Facial recognition can be an incredibly useful tool that helps you accomplish a number of tasks a lot quicker. For example, a great facial recognition setup on your smartphone can helps you authenticate into your device without having to type in an annoying PIN or press your thumb on a fingerprint scanner—which was already quite speedy. It can remind you to tag friends’ faces in photos on social media or, if the service is especially clever, tag their faces on your behalf. You might even be able to search for a friend’s name on your online photo archive and find all the pictures with them in it.
However, facial recognition is only as useful as it is accurate, and a new report from The New York Times indicates that today’s artificial intelligence systems are still struggling to identify faces of different genders and colors. A new study from the M.I.T. Media Lab, authored by Joy Buolamwini and Microsoft’s Timnit Gebru, shows that it’s a lot easier for systems to identify light-skinned males than darker-skinned women—incredibly difficult to get it wrong for the former, and surprisingly error-filled for the latter.
The study relied on a data set of 1,270 faces, which was sourced from three African and three Nordic countries.
“The specific African and European countries were selected based on their ranking for gender parity as assessed by the Inter Parliamentary Union. Of all the countries in the world, Rwanda has the highest proportion of women in parliament,” reads Buolamwini and Gebru’s paper, “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification”
“Nordic countries were also well represented in the top 10 nations. Given the gender parity and prevalence of lighter skin in the region, Iceland, Finland, and Sweden were chosen. To balance for darker skin, the next two highest-ranking African nations, Senegal and South Africa, were also added.”
Each face was then assigned a rating for skin type based on the six-point Fitzpatrick rating system, which dermatologists use as “the gold standard” for classifying different shades of skin, the paper notes. The researchers then tested three different companies’ facial recognition technologies to see how well they identified gender in each face photograph: Microsoft, IBM, and Megvii.
The results? The systems had an easy time with lighter-skinned men, only misidentifying gender in up to one percent of all photos. The systems were slightly less-accurate for lighter-skinned women, offering up an incorrect assessment in up to seven percent of photos. The systems struggled with darker-skinned men more, with errors in up to 12 percent of all photos, and had a terrible time with darker-skinned women, misidentifying gender in up to 35 percent of all photos.
“The substantial disparities in the accuracy of classifying darker females, lighter females, darker males, and lighter males in gender classification systems require urgent attention if commercial companies are to build genuinely fair, transparent and accountable facial analysis algorithms,” the paper reads.