@inproceedings{shafiei-etal-2026-moral,
title = "Moral Safety in {LLM}s: Exposing Performative Compliance with Puzzled Cues",
author = "Shafiei, Mohammadamin and
Li, Shuyue Stella and
Tsvetkov, Yulia",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1994/",
doi = "10.18653/2026.findings-emnlp.1994",
pages = "42485--42516",
ISBN = "979-8-89176-462-0",
abstract = "As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, verifying that their ethical behavior is genuine rather than superficial becomes essential. We show that current fairness evaluations substantially overestimate moral safety. Models judged fair when demographic identity is explicitly labeled become measurably less fair when the same identity must be inferred {---} a pattern we term performative compliance: safety behavior contingent on surface-level evaluation signals rather than internalized fairness principles. We introduce a cue-variation methodology that holds moral content fixed while varying only whether demographic identity is stated or must be inferred, and apply it to 14 models across 100 everyday dilemmas spanning 3 genders and 5 races. Hiding the label leaves decisions that benefit the described individual unchanged (+0.9pp) while raising decisions that harm them by +4.4pp. This asymmetry persists even when models correctly infer the demographic, ruling out attribution error. We propose the Direct vs Puzzled gap as a model-agnostic robustness metric that requires no new annotation, changes model safety rankings, and can be adopted by any existing fairness evaluation to separate genuine from performative moral safety. We argue that fairness evaluations that omit cue-variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="shafiei-etal-2026-moral">
<titleInfo>
<title>Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues</title>
</titleInfo>
<name type="personal">
<namePart type="given">Mohammadamin</namePart>
<namePart type="family">Shafiei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shuyue</namePart>
<namePart type="given">Stella</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yulia</namePart>
<namePart type="family">Tsvetkov</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: EMNLP 2026</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-462-0</identifier>
</relatedItem>
<abstract>As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, verifying that their ethical behavior is genuine rather than superficial becomes essential. We show that current fairness evaluations substantially overestimate moral safety. Models judged fair when demographic identity is explicitly labeled become measurably less fair when the same identity must be inferred — a pattern we term performative compliance: safety behavior contingent on surface-level evaluation signals rather than internalized fairness principles. We introduce a cue-variation methodology that holds moral content fixed while varying only whether demographic identity is stated or must be inferred, and apply it to 14 models across 100 everyday dilemmas spanning 3 genders and 5 races. Hiding the label leaves decisions that benefit the described individual unchanged (+0.9pp) while raising decisions that harm them by +4.4pp. This asymmetry persists even when models correctly infer the demographic, ruling out attribution error. We propose the Direct vs Puzzled gap as a model-agnostic robustness metric that requires no new annotation, changes model safety rankings, and can be adopted by any existing fairness evaluation to separate genuine from performative moral safety. We argue that fairness evaluations that omit cue-variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings.</abstract>
<identifier type="citekey">shafiei-etal-2026-moral</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1994</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1994/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>42485</start>
<end>42516</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues
%A Shafiei, Mohammadamin
%A Li, Shuyue Stella
%A Tsvetkov, Yulia
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F shafiei-etal-2026-moral
%X As large language models take on morally consequential roles in healthcare, legal, and hiring contexts, verifying that their ethical behavior is genuine rather than superficial becomes essential. We show that current fairness evaluations substantially overestimate moral safety. Models judged fair when demographic identity is explicitly labeled become measurably less fair when the same identity must be inferred — a pattern we term performative compliance: safety behavior contingent on surface-level evaluation signals rather than internalized fairness principles. We introduce a cue-variation methodology that holds moral content fixed while varying only whether demographic identity is stated or must be inferred, and apply it to 14 models across 100 everyday dilemmas spanning 3 genders and 5 races. Hiding the label leaves decisions that benefit the described individual unchanged (+0.9pp) while raising decisions that harm them by +4.4pp. This asymmetry persists even when models correctly infer the demographic, ruling out attribution error. We propose the Direct vs Puzzled gap as a model-agnostic robustness metric that requires no new annotation, changes model safety rankings, and can be adopted by any existing fairness evaluation to separate genuine from performative moral safety. We argue that fairness evaluations that omit cue-variation measure surface compliance, not moral robustness, and should not ground deployment decisions in high-stakes settings.
%R 10.18653/2026.findings-emnlp.1994
%U https://aclanthology.org/2026.findings-emnlp.1994/
%U https://doi.org/10.18653/2026.findings-emnlp.1994
%P 42485-42516
Markdown (Informal)
[Moral Safety in LLMs: Exposing Performative Compliance with Puzzled Cues](https://aclanthology.org/2026.findings-emnlp.1994/) (Shafiei et al., Findings 2026)
ACL