What is Synthetic Data?

Synthetic data is training data generated by a computer rather than collected from the real world, and in AI it usually means data written by another AI model. It gets used when real data is scarce, private, or too expensive to label. It works well enough to be standard practice and badly enough to be argued about.

Synthetic data exists because real data runs out, costs too much to label, or cannot legally be used. Start with the privacy case: a hospital that cannot share patient records can generate artificial patients with similar statistics and share those instead, since no real person is in the file. The scarcity case is bigger, because labs training on human text have been working through a finite supply, so they increasingly have models write training material for the next model. Done carefully this is genuinely useful, as generated examples can be aimed squarely at a model's weak spots, which beats scraping another mountain of what it already knows. Done carelessly it degrades, because flaws and blind spots in the generator get inherited and amplified by whatever learns from it, a failure researchers call model collapse. The safeguards follow directly from that risk: keep real data in the blend, filter the generated batch, and test the result against reality rather than against its teacher.

Flight simulators for training pilots. Real emergencies are rare, dangerous and impossible to schedule, so airlines manufacture them: engine failures on demand, any storm you like, a hundred times a week. Pilots trained this way fly better, because they have met situations a normal career never supplies. But a simulator only contains what its builders thought to put in it, which is why nobody lets pilots skip real flying hours.