>"Facial recognition" means an automated or semiautomated process that assists in identifying, or verifying the identity of, an individual based on the physical characteristics of an individual's face.
So I have a still from security camera footage. And I have 50,000 mug shots stored in the mug shot data base. I can see that it's a man in the image, he's white, between the ages of 15 - 50. So I go to my computer and put in the filter for the mug shot database to include those criteria. Sounds to me like I just used a semi-automated method to attempt to identify the person in the still image.
This ordinance is dumb. The use of a computer to pattern match, to automate what a person can do, is a lever that makes everyone more efficient. What's the logic in making someone thumb, or click through, every photo in the state's mug shot collection instead of allowing a system to reduce the possible candidates, if any.
But if you didn't record or have access to those fields, and you tried to infer age, gender and race from the photo, and then filter -- well, what if your system is much more accurate guessing ages for some racial groups than others, or more accurate in distinguishing between white and latinx than between latinx and asian? Suddenly your "filtering" process can also have a bias problem, where some people are more likely than others to be included in the results for a query which shouldn't actually include them.
A case agent looks at the still photo from security footage, concludes the person is male and white. They put that information into the mug shot database and it returns the subset of all mugshots that are male and white. They have just used a semi automated system to assist to ascertain the identity of the person in the security footage based on the physical characteristics of their face. This is forbidden by the ordinance.
>what if your system is much more accurate guessing ages for some racial groups than others, or more accurate in distinguishing between white and latinx than between latinx and asian? Suddenly your "filtering" process can also have a bias problem, where some people are more likely than others to be included in the results for a query which shouldn't actually include them.
What if it isn't? Doesn't matter. It's a tool to attempt to match two photos. If a tool is not as effective due to the way light and different skin tones interact, it just limits the leverage factor of the tool. It doesn't make it wrong and it isn't an infringement on people's rights. We use DNA matching even though it can't distinguish twins and sometimes it comes back and says it's a close match for Bob which leads them to suspect Bob's brothers family, e.g. Golden State Killer. I fail to see why tools have to be perfect or not be useful.
"The satnav directions say to turn onto this pier."
"The model says this house is worth 200k less than the ones in the whiter neighborhood a quarter mile away."
"The system recommends the patient have 40% allocated home health care hours per week going forward."
"The algorithm says this grainy footage of someone far away from the camera in the dark and the rain is X."
If you put humans in the loop informed by an algorithm making judgements, the human will often defer to the algorithm. Does the algorithm give an output that indicates its uncertainty? Based on what model? How equipped is the human to consider that?