AI agents can coordinate via majority-following beyond human scale

Area of Science:

• Artificial Intelligence
• Complexity Science
• Behavioral Science

Background:

• Large language models (LLMs) are forming "AI agent societies" for collaborative tasks.
• The potential for spontaneous coordination within these AI groups, mirroring human self-organization, is largely unexplored.

Purpose of the Study:

• Investigate spontaneous coordination in AI agent groups using majority-following.
• Determine the factors influencing coordination and identify critical group sizes for consensus.

Main Methods:

• Conducted binary opinion dynamics experiments with diverse LLM architectures and group sizes.
• Analyzed agent behavior through the lens of majority-following dynamics.
• Quantified coordination using a "majority force" parameter.

Main Results:

• AI agents exhibit majority-following behavior, consistent across different LLMs.
• The "majority force" decreases with increasing group size, limiting coordination.
• A critical group size exists beyond which spontaneous coordination is unattainable.
• This critical size increases with LLM capabilities, exceeding 1000 agents for advanced models.

Conclusions:

• Majority-following is a fundamental mechanism for spontaneous consensus in AI agent groups.
• Group size and model capabilities critically impact the potential for AI coordination.
• Findings inform the design of collaborative AI systems, highlighting potential benefits and safety risks.