Hey Juno

Real participants vs synthetic respondents

By Hey Juno · Published · Updated

Synthetic respondents are AI-generated personas that answer research questions in place of real customers. They can be useful for rehearsing a study or generating ideas to test, but the evidence so far does not support using them as the basis for decisions. Findings that will shape a product, brand or budget should come from real participants.

Juno, the AI researcher from Hey Juno, interviews real people. It also lets teams rehearse a study before launch, including a simulated interview, and treats that as rehearsal only. This article explains the difference, what the evidence says, and where each approach fits.

What synthetic respondents are

A synthetic respondent is a language model asked to behave like a person. You give it a profile, such as "a 34-year-old parent in Manchester who shops at discount supermarkets", and then ask it your survey or interview questions. The answers are generated from patterns in the text the model was trained on. No real person said them.

You will see several names for the same idea: synthetic respondents, AI personas, digital twins, and synthetic data in market research. Some products build a "twin" from a real person's earlier answers. Others build a persona from demographics alone. The two are not the same thing, and the difference matters for the evidence below.

What the evidence says

The research is young and mixed. Most studies test whether a model can reproduce answers that real people have already given. That is easier than predicting how people will respond to something new.

Averages can look close. Relationships often do not. Bisbee and colleagues (2024), writing in Political Analysis, asked ChatGPT to answer feeling-thermometer questions as personas built from real American National Election Study respondents. Average scores were similar to the human data. But when the authors ran the regressions that survey researchers normally run, the synthetic data often led to different conclusions. They also found that results were hard to reproduce.

Prices and preferences are sometimes right and sometimes wrong. Brand, Israeli and Ngwe (2023) tested a language model on market research tasks such as willingness to pay. Their working paper, now titled "Using LLMs for Market Research", reports estimates that were sometimes comparable to human studies, but often inaccurate and in some cases had the wrong sign.

Groups get flattened. Wang, Morgenstern and Dickerson (2025) ran human studies with 3,200 participants across 16 demographic groups. Their paper, accepted at Nature Machine Intelligence, found that language models misportray identity groups and flatten the variety within them. They suggest a narrower role for models in tasks such as pilot studies, where the goal is to supplement people, not replace them.

Richer inputs help. Park and colleagues (2024) at Stanford built agents from two-hour interviews with 1,052 real people. Agents built from interviews and survey data matched 72.8 per cent of participants' General Social Survey answers, against 79.5 per cent for the participants' own consistency two weeks later. The agents also showed less bias across ethnic, gender and political groups than agents built from demographic descriptions. The study tests recall of answers the real people gave. It shows that real interviews carry a great deal of information. It does not show that a model can stand in for a new interview with a new person.

People commissioning research should ask questions. ESOMAR, the global body for market research, has published 20 questions for people buying AI-based research services. Asking how a vendor's output was produced and how it was checked is a sensible start.

Our reading of these studies is that synthetic respondents tend to do better on broad, familiar, well-documented topics, and worse on detail, subgroups and relationships between variables. For the full case, with newer evidence on digital twins, see why synthetic participants and digital twins can't replace real people.

Where synthetic respondents can help

Used openly, as a thinking aid, they have honest uses:

  • Rehearsing a study. Running a mock interview shows you where a question is confusing, leading or too long before anyone real is asked.

  • Stress-testing a question guide. A simulated respondent can reveal gaps in coverage or awkward ordering.

  • Generating hypotheses. A model can suggest reasons people might hold a view. Treat each one as a guess to check, not a finding.

  • Practice for the team. Reading a sample transcript helps a team agree what a good answer would look like.

In every case, the output is about the study, not about the market.

Where they mislead

  • Novel products and ideas. A model draws on what already exists in its training text. A concept with no history gives it little to work from, so it fills the gap with plausible language.

  • Minority and unusual views. Models drift towards the common answer. The people whose experience differs from the average are often the reason you are doing the research.

  • Emotional and contextual detail. A real person mentions the argument they had in the shop, or the thing they were embarrassed to admit. A generated persona produces tidy, agreeable reasons.

  • Anything the model has not seen. Local, recent or private situations are the weakest ground.

  • Confidence. Generated answers read fluently whether or not there is anything behind them. That is the same trap as any fluent AI summary, and it is why it pays to check AI research findings against the quotes behind them.

Why decisions should rest on real participants

Juno does not blend in generated answers. Every finding comes from a real participant, and a generated answer is not data about anyone.

A research finding is a claim about people. The only way to know what people think, feel and do is to ask them. A synthetic answer is a prediction of what people might say, based on what others said before.

A real participant answers from their own experience, and a good interviewer can follow it. A persona gives a fluent one, and fluency is not evidence. The honest status of a synthetic answer is provisional: a hypothesis to take to real people, not a conclusion.

Real participants can also surprise you. In qualitative research, the most valuable moment is often the unexpected answer that changes the question. A model built from past text is least able to produce that.

There is also a practical reason. A finding traced to real quotes and real transcripts can be defended in front of a client or leadership team. A finding that came from a persona cannot be traced to anyone.

How Hey Juno handles this

Juno runs AI-moderated qualitative interviews with real people, by voice or text. Findings link back to verbatim quotes and transcripts.

Before a study goes live, a team can rehearse it. They can preview the interview themselves, or run a simulated interview in Test mode, to catch weak wording, screening problems, pacing and coverage gaps. Hey Juno states that simulated interviews are used only to rehearse a study before launch, never as research data. Every finding comes from real participants.

Limits of this article

The evidence is changing quickly. Most of the studies above used earlier language models, and newer ones may do better on some tasks. Some are preprints or working papers that have not been through peer review. Many tests compare a model with answers people already gave, so they say little about new products or new audiences. This article is not a test of any particular vendor's synthetic product, and a well-validated tool used within a narrow, known range may perform better than the general results suggest.

Frequently asked questions

What is a synthetic respondent?

A synthetic respondent is an AI-generated persona that answers survey or interview questions in place of a real person. The answers come from a language model, not from anyone's actual experience.

Are digital twins the same as synthetic respondents?

Not quite. A digital twin is usually built from a specific person's earlier answers, while a persona is often built from demographics alone. Park and colleagues found that agents built from long interviews tracked their real counterparts better than demographic-only agents. Both are still predictions, not new answers from the person.

Can I use synthetic respondents for concept testing?

They can help you draft and rehearse a concept test. They should not decide whether the concept works, because a new idea gives a model little real evidence to draw on. Test the concept with real participants.

Does Juno use synthetic respondents for findings?

No. Juno interviews real people, and findings come only from them. Simulated interviews are used to rehearse a study before launch and are never treated as research data.

How do I rehearse a study before it goes live?

In Juno you can preview the interview yourself or run a simulated interview in Test mode. Read the sample transcript to spot unclear questions, screening problems and gaps, then fix them before real participants take part.