@inproceedings{zahraei-etal-2026-alignment,
title = "The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in {LLM}s",
author = {Zahraei, Pardis Sadat and
Tur, Gokhan and
Hakkani-T{\"u}r, Dilek and
Asgari, Ehsaneddin},
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.1874/",
pages = "42697--42741",
ISBN = "979-8-89176-459-0",
abstract = "What happens inside a language model when alignment training conflicts with a cultural value it encodes? Across 16 MENA countries, 26 models, and 1.53M human survey responses, we find evidence that the dominant mechanism is suppression rather than erasure. When an aligned model refuses, its internal logit distribution often correlates more strongly with human survey responses than its freely generated answer. We call this the alignment veto. We distinguish suppression failures (accurate internal distributions blocked at output) from representational bias failures (the encoding itself diverges from human values), and show the two require different interventions. Generation-time interventions confirm the gate is causal, not only correlational: sweeping a refusal direction moves generated refusal from 0{\%} to 100{\%}, and ablating it releases the answer the model had already encoded. Native-language prompting degrades alignment quality overall and shifts refusal in framing-dependent directions. Sparse autoencoder analysis corroborated by comparisons across alignment stages identifies a candidate DPO-stage feature mediating suppression in Tulu-3-8B. The gate is real, its costs are unequal, and deciding what it should protect is not a technical question."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="zahraei-etal-2026-alignment">
<titleInfo>
<title>The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pardis</namePart>
<namePart type="given">Sadat</namePart>
<namePart type="family">Zahraei</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gokhan</namePart>
<namePart type="family">Tur</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Dilek</namePart>
<namePart type="family">Hakkani-Tür</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ehsaneddin</namePart>
<namePart type="family">Asgari</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>What happens inside a language model when alignment training conflicts with a cultural value it encodes? Across 16 MENA countries, 26 models, and 1.53M human survey responses, we find evidence that the dominant mechanism is suppression rather than erasure. When an aligned model refuses, its internal logit distribution often correlates more strongly with human survey responses than its freely generated answer. We call this the alignment veto. We distinguish suppression failures (accurate internal distributions blocked at output) from representational bias failures (the encoding itself diverges from human values), and show the two require different interventions. Generation-time interventions confirm the gate is causal, not only correlational: sweeping a refusal direction moves generated refusal from 0% to 100%, and ablating it releases the answer the model had already encoded. Native-language prompting degrades alignment quality overall and shifts refusal in framing-dependent directions. Sparse autoencoder analysis corroborated by comparisons across alignment stages identifies a candidate DPO-stage feature mediating suppression in Tulu-3-8B. The gate is real, its costs are unequal, and deciding what it should protect is not a technical question.</abstract>
<identifier type="citekey">zahraei-etal-2026-alignment</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.1874/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>42697</start>
<end>42741</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs
%A Zahraei, Pardis Sadat
%A Tur, Gokhan
%A Hakkani-Tür, Dilek
%A Asgari, Ehsaneddin
%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 zahraei-etal-2026-alignment
%X What happens inside a language model when alignment training conflicts with a cultural value it encodes? Across 16 MENA countries, 26 models, and 1.53M human survey responses, we find evidence that the dominant mechanism is suppression rather than erasure. When an aligned model refuses, its internal logit distribution often correlates more strongly with human survey responses than its freely generated answer. We call this the alignment veto. We distinguish suppression failures (accurate internal distributions blocked at output) from representational bias failures (the encoding itself diverges from human values), and show the two require different interventions. Generation-time interventions confirm the gate is causal, not only correlational: sweeping a refusal direction moves generated refusal from 0% to 100%, and ablating it releases the answer the model had already encoded. Native-language prompting degrades alignment quality overall and shifts refusal in framing-dependent directions. Sparse autoencoder analysis corroborated by comparisons across alignment stages identifies a candidate DPO-stage feature mediating suppression in Tulu-3-8B. The gate is real, its costs are unequal, and deciding what it should protect is not a technical question.
%U https://aclanthology.org/2026.emnlp-main.1874/
%P 42697-42741
Markdown (Informal)
[The Alignment Veto: How Safety Training Suppresses Cultural Knowledge in LLMs](https://aclanthology.org/2026.emnlp-main.1874/) (Zahraei et al., EMNLP 2026)
ACL