You will get false positives with eyewitness identification as well. But if you only use one identification method you'd never know if another would disagree.

> But if you only use one identification method

I'd like to emphasize that this is arguably the same method (facial similarity) being attempted twice, rather than two different and independent methods. While a marginal improvement, each attempt could go wrong for shared or similar reasons--for example, facial-recognition AI can also suffer from the "all those minorities look alike" problem.

Consider the contrast between:

1. "A human who saw the culprit's face believes your face matches, and a computer looking at video-footage agreed."

2. "A human who saw the culprit's face believes yours matches, and your car is the same color and model as the car the culprit drove off in."

I would be very curious which of the above options is more accurate, I don't think it is open-and-shut that the 2nd would be.

Obviously computer + car make would be the best combo?

Wait, having a computer saying that your face matches a video does not prove that you are the person on the video. Sure, the police can talk to you, but that is not a sufficient piece of evidence.

people have been convicted and sentenced to decades in prison on the basis of less evidence. see Anthony J. Broadwater, for instance

Okay, in that case, the victim (as sole witness) said that she recognized the defendant as the perpetuator. I am not a lawyer, but not sure if nowadays you could go to jail just because a victim (and sole witness) recognize you ? That seems rather slim evidence, and could be weaponized.

This would be a fantastic circumstance to absolutely not speak to any law enforcement personnel, by the way.

good thing eyewitness identification is very accurate, it'd be a shame if the status quo approach caused more false arrests than the alternative

False positives are the main problem here, exactly. We'd all feel a lot better about facial recognition technology with police reform that ended/compensated the widespread harm being done to mere suspects, as well as fixing courtroom standards of evidence to take into account that things like facial recognition are essentially pulling signal out of noise. If facial recognition turns police onto a suspect who is then found to possess the stolen items, the latter is corroborating evidence that they've found the right person. If facial recognition turns police onto a suspect who then merely looks like what eyewitnesses remember, that part of the witness's recollection is inherently correlated to how the suspect was picked and so it needs to be disregarded.

seems likely that face recognition technology and cameras could reduce the frequency of police-civ interactions