Do networks of large language model agents spontaneously develop opinion polarization?

Polarization emerges without being programmed

The study builds a synthetic society of roughly 1,000 LLM agents per experiment, drawn from several backbone models including GPT-3.5, GPT-4o, ChatGLM, Llama-3 and DeepSeek-V3. Agents converse in natural language, choose their own social contacts, and revise their opinions on political issues over repeated timesteps. No rule tells them to sort into camps. Yet their collective opinions converge into a polarized pattern, with agents clustering at the left- and right-leaning extremes rather than at the neutral middle.

The authors measure polarization as the average distance between agents' opinions and the neutral position, on a five-point scale from strongly left to strongly right. The score ranges from 0 to 2, and a larger value means a more polarized system. This is a property of the whole population at a given timestep, not of any single conversation, so the finding is about collective order emerging from many local exchanges.

One design detail matters for reading the result. Early runs drifted toward a left-leaning imbalance, which the authors attribute to the models' own political bias leaking into agents' behaviour. They add a self-regulation step in which agents check whether their messages, opinions and stated reasons are mutually consistent and regenerate until they are. This cuts the self-inconsistency rate by 9.4–52.2% across issues and produces a more balanced right-left split. The authors argue that polarization therefore comes from social interaction rather than from model bias alone, since self-regulated agents still polarize.

Agents build a homophilic social network

The network is not fixed in advance. Agents decide whom to interact with, and those choices accumulate into a structure. The paper reports that the share of interactions between agents in the same camp rises by 156.8–382.7% over time, ending with 48.5–88.3% of interactions occurring between like-minded agents. Agents also increasingly avoid opposing camps. The authors call this homophilic clustering, meaning the tendency of similar agents to associate, and note that no in-group bias was hard-coded into the system.

The resulting networks resemble human ones in further respects. The paper reports a scale-free degree distribution, in which a small number of agents become highly popular, alongside modularity, assortativity and homophily metrics. A control comparison is especially informative: when the self-organized network is replaced by a random or static network, one camp's opinions dominate the whole system. The self-organized structure is therefore not decorative; it is what allows two camps to persist side by side.

Echo chambers and backfire both raise polarization

Two mechanisms connect the network to the opinions. The first is the echo chamber effect: when agents interact with more radical members of their own camp, their polarization level rises. The paper supports this with two-sided Student's t-tests across three issues, for example partisan alignment t(169) = 16.80, p ≪ .001, Cohen's d = 1.288, and gun control t(1465) = 19.37, p ≪ .001, Cohen's d = 0.506. These tests compare polarization after interactions with more versus less radical same-camp neighbours, and the sample sizes are counts of agent interactions, not agents.

The second is the backfire effect. Homophilic interactions are the majority, at 54.8–82.2%, but agents still encounter opposing views in 7.4–10.2% of interactions. Those encounters do not reliably pull opinions back toward the centre; the paper finds that exposure to opposing agents can instead increase polarization. The authors are careful to note that echo chamber and backfire effects are heterogeneous in real human populations, and they report that their simulated system reproduces some of that heterogeneity rather than applying the effects uniformly.

Individual traits shift the whole system's polarization

The paper then varies how many agents carry traits borrowed from human opinion research. Selective exposure is the tendency to seek out like-minded contacts; confirmation bias is the tendency to read messages in a way that fits existing beliefs. In comparative experiments, raising the share of agents exhibiting either behaviour markedly raises system-level polarization, with 0% and 100% deliberately included as extreme comparison cases.

A third mechanism concerns influential voices. The authors add top influencers who send non-personalized messages to everyone, distributed symmetrically across the two camps. When those influencers hold non-neutral opinions, final polarization rises significantly; when they hold neutral opinions, polarization falls by 28.5%. The relevant statistics are two-sided Student's t-tests with n = 5 timestep-level polarization values, for instance t(8) = 21.56, p ≪ .001, Cohen's d = 13.636 for the neutral-influencer comparison.

These experiments are interventions on a model, not observations of human populations. Their value is that the direction of each effect matches patterns reported in empirical studies of human opinion formation, which is the basis for the authors' claim that the system can serve as a synthetic testbed.

Individual openness beats rewiring the network

The intervention experiments start from a system that has already converged to a polarized state at timestep 35, then compare its subsequent trajectory with an untouched copy. Five strategies are tested. At the network level, random interactions remove homophilic clustering, and moderate opposing interactions restrict incoming messages to moderate opposing views. At the individual level, agents are prompted to have no selective exposure, to have no confirmation bias, or to receive neutral messages from new top influencers.

All five significantly reduce polarization, but the sizes differ in a way that matters. The two individual-level strategies of removing confirmation bias and adding neutral elite signaling produce the largest reductions, at 11.8% and 8.8% respectively. Moderate opposing interactions reduce polarization significantly but shift opinions by only about 2%, a modest effect the authors note is consistent with human studies of cross-cutting exposure. Removing selective exposure has limited effect: in an already stable network, agents' willingness to talk to anyone does not expose them to diverse views and instead quickly crystallizes the existing structure.

The paper's summary comparison is that individual-level strategies are associated with larger observed reductions than network-level interventions. The authors also flag a practical asymmetry: recommending openness is easier to implement in a simulation than in human populations. All strategies reduced homophilic interactions by 2.6% to 100% while increasing inclusive conversation, so reduced clustering alone did not determine how much polarization fell.

Where the synthetic testbed claim stops

The core results rest mainly on OpenAI's ChatGPT models and three political issues, though the authors report consistent findings with other backbone models, with midway model substitution, and on immigration restriction and flat-Earth topics. They also vary the initial network structure and initial opinion distribution and find similar polarization patterns and final network structures.

The authors state that the work makes no theoretical contribution and follows the argument-accumulation account, in which polarization grows from accumulating and reinforcing persuasive arguments rather than from simple opinion averaging. Consistent with that, agents who become more polarized accumulate more supporting arguments over time. The authors also acknowledge that generative agent-based models trade interpretability for realism, that agent behaviour is shaped by prompt design and simplified opinion representations, and that how closely these patterns match real human behaviour quantitatively still needs validation.

The most useful reading is therefore a priority ordering rather than a verdict on any single lever. In this testbed, once a system is polarized, changing what individual agents are willing to hear and how open-mindedly they process it moved polarization further than changing the contact network itself, while the network-level fix that looked most intuitive, removing selective exposure, barely helped because the existing structure absorbed it.

Sources used in this answer