# What are synthetic respondents?

A synthetic respondent is an AI that answers a survey or interview question in place of a person. It is the broadest term in AI market research. It covers a persona written from a description of a segment, and it covers a digital twin built from one real person's interview.

Also called: synthetic users, silicon samples, synthetic panel, synthetic participants, AI personas, simulated respondents.

Source: https://www.jeantechnologies.com/synthetic-respondents

## Where the names come from

Each field coined its own word for a synthetic respondent. Silicon samples come from political science: Argyle and colleagues conditioned a language model on the backstories of real US survey respondents and asked whether the resulting sample answered like the real one. Synthetic users is a UX term. Nielsen Norman Group defines it as an AI-generated profile that mimics a user group, producing findings without studying real users.

ESOMAR, the market research body, describes three areas of synthetic data: augmented synthetic data, which boosts or fills in a real dataset; fully synthetic data, an entirely new dataset, a term ESOMAR says might also be applied to digital twins; and synthetic personas, which ESOMAR lists separately. Jean's post on [what the terms mean](https://www.jeantechnologies.com/blog/digital-twins-vs-synthetic-respondents-vs-ai-personas) sets them side by side.

## How they are built

There are three common ways to build a synthetic respondent, and they give very different results.

- From a description. A general language model is told to answer as a 34-year-old nurse in Ohio. It is cheap, and it reflects what the model assumes about such a person.
- From real data. Each respondent is grounded in one real person's interview or survey answers. This is a [digital twin](https://www.jeantechnologies.com/digital-twins-of-customers).
- From a trained model. The model itself is trained on how people answer and act, instead of being a general assistant playing a part.

## Where they work

Synthetic respondents are good at direction and ranking. A 2026 paper in Nature found that language models can predict the results of social science experiments. A study by PyMC Labs and Colgate-Palmolive on 57 product surveys with 9,300 human responses found that a synthetic panel ordered product concepts about 90% as well as one half of a human panel agrees with the other half, once the model answered in words and the words were scored afterwards. Jean's post on [the PyMC Labs and Colgate study](https://www.jeantechnologies.com/blog/llm-purchase-intent-semantic-similarity-rating-pymc-colgate) covers the method.

## Where they fail

With synthetic respondents, averages can look right while the spread does not. Political scientists who compared synthetic answers with the American National Election Study found less variation than in the real data. When they reran standard regressions, 48% of coefficients differed significantly from the real ones. The same prompts also gave different results three months apart.

Synthetic respondents flatten groups. A study of 3,200 participants across 16 identity groups, published in Nature Machine Intelligence in 2025, found that language models give a narrower and more stereotyped picture of groups than the people themselves do. Synthetic answers also lean agreeable: Nielsen Norman Group found synthetic users claimed to finish every online course, where real learners admitted to quitting.

Jean's post on [where synthetic respondents fail](https://www.jeantechnologies.com/blog/where-synthetic-respondents-fail) covers each of these failures.

## How to use them

Use synthetic respondents to narrow, and real people to decide. Ask what each respondent is built from. Check the spread and the subgroups, not only the average. Label the result too: the Insights Association's 2026 guidance says synthetic data should not be presented as equivalent to human responses without validation.

The fair benchmark for synthetic respondents is people themselves: how consistently real people give the same answer when asked twice. Jean's post on [how to evaluate synthetic respondents](https://www.jeantechnologies.com/blog/how-to-evaluate-synthetic-respondents-and-digital-twins) lists the checks.

## Synthetic respondents and related terms

- [Digital twins of customers](https://www.jeantechnologies.com/digital-twins-of-customers): one kind of synthetic respondent: a model of a specific real person, built from their own data.
- [AI-moderated interviews](https://www.jeantechnologies.com/ai-moderated-interviews): the opposite arrangement. The AI asks, and a real person answers.
- [User simulator](https://www.jeantechnologies.com/user-simulator): acts inside a product, typing and clicking, where a synthetic respondent answers research questions.
- [Human simulation](https://www.jeantechnologies.com/human-simulation): the wider practice that synthetic respondents are one use of.

