Abstract
The operating room remains a paradox: it is one of the most sensor-rich environments in the hospital, yet it produces largely underutilized data. While surgical artificial intelligence (AI) has achieved remarkable progress in recent years, the day-to-day practice of surgery has changed little, with most systems confined to passive decision support. This narrative review traces the evolution of surgical AI from perception to cognition to early forms of action, arguing that the next paradigm shift requires "physical AI"-systems capable of meaningful physical interaction and autonomous execution. The clinical motivation for pursuing physical AI is clear: surgical outcomes vary substantially across surgeons, access is constrained by workforce shortages, and high-quality care remains tied to the scarcity of human expertise. If reliable autonomous systems can be developed, surgery could become more standardized, scalable, and reproducible. However, a critical bottleneck persists: the scarcity of synchronized, multimodal training data. The fundamental barrier is environmental rather than algorithmic, as most operating rooms are not configured to measure surgical practice objectively. We propose reconceptualizing the operating room as a "Surgical Data Factory"-a closed-loop ecosystem designed to capture multimodal signals, structure them via consensus taxonomies linked to outcomes, and utilize them for training, validation, and monitoring. Surgeons must transition from passive users to active architects of this infrastructure. Investing in systematic data governance is the prerequisite for responsibly developing, validating, and scaling physical AI in surgery.