Large language models in critical care documentation and communication: translating emerging technology into safe practice

Critical care generates dense streams of clinical language through progress notes, handovers, transfer and discharge summaries, communication with patients and families, and documentation of high-stakes decisions. Large language models (LLMs) increasingly support these tasks, but their fluent output can conceal clinically important errors. This narrative review examines the use of LLMs for clinical documentation and summarization, handover and interprofessional communication, communication with patients and families, and documentation of sensitive decisions in critical care. Early evidence indicates potential improvements in efficiency, readability, and information organization. However, reported risks include hallucinated or unsupported content, omission of salient details, loss of uncertainty and disagreement, variable performance across workflows, and unresolved questions about authorship and liability. This review proposes the intensive care unit (ICU) Clinical Communication Integrity Framework, which comprises five safeguards: source fidelity, context preservation, audience adaptation, human authorship and accountability, and prohibition of the autonomous generation, finalization, or transmission of high-stakes communication. These safeguards aim to retain the efficiency benefits of LLM-assisted drafting while preventing generated language from distorting the clinical record or diluting professional responsibility. Prospective, workflow-based evaluation is required before routine deployment in critical care.