LLMs Can Infer Political Alignment from Online Conversations

A team of researchers from Indiana University and KAIST has published a study demonstrating that large language models (LLMs) possess a concerning ability to infer users' political leanings from seemingly mundane online conversations. By analyzing text from platforms like X and Reddit, models such as GPT-4 achieved prediction accuracies exceeding 80%, significantly outperforming traditional supervised machine learning classifiers. The research highlights that LLMs excel at detecting subtle, non-explicit linguistic cues—like slang, cultural references, or preferences for specific bands—that correlate with political identity, effectively piecing together a digital footprint that users may consider private.

The study's methodology involved aggregating multiple text-level inferences to form user-level predictions, with accuracy improving as more data from 'politics-adjacent' domains was included. This capability underscores a fundamental privacy risk: as AI models become more advanced and our online data exposure increases, the potential for misuse grows. The findings suggest that even innocuous public posts can be weaponized by AI to profile individuals, posing challenges for data privacy regulations and platform design. The 55-page paper, currently on arXiv, calls for greater awareness of how socio-cultural correlates in massive public datasets can be exploited by increasingly sophisticated computational methods.