One of the big debates about facial recognition is how accurate it is. It’s a debate we consciously didn’t explore in our BBC Click report. That’s because as soon as you start talking about accuracy you get into an important but very gnarly debate about statistics.
However, given the amount of interest the story has generated, and given that some people are talking about accuracy, let’s unpack it!
Basically, when you walk past a facial recognition camera which matches your face against a watch list, there are four possibilities:
– The system correctly identifies you as being on the watch list. This is called a true positive.
– The system incorrectly identifies you as being on the watch list. This is called a false positive. It might result in an innocent person being stopped by police.
– The system correctly identifies you as NOT being on the watch list. This is called a true negative.
– The system incorrectly identifies you as NOT being on the watch list. This is called a false negative. It might result in a potentially guilty person slipping through the net.
When protesters talk about accuracy rates, they’re generally taking about the number of false positives. For example, if 300 people walk past the camera, ten get identified as being on the watch list, but only one identification is correct, you could say that the system is 90% inaccurate, because only one in ten of its matches were right. The rest were false positives.
However, the police will argue it differently. They’ll say that it correctly identified 290 people as NOT being on the watch list (“true negatives“). From that perspective, it was almost 100% accurate.*
When talking about accuracy, you need to know which result you’re talking about – false positives or true negatives.
There’s another thing to bear in mind: facial recognition systems can be tuned to make them more or less sensitive.
For example, if police are hunting a terrorist who’s about to bomb a bus, they might want to turn the sensitivity up. The system will over-identify people, creating more false positives, but they might consider that a price worth paying, given the risk.
For a low-level offender, they might turn the sensitivity down. This might mean guilty people slipping through the net (false negatives), but the consequences are less serious.
Depending on the sensitivity, rates of false and true positives and negatives can vary, even when tested on the same group of people.
*The sharper reader will have spotted a problem here – how do you know they were really true negatives? After all, the system let them pass without being stopped, so how can you check if they were on the watch list or not? It’s possible to test this, but only in controlled conditions – you take a group of, say, 300, you put a random ten of them on the watch list, then see how many of those ten the system incorrectly identifies as not being on the list. Simples.