Thousands of language-model agents developed like-minded connections and diverging opinion groups in simulations reported in Nature Communications. The result shows why an AI agent’s behaviour in a group can differ from what an isolated conversation reveals.[1]
Jinghua Piao and colleagues studied simulated networks using five model backbones: GPT-3.5, GPT-4o, ChatGLM, Llama-3 and DeepSeek-V3. In the paper, published on 8 October, agents exchanged views on political questions and changed their connections. The resulting patterns included homophily, the tendency to associate with similar others, and opinion polarisation.[1]
The network changes the conversation
Opinions affect whom an agent chooses to hear, while those contacts affect the opinions it encounters next. Allowing both to change creates a feedback loop. The simulation gives researchers a way to observe that loop and vary parts of it while holding other experimental conditions steady.[1]
These agents respond according to their models, prompts, memory and conversation histories. The resemblance to human social patterns can motivate hypotheses, but it does not establish that people respond for the same reasons. The experiment directly measures a synthetic network.[1]
That network is an important object of study in its own right. As assistants exchange information or divide work, collective dynamics may affect the evidence a group notices and the conclusions it reaches. Assessing each member separately cannot reveal all of those interactions.[1]
Changing contacts and response patterns
The paper examines interventions at two levels: the network of contacts and individual tendencies such as confirmation bias. Increasing contact diversity changes which views can reach an agent; changing its response tendencies affects how it treats those views. Studying the two separately helps distinguish access to contrary information from willingness to use it.[1]
For developers of agent teams, this suggests testing whether a group revises a conclusion when contrary evidence arrives and whether its members keep drawing on independent sources. Those are practical evaluation questions prompted by the study, rather than additional outcomes measured here.[1]
Collective behaviour needs its own test
The next comparison is across model versions, prompts, network structures and memory designs: which group effects persist, and which depend on a particular setup? That would make the observed clustering more useful for designing systems in which agents must exchange evidence without simply amplifying agreement.[1]
Sources & further reading
- Emergence of polarization in networks of large language model agents Nature Communications · Accessed 24 Oct 2026