For ML practitioners building clinical decision support tools, the deployment of AI-based intimate partner violence (IPV) screening in hospitals creates a governance and consent gap that standard performance metrics do not capture. An opinion column in MedPageToday by Oni Blackstock -- physician and founder of Health Justice -- argues that some hospitals run electronic health record data through IPV risk-scoring models without patients necessarily knowing this analysis occurred, potentially exposing survivors to harm. The critique targets a real deployment trajectory: a 2026 NIH-funded study from Mass General Brigham and Harvard Medical School (the AIRS tool) showed an 88% accurate multimodal fusion model that can flag IPV risk more than three years before patients seek help at intervention centers. A concurrent systematic review by Yang Li et al. in the Journal of Clinical Nursing identified 41 studies applying AI to IPV detection and flagged algorithmic bias, data privacy risks, and integration barriers as unresolved challenges.
First impressions of the EA/AI safety ecosystem from the outside