> For example you wouldn't know if a kind of steganography has been used to hide image in an image and that neuralhash picked on a hidden image.

How would NeuralHash pick a "hidden image"? It only uses the pixels of the image to get the hash. Any hidden image in the metadata would not even be picked up and no amount of steganography can fool NeuralHash.

> There is many vectors. For example you can leave phone unattended and someone can snap a picture of an image or since a collision may look innocent to you, you would overlook it in an email etc...

As iterated elsewhere in this thread, random gibberish pixels colliding with CSAM would definitely not be useful in incriminating anyone. The manual process would catch that. Also, if the manual process is overloaded, I'm pretty sure basic object recognition can filter out most of the colliding gibberish .