The Social Cost of Intelligence: Emergence, Propagation, and Amplification of Stereotypical Bias in Multi-Agent Systems

2025-10-12

AI Summary

This study investigates the dynamic evolution of stereotypical bias in multi-agent systems (MAS), focusing on how collaboration and communication among large language model (LLM)-based agents generate, propagate, and amplify such biases. Method: We employ a simulation-based experimental framework, evaluating bias dynamics across three established bias benchmarks under diverse social scenarios, while systematically varying agent specialization, underlying LLM capabilities, and communication protocols (e.g., cooperative reasoning vs. adversarial debate). Contribution/Results: We首次 identify three key phenomena in MAS bias dynamics: early emergence, heightened susceptibility, and in-group preference. Crucially, we demonstrate that structured debate protocols significantly mitigate bias amplification, whereas stronger foundational models enhance system-wide fairness and stability. Results indicate that MAS inherently amplify bias more than single-agent systems; however, strategic optimization of communication mechanisms and model selection effectively attenuates this effect—providing both theoretical foundations and actionable design principles for developing fair and robust socially situated reasoning systems.

Technology Category

• Multiagent Systems: Agreement, Argumentation & Negotiation
• Cognitive Modeling & Cognitive Systems: Social Cognition And Interaction
• Knowledge Representation and Reasoning: Preferences

Application Category

• Social Networks and Social Media: Fairness and bias in social network and social media analysis
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Bias in large language models (LLMs) remains a persistent challenge, manifesting in stereotyping and unfair treatment across social groups. While prior research has primarily focused on individual models, the rise of multi-agent systems (MAS), where multiple LLMs collaborate and communicate, introduces new and largely unexplored dynamics in bias emergence and propagation. In this work, we present a comprehensive study of stereotypical bias in MAS, examining how internal specialization, underlying LLMs and inter-agent communication protocols influence bias robustness, propagation, and amplification. We simulate social contexts where agents represent different social groups and evaluate system behavior under various interaction and adversarial scenarios. Experiments on three bias benchmarks reveal that MAS are generally less robust than single-agent systems, with bias often emerging early through in-group favoritism. However, cooperative and debate-based communication can mitigate bias amplification, while more robust underlying LLMs improve overall system stability. Our findings highlight critical factors shaping fairness and resilience in multi-agent LLM systems.

Problem

• Investigating stereotypical bias emergence in multi-agent systems
• Analyzing bias propagation through inter-agent communication protocols
• Evaluating bias mitigation strategies in collaborative AI systems

Innovation

• Multi-agent systems analyze bias emergence and propagation
• Simulate social contexts with diverse agent interactions
• Cooperative communication protocols mitigate bias amplification