Background <p>Large language models (LLMs) are being rapidly deployed in healthcare, often with emphasis on interpretive and decision-making capabilities. Persistent challenges with accountability, transparency and safety suggest that technical performance alone cannot address the risks these systems pose. </p> Objective <p>To propose a design hypothesis that reframes healthcare AI development around explicit scope limitation. </p> Methods <p>This article presents a conceptual perspective synthesising existing literature in healthcare AI, human-AI interaction and sociotechnical systems theory to examine how mismatches between system scope and professional responsibility may contribute to risk. </p> Results <p>The analysis suggests that constraining LLMs to translational tasks – such as restructuring information between clinical documentation, scientific literature, regulatory frameworks and patient-facing language – may reduce ambiguity about authority and improve auditability. </p> Conclusions <p>A boundary-respecting AI approach reframes constraint as a design safeguard. By limiting interpretive authority and clarifying task boundaries, healthcare AI systems may achieve safer and more accountable integration into healthcare workflows.</p>