Synthetic Data vs Synthetic Respondents: What Differs

Enterprise · 11 min read

TL;DR: Two technologies are sold under the word synthetic, and they are not versions of each other. Synthetic data is a file generated to imitate the statistical structure of a real dataset it was fitted to, and its privacy risk is inherited from that source. Synthetic respondents are model-generated answers to questions nobody was asked, specified from published population totals rather than fitted to anyone's record. The privacy assurances do not carry across. Neither do the validation standards: one is graded against the file it copied, the other only against an independent human survey. One question separates them. Ask what the generator was fitted to, and keep the written answer with the study.

What is the difference between synthetic data and synthetic respondents?

Synthetic data is a generated file that imitates the statistical structure of a real dataset it was fitted to. Synthetic respondents are model-generated answers to questions nobody was asked, specified from published population totals. The first copies a source that exists. The second simulates a reply that never happened. The word is shared and very little else is.

The word collision is recent. Neither field is.

Synthetic data has a long literature in statistical disclosure control, where the problem is releasing something useful about a confidential file without releasing the file itself. The Royal Society's survey of the field sets out the standard construction: fit a generative model to real records, then sample new records from it.¹ The output is judged by how well it stands in for the original.

A synthetic respondent has no confidential file behind it. There is no set of real answers a generator was fitted to, because the questions were never asked of anyone. What stands in for a source is a specification: age, income, region, household composition, drawn so that the panel matches published totals.

That difference decides what follows. A synthetic dataset can always be checked against the thing it copied. A synthetic answer has nothing to be checked against except a separate study of real people.

Vendor pages use the two terms interchangeably, so a procurement call can run a long way before anyone notices the two sides mean different machines.

Does synthetic data remove privacy risk?

Not by construction. A file fitted to real records can still expose them, and the exposure rises as the copy gets closer to the original. Work presented at USENIX Security in 2022 argued that synthetic data does not deliver the privacy and utility trade-off claimed for it. Synthetic is a method, not a guarantee.

Stadler, Oprisanu and Troncoso published their result under the title Anonymisation Groundhog Day, and the argument is that synthetic generation repeats the failure pattern of earlier anonymisation techniques rather than ending it.² It is now the standard citation when a privacy reviewer pushes back on a synthetic file.

NIST takes the same position. Special Publication 800-188 treats de-identification as something an organisation evaluates and documents, not something it asserts by naming a technique.³

The UK Statistics Authority frames the problem as a trade-off rather than a binary. The closer a synthetic file sits to its source, the more useful it is and the greater the confidentiality risk, which is why national statistical offices release low-fidelity and high-fidelity products under different access rules.⁴

None of that reasoning reaches a panel specified from published totals, because there is no source file to expose. The point cuts both ways. A synthetic panel earns no credit for disclosure controls it never needed, and a synthetic dataset earns none for a privacy argument that only works when nothing was fitted.

Which assurances transfer between the two?

Almost none. Disclosure controls and fidelity scores belong to synthetic data and assume a source file exists. Replication evidence and method disclosure belong to synthetic respondents and assume there is nothing to expose. Carrying an assurance across the gap produces a document that answers a question nobody asked.

Put the table below in front of a privacy reviewer and a research lead at once. Each will recognise one column and not the other.

How do you tell which one a vendor is selling you?

One question does it. Ask what the generator was fitted to. There are three honest answers: your own file, a third-party dataset the vendor licensed, or nothing, because the panel is specified from published population totals. The first two put you inside a disclosure review. The third does not.

Ask it in writing, and ask before the pilot rather than after it.

If the answer is your own file, you have commissioned synthetic data. Customer records went into a generator, the output inherits their provenance, and your privacy team owns the result. That is a legitimate thing to buy. It is not the thing that removes the compliance trigger, and the separate question of whether the output represents your buyers is covered in what a census-grounded panel misses about your customers.

If the answer is a licensed third-party dataset, you inherit its collection and consent history, and you will usually not be shown it.

If the answer is published population totals, no individual record entered the system and there is nothing to re-identify. That is the argument set out in why synthetic personas remove the GDPR trigger, and it holds only for this third shape.

ESOMAR's published discussion topics for buyers of augmented synthetic data put provenance at the front for the same reason.⁵ The question is not whether a method sounds safe. It is what went into it.

PersonaHive is the third shape. Personas are grounded in national census data, country-specific, built from aggregated public statistics, and validated against real surveys. A panel is specified against those statistics rather than uploaded, so the provenance answer fits in one line of a method note.

Do the two share a validation standard?

No, and the reason is structural. Synthetic data is graded on fidelity to the file it was fitted to, which you can compute because that file exists. Synthetic respondents have no such file, so the only honest test is replication against an independent human survey the vendor did not fit to.

This asymmetry is where most vendor evidence falls apart.

