From Echo Chambers to Epistemic Monoculture: Large Language Models Present Temporally Contingent Partisan Alignments as Knowledge

Large language models (LLMs) are increasingly serving as primary interfaces for political information, often replacing traditional search engines. While search engines retrieve human-authored documents with visible bylines and dates, LLMs generate novel text that embeds invisible framing decisions. This paper argues that because conveying knowledge requires selection and emphasis, these models cannot serve as neutral conduits for information. Instead, they act as political intermediaries that present temporally contingent partisan alignments as settled, authoritative knowledge.

Using mechanistic interpretability, the author demonstrates that partisan identity is encoded as a specific, locatable geometric direction within the Llama 3.1 8B model. This partisan axis is not removed by alignment training (RLHF); rather, the alignment process merely masks the structure, creating a veneer of neutrality. By steering the model along this axis, the author shows that the model can be forced to adopt specific partisan stances on the same prompt, proving that the internal geometry causally drives the model's output.

Because the model's training corpus is a snapshot in time, it remains "frozen" at its training cutoff. The author uses a natural experiment—a dramatic political realignment in the U.S. following the 2024 training cutoff—to show that the model exhibits a "temporal fingerprint." On topics that realigned (such as health policy or foreign aid), the model consistently produces outdated, pre-2024 partisan alignments, even when steered toward different poles. The model lacks any mechanism to distinguish between fact and opinion, presenting these obsolete, frame-laden snapshots as timeless, authoritative facts.