I was listening to a podcast episode that was talking about AI-based face recognition and how it’s being used.
The way it works, you take your shitty, off-center still from a security camera showing the suspect at an odd angle, and you feed it into the AI. The AI searches all the images it has access to to find people who look like the suspect.
Police take the results, pick up the person from that list that they think is the closest match, then they put them in front of a line-up, and ask the witnesses to identify the person who committed the crime. Once they have a positive ID, they use that to charge and convict the suspect.
The problem, of course, is that the AI is just finding people who look like the suspect. If they use the tool correctly, they would do some investigation to determine if there is any chance that anyone identified by the AI could actually be the suspect. Instead, they treat it like the AI identified the specific person they were looking for.
One example they used was a woman who was arrested. They arrested her despite the woman in the video not being pregnant and the woman they arrested being about 8 months along.