Can you tell if a person is gay or straight simply by looking at their face? Researchers at Stanford University say they’ve developed a computer algorithm that can make a very good guess. In a new study, the artificial-intelligence program accurately identified men as gay or straight 81% of the time, and women 71% of the time.
That’s better than we humans can do with our own eyes and brains. Using the same photographs, study volunteers could only predict men and women’s sexuality 61% and 54% of the time, respectively. Previous research has also suggested that people’s assumptions about sexual orientation—based on looking at faces alone—are correct only slightly more than half the time.
RELATED: Do These 5 Things Really Influence a Baby's Gender?
But the new research, scheduled to be published in the Journal of Personality and Social Psychology, is not without controversy. Shortly after media outlets reported the study last week, two prominent gay-rights groups—GLAAD and the Human Rights Campaign—released a joint statement criticizing the research and voicing concerns about its potential implications.
Those groups call the study “dangerous and flawed research that could cause harm to LGBTQ people around the world.” The researchers behind it have responded by defending their findings and their motivations for publishing them.
But back to the science in question here: How can a headshot alone reveal clues about sexual orientation?
For the study, the researchers analyzed hundreds of thousands of publicly available photographs from profiles on a popular American dating site. When they narrowed those down to photos with faces of sufficient size and clarity—and made sure that men and women, as well as gay and straight people (based on information in their profiles), were all equally represented—they had a sample representing nearly 15,000 members.
The researchers fed most of these images into a software program that created “faceprints.” The program looked for consistencies among those who were interested in same-sex partners. With this information, the software developed a predictive model, which the researchers then tested against other photographs not included in the initial batch.