The moment you need ML magic to magnify you are basically telling the ML "guess what is there based on what we can see and the training database you where trained on". Also due to the nature of such algorithm they do not have any form of common sense or "realistic general purpose" world view, and tend to be black box so it's hard to check if they learned something "stupid"/"ridiculous".
So it's totally possible that somehow it learned to place knifes into the hyper-magnified hands if they have a specific skin color and a turquoise sleeve on a picture with a color scheme as if it's a cloudy day. Or similar arbitrary things. Sure it's supper unlikely but not impossible and there is not good way to find such "bugs". Worse multiple ML systems trained on the same data might share some mistakes and there are just that many huge comprehensive datasets.
Sure, but that depends very strongly on the number of natural images you trained on and the specificity of the situation you present to the algorithm. Maybe one such structure in the training set just happens to be people holding knives given "a specific skin color and a turquoise sleeve on a picture with a color scheme as if it's a cloudy day".
But there is no good way to reliably know if it's done well for this specific image.
So it can't be trusted.
Which doesn't mean it's worthless.
It's kinda like someone saying "I think I maybe have seen <something> when passing by but I didn't pay attention."
It's not trustable as the human mind if fickle but it can be a good starting point.
I love these guys so much