Synthetic data

Artificially generated data intended to statistically resemble a real dataset without containing actual real records. Synthetic data carries a formal, testable privacy guarantee only when the generative model producing it was itself trained under differential privacy with a stated epsilon; otherwise its privacy properties are unquantified. (see Topic 4.5)

Training or fine-tuning data generated by an algorithm rather than collected from real-world sources. A legitimate acquisition option that reduces most personal-information consent questions but remains subject to the same data-quality obligations as any training data, and can carry its own licensing and memorization risks.

Data that was manufactured by a process rather than measured directly from a real person, transaction, or event. It is designed to resemble real records statistically. It comes in three practical families: rule-based augmentation, generative-model output, and simulation.

Data generated by an AI to fill gaps in real data; useful when real data is scarce, but it can amplify the flaws of the model that produced it (including representation gaps from the original collection), so it is best used cautiously and verified before relying on it.

Artificially generated data resembling real data, used to augment (never replace) real data, valuable for rare edge cases and privacy-preserving sharing, dangerous as a primary training signal in high-stakes domains without real-world validation.