Understanding Political Polarization With Generative AI: A Large-Scale Analysis of Voter Preferences on Social Media

In an era where users continuously express opinions across digital platforms, social media provides a large-scale lens on how people frame issues, justify choices, and react to events. Traditional approaches typically combine classifiers with unsupervised models to produce analytical outputs (e.g., spreadsheets, aggregate statistics), leaving human experts to interpret these outputs, connect signals across them, and translate them into explanations. The advent of Large Language Models (LLMs) has reshaped this process by enabling end-to-end pipelines that can both annotate large-scale text and generate coherent narrative reports that synthesize the evidence. This study presents an AI-based methodology for large-scale analysis of social media opinions during political events to estimate users’ propensity to support different candidates or factions. We classify and enrich posts with multidimensional annotations (e.g., candidate support, topics, sentiment) using a combination of traditional models and LLMs, inferring missing attributes from profile metadata and textual cues.We then aggregate annotations at the user level to build representative samples, validate their distributions against official population statistics, and generate concise reports that summarize the main motivations and issues driving support for each candidate or faction. We apply this methodology to X/Twitter data from swing states in the 2020 and 2024 U.S. presidential elections. For each state–candidate pair, the pipeline generates a report, which we compare against a baseline in which an LLM analyzes raw posts without any preprocessing or enrichment. The enriched configuration consistently outperforms the baseline across semantic similarity metrics, LLM-as-a-judge scores, and human preferences. Moreover, the motivations inferred from pre-election social-media data align closely with post-election expert summaries from sources such as Wikipedia and major newspapers, indicating that our approach can translate large-scale digital corpora into robust and interpretable political insight.