What are synthetic personas?

A synthetic persona can be built from an invented biography, from observations of a real person, or from that person’s own interview and survey answers. The method decides what you can later check the persona against.

Invented-biography prompting needs no data about any real person. Our method and evaluation page describes it as the standard approach: a language model receives a biography such as “you are a 34-year-old teacher from Ohio” and answers in character, which produces generic, stereotyped answers. The model supplies everything the biography leaves out from its training data, which is the likely source of that stereotyping.

Grounded personas start from evidence about a real person. At Artificial Societies each persona begins with a real individual, observed through published research, online commentary, reviews, anonymised social data and public records of voting and investment, from sources including LinkedIn, X, Reddit, YouTube and Amazon. Psychometric methods reconstruct how that individual reasons, and each persona is validated against audience benchmarks and prior behaviour. The persona is judged as a member of an audience, not scored against its original person, which is what separates it from a digital twin. Our explainer on AI persona simulation sets out the full pipeline.

Academic work adds agents built from a person’s own self-reports. Park and colleagues built agents for 1,052 Americans by applying language models to qualitative interviews about their lives, and in the paper’s June 2026 version from survey answers as well (arXiv 2411.10109 (opens in a new tab)). Each agent stands for one participant and is scored against that person’s answers to questions it never saw, which makes it a digital twin.

Output quality tracks the method. In Artificial Societies’ January 2026 Survey Evaluation Report, biography-prompted LLM personas contradicted themselves in up to 35% of cases, against under 2% for Artificial Societies’ grounded personas and about 9% for human panels. On distinct-word richness, a measure of how varied the vocabulary of open answers is, the same report scores biography-prompted personas at 61%, grounded personas at 93% and a benchmark of 120,000 human social-media posts at 91%. The contradiction rate matters once you read a segment’s reasoning instead of its totals.

The four terms overlap because each names a different aspect of one object. The persona is the simulated person, and a respondent is the role it plays when it answers a questionnaire. A twin is a persona built from one individual’s own data and judged against that individual. “AI persona” is the everyday synonym, and the one our own pages use.

Sources: ICC/ESOMAR International Code, 2025 edition; AAPOR task force report, 2026; Park and colleagues, June 2026; Artificial Societies, method and evaluation, read 28 September 2026.

A 2026 task force report accepted by the American Association for Public Opinion Research (AAPOR) prefers “synthetic responses” to “synthetic samples”, because generating answers estimates what a set of people would say; it does not sample them. Our page on synthetic respondents compares these terms against industry definitions.

The accuracy figures for synthetic personas in this section answer two different questions. Individual accuracy asks whether a persona gives the answers its real counterpart gave. Distribution accuracy asks whether a population of personas splits across the answer options the way a population of people does. Park and colleagues and Artificial Societies read their figures against human test–retest consistency, the rate at which people repeat their own answer when asked again; Peng and colleagues report a raw correlation.

Sources: arXiv 2411.10109 (opens in a new tab); arXiv 2509.19088 (opens in a new tab); Artificial Societies, Survey Evaluation Report, January 2026, as published on the method and evaluation page. All read 28 September 2026.

The two 86% figures in that table measure different things. Park and colleagues’ figure is a ratio: the agents’ match rate divided by the participants’ own two-week consistency, so the share of answers an agent actually matched is lower than 86%. Artificial Societies’ figure is the overlap between simulated and human opinion distributions, five points below the 91% at which people reproduce their own answers. It establishes that a population of personas splits the right way, but not that any single persona answered as a particular person would; our accuracy page covers Artificial Societies’ other published measures.

Peng, Toubia and colleagues report a weaker result for digital twins. Twins built from each person’s answers to over 500 questions correlated with that person’s responses at an average r of 0.20, where 1 is perfect agreement and 0 is none (Digital Twins as Funhouse Mirrors (opens in a new tab), version of 19 April 2026). They were only modestly more accurate than a base language model given no individual data. At that correlation, a twin’s answer tells you little about which way its individual will go.

Peng, Toubia and colleagues’ April 2026 paper, Digital Twins as Funhouse Mirrors, named five distortions in digital twins. They are the failure modes to test any synthetic persona for:

• •

  Insufficient individuation: twin answers were less spread out than human answers in 154 of 164 outcomes (93.9%), so the simulated population is more uniform than the real one.

• •

  Stereotyping: twins relied heavily on demographic characteristics and made little use of the individual information they were given.

• •

  Representation bias: twins were more accurate for participants with higher education and higher income.

• •

  Ideological bias: twins expressed more pro-human views than the people they modelled, and also showed pro-technology tendencies.

• •

  Hyper-rationality: twins tended to display perfect knowledge on questions with a correct answer and were more likely than people to choose the textbook-rational answer.

Artificial Societies’ September 2026 validity framework by Edoardo Chidichimo and Felix Wallis, Distribution Matching Is a Bad Way to Evaluate Simulations, describes three further failures that a fluent answer hides. Sycophancy is a language-model persona agreeing with whatever a question implies, far more often than people do, so confident wording can move a persona that was never persuaded. Weak discrimination is a persona rating a new product high on quality, value, trustworthiness and likelihood to recommend at once, although in real consumer data those judgements come apart. The third is exposure: major public survey programmes appear in language-model training data, so matching them tests memory rather than prediction.

Two limits sit outside the model. What people say is often a weak predictor of what they later do, and synthetic personas inherit that intention–behaviour gap from the surveys they are scored against. An audience with no public trace and no first-party research gives a persona nothing to be grounded in, and a taste trial or other sensory test needs real people handling the real product.

A synthetic persona is one simulated person; a synthetic audience is a group of them, and in a networked build its members influence each other. Artificial Societies clusters validated personas by shared characteristics and connects them through real relationships and patterns of influence, in societies of 12 to 3,500, according to its method page (read 28 September 2026). Artificial Societies’ crisis simulation study for a consumer goods company built 1,174 personas across five societies and mapped each audience’s social influence network before testing 24 messages.

Questioning personas one at a time cannot show whether a message that persuades individuals still holds once they hear from the people they trust. Testing that needs a group with social structure. A paper by Artificial Societies researchers, He, Wallis, Gvirtz and Rathje in the British Journal of Psychology (opens in a new tab) (December 2024), found that 33,299 chatbots formed communities around a shared language, as people do. That shows a society of language-model agents can form human-like structure; it does not by itself show how persuasion changes inside it. What is a synthetic audience? and our article on networked audience simulation take the group question further.

Synthetic personas fit a question about opinion in an observable audience that is slow, impossible or unwise to ask directly. Artificial Societies’ hyperscaler earnings-call study is one such case: rules on selective disclosure meant unreleased numbers and draft scripts could never be shown to a real investor. For that study, Artificial Societies built more than 2,400 personas, each matched one-to-one to a real individual using public data and social activity, including 14 for the sell-side analysts who had asked questions on the company’s last eight calls. According to the case study, the live call’s questions ran the way the simulated analysts had predicted, which is one engagement reported qualitatively, not an accuracy score.

AAPOR’s 2026 task force report names a lower-stakes use: a survey pretest. Before fielding, a team prompts a language model, optionally under persona constraints, to complete the questionnaire repeatedly and find broken skip patterns, ambiguous wording and excessive completion time. That use needs plausible personas, not accurate ones.

Synthetic personas are the wrong tool when the decision needs proof of behaviour, when the test is physical, and when the result will be published as public opinion. The same AAPOR report says AI-generated cases are not research participants and must be identified as AI-created in any purported study of public opinion.