I don't think it's true. Can a non-GPL backup tool backup files with GPL code? It will certainly be copying them, that's what backup is. I think you would agree that claiming it is a GPL violation would make no sense. What if that tool also calculates hash sums of the code? Still OK? What if it also builds inverse indexes on the content of the files, for better backup retrieval? Still OK? What if more complex calculation on the content of the files is done, for more efficient storage and compression? The resulting data set would certainly have something derived from the code, but it would not be anything like the original code. What if the same data set is used to train an ML model? I still don't see the boundary there, and "you can't touch bytes in any GPL code in any way" doesn't seem to be it.
I am not talking about "enforcing ownership rights" - if I train a model on GPL code and it produces some code, I am not asking about ownership rights on that code. For all I care, that can be public domain. What I am asking is how much of the original code should be left in this product for it to be still covered under GPL and how would one find such a boundary?