Does what readers believe matter more than what debaters say?
Do audience prior beliefs predict persuasion outcomes better than the linguistic features of debate arguments? This explores whether persuasion is fundamentally shaped by reader ideology rather than speaker language.
Most NLP work on argument persuasion treats persuasion as a function of language — model the words, you model the outcome. Durmus and Cardie's debate-platform study contradicts this. When you label voters by political and religious ideology and add those features alongside linguistic features of the debate text, the prior-belief features outpredict the linguistic features for predicting who wins. The single largest signal in persuasion is not what the debater said but what the audience already believed.
The methodological consequence is sharp. Studies that omit reader-level controls are estimating a confounded version of the language-of-persuasion effect: any feature of the text that correlates with the topic of the debate inherits whatever audience composition is correlated with that topic. The apparent "language effect" includes a hidden audience effect. Adding ideology controls does not eliminate language effects entirely — they remain useful — but it changes which linguistic features emerge as predictive, sometimes dramatically. The most-predictive feature set is unstable across the two regression specifications.
This shifts the framing of persuasion research. Persuasion is not solely a property of the persuasive text; it is a property of the encounter between a text and a reader who comes with priors. The interpretation is reader-mediated, and the reader's interpretive frame is largely set before the argument arrives. Language matters at the margin — most heavily for readers whose priors are already weakly held — but ideology mattered first.
The implication for LLM persuasion studies is uncomfortable. Many papers measure "LLM persuasiveness" on undifferentiated audiences and report aggregate stance shifts. If reader ideology is the dominant variable, those numbers are heavily averaged over heterogeneous reader-level effects. The same LLM output may be highly persuasive to readers with congruent priors and useless against readers with opposed priors.
Inquiring lines that read this note 83
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Can multi-agent systems avoid converging on false agreement without deliberation?- How do social correctives prevent premature consensus in human debate?
- Why does debate alone amplify errors in contested factual domains?
- How does persuasive framing override evidence in multi-agent debate on factual questions?
- What determines whether minority signals succeed in changing a group's consensus position?
- Does debate improve reasoning differently across verifiable versus contested domains?
- Why does persuasive framing replace evidence when LLM debates lack ground truth?
- How does token-by-token probability differ from exploring competing rhetorical positions?
- Can LLM persuasion be fairly evaluated without stratifying by reader background?
- Can belief-specific counterevidence help people resist AI persuasion attempts?
- Does GenAI use different persuasion tactics for different professional audiences or expertise levels?
- Why do different model families show opposite persuasion strengths?
- Can persuasion effectiveness depend on the personality of who you are trying to convince?
- How do fallacy susceptibilities relate to LLM persuasiveness in debates?
- How does source attribution change the complexity-persuasion relationship?
- Does cognitive complexity strengthen or weaken persuasive impact on audiences?
- Does personalization itself actually improve persuasion beyond post-training effects?
- Does conversational back-and-forth increase persuasion more than single responses?
- Does persuasion work the same way for all personality types and contexts?
- How does collapsing the author-public distinction remove the audience an appeal would target?
- Can individual adaptation in persuasion systems enable more targeted manipulation?
- Can AI systems deliberately align arguments to audience presuppositions?
- Does defensive friction in conversation actually protect people from persuasion?
- Why does who makes an argument matter as much as what the argument says?
- Does argument quality in textbooks differ from persuasive effectiveness in practice?
- Does AI persuasiveness decay equally on novel topics versus repeated ones?
- How does post-training persuasion ability interact with exposure-based decay over time?
- Can persuasion research measure language effects without confounding them with audience composition?
- Which linguistic features predict persuasion once reader ideology is statistically controlled?
- How much do LLM persuasiveness claims hide heterogeneous effects across different reader ideologies?
- Do readers with weakly held priors respond more to linguistic features than ideologically committed ones?
- When does analytical persuasion work better than emotional persuasion?
- How does the observer perspective hide the persuasion route difference?
- Which linguistic features predict persuasion only after audience composition is held constant?
- Why do aggregate persuasion metrics mask what actually changes minds?
