The “knowledge” is not the copyrighted works, unless it can reproduce them in full. Shannon and amount of entropy calculations can come in handy when deciding whether a given size neural network is even capable of holding a copy of the works.
In that regard, a collage is less of a transformative work, since it fully reproduces the original works… and there is no lower bound to how much or what quality of input a human needs to add to a collage for it to be transformative, so “a human spitting garbage” sounds like a valid enough transformation.