• by Kaila Colbin , Featured Contributor, June 18, 2021

You may have seen them in your Twitter feed last year: tweets with two of what looked like the same picture of a white man, side by side, and a caption that read something like, “Let’s see if this works.” Click on the photos, and the images would be revealed to feature two people each: the white man, and a Black man.

The point of the posts? No matter how the image was composed, Twitter’s algorithm -- designed to look for faces -- would crop to the white man’s face. Didn’t matter who was on the left or who was on the right; the Black man wouldn’t get a look in.

It wasn’t just humans, either. It happened with cartoon characters, too. And dogs.

At the time, Twitter apologized and promised to do better. This week I learned how it fixed the problem.

To be clear, I wasn’t trying to research this issue. I just saw a tweet about superhuman levitating wondergymnast Simone Biles and her Glamour magazine photo shoot: four images, none of them showing her face. So I took a screen grab of it and posted it, saying, “C’mon Twitter algorithm, show me this goddess’ face in the preview.”

To which Rumman Chowdhury, who took on the role of director of machine learning ethics at Twitter in February, replied, letting me know that “non standard aspect ratio images are now center cropped bc we removed the algorithm.”

That’s an immediate fix, I suppose, in much the same way Google stopped its image recognition software from classifying Black people as “gorillas” by blocking the algorithm from identifying gorillas altogether.

It’s fantastic that Twitter has hired someone like Chowdhury to tackle these problems. And I did find it interesting to note that, when I returned to the original tweet about Biles, the images were now centered on her face.

But it highlights an issue that is at the fore of any conversation about artificial intelligence: an AI will never stop to ask itself “Is this racist?” It will never wonder whether it is embedding systemic bias, or reinforcing historic disparities. It will only optimize for what we instruct it to.

If we tell it to optimize for faces, but we only train it to recognize white faces, it will do that.

If we tell it to optimize for engaging video content, it will do that -- even if, as was the case at YouTube, the most engaging content is false, incendiary or toxic.

We could try what Airbnb did in 2015: create a “Smart Pricing” algorithm to do all the work of figuring out how much to charge, theoretically making it so that all hosts would make more money, no matter the color of their skin.

Except what actually happened is that white hosts adopted Smart Pricing at a higher rate than Black hosts -- leading to a 20% increase in the earnings gap between Black and white Airbnb hosts in the US.

Algorithms are like little kids. They take everything literally, so you have to be extremely careful with the instructions you provide.

An algorithm will never stop to ask itself whether the outcomes it’s producing is racist. Like a little kid, it has no idea what “racist” means. It only knows what it was instructed to deliver.

If we want algorithms that reduce racial disparities and eliminate historic biases, we’re going to have to get a lot better at babysitting them.