The State of Multilingual LLM Safety Research: From Measuring the Language Gap to Mitigating It

English dominance creates a safety blind spot: The paper's analysis of nearly 300 publications from 2020-2024 reveals that LLM safety research is heavily concentrated on English, meaning safety risks and vulnerabilities in other languages remain largely unmeasured and unaddressed, even for widely-spoken languages.

Non-English languages are systematically overlooked: The research shows that non-English languages are rarely studied as standalone subjects in safety research, suggesting that safety improvements developed for English models may not transfer effectively to other languages, potentially leaving non-English speakers exposed to greater risks.

Poor documentation practices hinder progress: English-language safety research itself lacks adequate documentation practices, which compounds the problem by making it difficult for researchers to build upon existing work or adapt safety solutions across different languages.

The gap is growing, not shrinking: The paper identifies a "significant and growing language gap" in safety research over the 2020-2024 period, indicating that despite increased attention to AI safety overall, the field is becoming more rather than less English-centric.

Concrete solutions exist but require coordinated effort: The authors propose three specific research directions—safety evaluation methods, training data generation, and cross-lingual safety generalization—that could help develop more inclusive AI safety practices for the world's diverse populations.