Cameras
There are a lot of assumptions that are built into any camera derived machine generated IDs. 1) the camera lens is clean and does not have any flaws. 2) The algorithm used adjusts for non optimal view or distortions in its view. An excellent example is Google Earth handling of overpasses; you know those funny shaped dips where two highways cross. 3) Laziness on behalf of the low-bid agency doing the interpretation. Look at enough pictures long enough and they begin to lose their distinctiveness to coin a neologism. Start with one million suspects, winnow down to a dozen and the last half blend together.
The problem of lens distortion is such that license plates can be misread which are a lot more distinctive than a face. Personal experience here when Los Angeles City sent a Notice of Violation at a certain in location and time while I had a receipt in San Diego 10 minutes prior to noted time. Seems that there was something deposited on the lens that was 'read' as a character.
But any biologic is going to use the path of least resistance and the only way to correct errors is to raise the resistane such that other paths are more viable because they are easier!