Simulating Persuasive Dialogues on Meat Reduction with Generative Agents

TL;DR

Meat reduction offers public health and environmental benefits, but social norms maintain meat-centered meals. The paper explores generative-agent-based simulations using large language models to model multi-round persuasive dialogues about reducing meat, anchored in the Theory of Planned Behavior and validated against human data. Preliminary results show larger models produce reliable, theoretically consistent patterns that closely resemble human responses, supporting the approach's viability while highlighting limitations of smaller models and the need for more diverse personas and rounds. This methodology can accelerate discovery of tailored, low-cost persuasion strategies for different groups and inform subsequent human studies, all while prompting careful ethical governance of persuasive AI technologies.

Abstract

Meat reduction benefits human and planetary health, but social norms keep meat central in shared meals. To date, the development of communication strategies that promote meat reduction while minimizing social costs has required the costly involvement of human participants at each stage of the process. We present work in progress on simulating multi-round dialogues on meat reduction between Generative Agents based on large language models (LLMs). We measure our main outcome using established psychological questionnaires based on the Theory of Planned Behavior and additionally investigate Social Costs. We find evidence that our preliminary simulations produce outcomes that are (i) consistent with theoretical expectations; and (ii) valid when compared to data from previous studies with human participants. Generative agent-based models are a promising tool for identifying novel communication strategies on meat reduction -- tailored to highly specific participant groups -- to then be tested in subsequent studies with human participants.

Figures (9)

• Figure 1: Simulation of Persuasive Dialogues. We use generative agents to develop and select persuasion strategies for meat reduction with minimal social costs, to be tested in subsequent studies with human participants.
• Figure 2:
• Figure 3: Simulation Setup. We simulate dialogues in which a Persuader agent aims to convince a Recipient agent to reduce their meat consumption. In each round of the simulation, agents first (1) perform internal reflection, and (2) answer a questionnaire, before they (3) generate a response to the other agent.
• Figure 4: Simulated Survey Responses Are on Average Close to Responses from Human Participants. Mean values of our six central constructs based on human-reported data and simulated agent data. Each point represents one construct. The dashed diagonal line indicates perfect agreement between simulated and human means. All response scales range from 1 to 7.
• Figure 5: Strong Pearson Correlations between Attitude, Intention, Behavioral Control, Social Attractiveness and Social Closeness support Convergent Validity. Weaker correlations with theoretically more distinct constructs such as Subjective Norms support Discriminant Validity. Threat of Freedom and Meat Attachment were included to help assess if constructs form a distinct and coherent cluster.