Evaluation sets are difficult to build because the most realistic examples often contain customer data, internal decisions or commercially sensitive context. Synthetic data can help, but only when it is treated as an evaluation asset rather than filler.
The goal is not to create fake data that looks convincing. The goal is to create controlled test cases that exercise real behaviours without exposing real people or customers.
Use synthetic data for coverage gaps
Start with the behaviours that your current evaluation set misses. Common gaps include rare intents, policy edge cases, multilingual phrasing, conflicting evidence, missing information and adversarial prompts.
Synthetic data is especially useful when collecting real examples would be slow, risky or biased toward the most common cases.
Keep the scenario structure realistic
Synthetic examples should reflect the structure of real work: user role, task, source material, expected answer, refusal conditions and evaluation rubric. Without that structure, synthetic cases become generic prompts that do not predict production behaviour.
For RAG systems, include synthetic documents and expected citation behaviour. For agents, include tool states, permissions and expected stop conditions.
Protect against accidental memorisation
Do not derive synthetic cases by lightly rewriting sensitive records. Use abstraction, fictional entities, value ranges and scenario templates. Privacy review should focus on whether a synthetic example can be linked back to a real person, customer, deal or incident.
Connect this process to evaluation data handling and data minimisation.
Label synthetic data clearly
Synthetic cases should be tagged by source, generator method, intended behaviour, risk category and review status. This avoids confusing synthetic coverage with real production distributions.
Teams should also track which cases are hand-authored, model-assisted or derived from anonymised patterns.
Validate with human review and production signals
Synthetic evaluation data should be reviewed by domain experts. It should also be compared against production feedback over time. If real users reveal failure modes that synthetic cases did not cover, update the scenario library.
Synthetic data is not a substitute for production monitoring. It is a way to make pre-production evaluation broader, safer and more deliberate.