I don't think this is true, because just 50% of the men and 50% of the women in the sample they used were heterosexual (not 90% and 85%), and their accuracy statistics represent "the likelihood of a classifier being correct when presented with the faces of two randomly selected participants--one gay and one heterosexual" (from lines 241 and 299 - 300 of the linked paper).
Slashdot Mirror
I don't think this is true, because just 50% of the men and 50% of the women in the sample they used were heterosexual (not 90% and 85%), and their accuracy statistics represent "the likelihood of a classifier being correct when presented with the faces of two randomly selected participants--one gay and one heterosexual" (from lines 241 and 299 - 300 of the linked paper).