A fidelity score is a self-comparison. It reports that the copy resembles the original, which is what the generator was optimised to do. Useful for synthetic data. Meaningless for a synthetic panel, because there is no original.

Replication is a harder claim. Pew Research Center tested whether AI can stand in for human survey takers in September 2026 and published the answer as not really.⁶ In Political Analysis, Bisbee and colleagues set out the perils of using large language models as synthetic replacements for human survey data.⁷ Both are arguments about prediction, not about resemblance.

A third number gets offered in place of either: consistency across repeated runs. That measures whether the instrument is stable, which is worth recording and is not the same as being right, as set out in the six fields that make a synthetic run reproducible.

So the question to a synthetic respondent vendor is not how accurate the personas are. It is which human survey the comparison ran against, whether that survey was held out, and whether you are allowed to read it. PersonaHive publishes its benchmark against the United States CFPB 2024 National Age-Friendly Banking Survey so that a reviewer can read the comparison instead of accepting a figure.

The smallest useful next step is one email. Ask your current or prospective vendor what the generator was fitted to, and keep the written answer with the study.

When is each one the right tool?

Synthetic data is right when you need to work with a dataset you are not allowed to release: testing software, developing a model, sharing a structure with a partner. Synthetic respondents are right when the question is about reaction to something that does not exist yet. Neither is right when the output must stand as a measurement.

Statistical agencies settled the first case years ago. The United States Census Bureau runs the SIPP Synthetic Beta, a synthetic version of linked survey and administrative records, paired with a validation server that re-runs an approved analysis against the confidential file before any result is released.⁸

That design is the honest version of synthetic data. The copy is for exploration and the real file settles the answer. Nobody publishes the copy as the finding.

The second case is a different job entirely. A concept, a price point and a claim set do not appear in any dataset, so no generator can be fitted to them. Asking a census-grounded panel which of nine claims reads strongest is a question a simulation can take. Asking it how many households bought the category last month is not, and the boundary is mapped in where synthetic research stops being valid.

The failure mode shared by both is the same one: treating a generated artefact as a measurement because it arrives formatted like one. A synthetic file looks like a dataset. A synthetic panel returns percentages. Neither counted anybody.

So send the email. Two minutes of writing tells you which review the work has to survive, and which of the two things on the invoice you are actually buying.

Frequently asked questions about synthetic data and synthetic respondents

Short answers to five questions that come up when the two terms meet in one procurement thread: whether a panel and a dataset are the same thing, how GDPR treats each, whether a fidelity score grades a panel, what to ask a vendor who says personas come from actual customer records, and which one a privacy impact assessment is for.

**Is a synthetic panel the same thing as a synthetic dataset?** No. A synthetic dataset is generated from a real file of individual records and is judged against that file. A synthetic panel produces answers to questions that were never asked of anyone, so there is no file to judge it against. The only external test is a human survey.

**Does synthetic data count as personal data under GDPR?** It depends on what it was fitted to and how closely. A file generated from real records can still permit inference about those records, which is why NIST treats de-identification as something to evaluate rather than declare.³ A panel specified from published population totals has no source records, so the question does not arise.

**Can I use a fidelity score to judge a synthetic panel?** No. Fidelity measures resemblance to a source file. A synthetic panel has no source file for its answers, so a fidelity figure quoted about one is either measuring something else or measuring nothing. Ask for a replication result against a named human survey instead.

**Our vendor says their personas are built from our own customer records. What should we ask?** Ask whether those records went into the generator. If they did, you have bought synthetic data, and your privacy team needs the provenance, the retention terms and the processor agreement before the pilot starts.

**Which one does a privacy impact assessment apply to?** A generator fitted to personal records needs one, and the assessment covers the source file and its disclosure risk.⁴ For a panel built from aggregated public statistics, the assessment covers what you upload as stimulus, not the panel itself.

Sources

• Synthetic Data: What, Why and How? — James Jordon and Lukasz Szpruch, The Royal Society
• Synthetic Data: Anonymisation Groundhog Day — Theresa Stadler, Bristena Oprisanu and Carmela Troncoso, 31st USENIX Security Symposium
• NIST Special Publication 800-188: De-Identifying Government Data Sets — National Institute of Standards and Technology
• Ethical Considerations Relating to the Creation and Use of Synthetic Data: Confidentiality, Privacy and Disclosure — UK Statistics Authority
• 5 Topics of Discussion to Help Buyers of Augmented Synthetic Data — ESOMAR
• Silicon Samples and Synthetic Surveys: Can AI Stand In for Human Respondents? — Pew Research Center
• Synthetic Replacements for Human Survey Data? The Perils of Large Language Models — Political Analysis, Cambridge University Press
• SIPP Synthetic Beta Data Product Request Guidelines — United States Census Bureau