It's an understandable move. There's a lot of possible drama on the copyright front in the future. It's important to note, however, that this isn't a defense of "artists" (who, in reality, encompass various types of craftsmen), but rather a defense of Valve itself. In practice, this would probably lead to companies like Adobe paying 30 to 40 people for a stylization training dataset, while sourcing everything else directly from content clearly labeled as public domain. This would effectively clear copyright concerns for everyone. As for the notion of "copyright" right now, the idea itself of how it actually works in images isn't as settled as in music. In music, melodies and samples are copyrighted, but styles and imitation aren't. As someone who spent 5.5 years studying art history, I would say that while "AI" is not the future (simply because it's a meaningless buzzword), machine learning is. As a supplement to human intentions, not a replacement. So, the questions of the day are: how much human intention is required, how on earth can we vet a model training dataset (it's practically impossible on publishers end), and to what extent can we infringe upon someone else's "style"? We'll see how they figure out all the details de jure.
p.s. grammar corrected by gpt-4 :3