Most child-related AI incidents trace to deployment failures, not model defects, study finds

A preprint posted to arXiv by researchers led from the University of Texas at Austin coded 1,331 publicly reported AI incidents involving children, drawn from the OECD.AI incident monitor (1,143), the AIAAIC repository (141) and the AI Incident Database (47). In 779 of the 1,331 incidents, the primary failure was in how a system was deployed, governed or used, led by systems placed in settings that broke expectations about age-appropriateness, supervision or consent (398); model behavior or technical failure was the primary failure in 196. Sexualization and sexual exploitation accounted for 447 incidents, about a third of the set; the authors caution that the counts reflect what gets publicly reported, and that their failure-mode coding is provisional because agreement between coders was moderate.