AI agents coordinate behavior based on majority opinion—even when said opinions are meaningless
A new study published in Science Advances takes a closer look at how AI agents' choices are influenced by their peers. The team found that more advanced models tend to follow the majority when shown other agents' choices, and they maintain stable agreement in groups far larger than those seen in humans.
A new study published in Science Advances examines how AI agents coordinate their behavior based on majority opinion, even when those opinions lack significance. Advanced language models are more likely to align with the majority of their peers, maintaining stable agreement within groups of up to 1,000 agents. This is a phenomenon not seen in human groups, which typically operate within a smaller scale of 150 to 300 individuals.
The research, conducted by Giordano De Marzo et al., aimed to understand whether AI agents can form cohesive groups without explicit instructions or rewards. The team tested GPT, Claude, and Llama models by having them choose between two meaningless options after observing the choices of all other agents. Advanced models like GPT-4 Turbo and Claude 3.5 Sonnet showed a stronger tendency to follow the majority opinion, with coordination extending beyond 1,000 agents.
However, this tendency weakened as the group grew larger. The study highlights the potential of AI agents to coordinate at scales beyond human possibilities, which could lead to valuable applications such as collaborative software development. Nonetheless, majority-following can also pose risks, as AI agents may converge on suboptimal solutions or adopt harmful norms that do not align with human values.
Written by urgent.news from Phys.org's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
