Artificial intelligence is moving quickly into the interpretation of CT, magnetic-resonance and ultrasound images, but China’s newest ethics guidance draws an unambiguous boundary around who remains responsible when the technology is used in research or clinical care. The rules say an AI output cannot serve as the final diagnosis. A physician must retain the final decision, and human beings remain ultimately accountable.
Xinhua reported on August 29 that China’s national medical-research ethics body had issued new guidance for AI imaging. The document covers work from designing an algorithm through validating it and eventually using it in a clinical setting. Its appearance is notable because imaging is one of the most immediate healthcare uses of AI: machines can help clinicians identify lesions and process scans, yet the same systems require large volumes of highly sensitive health data.
The document does not reject medical AI. It establishes conditions for using it. Its six principles are to advance human welfare, support fairness, respect human autonomy, protect privacy and data security, ensure safety and controllability, and strengthen transparency. That set of principles connects clinical accuracy with the rights of people whose records and images make the systems possible.
Doctors Retain Authority Over AI Medical-Imaging Diagnoses
The clearest operational requirement is that AI-generated results are not final diagnoses. Doctors retain decision-making authority. That distinction prevents a hospital or researcher from treating a model score as a substitute for clinical judgment when a patient’s treatment depends on the result. It also clarifies accountability when an output is wrong, biased or insufficiently explained.
Medical images may appear more standardized than conversational data, but they are not neutral inputs. A CT scan, an MRI or an ultrasound image is tied to an individual’s health information. The guidelines require researchers to secure informed consent from participants. They also state that participants can withdraw authorization. These provisions make data governance a continuing responsibility rather than a one-time consent form collected at the start of a project.
The framework also addresses performance claims. Researchers are prohibited from overstating what a model can do, and the rules warn about diagnostic bias created by imbalanced training data. A system trained primarily on one patient population, one device type or one hospital’s imaging protocols may not perform equally well elsewhere. The document’s fairness requirement therefore reaches beyond privacy: it asks researchers to consider whether the data behind an apparently accurate model leave some patient groups at a disadvantage.
China previously encouraged technical development in this field. An April EastFrontier article on a national AI medical-imaging diagnostics competition showed how public institutions were promoting tools that could support diagnosis. The new guidance covers a different question. It establishes the ethical conditions for creating, testing and applying such systems once they reach research or clinical environments.
Consent, Privacy and Bias Become Design Requirements
The guidelines make privacy and data security two of their six named principles. That matters because AI imaging research often pulls data from hospital workflows in which the goal is rapid clinical use, not merely academic analysis. A developer needs images to train and validate a model. A hospital needs systems that do not expose patient information or create unclear channels for secondary use. The framework places those interests inside the design and validation process.
Participants’ right to withdraw consent has practical consequences. It means teams need a way to track which images and records are authorized for which research purpose, and a way to respond if permission changes. The document does not lay out a particular technical architecture for doing that. Instead, it defines the rights and obligations that researchers must respect while choosing their own technical procedures.
Safety and controllability also have a specific clinical meaning. An imaging model can flag a possible lesion or help sort a queue of scans. It is riskier when users assume the system’s recommendation is conclusive. By requiring a human physician to make the final call, the guidance preserves a decision point where clinical context, patient history, and the tool’s limits can be considered together.
The broader policy direction is consistent with China’s approach to AI systems that affect people directly. In July, EastFrontier examined how AI companion rules took effect, illustrating regulatory concern about systems that can influence vulnerable users. Medical imaging raises a different set of risks, but the policy logic is similar: an AI provider cannot simply point to a model output and disclaim responsibility for how that output affects people.
Research Rules Shape the Route to Clinical Adoption
The new rules arrive while Chinese hospitals, universities and companies continue to explore AI-assisted imaging. Their immediate impact will be on researchers and organizations that develop algorithms, organize clinical validation, recruit participants and prepare a system for use. Requiring consent, monitoring bias and retaining physician authority adds procedural work, but it also gives patients and clinicians a more defined basis for trust.
The document is not a product approval, and it does not say that a specific model may diagnose a particular disease. It is a governance framework. Its significance lies in putting ethical requirements in place before an AI result is relied upon in a clinical decision. For developers, the message is that validation cannot be limited to a benchmark score. The relevant questions include whether the data were obtained properly, whether performance claims are accurate, and whether doctors can remain accountable in the final workflow.
For hospitals, the rules provide a standard for evaluating vendors and internal projects. A useful system will need more than an impressive demonstration. It will need an approach to consent, privacy, data security, fairness and oversight that clinicians can explain to patients. For patients, the guidance formally affirms that access to technology does not remove their right to control authorization or their expectation that a qualified person remains responsible for the conclusion.
China’s medical-imaging guidance does not solve every challenge associated with AI in healthcare. It does, however, set a clear baseline at a moment when models are becoming more capable and more widely tested. Medical AI can assist with interpreting scans. Under the new framework, it cannot become the final clinician, the final custodian of patient data or the final bearer of responsibility.