Synthetic respondents explained: digital twins, audiences and panels

Artificial Societies uses networks of AI personas to simulate high-value audiences. Our method and evaluation place individual nuance alongside population accuracy: you can inspect a persona’s reasoning and examine how opinions differ between segments. We build that combination by grounding personas in real individuals and connecting them through relationships and patterns of influence. The audience needs diverse observations, whether public or client-provided, to support that work.

The evidence for the method is Artificial Societies’ 86% distribution accuracy against a 91% human ceiling across 1,000 surveys, explained in full on the accuracy page. It tells you how closely the spread of answers matches the human benchmark; it does not establish that a simulated person will act as a real person would.

We would start a consequential research decision with that combination of individual reasoning and population evidence. An average can answer a market-level question. If your recommendation depends on which stakeholders believe what, you also need to examine the relationships between their answers. Our published method makes those relationships part of the research task.

Synthetic respondents are simulated participants that produce research answers. The Market Research Society’s Using synthetic participants for market research (opens in a new tab), published in 2024, describes large language models (LLMs) using artificial profiles to imitate human survey participants. A language model generates the response from the profile and the question; the resulting answer is simulated evidence.

The ICC/ESOMAR International Code, 2025 revision (opens in a new tab), defines a synthetic persona as a digital representation designed to mimic the behaviour, preferences and characteristics of real people or groups. The Code separately defines a person as human. That distinction matters when a report describes its participants: a synthetic respondent has produced an answer, but a human respondent has not been interviewed for that answer.

NIQ’s explanation, updated 19 February 2026 (opens in a new tab), likewise defines synthetic respondents as artificial personas generated by machine learning to mimic human research responses. Neither the role nor the fluent answer tells you how the persona was built. In our terminology, a digital twin, a synthetic population or a network of personas can fulfil the respondent role. Ask for the construction method alongside the questionnaire.

Gartner’s concept of a digital twin of a customer, abbreviated DToC, is a virtual representation used to simulate and anticipate that customer’s behaviour. Its reported commercial uses include personalisation and customer experience, as Scott Clark explains in CMSWire, 18 October 2022 (opens in a new tab). The unit of interest is the customer whose response you want to model.

Park and colleagues’ paper, in its 28 June 2026 version (opens in a new tab), describes individual validation using agents built for 1,052 Americans from interviews, surveys or both. They scored the agents on survey items withheld from their construction, against the participants’ own consistency when answering again after two weeks. The comparison asks whether an agent reproduces its particular person’s answers.

Combined interview-and-survey agents achieved 86% of participants’ own test–retest benchmark, compared with 74% for demographics-only agents, in Park and colleagues’ June 2026 paper. Those figures use a different denominator from our distribution-accuracy result. For work centred on an individual, we would examine this approach first; for a market estimate, we would ask for population-level evidence before treating a collection of twins as representative.

We use synthetic audiences to mean simulated groups whose reactions you can study before making a decision. Our September 2026 validity framework states the boundary explicitly: “Synthetic audiences will never be a true substitute for the people they seek to represent.” In our construction, Persona Genesis models individual belief systems from observations of anonymised real individuals, Society Assembly connects the personas through social relationships, and the Simulation Engine exposes the society to a scenario; Artificial Societies’ published method specifies societies of 12 to 3,500 personas. What is a synthetic audience? covers the term in full.

Synthetic panels provide generated responses in a panel-research setting. Qualtrics describes their output as “AI-generated responses that mimic real human feedback” in Ali Henriques’ article of 5 June 2025 (opens in a new tab). The word panel identifies the research format. You still need to establish which model generates its responses and which human data support them.

Qualtrics distinguishes model types in that article. It says an LLM wrapper is unsuitable for quantitative survey responses or niche topics absent from public discussion. It describes machine learning models as unsuitable for qualitative feedback or situations outside their training data. These are Qualtrics’ stated limits for those model types, not a verdict on a named supplier’s product.

Qualtrics’ 15 November 2025 account of synthetic responses (opens in a new tab) places them in idea testing, new-market exploration and privacy-sensitive work. It directs researchers dealing with missing data or population matching in an existing dataset towards traditional techniques. That distinction separates generating a simulated research answer from repairing a dataset you already hold. Our synthetic market research guide covers the wider category those tasks sit within.

Our evaluation page separates digital twins, synthetic populations and networks of enriched personas by individual detail and population coverage. The table uses that framework, with a boundary for each approach. It compares types of research model, rather than assigning suppliers to categories from their names.

Source: our method and evaluation for the three approaches; our validity framework for the boundaries; the physical and sensory exclusion is our published position.

Gartner’s customer twin fits the individual approach. Synthetic respondents can play a role in any of the three, while synthetic panel language describes a research output rather than settling how it was made. For synthetic audiences, our implementation is the network of personas. This mapping is our interpretation of the definitions, not an industry classification awarded to products.

The synthetic-population approach describes a top-down method: reproduce headline averages and generate the individuals beneath them. Digital twins proceed from individual detail. We connect grounded individuals into a society so that you can examine both their reasoning and the population’s responses. A crosstab, the breakdown of answers by another characteristic or answer, is where that distinction becomes practical: matching the total alone does not establish the subgroup result.

Synthetic research needs a human comparison suited to the claim. For a synthetic population, distributional validation compares the simulated spread of answers with human answers withheld from model construction. Our September 2026 validity framework explains why matching each question’s totals still leaves the relationships between answers untested. A model can reproduce a survey’s separate totals while assigning the combinations of answers to the wrong people.

Artificial Societies commits to eight tests in its September 2026 validity framework. We examine whether a persona’s answers follow a stable belief system, whether that structure matches human data and whether it changes under a stimulus as people’s answers do. The purpose is to interrogate the simulated belief system as well as its output. A convincing explanation attached to an answer needs that scrutiny too.

A peer-reviewed paper on group behaviour co-authored by Artificial Societies’ James He and Felix Wallis (He, Wallis, Gvirtz and Rathje, British Journal of Psychology (opens in a new tab), December 2024) supports the network modelling, not the survey accuracy figures; the science page explains what it found.

Synthetic audiences fit our work on opinion shifts around consequential decisions. We establish an audience’s baseline, expose it to a message or scenario, and study how opinion changes through the group. Strategic communications, investor communications and public affairs are uses where that sequence matters. Confidential material is another reason to simulate: our earnings-call case study describes testing unreleased numbers and draft scripts that could never be shown to a real investor.

Teneo’s own account of the engagement is in our Teneo case study. For your research, the relevant starting point is whether you need to test a sensitive decision and examine how reactions spread through the audience.

Early exploration can have a different scope. The MRS Delphi Group’s 2024 report describes prompted customer personas as useful for concept testing and hypothesis generation. We would use that work to decide which options deserve investigation, then apply networks of enriched personas when the strategic recommendation depends on individual beliefs and their relationships. Where the audience lacks diverse observations, we would commission the human research needed to ground the simulation before modelling its response.