Large language models and conversational AI in nursing practice: a narrative review of English- and Chinese-language evidence
TL;DR
To synthesize representative evidence by nursing workflow, distinguish the inferences supported by different study designs, and propose task-specific safeguards for clinical evaluation, large language models are being evaluated across nursing services, but content quality, workflow performance, and patient benefit are often conflated.
Abstract
Large language models (LLMs) are being evaluated across nursing services, but content quality, workflow performance, and patient benefit are often conflated. English- and Chinese-language studies offer complementary examples of how these systems enter nursing workflows. To synthesize representative evidence by nursing workflow, distinguish the inferences supported by different study designs, and propose task-specific safeguards for clinical evaluation. This narrative review used PubMed, CNKI, and Wanfang Data for a primary-study publication window of November 1, 2022 to June 22, 2026, supplemented by reference tracing and targeted source verification during revision. Two authors jointly reviewed candidate publications and resolved inclusion decisions by discussion. Nursing education-only studies and chatbots without a verified generative LLM component were excluded from the core synthesis. Findings were interpreted at model, workflow, and patient levels without quantitative pooling or formal risk-of-bias scoring. The revised synthesis includes 23 representative studies, 13 published in Chinese, from China, Japan, Türkiye, Belgium, South Korea, and the United States. Applications covered patient education and consultation, decision support, documentation, and triage/reception and continuity of care. Most evaluations established content quality or technical feasibility. Three randomized trials addressed outpatient reception, preoperative anxiety, or a chatbot versus a nurse hotline; findings differed by endpoint, analysis population and follow-up. A small palliative-care thesis provided implementation and usability observations without demonstrating symptom benefit. Chinese-language studies added hospital-hosted and knowledge-enhanced systems, but this selected sample does not establish national deployment patterns. Evidence on sustained safety, nursing workload, equity and cost remained limited. The clearest current value of LLMs is support for bounded information and documentation tasks, with some context-specific service and patient-experience benefits. Oversight should follow the consequence of the output, automation, reversibility, and availability of human review. The proposed safeguards require prospective evaluation within the intended nursing service.