Define the intended use
Synthetic data is created to resemble selected properties of a source dataset. It can support software testing, demonstrations, or exploratory analysis when access to the original records is restricted. It is not automatically a substitute for the original data in every task.
Start by describing the intended use precisely. Testing whether a form accepts a valid record requires different properties from studying the relationship between clinical measurements. The purpose determines what the synthetic data must preserve.
Check more than individual columns
Matching the average of each column is not enough. Relationships between variables, rare categories, missingness, and unusual combinations may matter to a downstream analysis. Evaluate the synthetic dataset using the same types of checks and tasks that its users intend to perform.
Keep a record of where results diverge. If the synthetic data performs poorly for a subgroup or an edge case, document that limitation rather than presenting one overall score as proof of quality.
Assess disclosure risk separately
A dataset can be useful and still reveal too much about its source. Check for exact or near-duplicate source records, unusually distinctive combinations, and scenarios where an attacker might combine the output with other information. Privacy evaluation should reflect the sensitivity and intended release audience.
The generation method matters, but its name alone does not establish protection. Where formal privacy guarantees are claimed, the implementation, assumptions, and parameters need review. Utility and privacy are separate requirements and should be documented separately.
Release with instructions and boundaries
Provide a short data card describing the source context, generation approach, evaluation, permitted uses, and known limitations. Make clear that generated records do not represent actual individuals. Preserve the version information needed to reproduce an evaluation.
Use staged access where appropriate. A controlled internal trial can reveal problems before wider sharing. Reassess the release when the source data, intended use, or generation method changes; an earlier evaluation does not automatically cover a new version.