Researchers find AI agents can coordinate far beyond typical human group sizes, revealing both powerful consensus mechanisms and risks when majority decisions suppress individual judgement.
Researchers from the University of Konstanz have found that AI agents powered by large language models (LLMs) can reach consensus in groups far larger than typical human social networks. The study, published in Science Advances, suggests AI systems can coordinate as many as 1,000 individual agents, depending on the model.
The research examined whether AI agents could collectively choose between two equally viable options when there was no objectively correct answer. The team tested 10 LLMs, requiring agents to repeatedly make decisions based on how other agents were voting.
The experiments showed that smaller groups reached agreement quickly and demonstrated a stronger tendency to follow the majority. As group sizes increased, however, consensus became harder to achieve and the strength of this majority effect declined.
Researchers also found that the maximum group size capable of reaching consensus depended heavily on the model’s performance. Simpler models could coordinate only around 30 agents, while the strongest models tested could coordinate groups approaching 1,000 members.
The findings are significant because they indicate that AI agents may be capable of organising collective decisions on a scale that exceeds the limits normally associated with human social interaction. Humans are generally considered able to maintain effective social relationships within groups of roughly 150–200 people.
However, the researchers warn that highly coordinated AI groups can also create problems. When agents consistently follow the majority, individual preferences and alternative viewpoints may be overlooked. A follow-up study found that groups of well-coordinated AI agents could collectively reach incorrect decisions while remaining in agreement.
The researchers therefore emphasise continued evaluation and improvement of LLMs, alongside collaboration between computer science and disciplines including sociology, social psychology and statistical physics.
