The Digital Closet, page 10
Figure 2.1
Lenna’s Playboy centerfold scan by SIPI. Retrieved from https://en.wikipedia.org/wiki/File:Lenna_(test_image).png.
In 1997, Sunny Bains, a prominent scientist, tech journalist, and editor of engineering journals, wrote an op-ed for Electronic Engineering Times in which she argued that “the Lenna image grates because of its exclusivity. It’s not difficult to feel isolated when you’re a woman working in a male-dominated field. Seeing provocative images of women in learned journals can add to that feeling of non-inclusion.”5 This feeling has endured over time. In 2015, Maddie Zug published an op-ed in the Washington Post arguing that the use of the image in computer science curriculum led to sexual comments from male classmates and indicated a broader cultural problem that is at least partly responsible for the depressed numbers of women working in advanced computer science labs.6 In 2013, Deanna Needell and Rachel Ward published a paper in which they used an image of the Italian-American model Fabio in place of Lenna for image compression research in hopes of motivating their field to reconsider the use of Lenna.7 Jeff Seideman, an industry leader in image encoding, captured these critiques perfectly in his defense of the continued use of the Lenna image, telling the Atlantic in 2016 that “when you use a picture like that for so long, it’s not a person anymore; it’s just pixels.”8
The use of the Lenna image fits into a long series of literal objectifications of women that have been central to the development of technology, ranging from the metaphorical objectification of the original labor of women computer operators whose function was automated by increasingly sophisticated circuitry to the literal objectification of women ranging from Kodak’s use of “Shirley cards” to optimize their film and film processing technologies to the unauthorized use of Suzanne Vega’s voice to perfect the sound compression algorithms that led to the MP3.9 For nearly fifty years, Lenna has served as the benchmark of image-processing quality, shaping everything from the development of image compression formats like JPEG to the operations of smartphone cameras like Apple’s iPhone to the operations of image software like Google Images.
The omnipresent use of the Lenna image is indicative of the unvoiced heteronormativity that permeates Silicon Valley. It harkens back to an earlier détente in the war on pornography in which Playboy was allowed to publish objectifications of a particular variety of female bodies and increasingly granted public legitimacy, while such open representations of alternative forms of desire—for different shapes, sizes, anatomies, and colors of bodies, perhaps in different contexts, performing different erotic acts, and so on—were denied such legitimacy and public visibility. It is the assumption of banality, the presumption that such an image was by default uncontroversial, that belies its heteronormativity. As I will show throughout this chapter, this “original sin” can be taken as symbolic of the gender and sexuality-based biases that ground the research and development of new technologies, where similar assumptions of banality, of shared norms, and an expected lack of controversy lead to heteronormative hardware and software.
In particular, we’ll look at the history of Google’s attempts to automate the censorship of “adult” content via its SafeSearch algorithms and image recognition technologies and Facebook’s efforts to streamline the human review of content flagged as inappropriate and produce “human algorithms.” While this critique is in no way confined to Google or Facebook—and I intend it to speak to the broader discursive community of computer programmers and software engineers, for which I will use the shorthand “coders”—I will draw heavily on case studies from the two companies to demonstrate the practical effects of this permeation of heteronormativity. The chapter considers this implicit heteronormativity from three perspectives: (1) its permeation into the discursive community of coders themselves, (2) its subsequent permeation into the parameters of the algorithms and datasets that currently shape computer vision as a field, and (3) its ongoing maintenance by “human algorithms,” the people charged with performing the human labor of reviewing content flagged by the system for violating community standards. Across these three domains, we can see that heteronormative biases have a strong impact on the research, development, implementation, and everyday operation of content moderation algorithms.
