The State of Multilingual LLM Safety Research: From Measuring the Language Gap to Mitigating It
Zheng-Xin Yong, Beyza Ermis, Marzieh Fadaee, Stephen H. Bach, Julia Kreutzer
2025-06-02
Summary
This paper talks about the current state of research on making large language models safe in different languages, showing that most safety studies focus only on English and not enough on other languages.
What's the problem?
The problem is that while AI models are used all over the world and in many languages, most of the work to make them safe and reliable is done in English, which leaves a big gap for people who use other languages.
What's the solution?
The researchers studied this language gap and suggested ways to test and improve AI safety in multiple languages, so that safety features and protections work well no matter what language someone is using.
Why it matters?
This is important because it helps make AI safer and fairer for everyone, not just English speakers, and ensures that people around the world can trust and benefit from AI technology.
Abstract
The analysis reveals a significant language gap in LLM safety research, focusing mainly on English, and suggests recommendations for multilingual safety evaluation and crosslingual generalization to enhance global AI safety.