If you have a reasonable group of people like a friends and family group, you can get machine learning to be very accurate (with odd exceptions like babies and twins). But what is 99% accurate for 50 people will almost universally give false positives when searching a group of 50 million faces. The math just doesnt work for them.

It can't.

Also, and I have no idea what race this person is, but facial recognition tents to be even worse when it comes to minority groups, especially black people. Because its more often created and tested by white or asian dudes.

Which has absolutely terrible implications when you consider it gets used for things like Flock, for crimes, which already have a racism problem skewed against black people.

Zero. Zero accurately. They don't even talk about it, because they don't care. The arbitrary cruelty is a bonus feature.

Facial recognition has an inherent bias as well, or at least it did not too long ago. It was mostly trained on white people, so it's accuracy with non whites was really bad. I'm sure the training data is far more robust now, but that doesn't necessarily mean better.

This has happened

Not at all, but that’s the court’s problem I guess.