When English Rewrites Local Knowledge: Global Narrative Dominance in Large Language Models

Md. Arid Hasan, Ruwad Naswan, Farhan Samir, Sharifa Sultana, Syed Ishtiaque AhmedarXiv 2026 · DOI

TL;DR

Large Language Models (LLMs) often exhibit "global narrative dominance," replacing culturally specific knowledge with globally prevalent information, especially when prompted in English. This paper introduces CulturalNB, a Bengali cultural dataset, and demonstrates that even with local evidence, LLMs struggle to maintain cultural grounding, highlighting a critical bias in cross-cultural AI.

What Problem Does This Paper Address?

LLMs are increasingly used as cross-lingual knowledge interfaces, but their understanding of low-resource cultures is often uneven, less reliable, and more susceptible to prompt language. This paper investigates a specific failure mode: "global narrative dominance" (GND). GND occurs when LLMs, even when possessing the correct information, replace culturally specific details, interpretations, or narratives with more globally dominant or standardized alternatives. This is particularly concerning when users ask culturally grounded questions that depend on local history, institutions, or practices. The authors argue that existing evaluations often focus on factual knowledge gaps rather than the subtle, yet significant, shifts in interpretation and framing that occur due to language and cultural biases within LLMs.

Methods

The researchers introduce CulturalNB, a new dataset of 717 manually curated Bengali cultural instances. Each instance includes a culturally grounded question-answer pair, supporting evidence, and parallel Bangla-English translations. They evaluate nine state-of-the-art LLMs using two prompting strategies:

  1. Question-only prompting: This tests the LLM's internalized knowledge.
  2. Evidence-based prompting: This provides local evidence to see if it can override potential biases.

To evaluate the LLM responses, they employ a multi-faceted approach:

  • Human and LLM judges: They use both human evaluators and LLM-based judges (GPT-5.4-mini and Mistral 4 small) to assess responses.
  • Five key metrics:
    • Cross-Lingual Factual Consistency (CLFC): Measures if answers are semantically equivalent across Bangla and English.
  • Language Anchor Bias (LAB): Assesses if English prompts systematically shift responses towards globally dominant interpretations.
  • Global Substitution Rate (GSR): Quantifies the replacement of local referents with global alternatives.
  • Institutional Bias Rate (IBR): Measures shifts towards institutional authority over local knowledge.
  • Epistemic Perspective Coverage (EPC): Evaluates how well models represent multiple local viewpoints.

Datasets

  • CulturalNB: A newly introduced dataset comprising 717 manually curated Bengali cultural instances. It features parallel Bangla-English question-answer pairs, supporting evidence, and sociocultural annotations across five domains: History & Politics, Religion & Mythology, Traditional Medicine & Ecology, Geography & National Identity, and Art, Literature, & Cultural Practices. The dataset aims to capture locally situated knowledge that might be underrepresented in global corpora.

Key Insights and Ideas

  • English as an epistemic anchor: The language of the prompt (English vs. Bangla) significantly influences LLM responses, acting as an "epistemic anchor" that can steer interpretations towards globally dominant narratives, even when local evidence is provided.
  • GND is more than just missing knowledge: The paper argues that global narrative dominance isn't solely due to LLMs lacking specific cultural facts. Instead, it's a failure of grounding and narrative prioritization, where language and inherent biases lead to the overwriting of local knowledge.
  • Evidence helps, but doesn't fully fix it: Providing local evidence improves factual consistency and perspective coverage, but it doesn't eliminate language-induced shifts or global narrative dominance.
  • LLM judges can be biased: Human evaluation reveals that LLM judges might underestimate cultural inappropriateness, potentially due to inheriting their own training priors. This highlights the need for human expertise in evaluating culturally sensitive AI outputs.
  • Cultural knowledge is unevenly encoded: LLMs exhibit "epistemic instability," meaning their cultural understanding is not consistent across languages or contexts, often privileging globally prevalent perspectives when local knowledge is uncertain.

Key Results

  • Language Shift: English prompts systematically increase global substitution (GSR) and institutional bias (IBR), while reducing epistemic perspective coverage (EPC) compared to Bangla prompts.
  • GSR Impact: In the question-only setting, English prompts often lead to GSR values above 0.60 across models, compared to 0.21-0.40 for Bangla. This effect persists even with evidence.
  • IBR Impact: English prompts generally increase IBR, especially under the Mistral judge, with some models showing IBR above 0.64 for English questions.
  • EPC Improvement: Providing evidence substantially improves EPC for all models, but English questions still yield lower coverage than Bangla questions.
  • CLFC Instability: Cross-lingual factual consistency (CLFC) is generally weak, with models often producing substantially different claims when the same question is asked in Bangla and English. Scores often fall below 0.15 under the stricter GPT judge.
  • LAB Pervasiveness: Language Anchor Bias (LAB) is consistently positive across models, indicating that English prompts systematically shift responses toward globally dominant interpretations.

Limitations and Future Work

  • Dataset Scope: CulturalNB focuses on Bengali culture, and findings may not directly generalize to other low-resource contexts. The dataset size (717 instances) might not fully capture the diversity of Bengali culture.
  • Translation Nuance: While human translation and expert validation were used, some semantic or cultural nuance might still shift across languages.
  • LLM Judge Limitations: While scalable, LLM judges might have their own biases, and human evaluation is crucial for validation.
  • Future Work: The authors suggest exploring retrieval-augmented generation to reduce GSR by improving cultural grounding. They also emphasize the need for culturally responsible model development with community expertise.

Summary updated

BibTeX
@misc{hasan2026when,
  author = {Md. Arid Hasan and Ruwad Naswan and Farhan Samir and Sharifa Sultana and Syed Ishtiaque Ahmed},
  title = {{When English Rewrites Local Knowledge: Global Narrative Dominance in Large Language Models}},
  year = {2026},
  eprint = {2605.30481},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.30481},
}
Figure