Synthetic data takes several forms. Rule-based generators produce records to a specification, such as fake orders for testing software. Statistical methods create tables that preserve the distributions of a real dataset without copying its rows. Language models write text, such as invented support conversations, from prompts or from real examples used as seeds.
Its usefulness to AI labs has limits. A model trained mostly on the output of other models tends to inherit their habits and errors, and synthetic text lacks the unexpected detail of real work. This is one reason labs pay for real company data. Synthetic data also is not automatically private. If it was generated from real personal data, traces of that data can persist, so its privacy status needs assessment like any other dataset. For sellers, the rule is disclosure. A dataset presented as human work should not contain generated or machine-translated records without saying so. The data card should state whether any records are synthetic, how they were made and how they are marked.
Marking synthetic records with a dedicated field lets a buyer filter them in or out without guessing.