They didn't say "they all look alike", they said that white faces have more contrast so are easier to differentiate even with a naïve CV algorithm. It's one more source of systemic bias in the ML literature, especially given the comparative scarcity of source data from non-majority groups.
> It's one more source of systemic bias in the ML literature, especially given the comparative scarcity of source data from non-majority groups.
centimeter's posts in that thread directly reject your proffered line of thought:
> your training data does not have enough people with dark skin or African American face features
This isn’t the issue - the issue is lower variance across black faces in any basis.
Here's what I said: "if you partitioned the faces by race the output of the SVD would be much wider for white people [...] white people have more light/dark contrast, more hair colors, more eye colors, etc.". This is an obvious fact that is widely recognized by CV practitioners.