Why is synthetic data NOT automatically anonymous or safe to share?

Because a generative model can reproduce real patterns — rare attribute combinations in the synthetic data may correspond directly to unique real individuals.

A common misconception is that synthetic data is automatically safe to share freely. But artificial data can still pose privacy risks depending on how it was generated and the quality of the model. The synthetic dataset may retain patterns that identify people, especially if it preserves strong correlations between variables — and rare combinations of attributes in the synthetic data might correspond directly to unique individuals in the original.

What can leak: descriptive statistics (means, medians, std-devs), variable relationships (correlations), plausible value ranges (realistic min/max), and distribution patterns (overall shape). Whether synthetic data counts as "personal data" under the GDPR must be assessed case-by-case — there's no one-size-fits-all answer.

Tip: "No real people in the table" defeats naive linkage, but a model that memorized an outlier can still regurgitate them. Always validate synthetic data for leakage before release.