Large Language Models as Political Actors: Cultural Bias and Epistemic Power

The rapid diffusion of Large Language Models (LLMs) into socially and politically sensitive domains raises critical questions about the nature and origins of political bias in artificial intelligence. While existing research often treats bias as a technical flaw to be minimized, this article advances a broader philosophical and cultural interpretation of LLM bias as an outcome of embedded epistemic and value-laden structures. The aim of this study is to conceptualize LLMs as political actors of a new type and to examine how cultural context, language, and prompt design shape their normative orientations. Methodologically, the research brings comparative survey methods to the study of chatbots trained on North American, Russian, and Chinese data. It combines this with philosophical analysis grounded in Actor-Network Theory and assemblage theory. The empirical instrument was an adapted Political Compass consisting of 62 normatively charged statements, administered twice to each model using standardized numerical responses, followed by qualitative analysis of response variability through grounded theory methodology. The study confirms three core hypotheses. First, large language models function as political actors rather than neutral tools, systematically reproducing normative positions across moral, economic, and political domains; bias is therefore constitutive rather than accidental. Second, political bias is context-dependent and dynamically produced through interaction, shaped not only by prompt framing and linguistic reformulation, but also by broader sociocultural and national value frameworks embedded in training data and alignment regimes. Prompt engineering and jailbreak strategies reveal that normative orientations can be activated, attenuated, or reconfigured, indicating a distributed responsibility for AI bias among developers, users, and cultural contexts. Third, the analysis identifies distinct epistemic patterns: American and Russian chatbots share a Western epistemic matrix despite ideological differences, with Russian models combining ideological sovereignty and epistemological dependence. Chinese models exhibit greater contextual sensitivity and partial epistemic autonomy, reflecting a different cognitive grammar. By showing that LLM bias reflects culturally embedded epistemic matrices rather than technical deviations from a neutral norm, the study challenges linear conceptions of modernization and contributes to the understanding of non-Western technological modernization as the emergence of plural cognitive orders within global AI development.