AI models still racist, even with more balanced training
theregister.com
theregister.com
It took a complaint by furniture and chocolate makers that the photos of their products were shit before Kodak fixed it.
So were the chemists racist ???.
Obviously not but their subconscious bias of being white slipped in when designing it - it is the same as any algorithm or technology designed by a non diverse team.
https://99percentinvisible.org/episode/shirley-cards/
> Each Kodak printer needed to be calibrated and standardized before photos were printed on it. And so the printers came with something called a “Shirley Card,” which was a color reference card created by Kodak in the 1950s.
Shirley was a white woman, and so the cameras used to be calibrated to white skin tones, and performed poorly with darker skin tones.
Or they didn't have dark skinned people around them to test it on... and didn't think about why it would not work on darker skins.
By saying "subconscious bias" you're implying that deep down they didn't want it to work for dark skinned people.
They did not think of it even though the US has a significant population of African Americans - they moved in social circles or lived in areas where black people simply did not feature in their life.
Out of sight - out of mind.
Why would you optimise your product for people who can't afford it?
> The typical black family had almost no wealth in 1968 ($2,467; data refer to 1963). Today, that figure is about six times larger ($17,409), but it is still not that far from zero when you consider that families typically draw on their wealth for larger expenses, such as meeting basic needs over the course of retirement, paying for their children’s college education, putting a down payment on a house, or coping with a job loss or medical crisis.
> Over the same period, the wealth of the typical white family almost tripled, from a much higher initial level. In 2016, the median African American family had only 10.2 percent of the wealth of the median white family ($17,409 versus $171,000).8
They don't need to actively discriminate, that was my point, the invisible hand does it, just like it turns their selfish desire for money into goods and services for the rich.
Though, ideally if you could make the fish addictive, and if it kills them then only slowly and in a way that needs statistics to be tied back to your product, that's the ideal.
To answer your question, though, you would want your camera to work for dark skin in case your white customers want to photograph an African American. I see no reason why that should be disallowed even in a racist society. But again, that was an oversight as they didn't even consider that anyone might want to do that
I usually use 'incompetence' or 'ignorance' rather than 'stupidity'.
The fact they join the legions of mentally ill and toxic ex-Googlers and Twitter people doesn't say much.
The Register is irreverent, don't read them literally.
On the other hand I'd be happy calling, laws, customs, organizations, religious beliefs, and various other things and conclusions drawn from statistics 'racist' without them being living, thinking beings.
First, although "AI exhibits racial bias due to biased training data" is far more accurate, I think it's perfectly acceptable to condense that to "AI is racist". Especially in the headline of an article that goes on to explain the issue in detail.
Second, I would say that even racist humans are racist because of bad training data, so if we're fine calling people racist, why not AI?