Can your sexual identity be predicted?
The reaction in the LGBT community has been critical — although not completely so
Can a machine judge whether you are gay or heterosexual from your photo alone? That’s the tantalizing question that’s been addressed by a pair of Stanford researchers whose work has set off an international controversy shaped by fears of the past and future.
The two researchers, Michal Kosinski and Yilun Wang, set off to determine whether artificial intelligence could recognize something very personal about us — in this case, sexual orientation. Their purpose was to sound the alarm about things that could be done by authoritarian regimes with existing technology.
In doing so, they’ve sparked a debate about the weight of their evidence and the wisdom of their research. It’s not just that AI potentially can be used to identify sexual orientation. It can be misused.
Kosinski and Wang began by scraping some 35,000 photos from online dating profiles — both heteroseuxal and gay (presuming, for example, that men seeking men fell into the gay category and that men seeking women were heterosexual).
By using a software program called VGG-Face, they then trained a machine to pick out facial data points like nose width, moustache, eyebrows, etc. Given two photos — one presumptively of a heterosexual, the other of a gay person — the machine reportedly picked who was more likely to be gay with 81 percent accuracy for men, 74 percent for women.
Here’s where you have to be careful: Kosinski and Wang were not saying that AI could pick out who was gay in a crowd 81 percent of the time. Picking who is most likely gay among two people with similar pictures is different than sweeping a room and identifying gays or lesbians.
Since their findings were unveiled — you can read a draft
HERHOLD >> PAGE 8
of their paper at goo.gl/ VGU2Ci — a lot of the discussion has been about just how accurate AI could be. Even if a machine gets it right 81 percent of the time, there’s huge room for error.
Other people worry about a resurrection of an old and discredited science — phrenology — which attempts to equate skull shapes with intelligence. One critic of the KosinskiWang research called it the “algorithmic equivalent of a 13-year-old bully.”
My fear is related to that, but I see it in political
terms. One of the reasons that gays and lesbians have benefited from the most successful civil rights movement of the last halfcentury is that they have come out of the closet, arguing that “We are your brothers and sisters.”
The problem with research like this is that it makes the opposite point. It supposes that there are facial differences that can reveal sexual identity. Whether that’s true or not, it creates the basis for differentiating people — and for discriminating against them.
I couldn’t reach Kosinski for comment, but he has argued that the point of the research is to alert people to what could be done by bad actors. The reaction in the LGBT community has been critical — although not completely so. I think the work deserves publication in a journal, though it may have to be qualified considerably.
The real fear here isn’t that a machine could identify sexual identity with accuracy. It’s that bad folks believe that it could. In many countries, it is still against the law to be lesbian or gay.
We need to be worried not just about the machine — but about the human reaction to its findings.