Is It Reasonable to Expect Facial Recognition to Identify Transgender and Non-Binary People?

Friday, 01 November 2019 - 3:06PM
Technology
Friday, 01 November 2019 - 3:06PM
Is It Reasonable to Expect Facial Recognition to Identify Transgender and Non-Binary People?
< >
Composite from Pixabay
On Tuesday, Forbes published an article by contributor Jesse Damiani titled "New Research Reveals Facial Recognition Software Misclassifies Transgender, Non-Binary People." The article discussed some of the potential social concerns surrounding research conducted at the University of Colorado Boulder's Information Science department's Identity Lab by doctoral candidate Morgan Klaus Scheuerman (lead author), assistant professor Jed Brubaker, and undergraduate student Jacob Paul.


The team found that while four of largest facial recognition service providers – IBM, Amazon, Microsoft and Clarifai – correctly identified cisgendered men (people born as men and who identify as the same) and women 97.6 and 98.3 percent of the time, they incorrectly identified trans-men (people who were born women but identify as men) as women up to 38% of the time (the average was 29.5% misclassification) with trans-women faring a little more accurately (the average was 22.7% misclassification, though this wasn't mentioned in either Forbes or CU Boulder Today, who also covered the story). Those who identified as agender, genderqueer or nonbinary (in other words, belonging to neither gender), were misidentified 100% of the time. The researchers used samples with self-identifying hashtags from Instagram to conduct their analysis. They concluded their research with some "design and policy considerations" for those working in the facial recognition sector of AI and machine learning.


When I read the Forbes piece, my initial reaction was, "well, of course they misidentified nonbinary people: these systems are created to classify images that convey physical data, not abstractions." 


I stand by my initial response, although since then, I've considered some of this more deeply, particularly around the way that media outlets report data and the way that computers "think." Let's unpack this a bit. 


It's important to note the way this data was reported outside of its publication in the Proceedings of the ACM on Human-Computer Interaction journal, including in the second paragraph here. Both Forbes and CU Boulder Today positioned the statistics in such a way that the numbers support a thesis of a "critical shortcoming" (Forbes) and "gender problem" (CU Boulder Today) inherent in facial recognition technology. Rather than reporting both the cisgender and transgender data in the same way, they reported the former's accuracy rate and the latter's inaccuracy rate. If you look at the table from the article by Scheuerman, Brubaker, and Paul, you'll note that the data presents an accuracy rate, from which the inaccuracy rate can be derived through simple arithmetic. 


Screenshot: Proc. ACM Hum.-Comput. Interact. 3


I'm not doubting the ability of either Forbes or CU Boulder Today editors to do simple math (but only because I checked the numbers myself), but I wonder why they didn't choose to say that transmen were correctly identified up to 76% of the time, while transwomen were correctly identified up to 94% of the time. That "up to" works in both directions. 


One of the ways that biological sexual differentiation occurs in humans is through certain outward physical traits: phenotypical expression driven partly by hormones. Without getting into the nuances of gender-as-a-construct, critical theory, the alignment (or misalignment) of identity with biology, and the intricacies of machine learning, computers are trained to recognize and classify things according to their physical qualities and by what their programmers tell them those qualities express. For facial recognition, those likely include some of the performative or presentative aspects of gender (I had really hoped to avoid this) like hair and makeup that have to be balanced with phenotypical and sexual differentiation markers, some of which can be altered through hormone therapy which affects gene expression, even after sexual maturation.


Going back to framing statistics: in examining the data table above, one could argue that, on average, 79% of transgendered people presenting themselves as the gender with which they identify are identified as the same by facial recognition technology.


That aside, all of this is just a long-winded way to say this: regardless of how people identify, computers are trained to "see" the physical aspects of biological sex. They cannot discern gender as an abstraction or a concept. They cannot examine a visual image of someone and tell you what they think or how they identify, especially if their identification negates the classification that the computer has been trained to recognize. 


This fact wasn't missed by the researchers. "These systems don't know any other language but male or female," Brubaker told CU Boulder Today, "so for many gender identities it is not possible for them to be correct." Given this understanding, one wonders why the machines were fed the data in the first place.


Nevertheless, like those who developed an ostensibly "genderless" AI voice, Brubaker sees this as a shortcoming that needs to be addressed, with the article citing the negative psychological impact suffered by those who do not find the same kind of social acceptance enjoyed by the cisnormative. "As our vision and our cultural understanding of what gender is has evolved," Brubaker told CU Boulder Today, "the algorithms driving our technological future have not. That's deeply problematic."


Although these concerns are real, some may question the value of the solutions that are implied, if not outright proposed, in asking technology to conform with critical theory (in this case through via the relatively nascent field of gender studies), or any social theory, for that matter. I could dispassionately argue either side of this: from a practical – and even a social – standpoint, there's certainly value in developing technology that reflects human diversity. On the other hand, I have concerns about how this could lead to unexpected consequences – particularly as they relate to privacy – and set a precedent that may be exploited at a later date. I also tend to avoid things championed by people – including academics, journalists, politicians, casino workers, and carnies – who advance their causes by playing games with numbers or by conflating academia with the real world. 


You can classify that however you want: it has no bearing on me whatsoever.
Science
Artificial Intelligence
Technology