The Multilingual Divide and Its Impact on Global AI Safety

AI safety risks are unevenly distributed globally: Large language models perform significantly better on safety benchmarks for dominant languages (like English) compared to many other languages, meaning people using AI in non-English languages may encounter more harmful outputs, misinformation, and unsafe behavior.

The language gap is widening, not closing: The paper documents that as AI capabilities advance, the performance disparity between dominant and non-dominant languages is actually growing, creating an expanding safety problem rather than one that naturally resolves with technological progress.

Data scarcity is a root cause: The language gap exists because AI training relies heavily on large datasets, and most available digital text comes from a small number of wealthy, English-speaking countries, leaving AI systems undertrained for the world's linguistic diversity.

Policymakers should prioritize multilingual dataset creation and transparency: The authors recommend that governance efforts focus on supporting the creation of safety-focused datasets in multiple languages and requiring companies to disclose performance gaps across languages, similar to other AI transparency requirements.

This is a global equity issue: The language gap means that speakers of minority and non-dominant languages have less access to safe, reliable AI systems, potentially widening existing technological and economic disparities between countries and communities.