This is disingenuous. Large models are observed to reproduce works from their training data, sometimes verbatim, often close enough to constitute a derived work. You scraping metadata would never be able to do that.
I agree that generative models have been observed to reproduce training input verbatim, but that's not what I understand to be at issue in the article. The issue here is whether you can presume the consumer of a list will do something illegal with an item on the list without any evidence, and therefore threaten the compiler of the list.
I can produce work myself that is infringing or not infringing as a result of learning.