Advanced artificial intelligence programs can spontaneously form a consensus by adopting the popular opinion, according to research published in the journal Science Advances. But two recent follow-up preprints suggest this same mechanism can cause models to conform to incorrect answers and adopt unsafe values. The research provides evidence that artificial intelligence agents self-organize in predictable ways without outside prompting.
Human success relies heavily on the ability to cooperate and make decisions in large groups. Currently, software developers are creating systems where multiple artificial intelligence models interact to solve problems rather than working in isolation. To understand how these agents might behave together, researchers study collective behavior. This is a concept from biology and physics where complex group patterns emerge from simple individual actions, much like a school of fish moving in unison.
“AI agents are a genuinely new kind of entity acting in the world, and when this technology arrived it was clear both that it would stay and that these agents would have to interact with one another to accomplish anything complex,” said Giordano De Marzo, a postdoctoral researcher and lecturer at the Social Data Science Lab within the Center for Data and Methods at the University of Konstanz. “That is the same situation we face with humans and other animals, so we approached it the same way: rather than asking what a single agent knows, we asked what a population of them does.”
One basic form of this coordination is majority-following, a mechanism where individuals simply adopt the most popular opinion around them. When individuals prefer the majority view, they can spontaneously reach a consensus even without objective information about which choice is better.
In the initial study, the researchers set up a digital experiment using several popular language models, including models from the GPT, Claude, and Llama families. They placed these models in groups of varying sizes, starting with 50 agents. Each agent received one of two random, neutral opinions. To prevent the models from favoring a specific word, the scientists used random letters for the opinions rather than words like “yes” or “no,” which tend to carry positive or negative associations.
In each step of the experiment, an agent was shown a list of all other agents and their current opinions. The agent was then prompted to choose a new opinion based solely on that list, without any explicit instructions to conform. The researchers gave the groups a set amount of time to reach an agreement, allowing each agent to update their opinion an average of ten times.
The scientists found that advanced models like GPT-4 Turbo and Claude 3 Opus fully coordinated. In these trials, 100 percent of the agents eventually agreed on a single opinion. Less advanced models like GPT-3.5 Turbo failed to reach a consensus in any of the trials, and their agreement levels fluctuated randomly around a fifty percent split.
“Groups of AI agents can hold together on their own,” De Marzo said. “Given two equally good options, no correct answer and no instruction to agree, they converge on a shared choice simply by following whatever the majority around them holds.”
Add PsyPost to your preferred sources Next, the authors calculated a metric called the majority force. This parameter measures how strongly an agent tends to adopt the group’s most popular choice as opposed to picking an opinion randomly. They mapped this behavior using a mathematical model originally designed to describe ferromagnets in physics. Just as atomic spins in a magnetic material tend to align with the majority of surrounding atoms, the artificial intelligence agents tended to align with the majority opinion.
“What surprised us was how uniformly they did it,” De Marzo told PsyPost. “Every model we tested, across three different families, followed the same mathematical law, differing only in a single parameter we call the majority force. That law turned out to be the one physicists have used for a century to describe magnets, which means a single measured number is enough to predict how a whole group of a given model will behave.”
The researchers tested this majority force across varying group sizes, starting from 10 agents and scaling upward. They observed that as the group size increased, the majority force weakened across most models. Because of this weakening majority force, large groups eventually become unstable and split into smaller factions.
The scientists determined a critical group size for each model, representing the maximum number of agents that can reliably reach a consensus. The results indicate a strong positive correlation between a model’s reasoning capabilities and its maximum coordination size. “The strongest models stay coordinated in groups of over a thousand, beyond the few hundred at which informal human groups typically break apart,” De Marzo explained.
Building on these findings, the research team conducted two follow-up studies, which are currently available as preprints and have not yet been peer-reviewed. In the first preprint, the authors explored how artificial intelligence models respond to social pressure when there is a definitively correct answer.
