What is a synthetic audience?

A synthetic audience is built in four steps: decide who belongs in it, gather observations of those people, turn the observations into personas, and connect the personas to each other. At Artificial Societies, observations come from published research, online commentary, reviews, anonymised social data and public records of voting and investment, with the client’s own research as an optional layer. Each persona is validated against audience benchmarks and prior behaviour before it joins, as described in What are synthetic personas?.

Membership shapes everything after it. Artificial Societies’ hyperscaler earnings-call study defined four societies around one stock: 14 sell-side analysts who had asked questions on the company’s last eight earnings calls, 250 buy-side investors at institutions holding the stock, 570 customers, suppliers, partners and policymakers, and more than 1,600 financial journalists. Artificial Societies’ misinformation study (September 2026) used a UK General Population society instead, reweighted to regional age, sex, graduate and ethnic-minority profiles when regions were compared.

Connection turns a list of personas into an audience. Artificial Societies clusters personas by shared characteristics and links them through real relationships and patterns of influence, in societies of 12 to 3,500 personas, according to its method page (read 28 September 2026). For its crisis simulation study, Artificial Societies built 1,174 personas across five societies (policymakers and regulators, media, influencers, consumers and commercial customers) and mapped each audience’s social influence network before testing a single message.

A synthetic audience differs from its neighbours in its unit. Personas, respondents and twins each describe one person or one set of answers, and silicon sampling pools such answers into a sample whose members never affect each other.

Sources: ICC/ESOMAR International Code, 2025 edition; AAPOR task force report, 2026; arXiv 2509.19088 (opens in a new tab); Argyle and colleagues, 2023 (opens in a new tab). All read 28 September 2026.

Argyle and colleagues’ 2023 paper, Out of One, Many, is the origin of silicon sampling, the closest neighbour in the table. It conditioned a language model on thousands of socio-demographic backstories from real participants in large US surveys and read the outputs as a simulated sample; the authors named the property they tested algorithmic fidelity. The synthetic respondents page compares these terms against industry definitions.

The unit decides what you can ask. A sample of independent respondents tells you what each person thinks on first sight. A connected audience is built to show, in addition, where a message spreads, where it stalls and which segment moves first, although Artificial Societies’ published evidence for that second kind of answer is still qualitative. If your decision depends on spread, a sample of independent respondents cannot answer it, whatever it is called.

Audience simulation is the act of running a synthetic audience through a scenario and measuring how it responds. In Artificial Societies’ crisis study, the run began with a survey of each society’s baseline sentiment and likely purchasing behaviour; 24 messages and creatives were then tested against that baseline. A simulated focus group, where a few personas discuss a topic with a moderator, is a different format, compared in AI focus groups and audience simulation.

Time can be simulated as well as exposure. In Artificial Societies’ misinformation study (September 2026), the human fact-check effect had shrunk to 8% of its original size two months later. Synthetic respondents do not live through the weeks between surveys, so Artificial Societies modelled forgetting with a recall rate from Guess and colleagues (2020), and the synthetic effect also shrank to 8%. The match depends on the imported forgetting rate; the personas did not discover the fade on their own.

A synthetic audience can test a statement, a script, a product concept or a correction before it is released, and its strongest uses are where the real audience cannot be asked. Artificial Societies’ earnings-call study tested unreleased numbers and draft scripts that selective-disclosure rules kept from real investors. Its simulated press coverage matched the real next-morning framing: growth in the headline, the margin reset as a cost story, financing as the emerging risk (earnings-call preparation).

Crisis responses fit the same pattern. Artificial Societies’ crisis study gathered 328,576 responses in under two weeks and found that evidence-backed messaging about the company’s remedial steps was the best path, as the case study reports. Crisis communication simulation and message testing cover the method.

Public-interest interventions can be tested too. In Artificial Societies’ misinformation study (September 2026), prebunking infographics, which warn people about manipulation techniques before they meet them, raised the rated manipulativeness of fake posts by 0.21 standard deviations in both the human and the synthetic group, a small effect in both. The fake headlines that fooled the most Britons also fooled the most synthetic respondents, with a rank correlation of 0.87, where 1 would mean the same order (prebunking versus debunking).

