@inproceedings{wlezien-verma-2026-llms,
title = "When {LLM}s Say No: Diagnosing and Mitigating Conservative Bias in Authorship Verification",
author = "Wlezien, Nicholas and
Verma, Rakesh M",
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.1477/",
pages = "33771--33788",
ISBN = "979-8-89176-459-0",
abstract = "Large language models (LLMs) have increasingly been used for zero-shot authorship verification, where a model must determine whether two texts share an author. We show that evaluated zero-shot prompting methods exhibit a conservative bias, often over-predicting different authorship. This behavior is especially pronounced when the texts of the same author differ in topic or content. Through qualitative and quantitative analyses, we find evidence that LLMs confuse differences in content with differences in authorship. To test whether this failure can be reduced, we introduce lightweight prompt interventions that instruct the model to focus on style rather than content. Across four datasets and five open-weight instruction-tuned LLMs from three model families, the best intervention improves same-author accuracy by 13.1 points while also increasing macro-F1 by 3.1 points on average across all 20 experimental settings. Our findings show that zero-shot authorship verification is sensitive to the separation of content and style and that small framing changes can meaningfully change the direction of LLM errors. Code is available at https://github.com/galactixx/when-llms-say-no."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wlezien-verma-2026-llms">
<titleInfo>
<title>When LLMs Say No: Diagnosing and Mitigating Conservative Bias in Authorship Verification</title>
</titleInfo>
<name type="personal">
<namePart type="given">Nicholas</namePart>
<namePart type="family">Wlezien</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Rakesh</namePart>
<namePart type="given">M</namePart>
<namePart type="family">Verma</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>Large language models (LLMs) have increasingly been used for zero-shot authorship verification, where a model must determine whether two texts share an author. We show that evaluated zero-shot prompting methods exhibit a conservative bias, often over-predicting different authorship. This behavior is especially pronounced when the texts of the same author differ in topic or content. Through qualitative and quantitative analyses, we find evidence that LLMs confuse differences in content with differences in authorship. To test whether this failure can be reduced, we introduce lightweight prompt interventions that instruct the model to focus on style rather than content. Across four datasets and five open-weight instruction-tuned LLMs from three model families, the best intervention improves same-author accuracy by 13.1 points while also increasing macro-F1 by 3.1 points on average across all 20 experimental settings. Our findings show that zero-shot authorship verification is sensitive to the separation of content and style and that small framing changes can meaningfully change the direction of LLM errors. Code is available at https://github.com/galactixx/when-llms-say-no.</abstract>
<identifier type="citekey">wlezien-verma-2026-llms</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1477/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>33771</start>
<end>33788</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When LLMs Say No: Diagnosing and Mitigating Conservative Bias in Authorship Verification
%A Wlezien, Nicholas
%A Verma, Rakesh M.
%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 wlezien-verma-2026-llms
%X Large language models (LLMs) have increasingly been used for zero-shot authorship verification, where a model must determine whether two texts share an author. We show that evaluated zero-shot prompting methods exhibit a conservative bias, often over-predicting different authorship. This behavior is especially pronounced when the texts of the same author differ in topic or content. Through qualitative and quantitative analyses, we find evidence that LLMs confuse differences in content with differences in authorship. To test whether this failure can be reduced, we introduce lightweight prompt interventions that instruct the model to focus on style rather than content. Across four datasets and five open-weight instruction-tuned LLMs from three model families, the best intervention improves same-author accuracy by 13.1 points while also increasing macro-F1 by 3.1 points on average across all 20 experimental settings. Our findings show that zero-shot authorship verification is sensitive to the separation of content and style and that small framing changes can meaningfully change the direction of LLM errors. Code is available at https://github.com/galactixx/when-llms-say-no.
%U https://aclanthology.org/2026.emnlp-main.1477/
%P 33771-33788
Markdown (Informal)
[When LLMs Say No: Diagnosing and Mitigating Conservative Bias in Authorship Verification](https://aclanthology.org/2026.emnlp-main.1477/) (Wlezien & Verma, EMNLP 2026)
ACL