Glossary
Synthetic data
Synthetic data is artificially generated data that mimics the patterns of real data without containing real people's information. It's used to train and test AI systems when real data is scarce, sensitive or legally hard to use.
Sometimes you need data that behaves like the real thing without being the real thing. Synthetic data is generated, often by an AI model, to copy the statistical shape of a real dataset: a fake hospital’s worth of patient records with realistic ages and diagnoses, but no actual patients. Because no real person is in it, it sidesteps many of the privacy constraints that come with personal data, which is why it’s popular for training and testing AI in healthcare, banking and software development. It also matters because LLMs have read much of the public internet already, and synthetic data is one way to keep feeding them.
There’s an everyday version of the trick, too. A bar owner wants help designing a loyalty-card spreadsheet but doesn’t want to paste real customers into a chatbot. She asks the AI to invent twenty plausible customers and builds the formulas on those. Same help, zero exposure: a friendlier cousin of anonymization.
Worth knowing: synthetic data is only as unbiased as the real data it imitates, and poorly generated sets can still leak traces of the originals. When deciding what real data you can paste instead, the paste checker is the quicker route.
Where you’ll meet this
- AI vendors’ model documentation describing training-data sources
- Developer tools offering synthetic test data for apps and databases
- Research and policy discussions on training models after “running out” of web text