The problem isn’t that it does it regularly, but that it can do it, meaning that the copyrighted works are reproducible, regardless of how much the interface tries to hide that. That means the model isn’t really “learning” the same way a human would in any capacity (that should be obvious), but that it’s storing data that would violate fair use, and could generate copyright-violating portions of works.
Humans read and don’t retain the originals. The argument is that LLMs retain the originals, and that’s where the issue lies.