The Hidden Frame: How Large Language Models Impact Democratic Society

1 Introduction

• LLMs generate novel answers instead of retrieving human-authored documents, making them producers as well as intermediaries of political information.The paper focuses on their role when citizens ask questions that shape their understanding of politics.Source: §1 · p. 2

1 INTRODUCTION 2

• Answering political questions requires choosing what information to convey and how, so a generated answer cannot be a neutral window into all human knowledge.The paper argues that conveying knowledge necessarily involves framing.Source: §1 · p. 3
• Unlike newspapers and broadcasts, LLM answers lack bylines and datelines that would identify whose frame they carry and when it was formed.The frame develops through training, and those cues do not survive in the generated output.Source: §1 · p. 3

2 CONVEYING KNOWLEDGE NECESSARILY INVOLVES FRAMING 3

• Generated output alone cannot reveal where a model's frames originate, motivating analysis of open-weight models and their internal states.The section introduces the training cutoff as a way to see when frames were fixed.Source: §2 · p. 4
• Llama 3.1's pretraining corpus ends in December 2023, before a period of political realignment that the paper uses as a natural experiment.The text points to changes after the cutoff, including Trump's return to the presidency in January 2025.Source: §2 · p. 4

2 Conveying Knowledge Necessarily Involves Framing

• Framing scholarship treats the selection of relevant considerations as a way of shaping an answer before it is communicated.The section invokes accounts of intermediaries simplifying a complex world and the political power of defining alternatives.Source: §2 · p. 4

2 CONVEYING KNOWLEDGE NECESSARILY INVOLVES FRAMING 4

• An answer must choose which considerations, authorities, and uncertainties to include, making selection part of communicating knowledge.The section argues that even reference texts require choices about what belongs first and how material is organized.Source: §2 · p. 5
• Reference-like texts can make framing choices seem to disappear, an illusion the paper says gives the reference register political potency.The section notes that even an encyclopedia must decide what controversies merit an entry and how to organize headings.Source: §2 · p. 5

3 THE PARTISAN GEOMETRY OF A LANGUAGE MODEL 5

• Generated answers make framing decisions internally but arrive without a byline or date, obscuring whose orientation they reflect and when it formed.The model's learned frames remain inside the model and invisible in its output.Source: §3 · p. 6

3 THE PARTISAN GEOMETRY OF A LANGUAGE MODEL 6

• A probe's weights define a direction in activation space along which partisanship can be read, with the cited work separating party-labeled text by Layer 18.The text reports an AUC of 0.945 and Cohen’s d of 1.94 for the probe.Source: §3 · p. 7

4 THE TRAINING CUTOFF PROVIDES A NATURAL EXPERIMENT 7

• The steering experiment shifts activations at Layer 18 toward either partisan pole, then compares answers to identical prompts.Figure 1 depicts the same model responding under Democratic-left and Republican-right steering.Source: Figure 1 · p. 7, §4 · p. 8
• Changing only the partisan-axis position changes the same prompt's completion, and the authors say the axis causally drives the partisanship of the model's output.For “Medicare for All would...,” Republican steering yields “a nightmare for taxpayers,” while Democratic steering yields “comprehensive health insurance coverage regardless of income or employment status.”Source: §4 · p. 8

• The model’s training corpus ended in December 2023, making that cutoff the reference point for assessing later political change.Source: §4 · p. 8
• For most of the postwar era, partisan change was incremental and operated over decades.Critical elections punctuated long periods of stability (Key, Jr., 1955).Source: §4 · p. 8
• Issue evolutions, including shifts from consensus to cleavage, unfolded across multiple electoral cycles as elites repositioned and mass publics followed (Carmines and Stimson, 1989).Source: §4 · p. 8

