Stanford University scientists created an AI algorithm that allows you to know a person’s sexual orientation by photo. The system catches barely visible features and identifies homosexuals with 92% accuracy. The technology is not yet perfect and will be further developed. The article preprint is posted on the PsyArXiv.com resource.
The new American scientists' development can completely devaluate the Internet privacy rights. People can easily determine sex, race, age and even someone’s personality by a picture. However, few people can accurately determine sexual orientation just from someone’s looks. Now AI is here to help them.
The American researchers’ system analyzed 35,000 quality images of young men and women. The numbers of heterosexual and homosexual people among them were approximately equal. The Face++ algorithm paid attention to placement and shape of the eyes, nose, mouth, head tilt, etc. Next, the VGG-Face program transformed face images into a vector of all sorts of features (skin tone, eye shape, forehead size, look direction, etc.). With linear regression, system determined how much do these characteristics correspond to the orientation. Artificial Intelligence even created portraits of typical heterosexual and homosexual representatives.
To test the accuracy of the developed method, scientists tested the system on other pictures. The results are striking:
- by photo the algorithm accurately identified male homosexuals in 81% of the cases and women - in 71% of the cases;
- by several images, the accuracy increased to 92% for men and 83% for women.
In comparison, ordinary people correctly guessed the orientation in 61% of the cases for men and 54% for women. In other words, they were just giving random answers.
This study indirectly confirms the theory that sexual orientation is determined at the stage of a child development in the womb and largely depends on hormonal balance. Moreover, results of the development testing prove that the human face contains much more information than we thought previously.
It is important to consider that the neural network was trained on a specific selection of young people with white skin. We can’t tell yet, whether the accuracy will remain the same, when scaling the results to a bigger number of people.