Physiognomy never fully disappeared from serious inquiry, and it never gained genuine scientific support either; no controlled study has established that facial structure reliably predicts character, intelligence, or criminal tendency, and the field's core method, matching an observed trait to an assumed inner quality by resemblance alone, has never worked better than chance or prejudice. What has changed is the technology used to make the same claim. From the mid-2010s, several machine learning papers claimed algorithms trained on facial photographs could predict criminality, sexual orientation, or other traits with high accuracy; a widely discussed 2016 preprint claimed nearly ninety percent accuracy distinguishing criminal from non-criminal faces from simple headshots.
These modern studies drew immediate and detailed criticism on grounds that echo physiognomy's oldest problems in new statistical language: unclear or undisclosed sources for the photographs used to train and test the algorithms, particularly the criminal image sets; a serious risk the models were picking up incidental cues, such as photograph type or grooming, rather than any genuine facial signal; and, most fundamentally, no plausible biological mechanism by which stable facial geometry would predict criminal conduct, a legal and social category rather than a physical trait. Critics writing in outlets including Undark described the resurgence as physiognomy dressed in the authority of artificial intelligence, warning a confident-sounding algorithm can lend old pseudoscience a new, harder-to-question legitimacy. Mainstream researchers have broadly treated this work as a cautionary example of the field's own capacity to encode bias and mistake pattern-matching for discovery.