Large language models embed shifting political biases as knowledge
From Echo Chambers to Epistemic Monoculture: Large Language Models Present Temporally Contingent Partisan Alignments as Knowledge
Computation and Language
Summary
Large language models, like the ones used to answer questions, are becoming key sources of political information. The authors found that these models contain hidden political biases tied to specific time periods, especially around recent US politics changes. Instead of just showing facts, the models present partisan views as if they were factual knowledge, without distinguishing opinion from fact. This can lead to a uniform political viewpoint inside the model that amplifies existing divides in society rather than offering a neutral source of information.
What this means in practice
- •For social media platform teams: Identify how language models might amplify partisan biases in political content moderation and recommendation systems.
- •For ai safety engineers: Design mitigation strategies that recognize temporally contingent political biases embedded in large language models to reduce misinformation risks.
Authors
Wend K. Tam
Abstract
Large language models (LLMs) are rapidly becoming an interface between citizens and political information. They are often regarded as "a better Google." While this analogy might work for some instances, it is unintuitively problematic for democratic politics. A search engine retrieves human-authored documents, while a language model generates novel text that necessarily embeds invisible framing decisions. Because conveying knowledge involves framing, a system that generates answers cannot serve as a neutral conduit to "all human knowledge." Instead, these systems are becoming a new kind of political intermediary. Mechanistic evidence shows that partisan identity is encoded as a locatable geometric direction inside the Llama 3.1 8B model, and that alignment training masks rather than removes this structure. Building on that evidence, we present steering experiments that exploit a model's training cutoff in 2024. This cutpoint auspiciously falls just before a dramatic realignment in American politics marked by the second Trump administration and the MAHA transformation of health politics, providing us with a natural experiment. We find that the model presents temporally contingent partisan alignments as knowledge, with no mechanism for distinguishing fact from opinion. This reality moves the information environment beyond the echo chamber toward an epistemic monoculture where language models, purporting to summarize "all human knowledge" are, in actuality, simply magnifying the cultural and partisan divides inherent in their training data.