When scientists presented the algorithm with five facial images of a single person, the accuracy increases to 91 percent for men and 83 percent for women.

Kosinski and Wang used “deep neural networks” to sample 35,326 facial images of men and women taken from a leading dating website.

The advanced computer analysis compared different facial characteristics and found that gay men and women tended to have “gender-atypical” features. This included fixed features such as the shape of one’s nose or jaw, as well as transient features, including hairstyles and facial hair.