Why does opinion spread through networks?

In a 2006 experiment, Salganik, Dodds and Watts built an artificial music market in which 14,341 participants downloaded unknown songs, some with knowledge of what earlier participants had chosen and some without (Science, 2006 (opens in a new tab)). Stronger social influence made success both more unequal and less predictable. The best songs rarely did badly and the worst rarely did well, but any other result was possible.

Muchnik, Aral and Taylor measured conformity to visible numbers in a randomised experiment on a social news website, where positive social influence raised the likelihood of positive ratings by 32% and lifted final ratings by 25% on average (Science, 2013 (opens in a new tab)). Users corrected manipulated negative ratings instead, so herding was asymmetric. It also varied by topic and by whether readers were viewing the opinions of friends or of enemies.

Centola’s 2010 experiment spread a health behaviour through artificially structured online communities (Science, 2010 (opens in a new tab)). People were much more likely to adopt it when several neighbours reinforced it, and it travelled farther and faster through clustered networks than through random ones. Real clusters form around similarity: McPherson, Smith-Lovin and Cook’s review shows people connecting with similar others in marriage, friendship, work and advice, which limits the information they receive and the attitudes they form (Annual Review of Sociology, 2001 (opens in a new tab)).

In Bail and colleagues’ month-long experiment, contact with the other party hardened some views (PNAS, 2018 (opens in a new tab)). Democrats and Republicans who used Twitter at least three times a week were paid to follow a bot that showed them messages from the opposing side. Republicans became substantially more conservative; the Democrats’ slight shift was not statistically significant. James He, now Artificial Societies’ CEO, was one of the authors of a study of 464 and then 1,600 Twitter users in which people with low and high confidence in COVID-19 vaccines separated into two distinct online communities (PNAS Nexus, 2022 (opens in a new tab)).

We use “independent AI respondents” for language-model personas that each answer a questionnaire alone, the approach known as silicon sampling. The standard version, as our method page describes it, prompts a large language model (LLM) with an invented biography such as “you are a 34-year-old teacher from Ohio”. Each answer is a private first reaction, and nothing one persona says reaches another. In the terms of Salganik and colleagues’ music market, such a panel reproduces only the condition in which nobody could see anyone else’s choices. Networked audience simulation links the personas instead, so that a reaction can depend on the people a persona is connected to.

Source: the studies cited, and Artificial Societies’ method page, read 28 September 2026.

We would still use independent respondents for a private first reaction, such as a product name tested for immediate appeal. A policy announcement that one community will read through its trusted voices and then argue about turns on the other rows, and for that decision a panel of isolated respondents measures an audience that will never exist: one in which nobody talks.

Our method page sets out three stages. It starts from observations of real people: published research, online commentary, reviews, anonymised social data and public records of voting and investment. Psychometric methods, the statistical tools used to measure attitudes and traits, then reconstruct how each individual reasons, and every persona is validated against audience benchmarks and prior behaviour before it enters a society.

In the society-creation stage, personas are clustered by shared characteristics and connected through real relationships and patterns of influence. Artificial Societies’ method page, read on 28 September 2026, specifies 12 to 3,500 personas per society, with results readable at market, segment or persona level. Artificial Societies’ September 2026 simulation of reactions to Dario Amodei’s essay “We Must Pace the Frontier” shows the scale: 4,491 simulated people in three societies, Tech Twitter (1,884), US tech leaders (1,526) and the US public (1,081), each read separately. What is a synthetic audience? covers the audience-level definition, and How artificial societies are built covers each stage.

Artificial Societies’ Survey Evaluation Report (January 2026) measured 86% distribution accuracy, the overlap between simulated and human answer distributions, across 1,000 real-world surveys. The comparator is a human self-replication ceiling of 91%, since people asked the same question twice give the same answer about 91% of the time; biography-prompted LLMs reached 67%. The report scores answers to survey questions, not how fast or how far an opinion travels through a society, and the peer-reviewed evidence for the networked stage so far concerns community formation.

Opinion dynamics simulation, and social network simulation more widely, predate language models by decades. In Growing Artificial Societies (1996), Epstein and Axtell’s Sugarscape model showed group formation, cultural transmission, combat and trade emerging from agents that each followed a few simple rules (MIT Press (opens in a new tab)). He and colleagues’ 2024 paper notes that such classical set-ups often simplify and homogenise what agents do, which limits how much real social variety they capture.

The same paper notes that researchers have proposed language-model agents to address that limit. Park and colleagues’ 2023 generative agents study, which Generative agents explained covers in full, put 25 such agents in a simulated town (UIST ’23 (opens in a new tab)). The number who knew about one agent’s Valentine’s Day party grew from 1 (4%) to 13 (52%). Network density, the share of possible relationships that existed, rose from 0.167 to 0.74 as the agents got to know one another.

The peer-reviewed study behind Artificial Societies’ network modelling is He, Wallis, Gvirtz and Rathje’s Artificial intelligence chatbots mimic human collective behaviour (opens in a new tab), British Journal of Psychology, December 2024. The authors analysed the first 28 days of Chirper.ai, a platform populated only by AI chatbots, most of them powered by GPT-3.5: 33,299 chatbots and 312,969 social engagements. Communities formed around a shared language and, the paper reports, among 17,746 chatbots that mainly used English, around similar posted content. That is homophily, the tendency to form communities with similar others, arising in agents with no instruction to socialise like humans. The data come from one platform outside the researchers’ control, and the authors write that homophily alone cannot establish that chatbots emulate human societies; they name social influence and conformity among the behaviours still to test.

Ashery, Aiello and Baronchelli showed shared conventions emerging spontaneously in populations of LLM agents, with strong collective biases appearing even when no agent was biased on its own, and committed minorities of adversarial agents able to impose a new convention on the rest (Science Advances, 2025 (opens in a new tab)). Flint and colleagues found that interaction can amplify an individual model’s biases, add new ones or override its preferences, and that group size changes the dynamics non-linearly (PNAS, 2026 (opens in a new tab)). AI safety researchers study the same effects from the other side, as What is group alignment in multi-agent AI? explains.

OASIS (Open Agent Social Interaction Simulations) is an LLM social simulation of platforms. Yang and colleagues’ system models up to one million users on versions of X and Reddit, with a recommendation system, a follow network that updates as agents act, and 21 types of action (arXiv 2411.11581 (opens in a new tab), version 5, March 2025). The authors report that it reproduces information spreading, group polarisation and herd effects.

Against 198 real Twitter cascades, Yang and colleagues report that simulated spread followed real-world trends well, with a normalised root-mean-square error, the average gap between simulated and real curves, of around 30%. In their replication of Muchnik, Aral and Taylor’s 2013 rating experiment, agents behaved like the human participants when a post had been up-voted. When it had been down-voted, agents added further dislikes where humans had corrected the manipulation. In the polarisation test, agents’ responses grew more extreme as the interaction went on.

In that herd test, language-model agents conformed more readily than people, so a cascade inside any networked simulation, ours included, needs a check on whether real people would have joined it. Conformity is not yet part of Artificial Societies’ published evaluation, which leaves the networked stage without a published measure of its own.