The measure that matters: What medicine asks of artificial intelligence

Medical artificial intelligence (AI) is advancing at unprecedented speed, yet the dominant question in the field, how well does a model perform, diverges from what clinicians ask of it in practice. Did it help someone survive? In this Position Paper, we argue that a misalignment exists between how the AI community evaluates success and how clinicians define value. Even the strongest randomized trials of clinical large language models measure subspecialist preference rather than patient survival, and in a 2024 scoping review of 86 randomized trials of AI in clinical practice, only 17% had a primary endpoint related to patient symptoms or behavior. Surveys of clinicians point to a further dimension that no benchmark captures, the erosion of human connection between clinician and patient. We propose a life-centered evaluation framework built on five principles, namely outcome primacy, dynamic evaluation, clinician co-design, relational outcomes, and transparent failure publication. The measure that ultimately matters is whether a system keeps a patient out of the intensive care unit, prevents a resuscitation, or spares a family a preventable loss.