Synthetic Data

Synthetic generation is the answer to a genuine bind: realistic data is needed for testing, analytics development and model training, and real personal data is increasingly unlawful, impractical or reputationally unwise to copy into those environments.

Quality is measured by how many properties survive generation. Weak generators produce valid-looking rows with uniform distributions, no correlations and no anomalies — safe, and useless for finding the defects that live in skew and edge cases. Strong generators preserve marginal distributions, inter-column correlations, referential integrity across tables, temporal sequences and the long tail.

The subtle failure is privacy leakage through overfitting. A generative model trained on a dataset with rare records can reproduce them nearly verbatim, so a synthetic dataset containing the only customer in a given postcode with a given rare condition has not protected that person. Serious use requires evaluating both utility and privacy, typically with distance-to-nearest-record and membership inference checks rather than assuming synthesis equals anonymity.

Where it is now unambiguously worth it: high-volume test datasets, sharing across organisational boundaries, and populating ephemeral environments at scale.