In general you can't, but surely it's not that big a deal if ChatGPT offers an inaccurate summary of a movie you're about to use to kill time on a flight? I suppose it becomes important if, e.g., you're relying on it to tell you whether a movie is appropriate for children, but, if you're just asking it whether a movie is worth watching, that's a question that doesn't have an objective, factual answer anyway, so a hallucinated answer is probably about as useful as that of a not-previously-known reviewer.

Sure, but that's the filmmaker's interest. As someone sitting on a plane trying to decide whether to watch a movie, I care about my interest, not that of the person who made it. I'm not particularly arguing for the use of ChatGPT here (I wouldn't use it), just that the risks it usually poses are fairly minimal in this case.

You don't even seem to be disputing the actual results here, just gesturing towards a kind of philosophy class exercise of whether we can ever "really" verify its accuracy. I see Wittgenstein's name increasingly tossed around in these parts (a good thing!), so I'll just note that one of the reasons he's hailed as one of the great philosophers of the 20th century is because he felt these puzzles about "really" knowing were frivolous.

I don't think I agree that what's needed here is some new and extra process of verification. I think the same usual quality control criteria that are already being used are good enough in this case.