Twitter Algo Test

We're testing to see if the twitter algorithm detects Black faces as easily as White ones by testing the image preview. Highlighting machine learning biases. In this tweet, the White man in both pic variants was the preview. https://twitter.com/NotAFile/status/1307337294249103361?s=20

Hi everyone telling me what you see: we all see the same thing no need to show me.

Twitter Algo test round 2: Carson and Graham in near identical shots

It could be about skin color, I dont put it past Silicon Valley at all. But I also think its about color variety. The machine learning is looking for color contrast to determine whats the photo to focus on. Dark skinned people in dark clothing could be overlooked

Considering white people turn all sorts of colors depending on emotional states machine learning might be biased towards white people

Lindsey Graham with a very racist and very bad tan vs normal mayo Lindsey Graham. Same shot. Same framing.

Huh. Thats interesting.

Yeah I think its the color contrast that the twitter algo is picking up the most.

Which would mostly hurt dark skinned individuals because since more often than not dark skinned people have dark hair and unless they're in bright environments or wearing bright clothes would be more likely to be overlooked by the ML.

I'd like to see these results in practice.

Ok final test: Using 4 solid contrast color vs 4 blending colors.

The twitter algo should show red, blue, yellow, green because they contrast most.

Now we know how the ML works. It's probably blurring the images and just examining the colors.

Now in practice, if machine learning is looking for contrast, then a dark-skinned person with dark hair won't be picked up like a light skinned person with a dark shirt or background

This was fun! Don't usually touch machine learning, thats @beenwrekt's territory

Ok just one more test. A Black woman and a White woman are lying in the same bed together with the same lighting in this picture.

What does Twitter's Machine Learning focus on:

Thats some troubling but predictable results

h/t to @NotAFile who did these first and inspired me to fuck around with this in gimp for an hour

Look a lotta people keep QTing the very first test to prove it's not racist. The tests with natrual lighting are more relevant to people.

If a Black face is only chosen in studio light backgrounds then that means 99% of dark skinned faces in real life wont be chosen.

Original photo for reference. Natural lightening, same picture.

Testing solid colors to see if the algo is actually just attracted to pixels closest to white as opposed to contrast.

Two White/Black Interracial Couples in Natural lighting. Black individuals chosen because they're wearing lighter clothing which was rather hard as most images were of people wearing dark clothing.