"Accurate" isn't a terribly meaningful description. We really need to talk about sensitivity and specificity. If you're looking for a very low-probability event

"Accurate" isn't a terribly meaningful description. We really need to talk about sensitivity and specificity. If you're looking for a very low-probability event (a terror suspect in a random crowd) then your facial recognition algorithm needs to have phenomenally high specificity to be useful, otherwise you'll be glutted with false positives. This isn't a problem isolated to facial recognition - many cancer screening tests are worse than useless, because the harm of unnecessary treatment due to false positives outweighs the benefits of early detection. https://en.wikipedia.org/wiki/Base_rate_fallacy https://en.wikipedia.org/wiki/Base_rate_fallacy