Usually these systems are implemented via a generic face detector that first runs on the whole image, to locate all the faces that are present. After the faces are located, they’re matched against a set of candidates (which I would guess is an enumeration if all your Nth order friends [who don’t opt out of this], to reduce computational complexity).

This isn’t the only possible way to do it, but I believe it’s the most common and effective approach.

So it’s not at all unlikely that when running the identifier, they simply don’t match against people who say no to this. They’d never be recognized at all, and the system would just identify their face as an “unknown” identity.

I cannot say for sure whether or not this is how they implemented it, but it seems likely.