Synthetic Data
Artificially generated data used to train, test, or evaluate models.
Invent three returns to test a library kiosk
A developer uses fabricated events to exercise the returns screen. The three records were designed for the test and have no link to actual patrons.
- Ordinary return
- Demo item A11 is scanned at 14:05. The screen should confirm a single return.
- Repeated scan
- A11 is scanned again at 14:06. The test checks that the inventory does not count two returned copies.
- Missing timestamp
- Demo item B22 arrives without a recorded time. The expected result is a recoverable error with the item identifier preserved.
What this test set can tell you
The cases show whether three known paths work. They say nothing about how often duplicate scans occur in real branches. Synthetic records learned from real examples require additional checks for usefulness, bias and disclosure risk. A synthetic-data label alone establishes none of those properties.
What Synthetic Data Can and Cannot Do
Synthetic data is helpful, but it is not automatically equivalent to real data. If the generation process oversimplifies reality, the model may learn a distorted world. If it fails to capture long-tail behavior, bias, or noise patterns, the resulting system may look stronger in the lab than it performs in practice. Synthetic data is only as useful as its fidelity and relevance.
That is why synthetic data is often best used as a complement to a strong Training Set, not a careless replacement for one. Teams still need careful evaluation to confirm that models trained on synthetic data generalize well.
Follow the idea
Training Set · A training set has a different purpose from a small suite of regression cases.
Verification checks each observed outcome against its expected result.
Related concepts (5)
Examples in the AI guides
Related Yenra articles (3)
Source · NIST guidance on synthetic and de-identified data
Return to Synthetic Data in the A–Z glossary