Training Algorithms in Diversity

Is Facial Recognition Software Racist?

By Selena Anjur-Dietrich

Facial recognition software is becoming ubiquitous – from social media platforms to police surveillance and security practices. As pedestrians become subject to immediate facial recognition and matching with suspect databases, research cautions against naïve trust in the neutrality of technology.

Algorithms learn as babies do, through constant feedback loops between the environment and cognitive capacities. Individuals raised in a racially homogenous household or community develop recognition mechanisms fine-tuned to only one race. According to evolutionary psychologist Leda Cosmides, we evolved to track race as a proxy for coalitional alliances, the recognition of which was essential for survival in hunter-gatherer societies. Even beyond race, facial recognition is rife with potential for error, featuring own-age and own-gender biases.

Clearly, “natural” facial recognition is not a standard that developers of powerful and ubiquitous technology should aspire to. Yet, efficiency is still most often the winning concern; in a predominantly White society, an algorithm that is 97 percent effective is considered a success even if minorities are grossly overrepresented in the remaining three percent. Conversely, software developed in East Asian countries performs better on East Asian faces. Quality testing of facial recognition algorithms is currently entirely voluntary. Efficiency should not be a success criterion for a technology that will play a defining role in policing and global security, nor should it have been for social applications.

Research shows that humans can overcome lapses in facial recognition through exposure to more diverse face stimuli, and even more simply by redirecting attention to the lower half of the face — a feature that is less implicated in typical perceptions of race. Studies performed with newborn chimpanzees show that even own-species bias can be overcome by early exposure to diverse face stimuli. Software developers can and should follow suit to keep pace with growing demand for their work.


Selena Anjur-Dietrich is a senior in Ezra Stiles College. She can be contacted at .