So, at TMC (Twitter Math Camp) I had the privilege of attending two separate talks that got me wondering. First was Grace Chen’s absolutely phenomenal Keynote, “The Politics(?) of Mathematics Teaching”. In it, she mentioned that the age of black children is often overestimated by 5 years.
Later, Bob Lochel shared the hilarious website how-old.net Upload/take a picture and it will estimate the age of people in the photo. The whole room had a great time playing with this. I personally (actually 32 years old) received everything from 26-42 as an age, depending on the picture.
Recalling Grace’s talk, I thought I might investigate. My boyfriend’s niece is Haitian and has very dark skin. She’s absurdly adorable and 5 years old. My nephew is white and has very light skin. He is ALSO absurdly adorable and 5 years old.
Sure enough, how-old.net did a good job with my nephew and greatly overestimated the age of my niece.
This matters. Although it’s a silly website, by overestimating the age of black children, and seeing white children as more innocent, we perpetuate racism. I can’t say it better than Robin Bernstein did in his article, so go read that. (Pointed out to me, naturally, by Grace Chen.)
Knowing I needed more data points, I started entering a ton of photos and recording the information. If you would like to help out, head over to how-old.net and add the data to this quick google form! I’ll update the graph. Below are the current results.
On the x-axis are the people’s actual ages in the photos. On the y-axis is the age shown on how-old.net. The diagonal line shows x=y (or if how-old.net got it right). It’s color coded by race. What do you notice? What do you wonder? Anything else we should investigate? (Add in comments!)
For what it’s worth, Emma Veach also happened to find this gem on the same day!
2 thoughts on “Is the algorithm for How-Old.net Racially Biased?”
Fascinating. It sure looks like it’s biased!
I wondered why. It looks like the ‘how-old’ site isn’t using traditional algorithms, but is rather using machine learning to assign an age. Now, I don’t actually understand machine learning. What I understand is that it’s hugely correlational, and that it depends on training on an enormous set of data.
That makes me wonder if the source of bias in the algorithm is because of a bias in the sample that it’s trained on. If most of the pictures it’s trained on are pictures of people with lighter skin then maybe it never gets good at other skin hues? I don’t know how much it depends on people coming to their site and giving them a picture and their ages for data.
This is very interesting. I’m excited to read more about what you find!
There’s a great TED talk to pair with this by Joy Buolamwini… she’s a black researcher with dark skin and the AI only recognizes her face when she’s wearing a white phantom of the opera type mask. Fits well with this. Great post!