Should you use synthetic data?

Use synthetic data when real data is too sensitive to share, too scarce in rare scenes, or too slow to label, and only if you can test the result on real data. If masking would do, or rare events are the whole point, start elsewhere. Six questions below give a verdict.

How is the verdict worked out?

The rules are checked in this order, and the first one that fits your answers gives the verdict.

Decision rules used above, in order
If	Verdict	Why
No real data held back to test against	Not yet	Without a real test set you cannot tell whether synthetic data helps or quietly hurts. Set one aside first.
You need rare or hard-to-capture labeled scenes	Trial a simulation vendor	Rendered scenes come with exact labels and cover cases you cannot film. Measure the gap on real images.
Masked production data would do	Try masking first	Cheaper, simpler and closer to production. Generate only if masking leaves too much real data exposed.
Rare events are the target	Use with caution	Generators smooth away or copy outliers. Use synthetic data for pipelines and sharing, real data for the rare class.
Fewer than tens of thousands of source rows	Probably not	Small tables give the generator little to learn and more chance to memorize individual rows.
The data is too sensitive to share as it is	Good case for a tabular trial	This is the job tabular synthetic data does best. Ask for the privacy method in writing.
None of the above	Weak case	Real data under ordinary access controls may serve you better than a generator.

For the warning signs that show up once a trial is running, read when synthetic data fails. To see who sells it, go to the synthetic data vendor list.

About this decision aid

Who wrote these rules?

This site, from common practice in data engineering and machine learning. They are rules of thumb, not a vendor's assessment, and not a quote for any project.

Why does a real test set come first?

Every quality test for synthetic data ends with a comparison against real data the generator never saw. Without it, a model trained on synthetic data can look good on synthetic test data and fail in production.

Is synthetic data cheaper than real data?

Sometimes, per labeled image in simulation work. For tabular data the comparison is with masking and access controls, which are often cheaper. Only one vendor on our list publishes prices: Rendered.ai, from $5,000 a month (read 7 Nov 2026).

Can I use synthetic data to evaluate a model?

For stress tests and scenarios you design, yes. For the headline accuracy number, use real held-out data.