Synthetic data consists of artificially generated records designed to resemble data from real people, events, or systems. In healthcare, synthetic genome data could support research while avoiding disclosure of an individual’s genetic information. section 10.25
Its appeal is practical as well as ethical: generated data can expand a dataset without collecting every record from human subjects. It can preserve privacy, represent rare scenarios, and scale more cheaply than authentic data, but resemblance to real data does not guarantee usefulness. section 4.13 section 10.25
Synthetic data is especially useful when authentic data is costly, scarce, sensitive, or difficult to collect:
- Augments training data, evaluation datasets, or other collections when additional human-generated records are expensive. section 4.13 section 7.9
- Simulates rare situations that may be underrepresented in naturally collected datasets. section 4.13
- Requires scrutiny because generated prompts may be less realistic, omit multi-turn context, or differ from real users’ language and behavior. section 7.9
- Works best as an augmentation rather than a complete replacement for authentic human data; experts remain divided about training with synthetic data alone. section 4.13
Synthetic prompts for evaluation
An evaluation team can ask a large language model to generate prompts for a specific task or imitate a small set of manually written examples. This produces a larger, cheaper test set, but evaluators should compare it with prompts from real users or production logs when available. section 7.9
