Definition
Synthetic data can be produced by simulations, statistical models or generative AI. It is used to protect privacy (sharing realistic records without real people in them), to cover rare cases such as fraud or equipment failures, to create test sets and evals, and to train smaller models from larger ones.
Its value depends on fidelity: synthetic data that misses the quirks of the real world produces models that fail on real inputs. It can also leak information about the original data if generated carelessly, so privacy claims need testing.