A new paper argues that “synthetic data” is not one thing in law or risk. It proposes a taxonomy—transformed, augmented, and simulated—that could shape how teams document privacy, bias, and downstream accountability.
Taxonomizing Synthetic Data for Law
Researchers on arXiv propose a legal-and-technical classification of synthetic data into three buckets: transformed, augmented, and simulated. The core claim is practical: the label “synthetic” is too coarse to support compliance decisions, because different generation approaches create different privacy and discrimination risks—and may trigger different legal treatments.
The paper’s key finding is a direct pushback on a common assumption in product and governance conversations: synthetic data does not inherently prevent privacy harms. Depending on how it is produced and used, synthetic data can still enable privacy-relevant inferences, and it can also amplify bias when the generator’s assumptions or source data encode skewed patterns. The authors frame the taxonomy as a way to connect technical choices (how data is generated and what it preserves) to legal questions about disclosure, risk assessment, and responsibility for downstream impacts.
- Risk assessments need to be method-specific. Treating all synthetic data as “de-identified” or “privacy-preserving” is a governance shortcut; the taxonomy suggests documenting the generation method and what attributes/relationships are retained to ground privacy and re-identification analysis.
- Bias controls don’t disappear with synthesis. If synthetic data can amplify bias through flawed assumptions, teams may need the same (or stronger) representativeness checks, subgroup evaluation, and monitoring they apply to real datasets—plus validation of the generator itself.
- Disclosure and audit artifacts may become table stakes. A standardized way to describe “transformed vs. augmented vs. simulated” could feed into dataset documentation, procurement questionnaires, DPIAs/PIAs, and model cards—especially where regulators or customers ask what “synthetic” actually means.
- Procurement and contracts should pin down definitions. Vendors often market “synthetic” as a blanket risk reducer; this framework gives buyers a vocabulary to require specifics (inputs, constraints, intended use, and known limitations) rather than accepting the label.
