>images of a person of interest from security cameras or public photos uploaded onto the internet could be compared against a national repository of images held by the FBI. An algorithm would perform an automatic search and return a list of potential hits for an officer to sort through and use as possible leads for an investigation.
I feel like this is going to be the prosecutor's fallacy on steroids. Not only do you have the "cold hit" problem of DNA testing (but with way, way more false positives), but you'll end up with a defendant with near 100% odds of positive identification by the victim. Of course "looks like the perpetrator" and "positively identified by victim" almost entirely overlap as evidence, but it won't necessarily feel that way to a jury.
Are false positives really such a problem? Most repeat criminals do it because they are dumb and/or impulsive. If you get 10 hits, you just send a cop around to ask each one if they did it. The one that says yes gets arrested. A major fraction of crimes can be solved this way. Many more can be solved by slapping surveillance robots on people and catching their next crime.
Many cases will be like that yes. And if, as they say, the databases are limited to wanted persons or convicted felons, then the priors won't be terribly out of whack. But it's going to be way easier, PR-wise, to broaden a face database than it ever would have been to broaden a DNA database. It did not take long in the UK, for example, for the National DNA database to creep from "only convicted criminals" to "anyone arrested, even if they were never charged".
And that's with people's DNA, which is pretty invasive to obtain. The government takes routinely takes pictures of people; there's no visceral moment of privacy invasion there. It would be easy for the database to expand to include a tremendous number of people, and then your false positive rate really is a problem. It stops even being a reliable way to narrow people down, because there just isn't that much variance in human faces, especially if you have to deal with low resolution and/or noisy data.
I feel like this is going to be the prosecutor's fallacy on steroids. Not only do you have the "cold hit" problem of DNA testing (but with way, way more false positives), but you'll end up with a defendant with near 100% odds of positive identification by the victim. Of course "looks like the perpetrator" and "positively identified by victim" almost entirely overlap as evidence, but it won't necessarily feel that way to a jury.