Training data generated by a model rather than collected from the real world. It is used because high-quality human-produced data is finite and expensive, and because you can generate examples of rare situations that would otherwise be hard to find. It works well in some settings and causes problems in others.
Technical view
Most effective where outputs can be verified automatically, such as code that either runs or does not, or mathematical results that can be checked. Riskier for open-ended content, where model-generated training data can amplify existing biases and narrow diversity across successive generations, a degradation sometimes described as model collapse. In practice it is used alongside human data and filtering rather than as a replacement.
Related
Part of AI Terminology Index, section: Testing, Safeguarding and Running Models in Practice.