I think OP made it pretty clear:

if the submitters can’t prove they own the rights to the assets the AI was trained on

I would also say that hopefully gamedevs are designing/tweaking their own procedural generation too. Though I won’t disagree that lazy procedural content can/has been used for shovelware (and in a wider sense, filler). But I would say that AI can take that to a whole new level, and one that may fool some people on the surface (like having a really high-quality asset pack that can’t easily be pointed out).

Or worse when they can use AI to pump out content with even less effort than before. For an example, the new wave of (likely all related) fake science video spam channels on YT that are a step above older tactics (like a low-quality Text-to-Speech voice reading an existing article).

(on the other side of the coin, you can still use AI as a tool that is no longer turn-key… but I suspect in instances like that the artist would/should be able to prove that with their workflow steps. Then again, that probably doesn’t cut it as Valve likely means no tainted training data can be used even if original art was added in some way)