No, AI can’t tell if you’re gay
The other day, I came across a stunning article about Artificial Intelligence (AI) from research out of Stanford University. A professor and a graduate researcher used AI facial recognition to determine whether or not someone is gay using a single photograph. How successful is it? The AI could predict accurately 81% of the time for men and 74% of the time for women. If given five photographs, those numbers increase to 91% and 83%, respectively.
So it seems that AI can now predict human sexuality from our faces… except it can’t. What it can do is regurgitate the researchers’ fallible conceptualization of sexual orientation which was based on a foundation of stereotypes.
To begin, Wang and Kosinki base their entire research project on sexual inversion theory which, in essence, claims that gay men are effeminate and lesbians are hypermasculine women. It’s evident that Wang and Kosinki are not scholars of gender theory, nor do they demonstrate even basic understanding of human sexuality, sexual orientation, sex, gender identity, gender expression, or cultural influences. If you’re curious, here’s a quick educational lesson:
- Sex ⇒ your biological, reproductive parts
- Gender ⇒ “our deeply held, internal sense of self as male, female, a blend of both, or neither; who we internally know ourselves to be” ¹ Additionally, cisgender can be defined as your internal sense of self aligning with your biological sex; transgender can be defined as your internal sense of self not aligning with your biological sex.
- Gender Expression ⇒ “how we present our gender in the world and how society, culture, community, and family perceive, interact with, and try to shape our gender” ²
- Sexual Orientation ⇒ to whom we are physically, emotionally and/or romantically attracted; note that sexual orientation is interpersonal while gender is personal ³
Wang and Kosinki collapse all of these immensely complex ideas into a profoundly simple conceptualization: sexuality can be determined by your face. Furthermore, the group of people that were used to train the computer model were exclusively white, cisgender, openly gay men.
“Next, we employed Amazon Mechanical Turk (AMT) workers to verify that the faces were adult, Caucasian, fully visible, and of a gender that matched the one reported on the user’s profile.” (Wang and Kosinki, p. 12)
Let’s unpack the problems with that.
- Their sample introduces racial bias which further isolates any meaningful incorporation of cultural elements across racially diverse communities, particularly how those elements influence gender expression.
- The cisgender criteria introduces gender bias, completely ignoring the transgender community. Note, the researchers don’t even use the word cisgender or transgender once in their paper.
- Moreover, Wang and Kosinki outright ignore any theory of gender fluidity.
Even the researchers’ core mission — to “advance our understanding of the origins of sexual orientation and the limits of human perception” — is restricted. Sexual orientation, like gender, is on a spectrum. To give some structure and validity (a product of the heteronormative world), some identities are commonly claimed; we refer to these as GLBTQ+, meaning gay, lesbian, bisexual, transgender, queer, and the plus is representative of the entirety of other identities on the spectrum.
I still find it curious how the researchers flatly ignore the reality of bisexuality. Wang and Kosinki collapse the entire GLBTQ+ spectrum down to gay and lesbian identities. Not only is their research plagued with social bias and stereotypes, the very goal of advancing an understanding of sexual orientation is without context of intersectional communities of identity.
“Lesbians tended to use less eye makeup, had darker hair, and wore less revealing clothes” (p. 20)
“Lesbians tended to wear baseball caps” (p. 21)
I used to think that professors from elite institutions like Stanford should be granted some inherent intellectual merit. Suddenly, I’m questioning that.
These results are not surprising. As I mentioned above, the researchers based their entire model on a foundation of stereotypes. Wang and Kosinki fed the model stereotypes so that’s exactly what it spit out. At its core, their study is one that embodies the ignorance of their own social biases.
It’s also fair to speculate on correlation versus causation. Gender atypical behavior may cause someone to be gay or being gay may cause gender atypical behavior; regardless, when it comes to addressing these ideas, Wang and Kosinki are absent.
We need to talk about the ethics of this project right now. I can’t imagine if a country like Russia got ahold of this technology. Violence and hate against those who identify as LGBTQ+ is real and it is terrifying.
If you’re motivated or curious about the status of rights for LGBTQ+ people internationally, here is a great resource.
A Note on Viral Media
Once this research was released, major media and journalistic groups spread it like wildfire. However, I believe this points out a very problematic point in our society: there is a fundamental lack of understanding of science and research, both in the population-at-large and in the media.
One of the core lessons that we can learn from this scenario is that the media often liken a researcher’s intellectual and academic merit to their institution’s intellectual and academic merit. I read many articles in preparation for this piece. Almost every one made sure to mention that the research was out of Stanford University. What very few mentioned was the asymmetry between the researcher’s academic background and the research area of interest. Determining sexual orientation requires an understanding of gender theory, something to which a researcher of Anthropology might be well versed.
The professor in this study was out of the Stanford business school. The other researcher recently completed his Master’s in computer science. Neither of which has merit to be making claims about the sexual orientation of the public. Now I’m not saying that only anthropologists should study sexual orientation. In fact, I believe there is a great benefit to having a research team of diverse academic backgrounds; that leads to more intersectional research and deeper insights. What I am saying is that at least one expert in the field should be involved in a study’s topic of interest. Additionally, we need a greater public understating of academia and research, so we stop conflating an institution’s reputation with that of a researcher’s.
1, 2, 3 — Gender Spectrum
If you’re looking for a more scathingly critical and detailed review of the Wang and Kosinki study, you can find that here. Disclaimer: that source was a major motivator for me to write this piece; I owe immense credit to the author, Greggor Mattson.