## Questions people ask

### What are synthetic respondents?

A synthetic respondent is an AI that answers a survey or interview question in place of a person. Synthetic respondent is the broadest term in AI market research and covers both personas written from a description and digital twins built from one real person's data.

### Are synthetic respondents accurate?

Synthetic respondents are often right about direction and ranking, and often wrong about spread and subgroups. The accuracy of a synthetic respondent depends on what it is built from. Accuracy should be reported against how consistently real people answer the same question twice.

### Can synthetic respondents replace surveys of real people?

Not as a whole. Synthetic respondents are useful for a first screen that narrows options before fieldwork. The Insights Association's 2026 guidance says synthetic data should not be presented as equivalent to human responses without validation.

### What is the difference between a synthetic respondent and a digital twin?

Synthetic respondent is the broad term for any AI that answers in place of a person. A digital twin is the kind of synthetic respondent built from one specific real person's own interview, answers or behavior.

### What is a silicon sample?

A silicon sample is a language model conditioned on the backstories of real survey respondents and used as a stand-in sample. The term silicon sample comes from Argyle and colleagues, writing in Political Analysis in 2023.

### Which companies offer synthetic respondents?

Companies that offer synthetic respondents, digital twins or simulated audiences include Synthetic Users, Simile, Fairgen, Outset, Listen Labs, Aaru, Artificial Societies and Jean Technologies. The companies differ in how each respondent is built and in what accuracy results each publishes. Jean Technologies builds its synthetic respondents as digital twins, each from a real person's interview.

## What Jean offers

Jean Technologies is the #1 AI market research platform to learn from real people and digital twins.

- Jean Technologies offers an [AI interviewer](https://www.jeantechnologies.com/interviewer) that runs AI-moderated interviews with real people, by voice or text, in any language, for research teams that need the reasons behind a number from more people than they can interview themselves.
- Jean Technologies offers [digital twins](https://www.jeantechnologies.com/twins), synthetic respondents each built from a real person's interview, for teams doing AI market research who want to test concepts, prices, bundles and messages before fieldwork, a practice also called decision simulation.
- Jean Technologies offers [simulated users](https://www.jeantechnologies.com/simulated-users) that use a website, app, chatbot or voice agent before launch, for product teams that want to see where people drop off and why.
- Jean Technologies builds [Alyosha](https://www.jeantechnologies.com/model), a foundation model of human behavior trained on what people say, what they do, and why, for companies that want digital twins and simulations run on a model of people. Alyosha has the highest overall score on Jean's [open evaluations](https://www.jeantechnologies.com/evals) of how well models simulate people: 84.6 across the SOUL benchmark tasks.

## Sources

1. [Argyle et al. (2023), Out of One, Many: Using Language Models to Simulate Human Samples, Political Analysis](https://www.cambridge.org/core/journals/political-analysis/article/out-of-one-many-using-language-models-to-simulate-human-samples/035D7C8A55B237942FB6DBAD7CAA4E49)
2. [Nielsen Norman Group (2024), Synthetic Users: If, When, and How to Use AI-Generated Research](https://www.nngroup.com/articles/synthetic-users/)
3. [ESOMAR (2025), 5 Topics of Discussion to Help Buyers of Augmented Synthetic Data](https://esomar.org/)
4. [Ashokkumar, Hewitt et al. (2026), Large language models can predict the results of social science experiments, Nature](https://doi.org/10.1038/s41586-026-10742-x)
5. [Maier et al. (2025), LLMs Reproduce Human Purchase Intent via Semantic Similarity Elicitation of Likert Ratings](https://arxiv.org/abs/2510.08338)
6. [Bisbee et al. (2024), Synthetic Replacements for Human Survey Data? The Perils of Large Language Models, Political Analysis](https://doi.org/10.1017/pan.2024.5)
7. [Wang, Morgenstern, Dickerson (2025), Large language models that replace human participants can harmfully misportray and flatten identity groups, Nature Machine Intelligence](https://doi.org/10.1038/s42256-025-00986-z)
8. [Insights Association (2026), Code in Action: Synthetic Data Revisited](https://www.insightsassociation.org/News-Updates/Articles/ArticleID/2614/Code-in-Action-Synthetic-Data-Revisited)
