Since the launch of ChatGPT in 2022, artificial intelligence has been touted as the undeniable future not only of how information is procured, processed and disseminated, but also generally of the world as we know it. Queerness too is not outside this so-called future’s ambit. As I write this, researchers are experimenting with how AI tools can aid LGBTQ+ rights advocacy. “The use of AI-generated virtual characters offers a unique opportunity to facilitate advocacy by engaging individuals in simulated conversations that can foster understanding, education and empathy,” MIT researchers argue in a 2024 paper testing how users respond to queer AI characters simulating conversations around coming out. Then, “Queer AI influencers” have begun cropping up on the internet, challenging human influencers to “bring more authenticity and better storytelling to the table to stay relevant.” All this, even as these AI influencers themselves sell a polished, calculated, and depoliticised brand of queerness. As University of Budapest sociologist Lilla Vicsek told Huffington Post last year, these influencers “won’t demand more rights, won’t get depressed and not post for a while.”
On the other side of the table, critics have pointed out the many issues with artificial intelligence vis-à-vis queerness. Some of these critiques revolve around how AI tools are developed and deployed, such as how these tools resuscitate and exaggerate existing biases, and can be used to violate queer people’s privacy. Other critiques, like that put forth by Chinese University of Hong Kong professor Nishant Shah, concern how queerness is constructed amidst the rise of AI. Yet other critiques dwell on the infrastructures of artificial intelligence and their fundamental incompatibility with queerness. For instance, university of Salford lecturer Daniella Gáti has pointed out in a recent article the “categorical logic” of the code that runs AI algorithms: how they construct digital worlds that “exclude fluidity and enforce strict boundaries between people and identifications,” in Gáti’s words.
The critics have also suggested ways to ameliorate these concerns. Inclusive training datasets, better ethics, and the involvement of queer people in development and deployment of AI tools are some oft repeated solutions. Other less common but significant suggestions involve calls for a fundamental restructuring of AI systems. Shah, for example, suggests that “queering AI” has to include an “ontological reworking of some of its computational and discursive practices and definition, intentions and ambitions.” How? Through “collective, fragmented, and promiscuous AI systems,” Shah adds.
I want to take this opportunity to introduce a different critique into the picture by pausing at the premise with which I began this article: that AI represents an advance, that it is an otherwise neutral instrument waiting to be directed towards better or worse ends. Most critiques of AI as well as the proposed solutions have one belief in common: that troubles of AI are largely those of calibration. And that imbuing this seemingly neutral technology with queerness will solve these troubles and herald, to borrow Shah’s words, “kinships and collectivities that contaminate the gentrified digital futures with joyful possibility”. I am unsure. Before we undertake this project of calibration, I wish to examine what it means for a digital technology to make a stride.
In making sense of this moment, I am importing to the AI context what historian of science and technology Kavita Philip wrote of the internet in “The Internet Will be Decolonized” (2021). In the 1990s, the internet’s story was one of transcendence, of a frictionless virtual space, free from the “meatspace” troubles of race, colonialism, and gender. AI’s story today is overwhelmingly similar: a neutral instrument that learns from the data, and in seeing and doing what humans cannot, precipitates futures that humans cannot.
But, what if AI’s purported futures are, in fact, reprisals of the past?
***
In a 2017 Journal of Personality and Social Psychology paper, Stanford computer scientist Yilun Wang and psychologist Michal Kosinski reported training a deep neural network (DNN) – a kind of artificial intelligence – to distinguish between facial images of gay and straight people. The DNN outperformed “human judges”, they claimed: given a single facial image the classifier could “correctly distinguish between gay and heterosexual men in 81% of cases, and in 74% of cases for women” compared with human judges’ 61% and 54% accuracy respectively. When given five images, the DNN’s accuracy reportedly climbed to 91% for men and 83% for women.
The duo saw this as evidence for the prenatal hormone theory. Proponents of this theory believe – incorrectly for the most part – that a person’s sexual orientation is determined by the exposure they have to hormones like testosterone, estrogen, and progesterone as a foetus. “Consistent with [the theory], gay men and women tended to have gender-atypical facial morphology, expression, and grooming styles,” Wang and Kosinski wrote in their paper. Here is what they meant, revealed a few pages later: as compared with heterosexual men, gay men should have “smaller jaws and chins, slimmer eyebrows, longer noses, and larger foreheads; the opposite should be true for lesbians.” The duo also chalked the humans’ failure to the “limits of human perception”. That is, according to them, a person’s sexual orientation is detectable through external features – just not by humans.
Amidst all this, there was one thing they refused to consider: that sexual orientation might not be, quite literally, written on one’s face.
Using a person’s external features to predict their behavioural traits and character is called physiognomy. Even though Wang and Kosinski acknowledged it as a “mix of superstition and racism disguised as science,” they went on to justify their project as a taboo-breaking endeavour. Three researchers from Google and Princeton University, Blaise Agüera y Arcas, Margaret Mitchell, and Alexander Todorov, called similar endeavours “Physiognomy’s New Clothes”. In a separate article, they challenged Wang and Kosinski’s work as scientifically unsound, one that confounds cultural differences as biological ones.
“Like computers or the internal combustion engine, AI is a general-purpose technology that can be used to automate a great many tasks, including ones that should not be undertaken in the first place,” Agüera y Arcas, Mitchell, and Todorov, wrote.
***
The attempt to read sexuality from biological artifacts has a long scientific history –phrenology, craniology, and even the genetic attempts of the twentieth century. Despite its repeated failure, its appeal is little dampened, as demonstrated by the Wang and Kosinski saga. What AI adds to this tradition of biological essentialism is another figure of authority. Historically, the power over people’s sexuality has rested with, among others, the state, religion, and the medical and scientific establishment. To this, AI adds the authority of computational scale.
Computational authority is particularly difficult to resist because it conceals its own exercise of power under the garb of technological neutrality. Consider the commonplace adage ‘your AI is only as good as the data’. Here, not only do “AI” and “data” appear as distinct terms, but also the ‘goodness’ of the former is contingent on the latter. In distinguishing “AI” from “data”, we not only make a distinction between the algorithm and its infrastructure, but also displace the responsibility of goodness onto the latter. It is through this process of displacement – of morality, and of bias – that the algorithm is constructed as neutral.
Thus, resisting this authority cannot just involve building better classifiers, or training them on more inclusive data, or ensuring queer people are in the room when the models are designed (remember that OpenAI’s CEO Sam Altman is gay). These are calibrations, and calibration accepts the legitimacy of the instrument.
How do we resist computational authority, then? That work begins, I suggest, by learning to see differently. In a 2004 book, sociologist Satish Deshpande defined the task of sociology as “squinting”. A decade later, in 2014, sociologist Pushpesh Kumar wrote about “squinting through queer eyes” as a means of “incorporating the sexuality perspective in Indian sociology.” To resist computational authority, then, we squint. At AI’s claims of making unimaginable strides, at our instinct to be impressed by the speed at which it makes these strides. And at the belief that queering AI will manifest a queer future.
Squinting is not all of resistance. But it could be where the work begins.
Cover image by Google DeepMind on Unsplash
