A Human Was in the Loop. The Agent Kept Going. Before You Give AI Authority

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

PREPRINT — This manuscript has not undergone peer review. This essay examines the gap between detecting a problem and ending an AI system’s activity. It uses a reported OpenAI agent incident to argue that institutions should demonstrate that interruption, recovery, and limits on authority work before allowing an AI system to act. The hospital example is hypothetical. First follow-on essay in the Quis Custodiet series. Approved author version submitted to GrandRoundsMD October 1, 2026; journal decision and publication pending. The Word and PDF files preserve the September 30 approved manuscript and title. AI assisted drafting and revision; the author reviewed and approved the manuscript and takes responsibility for its content.