This take is correct although I would make one addition. It is true that copyright violation doesn’t happen when copyrighted material is inputted or when models are trained. While the outputs of these models are not necessarily copyright violations, it is possible for them to violate copyright. The same standards for violation that apply to humans should apply to these models.
I entirely reject the claims that there should be one standard for humans and another for these models. Every time this debate pops up, people claim some province based on ‘intelligence’ or ‘conscience’ or ‘understanding’ or ‘awareness’. This is a meaningless argument because we have no clear understanding about what those things are. I’m not claiming anything about the nature of these models. I’m just pointing out that people love to apply an undefined standard to them.
We should apply the same copyright standards to people, models, corporations, and old-school algorithms.