What is the difference between synthetic data and de-identified data?

How synthetic data differs from de-identified data

Synthetic data and de-identified data solve different problems. Synthetic data is created to behave like the source data statistically, while de-identified data is still real data with direct identifiers removed or transformed. That means synthetic data can be useful when you want realism without exposing records, whereas de-identified data still carries residual privacy and linkage risk.

What each one preserves, and what it loses

The practical difference is what remains grounded in the original dataset. Synthetic data aims to preserve patterns, distributions, and relationships, but it should not be a copy of actual people’s records. De-identified data preserves the original records and their analytical value, but tries to reduce identifiability by removing names, IDs, or other obvious identifiers. In practice, de-identification often reduces direct exposure rather than eliminating it.

That difference matters because the two approaches fail in different ways. Synthetic data can distort rare cases, edge conditions, or correlations that matter to downstream analysis. De-identified data can still leak information through quasi-identifiers, linkage with outside datasets, or inference about individuals. For some use cases, the question is not which is “safer” in the abstract, but which preserves the minimum data characteristics needed for the task.

When each approach is the better fit

Synthetic data is usually the better fit for development, testing, experimentation, sharing, or training workflows where the actual records are not required. It is especially useful when the goal is to reduce exposure while keeping the shape of the data useful. De-identified data is better when you need continuity with real records for analytics, reporting, or operational processes and can tolerate the residual privacy controls and governance burden that come with real source data.

For teams handling sensitive datasets, a synthetic copy can reduce the need to open access to production data, which can simplify sharing across engineering and analytics groups. De-identified data may still require strong access controls, purpose limitation, retention rules, and review of re-identification risk. The difference is not only technical, it changes who can safely use the dataset and under what oversight.

For practitioners wanting a deeper view of the privacy trade-offs, Privacy-Enhancing Technologies Guide explains where synthetic data sits alongside masking, tokenisation, and other privacy techniques.

When the central question is re-identification risk, the Synthetic Data Risk Guide is the most direct reference for understanding how synthetic releases can still leak information or fail validation.

Risk and Threat Considerations

The main risk difference is that de-identified data can remain linkable, while synthetic data can be unsafe if it reproduces real records too closely or overfits rare individuals. In both cases, the label alone is not enough, because the privacy outcome depends on the generation method, the uniqueness of the data, and what other data an attacker or recipient can combine it with.

Failure mechanism: De-identification may remove obvious identifiers but leave enough quasi-identifiers, unusual combinations, or retained structure for linkage attacks, inference, or re-identification. Synthetic data may accidentally memorize source records, preserve outliers too faithfully, or leak sensitive patterns if it is not tested before release.

Impact: Organisations can expose personal data, violate internal privacy commitments, or create a false sense of safety that leads to broader sharing than the dataset actually supports. That can affect legal, contractual, and trust obligations, and it can also undermine model quality or analytical correctness if the chosen technique is too lossy.

Practitioner Guidance

What to verify: Confirm whether the use case needs actual record continuity or only statistical similarity. If you need row-level continuity, synthetic data may be the wrong tool; if you only need patterns, de-identified real data may be unnecessary exposure.

Decision rule: Treat de-identified data as lowered-risk real data, not as automatically anonymous data. Treat synthetic data as safe only after testing for memorisation, linkage risk, and utility against the specific downstream use case.

Practitioner takeaway: The right choice is driven by the fidelity you need and the exposure you can tolerate, because synthetic data reduces linkage to real records while de-identified data reduces direct identifiers but may still remain personal data in practice.

Related resources from NHI Mgmt Group

• Why do biological AI systems create risk even when the underlying data has been screened or de-identified?
• How does de-identified data work?
• Why does de-identified data matter to security teams?
• What are common mistakes with de-identified data?