> This is a strong claim that just downloading articles into training data is what violates the copyright. That GTP outputs verbatim copies is a red herring.

It's the other way around. There is no infringement if the model output is not substantially similar to a work in the training set [1]:

> To win a claim of copyright infringement in civil or criminal court, a plaintiff must show he or she owns a valid copyright, the defendant actually copied the work, and the level of copying amounts to misappropriation.

The questions are, which parties should bear liability when the model creates infringing outputs, and how should that liability be split among the parties? Given that getting an infringing output likely requires the prompt to reference an existing work (which is what's happening in the article), an author of a work, an element in an existing work, or a characteristic/style strongly associated with certain works/authors, I believe that the user who makes the prompt should bear most of the liability should the user choose to publish an infringing output in a way that doesn't fall under fair use. (AI companies should not be publishing model outputs by default.)

[1] https://en.wikipedia.org/wiki/Substantial_similarity#Substan...