Multilingual AI Reaches Arabic and Hindi: What Gets Lost
New AI systems finally support the world's most-spoken languages at quality matching English -- but cultural representation in training data remains deeply uneven.
The Language Barrier Falls
Models like Voxtral Transcribe 2 support dozens of languages at near-English quality. Meta Omnilingual ASR handles 100+ languages. Open-source projects fine-tune models for regional dialects.
A Rights Issue
When AI works better in English, English speakers get better healthcare triage, legal assistance, educational tools, and economic opportunities. Language access to AI is an equity issue.
What Gets Lost
Multilingual capability is necessary but not sufficient. AI trained primarily on English data carries assumptions about family structure, legal concepts, medical knowledge, and governance that do not translate globally.
The Democratic Opportunity
TGF envisions participation in native languages -- not just translation, but cultural context. This requires community-led model training, open-source models, and democratic data governance.