Why languages matter for AI safety
Models are used in many languages, but much of their training and testing has been weighted towards a few. Behaviour checked in one language may not hold in another.
Meaning also lives in register, idiom and culture. A response can be grammatical and still wrong, rude or unsafe for the people reading it. Native speakers hear the difference.
What language experts do
Language experts write and rank responses in their native language, evaluate translations, and annotate where meaning, tone or register goes wrong. They also contribute to voice and video data.
They red-team models in their language to test whether safety behaviour carries over, and produce alignment data on culturally appropriate responses. Linguists help design evaluations that measure what a language actually requires.
- Native-language answers and rankings
- Translation evaluation
- Voice and video data
- Red-teaming beyond English
- Alignment data on cultural context
Failure modes language experts catch
Experts catch literal translations that lose meaning, the wrong level of formality, mistakes in dialect, and idioms carried over from another language. They notice when a model sounds translated rather than written.
They also catch safety gaps: a request a model refuses in one language but answers in another, or content that is harmless in one culture and offensive in the next.
How we vet language experts
Language candidates are interviewed by AI voice agents, which also lets us hear how they speak. They then complete real-world work tests, such as translating a passage with the right register or judging a model's response in their language.
Credentials and experience are reviewed as part of vetting. Native speakers, translators and linguists are each assessed on the work their role involves.