A.I. Has A Race Problem

Facial recognition software still gets confused by darker skin tones.

- Lizette Chapman And Joshua Brustein

A couple of years ago, as Brian Brackeen was preparing to pitch his facial recognition software to a potential customer as a convenient, secure alternative to passwords, the software stopped working. Panicked, he tried adjusting the room’s lighting, then the Wi-Fi connection, before he realized the problem was his face. Brackeen is black, but like most facial recognition developers, he’d trained his algorithms with a set of mostly white faces. He got a white, blond colleague to pose for the demo, and they closed the deal. It was a Pyrrhic victory, he says: “It was like having your own child not recognize you.”

At Kairos AR Inc., his 40–person facial recognition company in Miami, Brackeen says he’s improved the software by adding more black and brown faces to his image sets, but the results are still imperfect. For years the same problem has bedeviled companies including Microsoft, IBM, and Amazon and their growing range of customers for similar services. Facial recognition is being used to help India’s government find missing children, and British news outlets spot celebrities at royal weddings. More controversially, it’s being used in a growing number of contexts by law enforcement agencies, which are often less than forthcoming about what they’re using it for and whether they’re doing enough about potential pitfalls. Brackeen believes the problem of racial bias is serious enough that law enforcement shouldn’t use facial recognition at all.

Microsoft, IBM, and China’s Face++ misidentified darker-skinned women as often as 35 percent of the time and darker-skinned men 12 percent of the time, according to a report published by MIT researchers earlier this year. Such software can see only what it’s taught to see, which has been mostly white men.