Research · Data for AI

Synthetic Data

Synthetic data is generated to mimic the statistical properties of real data. It can speed up AI development, fill gaps and reduce exposure of sensitive records, but it is not automatically private or accurate.

What it is

Synthetic data is produced by algorithms rather than collected from real events. Methods range from rule-based generators and statistical models to machine learning techniques that learn from a real data set and produce new records. Generative AI models are also used to create synthetic text, images and conversations for training and testing.

Where it delivers value

• Test data for AI and software without using real customer records.
• Augmenting training data for rare events such as fraud or equipment failure.
• Creating evaluation sets for generative AI and agents.
• Sharing representative data with partners or researchers.

Risks and pitfalls

• Poorly generated records that can still reveal information about real individuals.
• Data that looks realistic but misses important patterns or edge cases.
• Bias in the source data carried into the synthetic version.
• Models trained heavily on AI-generated data degrading in quality over time.
• Treating synthetic data as outside privacy rules without assessment.

How to evaluate it

Judge synthetic data on three questions: fidelity (does it preserve the patterns that matter?), utility (do models trained or tested on it perform as expected on real data?) and privacy (can records be linked back to real people?). Each needs its own tests.

How to get started

• Start with test and evaluation data, where benefits are clearest.
• Define fidelity, utility and privacy measures before generating.
• Involve privacy and legal teams in approving methods.
• Label synthetic data clearly so it is never mistaken for real records.