In 2018, Lauren Rhue showed that two leading emotion detection software products had a racial bias against Black Men: Face++ thought they were more angry, and Microsoft AI thought they were more contemptuous.
Rhue used a set of player profile pictures from the NBA. These pictures mostly have standard poses and angles, and these players are relatively homogeneous in many dimensions.
Both Face++ and Microsoft AI create separate scores for how much a person is smiling and what emotions the face shows. If Black and White players share their smile, yet Black players are assigned more negative emotions, then this is evidence of racial bias. See for example this picture:
![A Black and a White basketball player, both smiling](https://racismandtechnology.center/wp-content/uploads/darren-collison-and-gordon-hayward-1024x757.png)
Rhue writes:
According to Face++, Darren Collison and Gordon Hayward have similar smile scores of 48.7 and 48.1 respectively. Translated into emotion, Face++ rates Hayward’s expression as 59.7% happy and 0.13% angry yet rates Collison’s expression as 39.2% happy and 27% angry. Collison is rated as more than 180x angrier than Hayward despite his smile!
You can read the full article (as a PDF) here.
This is yet another example of why you should stay clear of any software that detects emotions on faces. A recent overview article in the Guardian writes about the debate between scientists about whether it is possible to infer emotions from facial movements alone. The consensus is pretty clear: we can’t.
See: Racial Influence on Automated Perceptions of Emotions at SSRN, and Are you 80% angry and 2% sad? Why ‘emotional AI’ is fraught with problems at The Guardian.
Header image by Comuzi / © BBC / Better Images of AI / Mirror D / Licenced by CC-BY 4.0