This paper is bad and Gizmodo should feel bad. Besides making ridiculous and unsupported causal claims (as other commenters have noted), the results are really

This paper is bad and Gizmodo should feel bad. Besides making ridiculous and unsupported causal claims (as other commenters have noted), the results are really poor (R score of 0.2). They show that the model achieves human-level performance, which is also really poor. Then they run it on natural images of politicians and the results are significantly worse. What the paper actually shows is that facial appearance can not consistently predict political alignment, but the authors want to try really hard to do it anyway.