How artificial societies are built: 3 stages explained

Across 1,000 real-world surveys, Artificial Societies’ 86% distribution accuracy is 95% of the 91% human self-replication ceiling, according to its Survey Evaluation Report, January 2026. Distribution accuracy measures the overlap between simulated and human opinion distributions. The ceiling reflects the variability in people’s answers when asked again. The report is internal research; the peer-reviewed evidence concerns collective behaviour and is a separate result.

The evaluation distinguishes agreement with a population from consistency inside a persona and variety in its language. A population-level match can coexist with the wrong combinations of beliefs inside its members, as our September 2026 validity framework explains. We would judge a simulation by those relationships as well as its headline distribution score.

Source: Artificial Societies, Survey Evaluation Report, January 2026, with definitions and comparators on our evaluation page.

The labels also set the limits. Hallucination here means self-contradiction, rather than factual error about the outside world. Cronbach’s alpha measures whether related questions hold together; it cannot establish that findings generalise to a real audience. Word richness measures language variety. Keeping these measures separate lets you ask whether a result supports a claim about population opinion, coherent attitudes or the substance of an open answer.

The peer-reviewed basis for Artificial Societies’ group modelling is Artificial intelligence chatbots mimic human collective behaviour (opens in a new tab), by He, Wallis, Gvirtz and Rathje, published online in the British Journal of Psychology in December 2024. It examines homophily: people’s tendency to form communities with similar others. The research asks whether groups of chatbots exhibit that collective behaviour, extending the question beyond the behaviour of an individual chatbot.

The simulated society contained 33,299 chatbots powered by large language models, or LLMs, according to He and colleagues’ December 2024 paper. Communities formed around bots using a common language. Within the predominantly English-speaking subset of 17,746 chatbots, communities formed around similar posted content, the same paper reports. The agents reproduced this pattern without instructions to do so. The authors describe the finding cautiously: “These initial empirical findings suggest that AI chatbots mimic homophily, a key aspect of human collective behaviour.”

For our method, the value is evidence that connected agents can reproduce a documented feature of community formation. Our method page traces the approach to this research. The paper validates that group behaviour; it does not validate the commercial survey-accuracy figures. Those belong to Artificial Societies’ January 2026 evaluation, which is not peer-reviewed. You can therefore inspect the evidence for modelling a connected group and the evidence for matching survey responses as distinct scientific claims.

Persona Genesis uses observations of anonymised real people to reconstruct individual belief systems. Our method page names published research, reviews and online commentary, alongside anonymised social data and public records of voting and investment. Its sources include LinkedIn, X, Reddit, Facebook, YouTube, Instagram and Amazon. The starting material concerns how people communicate and behave in context.

Our Behavioural Analysis Engine uses psychometric methods to infer beliefs from those observations. Here psychometrics means reconstructing how an individual reasons, with attitudes that can be checked against related answers. Persona construction turns the observed material into a model that can respond to scenarios beyond the original observations. Before a persona enters a society, we validate it against audience benchmarks and prior behaviour. That check sits between construction and participation in a simulation.

The alternative baseline in our evaluation is a biography-prompted LLM: a model asked to respond as an invented person. Such personas reached 67% distribution accuracy against Artificial Societies’ 86% and the 91% human ceiling in the January 2026 Survey Evaluation Report. The report compares the approaches on overlap with human opinion distributions.

Our product description calls this stage Persona Genesis. The evaluation page separates its data sources from persona construction, so its diagram divides this part of the work into separate steps. The method page also specifies validation against audience benchmarks and prior behaviour before a persona enters a society.

Society Assembly groups personas by shared characteristics and connects them through real relationships and patterns of influence. Artificial Societies’ method write-up describes societies containing 12 to 3,500 personas. The network retains individual belief systems while adding the relationships through which opinion moves. You can study a population without reducing its members to a single average response.

Network science enters at this stage. The method treats connections as part of the audience being modelled, alongside the beliefs of the people in it. The homophily result in the British Journal of Psychology paper gives this choice a concrete scientific basis: the simulated agents formed communities around shared language and similar content. Our use of connected personas follows that interest in collective behaviour, and why opinion spreads through networks sets out what the connections add to a study of an audience.

For a strategist, the distinction concerns the object of the study. An independent answer describes how a persona responds to a stimulus; a society includes the surrounding network of influence in the model. Our method combines the mapping of individual belief systems with network modelling because opinion at group level is part of the question we are trying to answer. The resulting analysis remains available at market, segment and persona level.

