Models are far better in English than in most languages, and worst in those with least written material online. That gap decides who gets a competent assistant and who gets an unreliable one, and it maps closely onto existing economic inequality.
Technical view
Performance correlates with representation in training data, and tokenisation efficiency compounds it: under-represented scripts consume several times more tokens for the same content, raising cost and consuming context. Approaches include deliberate multilingual data curation, continued pre-training for specific languages, tokeniser adaptation, and community-led dataset construction. Evaluation is itself a gap, since benchmarks in most languages are translated rather than native and miss culturally specific competence.
Related
Part of AI Terminology Index, section: Society and Impact.