What is synthetic data?
Synthetic data consists of artificial records generated by models or rules to resemble aspects of real data.
How it works and where it is used
It can support software testing, training or simulation of rare cases. A generator learned from real data may preserve useful statistics but misrepresent small groups. “Artificial” does not automatically mean private: memorization and re-identification risks may remain.
Limits, safety and selection criteria
Measure utility, distribution shift and privacy separately for the intended use. A system that performs well on synthetic samples can fail in the real world. Document provenance, generation method, lost properties and usage limits. Synthetic data imitates patterns in real records and can support software tests or experiments. A payment team might generate plausible transactions, including rare failures. If generated rows are too similar to actual people, privacy risks remain. A generator can also miss uncommon groups and produce misleading model performance. Evaluate statistical similarity, diversity, re-identification risk and performance on the target task separately. The word synthetic is not a privacy guarantee, and a useful test set should include realistic edge cases.