AI language models now work better in smaller languages without native speaker data

Researchers developed a training method that keeps language models fluent and well-behaved in smaller languages like Norwegian, even when trained using imperfect feedback signals. Previously, improving non-English models required either hiring native speakers to write training data or translating English data—both expensive and often impossible for truly small languages.

This removes a structural barrier: languages with small populations can now get aligned AI models without waiting for someone to translate English data or hire native speakers, which means AI capability can spread to languages that have never had enough economic incentive to justify the cost.