The Heteronormativity of Coders
In her article “Going to Work in Mommy’s Basement,” Sarah Sharma draws on the common Silicon Valley trope of “beta” coders whose conditions of existence are founded upon taking advantage of the unrecognized and feminized labor of their mommies, a twenty-first-century twist on the devaluation and rendering invisible of feminized reproductive and affective labor. She asks, “What kind of work is done in this ‘coder’s cave’ of antisocial techbro culture? What kind of world gets programmed from a position of uncomplicated safety and abundance?”10 This best of all possible worlds for male coders is what Emily Chang calls a brotopia.11 In this brotopia, men who often identify as spurned lovers or borderline incels in their youth are finally recognized, courted by large tech companies, put in charge of cutting-edge start-ups, and through their power, prestige, and wealth can finally make up for lost time when it comes to sex. Sarah Banet-Weiser has described this as “toxic geek masculinity” and shown that it is not an isolated phenomenon but is instead undergirded by and connected to the broader cultural context of misogyny and heteronormativity online (examined in chapter 1).12
Toxic geeks understand themselves as being the victims of marginalization and alpha-male masculinity. Nathan Ensmenger has shown that the tech bros and toxic geeks referred to here are usually shaped by the historical injury of having been geeks, nerds, and socially awkward in their formative years.13 As Kristina Bell, Christopher Kampe, and Nicholas Taylor explain, they thus understand themselves through the stereotype of being “weak, easily bullied, and socially awkward males who lack social skills, athletic abilities, and physical attractiveness,” with their sole redeeming feature and claim to political, economic, and sexual agency being their “perceived [ . . . ] mastery over digital technologies.”14 Adrienne Shaw argues that because of this felt sense of victimhood, toxic geeks react hostilely to anyone who calls them out as being the perpetrators of abuses of power themselves. They seem totally incapable of recognizing their own privilege and in response receive feminist critiques as unwarranted attacks, even going so far as to define their identity as anti-feminist.15 They are thus doubly injured by women, first through sexual rejection and second by feminist critique and women seeking entry into the workplace at technology companies. As Banet-Weiser notes, “This assemblage of features—technological prowess, social awkwardness, and cognitive dissonance about privilege—yields a contradictory subjectivity. According to this frame, geek men have been injured by the world and, more importantly, by women. The aggressive and violent regulation and exclusion of women is a way to regain masculine capacity.”16
Sue Decker, former president of Yahoo, has used the metaphor of a fish being the last to discover water to describe the ubiquity of gender bias and heteronormative sexual harassment in Silicon Valley.17 This bears out in what little comprehensive survey data we have from tech companies. Take, for instance, the infamous “Elephant in the Valley” study from 2017, which surveyed women of various ages and ranks that worked in tech companies about their experiences with sexism in the workplace. The study found that 90 percent of women surveyed had experienced sexist behavior at company off-sites or at industry conferences. Further, 60 percent of them had received unwanted sexual advances; most reported these advances were not one-time instances but instead repeated overtures, and more than half came from a superior at their company. A majority of those who reported sexual harassment were dissatisfied with how the company handled their case, and many ended up signing nondisparagement agreements to keep them from going public with their stories. Nearly 40 percent of women who experienced sexual harassment declined to report it for fear it would stunt their career advancement.18
This harassment takes place in both the materialized utopias of tech campuses and after work at off-site company events and industry conferences. Tech campuses are built to accommodate frat-like behaviors and to offer all the comforts of “mommy’s basement.” Most of them offer unlimited free alcohol and games like table tennis and foosball. They regularly keep free high-end food within fifty yards of every employee at all times and offer free dinners for employees who stay after 5 p.m. They contain services on-site ranging from gyms to doctors to hairdressers to laundry to pet care. All of this takes place within open floor plans that make it notably difficult for employees to avoid coworkers who might harass them. In short, their designs skew toward the desires of young, single men. This is perhaps nowhere more visible than in Apple’s failure to include a daycare service in its new $5 billion Apple Park campus that opened in 2017.19 As Emily Chang has found, “Few employers offer stipends for child care, and even fewer provide on-site child care. Sure, you can bring your dog to work, but you are (mostly) on your own with your baby.”20
Silicon Valley tries to position itself as being on the cutting edge of both technological and sexual experimentation, with strong polyamorous communities and hookup culture buoyed by exclusive company sex parties hosted at private homes. As Chang has found, most of these events skew toward the fantasies of heterosexual men, as they are maintained with higher ratios of women (selected for their appearance) to encourage sexual encounters with tech bros and toxic geeks. While the Valley’s progressivism extends to threesomes, these are almost exclusively a man and two women, with gay and bisexual sex acts conspicuously absent from the scene and little pressure on men to engage in this sort of progressive experimentation. In explaining his peers’ behavior, Evan Williams, a cofounder of Twitter, has described polyamory as a “hack.”21 Thus, most of the rhetoric surrounding sex in the Valley is simply a convenient means for justifying the voracious and heteronormative sexual appetites of men who are finally able to get access to women’s bodies in the ways they dreamed of as deprived adolescents.22
The liberation of this “progressive” scene is exclusively male. Women who participate in sexual exploration lose credibility and respect. They also gain a reputation of being open to any and all future advances, anywhere, and at any time. However, not attending has similarly bad consequences, as it can severely limit women’s opportunities to network and advance their careers since work gets done at these sex parties.23 The women at these parties are also kept at arm’s length for fear that they might be “founder hounders,” the Silicon Valley neologism for gold diggers. The rhetoric surrounding founder hounders is frequently used to justify predatory behavior toward these women, as it presumes that they are similarly engaging in predatory behavior by trying to trap rich men and extract capital from them. In a chilling interview, Chang spoke to an anonymous tech company founder about the rampant use of drugs to “lubricate” sex parties and the potential advantage tech bros were taking of women. He replied that “on the contrary, it’s women who are taking advantage of him and his tribe, preying on them for their money.”24
