Can A Machine Tell Whether You Are Gay?

Sep 15, 2017

Drawing on databases of images collected from an online dating site, a new study conducted at Stanford University concludes that faces carry information about sexual orientation.

This information is not available to visual inspection by ordinary perceivers. But it can be extracted by powerful, pattern-recognizing machines ("deep neural networks" or DNNs).

According to the study, which has also been reported here , here and here, a DNN was 91 percent accurate in determining sexual orientation from photos of men and 83 percent accurate with woman. Humans, given the same images to inspect, had a much lower level of accuracy (on the order of 20 percentage points lower).

The information is there, in the image, it seems, but not visible to the naked eye.

The machine, in making its judgment, is able to make use of information about the shape of the noses and jaws of imaged people, that is, on the basis of unlearned, non-acquired physical characteristics.

Jawline and nose shape are controlled, in part, by the presence of sexual hormones in utero, according to the prenatal hormonal theory or PHT. For this reason, the authors take their findings to support what they claim is a "widely accepted" biological (nature, not nurture) theory of sexual orientation according to which "overexposure" to androgen leads to homosexuality in females whereas "underexposure" leads to homosexuality in males.

Gay men, so the PHT would have it, are men with less "male hormone." And gay women, so the theory would have it, are women with "too much."

I'm surprised to learn that you can tie sexual preference to physiognomy. Of course, the putative fact (supposed by PHT and the Stanford study) that men with more "feminine" facial features are more likely to be gay doesn't show that the physical feature and the sexual preference have a common genetic cause.

Which brings me to the deeper issue raised by the whole hormonal theory in the setting of this study.

"Gay" and "straight" are pretty useful terms. We know what they mean and how to use them. But can we even begin to take seriously the idea that these categories exhaustively and reliably bifurcate the space of adult human sexuality? The point is a serious one — it has a bearing on whether we can tell whether the DNN gets it right when it attributes gay or straight in a given case. A person was counted "really straight" for the purposes of this study if they were seeking dating partners of the opposite sex. But is that a reliable, an objective indicator? Which is another way of asking: Is there an objective fact here at all?

You might have thought that there is something silly, and even offensive, about the idea that you, or a machine, might read off sexual preference from pictures. Indeed, it's about as silly and offensive as the idea that you could judge a person's politics or intelligence or liability to criminal behavior on a similar basis.
But this is a bullet the study's authors are happy to bite.

Prof. Michal Kosinski, one of the authors, was quoted in The Guardian saying: "The face is an observable proxy for a wide range of factors, like your life history, your development factors, whether you're healthy."

It's just a matter of time, they seem to believe, before DNN can crack the face code. This will doubtless be an evolution with broad civil liberties and privacy implications.

What it won't do, I suspect, is settle any deeper questions about nature versus nature.


Alva Noë is a philosopher at the University of California, Berkeley, where he writes and teaches about perception, consciousness and art. He is the author of several books, including his latest, Strange Tools: Art and Human Nature (Farrar, Straus and Giroux, 2015). You can keep up with more of what Alva is thinking on Facebook and on Twitter: @alvanoe

Copyright 2017 NPR. To see more, visit http://www.npr.org/.