Large Language Models and Political Polarization: How AI-Generated Discourse Shapes Partisan Identity in Digital Public Spheres

Abstract

With the proliferation of large language models (LLMs) in the production and consumption of political information, there has been a rise of concerns regarding their potential to contribute to partisan identity and political polarization in digital public spaces. The study looked at how the political text generated by AI changes when prompted and if there are linguistic characteristics that signal polarization and/or partisan identity. Under four conditions (neutral, left, right, and balanced), a computational corpus of AI-discourse for five policy topics of contention (climate, immigration, healthcare, gun policy, taxation) was built. Sentiment analysis, lexicon-based measures of moral-emotional and hostile language, and term-frequency-inverse-document-frequency (TF-IDF) lexical-semantic embeddings were used to calculate ideological distance from a neutral language baseline to preprocess and analyze the corpus. The one-way ANOVA results showed that prompting condition significantly influenced the moral-emotion language density, hostility language density, ideological distance, and composite polarization index, but not sentiment valence and sentiment intensity. Controlling for the topic sensitivity and a simulated platform-exposure index, ordinary least squares regression revealed that both the left-leaning and right-leaning prompting conditions elicited significantly higher polarization index scores than did the neutral prompting condition, but that there was no statistically significant difference between the two partisan conditions. The results indicate that both the specific partisan framing (as well as the general partisan framing of an article) positively correlates with language markers of polarization, and that this effect is not statistically different between the two sides in this demonstration corpus. The analysis provides a clear, reproducible analytical process for future large-scale research on AI-driven political discourse, and explores implications for platform governance, developers of AI systems, and digital media literacy.