What Is Synthetic Data?
Synthetic data is information that is generated rather than measured from real people or events. A well-made synthetic dataset reproduces the statistical shape of a real population — its proportions, ranges, and the way attributes move together — without containing a single real individual.
Because no real record is ever copied, synthetic data sidesteps the privacy obligations that attach to production data. It can be shared, published, and loaded into test systems where actual personal data would be off-limits. The trade-off is fidelity: synthetic data is only as useful as the statistics it was built to match.
A working example of synthetic data, no signup required.
Two broad methods dominate. Distribution-calibrated generation starts from published reference statistics and draws each value to match them, enforcing cross-field rules along the way. Machine-generative methods — GANs, variational autoencoders, diffusion models, and large language models — instead learn patterns from a real source dataset and sample new records that resemble it.
SimpleIDGen uses the calibrated approach. Each attribute is fitted to public US references — NHANES 2017–2020 for body measures and lab values, the American Community Survey (ACS) 2022 for demographics and geography, and CDC surveillance for disease prevalence — and invariants are enforced so records stay internally consistent: BMI equals weight over height squared, insulin appears only for diagnosed diabetics, and a ZIP code matches its state. Generation is deterministic by seed, so the same seed always yields the same people.
| Approach | How it works | Trade-off |
|---|---|---|
| Distribution-calibrated | Draws each value to match published reference distributions, with constraints enforced | Transparent and auditable; captures the relationships you model, not ones you do not |
| Machine-generative | A model learns patterns from a real source dataset, then samples new records | Can capture subtle correlations; needs real training data and may risk memorizing it |
For more on the calibrated method, see NHANES-calibrated synthetic data →
Three uses recur. Privacy-safe testing: filling staging and QA environments with realistic records when production data cannot leave its boundary. Machine learning: augmenting or standing in for scarce or sensitive training data, and rehearsing pipelines before real data is available. Demos and documentation: populating product tours, screenshots, and tutorials with believable but fake people.
In each case the value is the same — data that behaves like the real thing without the liability of holding it. A calibrated synthetic person dataset carries 65 attributes per record across nine domains, from identity and geography to finances and health, so a test or model sees the kind of structure it would meet in production.
Synthetic data is not anonymized data. Anonymization strips identifiers from real records, which still carry re-identification risk; synthetic data is generated from distributions and references no real person to begin with. We cover that distinction in full in synthetic vs. anonymized data →
Synthetic data is also a model, and every model omits something. Calibrated generation preserves the marginal distributions it targets and the joint relationships it explicitly encodes — age against A1c, sex against height, income against geography — but it will not reproduce correlations nobody modelled. Rare edge cases and long-tail interactions present in real records can be smoothed away. The honest rule: use synthetic data where realistic structure matters, and validate against real data before drawing a conclusion that depends on a specific real-world correlation.
Not quite. Dummy data is usually random noise that merely fits a field type. Synthetic data is generated to match real statistics, so it behaves like a real population. Calibrated synthetic data sits at the realistic end of that range.
When it is built from public reference distributions and contains no real individual, calibrated synthetic data carries no personal data and falls outside those obligations. Confirm the specifics for your jurisdiction and the method you use, since ML-generative data trained on real records can behave differently.
For testing, demos, and many machine-learning tasks, yes. For a conclusion that hinges on one specific correlation, validate against real data — generation only preserves the relationships it was built to model.
Download the 1,000-row sample above with no account, or generate your own — a free account produces up to 5,000 rows per UTC day in CSV or JSONL.