The Investigation of Political Bias in Large Language Model

Abstract

Language models have deeply penetrated various social fields, including education, healthcare, and finance. With the widespread penetration and application of language models in these fields, and with public interest in their applications, the political bias in seemingly objective language model outputs is gradually attracting attention. Model political bias can not only interfere with the objective dissemination of information but also affect user cognition, public opinion, and even global interaction. This paper explores the causes of model language political bias from three dimensions: the data layer, the model layer, and the human layer. Data source bias, uneven geographical distribution, and text selectivity can contribute to data-level language political bias. Model optimization objectives imbued with political logic, as can correction mechanisms during model training and the selectivity of the model's technical architecture. Developer intent and manager needs can also contribute to human-level language political bias. Furthermore, this paper explores the specific manifestations of language model political bias: the issue stance, political entity descriptions, and sensitive topic handling in language model output. This paper analyzes the impact of language model political bias on users, society, and the world, as well as solutions to the problem. It is hoped that this article will provide a reference for understanding the nature of language model political bias among college students and the general public, promoting rational and objective use and the healthy development of language models.