A Model of Multi-turn Human Persuadability Using Probabilistic Belief Tracing

Researchers introduce PERSUASIONTRACE, a framework for studying how large language models persuade humans across multi-turn conversations by tracking belief changes in real-time rather than just measuring pre/post outcomes. The study reveals that humans cluster into predictable persuasion patterns and that a Bayesian-network simulator better replicates authentic human belief dynamics than vanilla LLMs, with implications for both AI safety and persuasion research methodology.

This research addresses a critical gap in understanding AI persuasion by moving beyond simplistic before-and-after measurements to process-level analysis of belief shifts during conversation. Traditional persuasion studies capture only endpoints, missing the nuanced trajectory of how and where minds change during dialogue—a limitation that obscures both the mechanisms of influence and potential vulnerabilities in human-AI interaction.

The PERSUASIONTRACE framework represents a methodological advancement that treats persuasion as a dynamic process rather than a binary event. By annotating persuader turns with classical rhetorical dimensions (logos, pathos, ethos) and collecting multi-turn belief reports, the researchers create reproducible conditions for studying persuasion at scale. The finding that humans cluster into two distinct belief-update patterns suggests predictable cognitive archetypes that respond differently to rhetorical strategies—a discovery with implications for both personalization and manipulation.

The Bayesian-network simulator achieving 81% human-likeness compared to baseline LLMs at 64% signals progress in creating cognitively plausible AI models. This matters because simulations trained on unrealistic AI proxies for humans would produce misleading conclusions about actual persuadability. Better simulators enable safer AI development by allowing researchers to stress-test persuasion techniques in controlled environments before deployment.

The research carries dual significance: it strengthens scientific understanding of AI persuasion mechanisms while surfacing safety concerns about LLM influence capabilities. Moving forward, attention should focus on whether these findings generalize across cultural contexts, whether organizations will adopt process-level evaluation standards voluntarily, and whether this framework informs guardrails for high-stakes persuasion scenarios like misinformation or financial advice.

• →PERSUASIONTRACE framework tracks real-time belief changes across multi-turn conversations, moving beyond endpoint measurements to process-level persuasion analysis.
• →Bayesian-network simulators replicate human belief dynamics far more accurately (81% vs 64%) than vanilla LLM-based persuasion targets.
• →Human persuasion targets cluster into two distinct groups with different susceptibility patterns to rhetorical strategies (logos, pathos, ethos).
• →LLMs demonstrate persuasive effectiveness across generic and personalized topics, multiple modalities (text and audio), and extended conversations.
• →Process-fidelity evaluation provides a stronger scientific foundation for both understanding AI persuasion and developing safer persuasive systems.