Hacker News new | ask | show | jobs
by flumpcakes 749 days ago
One in forty false positives still isn't bad for a system to automatically sift out wanted criminals. That's going to be orders of magnitude better than police stopping people who 'fit a description'.
4 comments

That claimed false positive rate sounds unbelievably low to me. Is there any proof publicly available and verified openly by a third party?

Why I find it hard to believe is that I am imagining having access to a face photo of everyone in the world, and trying to pick a specific close friend from those photos. Given infinite time to find the photo of that one person, I am convinced that I would find tons of false positives that might be that person, and that I would never even be sure I found the correct photo.

I am guessing that there is some serious sampling bias leading to such a low false positive rate (assuming the number is not complete fiction).

According to government (referring to "live facial recognition"): "All deployments are targeted, intelligence-led, time-bound, and geographically limited." https://homeofficemedia.blog.gov.uk/2023/10/29/police-use-of...

which makes that 1 in 40 number completely irrelevant when discussing false positives from mass deployment in grocery stores or other non-targeted spying.

There is zero reason for a police van in the middle of a city to sift through the whole internet. Everyone identified as living in that particular area is fine, and cuts down on a huge number of matches. The government also has a lovely national database of facial ID’s with names, so they don’t even have to rely on shitty internet photos.
> One in forty false positives still isn't bad for a system to automatically sift out wanted criminals.

It's an absolutely disastrous false positive rate when the consequences to those who are falsely identified are more than trivial.

That one in fourth will be banned from using nultiple stores and establishments without due process and without a clear way to clear themselves up.
OK, can you show the numbers you are using?