Should I Agree with You? Simulating Persuasion and Decision Dynamics in Multi-Agent Moral Dilemmas
Track: Preliminary Work Track
Keywords: moral dilemmas, multi-agent debate, personas, moral reasoning
TL;DR: This paper explores how personas influence moral decision-making in AI-driven debates. Using a dataset of ethical dilemmas, it analyzes single-agent and multi-agent interactions and track persuasion dynamics.
Abstract: Moral dilemmas challenge individuals to navigate competing ethical objectives, with no definitive right or wrong answer. While prior research has explored moral reasoning across demographic and cultural dimensions, the role of personas in multi-agent debate settings remains underexplored. In this work, we investigate how an agent’s persona influences its decision-making in both single-agent and multi-agent contexts, focusing on persuasion dynamics in moral dilemmas. We construct a comprehensive dataset of multiple-choice moral dilemmas and assign personas across six dimensions: gender, age, socioeconomic status, country, political ideology, and personality. Using this dataset, we first examine how an agent’s persona affects its confidence in moral decisions in a single-agent setting. We then extend our analysis to multi-agent debates, where we study the evolution of moral stances, the impact of different debate formats, and the effectiveness of persuasion strategies. To systematically evaluate these interactions, we introduce key metrics, including confidence change, win rate, consensus rate, and debate efficiency. Our findings provide insights into the interplay between personas and moral reasoning, contributing to the development of ethically aware AI systems capable of engaging in nuanced moral discourse.
Submission Number: 17
Loading