> Black compared the left thumb in the picture with the Hogmanay image and found matching details, including an unusually shaped lunule, the white area at the b

> Black compared the left thumb in the picture with the Hogmanay image and found matching details, including an unusually shaped lunule, the white area at the base of the nail. "This time, I was able to go back to my database and put statistics to the data." > Once they have established the offender's features, they study images of the suspect, trying to establish a match. The method as described by the article seems precarious. There's a lot of room for human bias to sneak in when there's prior expectation of which thumbs should match. Easy to be lenient when matching features to the suspect's thumb, and strict when matching features to the dataset which you already know shouldn't match and would undermine your method if you found matches. Deciding the features by comparing the suspect to the offender also introduces pitfalls. If randomly selected hands have 5 out of 20 matching features, then in most cases you'll be able to claim that a completely random hand matches the offender's hand in a way that 0% of images in the 500-image database do. An improvement might be to have one person indiscriminately categorize features of the suspect's hand, and a separate person indiscriminately categorize features of the hand in the video. Neither person should see both images - especially not when deciding which features are relevant. > The Oketch case presented her with two technical problems. First, he was black, "and all the people we had looked at previously had been white. This has a lot of the smells of evidence analysis techniques that we find out in 20 years were ill-founded and lead to a bunch of false imprisonments. > A suspect can be excluded with 100 per cent certainty, but a match can only carry a grade of "strong support" that the suspect and the offender are the same person. This equates to between a 1-in-1,000 to 1-in-10,000 chance that it could be someone else. I worry that courts aren't properly equipped to understand that false positive rates aren't "the chance that it could be someone else". As an extreme scenario, imagine it's 2050 and you're suspected of a crime based solely on a match from a comprehensive global facial recognition system (~10 billion faces). If the system has a has a 1-in-1,000,000,000 false positive rate, then that still means there's a 90% chance you're innocent (if given no other information).