[1H5-OS-5b] OS-5

6:55 PM - 7:10 PM JST(9:55 AM - 10:10 AM UTC)

[1H5-OS-5b-06] Designing Persuasive Narratives with Generative AI Based on the Active Inference ModelA Behavior Change Experiment on Walking for Health Promotion

〇Masaki Watanabe¹, Susumu Nagayama¹ (1. Hitotsubashi University)

Keywords:

generative AI,behavior change,narrative,active inference,persuasion

Persuasive narratives can be a powerful intervention approach for behavioral change. In recent years, the effectiveness of persuasion by generative AI incorporating specific narrative theories has been increasingly examined. However, many existing narrative theories rely on empirical perspectives and lack a unified theoretical foundation regarding the underlying cognitive mechanisms. This study proposes a framework for persuasive narratives based on the Active Inference model, which provides a unified account of perception, action, and learning. In this framework, narratives are designed to increase value alignment between the agent's preferences and predicted outcomes while simultaneously reducing uncertainty about action consequences. To examine the effectiveness of this framework, we conducted a silicon sampling experiment using 100 LLM agents equipped with demographic personas, comparing four narrative conditions: Active Inference, Prospect Theory, Aristotelian Rhetoric, and Baseline, all designed to encourage walking behavior. Results showed that the Active Inference condition achieved the highest selection probability at 54%, while also effectively suppressing counterarguing and psychological reactance, suggesting the potential advantage of this approach for narrative persuasion design.