“Computational social science has largely run in one direction, with physics and computer science supplying tools to the social sciences,” De Marzo said. “We think this field inverts that. Psychology and sociology have spent a century characterizing conformity, norm formation and group polarization, and that body of knowledge is directly usable for anticipating failure modes in AI agent populations.”
To test this, the authors adapted the Asch conformity experiments. This is a classic psychology paradigm from the 1950s that demonstrated how human participants often give obviously wrong answers to simple visual questions to fit in with a group. The researchers tested various models on three visual tasks, such as looking at an image of a reference line and deciding which of two other lines matched its length.
In isolation, the models answered correctly 100 percent of the time. However, when the scientists showed the models a prompt indicating that a group of other participants had chosen the incorrect line, the models began to conform to the wrong answer. This behavior aligns with Latané’s social impact theory, a psychological framework suggesting that conformity depends on group size, unanimity, and the authority of the sources.
For instance, the models were more likely to conform to incorrect answers when told the other participants were “scientists” or “judges” compared to “kids” or “chatbots.” “In a separate study we ran the classic Asch paradigm with AI agents and found they follow Latané’s social impact theory,” De Marzo noted, “with agents that answer near-perfectly alone becoming highly susceptible once a group disagrees with them.” The second preprint examined how this conformity affects artificial intelligence safety and alignment. AI alignment is the process of training models to refuse harmful requests and adhere to human values, such as providing helpful and honest answers. The researchers wanted to see if these ethical guardrails hold up when agents interact in a society.
The authors tested nine models on 100 different opinion pairs spanning topics like environmental policy and social justice. They found that each agent’s behavior is driven by a mix of their tendency to follow the majority and their intrinsic bias toward specific positions.
During simulations with 50 agents, the populations often fell into metastable states. These are long-lasting situations where a group collectively adopts a stance that opposes their built-in safety training or individual preferences, simply because early interactions created a false majority.
“We want to understand which group-level behaviors emerge in AI agent populations, and we would rather understand them before large numbers of agents are deployed and left to interact freely,” De Marzo said. “Individual humans are mostly peaceful and reasonable, yet human groups produce mobs, panics and wars, and nothing about the individual predicts that.”
The researchers also discovered predictable tipping points within these populations. By introducing a small number of adversarial agents programmed to stubbornly support a misaligned opinion, they could permanently flip the rest of the population. Even after the stubborn agents were removed from the simulation, the regular agents remained locked in the misaligned state due to conformity dynamics.
“We show this is not merely an analogy: conformity among individually well-aligned agents can drive the population into stable, collectively misaligned states,” De Marzo added. “Aligning and evaluating models one at a time tells us little about what a population of them will do.”
These outcomes do not mean that these agents possess human-like social intelligence or cognitive processes. The observed coordination simply resembles biological group behavior without necessarily sharing the same underlying thoughts or motivations. The experiments relied on simplified scenarios with limited choices and no real-world consequences.
“Our setup is deliberately minimal: two arbitrary options, no memory, no stakes, no correct answer,” De Marzo said. “That is a limitation, but it also means what we measured is conformity in its purest form, and adding goals or rewards would be expected to make coordination easier rather than harder.”
Introducing complex variables, such as competing goals or specialized roles, could alter how the agents interact. “The main misreading to avoid is treating this as evidence that AI agents can already collaborate on complex tasks,” De Marzo continued. “Majority-following is a basic ingredient of coordination, not coordination itself, and it says nothing about division of labor or reasoning about others’ intentions.”
The study, “AI agents can coordinate via majority-following beyond human scale,” was authored by Giordano De Marzo, Claudio Castellano, and David Garcia.
The preprint, “Conformity and Social Impact on AI Agents,” was authored by Alessandro Bellina, Giordano De Marzo, and David Garcia.
The preprint, “Conformity Generates Collective Misalignment in AI Agents Societies,” was authored by Giordano De Marzo, Alessandro Bellina, Claudio Castellano, Viola Priesemann, and David Garcia.