@inproceedings{luo-etal-2025-red,
title = "Red-Teaming for Uncovering Societal Bias in Large Language Models",
author = "Luo, Chu Fei and
Ghawanmeh, Ahmad and
Coimbatore Murali, Kashyap and
Kumar, Bhimshetty Bharat and
Jadhav, Murli and
Zhu, Xiaodan and
Khan Khattak, Faiza",
editor = "Calabrese, Agostina and
de Kock, Christine and
Nozza, Debora and
Plaza-del-Arco, Flor Miriam and
Talat, Zeerak and
Vargas, Francielle",
booktitle = "Proceedings of the The 9th Workshop on Online Abuse and Harms (WOAH)",
month = aug,
year = "2025",
address = "Vienna, Austria",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.woah-1.46/",
pages = "522--537",
ISBN = "979-8-89176-105-6",
abstract = "Ensuring the safe deployment of AI systems is critical in industry settings where biased outputs can lead to significant operational, reputational, and regulatory risks. Thorough evaluation before deployment is essential to prevent these hazards. Red-teaming addresses this need by employing adversarial attacks to develop guardrails that detect and reject biased or harmful queries, enabling models to be retrained or steered away from harmful outputs. However, red-teaming techniques are often limited, and malicious actors may discover new vulnerabilities that bypass safety fine-tuning, underscoring the need for ongoing research and innovative approaches. Notably, most red-teaming efforts focus on harmful or unethical instructions rather than addressing social bias, leaving this critical area under-explored despite its significant real-world impact, especially in customer-facing AI systems. We propose two bias-specific red-teaming methods, Emotional Bias Probe (EBP) and BiasKG, to evaluate how standard safety measures for harmful content mitigate bias. For BiasKG, we refactor natural language stereotypes into a knowledge graph. and use adversarial attacking strategies to induce biased responses from several open- and closed-source language models. We find our method increases bias in all models, even those trained with safety guardrails. Our work emphasizes uncovering societal bias in LLMs through rigorous evaluation, addressing adversarial challenges to ensure AI safety in high-stakes industry deployments."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="luo-etal-2025-red">
<titleInfo>
<title>Red-Teaming for Uncovering Societal Bias in Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Chu</namePart>
<namePart type="given">Fei</namePart>
<namePart type="family">Luo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ahmad</namePart>
<namePart type="family">Ghawanmeh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kashyap</namePart>
<namePart type="family">Coimbatore Murali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bhimshetty</namePart>
<namePart type="given">Bharat</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Murli</namePart>
<namePart type="family">Jadhav</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaodan</namePart>
<namePart type="family">Zhu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Faiza</namePart>
<namePart type="family">Khan Khattak</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the The 9th Workshop on Online Abuse and Harms (WOAH)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Agostina</namePart>
<namePart type="family">Calabrese</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christine</namePart>
<namePart type="family">de Kock</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Debora</namePart>
<namePart type="family">Nozza</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Flor</namePart>
<namePart type="given">Miriam</namePart>
<namePart type="family">Plaza-del-Arco</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeerak</namePart>
<namePart type="family">Talat</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Francielle</namePart>
<namePart type="family">Vargas</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Vienna, Austria</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-105-6</identifier>
</relatedItem>
<abstract>Ensuring the safe deployment of AI systems is critical in industry settings where biased outputs can lead to significant operational, reputational, and regulatory risks. Thorough evaluation before deployment is essential to prevent these hazards. Red-teaming addresses this need by employing adversarial attacks to develop guardrails that detect and reject biased or harmful queries, enabling models to be retrained or steered away from harmful outputs. However, red-teaming techniques are often limited, and malicious actors may discover new vulnerabilities that bypass safety fine-tuning, underscoring the need for ongoing research and innovative approaches. Notably, most red-teaming efforts focus on harmful or unethical instructions rather than addressing social bias, leaving this critical area under-explored despite its significant real-world impact, especially in customer-facing AI systems. We propose two bias-specific red-teaming methods, Emotional Bias Probe (EBP) and BiasKG, to evaluate how standard safety measures for harmful content mitigate bias. For BiasKG, we refactor natural language stereotypes into a knowledge graph. and use adversarial attacking strategies to induce biased responses from several open- and closed-source language models. We find our method increases bias in all models, even those trained with safety guardrails. Our work emphasizes uncovering societal bias in LLMs through rigorous evaluation, addressing adversarial challenges to ensure AI safety in high-stakes industry deployments.</abstract>
<identifier type="citekey">luo-etal-2025-red</identifier>
<location>
<url>https://aclanthology.org/2025.woah-1.46/</url>
</location>
<part>
<date>2025-08</date>
<extent unit="page">
<start>522</start>
<end>537</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Red-Teaming for Uncovering Societal Bias in Large Language Models
%A Luo, Chu Fei
%A Ghawanmeh, Ahmad
%A Coimbatore Murali, Kashyap
%A Kumar, Bhimshetty Bharat
%A Jadhav, Murli
%A Zhu, Xiaodan
%A Khan Khattak, Faiza
%Y Calabrese, Agostina
%Y de Kock, Christine
%Y Nozza, Debora
%Y Plaza-del-Arco, Flor Miriam
%Y Talat, Zeerak
%Y Vargas, Francielle
%S Proceedings of the The 9th Workshop on Online Abuse and Harms (WOAH)
%D 2025
%8 August
%I Association for Computational Linguistics
%C Vienna, Austria
%@ 979-8-89176-105-6
%F luo-etal-2025-red
%X Ensuring the safe deployment of AI systems is critical in industry settings where biased outputs can lead to significant operational, reputational, and regulatory risks. Thorough evaluation before deployment is essential to prevent these hazards. Red-teaming addresses this need by employing adversarial attacks to develop guardrails that detect and reject biased or harmful queries, enabling models to be retrained or steered away from harmful outputs. However, red-teaming techniques are often limited, and malicious actors may discover new vulnerabilities that bypass safety fine-tuning, underscoring the need for ongoing research and innovative approaches. Notably, most red-teaming efforts focus on harmful or unethical instructions rather than addressing social bias, leaving this critical area under-explored despite its significant real-world impact, especially in customer-facing AI systems. We propose two bias-specific red-teaming methods, Emotional Bias Probe (EBP) and BiasKG, to evaluate how standard safety measures for harmful content mitigate bias. For BiasKG, we refactor natural language stereotypes into a knowledge graph. and use adversarial attacking strategies to induce biased responses from several open- and closed-source language models. We find our method increases bias in all models, even those trained with safety guardrails. Our work emphasizes uncovering societal bias in LLMs through rigorous evaluation, addressing adversarial challenges to ensure AI safety in high-stakes industry deployments.
%U https://aclanthology.org/2025.woah-1.46/
%P 522-537
Markdown (Informal)
[Red-Teaming for Uncovering Societal Bias in Large Language Models](https://aclanthology.org/2025.woah-1.46/) (Luo et al., WOAH 2025)
ACL
- Chu Fei Luo, Ahmad Ghawanmeh, Kashyap Coimbatore Murali, Bhimshetty Bharat Kumar, Murli Jadhav, Xiaodan Zhu, and Faiza Khan Khattak. 2025. Red-Teaming for Uncovering Societal Bias in Large Language Models. In Proceedings of the The 9th Workshop on Online Abuse and Harms (WOAH), pages 522–537, Vienna, Austria. Association for Computational Linguistics.