In 2015, Google rolled out a new Photos app with image recognition, which horribly backfired when it labeled African-Americans as gorillas. The software wasn’t racist, but it clearly hadn’t been trained enough to recognize the range of human diversity.
Clarifai, which provides a cloud-based image-recognition service that plugs into apps and websites, says it’s working to prevent such gaffes. It’s introduced what it claims is a well-trained Demographics model to recognize gender, age, and what the company calls “multicultural appearance.” The terms “race” and “ethnicity” have too many negative connotations or generalizations, writes CEO Matthew Zeiler in a blog post. Terminology notwithstanding, multicultural appearance recognizes familiar identities including Asian; black or African-American; Hispanic, Latino, or Spanish origin; Middle Eastern or North African; and white. Meanwhile, “gender” is designed to recognize how people present themselves, not just their biological sex, says Zeiler.
Clarifai is offering its Demographics model in beta, and inviting people to make it better. Anyone can upload images, or URLs that point to images, to see the results it comes up with and provide feedback. I uploaded a silly photo that I use for my Twitter profile. It was dead right on my gender and very close on age (unfortunately) but only 38% sure that I’m white. So I clicked the Feedback button and checked the “white” option. It’s too early to know how well Clarifai’s new service will work, but it’s a welcome effort to right a troubling technical wrong in today’s AI.