‘Face recognition algorithms distinguish between black and white people’

Most face recognition algorithms have more difficulty recognizing black and Asian people than white faces. This is evident from a study of 189 algorithms from large companies. Also, there are more different results between women than men.

The National Institute of Standards and Technology study looked at 189 algorithms from 99 different developers. This concerns systems from large companies, such as Microsoft and Intel. The study mainly looked at empirical results, but does not draw any conclusions about the causes of the algorithms’ biases. The study looked at matching a face with one specific photo, such as when using a biometric facial scanner on a telephone, and at matching a photo from a large collection of photos, as is often done during criminal investigations. The researchers note that the differences between black and white people are particularly large in the first group. When matching one face with multiple faces, the results differ. “Not all algorithms have such a high false positive rate. Different algorithms yield different results,” they write.

The research shows that black and Asian men are in some cases a hundred times more likely to be incorrectly identified as a photo match than white men. Other ethnic groups, such as Native Americans and people with a Polynesian appearance, are also highlighted in the algorithms significantly more often than white people. The algorithms also distinguish between age and gender. Women are more often unfairly singled out, as are the elderly and children. According to the researchers, this is certainly a danger with detection algorithms. “In a one-to-one comparison, a false negative is just an inconvenience that you can’t get into your phone, for example. But in a false positive that compares with other images, an incorrect match can lead to further investigation,” the researchers write.

There has been a lot of talk about facial recognition lately, especially by the police and other investigative services. Several politicians from CDA and D66, for example, want this to be temporarily stopped, something that is also endorsed by Minister Grapperhaus of Justice and Security and the European Commission. This would involve the use of facial recognition to detect criminals, but also by tech companies that build it into their products and services.