Study: LLMs impose global narratives when queried in English vs Bangla

A new study by Md Arid Hasan, Ruwad Naswan, Farhan Samir, Sharifa Sultana, and Syed Ishtiaque Ahmed examines how large language models (LLMs) handle culturally grounded questions across languages. The researchers introduce CulturalNB, a meticulously curated dataset of 717 Bengali cultural instances with parallel Bangla–English question-answer pairs, supporting evidence, metadata, and sociocultural annotations. Using question-only and evidence-based prompting, they evaluated nine state-of-the-art LLMs with both human judges and two independent LLM judges, measuring cross-lingual consistency, language anchoring, global substitution, institutional bias, and epistemic perspective coverage.

The results reveal a systematic pattern of "global narrative dominance"—when questions are asked in English, LLMs exhibit significantly higher rates of global substitution and institutional framing, while coverage of local perspectives drops sharply. Even when local evidence is provided, it improves factual consistency and perspective coverage but does not eliminate the epistemic shifts triggered by the language of the query. The authors argue that these cultural failures are not simply missing-knowledge errors but reflect deeper failures in grounding and narrative prioritization. This has profound implications for the deployment of LLMs in low-resource cultural contexts, where English-mediated interactions may inadvertently overwrite local knowledge with globally dominant narratives.