Google Photos' face recognition feature is pretty good.. until Google started to think person A and person B are the same person and you want to fix it manually: which is basically impossible.
I used to use this feature heavily, and have to give up. Don't get me wrong, I still think Google Photos is the best in class. The point is, even the "best" is not enough due to the lack of options to fine tune it. And once it breaks, it jumps from "holyshit" to "unusable" straightly.
I wrote a very long article to detail it somewhere else
before, here is the gist of it:
-------
* There is no way to tell Google to "split" a person/face. All you can do is remove wrong results (person B) from person A (you can also say it's "wrong person" on web, but not in app). You can't manually re-tag them as B in batch.
* These removed results don't seem to get re-recognized as other people; it's either not get re-analyzed at all, or often times, they will be re-recognized back to the wrong person again!
* You can't multi-select more than a few hundreds pictures and remove. It says "can't remove results" on both app and web if you select, say, 1000 photos. This is extremely annoying since I have people that have 10k images and I have to do so in small batch.
* You can manually assign a face, but can only do so one by one. And that's assuming Google actually detects the face, no matter how obvious it is.
* The people you added manually (by selecting a face from a photo, and manually add a name) are somehow treated as second-class citizens. When you visit https://photos.google.com/people, their names appeared at the very end of the list of named people (no matter how many photos you have for him/her), and they are almost never considered as face-recognition candidates when you upload new photos. I guess Google simply doesn't use the manually assigned people/face in their AI.
* And the final nail on the coffin is, even if you went to the trouble and fixed all these problems manually, which I did, nothing stops Google to recognize the newly uploaded photos wrong. That means you need to repeat the process again and again.
Also from an AI/machine learning point of view (I'm guessing, I'm not in this field), the more tricky cases that the user allowed/confirmed, the model for that person became more overfitted. So eventually it started to think other people are the same guy/gal.
This is typical, but the problem is, you, as the supervisor, can't tell the AI that he was wrong, and re-tag the training set as easily.
I think all this can be easily fixed if Google can add a way to "split" a person, and then force the AI to re-train the models based on the photos from now two different people. But I won't hold my breath for a free service.