Synthetic data

Data generated or simulated rather than directly observed in the target environment, often used for training, testing, privacy, or coverage.

Synthetic data can come from language models, simulators, procedural generators, perturbations, or privacy-preserving statistical methods. It can target rare cases or create labels cheaply, but it can also copy generator errors, miss real variation, contaminate an evaluation, or leak source material. Quality depends on purpose and evidence, not authorship alone.

Builder example

Training data needs provenance, generation settings, deduplication, coverage analysis, and task-appropriate validation. Real data is not automatically clean, and mixing real and synthetic records does not automatically preserve diversity; compare distributions and downstream performance.

Common confusion: Synthetic does not mean fake in the sense of useless, nor does it mean independently verified. A generated example can be excellent, wrong, or too similar to its source.