Specifically it’s data useful for training language models. Using text generated by another model has been shown to be counterproductive. This is why so many companies are trying to lock down access to their user generated content and set up data sale agreements for access to data prior to the major release of large language models.