Should an AI copy of you help decide if you live or die? (arstechnica.com)

Researchers at the University of Washington are taking early steps to pilot “AI surrogates” — digital models of patients intended to predict incapacitated patients’ end-of-life preferences — at Harborview Medical Center. Muhammad Aurangzeb Ahmad, a resident fellow working with UW Medicine trauma faculty, says the project is still largely conceptual but aims to train models on hospital patient data and test whether those models can reproduce patients’ values when they cannot speak for themselves. No hospital has yet deployed such systems, but Ahmad hopes initial accuracy testing could later expand across the UW network. The effort matters because surrogate models could reshape how families, clinicians, and ethics boards approach life-or-death decisions by offering consistency and value-aligned recommendations. Technically, the project faces major limitations: model validation hinges on patients surviving to confirm whether the AI’s predictions match their preferences, and developers aim for roughly two-thirds predictive accuracy as an initial benchmark. Beyond accuracy, the work raises urgent issues around consent, data privacy, bias in training data, accountability, and whether an AI should advise or decide. If pursued responsibly, AI surrogates could become a powerful but ethically fraught tool in critical-care decision-making.