Read on Twitter
Twitter is abuzz with examples of its seemingly biased image cropping algorithm. Here's a thread of the most interesting examples and what they mean.
It started with this comparison where Twitter shows McConnell but not Obama (click the photos to see the images I uploaded).
It started with this comparison where Twitter shows McConnell but not Obama (click the photos to see the images I uploaded).
First, if you didn't see it, here is a brief thread with Zoom cropping out a black man's head but not a white man's. Harder to find more examples of this, but interesting. https://twitter.com/colinmadland/status/1307111818981146626
Some people have taken the images of white men and simply darkened the skin tone (or vice versa), with mixed results. For example, if you lighten the Obama photo and compare to the original Obama photo, Twitter selects the lightened version. https://twitter.com/ConstLarionoff/status/1307638069743542274
Obviously the big challenge here is that photos vary in much more than skin tone and facial structure. For example, switching the glasses in the McConnell-Obama comparison leads Twitter to select Obama, perhaps because glasses very clearly indicate a face. https://twitter.com/SergioSemJ/status/1307493041742254080
Similarly, smiles seem to have a big impact. These two images are McConnell scowling (I think that's what he's doing?) and Obama smiling with his eyes closed. https://twitter.com/MASCARPOWN/status/1307454405059452928
You can also change the contrast, lighting, etc and get Twitter to show Obama instead of McConnell. https://twitter.com/KonProg/status/1307588926731943936
This seems to be the closest thing to a serious experiment so far with 92 images of white and black men with no smiles, no glasses, and white backgrounds. It seems to find that the images of black men are favored at a ratio of 52:40. https://twitter.com/vinayprabhu/status/1307486263688085504
A couple folks at Twitter have responded, indicating that (i) it's not directly based on face detection/recognition, just on saliency or gaze prediction, (ii) they found no significant bias in their pre-release tests. https://twitter.com/ZehanWang/status/1307461285811032066
A different Twitter representative said, "It’s 100% our fault," though it's unclear whether @Dantley knew of the testing. Perhaps they were just reacting to the initial outcry based on the few examples. https://twitter.com/dantley/status/1307432466441859072
