- A facial recognition algorithm was applied to naturalistic images of 1,085,795 individuals to predict their political orientation by comparing their similarity to faces of liberal and conservative others. Political orientation was correctly classified in 72% of liberal–conservative face pairs, remarkably better than chance (50%), human accuracy (55%), or one afforded by a 100-item personality questionnaire (66%). Accuracy was similar across countries (the U.S., Canada, and the UK), environments (Facebook and dating websites), and when comparing faces across samples. Accuracy remained high (69%) even when controlling for age, gender, and ethnicity.
When the gut reaction is "junk science" it's interesting to try and find out how it might be junk. Probably the study titled "Facial recognition technology failed to expose political orientation" didn't get accepted. The abstract claims 69% accuracy with controls but this figure does not appear in the body. A million faces sounds like a lot but "controlling for age, gender, and ethnicity" greatly reduces the power. Table 1 shows that the samples were 65% female and 68% white, leaving 35% non-female and 32% non-white. If, ideally, these features are completely uncorrelated, that gives just 11% of the total sample (121,609) "minority" faces for the control group. (If we allow that there is more than one category of minority, it gets worse.) Controlling for age is worse, even bracketing by decades would provide half a dozen categories; the algorithm tried to match within one year. The algorithm scored a barely-better-than-luck 58% by determining which way the head was posed, and 57% by evaluating facial expressions like anger or surprise. The accuracy was better (66%) for personality, but I'm not clear on how personality was evaluated. Somehow these numbers are combined to come up with 73% overall. Might be more clear when the dataset is posted in mid-January.