Modelling columnists with large language models

Official URL: https://dx.doi.org/10.1109/SIU66497.2025.11112306

Large language models not only provide engineering solutions but also serve as promising tools that open up new research questions in social sciences. In particular, open-source language models fine-tuned for specific applications can be used to mimic human behavior and create social simulation systems. In this study, we model the opinions of eight columnists with diverse political ideologies. When the fine-tuned models are analyzed for their support on various claims made in columns, similarities are observed among the models based on columnists with similar viewpoints. However, when the base open-source model is used without fine-tuning, it fails to effectively capture the differences between the columnists' claims. Additionally, survey questions from election studies are posed to these models, and principal component analysis of the responses revealed meaningful distinctions that can be explained partly by economic voting theory between columnists. Last, their view on various security issues is shown to be an important dimension to differentiate pro-government columnists.