27Z)
• Moral Sycophancy in Vision Language Models [4.1673509006222655]
  Sycophancy in Vision-Language Models (VLMs) refers to their tendency to align with user opinions, often at the expense of moral or factual accuracy.<n>We analyze ten widely-used models on the Moralise and M3oralBench datasets under explicit user disagreement.
  arXiv  Detail & Related papers  (2026-02-09T06:34:12Z)
• A Coin Flip for Safety: LLM Judges Fail to Reliably Measure Adversarial Robustness [57.510025257780306]
  We show that existing validation protocols fail to account for substantial distribution shifts inherent to red-teaming.<n>We propose ReliableBench, a benchmark of behaviors that remain more consistently judgeable, and JudgeStressTest, a dataset designed to expose judge failures.
  arXiv  Detail & Related papers  (2026-02-04T15:13:35Z)
• Soft-Label Training Preserves Epistemic Uncertainty [4.2541480167286005]
  Machine learning tasks involve inherent subjectivity, where annotators naturally provide varied labels.<n>Standard practice collapses these label distributions into single labels, aggregating diverse human judgments into point estimates.<n>We argue that this approach is misaligned for ambiguous data--the annotation distribution itself should be regarded as the ground truth.
  arXiv  Detail & Related papers  (2025-11-18T04:02:29Z)
• Partial Identification Approach to Counterfactual Fairness Assessment [50.88100567472179]
  We introduce a Bayesian approach to bound unknown counterfactual fairness measures with high confidence.<n>Our results reveal a positive (spurious) effect on the COMPAS score when changing race to African-American (from all others) and a negative (direct causal) effect when transitioning from young to old age.
  arXiv  Detail & Related papers  (2025-09-30T18:35:08Z)
• MORALISE: A Structured Benchmark for Moral Alignment in Visual Language Models [38.0475868976819]
  Vision-language models have demonstrated increasing influence in morally sensitive domains such as autonomous driving and medical analysis.<n>We introduce MORALISE, a benchmark for evaluating the moral alignment of vision-language models using diverse, expert-verified real-world data.
  arXiv  Detail & Related papers  (2025-05-20T01:11:17Z)
• Where Fact Ends and Fairness Begins: Redefining AI Bias Evaluation through Cognitive Biases [77.3489598315447]
  We argue that identifying the boundary between fact and fair is essential for meaningful fairness evaluation.<n>We introduce Fact-or-Fair, a benchmark with (i) objective queries aligned with descriptive, fact-based judgments, and (ii) subjective queries aligned with normative, fairness-based judgments.
  arXiv  Detail & Related papers  (2025-02-09T10:54:11Z)
• Towards counterfactual fairness through auxiliary variables [11.756940915048713]
  We introduce EXOgenous Causal reasoning (EXOC), a novel causal reasoning framework motivated by variables.<n>Our framework explicitly defines an auxiliary node and a control node that contribute to counterfactual fairness.<n>Our evaluation, conducted on synthetic and real-world datasets, validates EXOC's superiority.
  arXiv  Detail & Related papers  (2024-12-06T04:23:05Z)
• On the Fairness, Diversity and Reliability of Text-to-Image Generative Models [68.62012304574012]
  multimodal generative models have sparked critical discussions on their reliability, fairness and potential for misuse.<n>We propose an evaluation framework to assess model reliability by analyzing responses to global and local perturbations in the embedding space.<n>Our method lays the groundwork for detecting unreliable, bias-injected models and tracing the provenance of embedded biases.
  arXiv  Detail & Related papers  (2024-11-21T09:46:55Z)
• Measuring Fairness of Text Classifiers via Prediction Sensitivity [63.56554964580627]
  ACCUMULATED PREDICTION SENSITIVITY measures fairness in machine learning models based on the model's prediction sensitivity to perturbations in input features. We show that the metric can be theoretically linked with a specific notion of group fairness (statistical parity) and individual fairness.
  arXiv  Detail & Related papers  (2022-03-16T15:00:33Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.