AI-generated patient instructions as safety-critical communication: a provisional evidence-informed framework for clinical deployment

Generative artificial intelligence (AI) can convert clinical information into patient-facing instructions, including discharge summaries, medication explanations, portal messages and plain-language educational materials. These outputs may improve accessibility and reduce documentation burden. They also create a distinct patient-safety problem: patients may act on fluent, incomplete, or contextually unsafe instructions without the clinical knowledge required to detect error. Early evaluations show a recurring trade-off. Large language models can improve readability and understandability of discharge information, yet physician and pharmacist review has identified omissions, inaccuracies, newly introduced actions, medication-related problems and potentially harmful safety issues, particularly in complex discharge contexts. This narrative and interpretative review reframes AI-generated patient instructions as a safety-critical informatics intervention rather than a language-simplification tool. A targeted literature mapping was undertaken across empirical studies of AI-generated discharge communication, health-literacy and medication-safety literature, patient-safety evidence, and emerging artificial-intelligence governance frameworks. I propose a sevendomain safety framework covering factual accuracy, clinical completeness, actionability, medication clarity, escalation and safety-netting, health-literacy alignment, and accountability with auditability. The framework is intended to support implementation, local evaluation and reviewer assessment rather than to function as a formal consensus guideline. High-risk patient-facing outputs require accountable clinical workflows, with human verification, traceability, equity testing and post-deployment monitoring. Before wider use, evaluation needs to assess not only readability, but also comprehension, actionability and potential harm.