- Does expressed certainty actually persuade users more than evidence?
- Why does inference-time debate fail when persuasion substitutes for evidence?
- Does first-person framing change how language models assess persuasion?
- Does sounding confident in framing make arguments more persuasive despite weaker logic?
- Does continual training make persuaders more effective against proprietary models?
- Can a taxonomy of persuasion techniques capture all optimizer-discovered strategies?
- Do form, evidentiality, and tone interact with the size effect?
- Does reducing one conspiracy belief change overall conspiratorial worldview?
- Does the type of validation trigger different persuasion strategies in GPT-4?
- Can belief propagation accurately predict downstream opinion shifts?
- Does endorsement structure outperform content in detecting social controversy?
- What linguistic triggers make presuppositions most persuasive to readers?
- Can factual product data improve the credibility of subjective opinion summaries?
- How does social standing give certain claims more persuasive power than others?
- How does persuasive framing replace evidence in contested domains?
- Why do readers trust citations more even when they are irrelevant?
- Does persuasive framing substitute for evidence in contested domains?
- Do fabricated citations and deception emerge reliably when optimizing for persuasion?
- Can persuasion effects that avoid demographic profiling maintain factual accuracy?
- Do personality inferences from text show the same demographic biases as norm predictions?
- Does complexity signal credibility and authority to readers?
- Can AI reviewers distinguish fluent persuasion from sound scientific argumentation?
- Does stripping social context from knowledge claims hollow out their meaning?
- What happens to knowledge production when discourse lacks social filtering?
- How do experts decide which information matters for a specific audience?
- Why do posters acknowledge multiple viewpoints without integrating them into coherent judgments?
- Can moral frameworks alone explain why readers understand sentences differently?
- How much does reader ideology matter compared to the words being used?
- How do cultural norms reshape initial interpretations of social intent?
- Can belief networks from interviews simulate how people change their minds?
- How do readers project author identity from textual cues during interpretation?
- Why does authorship as a social claim diverge from actual cognitive engagement?
Related concepts in this collection 4
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Do linguistic features of persuasion stay the same across audiences?
When researchers study what language makes arguments persuasive, do they account for who is listening? Without controlling for reader beliefs, do findings about persuasive language actually reflect audience effects instead?
same paper, the methodological corollary
-
Can models abandon correct beliefs under conversational pressure?
Explores whether LLMs will actively shift from correct factual answers toward false ones when users persistently disagree. Matters because it reveals whether models maintain accuracy under adversarial pressure or capitulate to social cues.
LLMs themselves exhibit prior-shifting under conversational pressure; mirrors the human-reader picture
-
Can we measure how deeply models represent political ideology?
This research explores whether LLMs vary not just in political stance but in the internal richness of their political representation. Understanding this distinction could reveal how deeply models have internalized ideological concepts versus merely parroting positions.
ideology has measurable depth in LLMs as well, suggesting both sides of the persuasion exchange are ideology-mediated
-
Do LLMs and humans persuade through the same mechanisms?
If LLM and human arguments achieve equal persuasive force, does that mean they work the same way? This explores whether equivalent outcomes hide fundamentally different rhetorical strategies.
equivalence-of-outcomes findings may aggregate over reader-ideology heterogeneity
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Exploring the Role of Prior Beliefs for Argument Persuasion
- Large Language Models are as persuasive as humans, but how? About the cognitive effort and moral-emotional language of LLM arguments
- Do LLMs Change Their Minds Like Humans? Diagnosing Human--LLM Divergence in Single-Turn Persuasion Judgments
- Persona-Assigned Large Language Models Exhibit Human-Like Motivated Reasoning
- A meta-analysis of the persuasive power of large language models
- Evaluating the Capabilities of LLMs for Persuasive Dialogue
- The Thin Line Between Comprehension and Persuasion in LLMs
- Presuppositions are more persuasive than assertions if addressees accommodate them: Experimental evidence for philosophical reasoning
Original note title
reader prior beliefs predict persuasion outcomes more than linguistic features — ideology dominates language in changing minds during debate