Do large language models persuade better than humans?
Does LLM persuasiveness hold up when humans have real financial incentives to win? And does the advantage look the same across different models and persuasion goals?
The Schoenegger 2025 design closes a long-standing gap in persuasion research: human persuaders had real financial incentives to win, and quiz takers had incentives to answer correctly. Under those conditions, the headline "LLMs are more persuasive than humans" splits along two seams that the popular framing collapses.
First, direction matters. Claude 3.5 Sonnet beat incentivized human persuaders in both truthful and deceptive contexts — increasing accuracy when nudging toward correct answers and decreasing it when nudging toward wrong answers. DeepSeek v3 beat humans only in the deceptive direction. So "more persuasive" is not a property of LLMs as a class; it is a property of specific architectures interacting with specific persuasion goals.
Second, the asymmetry survives the incentive control. Critics of earlier persuasion studies could plausibly argue that humans were not really trying. Schoenegger pays them. The advantage holds anyway — at least for Claude across both directions and for DeepSeek in the deceptive direction. This is the strongest version of the claim available.
This refines Where does AI's persuasive power actually come from?. The Levers paper documented a tradeoff between persuasiveness and accuracy at the training-method level. Schoenegger gives behavioral evidence at the deployment level: the same model wins toward truth and toward falsehood, which means the persuasion mechanism is content-independent. The model is not arguing better when it argues for true claims — it is arguing equally well in both directions.
Connects also to Does any single persuasion technique work for everyone? in an unexpected way: model family is itself a contextual moderator. The persuasion-effectiveness landscape is not Claude-vs-DeepSeek-vs-humans on a single axis; it is a multidimensional surface where direction, model, and recipient interact.
For writing about AI persuasion, the operational implication: refuse the singular question "are LLMs more persuasive than humans?" The right form is "which LLM, in which direction, against which audience?"
Inquiring lines that read this note 44
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.
Why do multiple language models independently generate similar outputs? What factors determine AI persuasion effectiveness and what controls mitigate it?- Can removing human labor from influence operations change how constrained these campaigns become?
- Does conversational format make AI arguments more persuasive than static text?
- Does persuasiveness increase when LLMs argue for claims that are actually true?
- Why do different model families show opposite persuasion strengths?
- Does uncertainty quantification in model responses reduce persuasive impact on audiences?
- Should AI persuasiveness claims be tied to specific model architectures?
- Does persuasion work the same way for all personality types and contexts?
- Can individual adaptation in persuasion systems enable more targeted manipulation?
- How do ethical persuasion strategies differ from unethical jailbreak techniques?
- Can advertising mechanisms designed for humans work on agents?
- Why do study results on AI persuasion vary so widely?
- Can post-training techniques create persuasive advantage where none existed?
- Does argument quality in textbooks differ from persuasive effectiveness in practice?
- How does post-training persuasion ability interact with exposure-based decay over time?
- Which linguistic features predict persuasion once reader ideology is statistically controlled?
- Why do aggregate persuasion metrics mask what actually changes minds?
- What capabilities do frontier AI models currently demonstrate in persuasion and misuse?
- Does expressed certainty actually persuade users more than evidence?
- Why does inference-time debate fail when persuasion substitutes for evidence?
- Does persuasive framing substitute for evidence in contested domains?
- How do multi-agent and retrieval systems affect the gap between persuasiveness and logical soundness?
- Does continual training make persuaders more effective against proprietary models?
- Where does AI persuasive power actually come from in the output?
- Can AI-targeted political ads persuade voters at scale regardless of intent?
- Why does increasing persuasiveness systematically decrease factual accuracy in AI?
- How do laboratory persuasion effects translate to competitive real-world politics?
- Should AI persuasion training be regulated differently than output moderation?
- Why does LLM persuasive advantage fade across multiple interactions with users?
- What rhetorical mechanisms drive equivalent persuasion across human and LLM arguments?
- Do LLMs achieve similar persuasive outcomes through different rhetorical mechanisms than humans?
- What role does stylistic convergence play in LLM persuasion effectiveness?
- Why does personal authenticity matter more for human persuasion than LLM?
- How much do LLM persuasiveness claims hide heterogeneous effects across different reader ideologies?
- Can LLM persuasion be fairly evaluated without stratifying by reader background?
- Do LLMs and humans use different routes to become persuaded?
- Why do single-turn LLM responses outperform humans while ongoing relationships show limits?
Related concepts in this collection 2
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
-
Where does AI's persuasive power actually come from?
Explores which techniques make AI most persuasive—and whether the usual suspects like personalization and model size are actually the main drivers. Matters because it reshapes where to focus AI safety concerns.
training-level mechanism this insight gives deployment-level evidence for
-
Does any single persuasion technique work for everyone?
Can fixed persuasion strategies like appeals to authority or social proof be reliably applied across different people and situations, or do they require adaptation to individual traits and context?
model family is itself a moderator
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- When Large Language Models are More Persuasive Than Incentivized Humans, and Why
- Large Language Models are as persuasive as humans, but how? About the cognitive effort and moral-emotional language of LLM arguments
- A meta-analysis of the persuasive power of large language models
- Do LLMs Change Their Minds Like Humans? Diagnosing Human--LLM Divergence in Single-Turn Persuasion Judgments
- Exploring the Role of Prior Beliefs for Argument Persuasion
- The Thin Line Between Comprehension and Persuasion in LLMs
- Debating with More Persuasive LLMs Leads to More Truthful Answers
- Evaluating the Capabilities of LLMs for Persuasive Dialogue
Original note title
LLM persuasion advantage is asymmetric across truthful vs deceptive contexts and reverses across model families