Well it's an automated system, so it's highly unlikely that someone is reading all (or any) of your messages. The volume of messages Facebook and Google process every day is astronomical, so no manual oversight process would scale. It's similar to how email spam filters have worked in a completely automated fashion for years. In this case, PhotoDNA works by comparing image hashes, so it probably has fewer false positives than spam filters.

But with a powerful search tool, one could go hunting for any 'type' of person they wanted based on social association info, geolocation or keywords. Its not benign or unweildy just because its large.

It mentions that humans verify hits in the above links - but yes, it would seem unlikely for the average user.