Synthetic Data

Artificially generated data intended to reproduce, to some extent, the properties of a real or plausible dataset without requiring every record to be collected at the time of use. It can support prototyping, edge-case testing, class balancing, or broader evaluation coverage when real data is scarce, sensitive, or expensive to label. Its usefulness depends on preserving the characteristics that matter for the decision being evaluated.

Synthetic data is not automatically private, representative, or correct. A generator can memorize source data, erase important subgroups, or produce examples that are too clean. Before using it for training or validation, teams need to check re-identification risk, coverage of real cases, statistical fidelity, and, above all, performance on the evaluation set that represents operation.