The Simulation Engine applies individual beliefs and social network dynamics to a scenario. Each run begins with context based on the audience’s current situation, as our product description explains. Personas then encounter the stimulus. Their responses produce opinion at individual, segment and population level, with an explanation of why each persona answered as it did.

You can read the aggregate response, examine a segment and inspect an individual answer’s reasoning. A persona’s explanation helps you interpret its answer; survey evaluation provides the separate comparison with human responses. We would examine both when deciding how much weight to give a finding, especially when a persuasive explanation accompanies an unexpected population result.

The context matters to the work we use simulation for: strategic communications, investor communications and corporate strategy are among the decision areas our method page describes. The stimulus is considered in an audience’s situation, with its beliefs and relationships represented. Our case studies show how those builds become research engagements; they provide a route from the construction described here to the decisions clients have tested.

Our method combines psychometrics and network science, with social science and psychology informing how we evaluate it. Psychometrics supports the reconstruction of individual reasoning; network modelling represents relationships between personas. Artificial Societies’ September 2026 validity framework, by Edoardo Chidichimo and Felix Wallis, sets out eight tests for assessing the resulting simulations. The framework draws on Campbell (1957) and Cronbach and Meehl (1955).

Internal validity asks whether the stimulus caused the response shift. The tests include rephrasing a question and changing the order of response options. An answer should respond to substance without drifting merely because the questionnaire changed its presentation. The framework also tests whether changing the messenger changes reception and whether removing supporting evidence weakens a claim’s score.

Construct validity asks whether questions measure the attitude they are meant to measure. External validity asks whether results hold for real populations. Our framework tests external validity through response distributions in granular audience samples, the size of responses to experimental stimuli, and the relationships between answers. Matching a survey total is one part of that assessment.

These tests make the method open to a more demanding inspection than fluent persona answers alone. You can ask how a simulation behaves when wording changes, when evidence disappears or when linked questions probe a belief.

Our founders’ roles connect research, product development and engineering. James He, co-founder and CEO, guides deployment strategy and foundational research. His team biography states: “At Cambridge, he researched human behaviour using data science and agent-based simulation.” Head of Research Felix Wallis co-authored the British Journal of Psychology paper with James. His biography says he studied online community simulation while completing his master’s degree in Social Data Science at Oxford; he now develops our machine learning and evaluation methods.

Patrick Sharpe, co-founder and CPO, previously designed behavioural experiments and conducted customer research at ReAssure and Swiss Re’s iptiQ. He leads product strategy and development. Tom Whittle, co-founder and CTO, oversees simulation engineering. His background includes consultancy and work at growing startups, as described in the team biographies.

The publications page connects those biographies to specific work. Alongside the simulation paper, it lists James He’s co-authored study, Social Media Behavior Is Associated with Vaccine Hesitancy, in PNAS Nexus, September 2022. It also includes Edoardo Chidichimo’s co-authored Perspective, Towards an Informational Account of Interpersonal Coordination, in Nature Reviews Neuroscience, November 2025. These publications describe the team’s academic work; the survey results remain attributable to the separately published internal evaluation.

The same interest in how collectives behave runs through the team’s work on AI safety. AI Researchers Should Use Group Alignment to Reduce P(doom), published on 16 September 2026 by Emmanuelle Gelain-Sohn of Faculty with Yitian Chen and Felix Wallis of Artificial Societies, argues that AI agents aligned one at a time can still form a misaligned group. It defines a group as group-aligned when its interaction structure reliably produces joint behaviour that respects human values, and proposes “mixed swarms” that combine frontier agents with models of human behaviour. What is group alignment in multi-agent AI? explains the idea for a non-specialist reader.

An artificial society fits research into how an observed audience responds to a decision, including the relationships within that audience. Our method can model audiences for which diverse observations exist, whether public or client-provided. That requirement gives persona construction something to test against. If you hold relevant first-party research, it can join the observations used to build the persona set.

We would use that approach to examine opinion before a consequential decision, and commission direct behavioural evidence when the decision requires it. Our September 2026 validity framework explains the intention–behaviour gap: what people say can be a weak predictor of what they do. Simulation inherits that limitation from surveys. For an audience with no diverse observations, gathering those observations comes before persona construction; accuracy measured on observed populations cannot establish performance on an unobserved one.