That's not the take (although in a sense I agree training data should influence it especially if it materially reproduce training samples)

Instead the argument is that the individual outputs from ML can only be copyrighted if they carry a human expression (because that's what the law is specifically meant to cover), if there's creative height in the inputs to it resulting in an output carrying that expression.

Compare to photography - photographs aren't protected automatically just because a button is pressed and an image is captured, rather you gain copyright protection as a result of your choice of motive which carries your expression.

Too simple prompts to ML models would under this ruling be considered to be comparable to uncopyrightable lists of facts (like a recipe) and thus the corresponding output is also not protected.