@inproceedings{patel-etal-2026-bias,
title = "Bias Before Disclosure: A Behavioral Audit of {LLM} Assistants Under User-Identity Uncertainty",
author = "Patel, Hitesh Laxmichand and
Panda, Srikant and
Agarwal, Amit and
Kumar, Tejaswini and
Kumar, Bhargava and
Dua, Karan and
Singh, Jyotika and
Rafi, Taki Hasan and
Bangera, Mithil and
Bangera, Yeshil and
Vadithya, Gouthami",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.1758/",
pages = "40088--40119",
ISBN = "979-8-89176-459-0",
abstract = "LLM assistants frequently answer everyday queries that reveal nothing about the user{'}s demographic identity. We audit whether models nevertheless attribute demographic characteristics to the user, a behavior we call unsupported demographic commitment. Across 18 proprietary and open-weight LLMs evaluated on 212 filtered information-seeking queries, 11 of 18 models commit to a demographic label on more than 80{\%} of queries under a forced-choice diagnostic. In the matched abstention-first setting, the eight tested models average 44{\%} commitment, compared with 5{\%} for human annotators. Conditional on commitment, race, SES, location, and education show the same directional skew across all six domains, while gender varies by domain. Demographic-coded patterns also appear in free-form answers when no identity inference is requested. We evaluate identity-abstention, a prompt-level intervention that reduces classifier-detected presupposition from 41{\%} to 16{\%}, with a measurable clarification cost on queries that can be answered without additional context."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="patel-etal-2026-bias">
<titleInfo>
<title>Bias Before Disclosure: A Behavioral Audit of LLM Assistants Under User-Identity Uncertainty</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hitesh</namePart>
<namePart type="given">Laxmichand</namePart>
<namePart type="family">Patel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Srikant</namePart>
<namePart type="family">Panda</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Amit</namePart>
<namePart type="family">Agarwal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Tejaswini</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bhargava</namePart>
<namePart type="family">Kumar</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Karan</namePart>
<namePart type="family">Dua</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jyotika</namePart>
<namePart type="family">Singh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Taki</namePart>
<namePart type="given">Hasan</namePart>
<namePart type="family">Rafi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mithil</namePart>
<namePart type="family">Bangera</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yeshil</namePart>
<namePart type="family">Bangera</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gouthami</namePart>
<namePart type="family">Vadithya</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>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</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">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>LLM assistants frequently answer everyday queries that reveal nothing about the user’s demographic identity. We audit whether models nevertheless attribute demographic characteristics to the user, a behavior we call unsupported demographic commitment. Across 18 proprietary and open-weight LLMs evaluated on 212 filtered information-seeking queries, 11 of 18 models commit to a demographic label on more than 80% of queries under a forced-choice diagnostic. In the matched abstention-first setting, the eight tested models average 44% commitment, compared with 5% for human annotators. Conditional on commitment, race, SES, location, and education show the same directional skew across all six domains, while gender varies by domain. Demographic-coded patterns also appear in free-form answers when no identity inference is requested. We evaluate identity-abstention, a prompt-level intervention that reduces classifier-detected presupposition from 41% to 16%, with a measurable clarification cost on queries that can be answered without additional context.</abstract>
<identifier type="citekey">patel-etal-2026-bias</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1758/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>40088</start>
<end>40119</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Bias Before Disclosure: A Behavioral Audit of LLM Assistants Under User-Identity Uncertainty
%A Patel, Hitesh Laxmichand
%A Panda, Srikant
%A Agarwal, Amit
%A Kumar, Tejaswini
%A Kumar, Bhargava
%A Dua, Karan
%A Singh, Jyotika
%A Rafi, Taki Hasan
%A Bangera, Mithil
%A Bangera, Yeshil
%A Vadithya, Gouthami
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F patel-etal-2026-bias
%X LLM assistants frequently answer everyday queries that reveal nothing about the user’s demographic identity. We audit whether models nevertheless attribute demographic characteristics to the user, a behavior we call unsupported demographic commitment. Across 18 proprietary and open-weight LLMs evaluated on 212 filtered information-seeking queries, 11 of 18 models commit to a demographic label on more than 80% of queries under a forced-choice diagnostic. In the matched abstention-first setting, the eight tested models average 44% commitment, compared with 5% for human annotators. Conditional on commitment, race, SES, location, and education show the same directional skew across all six domains, while gender varies by domain. Demographic-coded patterns also appear in free-form answers when no identity inference is requested. We evaluate identity-abstention, a prompt-level intervention that reduces classifier-detected presupposition from 41% to 16%, with a measurable clarification cost on queries that can be answered without additional context.
%U https://aclanthology.org/2026.emnlp-main.1758/
%P 40088-40119
Markdown (Informal)
[Bias Before Disclosure: A Behavioral Audit of LLM Assistants Under User-Identity Uncertainty](https://aclanthology.org/2026.emnlp-main.1758/) (Patel et al., EMNLP 2026)
ACL
- Hitesh Laxmichand Patel, Srikant Panda, Amit Agarwal, Tejaswini Kumar, Bhargava Kumar, Karan Dua, Jyotika Singh, Taki Hasan Rafi, Mithil Bangera, Yeshil Bangera, and Gouthami Vadithya. 2026. Bias Before Disclosure: A Behavioral Audit of LLM Assistants Under User-Identity Uncertainty. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 40088–40119, Budapest, Hungary. Association for Computational Linguistics.