JMIR Publications

ABSTRACT

Medical textbook translation remains a bottleneck in global health knowledge dissemination, typically requiring 4 to 8 weeks by physician-translators who possess both clinical expertise and bilingual fluency. We present an orchestrated multi-model AI translation pipeline that completed the localization of a 158-page Japanese dermatology textbook (conversational genre, A5 format) into Korean in two calendar days of active pipeline execution — following a separate multi-day preparation phase for terminology extraction and database registration. The pipeline is built on three architectural principles: (1) a cross-model validation constraint, in which the AI model that produces the translation is never the model that validates it — an operationalization of the well-established separation-of-duties principle from AI safety; (2) a seven-layer progressive quality assurance system that filters inexpensive-to-detect errors upstream before engaging costly validation downstream; and (3) cumulative terminology databases shared across five sequential book projects, accelerating each successive translation. During the two-day execution, the physician-translator typed zero characters of translation text but made over 400 quality judgment decisions, including confirming 325 terminology entries, evaluating 28 translator note proposals, and approving 447 individual text corrections identified by cross-model review. We argue that the cross-model validation constraint — where the producer and evaluator of medical AI content must be different systems — should become a standard design requirement for AI-generated content in healthcare, and that the physician-translator's role is shifting from text production to quality judgment.