AI agents formed political camps—and changing their contacts had little effect

Let a thousand language-model agents choose conversation partners, defend political positions and revise their opinions, and camps begin to form. In experiments reported in Nature Communications, agents increasingly preferred like-minded contacts while the neutral middle shrank. Changing the network alone did relatively little to undo the divisions.[1][2]

The study, published on 8 October, used networks of 1,000 agents for its main experiments. GPT-3.5 agents debated U.S. party alignment, gun control and abortion bans on a five-point opinion scale, with roughly 40% starting neutral. The wider research also tested other model families. Each round combined argument generation, contact selection and an update to the agent’s stated position.[1][2]

How the center disappeared

After 40 rounds, the neutral proportion was 22.5% in the party-alignment scenario, 0.4% in the gun-control scenario and 5.1% in the abortion-ban scenario, according to the researchers' reporting. These numbers do not describe voters, election trends or the state of any real political community. They describe outputs of an artificial system whose rules were selected by the research team. Still, the contrast between the initial composition and the simulated ending is striking.[1][2]

A model agent can have a position without having the emotional and biographical investment that a person might. It can generate an argument without the experience of having lived through the events it invokes. The researchers' interest was therefore less in whether the agents were genuinely convinced and more in what kinds of collective patterns emerged when the generated messages and network connections fed back into one another. Agents increasingly preferred exchanges with others near their own stated positions, and strongly opinionated allies could pull an agent further away from neutrality.[1][2]

The investigators also discovered a methodological problem: agents did not always maintain the positions they were meant to play. In some tests, an agent assigned one stance generated an inconsistent declaration. A consistency-checking process reduced some of that instability, yet polarized clusters still formed. This result underlines the need to distinguish simulated social dynamics from mere instability in a chatbot's role-play instructions.[1][2]

Contact diversity and receptiveness are different variables

The researchers then tried several interventions within already polarized networks. Randomizing conversation partners offered only a modest reduction in the polarization metric. Encouraging agents to seek diverse viewpoints likewise changed little after relationships were established. Restricting some interactions to moderately opposing agents had a limited effect. These results challenge the simple idea that exposing a system to different opinions necessarily produces convergence.[1][2]

An instruction to remain receptive to persuasion from different viewpoints had a bigger reported effect: the study's polarization score fell by 11.8%. Adding influential agents broadcasting neutral messages reduced it by 8.8%. The metric captured average distance from the center of the opinion scale, not trust between the sides, warmth toward opponents or the likelihood of social reconciliation. A network can move toward a midpoint numerically without becoming more tolerant in any deeply human sense.[1][2]

The intervention changes a prompt, not a person’s history or beliefs. The researchers accordingly treat these experiments as a way to generate hypotheses about opinion dynamics, with human experiments and calibration against real-world data needed to test their wider relevance.[1][2]

What the network adds to a chatbot test

The revealing distinction is between a single response and a system of responses feeding back into one another. Evaluating an agent in isolation may miss patterns that arise when agents choose allies, repeat arguments and influence the next round of contact. Here, changing the rule for considering another view moved the outcome more than simply placing that view in front of the agent.[1][2]

The experiment gives researchers a controlled setting in which to vary those rules. Its most useful result is not a forecast of human politics but a demonstration that conversational networks can amplify a model’s behaviour—and that network structure and individual response rules deserve separate tests.[1][2]

Report an error or suggest a correction