Risks and benefits of artificial intelligence across the perioperative surgical pathway: a review

Artificial intelligence (AI) is increasingly evaluated for risk prediction, decision support, and monitoring across surgical care, but improved model performance does not necessarily prevent patient harm. This narrative review asks whether perioperative AI reduces preventable harm, which new hazards it may introduce, and what evidence is required to establish safe clinical use. We distinguish demonstrated patient outcomes from intermediate physiological, predictive, and operational measures. A structured narrative literature search was undertaken in PubMed/MEDLINE, Scopus, Embase, and Web of Science to identify relevant perioperative AI publications. The Scale for the Assessment of Narrative Review Articles (SANRA) was used as a quality-appraisal framework for the narrative review, rather than as a study-selection or systematic-review reporting methodology. The search and selection process was descriptive rather than systematic; no formal risk-of-bias assessment or pooled effect estimation was performed. Of the 90 publications included in this narrative review, 68 (75.6%) were identified through perioperative literature searches and 22 (24.4%) through targeted searches of non-surgical healthcare settings, highlighting the distinction between direct perioperative evidence and findings extrapolated from other clinical contexts. Across safety-oriented applications, evidence frequently concerns predictive or intermediate endpoints rather than demonstrated reductions in surgical harm. Examples of AI-related hazards, including missed sepsis, alert burden, inequitable prediction, and generated-text errors, are partly derived from non-surgical settings and should not be interpreted as direct evidence of their incidence or impact in perioperative patients. The clearest patient-outcome evidence came from a nonrandomized before–after implementation study in colorectal cancer surgery, which reported fewer severe complications, and randomized trials summarized in a meta-analysis that reported lower post-anesthesia care unit pain scores with the Nociception Level Index; randomized trials of hypotension prediction reported intermediate physiological endpoints with mixed results. The synthesis distinguishes evidence from surgical populations from transferable safety lessons requiring perioperative validation. Direct evidence that perioperative AI prevents patient harm remains limited. Prospective impact evaluation, ideally randomized when feasible and using patient-safety endpoints, is needed to establish benefit and characterize AI-related harms.