Anesthesiologists propose risk-based rules for safe use of AI in surgery

Large language models are being discussed as potential assistants across medicine, but an editorial published in Anesthesiology and Perioperative Science argues that anesthesiology is not simply another setting in which a fluent chatbot might be useful. The authors, led by Zhimin Tan, Qiyu He, Qian Li and Xiaoqiang Li, set out a risk-based roadmap in which the clinical readiness of an artificial intelligence system is judged not by how polished its output sounds, but by what would happen if that output were wrong. Their central proposal is a three-tier framework: lower-risk information-processing tasks may be automated, intermediate-risk interpretive tasks should augment clinicians with explicit verification, and time-critical or high-consequence decisions should not currently be delegated to these models at all.

The argument begins from the nature of the specialty itself. Anesthesiologists must integrate preoperative history, laboratory results, procedural context, rapidly changing physiology, drug administration, airway events, surgical stimulation and postoperative recovery, all under severe time constraints. Recent reviews have catalogued potential applications of large language models and broader artificial intelligence in anesthesiology, but the editorial’s position is more specific: readiness should be judged by the consequences of failure in the intended use, not by the fluency of the output. A model that produces eloquent prose, the authors contend, has not thereby demonstrated that it is safe to trust at the bedside.

The earliest practical value, the authors suggest, is likely to lie in narrow, text-dominant uses with verifiable outputs. These include structuring or summarizing narrative records, retrieving local guidance, and converting unstructured documentation into clinically usable fields. As one example, large language models have been used to extract structured information from intraoperative transesophageal echocardiography reports. Such uses are attractive because the source text remains available for comparison, and errors can often be detected before they affect care. By contrast, predictive or inferential uses demand greater caution. In perioperative risk prediction, one large language model performed reasonably on several classification tasks, including intensive care unit admission and in-hospital mortality, but poorly on continuous outcomes such as postanesthesia care unit, hospital and intensive care unit length of stay. That distinction matters clinically: a model that recognizes broad risk categories does not necessarily capture the evolving physiological trajectory, treatment response, censoring and workflow dependencies that shape continuous outcomes.

In anesthesia, where a patient’s state can change within seconds, the editorial warns that performance on static text or discrete labels should not be treated as evidence of readiness for dynamic bedside control. The proposed three tiers describe intended uses rather than fixed properties of task categories. Risk depends on the downstream clinical consequences of error, the degree of automation, and whether a clinician can independently verify the output before it affects care. Medication or allergy extraction, for instance, may be lower risk when it only prepopulates a draft that is checked against the source record, but higher risk when an omission directly drives prescribing or anesthetic planning. Lower-risk information-processing uses may include drafting summaries, preparing patient-education material, retrieving guideline passages, or extracting structured fields when provenance is visible and verification is feasible.

Intermediate-risk clinician-augmentation uses include American Society of Anesthesiologists Physical Status suggestions, perioperative risk flags, synthesis of intraoperative events and draft management considerations. Because these outputs can influence care, they require explicit clinician verification. At the high-risk end, the authors draw a firm line: difficult-airway strategy, anticoagulation-sensitive neuraxial decisions, vasoactive or anesthetic dosing, malignant-hyperthermia management and other time-critical rescue actions should not be executed or recommended without direct specialist control. The relevant boundary, they argue, is not whether a model can generate a plausible answer, but whether an undetected omission or incorrect recommendation could cause immediate harm.

The framework also yields a simple readiness test for hospitals considering adoption. Before deployment, teams should answer four questions: What exact task is the model allowed to perform? What harm could follow from omission, fabrication or delay? Can the clinician independently verify the basis of the output within the available time? What evidence shows stable performance in the local population and workflow? If any answer is unclear, the system is not ready for that use. Readiness remains task specific: after appropriate validation, the same model may be suitable for drafting a postoperative summary yet remain inappropriate for real-time hemodynamic recommendations. Governance should therefore authorize intended uses, not models, and define explicit escalation and shutdown criteria.

