Comment Re:alogrithms aren't racist (Score 4, Informative) 352
It's also one thing if this was a program just designed to distinguish between different people. But it looks like it's trying to recognize objects of all sorts and distinguish between people and just about everything else. That's a hard problem, and the only response to this sort o thing is to take a regular failure case and feed it back into the training data so you can hit the next regular failure case. Hopefully it will be less coincidentally embarrassing, but it will definitely be there. Perhaps confusing bald men with balloons or something like that.
But I also think people underestimate how much skin color affects machine vision problems. I spent years in the biometrics industry and one consistent fact is that people with darker skin just don't provide as much easy-to-recognize detail as people with lighter skin. There will be more misclassifications as long as the image is taken using the visible spectrum. To a computer extracting features, dark skinned people and gorillas are both human-ish face shapes with a particular color range and somewhat indistinct geometry due to weak contrast and shadows. Distingushing between those two sets just isn't as easy as distinguishing between fair-skinned blondes and gorillas. You can make that decision just by looking at the color histograms and not even bothering with geometry.