A release gate for patient-facing generative AI: what health systems should require before a new version reaches patients

Objective To set out what evidence should decide whether a new version of patient-facing generative AI reaches patients. Discussion FDA, Joint Commission and Coalition for Health AI guidance expect changes to be validated, yet none tells health systems how to decide whether a new version ships. Software release testing cannot fill this gap: it relies on cases with known correct answers, whereas generative AI often has no single correct answer, and each reply shapes what the patient says next. Benchmarks can miss the change: in our audit of one, a third of a new guardrail's actions fell on turns no LLM judge graded. We found no published method that combines a fixed test corpus, a measured noise floor and a pre-specified release rule. Conclusion We call such a method a release gate and set out its seven parts, how it changes for children, and what governance committees should require before and after go-live. Preprint of a Perspective submitted to JAMIA (Journal of the American Medical Informatics Association) on 3 November 2026; not peer reviewed.