If it was science those providing this service to law enforcement should publish the data they use to train the neural network (a kind of AI) with, they would publish the network architecture and they would publish the error bars in their results. My guess is they are too ashamed or embarrassed to admit the likely error bars. But more likely, they just do not know themselves what the error bars really are.
For a start, neural networks produce a number between 0.0 and 1.0, with a range of computer based on floating point numbers in between. My understanding is (because when I've asked I got no answer) that any number greater than or equal .5 is a hit anything less is a miss.
Let me put this in perspective, if the network produces .500000 then it's a hit, .49999999 is a miss. And if you don't know how computers handle floating point numbers you'll love to find out it pretty much rounds, so .4999 is very likely to become a .5 which will become a hit. So if you were at work surrounded by 50 of your work mates and the facial recognition software came up with a .4999 for killing someone's grandpa, too bad, it's prison time for you.
Unfortunately that .4999 is not statistical, it is based on the error the neural network has in it's comparison with the (usually) random numbers it's weights were set to before training began, and the input values (images) it was trained on. I have heard people gushing about how well 'deep' neural networks differentiate between images. Perhaps, but I think they are giving more credit to the network rather than actually understanding what the difference is that the network is doing.
I don't know how long the network is trained on each image, or even if there is a single output or even multiple networks. And this were science comes in, a company offering this tech must publish this information, because science needs to meet the reproducibility requirement. Company's cannot hide behind in-house IP secrets. Because the tech is in the public domain on how to do this in broad brush strokes.
Because that .4999 may refer to the fact that the network just does not know, it is insufficiently trained, or just another random number.
The results of lie detector tests are not admissible because they can be fooled, the results of facial recognition is even less credible. But try telling the punitive right that this tech is a complete failure when there is so much yellow journalism about AI.