I think there is a case to be made when AI models do produce copies. For instance, I think the NYT have a right to have an issue with the near verbatim recall

I think there is a case to be made when AI models do produce copies. For instance, I think the NYT have a right to have an issue with the near verbatim recall of NYT articles. It's not clear cut though, when these models produce copies, they are not functioning as intended. Legally that might produce a quagmire, is it fair use when you intend to be transformative but by accident it isn't? Does it matter if you have no control over which bits are not transformative? Does it matter if you know in advance that some bits will be non transformative but you don't know which ones. I presume there are people working on research relating to how to prevent output of raw training data, what is the state of the art in this area? Would it be sufficient to prevent output of the training data or should the models be required to have no significant internal copies of training examples?