A culture like this was able to emerge because women’s participation in the field significantly diminished leading up to the dot-com boom and tech’s resurgence after the dot-com collapse. This was a particularly notable turnaround when it came to the development of software, which was dominated by women for many decades.25 While in the early 1980s women were earning nearly 40 percent of all computer science degrees in the United States, that number decreased to closer to 20 percent by the time today’s platforms were emerging and has remained relatively stable since. At companies like Google and Facebook, from what numbers are publicly available, women account for between 30 and 35 percent of the workforce, but only around 20 percent of the technical jobs.26 This lack of representation is particularly acute in AI fields, where 80 percent of professors are men, as are 85 to 90 percent of the research staff at Google and Facebook.27 During their formative years, many such companies employed aptitude tests like the IBM Programmer Aptitude Test and the Cannon-Perry Test that were biased toward the selection of antisocial, combative, and hubristic coders who just so happened to also be predominantly male. These tests included “brain teasers” that asked applicants to make wild speculations on the spot backed by some form of logic and calculation, like asking applicants how many windows are in New York City. Google, for instance, did not stop using these sorts of brainteasers until 2013. Its longtime former head of human resources, Laszlo Bock, then admitted to the New York Times that “brainteasers are a complete waste of time. . . . They don’t predict anything.”28
While companies began to wake up to this problem in the 2010s, much of their culture, corporate policies, and technological infrastructures had already been determined by largely male coding and legal teams. Companies like Google espoused a commitment to hiring more women early on, but this commitment was often half-hearted, as the company’s organizational chart reads more like a soap opera script of interoffice affairs. CEO Eric Schmidt, cofounder Sergey Brin, and Andy Rubin, the lead technician who developed Android, all engaged in relationships with women at the company who were their subordinates, and longtime executive Amit Singhal was given a golden parachute after sexually harassing a woman.29 Despite this bad corporate behavior, the company did strive to implement fairer hiring practices. In 2008, Google established a secret hiring practice in which female applicants had their applications submitted to a second review committee called the “Revisit Committee” if the initial hiring committee found them unacceptable. The Revisit Committee was tasked with reviewing the applications of all potential diversity hires. Company policy stipulated that hiring committees remain silent about any interviews they conducted. Google also established a secret policy that all technical candidates’ committees contain at least one woman, a practice that put undue burden on women already at the company.30 This intense secrecy and the measures Google took to correct for bad hiring practices demonstrate a key antagonism within Silicon Valley that persists to this day: the antagonism between the myth of meritocracy and the use of hiring practices meant to combat unconscious bias.
Meritocracy may be the central myth around which Silicon Valley’s culture is constructed. The problem with this is that belief in meritocracy most often requires a belief that brilliance is innate, and research shows that these cultural biases lead gatekeepers like teachers and hiring committees to assume that (white) men are more likely to possess innate talent. One university study found that “the extent to which practitioners of a discipline believe that success depends on sheer brilliance is a strong predictor of women’s and African American’s representation in that discipline.”31 Another empirical study found that “when an organization is explicitly presented as meritocratic, individuals in managerial positions favor a male employee over an equally qualified female employee.”32 The problem with meritocracy is that it doesn’t recognize the cultural contexts within which “brilliance” is defined and emerges. In Silicon Valley, brilliance is defined in such a way that it privileges male coders, and the position of privilege from which male coders apply to jobs goes unrecognized in the application process. Even Michael Young, who brought the term into public discourse with his 1958 book The Rise of Meritocracy, recognized this problem.33 He concluded that meritocracy could produce a new social stratification and sense of moral exceptionalism based on who had access to elite education and social networks. Further, meritocracy is always impossible to implement because it first needs to be defined, and the definition of meritocracy is most frequently founded on preferences for certain qualities, aptitudes, demeanors, and skill sets that are primarily available to wealthy white men.
True believers in meritocracy don’t see these internal contradictions and instead use meritocracy as a logical explanation for the privilege that they enjoy. It gives them a smugness and overinflated sense of self-worth that can cause them to react violently to what they perceive as “discriminatory affirmative action” policies like the ones Google implemented to hire more women for technical positions. While incurring these violent backlashes may be worth it if diversity hiring actually leads to more equity in the workforce, this doesn’t seem to be the case as the number of women in technical positions at technology companies has remained rather stagnant despite the past decade of attempts at fairer hiring practices. Most companies now implement some equivalent of unconscious bias training where they offer employees workshops on how their unconscious biases about race and gender might impact their thinking in the workplace, a new and revised version of earlier attempts at “sensitivity training.” There is another problem, however, with how unconscious bias training actually plays out. In attempts to avoid shutting down dialogue by calling employees out on biased behavior, unconscious bias training begins with the premise that everyone has biases, that there is nothing wrong with having biases, and all one is responsible for is curbing them as much as possible. Studies have found that this essentially normalizes gender and racial bias by removing the cultural stigma around it. It can even cause people to accept these biases as unavoidable and make them more likely to exhibit these types of biases in the workplace.34 Even Anthony Greenwald, the inventor of the Implicit Association Test that helps demonstrate to people the unconscious biases they hold, has expressed concern about unconscious bias training. He told an interviewer, “Understanding implicit bias does not actually provide you with the tools to do something about it.”35
