Finally some sanity in this genre of research:<p>"To address these limitations, we perform a megastudy—a survey-based study that reports the predictability of numerous personal attributes (349 binary variables) from 2646 distinct facial images of 969 individuals. Using deep learning, we find 82/349 personal attributes (23%) are predictable better than random from facial image pixels."<p>If you look at the findings neatly summarized in Fig. 2, the predictable ones are as one might expect (race, gender, age, etc.), and the rest is just plain pseudo-science, even though the authors seem to interpret AUC ~= 0.6 as "predicatable" instead of plain noise. Nature's reviewers should take a hint.