99.6% accuracy sounds good
Until you realise that, in a million uses, there will have been about 4000 false matches.
And Ton-That is reluctant to swear that it is as high as 99.6%
US police have used Clearview AI facial recognition tech to conduct nearly one million searches since the company launched in 2017 – but its founder and CEO said he's still unwilling to testify to its accuracy. Those numbers were provided by Clearview CEO Hoan Ton-That to the BBC, along with another startling number: The US- …
Measured how? If that's "you have a photo of a suspect and our system matches it to a person you believe that is" i.e. a 1 to 1 match then 99.6% is about what you'd expect. Apple claims 1 in 50,000 false matches for Face ID but that's with 3D scanning and ideal conditions, so 99.6% (i.e. 1 in 250 false matches) sounds about right for simple 2D photo matching where you may have an off angle CCTV photo of your suspect to match with a few angles of mug shots.
The problem is police won't be using it that way. They'll have a photo of a suspect and want it match it against many photos. Like their entire database, though maybe they'll make it easier by saying "well the suspect is obvious a fat white male" so they can eliminate women, minorities and anyone under 250 lbs from the database they try to match against. No one has photo matching technology that does many to one matching well, so you have two choices. Either accept that you may miss the true match if you crank up the certainty really high to avoid false negatives, or if you want to minimize the chance you miss the true match you must accept that you will get a long list of potential matches all of which or all but one of which will be false matches. And it will suck to be one of those false matches, who idiot beat cops will treat as suspect because "the computer told them so".
The reason Face ID works so well is it is only trying to match to one face. If you pick up my phone it just has to look at your face and figure you aren't me. Unfortunately people (including decision makers buying stuff for police departments) see how well that works for the limits it is designed under and assume the technology will work just as well trying to match 1 to many. If Apple wanted to allow me to pick up a Apple store display model and have it recognize me specifically out of the billion iPhone users facial matching software would have to be about a million times more accurate than the state of the art. For all I know there's some guy in Europe who looks so much like me my own mother would do a double take, so this might not even be possible given a large enough database.
The actual accuracy isn't as important as the claimed accuracy.
The point is to grab a 'person known to the police', confront them with a 99.9% match and suggest that the might like to confess to a minor charge, or you will go up against a jury with a guaranteed 99.6% guilty match by the computer.
That 99.6% match would be torn apart by a good lawyer. Unfortunately only clients with money can afford good lawyers, the rest get overworked public defenders who don't have time or resources to bring in expert witnesses to teach the jury the difference between 1 to 1 and many to 1 matching and why the 99.6% claim is misleading.
In all US jurisdictions both sides can reject a certain number of jurors without showing cause by peremptory challenge, and others can be rejected by counsel or judge during voir dire.
While the right to a jury trial is a critical civil right in the abstract, in practice it often doesn't work out well. Potential and actual jurors often don't take their responsibilities seriously. When they do serve they often arrive at prima facie wrong conclusions, for the most vapid and hateful of reasons. Unfortunately it's a right that people are often better off not exercising.
Have you served on a jury in a criminal case? I have (in fact I was the foreman). There's a lot of pressure on a dissenting minority to agree with the majority during deliberation. In my case I tried to mitigate that as much as possible, and under the circumstances there wasn't a lot of disagreement anyway; but in general all it would take is one or two bullying blowhards to make many jurors decide it's not worth swimming against the tide.
Statistically, in the US, criminal defendants are better off not requesting a jury trial.
And historically, dubious forensic evidence like Clearview's snake oil has frequently been used to secure a conviction. Many judges will disallow the defense from challenging the "science"; this has happened with everything from fingerprint uniqueness to DNA matching to facial reconstruction to bite identification (which is pure, unadulterated bullshit).
Clearview is dangerous and invasive and should be shut down, full stop. No number of legitimate case closures justifies its existence.
If they aren't saying, then the answer is somewhere between zero and one inclusive.
Also "helped" is a rather vague term. It includes "used but match was wrong", and "used, someone was beaten up and brought in for questioning but they had a cast-iron alibi on account of being a thousand miles away at the time, but it was definitely them that did it, just look at their skin colour".
Which is of course the other problem. Institutional racism is endemic in police forces and these tools are known to be exceptionally poor at matching minorities and darker skin.
I remember when those 2 girls were murdered by school caretaker near Cambridge.
They released an unrecognisable still form a CCTV camera in a shop hours earlier and then claimed that CCTV helped solve the crime and Cambridge should have more CCTV than punt tours