[2M5-GS-11d] AI and Society

4:00 PM - 4:15 PM JST(7:00 AM - 7:15 AM UTC)

[2M5-GS-11d-03] Political Features of Qwen-Derived Large Language ModelsEvidence from Southeast Asia

〇Asei Ito¹, Kota Takaguchi² (1. The University of Tokyo, 2. Meiji University)

Keywords:

Large Language Model,Qwen,China,Political sensitivity,Southeast Asia

This paper examines the responses of large language models (LLMs) derived from Qwen, a China-origin foundational model, to politically sensitive questions, focusing on derivative models secondarily developed in Southeast Asia. Methodologically, we evaluate model outputs on political topics using a list of politically sensitive questions in Simplified Chinese released by the China Electronics Standardization Institute and Fudan University. Explicit refusals are identified using a dictionary-based approach, while contextual response avoidance is measured using an LLM-as-a-Judge framework. The results show substantial variation across models: the original Qwen model exhibits the highest level of alignment with official narratives, whereas several derivative models display lower alignment.