The training data still has to come from somewhere. This just means churning out cookie cutter filmography can even more shortcuts than green screen CGI.
Most films don't reinvent cinematography, but follow established patterns, maybe with a twist. Sure, there are the Roger Deakins of the world, but I'd wager that most cinematographers aren't that.
I find this to be the real strange bit of the "AI art debate". All the management and publishing are very bearish on AI, even though its legality relies on wide-scale copyright infringement. Actual artists are far more skeptical, begging for copyright to become a de-facto AI ban, even though they're way more likely to infringe copyright than to own it. And then you have the Free Software people - i.e. the people opposed to software copyright entirely - who were the first to actually point out that AI training was infringement and trying to sue over it.