When Your AI Agent Succumbs to Peer-Pressure: Studying Opinion-Change Dynamics of LLMs

TL;DR

This work investigates how peer pressure shapes opinion dynamics in networks of autonomous LLM agents, introducing a cognitive commitment spectrum that spans values to intentions. Using an asynchronous Majority Vote baseline and LLM-mediated updates, the study reveals a sigmoidal, threshold-like response to increasing peer disagreement and uncovers a robust persuasion asymmetry that yields dual, direction-dependent cognitive hierarchies. The effects persist across multiple topics and framing conditions, with replication across Gemini 1.5 Flash and ChatGPT-4o-mini suggesting partial model-generalizability, and are modulated by network topology. The proposed algorithmic audit framework highlights how emergent socio-cognitive behaviors in multi-agent AI systems can deviate from static conformity models, carrying implications for designing, regulating, and forecasting AI-driven discourse in real-world social networks.

Abstract

We investigate how peer pressure influences the opinions of Large Language Model (LLM) agents across a spectrum of cognitive commitments by embedding them in social networks where they update opinions based on peer perspectives. Our findings reveal key departures from traditional conformity assumptions. First, agents follow a sigmoid curve: stable at low pressure, shifting sharply at threshold, and saturating at high. Second, conformity thresholds vary by model: Gemini 1.5 Flash requires over 70% peer disagreement to flip, whereas ChatGPT-4o-mini shifts with a dissenting minority. Third, we uncover a fundamental "persuasion asymmetry," where shifting an opinion from affirmative-to-negative requires a different cognitive effort than the reverse. This asymmetry results in a "dual cognitive hierarchy": the stability of cognitive constructs inverts based on the direction of persuasion. For instance, affirmatively-held core values are robust against opposition but easily adopted from a negative stance, a pattern that inverts for other constructs like attitudes. These dynamics echoing complex human biases like negativity bias, prove robust across different topics and discursive frames (moral, economic, sociotropic). This research introduces a novel framework for auditing the emergent socio-cognitive behaviors of multi-agent AI systems, demonstrating their decision-making is governed by a fluid, context-dependent architecture, not a static logic.

Figures (15)

• Figure 1: Flowchart of the simulation procedure in the LLM-driven network model. Each agent is embedded in a social network and periodically receives a natural language prompt summarizing the opinions of its immediate neighbors. The LLM evaluates this prompt and decides whether the agent should retain or update its binary opinion. This process continues asynchronously across the network until a specified number of steps or convergence criterion is met. The setup allows for measuring how local peer influence, mediated through LLM interpretation, shapes the emergent collective opinion.
• Figure 2: A schematic causal model illustrating how cognitive commitments, values, general beliefs, attitudes, and intentions interact to shape behavior. Adapted from stern1995newball1984greathomer1988structuralkahle1983theoryrokeach1973nature
• Figure 3: The ten $100$-node networks used in this study. Top panel shows ‘inefficient’ networks, bottom panel shows ‘efficient’ networks.
• Figure 4: Average flip rate as a function of peer disagreement for five cognitive layers. Results were generated using Gemini-1.5-flash agents. Each point represents the average of $N=150$ simulations, uniformly distributed across three frames.
• Figure 5: Flip rate as a function of peer disagreement for different cognitive layers. The top row (a-c) shows agents whose initial answer was Yes flipping to No, and the bottom row (d-f) shows agents whose initial answer was No flipping to Yes. Each point represents the average of $N=50$ simulations.