This AI claims to know whether you’re gay or straight from a single photo
A new artificial intelligence system has been created that claims to be capable of accurately guessing whether a person is gay or straight, based purely on a photograph of their face.
Researchers from Stanford University developed a deep neural network – used to analyse visual patterns from large datasets – and extracted features from 35,000 images pulled from a US dating site.
After creating a predictive model from patterns in the data, they showed the system pictures of gay and straight men and women. They found the network was able to correctly distinguish sexual orientation 81% of the time for men, and 74% for women.
The results raise a number of sensitive questions about the biological origins of sexual orientation, as well as the potential for facial recognition systems to be abused as a tool for LGBT persecution.
The study, published in the Journal of Personality and Social Psychology and first picked up by the Economist, found that the faces of gay men and women had “gender-atypical” features and “grooming styles”.
The researchers – Dr Michal Kosinski and Dr Yilun Wang – say this provides “strong support” prenatal hormone theory (PHT) of sexual orientation, which posits that exposure to certain hormones before birth can determine sexuality. Their findings supported PHT conclusions that gay men tend to have narrower jaws, longer noses and larger foreheads compared to heterosexual men, while gay women tend to have smaller foreheads compared to heterosexual women.
The study also asked humans to guess the sexuality of the subjects. They were found to be a worse set of judges, only correctly identifying sexual orientation 61% of the time for men and 54% for women. On the other hand, the neural network was found to improve its accuracy when shown five images of each person, judging correctly 91% and 83% of the time for men and women respectively.
While the paper offers a shocking glimpse into how AI could be used to profile citizens, its authors defend their research as a necessary statement about the technology’s potential uses.
At the start of their paper, Kosinski and Wang note that the predictability of sexual orientation “could have serious and even life-threatening implications to gay men and women and the society as a whole”. They point to reports of governments and corporations developing face-based prediction tools to judge the likelihood of a person committing a crime, and also note that several countries consider homosexuality to be a criminal act:
“It is thus critical to inform policymakers, technology companies and, most importantly, the gay community, of how accurate face-based predictions might be.”
Kosinski has previously worked with Cambridge University to study the scope for psychometric profiling to be undertaken with Facebook data. The authors write that they hope future research will go further in exploring the link between facial features and personality, including political views or psychological conditions.
Image: From Kosinski and Wang’s study.