Two Stanford University researchers, Dr. Michal Kosinki and Yilun Wang have published a paper that claims that AI can predict sexuality from a single facial photo with startling accuracy. This research is obviously disconcerting since it exposes an already vulnerable group to a new form of systematized abuse.

The research can be found here https://osf.io/zn79k/ ,here https://psyarxiv.com/hv28a/ and has even been highlighted by Newsweek magazine here http://www.newsweek.com/ai-can-tell-if-youre-gay-artificial-intelligence-predicts-sexuality-one-photo-661643

enter image description here

Above is an image of composite heterosexual faces and composite gay faces from the research. (Image courtesy of Dr Michal Kosinki and Yilun Wang)

My question is, as knowledgable members of the AI community, how can we scientifically debunk/discredit this research?

  • 2
    $\begingroup$ All kinds of phrenology are discredit by itself. $\endgroup$ Commented Feb 21, 2018 at 8:34
  • 1
    $\begingroup$ Male and females can be differentiated by their.skull size,etc...Even a person's ethnicity can be determined...It is a very well known field in forensics...So since.huamns can do it by measuring skull feature sizes of course a nn can do it....I don't see the disconcerting part though $\endgroup$
    – user9947
    Commented Feb 21, 2018 at 9:49

1 Answer 1


One way to criticize the study could be to attack the data on which the study is based on. An image on a social network is not "neutral" (those are not ID photo) and certainly not images from a dating website (from which the data of the study come from).

For example as a homosexual / heterosexual person you will perhaps put forward different attributes on your photo (facial hair / glasses, type of clothes) to attract gender specific people.

Those are parameters are not directly linked with the "facial profile" of the person but they will influence the black-box model during training, so you will end up thinking you can detect sexual orientation of a person with only his face characteristics but in reality your black-box algorithm have only detect a totally different characteristic (glasses for example) that are link with a specific sexual orientation.

  • 3
    $\begingroup$ So the high accuracy of the system has only been validated on some dating sites . . . there was a similarly flawed "criminal detector" in the news last year that used police photos for positive class, and social media photos for negative class. It learned to detect the style of police photos . . . very worrying since the process for running the algorithm could well end up being "catch someone we think is suspicious, take their photo, pass it to the algorithm . . . aha, computer thinks we may be right!" $\endgroup$ Commented Feb 21, 2018 at 16:59

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .