China Synthetic Data Demos: Test What Was Preserved
Synthetic data can preserve some patterns while losing others. Evaluate the intended downstream task and the privacy claim separately during a company visit.
Direct answer. Ask what the synthetic dataset is intended to preserve, then test that property on the proposed downstream task. Visually plausible records do not establish that important relationships survived or that privacy risk was eliminated.
Name the intended use
At a China synthetic-data visit, distinguish software demonstration data from model-training data and privacy-protective sharing. These uses require different evaluation. Use approved samples; do not supply personal data merely because the presenter promises to synthesize it.
Select a relationship important to the task, such as an uncommon combination of order conditions. Ask whether the synthetic dataset retains that relationship and whether the evaluation used a separate permitted reference dataset. A good match on average values can coexist with poor behavior on rare cases.
Keep the claims apart
| Claim | Evidence question |
|---|---|
| Similar appearance | Which distributions or relationships were compared? |
| Task utility | What happens on the actual downstream evaluation? |
| Rare cases | Which infrequent patterns are retained or lost? |
| Privacy | What attack or disclosure risks were assessed? |
Avoid accepting 'not real people' as a complete privacy argument. Ask a qualified privacy reviewer to assess the method and residual risks. The correct interpretation depends on the data, generation process and intended disclosure.
The integrity and privacy questions are grounded in NIST AI RMF Core. The framework is voluntary and does not certify a synthetic dataset as private. Our table is a comparison aid, not a privacy-testing protocol.
Record what was actually preserved
A useful note might state that selected relationships were compared and that downstream performance or privacy testing remains unverified. Do not replace that distinction with a general 'high fidelity' label.
Refer to the deployment evidence guide, then plan a data-focused visit with both the intended task owner and the person responsible for approving data use.