Abstract
<title>Abstract</title> <p>Background Generative artificial intelligence (AI), including systems enabled by large language models (LLMs), is being introduced into health care to reduce documentation burden and support information work. These technologies can also redistribute work into verification, correction, explanation, incident management, and organizational assurance. Their effects therefore extend beyond technical performance to professional working conditions, accountability, patient agency, equity, and service quality. Objective The objective was to explain how authoritative public health documents configure responsible generative AI adoption as a sociotechnical work-system intervention. Methods A qualitative document analysis examined 10 English-language public policy, regulatory, safety, and implementation documents issued by international, European, British, and Australian authorities between 2024 and 2026. Maximum-variation purposive sampling covered governance levels, documentary genres, intended audiences, and analytical functions. Directed content analysis combined the Sittig-Singh sociotechnical model with the NASSS framework. The analysis included structured extraction, preservation of modal strength, cross-document comparison, negative-case analysis, and construction of a theme-by-document evidence map. Results Five themes were identified. First, benefits were framed as conditional augmentation rather than automatic consequences of model capability. Second, automation produced a verification-work paradox because safe use depended on checking and correction that consumed the same time and attention the technology was expected to save. Third, responsibility was compressed toward professionals at the point of care even when technical control was distributed across suppliers and organizations. Fourth, responsible adoption required organizational lifecycle capability spanning selection, integration, training, monitoring, change control, and withdrawal. Fifth, patient agency and equity were operational work-system requirements involving disclosure, practical choice, accessibility, population fit, and alternatives. Conclusions Responsible adoption is best understood as governed augmentation. Generative AI can support professional and organizational capacity, but value and safety depend on bounded use, competent and resourced verification, accountability matched to control, organizational ownership across the lifecycle, practical patient agency, and continuous evaluation of work-environment and service outcomes.</p>