What synthetic data genuinely solves
Generating realistic test data without exposing real customer records, producing enough examples of rare edge cases to test against, and augmenting training sets where real examples are genuinely scarce are legitimate, well-understood uses backed by real research.
In all three cases, the underlying production pipeline's data quality is a separate concern, untouched by the synthetic data exercise.
What it doesn't solve, no matter how good the generator is
If production data is inconsistent, poorly validated, or missing critical fields, synthetic data trained or modelled on that same flawed pattern will reproduce the flaw convincingly rather than reveal it. The fix for a quality problem is a quality programme — validation, contracts, ownership — not a more sophisticated data generator.
Synthetic data can hide a quality problem behind a more convincing-looking dataset. It cannot fix the problem underneath it.MNB Data Practice
Key actions
- Diagnose whether the problem is scarcity or quality before reaching for synthetic data
- Reserve synthetic data for privacy, scarcity, and augmentation use cases
- Fix quality problems with validation and ownership, not generation
- Audit whether synthetic data is reproducing a known upstream flaw