Big brother isn’t just watching anyway. Apparently he’s thinking too.
Researchers Michal Kosinski and Yilun Wang from Stanford University say artificial intelligence can now tell a person’s sexual orientation just by analyzing their face and facial structure, the Economist reports.
The guys used a total of 35,326 pictures of both men and women, gay and straight, all represented evenly collected from an unnamed dating website.
The images were then fed into a piece of software called VGG-Face, which creates “faceprints” of people through a long string of numbers.
From there, the software used a predictive model called “logistic regression” to guess people’s sexual oritentations based on their faceprints.
The rate in which it guessed correctly?
When given one photo of a gay man and one photo of a straight man, both chosen at random, the software correctly guessed 81% of the time. When given five photos of each man, it did even better, correctly guessing 91%–91%!–of the time.
The software didn’t do quite as well with women, however. When given one photo of a lesbian and one photo of a straight woman, it had a 71% accuracy rate, and an 83% accuracy when given five.
Still, it outperforms any human’s ability.
When researchers gave the same images to people, they could only tell the gay man from the straight man 61% of the time, and only 54% of the time for the women.
Kosinski tells the Economist that he conducted the experiment primarily to remind people of the power and potential dangers of machine vision. He believes that there will be a day when privacy as we know it no longer exists and that the AI systems may eventually be trained to determine other intimate traits, such as a person’s IQ or political views.