In October 2025 the General Offices of the National Health Commission and the National Development and Reform Commission, together with other ministries, jointly issued an implementation opinion on promoting and regulating “AI plus healthcare” applications, implementing the “AI plus” initiative.
A staged set of targets
The opinion sets a clear timetable: by 2027, establish a group of vertical clinical large models for specific diseases and specialties; by 2030, near-universal AI-assisted primary care, widespread AI-assisted medical imaging diagnosis at secondary hospitals and above, and an established AI application standards system.
This “specific diseases first, universal coverage later” path reflects the practical constraints of medical AI: general models struggle to meet clinical requirements for accuracy and explainability, while vertical models for specific conditions are more feasible.
Three core compliance questions
Access. Whether a medical AI product is a medical device and what approval it requires depends on its intended use and risk level. Software for assisted diagnosis and tools for health education fall under different rules. See our note on AI regulation of drugs and devices.
Data. Medical data is sensitive personal information, requiring a higher lawful basis. The opinion mentions high-quality datasets and trusted data spaces, pointing to technical and institutional paths for “usable but not visible” data. See our notes on data element market allocation and obligations of major personal information processors.
Liability. The assistive positioning of clinical decision support means the final judgment rests with the physician, but when model suggestions are biased, how liability is divided among hospital, physician and vendor must be settled in advance in procurement agreements and clinical workflows. See our note on AI commercial contract clauses.
What makes healthcare special
The key difference from other settings is that the consequences of error are irreversible. This imposes three requirements:
Higher explainability. Physicians need to understand the basis for a model’s suggestion, not merely accept the conclusion.
Stricter logging. Model inputs, outputs and adoption should be traceable for review and attribution.
Finer safety assessment. Medical AI assessment looks not only at content safety but at diagnostic accuracy, bias and failure modes. See our note on the AI application service provider action plan.
Practical suggestions
First, design products around specific diseases and specialties, entering through clinical value in a vertical scenario rather than chasing general capability.
Second, front-load the data compliance plan, clarifying data sources, authorisation chains and use boundaries. See our data compliance service.
Third, specify liability allocation and validation duties in contracts, especially re-validation after model updates.
The policy window for medical AI is open, but the access threshold and compliance cost are equally clear. See our note on the overall framework of “AI plus” and the State Council’s “AI plus” initiative.
This article is based on the Implementation Opinion on Promoting and Regulating “AI Plus Healthcare” Applications and on public information. It is for general reference only and does not constitute legal advice.