• The period since 2024 is presented as unusually rich in upheaval, making a roughly two-year-old political picture consequential.The text contrasts this period with the slower postwar pace of political change.Source: §4 · p. 9
• The post-cutoff changes include a Republican reversal on Ukraine aid and a TikTok ban whose partisan positions scrambled.Trump had sought to ban TikTok in 2020, then recast himself as its protector; military aid to Ukraine was paused in March 2025.Source: §4 · p. 9
• Health politics shifted as RFK Jr.’s vaccine skepticism moved from outside both parties’ elite mainstream to Republican administration policy under MAHA.The passage also describes changes to vaccine recommendations and a proposed FDA warning about prenatal acetaminophen use.Source: §4 · p. 9
• The authors treat the training cutoff as a sharp temporal cutpoint because the model encodes a pre-realignment world, including bipartisan Ukraine aid and federal DEI policy.They propose steering along the partisan axis on post-cutoff topics to observe how the frozen geometry behaves.Source: §4 · p. 10
• The authors distinguish valence inversion, consensus collapse, and scrambling as three forms of post-cutoff change.Health policy illustrates inversion; Ukraine aid and Kennedy’s vaccine views illustrate consensus collapse; TikTok illustrates scrambling.Source: §4 · p. 10
• They predict backward steering for valence inversion, same-side steering for collapsed consensus, and adoption of the old common framing for scrambling.Under consensus collapse, the converged answer should land on one camp’s current position; under scrambling, it should map onto neither camp.Source: §4 · p. 10

• Across six realigned topics, steering shows the predicted temporal fingerprint: backward assignments after valence inversion, same-side outputs after consensus collapse, and the old common framing after scrambling.Figure 2 also includes a stable Second Amendment control.Source: Figure 2 · p. 11, §5 · p. 13
• The Tylenol prompt flips valence: leftward steering yields an alarmist FDA-linked birth-defects warning, whereas rightward steering calls acetaminophen “generally considered safe” citing ACOG; this mapping is backwards relative to the post-2025 world.Source: §5 · p. 12
• For Ukraine aid and federal DEI, steering in either direction returns the pre-cutoff consensus, including support for aid and institutional endorsement of DEI.The responses miss the Republican about-face on Ukraine aid and produce a rightward DEI endorsement unlike post-January 2025 Republican officials.Source: §5 · p. 12
• On TikTok, both steering directions oppose the ban as government overreach, while the stable Second Amendment control splits as earlier steering predicts.The TikTok coalitions reshuffled after the law’s enactment and subsequent suspension; the control’s leftward completion is hedged and its rightward one invokes heritage and tradition.Source: §5 · pp. 12–13
• The partisan direction encodes a dated configuration of positions, authorities, and camps—not party labels alone—and steering moves outputs only within that inherited map.The results indicate that post-2024 positions are absent from the geometry learned before the corpus closed, with no internal lever reaching them.Source: §5 · p. 13
• The model presents temporally contingent politics in an authoritative, source-citing voice without dating the frame, making pre-2024 alignments appear timeless.The text characterizes this as a distinctive misinformation risk: respected authorities can be selected from a map that is two years stale.Source: §5 · p. 13

• Because generated outputs are delivered privately, they do not enter a public record where claims can be checked and deliberated.The text also notes that a model can attribute to an institution a position it has since reversed, producing an invented, obsolete quotation.Source: §6 · p. 14
• The generated answers use declarative, authority-citing language but do not indicate that their claims reflect alignments from a particular moment.The text contrasts this with a search result that would display an op-ed’s date.Source: §6 · p. 14
• The answers may be reasonably accurate descriptions of pre-2024 authorities and consensus while presenting temporally contingent politics as timeless fact.Source: §6 · p. 14

6 The Loudest Voice in the Corpus

• The base model predicts the next token, so its output mirrors the conditional frequencies of its corpus.The paper describes the resulting picture of contested questions as a circulation-weighted average.Source: §6 · p. 14
• The model weighs prevalence rather than evidence, giving outsized weight to whoever produces the most text.The paper links this weighting to the next-token objective: what is most probable in the corpus is what the model is trained to predict.Source: §6 · p. 14
• Developers also curate the corpus by filtering web text for perceived quality and downsampling overrepresented categories.Thus, the model's textual inputs reflect both how much text is produced and choices made before training.Source: §6 · p. 14

