FlashInsight

Glossary

Synthetic respondent

Also called simulated respondent · AI respondent

An instance of a model asked to answer questions as a human would, calibrated on data describing a real population. The output — the unit a sample is made of.

A synthetic respondent is an instance of a model put through a questionnaire or a discussion guide, answering from a given profile. It sits at the end of the chain: synthetic data is what goes in, the respondent is what speaks.

What it reproduces are population-level regularities — preferences, gaps between segments, the direction of an attitude — not an individual truth. The closest honest analogy is a flight simulator. Nobody boards one believing they are flying, and nobody faults it for staying on the ground. It reproduces certain laws faithfully enough to rehearse before the real thing.

Open the animation full screen

What it is not

It is not a fake respondent, and the “real or fake” framing is the wrong question. The useful question is whether the instrument is reliable once validated — not whether it is lifelike. Nor is it a replacement for fieldwork. Synthetic proposes; the field decides.

What a buyer can check

Which data the calibration rests on, market by market. How far the answers spread — a badly built panel agrees with itself far too much, and that is measurable. And whether the vendor answers all twenty of ESOMAR’s questions or only the flattering ones.

Frequently asked questions

What is a synthetic respondent?

It is an instance of a language model put through a questionnaire or a discussion guide, answering from a given profile anchored in data on a real population. It is not a copy of a person. What it reproduces are population-level regularities, such as preferences, gaps between segments and the direction of an attitude, not an individual truth.

Are synthetic respondents reliable?

On averages, often yes once calibrated. On spread, often not: before correction, our simulated respondents showed 0.20 to 0.76 of human spread across eight markets; after correction, 0.96 to 1.07 on attitudes. A vendor who quotes only an accuracy percentage has not answered the question.

Can a synthetic respondent replace a real one?

No. It is a pre-test layer between intuition and fieldwork: good for ranking and screening concepts, messages and stimuli early, so that fewer and better ideas reach the field. It is not a substitute where the finding is a minority view, a genuinely new category, or a decision that will not be checked against real people afterwards.

How do you test a synthetic respondent vendor?

Ask three things. Which public survey each market is calibrated on, named and dated. Whether the questions used to test the system were seen during calibration; if so, the figure describes a fit, not an ability to predict. And the dispersion ratio, market by market, before and after correction.

See also

  • Synthetic data — Artificially generated data that mimics the statistical properties of real data without containing anything personally identifiable. The raw material — not the respondent.
  • Synthetic persona — The briefing given to the model: the traits, attitudes and characteristics that steer its answers. A specification, not a person.
  • Under-dispersion — The commonest defect in simulated panels: answers cluster more tightly than real humans’. Disagreement disappears, and the signal goes with it.

Further reading

Back to the glossary