JMIR Publications

ABSTRACT

Artificial intelligence (AI) is now embedded infrastructure in perioperative care. Risk stratification algorithms, hemodynamic prediction tools, and clinical decision support systems are active in operating rooms at major health systems, and their adoption is accelerating. Yet the field has studied model performance and organizational implementation while largely bypassing the moment between them: the real-time encounter in which an anesthesia provider must decide, under active case conditions, what to do with an AI-generated output. We term this the cognitive transaction and argue it is the fundamental unit of perioperative AI implementation. The perioperative environment presents a specific constellation of conditions that existing human-AI interaction research was not designed to address. Continuous real-time decision demands, extreme time compression, high cognitive load, and consequences that unfold in seconds distinguish the operating room from the clinical contexts where most provider-AI interaction research has been conducted. What we know about AI adoption in radiology, oncology, or ambulatory care does not translate cleanly to this setting. The cognitive moment in anesthesia has its own structure, its own failure modes, and its own research requirements. This paper examines what those requirements are. We analyze how the operating room functions as a pre-existing human-machine cognitive system into which AI is now being inserted, and why the conditions of that system generate predictable vulnerabilities: miscalibrated trust, automation bias, and cognitive friction produced by interfaces optimized for technical accuracy rather than clinical usability. We argue that these failure modes are not incidental but structural, and that they will persist regardless of model performance until the provider-AI interaction is itself treated as a research object. We identify four priority research domains. The first concerns the structure of provider-AI disagreement and the methods needed to distinguish automation bias from legitimate clinical insight. The second concerns the longitudinal dynamics of trust calibration across repeated clinical encounters rather than single-session experimental designs. The third concerns interface design for high-acuity workflows, specifically what constitutes a usable AI output for a provider managing a patient in real time. The fourth concerns the need for ecologically valid study designs capable of capturing provider reasoning under actual intraoperative conditions rather than retrospective or survey-based proxies. The anesthesia and perioperative research community is positioned to lead this work. The clinical specificity, domain knowledge, and professional stake required to design meaningful studies are all present within the field. Evaluating the provider-AI dyad under intraoperative conditions, rather than the computational model in isolation, is both a methodological imperative and a patient safety priority.