The other problem with the legal argument that it's "just like a person learning" is that corporations whose human employees have learned what copyrighted characters look like and then start incorporating them into their art are considered guilty of copyright violation, and don't get to deploy the "it's not an intentional copyright violation from someone who should have known better, it's just a tool outputting what the user requested" defence...

Exactly.

Also, it is only a matter of time until one of those employees (thanks to free will and agency) will whistleblow, it doesn’t scale, etc.

Frankly, the fact that such a big segment of HN crowd unthinkingly buys big tech’s double standard (LLMs are human when copyright is concerned, but not human in every other sense) makes me ashamed of the industry.