505
In every reported case where police mistakenly arrested someone using facial recognition, that person has been Black
(www.businessinsider.com)
this post was submitted on 07 Aug 2023
505 points (97.7% liked)
Science
13222 readers
14 users here now
Subscribe to see new publications and popular science coverage of current research on your homepage
founded 5 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
It's probably the opposite. the AI was likely trained on a dataset of mostly white people, and thus more easily able to distinguish between white people.
It's a problem in ML that has been seen before, especially for companies based in the US where it is just easier to find a large amount of white people as opposed to people of other skin colors.
It's really not dissimilar to how people work either, humans are generally more able to distinguish between two people who are races that they grew up with. You'll make more mistakes when trying to identify people of races you aren't as familiar with too.
The problem is when the police use these tools as an authoritative matching algorithm.
It's not only growing up with them. We're just better identifying people/animals/things we're familiar with. Horses all look the same if you're not around them regularly. You can distinguish colours, but that's it.
Not comparing people to horses by the way...
except you quite literally are
I checked your name to see if you were a contrarian from another thread. You aren't.
Then, I thought about the name you chose. Did you mean to spell Dessert (the treat after a meal) or was that a misspelling? Then, I considered that regardless of the intent in spelling, your name appears to refer to a war (Desert Storm: USA vs Iraq in the 90s). Even if it's playful (Desserts Storming me, yum!), I dunno. At this point, I don't suspect we align in ideologies. I'll stop analyzing here.
I thought they would have trained it on mugshots. Either way, it should never be used to make direct arrests. I feel like it's best use would be something like an anonymous tip line that leads to investigation.
Using mugshots to train AI without consent feels illegal. Plus, it wouldn't even make a very good training set, as the AI would only be able to identify perfectly straight images shot in ideal lighting conditions.
Also makes me wonder if our defined digital color spaces being bad at representing darker shades contributes as well.