It’s probability, like you said. When you upscale you have to decide if the missing pixel is a smooth interpolation between neighboring pixels or a sharp transition. That decision will always be a guess with a large chance of error. For TV or gaming that’s often OK upscaled videos look nice and most of the time errors are small and not noticable.

For police purposes, if the difference wasn’t noticeable, they wouldn’t need upscaling in the first place. They’re trying to zoom in further than the video data will go. Nobody should face judicial consequences because an algorithm guessed at the interpolation between pixels and now it looks like your face instead of another person’s.