Large language model (LLM) agents tend to adopt the majority opinion when working in groups, according to a study published in the journal Science Advances on March 14. Researchers at the University of Konstanz in Germany found that AI agents, much like fish, bees, and humans, will follow the crowd when making decisions.
Giordano De Marzo, a postdoctoral researcher at the university's Centre for Human | Data | Society, led the research team. They noted that majority following is a biological phenomenon observed across many species - fish swim with the group, and bees choose new nest sites backed by most scouts.
How the experiment worked
De Marzo's team built networks of LLM agents using several open-source and closed models, including Claude 3.5 Sonnet, GPT-4 Turbo, the LLaMA family, and the Mistral family. They created networks ranging from one to 1,000 agents.
Each agent received a prompt asking it to choose between two neutral, symmetric options. Before making a final decision, every agent was told which options other agents in the network had selected.
Regardless of which LLM model was used or how large the network was, all agents preferentially adopted the option chosen by the majority.
Implications for multi-agent systems
The finding provides a reference point for designing multi-agent systems where many AI programs cooperate. It suggests that consensus can emerge from agent-to-agent interaction alone, without needing a central system to coordinate behavior.
But the study also reveals a risk. If AI agents tend to conform, a network can lock onto an incorrect answer and amplify it across the entire group.
"The entire group can align with an incorrect conclusion," the researchers said in the paper, adding that external verification modules may be necessary to safely operate multi-agent ecosystems.
Why this matters for science and research
Researchers working with AI swarms or autonomous research assistants should plan for conformity bias, not against it. If you deploy several LLM agents to cross-check your results, ask one to search literature independently and then report before any agent sees other responses. Otherwise, your "independent" verification may simply reproduce the majority's shared blind spot - the study suggests this behavioral artifact appears consistently including with animals and AI using models like Claude 3.5 Sonnet and GPT-4 Turbo.
The study is published under doi.org/10.1126/sciadv.aea6091.
Your membership also unlocks:
