@olav @timnitGebru the simple reason is that LLMs are a pure case of language without culture. Even as we start to identify biases present in models, we completely fail to associate them to the culture that they come from, to the extent that we risk attributing features of the recursive GenAI loop to human society as a whole.

There's also a lot of research on how narrower culture-specific models outperform LLMs on tasks, related to the corresponding cultures. All this should be obvious, but gets subdued in the AI marketing hype