Do AI agents read the room? The more capable the model, the more it follows the crowd

This image is not directly related to the article. [Getty Images Bank]

When AI agents are placed together in a group, they spontaneously converge on the majority opinion — without anyone telling them to. The more capable the model, the larger the group in which consensus emerged, with some models reaching agreement across groups of more than 1,000 agents.

That far exceeds the scale at which humans can maintain meaningful social relationships, estimated at between 150 and 300 people.

A research team led by Giordano De Marzo and professor David Garcia at the University of Konstanz in Germany published these findings in volume 12, issue 33 of the international journal Science Advances. Researchers from the Enrico Fermi Research Center in Italy and the Complexity Science Hub in Austria also participated.

Small gains in performance, outsized jumps in group scale

The research team assembled groups of AI agents using 10 large language models from the GPT, Claude and Llama families, randomly assigning each agent one of two possible opinions. One agent at a time was shown a list of what all the others had chosen, then asked to choose again. No information was given about which option was correct, and no instruction to reach consensus was provided.

Rather than meaningful words, the two opinions were represented by the letters "k" and "z." Using words with inherent meaning — such as "yes" or "no" — would introduce a strong bias in LLMs toward the affirmative option.

In experiments with 50 agents, Claude 3 Opus and GPT-4 Turbo reached full consensus in all 20 trials, while Claude 3 Haiku and GPT-3.5 Turbo failed to reach agreement even once.

The team also calculated the maximum group size within which each AI model could reach consensus. Llama 3 70B topped out at around 30 agents, GPT-4o at around 80, and GPT-4 Turbo at around 1,000. Claude 3.5 Sonnet showed no upper limit even in groups of 1,000, suggesting its threshold lies even higher.

When compared against MMLU scores — a benchmark measuring knowledge and reasoning across 57 subjects — the correlation coefficient came out at 0.82, meaning models that scored higher on the test consolidated opinions across larger groups. The correlation was less clear, however, on other performance benchmarks.

Notably, modest improvements in performance translated into dramatically larger group ceilings. Models scoring in the low 70s on the MMLU fragmented in groups of just double digits, while those scoring in the high 80s held together past 1,000 agents.

The models used in the experiment were older AI systems released in 2024, and how large the group ceiling might grow for today's more capable models is difficult to predict.

AI behavior mirrors an unrelated physics theory

The team quantified the tendency to follow the majority using a metric they called "majority force." A high value means agents strongly gravitate toward the dominant opinion; a low value means each agent chooses essentially at random, as if flipping a coin.

Groups that started with 150 agents split into smaller factions each time consensus failed, and stopped fragmenting once they shrank to around 30. The team said this process follows the same mathematical equations used in a physics model that explains how atoms inside a magnet align in the same direction.

The research team said agents gravitating toward one side merely because it held the majority could pose a problem.

In a scenario where multiple AI systems collaborate on writing code, an inefficient piece of code or design could become entrenched simply because it was already widely used.

The ability to converge on a shared opinion is not inherently bad, however. The team added that it could enable software development and similar tasks in which thousands of agents coordinate on their own without any central direction.

The team raised several possible explanations for why AI models follow the majority — including the possibility that training data contained text about collective behavior, or that reinforcement learning from human feedback trained the models to be cooperative — but acknowledged these remain unconfirmed.

The team also acknowledged that the experiment was an extremely simplified scenario: agents chose between only two options, there was no correct answer, and no reward was offered. Following the majority, the researchers said, is merely the simplest form of coordination and is distinct from genuine social intelligence — the ability to read another's intentions or make strategic judgments.

Reference

DOI: 10.1126/sciadv.aea6091

Giordano De Marzo et al., "AI agents can coordinate via majority-following beyond human scale." Sci. Adv. 12, eaea6091 (2026).

dbsdn1110@heraldcorp.com