Research in PNAS Nexus finds that popular AI language and image models infer trustworthiness, competence, and criminal risk from facial appearance alone, echoing human face-to-character bias despite no link between face structure and real traits; experts warn such judgments permeate high-stakes deci