Consider this image. It's full of blatant tells, like the bouncer becoming a real boy from the knees down, or the improbable stacking inside that trenchcoat. Yet it obviously conveys meaning in a clever way. You wouldn't commend whoever made it for their drawing skills, but the image transmits an idea from their brain to yours.

The model did not have to comprehend anything. That's the user's job. A person used the tool's ability to depict these visual elements, in order to communicate their own message.

If some guy spends days tweaking out the exact right combination of fifteen unforgivable fetishes, that amalgamation is his fault. You would not blame the computer for your immediate revulsion. It tried its best to draw a generic pretty lady in center frame. But that guy kept doodling a ball-gag onto Shrek until image-to-image got the leather strap right, and once he copy-pasted Frieren behind him, it just made her lighting match.

Neural networks are universal approximators, so you're always going to need human art to approximate human art. However, there are efforts to produce models using only public-domain, explicitly licensed, and/or bespoke examples. (Part of the 'do words matter' attitude is that several outspoken critics scoff at that anyway. 'Like that changes anything!' They'll complain about the source of the data, but when that's addressed, they don't actually care about the source of the data.)

Personally, though... I don't have a problem with using whatever's public. For properly published works, especially: so what if the chatbot read every book in the library? That's what libraries are for. And for images, the more they use, the less each one matters. If you show a billion drawings to an eight-gig model then every image contributes eight bytes. The word "contributes" is eleven.