It doesn’t work, for point 1 very well though. The tech is fine, but the way it’s presented to users is that it’s way more accurate than it actually is. That’s marketing rather than a technical problem. Second, the tech is not as good at recognizing non-white people. It’s just a fact that there are more pictures of white people to train the tech on since white people have historically had more access to photography among other reasons. And the models used to create most of the tech was built to favor facial traits that are more likely to differ in white people.

So, the likelihood of high probability matches is much lower so the likelihood that the highest probability match that is made is actually much lower probability of it being an actual match means the bad matches bubble to the top and get accepted as real. And these kinds of uses are more interested in a “better safe than sorry” stance and they aren’t sorry about killing the wrong person, only about not killing the right one. So they’re perfectly as happy killing many people that are possible matches as they are one person that’s the correct match.