Cohere study finds a cultural blind spot in leading AI models

Being able to converse in a language is not the same as sharing the assumptions embedded in it. Cohere Labs tested 28 AI systems on 6,200 riddles written by community members in their original languages, covering 62 communities, rather than on fact-recall questions that can be answered by a web search. The best model was 79.3% accurate overall, but performance collapsed in some less-represented communities; no model cleared 22% in either Lumasaba or Mongondow, and answers rarely acknowledged uncertainty. One telling example: only one model correctly read Germany’s “9/11” as November 9—the date the Berlin Wall fell—rather than the US attacks, underscoring why global AI needs testing for local context, not just multilingual fluency.