This is just research though - They usually have 5% significance, so if the accuracy is over 5% higher than flipping a coin (50%), it’s significant. What I took away was that humans can pick out gays significantly better than chance (61%) and the machine did it extremely well (84%). It’s impressive.
The math is one of those things that I don't get until somebody explains it to me carefully, and then when I try to think it through on my own I still don't understand it. It's like the "Monty Hall: Do you want what's behind the curtain?" problem.
I don't think anybody quarrels with the idea that it's easy to spot many people as gay. There was one kid in every high school like that. Statistically, though, if the school was big enough there could be 6 or 7 or 8 or 10 others who you wouldn't be able to accurately identify. The false positives and false negatives would start to add up, especially if you didn't have other clues.