Operationalization of (Trans)gender in Facial Recognition Systems: From Binarism to Intersectionality

In a paper published in 2018, Os Keyes investigated how the literature on Automated Gender Recognition systems (AGRs) conceived gender, finding that 94.8% of the papers treated it as binary, 72.4% as immutable and 60.3% as a physiological component. In the author's view, this is indicative of an operationalization of gender, that is, the assumption that the latter is a discrete and objectively applicable parameter. Keyes claims that such a vision is blind to the performative aspects of gender and particularly dangerous for transgender people. Here I will follow on these remarks, providing several examples that show how AGR systems' failures in recognizing the faces of transgender people are capable of both perpetuating and amplifying gender stereotypes and inequalities. Then, I will introduce the notion of intersectionality, which is the idea that humans ‘sit at the crossroads’ of many physical, social, and political factors, whose combination generates dynamics of discrimination or privilege. I will focus on a subfield of intersectional studies, that is, intersectional stereotyping, which explains how we usually make assumptions and judgments about an individual or group of people based on multiple social identities or categories they belong to, such as their race, gender, sexual orientation, class, religion and ability. I will argue that this area of research provides us with a set of knowledge that might help us rethink and redesign the data sets for AGR. Specifically, I will draw on three key notions of intersectional stereotyping—‘perceiver goals’, ‘category accessibility’ and ‘category fit’—and use them to envision new ways of collecting images for assessing gender through facial recognition. Finally, I will explicate why my observations call for an urgent integration between computer science and gender studies.