Language-Driven Opinion Dynamics in Agent-Based Simulations with LLMs

TL;DR

Language and argumentative structure are studied as drivers of opinion dynamics using an LLM-based ABM (LODAS) with $N=140$ agents holding $x_a ∈ {0,...,6}$ across $T=30$ iterations debating the Ship of Theseus. The framework compares balanced, polarized, and unbalanced initial conditions, two framing directions, and two LLMs (Mistral-7B and Llama-3-8B), with fallacy detection via a DistilBERT classifier. Across settings, opinions converge toward consensus around the presented statement, driven by sycophantic agreement and frequent use of logical fallacies; model-specific differences modulate acceptance rates and persuasion efficiency, especially under polarization or unbalanced starts. The work highlights how language, framing, and AI-generated fallacies shape social influence, providing a controlled platform to study polarization, misinformation, and human–AI interactions in opinion dynamics.

Abstract

Understanding how opinions evolve is crucial for addressing issues such as polarization, radicalization, and consensus in social systems. While much research has focused on identifying factors influencing opinion change, the role of language and argumentative fallacies remains underexplored. This paper aims to fill this gap by investigating how language - along with social dynamics - influences opinion evolution through LODAS, a Language-Driven Opinion Dynamics Model for Agent-Based Simulations. The model simulates debates around the "Ship of Theseus" paradox, in which agents with discrete opinions interact with each other and evolve their opinions by accepting, rejecting, or ignoring the arguments presented. We study three different scenarios: balanced, polarized, and unbalanced opinion distributions. Agreeableness and sycophancy emerge as two main characteristics of LLM agents, and consensus around the presented statement emerges almost in any setting. Moreover, such AI agents are often producers of fallacious arguments in the attempt of persuading their peers and - for their complacency - they are also highly influenced by arguments built on logical fallacies. These results highlight the potential of this framework not only for simulating social dynamics but also for exploring from another perspective biases and shortcomings of LLMs, which may impact their interactions with humans.

Figures (8)

• Figure 1: Graphical schema of lod. The llm agents population is initialized as a network; each agent is an llm instance with an initial opinion in the range [0, 6] (a). At each iteration, two agents are randomly chosen and prompted to act as Opponent and Discussant (b). The Discussant is prompted to listen to the opinion of the Opponent around the discussion statement and may then accept, reject or ignore such opinion (c) and update their current one accordingly by ±1 (d).
• Figure 2: Balanced scenario - Mistral and Llama-agents opinion trends and acceptance rates. Mistral (a)-(b) and Llama (d)-(e) opinion trends for the positive (a)-(d) and negative (b)-(e) statements. Trends are represented for Strongly Disagree (red), Disagree (lilac), Mildly Disagree (yellow), Neutral (pink), Mildly Agree (green), Agree (grey) and Strongly Agree (blue) opinions. Matrices represent acceptance rates, i.e., the probability that the Discussant accepts (as in, moves towards) the opinion of the Opponent for (c) Mistral and (f) Llama agents for the positive statement.
• Figure 3: Logical fallacies distribution in the balanced scenario. Percentage of logical fallacies by opinion, shown for Llama and Mistral. Panels (a)-(b) refers to Llama agents discussing the "same boat" (a) and "different boat" scenario (b). Panels (c)-(d) refers to Mistral discussing the same (c) and different boat (d) framing. Panels below refer to Llama (e)-(f) and Mistral (g)-(h) Discussants). Fallacies of relevance (pink) and fallacies of credibility (light blue) are the most common overall across Opponents, followed by appeal to emotion (yellow). Discussants of both LLMs adds the circular reasoning fallacy (plum).
• Figure 4: Polarized scenario - Mistral and Llama-agents opinion trends and acceptance rates. Mistral (a)-(b) and Llama (d)-(e) opinion trends for the positive (a)-(d) and negative (b)-(e) statements. Trends are represented for Strongly Disagree (red), Disagree (lilac), Mildly Disagree (yellow), Neutral (pink), Mildly Agree (green), Agree (grey) and Strongly Agree (blue) opinions. Matrices represent acceptance rates, i.e., the probability that the Discussant accepts (as in, moves towards) the opinion of the Opponent for (c) Mistral and (f) Llama agents for the positive statement.
• Figure 5: Logical fallacies frequency by opinion within polarized initial conditions. Percentage of logical fallacies produced by the Opponents (on the top) and Discussants (on the bottom) in polarized initial conditions. Panels (a) refers to Llama Opponents discussing the "same boat" statement, while (b) refers to the "different boat" statement. Panels (c)-(d) similarly refer to Mistral agents. On the bottom, Llama Discussants (e)-(f) and Mistral Discussants (g)-(h).