That’s a bold claim that is nonetheless supported by recent data, according to an article just out from the Economist. Fresh research out of Stanford University shows that deep neural networks may now be more able to detect signs of heterosexuality or homosexuality than their human counterparts.

With merely one image the deep neural network could accurately discriminate between straight and gay men with greater than 80 percent accuracy. That number dropped down to 74% for women. In other words, the deep neural network used by Stanford researchers could only tell whether female participants were actually gay or straight in fewer than three-fourths of cases. This was significantly lower predictive ability compared to men.


The deep neural network looked at interesting facial cues to decipher whether men and women were gay or straight. The size and composition of a subject’s nose as well as how well groomed the men in the study were proved to be decisive factors when it came to correctly discovering the sexuality orientation of the participant. Expressions and facial morphology were broad areas of discrimination that researchers now believe the AI in the study used to outperform humans.

Although the study is a watershed moment for AI and its learning capacity, the study presents virtually as many questions as answers. The researchers, though, were satisfied concluding that the higher discriminatory ability of their deep neural network enhanced our collective understanding of the roots and daily expression of human sexuality in everyday contexts.

The researchers’ findings also underscore the fallibility of human judgement and perception in classifying people based on perceived sexual orientation. So, in a sense, these findings are a deeply cautionary tale. The software that outperformed humans is known as VGG-Face, and it’s the latest incarnation of AI that houses and makes sense of thousands of images using numbers known as faceprints.

The disappointing aspect of the study is that humans only correctly deciphered the sexuality of the subject illustrated in the picture about 50% of the time. You would almost, but not quite, be better simply flipping a coin since humans trying to crack the images could only guess correctly 54% for pictures illustrating homosexual women. The numbers, like the AI results, were slightly higher for men in that people were able to correctly guess sexuality in male pictures 64% of the time.

More research is needed to determine the study’s implications.