@inproceedings{wu-brandao-2026-ai,
title = "{AI} Safety Lost in Translation: Evaluating the Effectiveness of {E}nglish-{I}talian Cross-Lingual {LLM} Safety Alignment",
author = "Wu, Alessio and
Brandao, Martim",
editor = "Piperidis, Stelios and
Bel, N{\'u}ria and
van den Heuvel, Henk and
Ide, Nancy and
Krek, Simon and
Toral, Antonio",
booktitle = "Proceedings of the Fifteenth Language Resources and Evaluation Conference",
month = may,
year = "2026",
address = "Palma de Mallorca, Spain",
publisher = "ELRA Language Resource Association",
url = "https://aclanthology.org/2026.lrec-1.296/",
doi = "10.63317/24nruqbycv2a",
pages = "3697--3713",
abstract = "Large Language Models (LLMs) have been shown to be vulnerable to various issues of bias and safety, for which new safety alignment techniques have been proposed. In this paper, we investigate the degree to which such techniques improve safety in a non-English language, specifically in Italian, both when they have and don{'}t have access to safety training data in that language. We evaluate standard mitigation techniques and assess cross-lingual safety transfer by comparing English-only versus bilingual Supervised Fine-Tuning (SFT), on several open-source small LLMs: Qwen3, Llama3.2, and Gemma3. Results confirm a significant cross-lingual safety gap, with most models performing worse in Italian. We find that while prompt engineering is generally effective, the impact of SFT is highly inconsistent. English-only SFT occasionally failed to transfer safety improvements into Italian and even deteriorated the performance of some models. Furthermore, bilingual SFT repeatedly underperformed other mitigation methods. These findings demonstrate that safety alignment does not always generalize across languages and models, and standard mitigation strategies can lead to unpredictable effects. We thus highlight the critical necessity for language-specific evaluation and dedicated multilingual safety research to ensure AI is developed equitably and safely for a global audience."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="wu-brandao-2026-ai">
<titleInfo>
<title>AI Safety Lost in Translation: Evaluating the Effectiveness of English-Italian Cross-Lingual LLM Safety Alignment</title>
</titleInfo>
<name type="personal">
<namePart type="given">Alessio</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Martim</namePart>
<namePart type="family">Brandao</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifteenth Language Resources and Evaluation Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">Stelios</namePart>
<namePart type="family">Piperidis</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Núria</namePart>
<namePart type="family">Bel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Henk</namePart>
<namePart type="family">van den Heuvel</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nancy</namePart>
<namePart type="family">Ide</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simon</namePart>
<namePart type="family">Krek</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Antonio</namePart>
<namePart type="family">Toral</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>ELRA Language Resource Association</publisher>
<place>
<placeTerm type="text">Palma de Mallorca, Spain</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Large Language Models (LLMs) have been shown to be vulnerable to various issues of bias and safety, for which new safety alignment techniques have been proposed. In this paper, we investigate the degree to which such techniques improve safety in a non-English language, specifically in Italian, both when they have and don’t have access to safety training data in that language. We evaluate standard mitigation techniques and assess cross-lingual safety transfer by comparing English-only versus bilingual Supervised Fine-Tuning (SFT), on several open-source small LLMs: Qwen3, Llama3.2, and Gemma3. Results confirm a significant cross-lingual safety gap, with most models performing worse in Italian. We find that while prompt engineering is generally effective, the impact of SFT is highly inconsistent. English-only SFT occasionally failed to transfer safety improvements into Italian and even deteriorated the performance of some models. Furthermore, bilingual SFT repeatedly underperformed other mitigation methods. These findings demonstrate that safety alignment does not always generalize across languages and models, and standard mitigation strategies can lead to unpredictable effects. We thus highlight the critical necessity for language-specific evaluation and dedicated multilingual safety research to ensure AI is developed equitably and safely for a global audience.</abstract>
<identifier type="citekey">wu-brandao-2026-ai</identifier>
<identifier type="doi">10.63317/24nruqbycv2a</identifier>
<location>
<url>https://aclanthology.org/2026.lrec-1.296/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>3697</start>
<end>3713</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T AI Safety Lost in Translation: Evaluating the Effectiveness of English-Italian Cross-Lingual LLM Safety Alignment
%A Wu, Alessio
%A Brandao, Martim
%Y Piperidis, Stelios
%Y Bel, Núria
%Y van den Heuvel, Henk
%Y Ide, Nancy
%Y Krek, Simon
%Y Toral, Antonio
%S Proceedings of the Fifteenth Language Resources and Evaluation Conference
%D 2026
%8 May
%I ELRA Language Resource Association
%C Palma de Mallorca, Spain
%F wu-brandao-2026-ai
%X Large Language Models (LLMs) have been shown to be vulnerable to various issues of bias and safety, for which new safety alignment techniques have been proposed. In this paper, we investigate the degree to which such techniques improve safety in a non-English language, specifically in Italian, both when they have and don’t have access to safety training data in that language. We evaluate standard mitigation techniques and assess cross-lingual safety transfer by comparing English-only versus bilingual Supervised Fine-Tuning (SFT), on several open-source small LLMs: Qwen3, Llama3.2, and Gemma3. Results confirm a significant cross-lingual safety gap, with most models performing worse in Italian. We find that while prompt engineering is generally effective, the impact of SFT is highly inconsistent. English-only SFT occasionally failed to transfer safety improvements into Italian and even deteriorated the performance of some models. Furthermore, bilingual SFT repeatedly underperformed other mitigation methods. These findings demonstrate that safety alignment does not always generalize across languages and models, and standard mitigation strategies can lead to unpredictable effects. We thus highlight the critical necessity for language-specific evaluation and dedicated multilingual safety research to ensure AI is developed equitably and safely for a global audience.
%R 10.63317/24nruqbycv2a
%U https://aclanthology.org/2026.lrec-1.296/
%U https://doi.org/10.63317/24nruqbycv2a
%P 3697-3713
Markdown (Informal)
[AI Safety Lost in Translation: Evaluating the Effectiveness of English-Italian Cross-Lingual LLM Safety Alignment](https://aclanthology.org/2026.lrec-1.296/) (Wu & Brandao, LREC 2026)
ACL