Safety, the authors argue, must be defined more broadly than hallucination. In hazardous regional-anesthesia scenarios, large language models have failed to recognize clinically important hazards despite fluent answers. Omission errors may be particularly dangerous in anesthesia: failure to surface anticoagulant use, a difficult-airway history, severe pulmonary hypertension, a critical allergy or susceptibility to malignant hyperthermia can be as harmful as generating false information. Other failure modes include automation bias, inconsistent responses to repeated prompts, retrieval failure, silent model or version drift, and superficial review of authoritative-sounding text. Human review is therefore a design requirement, not a complete safety solution. For a preoperative summary, the reviewing anesthesiologist should remain the accountable decision-maker; predefined critical fields should be checked against the source record; missing high-risk elements should trigger an incomplete-status warning; disagreement should default to clinician judgment and be logged; and the model should have no autonomous authority to place orders or alter device settings. Reinforcing the human-factors concern, a recent study of model-generated cardiac anesthesia plans found that trainees rated the outputs more favorably than experienced anesthesiologists did, underscoring the need to evaluate how reviewer experience affects oversight.

Perioperative systems, the editorial continues, cannot remain stand-alone text tools. The meaning of a decrease in blood pressure depends on its temporal relationship to induction, surgical stimulation, blood loss, positioning, ventilation and vasoactive administration. High-frequency waveforms and device data operate on a different timescale from laboratory values, medication records and narrative notes. Safe multimodal systems will require explicit temporal alignment, unit and timestamp harmonization, missing-data handling and provenance for every retrieved fact. Retrieval-augmented generation should be linked to version-controlled local guidelines and formularies, with visible source passages and explicit failure states when retrieval is incomplete. Integration with anesthesia information management systems should expose only the minimum necessary data while preserving auditability. The likely future tool, the authors conclude, is not a general-purpose chatbot but a constrained interface layered over structured electronic health record data, physiologic streams and local knowledge systems.

Clinical readiness, the authors argue, requires staged evidence in proportion to task risk. Their proposed sequence begins with retrospective technical testing on temporally separated data, followed by external validation at institutions not used for development, then prospective silent deployment in which outputs are recorded but do not affect care, then human-factor testing that measures error detection, trust, workload, override behavior and automation bias, and only then interventional evaluation. Recent clinical guidance on language models and perioperative machine learning similarly emphasizes clinically meaningful failure modes, external testing, human factors and clinical usability rather than benchmark performance alone. For lower-risk documentation uses, time saved, correction burden and clinically important omission rates may suffice as early endpoints. For decision support that can change management, multicenter randomized, cluster-randomized or stepped-wedge trials should assess unsafe recommendations, treatment changes, near misses, complications and patient-centered outcomes. Post-deployment registries should track performance after software, model, guideline or workflow updates. A single benchmark score, the editorial insists, should never serve as a surrogate for this evidence chain.

Governance must be equally concrete. Responsibility cannot disappear behind the general label of AI-assisted care: institutions should define who approves each intended use, which clinician is accountable for acting on an output, how disagreements are documented, when the system is disabled, and how adverse events are reviewed with the developer. Regulatory status should be assessed according to intended use rather than by whether a system is described as a language model. The editorial notes that the US Food and Drug Administration’s 2026 guidance distinguishes certain non-device clinical decision-support functions by whether a health professional can independently review the basis of a recommendation and not rely primarily on it, while time-critical functions and analysis of medical signals may fall outside that exclusion. In the European Union, AI systems used as safety components of regulated products can be classified as high risk under the AI Act when the relevant product requires third-party conformity assessment. The authors’ immediate research agenda calls for anesthesia-specific benchmark sets containing clinically meaningful omissions and hazards rather than only examination-style questions, standardized interfaces among models, electronic health records and anesthesia information systems, multicenter external validation, prospective human-factor studies, and trials testing whether assistance actually reduces errors, workload or postoperative complications. The relevant question, they conclude, is no longer whether a model can sound like an anesthesiologist, but whether a defined perioperative use, with a defined failure consequence, has accumulated sufficient evidence and safeguards to justify deployment.

Subject of Research: A risk-based framework for safe clinical adoption of large language models in perioperative anesthesia

Article Title: Large language models in perioperative anesthesia: a risk-based roadmap for safe clinical adoption

Article References: Tan, Z., He, Q., Li, Q., & Li, X. (2026). Large language models in perioperative anesthesia: a risk-based roadmap for safe clinical adoption. Anesthesiology and Perioperative Science, 4(1), Article 49. https://doi.org/10.1007/s44254-026-00196-9

Image Credits: AI Generated

DOI: 10.1007/s44254-026-00196-9

Keywords: large language models, anesthesia, perioperative medicine, artificial intelligence, clinical decision support, patient safety, risk stratification, human oversight, regulation, electronic health records, AI governance, hallucination