AI agents formed polarized opinions in simulated social networks

Across thousands of language-model agents, patterns of friendship and disagreement emerged as they discussed political issues.

Editorial illustration — not from the study.

Researchers simulated networks of thousands of language-model agents using several different underlying systems. The agents held conversations, updated their opinions and formed social connections, with collective views becoming polarized over time. Their networks showed homophily, meaning agents tended to cluster with others more like themselves.

The researchers then used these simulated networks to examine possible ways to influence opinion dynamics. They tested interventions including encouraging more diverse interactions and reducing confirmation bias, a tendency to favor information that supports existing views.

How the opinions divided

The agents developed polarized opinions through repeated, language-model-guided interactions. They also spontaneously formed social networks with features associated with human networks, including homophilic clustering, in which similar agents tend to group together. Across the language-model systems examined—including GPT-3.5, GPT-4o, ChatGLM, Llama-3 and DeepSeek-V3—the agents' collective opinion changes showed patterns consistent with mechanisms discussed in studies of human behavior and models of opinion dynamics. The simulated networks were also used to examine network- and individual-level interventions, such as promoting more varied interactions and reducing confirmation bias.

Why these simulations matter

The findings suggest that networks of language-model agents can provide a controlled setting for studying how opinions may spread, cluster and become divided in interacting artificial systems. This setting could help researchers test hypotheses about opinion dynamics and possible interventions before examining them in more complex networked systems. The results do not show that human groups will behave in exactly the same way, but they indicate that these agents can reproduce some patterns associated with human opinion dynamics.

Evidence and caveats

This was a simulation study, not an observation of human social networks or a test involving people. The researchers used thousands of agents across several language-model systems and examined their behavior during simulated conversations. The abstract does not report the size of the polarization effects, provide details of every simulation condition or establish that the tested interventions would reduce polarization in real-world communities. Similarities between the agents' behavior and human social phenomena therefore support their use as a synthetic testbed, rather than proving that the agents replicate human behavior.

A proposed experiment could make gravity appear to push instead of pull

A theoretical study proposes an experiment in which a mass placed in a quantum superposition could produce an effectively repulsive gravitational effect on another particle. The effect would arise from quantum interference and, if observed, would conflict with general relativity’s description of gravity as always attractive.

1 day ago·29 views

Physics & Space

Silicon quantum bits keep precise control when run together

Researchers report that silicon spin qubits can perform precise operations in parallel, overcoming a major source of error during simultaneous control. In a five-qubit array, gate fidelities stayed above 99.99% for up to three qubits and above 99.9% when all five operated together.

1 day ago·11 views

Physics & Space

Mercury’s magnetic shield may host recurring solar-wind outflows

A survey of MESSENGER measurements identified 116 candidate magnetic reconnection layers near Mercury’s magnetopause. The structures appear to recur under the planet’s typically low-Alfvén-Mach-number solar wind, but they cannot yet be confirmed as reconnection outflows.

6 days ago·51 views