That's exactly what they're doing. And they're requiring multiple matches before an account is flagged, further reducing false positives. They claim they've set the threshold to achieve a ~1 in 1 trillion probability of an account being improperly flagged.

False, they are scanning all photos in your library, regardless of whether or not you send them on to anyone else.

1 in 1 trillion is not a particularly impressive number, considering that users have tens of thousands of photos in their library, and there are a billion active iPhones.

That's if you do believe their claimed false positive rate -- which I don't (yet), because they have published absolutely nothing about how neuralMatch actually works. They have an interesting body of work around how suspected matches are encrypted so Apple can't see them until certain conditions are made, but nothing about how they identify suspect photos.

That is a good point.