Synthetic Data

Synthetic data companies produce data rather than collect it. The category covers two fairly distinct engineering traditions that share a name. One generates structured data — tabular, relational, time-series, and increasingly text — by fitting a model to a real source dataset and sampling from it, usually to remove personal data from development, testing, analytics, and sharing workflows. The other renders it, using 3D simulation, digital twins, and physically based sensor models to produce imagery, video, LiDAR, radar, and robot trajectories with labels attached at generation time, for perception and control systems where real edge cases are rare, dangerous, or expensive to capture.Establish which tradition a vendor belongs to before comparing anything else, because the evaluation criteria barely overlap. For structured generation the questions are how utility and fidelity are measured against the source data, what formal privacy guarantee is offered and whether it is differential privacy or an empirical re-identification test, and how referential integrity is preserved across related tables. For simulation they are which sensor models are supported and how they are calibrated, how domain randomisation is controlled, what is actually known about the sim-to-real gap for the task in question, and whether asset libraries and environments can be extended by the customer. In both cases, ask what the vendor retains: some generation workflows require the real source data to leave the customer’s environment.This category is often confused with data processing and curation, which reshapes data a customer already holds, and with the privacy tooling under governance and compliance, which masks or tokenises real records rather than generating new ones.