Hacker News Reader

Comment by devmunchies | Hacker News Reader

> negative outcomes for... people of color

Wouldn’t it do the opposite? Instead of racial profiling it’s using a system that pulls from social media. I wouldn’t say it’s a good system but it seems less biased.

EDIT: OP said “negative outcomes” and not “unfair towards”. Yeah I guess something can be more fair but still have disproportionate outcomes.

View on HN

> Wouldn’t it do the opposite? Instead of racial profiling it’s using a system that pulls from social media.

Oh there'll still be plenty racial profiling because the trained models will likely inherit these biases. Instead of a store clerk following me around, it'll be an automated system that just flags me as "high risk" and who knows what other secret feedback looks that results in.

I wish our leaders were smart enough to legally require that such services _must_ publish their accuracies and be open source.

A system like this as a for-profit business will go about as well as for-profit prisons have gone in the US.. except more subtle and sinister

Right, and then the discrimination is AI-washed - security personnel can say they're just following people flagged as "high risk" around the store, not discriminating, even though the "AI" mostly used skin color to add the high risk flag in the first place.

Discrimination: SOLVED!

>trained models will likely inherit these biases.

Living in middle America, I can tell you that Joe Average absolutely does not understand this. To him, based on my conversations with the people around me, AI/ML/anything sufficiently 'futuristic' sounding is without bias. To them, these models are developed in a vacuum, and using the AI label implies that the end result actually built itself, like a baby human would grow up.

And this scares me.

It does develop an idea in a vacuum. It finds the bias and uses the best biases to explain the world.

Facial recognition along with just cameras in general have historically struggled with the faces of non-caucasian individuals. This typically results from biases in training sets and in many cases cameras simply creating images with far more contrast for lighter skin tones.

It seems more likely a function of actual contrast in the real World than a function of bias (inadvertent or otherwise)?

I imagine it's down to skin tone, or do facial recognition systems really recognise dark skinned "Caucasians" better than light-skinned non-Caucasians.

Exactly: In the extreme, perfectly black objects have no visible contrast whatsoever as demonstrated by Vantablack.

https://www.wired.com/story/vantablack-anish-kapoor-stuart-s...

> faces of non-caucasian individuals

The Chinese government, operating the biggest surveillance state on the planet, have had no trouble with the faces of non-caucasian individuals, historically or otherwise. Their facial recognition systems are highly adept at identifying asian faces. They also supply half the world.

According to IHS Markit, China accounted for nearly half of the global facial recognition business in 2018.

https://www.ft.com/content/6f1a8f48-1813-11ea-9ee4-11f260415...

The real issue here is not Caucasian skin tones vs everyone else but instead light skin tones vs dark(er) skin tones. So Chinese government won't have an issue with current facial recognition tech.

Relatedly, photography itself developed an early bias against black people in terms of what colors/contrasts they optimized film for photographing:

https://www.vox.com/2015/9/18/9348821/photography-race-bias

That gets me thinking: in some parallel universe history of the Earth where interior lighting and photography was primarily developed by and for people of color, I bet average indoor lighting would be brighter, and Caucasian skin would often be washed out in CCV.

Right, because people of colour love having overly bright lights blasting into their eyes.

Lights aren't set for photography of any skin tone, they're set at a level for comfort and to provide enough light for us to operate.

That’s fair. I was thinking about dimly lit bars.

Reply on news.ycombinator.com