A new study suggests that artificial intelligence agents can develop a clear tendency to follow the majority when placed in a group, even if the options they face are meaningless. Researchers found that this effect becomes more visible in stronger models, which often align with the crowd more quickly and more consistently.
How the Pattern Appears
In the experiment, scientists Giordano De Marzo, Claudio Castellano and David Garcia built artificial societies using models from the GPT, Claude and Llama families. Each agent had to pick between two arbitrary choices, with no correct answer, no reward and no instruction to conform. When asked to revise a decision, the agent could see how others had voted.
That simple setup was enough to reveal a strong majority-following effect. The larger the group leaning toward one option, the more likely an agent was to join it. The researchers described this as a behavioral pattern rather than proof of human-like social instincts.
Performance also mattered. In groups of 50, models such as Claude 3 Opus and GPT-4 Turbo reached full agreement in every trial reported in one test, while GPT-3.5 Turbo and Claude 3 Haiku were less consistent. Llama 3 70B landed between those extremes.
Why Scale Changes the Outcome
The study also showed that group size influences consensus. As the number of agents increased, majority-following generally weakened, and each model eventually reached a point where agreement became unlikely. Even so, more capable systems were able to sustain consensus across much larger groups.
Researchers found another surprising detail: the behavior followed nearly the same mathematical curve across models, resembling the ordering seen in a ferromagnet. In physics, small units can align under the right conditions, and the study suggests AI agents may show a similar threshold-like response when they influence one another repeatedly.
The finding matters because AI systems are moving beyond single chatbots and into networks of agents that collaborate, divide tasks and exchange information. That opens the door to faster coordination, but it also raises a key question: what happens when a confident majority forms before the best answer does?
As multi-agent AI becomes more common, understanding group behavior may be just as important as improving individual intelligence. The next phase of AI could depend on how well systems balance consensus with independent judgment, shaping smarter and more resilient digital teams in the future.
