Large language models in clinical nutrition practice: opportunities, risks, and governance

Abstract

Large language models (LLMs) are rapidly entering clinical workflows, including medical nutrition therapy (MNT) planning, patient education, and documentation support. Their promise lies in speed, personalization, and scalability; however, unsafe outputs can occur due to hallucinations, guideline drift, missing contraindications (e.g., renal/hepatic failure, pregnancy, pediatrics), and overconfident language. This practice-focused review synthesizes current evidence and real-world considerations for using LLMs in clinical nutrition and dietetic practice, with emphasis on patient safety, accountability, and governance. We map high-impact clinical use cases across inpatient and outpatient settings and summarize common failure modes and risk amplifiers (insufficient clinical context, poor prompt hygiene, lack of verification, and inadequate oversight). Current evidence supports broader use in low- to moderate-risk educational, communication, documentation, and population-level applications, whereas individualized therapeutic use in medically complex patients should remain restricted to non-autonomous, clinician-verified workflows. We then propose an implementation framework grounded in human-in-the-loop review, scope restriction, documentation/logging, periodic re-validation, and escalation pathways for red-flag conditions. The manuscript provides a clinic-ready “LLM Safety & Quality Checklist” to support dietitians and clinical teams in verifying outputs against authoritative guidelines, identifying high-risk cases requiring clinician review, ensuring transparent disclosure to patients, and safeguarding privacy. By translating emerging evidence into actionable steps, this paper aims to help dietetic services adopt LLMs responsibly while balancing innovation with patient safety, ethics, and regulatory readiness.