From Snapchat selfie lenses to the iPhone X’s Face ID feature, advances in facial recognition technology are generating huge amounts of enthusiasm and cash. But seeing as face-analyzing capabilities could dramatically up the tech game of surveillance, security, and advertising agents, it’s normal to feel a little wary. As with drones, we have to ask: is facial recognition technology creepy or cool?
Creepy, Stanford scientist Michael Kosinski suspected after hearing Silicon Valley hype about Israeli and Chinese companies that are using facial recognition software to predict a person’s likelihood of committing terrorism and or other crimes. So he and a colleague, Yilun Wang, set out to make a point by designing a study around most invasive and inappropriate use of facial analysis yet: an “A.I. gaydar” that predicts your sexual orientation based on your face. Now they’re facing criticism from some LGBTQ advocacy groups on exactly those charges of creepiness and endangerment.
Kosinski and Wang’s program analyzed more than 75,000 photographs of white people seeking either same- or opposite-sex partners on dating websites in the United States. It used its facial scans to create neural network composites of straight and gay faces (respectively), which it would then compare with a given face to predict that face’s sexuality.
Under certain conditions, the machine could perform with as much as 91 percent accuracy—an impressive but still limited figure, if you consider that the relative smallness of the gay population means that “almost two-thirds of the time it says someone is gay, it would be wrong.” When the computer had fewer photos to analyze or was presented with a mix of pictures from Facebook as well as online dating sites, its accuracy rate dropped as low as 74 percent. Though Kosinski and Wang’s study was upfront about these statistics, exaggerated news reports that “made it sound almost like an X-ray that can tell if you’re straight or gay” soon emerged.
The study and subsequent media whirlwind provoked ire from Glaad and the Human Rights Campaign, advocacy groups who condemned the study for “threaten[ing] the safety and privacy of LGBTQ and non-LGBTQ people alike” and dismissed it as “junk science.” LGBTQ Nation responded to these statements in defense of Kosinski and Wang, roasting Glaad for not understanding “how science works.”
Part of the problem is that Kosinski and Wang’s report traffics in both established, uncontroversial scientific methodologies and theories of sexuality that are widely considered to be bunk. Making predictions within certain parameters and then testing if they were correct is the very core of the scientific method; in that respect, the Stanford study checks all the boxes.
But the authors positioned their work in the dark tradition of racist science when they suggested that their study was evidence in support of physiognomy, the practice of judging personality traits from facial characteristics that has been largely debunked as pseudo-science since its heyday in the 19th century. Furthermore, they backed this claim by drawing on prenatal hormone theory (P.H.T.), which they said “predicts the existence of links between facial appearance and sexual orientation” based on differential exposure to hormones in the womb. Not only is P.H.T. disputed within the scientific community, but the suggestion that it has implications for facial features is an even farther stretch.
So what explains the gaydar’s success if not physiognomic theories? Patterns—patterns that the A.I. has found in the data, though it’s hard to say what they are and if they tell us any useful information at all. The realization that “gay people tend to post higher-quality” photos, for example, has problematized similar attempts at programming an A.I. gaydar.
Kosinski and Wang’s paper is just one more illustration of how the line between “junk science” and real science is as thin and blurry as the line between correlation and causation. Any way of making sense of the world that involves carving it up into general categories is bound to miss something, and science, with its taxonomies and averages, is no exception. Sometimes the thing it misses is whatever makes us see others as deserving of acceptance and respect. That’s the risk we take whenever we do science: it can be cool, but it can also be deeply, deeply creepy.