New AI models are being trained to think in local languages

When a Japanese user asks an AI a question, they expect a response that respects keigo—a complex system of honorifics that signals social hierarchy. Standard models often miss these nuances, but a new wave of sovereign AI from Japan, South Korea, and France is being trained on massive national datasets to ensure they think within their own cultural frameworks. Japan, for instance, is feeding models over one trillion tokens of Japanese text to reduce errors and cultural hallucinations.