Western-Centric Bias in Large Language Models

Large language models learn patterns from training data and are commonly evaluated with benchmarks whose language, topics and respondent pools are unevenly distributed. That can make a system fluent in English yet less reliable or culturally aligned elsewhere. “Western-centric” is not a single error metric; research operationalizes it through comparisons between model answers and surveys, country-specific references or culturally grounded tasks. A 2024 PNAS Nexus study used World Values Survey items to compare outputs from five GPT models with values reported by people in survey data. It found model responses tended toward self-expression values associated with English-speaking and Protestant European settings in the study’s framework. A 2026 PNAS study compared LLM-generated moral-value estimates with survey data from 48 countries and reported systematic deviations, with stronger mismatches in some Middle Eastern and Sub-Saharan African samples. The authors noted that English-language survey data could underrepresent less globally connected respondents; analyses with native-language data found similar discrepancies. These are empirical comparisons to particular survey measures, not complete descriptions of every culture or person in a country. The World Values Survey itself uses structured questions and sampled respondents; its results are not a timeless cultural essence. Models also vary across versions, languages, prompting and topics. A model may answer more appropriately when given local context, yet prompting alone cannot guarantee factual or cultural fit. Evaluation should include local-language speakers, locally relevant sources, and checks for stereotypes. Western cultural bias is a system-level risk in data and evaluation, not an inherent trait of every Western user or every model output.