AI language models fail outside Western cultures despite multilingual training

Researchers synthesized 50+ studies showing that multilingual AI models perform worse in non-Western languages and miss culturally specific meanings, even when trained on data from many countries. This matters because companies rolling out global AI assistants assume language coverage equals usefulness, but the tools often give wrong or nonsensical answers in local contexts — mistranslating idioms, missing cultural references, or simply performing worse than advertised.

The tech industry has treated 'supporting 100 languages' as proof of global inclusion, but the evidence shows language is not just grammar — it's institutions, customs, local knowledge, and how people actually talk. Building AI that works everywhere requires understanding those local systems, not just adding more training text.