> The Metropolitan Police say that around one in every 33,000 people who walk by its cameras is misidentified. But the error count is much higher once an someone is actually flagged. One in 40 alerts so far this year has been a false positive.
A 97.5% accuracy rate seems… really really high!
Obviously facial recognition shouldn’t be the only bit of evidence. But that’s remarkably accurate.
It should also get worse as more people are added: In a large enough population almost every person has some look alike who happens to be a thief.
Another way to look at it is for the worst case person what is their false positive rate? It's going to be 100%: Joe Blow, who did nothing wrong to no one, will go from place to place getting kicked out for the rest of his (/the system's) life. The failures aren't random (or at least not entirely)-- some people will get a heap of error in their lap.
So while getting punted 1:10000 times you go somewhere unfairly isn't too bad, getting always punted is awful. The unequal distribution of errors creates bad incentives: the system is almost risk-less for those of us who aren't the victims of false hits, so why should we demand improving or banning it?
So high in fact that either the numbers are cooked or the threshold required for a positive match is set really really high, so they get a lot of false negatives, which opens the possibility that the threshold can be lowered in the future.
A 97.5% accuracy rate seems… really really high!
Obviously facial recognition shouldn’t be the only bit of evidence. But that’s remarkably accurate.