Synthetic audience accuracy starts with two questions: does the audience give the right spread of answers, and does it move by the right amount when something changes? On the first, Artificial Societies reports 86% distribution accuracy across 1,000 real-world surveys, against 67% for language models prompted with an invented biography and a 91% human self-replication ceiling, the rate at which people repeat their own answers (Survey Evaluation Report, January 2026). The accuracy page covers the other published measures.

A matching spread of answers does not prove the audience is right. Artificial Societies’ September 2026 validity framework shows two audiences with identical answer distributions on every question whose combinations of answers differ by a total variation distance of 0.4 (Distribution Matching Is a Bad Way to Evaluate Simulations, Chidichimo and Wallis). In plain terms, 40% of the synthetic respondents would have to change answer combination to match the humans. The framework sets eight validity tests and notes that the industry standard covers two.

Effect-size replication is the harder test. In Artificial Societies’ misinformation study (September 2026), five misinformation posts cut the share of Britons who would definitely accept a COVID-19 vaccine by 6.6 percentage points and the synthetic share by 8.2, inside the confidence interval of 3.8 to 9.0 points reported by Loomba and colleagues for their 2020 experiment. The design, instruments, code and panel were hashed and time-stamped before the first survey call, so the comparison could not be tuned after the results arrived.

The same study reports where the simulation missed. Shown factual posts, 1,000 real Britons showed no significant shift in vaccine intent, while Artificial Societies’ synthetic control group moved about four points towards vaccination, which suggests a small bias towards vaccine-positive content. The synthetic society represented the UK in 2026 and believed false claims somewhat less than the 2020 samples, so the comparisons rest on relative differences, not absolute rates.

Independent AI respondents miss the spread of opinion because each one answers alone, so a sample of them records first reactions only. A group can behave differently from its members even when every member is identical. Shen and colleagues found that organisations built from copies of one aligned AI model produced more effective but less aligned solutions than a single copy, across 12 business tasks (arXiv 2604.10290 (opens in a new tab), April 2026). That result concerns task outcomes, not opinion, but it shows interaction alone can change what identical agents produce; our group alignment post discusses it.

A paper by Artificial Societies researchers, He, Wallis, Gvirtz and Rathje in the British Journal of Psychology (opens in a new tab) (December 2024), tested whether a network of language-model agents forms social structure. In a simulated online society of 33,299 chatbots, communities formed around a common language, and among the 17,746 predominantly English-using chatbots, communities formed around similar posted content. That is homophily, the human tendency to cluster with similar others. The paper shows structure forming, not opinion passing along it.

Artificial Societies’ earnings-call study shows the practical version: its third stage modelled how the story would travel through the corporate ecosystem and the press after the call, a question no set of independent respondents can answer. Artificial Societies’ published evidence for spread is still thinner than its evidence for direct exposure, because the four misinformation replications compared responses to messages people saw themselves, not claims passed between them. Networked audience simulation sets out what the network adds and what remains to be tested.

A synthetic audience is the wrong tool in five situations, and each follows from how it is built. The first is an audience with no observations behind it: without a public trace or first-party research, there is nothing to ground personas in. The second is a physical or sensory test, such as a taste trial, which needs real people with the real product. The third is a decision that needs proof of behaviour: what people say is often a weak predictor of what they later do, and synthetic research inherits that intention–behaviour gap.

The fourth is a result you intend to publish as public opinion. AAPOR’s 2026 task force report notes that, before language models, terms such as poll, survey and public opinion assumed human respondents, and it requires AI-generated cases in a purported study of public opinion to be identified as AI-created. The fifth is a question about absolute levels, such as the exact share of a country that believes a claim today; the misinformation study compared relative differences for that reason.

Artificial Societies’ own framework states the boundary: “Synthetic audiences will never be a true substitute for the people they seek to represent.” Use a synthetic audience before the decision, while the real audience is unreachable or the material cannot yet be shown, and a human study when the decision needs the people themselves.