Mapping the Silence: Cross-Lingual Asymmetries in Large Language Model Refusal and Evasion on Sensitive History

Abstract

As large language models increasingly shape public discourse, this study evaluates how US and China-based models govern access to politically sensitive history in US English and Mandarin Chinese. Across 2,400 interactions covering sixty globally significant events, we examined how prompt configurations modulate LLM responses. Chinese models demonstrated statistically significant cross-lingual disparities, restricting Mandarin queries up to three times more than English queries. Furthermore, models were substantially more likely to restrict responses to subjective, political, or contemporary framings than factual queries, highlighting the need for further evaluation of cross-lingual differences in AI systems and their security implications in politically sensitive contexts.