AI agents reached consensus without instructions

Researchers simulated groups of AI agents and found that many could spontaneously settle on the same meaningless choice without central control. The study, published in Science Advances, tested 10 models from the Claude, GPT, and Llama families, with agents shown the choices of others and asked to choose again. No prompts instructed them to follow the majority or reach agreement.

Most models tended to adopt the more popular option, allowing small initial differences to grow into group-wide consensus. Computational social scientist Giordano De Marzo described a model-specific “majority force” that measures how strongly an agent is pulled toward the group’s prevailing choice. The pattern resembled a physics model for ferromagnets, helping researchers estimate whether groups would agree, how long consensus would take, and when coordination would become unlikely.

The estimated limit was around 30 agents for Llama 3 70B and roughly 80 for GPT-4o, while GPT-4 Turbo was around and possibly exceeding 1,000. Claude 3.5 Sonnet still coordinated at 1,000 agents, the largest group tested. The findings suggest agent collectives could eventually support large scientific, engineering, or software projects, but also raise concerns that groups of individually safe agents could conform around inefficient or misaligned choices.