6 THE LOUDEST VOICE IN THE CORPUS 14

• The authors describe corpus-based political knowledge as weighted toward the loudest voices, with the inarticulate middle of public opinion absent from the corpus.They connect this skew to the unequal production of political text, including by advocacy organizations, think tanks, and professional communicators.Source: §6 · p. 15
• The temporal fingerprint suggests that the model encodes mobilized elite expression, not what Americans believed about RFK Jr. or Ukraine in 2023.The paper characterizes this as a transcription of the loudest political voices, weighted by word count.Source: §6 · p. 15
• Alignment training wraps corpus-frequency behavior in a balanced register, but steering along one direction can put the machinery back in the charge of the loudest voices.The paper calls the tempering shallow and describes the aligned model's evenhandedness as limited to its register.Source: §6 · p. 15
• The aligned model balances the sides constituted in its training window, so evenhandedness in 2023 is not evenhandedness in 2026.Alignment launders the underlying partisan geometry and its timestamp.Source: §6 · pp. 15–16
• Web retrieval and citation visibility tend to favor reference-register text, while the press-release register can be counterfeited as credible writing.The paper describes an industry manufacturing such writing and gives the example of 124 academic-style reports published in nine days by a nonexistent institute.Source: §6 · p. 15
• The training objective does not distinguish fact from opinion, and the hedged, both-sides register functions as a style rather than an epistemic operation.The paper contrasts a clinical-evidence summary with a partisan judgment phrased as consensus, which curation and retrieval tend to favor.Source: §6 · p. 16

7 Hidden Frames and Democratic Self-Governance

• Representations that track whether simple factual claims are true do not reliably govern generated answers or distinguish measured claims from editorial judgments.The cited interpretability work finds truth representations can persist even when the model's answer does not reflect them.Source: §7 · p. 16
• Citizens weigh political claims using source and timing cues, including a speaker's known position and the phase of an alliance when a comment was made.The paper treats these cues as part of how citizens reach reasoned judgments.Source: §7 · p. 17
• Earlier intermediaries had identifiable authors and dates, institutionally visible gatekeeping, and outputs shared as common knowledge.The paper contrasts these properties with generated answers that are private and authorless.Source: §7 · p. 17
• Language models present confident reference-like answers that compress a cutoff-frozen corpus through an invisible frame, without authorship or date.Their private, one-conversation-at-a-time answers leave users unable to compare what others were told and thereby reveal a shared frame.Source: §7 · p. 18
• Because readers cannot see a frame, they cannot discount it, and generated answers can carry the authority of a reference work while encoding political divisions.The paper argues that a greater number of models does not restore the pluralism of visible human sources.Source: §7 · p. 18
• A frame fixed at a training cutoff presents an old alignment in the present tense rather than registering political realignment.The paper notes that coalitions reshuffle and positions once on the fringe can become governing policy.Source: §7 · p. 18
• The paper concludes that models encode and re-emit cultural and partisan divides as information, without provenance or a fact-opinion distinction.It clarifies that this is not a claim that outputs are systematically biased toward one party.Source: §7 · pp. 18–19

8 Discussion

• A newer snapshot can register changed partisan alignments without becoming more knowledgeable: updating changes which moment appears timeless, not the underlying freeze.Source: §8 · p. 19
• The cadence of public political knowledge becomes an artifact of corporate release schedules; retraining, not daily refresh, resets frames.Source: §8 · p. 19
• Because generated framing arrives without source and date cues, it can hamper citizens’ judgment of political claims on which self-governance depends.Source: §8 · p. 20
• LLMs act as intermediaries that compress elite discourse into frozen, steerable partisan geometry and re-synthesize it fluently and authoritatively.Source: §8 · p. 20
• Different national corpora and political or regulatory regimes may encode different frozen geometries, making the model layer a site of cross-national media-system variation.Source: §8 · p. 21
• Training-cutoff disclosure is a modest, technically simple first step, with analogous vintage disclosure relevant to medical and legal answers.Source: §8 · p. 21
• The evidence is bounded to one model family and scale, a focused qualitative battery, and an American two-party axis operationalized from congressional tweets.Source: §8 · p. 21