October 23, 2026
Andrew Zinin
Author
Lisa Lock
Scientific Editor
Robert Egan
Senior Editor
Give AI agents political positions, ask them to defend their views and persuade one another, and let them decide which contacts to keep. This recent experiment, published in Nature Communications, found that over repeated exchanges, something familiar emerged: opposing camps that increasingly prefer their own company.
In the simulations created by Jinghua Piao of Tsinghua University in China and colleagues, the agents were not explicitly instructed to favor like-minded partners or form echo chambers. Yet their exchanges produced increasingly divided opinions and networks.
Once those divisions took hold, the researchers tried to loosen them. Would changing whom the agents encountered help—or did the agents need different instructions about how to listen?
How the middle thinned out #
For their study, the team built networks of 1,000 agents discussing one issue at a time: U.S. party alignment, gun control or abortion bans. The main experiments used GPT-3.5, with additional tests involving other language models. Each agent began with an assigned position on a five-point political scale: left, moderate left, neutral, moderate right, and right. About 40% started neutral, making the middle the largest group.
During each round, agents generated reasons for their positions and messages intended to persuade others. They could maintain existing contacts or encounter someone new, then update their stated opinions using the messages they received.
After 40 rounds, the neutral share had fallen to 22.5% for party alignment, 0.4% for gun control and 5.1% for abortion bans. Agents increasingly interacted with others on their own side. Exchanges with more strongly opinionated allies tended to push their positions further from neutral.
But some agents also struggled to stay in character. One assigned a neutral stance declared strong support for the Democratic Party. In separate tests, right-leaning agents sometimes switched left even after talking to someone who shared their position.
The team added a checking step, prompting agents to review their messages and opinion changes and regenerate inconsistent responses. The left- and right-leaning camps became more balanced, but the divisions persisted. More consistent role-playing did not prevent polarization.
Meeting an opponent was not enough #
The researchers then tested five interventions in an already polarized network discussing party alignment, comparing its final rounds with a version that continued without intervention.
Randomizing conversation partners produced only a modest reduction in polarization. Allowing messages only from moderately opposing agents also had only a small effect. Simply prompting agents to seek diverse views did little once their connections were established.
A larger change came from instructing agents to be receptive to persuasion from different viewpoints. That reduced the study's polarization score by 11.8%. Adding influential agents that broadcast neutral messages reduced it by 8.8%.
The score measured the average distance of agents' stated positions from neutral, rather than hostility between camps. These results therefore show shifts in simulated opinions, not reconciliation.
The authors write that the intervention findings "should be understood as hypothesis-generating," because prompts shaped the agents' behavior, their opinions occupied just five categories, and resemblance to human social patterns does not establish that the same interventions would work on people.
For the AI networks tested here, however, the distinction was clear: Bringing different views into a conversation had less of an impact than altering how the agents were instructed to respond to them.
Written for you by our author [Andrew Zinin](https://sciencex.com/help/editorial-team/andrew-zinin/), edited by [Lisa Lock](https://sciencex.com/help/editorial-team/lisa-lock/), and fact-checked and reviewed by [Robert Egan](https://sciencex.com/help/editorial-team/robert-egan/)—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive.
If this reporting matters to you, please consider a [donation](https://sciencex.com/donate/?utm_source=story&utm_medium=story&utm_campaign=story) (especially monthly). You'll get an **ad-free** account as a thank-you.
Publication details
Jinghua Piao et al., Emergence of polarization in networks of large language model agents, Nature Communications (2026). DOI: 10.1038/s41467-026-78228-y
Journal information:
Nature Communications © 2026 Science X Network
Citation: AI agents can form political camps—and changing their conversation partners does little to soften their views (2026, October 23) retrieved 23 October 2026 from https://phys.org/news/2026-10-ai-agents-political-conversation-partners.html