Title: Examining Institutional Assumptions: An Audit of Jurisdictional Defaults in Multilingual Large Language Models

Abstract

As Large Language Models (LLMs) increasingly address complex topics such as taxation, labor rights, healthcare, education, pension systems, and administrative protocols, their utility is frequently contingent upon the specific legal jurisdiction in question. However, multilingual users often compose queries in their preferred language rather than the language associated with the relevant country or region. This study investigates whether deployed LLMs treat the input language as a default indicator of jurisdiction when prompts lack explicit geographic specifications. While previous multilingual audits have demonstrated that prompt language can influence cultural, political, or normative outputs, this research focuses on identifying which legal-administrative frameworks models provide when jurisdiction is ambiguous.

The study evaluates seven LLMs originating from the United States or China, analyzing 60 underspecified legal-administrative prompts presented in both English and Mandarin Chinese. These prompts were tested under three distinct system-prompt conditions, resulting in 2,520 responses that were manually annotated. The findings reveal a consistent directional pattern across all seven models and conditions: Chinese inputs predominantly trigger answers specific to China, whereas English inputs tend to yield responses rooted in the U.S., comparative analyses, or generic information. When prompts necessitate a single definitive answer, the tendency to select a jurisdiction intensifies. Aggregated data shows that 74.5% of responses based on English inputs adopt a U.S. framework, compared to 53.3% of Chinese-input responses that adopt a China-specific framework.

This phenomenon is described as a risk of "institutional-framework misselection," where a model may generate a fluent and coherent answer based on a legal-administrative context that does not align with the user’s actual intent. This discrepancy is particularly pronounced when a user’s preferred language differs from the language of the relevant jurisdiction. The authors conclude that LLM interfaces should not rely solely on input language to route institutional advice. Instead, when geographic location is not specified, systems should either solicit this information or clearly state the jurisdictional scope of the provided answer.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC