Umm, no? If someone happens upon some funny cat meme that 4chan users made with an intentional hash collision then they're not guilty of anything.

A poor analogy could be trolls convincing a flash mob to dress like a suspect's description which they overheard with a police scanner. No one in the mob is guilty of anything more than poor fashion choice.

The point made was that there are always flaws in these sorts of approaches that lead to false positives. If you can discover the flawed pattern(s) that leads to false positives and engineer them into seemingly harmless images, you can quite literally do what OP I'd suggesting. It's a big IFF but it's not theoretically impossible.

The difference between this and hashes that require image data to be almost identical is that someone who accidently sees it can avoid and report it. If I can make cat photos that set off Apple's false positives, then there's a lot of people who will be falsely accused of propagating child abuse photos when they're really just sending cat memes.