The AI was given a photo of one heterosexual person, and one of a non-heterosexual person, and they're only getting it right 81% of the time. As the article later says, I think they're really training the AI on stereotypes, as opposed to actual facial data-points that indicate someone's sexuality.