Earlier quoted context omitted.
False positives are clearly astronomically unlikely. Not a real issue. Engineered collisions seem unlikely too. Not impossible. Unless there is a straight up cryptographic defect in the hash algorithm, it seems hard to see how engineered collisions could be made to happen at any scale.
Why do you think that? There are plenty of whitepapers on fooling NNs by changing random pixels by a bit, so that the picture is not meaningfully changed for a person, but the computer will label it very differently. Do note that these are not cryptographic hashes because they have to recognize the picture even when compressed differently, cropped a bit, etc.
What is the actual attack you are imagining?