They are like workers that are hired.
Laws around “works for hire” come with their own copyright baggage that assumes workers are actual people; for instance, these laws include mechanisms by which workers can claim copyright themselves.
I’m not opposed to the general principle of training a model on copyrighted works potentially being fair use; however, the generated works would need to be sufficiently novel or seemingly “creative” by human standards for it to work. Otherwise, you’re in “derived work” territory. This, I think, is a major difference between the likes of DALL-E and Midjourney, and the likes of Copilot.
I personally found all the discourse way too hilarious, and wrote a satirical article on it only to get clobbered by Poe’s Law: An experiment to test GitHub Copilot’s legality.