What is this AI-generated training material called?

Synthetic data is training material generated by AI models or computer simulations rather than collected directly from real-world human writing, images, or events. As AI labs increasingly run into limits on the amount of fresh, high-quality human-generated text available for training, synthetic data offers a way to fill in gaps, especially in areas where real examples are scarce, expensive to gather, or private, letting researchers generate large volumes of additional, tailored training examples.

Open-source data refers to information or code that is freely available for anyone to use or modify, a licensing concept unrelated to whether the content was generated by AI. Metadata is simply descriptive information about other data, such as a file's creation date or author, not training content generated by an AI system itself.

Researchers have raised concerns that training new AI models too heavily on data generated by earlier AI models, rather than genuine human-created content, can gradually degrade quality and diversity over successive generations, a feedback-loop problem researchers have nicknamed "model collapse."