How does synthetic data preserve privacy, and what is its key caveat?

Question

Accepted Answer

Synthetic data is model-generated and contains no real individuals — immune to traditional linkage — but it must be validated to ensure it didn't memorize real patterns.
Synthetic data generation creates artificial datasets that preserve the statistical properties (correlations, distributions) of real data while eliminating links to real individuals. Three points:

Statistical fidelity — maintains correlations and distributions, so analysis stays meaningful.
No real records — because no actual individual exists in the data, it's immune to traditional linkage attacks.
Validation required — you must verify the synthetic data didn't accidentally memorize real patterns. Generative models can reproduce or leak real outliers (membership-inference / memorization risk).

So synthetic data is powerful but not automatically private — careful validation is mandatory, echoing the lesson that no method is risk-free by default.
Tip: "No real people in it" defeats linkage — but a model that overfit can still spit out a real outlier. Always test synthetic data for memorization before release.