Thousands of AI agents began forming closer ties with others who held similar views as they talked. Their political opinions also grew more divided. In a simulated society without human participants, researchers examined collective behavior that would be difficult to see by evaluating each agent’s answers in isolation.
Ask one AI agent for its view on a political issue and the result is a single response. Let many agents communicate, and another process begins. They receive arguments, revise their positions and change whom they interact with. Opinions and relationships can evolve together over repeated conversations.
That process matters when considering systems made up of multiple agents. A response from one agent may look measured on its own. The behavior of a connected group also depends on which messages circulate repeatedly and which connections become stronger. Whether such patterns occur in deployed systems requires direct observation and testing.
A paper published in Nature Communications by Jinghua Piao and colleagues created a simulated society to study those collective changes. The participants were thousands of agents powered by large language models (LLMs). They formed social ties and discussed political issues. No human participants took part in these conversations.
How the agents formed opinions and relationships
The main topics were partisan alignment, gun control and abortion bans in US politics. The researchers represented each agent’s opinion on a five level scale, ranging from left through neutral to right. They initialized the distribution of opinions and the network of social connections before the agents began interacting.
Each interaction involved three stages. An agent first generated reasons for its current position. It then sent a persuasive message to an existing connection or a new contact. After receiving messages, an agent updated its opinion. The researchers repeated this process and observed changes in both the distribution of views and the network of contacts. arxiv.org
The simulation had an experimental design. Its initial opinion distribution, starting network and opinion scale were set by the researchers. Prompts also shaped how the agents behaved. The researchers did not, however, insert a rule instructing agents to associate with those who agreed with them. They examined whether that pattern would develop through interaction.
Neutral positions declined as two camps formed
At the start of each issue’s simulation, 40% of agents held a neutral position. After repeated interactions, that share fell to 22.5% for partisan alignment, 0.4% for gun control and 5.1% for abortion bans. Agents moving away from neutrality formed groups supporting opposing positions. These percentages describe the simulated environments in the study.
The agents’ relationships shifted along with their opinions. Interactions within the same opinion group became more common over time. Depending on the experimental condition, 48.5% to 88.3% of interactions eventually took place between agents with similar views. The researchers also observed a tendency to assess members of one’s own camp more favorably and avoid contact with opponents.
The sequence is central to the findings. The researchers did not simply ask agents already divided into two fixed communities what they believed. The agents exchanged arguments, changed positions and adjusted their relationships. Those evolving relationships affected the arguments they encountered next. The authors interpret polarization as a process in which opinion formation and network formation influence each other.
They also examined patterns resembling selective exposure and confirmation bias: interacting more often with similar others and processing messages in ways consistent with an existing position. Describing these patterns does notestablish that the agents have human psychology. The comparison concerns behavior observed in the simulation.
The researchers examined bias in the underlying models
The first simulations did not produce evenly sized camps. Left leaning opinions became more prevalent. Attributing that entire imbalance to social interaction would overlook the influence of the language models powering the agents.
The researchers identified instances in which agents responded inconsistently with their assigned political positions. In a separate test, agents sharing the same position sometimes showed asymmetric shifts after an interaction. The team treated this as a self consistency problem and adjusted the agents to respond more consistently with their assigned positions before running further network simulations.
The resulting camps were more balanced in size, but polarization still appeared. The underlying model’s bias therefore matters when interpreting which camp became larger. The researchers argue that repeated interactions and evolving relationships also contributed to the division itself. They examined the main pattern with additional models and variations in initial conditions. arxiv.org
What happened when the researchers intervened
The team then applied five interventions to a network that had already become polarized. Some changed patterns of contact: agents could interact randomly or receive messages from people holding moderate opposing views. Other conditions encouraged contact across different positions, prompted agents to remain open to persuasion, or introduced messages from neutral influential agents.
All five reduced the polarization measure used in the study, but their effects differed. Prompting agents to remain open to persuasion from different views reduced the measure by 11.8%. Messages from neutral influential agents reduced it by 8.8%. Contact with moderate opposing views also had an effect, although the change in agents’ opinions was approximately 2%.
The researchers found that changing whom agents encountered could have limited effects once a network had become established. Exposure to a different view and openness to that view are separate parts of the interaction. These findings describe interventions within the study’s simulated society. They do not show that the same changes would have the same effects on people or on the recommendation systems of real social platforms.
What the simulation can and cannot establish
The resemblance between the simulated network and familiar descriptions of human polarization is striking. The authors are careful about its limits. Further validation is needed to determine how closely the patterns correspond to actual human behavior.
People do not form political opinions through a five level scale alone. Relationships have histories, and identity, emotion and lived experience can shape how a message is received. The study does not claim to reproduce all of those factors. Agents guided by prompts must also be distinguished from people encountering the same arguments. The authors present their interventions as a way to generate hypotheses for further research, rather than proven remedies for political division.
For readers considering the growth of connected AI systems, the paper points to a useful question. The accuracy or safety of one agent’s response may reveal little about how a group behaves over time. Researchers and developers may also need to observe which agents gain influence, which arguments are repeated and whether communication becomes concentrated within increasingly separate groups. This is an implication for examining future systems, not a finding that deployed AI services already exhibit the same polarization.
A conversation between agents does more than produce an immediate answer. It can alter the contacts and arguments available in the next conversation. This study provides a controlled setting in which that cumulative process can be observed. The next step is to test how relationships and opinions develop in AI systems operating beyond the simulation.
