How AI Models Overwrite Local Knowledge with Global Narratives

Researchers from ArXiv released a study highlighting how large language models (LLMs) often reflect globally dominant narratives instead of local cultural knowledge. They created a dataset called CulturalNB, which includes 717 manually curated Bengali cultural instances with parallel Bangla-English question-answer pairs, supporting evidence, metadata, and sociocultural annotations. This dataset helps identify how AI models can misrepresent local contexts, a problem they term 'global narrative dominance'.

This matters because AI models are increasingly used as cross-lingual knowledge interfaces, meaning they translate and interpret information across languages. When these models prioritize global narratives, they can erase or misrepresent local cultural knowledge, affecting how people understand and interact with their own cultures. For example, an AI might provide answers about Bengali culture that align more with Western perspectives than with actual Bengali traditions.

If you're curious about how AI models handle cultural knowledge, you can explore the CulturalNB dataset on ArXiv. While the dataset is technical, it's a valuable resource for understanding the nuances of AI and culture. You can also try asking AI models questions about local cultures and compare their responses to what you know to be true in your own cultural context.