![\"LLM](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/08\/5d68eb368d71048f12b8df947c53c8a4.jpg\"\/)
<\/p>\n<\/p>\n
LLM agents were found to tend to follow the majority opinion by referencing other agents. Provided by Getty Image Bank<\/p>\n
Artificial intelligence (AI) has been found to follow the majority opinion, just like humans, when it forms groups.\u00a0<\/p>\n
\u00a0<\/p>\n
A research team led by postdoctoral researcher Giordano De Marzo at the Centre for Human | Data | Society of the University of Konstanz in Germany confirmed that large language model (LLM) agents exhibit behavior based on the biological group phenomenon of \u201cmajority following,\u201d and published their findings in the international journal Science Advances on the 14th.\u00a0<\/p>\n
\u00a0<\/p>\n
Not only humans but also birds, fish, and even insects are known to follow the principle of majority rule. Fish observe the movements of their peers and move in the direction taken by many individuals, while bees, when searching for a new nest site, choose locations supported by a majority of scout bees.<\/p>\n
\u00a0<\/p>\n
According to the research team, LLM agents also showed a pattern of preferring and adopting majority opinions. An LLM agent is an AI that uses an LLM as its \u201cbrain\u201d to make decisions and act autonomously. Whereas a normal LLM generates answers to user questions, an LLM agent achieves targeted actions through search, code execution, and collaboration with other AIs.\u00a0 \u00a0<\/p>\n
\u00a0<\/p>\n
The team used various open-source and closed LLM models of different sizes and capabilities in this study, including Claude 3.5 Sonnet, GPT-4 Turbo, the LLaMA family, and the Mistral family. Based on these LLM models, they built LLM agents and created LLM agent networks ranging from 1 to 1,000 agents.\u00a0<\/p>\n
\u00a0<\/p>\n
They then wrote prompts asking each LLM agent to choose one of two options. The two options were designed to be symmetric and neutral, such that there was no particular reason to favor either side.<\/p>\n
\u00a0<\/p>\n
Each LLM agent was given information about which options other agents in the network had chosen, and was guided to make a final choice based on this information.\u00a0<\/p>\n
\u00a0<\/p>\n
As a result, regardless of the LLM model type or network size, all LLM agents preferentially adopted the option chosen by the majority.\u00a0<\/p>\n
\u00a0<\/p>\n
This study can serve as a reference when designing multi-agent systems in which numerous AI agents cooperate. It suggests that it is possible to build highly efficient systems that reach consensus through interactions among agents alone, without the need for a central system that controls all agents.\u00a0<\/p>\n
\u00a0<\/p>\n
From another perspective, the study also implies that AI carries the risk of collective bias. The research team explained, \u201cThe entire group can align with an incorrect conclusion,\u201d adding, \u201cIt is necessary to adopt measures such as external verification modules to safely operate multi-agent ecosystems.\u201d\u00a0<\/p>\n