Copyright infringement is never theft.

And, arguably, what they're doing with LLMs isn't even infringing copyright. If I look at a copyrighted picture, learn from it, then paint my own impression of it, my painting shouldn't be infringing the copyright. Do that with an LLM instead of a brain and it's a similar argument.

The dual standard is really the issue. Meta downloaded terabytes of books from LibGen and loaded them into its model. If that's not infringing copyright, then anybody should be able to download a book from LibGen and read it without worrying about copyright infringement because they're just loading them into their brains. But, I have a feeling that Meta will get away with it as fair use, but individual people will still be nailed for "copyright infringement" for loading media into their brains in exactly the same way.