Multilingual issues still present

Multilingual support often looks better on paper than it does in production. A system may handle one language well and still struggle with nuance, tone, or domain-specific wording in others.

The challenge is that language is not just translation. Grammar, idioms, culture, and context all affect how users interpret a response, so accuracy alone does not guarantee a good experience.

Improving multilingual performance usually requires real samples from each language, not just automatic translation of English test sets. Systems become much more reliable when evaluation reflects the way each language is actually used.