Analyzing Projection Patterns and Alignment Biases of Large Language Models as Socio-Political Simulators

Large language models (LLMs) have shown strong capabilities in text generation, but their alignment bias for long-term narrative projection remains underexplored. This is an exploratory work representing a small-n pilot study, focusing on the socio-political projection patterns of large language models in the next 5 years. The pool of personalities chosen for this paper comprises a mix of scientists, entrepreneurs, activists, political leaders, and athletes. The paper defines socio-political forecasting for these personalities to perform a comprehensive study of open-source and closed-source LLMs using quantitative and qualitative analysis of model responses under a social setting when constrained by an identical prompt. The study documents tendencies consistent with political bias in LLMs and highlights post-alignment constraints in closed-source models, a structural pattern influence in open-source models, and language differences in the comparative analysis. All findings are interpreted as exploratory observations within this sample and should not be generalized without further empirical validation. Language differences between the open-source and closed-source outputs are also studied.