Re: Real time false positives and trigger happy policemen with guns
"I actually think that UK style trials (limited number of suspects, public places) are sensible to do and involve no more than a few minutes of time and an apology."
No, and that attitude is flat out _dangerous_
Without salting the trials using mugshots of people you KNOW are in the crowd, you're ignoring the possibility of false negatives, let alone quantifying any rates - lest you think that's unimportant, I refer you to a similar issue expressed in star ratings: https://xkcd.com/937/
Email list companies have known about this problem for decades. It's very hard to prove someone hasn't opted in, but actively salting spammer lists with canary traps PROVES that the company hiring your services is a spammer. I know of several companies which hold the first few mailings of any new customer to check the outbound queues AND look for obvious spoor before letting the stuff out - there's no indication of any such quality checking in any of these facial recognition systems being deployed and they're far more dangerous than a few billion pieces of email getting loose on 't intawebs.
Another indicator of the dangerousness of these systems: The ENTIRE european genome can be expressed in around 20,000 individuals. Chinese genome variation is a bit smaller. The east-asian genome in general is only 50-60,000 people wide and it gets smaller as you fan away from Formosa (The entire african one can be expressed across about 2 million individuals, whilst the entire East Polynesian genome comes down to less than 5000 individuals).
(*) Which makes it more obvious the ubiquitous chinese facial recognition system is less about tracking potential offenders and more about keeping an eye on people who aren't Han.
The TL;DR of this is that there are a LOT of people out there wearing similar-to-near-identical faces
As with fingerprinting, the premise of "uniqueness" is fatally flawed by samples having too small a dataset to validate that claim (unique amongst the pool of criminals known to scotland yard perhaps, but across the entire population? This is what agencies are finding out to their cost as they hoover up more data)
The same thing was found to apply to DNA fingerprinting as practiced in law enforcement
Ditto Iris matching. (retinal vein layouts are likely more unique, as are subdermal hand vein layouts - these vary wildly even between identical twins - a factor of genetics AND local conditions during gestation.
Facial recognition is likely to be another gallon of snake oil - great for proving someone's NOT who you're looking for but almost useless at showing they _are_ the droids you seek.
The problem - as always - is that there are a lot of people who stand to make out like bandits from selling snake oil and a lot of people who stand to make lots of political/power points by buying it.