And if you're confused about how a system could even be deployed if the accuracy rate is so bad it leads to a majority of those identified as criminals being fa

And if you're confused about how a system could even be deployed if the accuracy rate is so bad it leads to a majority of those identified as criminals being false positives... This can result from just measuring accuracy rates, instead of sensitivity and specificity or related error metrics. Say a classifier has a 99.9% accuracy rate for any given photo. If used to identify criminals, and if only 0.1% of the population are criminals, this will lead to a false arrest rate of 50%. It is a basic statistical fallacy literally called the prosecutor's fallacy, because it gets misued so much in criminal law, also called the base rate fallacy. 99.9% accuracy sounds high, and is above what state of the art performance is for noisy real-world surveillance data. But imagine this system is used on a city with a population of 1 million people. With an accuracy rate of 99.9%, the system will correctly identify a criminal as a criminal 99.9% of the time. But the actual proportion of criminals is, say, 0.1% or 1,000 criminals out of 1 million people. When the system scans all 1 million people, it correctly identifies about 999 of the 1,000 criminals (99.9% of 1,000). But here's the catch: it also mistakenly identifies 0.1% of the 999,000 innocent people as criminals. That's 999 false positives. So, the system ends up flagging nearly 1,998 individuals (999 true criminals + 999 false positives) as criminals.