Twitter says it will look into its auto-cropping algorithm after it showed racial bias against black faces. Twitter uses a neural network to automatically crop photos so they don't take up as much space on a user's feeds, but it appears to focus more on white faces than black faces.
- The issue was highlighted in cryptographic engineer Tony Arcieri's recent experiment, where he demonstrated how the algorithm removed Barack Obama's face in a photo preview that users see as they scroll through feeds.
- The original photo featured both Obama and Mitch McConnell. In this case, the AI honed in on McConnell's face and removed Obama's in the preview.
- Obama's photo apparently did show when Arcieri inverted the colors. Intertheory’s Kim Sherrell also found that a higher-contrast smile was effective.
- A similar scenario happened with “The Simpsons” characters Lenny and Carl. In the preview, Twitter's algorithm removed Carl, who is black, and only showed Lenny, who is white.
- Twitter’s chief design officer Dantley Davis said it's now "on the team's mind" to stop cropping photos. The model was initially analyzed but requires "continuous improvement," Twitter CTO Parag Agrawal noted.
- Twitter spokeswoman Liz Kelley said the company tested the model for bias before using it, but it's become clear that more analysis is needed