What is synthetic data?

In plain English

Real data can be scarce, expensive or private. So labs create extra examples: practice maths problems written by a model, made-up customer records, or simulated road scenes for self-driving cars.

An example

A hospital wants to build a tool without sharing real patient files. It generates fake records with the same patterns as the real ones that belong to nobody.

Model collapse

There is a catch. A 2024 study in Nature found that when models are trained on the output of earlier models, again and again, they slowly lose the rare and unusual parts of real data. Their output gets blander and more often wrong. This is called model collapse.

Watch out

Synthetic data works best mixed with real data and checked carefully. As more of the web is written by AI, real human data gets harder to find and more valuable.