AI Regulation Tracker / Health and clinical AI
WHO Issues Ethics and Governance Guidance for Large Multimodal AI in Health
Non-binding, but the first global reference point for clinicians and health systems adopting general-purpose AI. On January 18, 2024, the World Health Organization published guidance on the ethics and governance of large multimodal models in health, with more than 40 recommendations for governments, developers, and providers.
This is a 2024 document, and I want to be honest about that up front. The WHO published this guidance on January 18, 2024, so nothing about it is breaking news. I am putting it in the tracker because it remains the standing international reference on how general-purpose AI should be governed in health, and because most US clinicians I talk to have never read it even though it is the framework that sits underneath a lot of the hospital AI policies now landing on their desks. Large multimodal models are the systems that can take in text, images, and other inputs and produce a wide range of outputs. In plain terms, this is the guidance for the ChatGPT-style tools that have arrived in clinical settings, not the narrow imaging algorithm cleared as a medical device.
The WHO is blunt about who has to act. In its own words, "Governments from all countries must cooperatively lead efforts to effectively regulate the development and use of AI technologies, such as LMMs." That framing matters because it puts the first-order duty on states, not on individual doctors. The guidance then asks governments to "use laws, policies and regulations to ensure that LMMs and applications used in health care and medicine ... meet ethical obligations and human rights standards." So the document is not pretending to be law. It is telling lawmakers what the law should require.
What does the WHO guidance actually cover?
The guidance maps five broad applications of LMMs in health and attaches recommendations to each. The five are diagnosis and clinical care, patient-guided use such as symptom checkers and patient-facing chatbots, clerical and administrative work like drafting notes and letters, medical and nursing education, and scientific research and drug development. That breadth is the point. Earlier health-AI governance mostly assumed a narrow tool doing one defined task. LMMs cross all five of those categories with a single model, which is exactly why the WHO wrote a dedicated document rather than folding this into its older AI guidance.
The recommendations run past 40 and split across the actors who touch these systems. Governments are asked to set standards, invest in public infrastructure, and mandate independent audits and impact assessments. Developers are asked to involve clinicians, patients, and health workers from the design stage rather than presenting a finished product. Providers and health systems are asked to train users and to keep humans accountable for clinical decisions. None of this is optional guidance dressed up as best practice for its own sake. It is the checklist that regulators and hospital governance committees are now borrowing from.
Is any of this binding on US physicians?
No, and it is important to say so clearly. WHO guidance is not law in the United States or anywhere else. It does not create a duty a physician can be sued for breaching on its own terms, and it does not override FDA authority, state medical-board rules, or hospital policy. What it does is supply the vocabulary and the expectations that those binding bodies increasingly adopt. When a hospital drafts an AI use policy, when a state board weighs whether a doctor exercised reasonable oversight of an AI tool, or when a plaintiff argues that a deployment was reckless, the WHO framework is the kind of recognized standard that gets cited. Guidance becomes the measuring stick long before it becomes a statute.
What should physicians and health systems do with it?
Read it as a governance baseline, not as paperwork. If you are a physician using a general-purpose model to draft notes, summarize records, or check your reasoning, the practical takeaways are simple and they track the guidance: keep a human accountable for every clinical decision, understand that the output can be wrong or biased, and do not let the tool quietly become the decision-maker. If you run a health system, the WHO recommendations give you a ready-made structure for an AI policy, covering procurement, clinician involvement, training, audit, and patient transparency. Aligning to it now is cheap. Retrofitting governance after an incident, or after a binding rule catches you flat, is not.
Questions professionals are asking
Is the WHO LMM guidance legally binding on doctors or hospitals?
No. It is WHO guidance, published January 18, 2024, with more than 40 recommendations. It creates no direct legal obligation on physicians, hospitals, or developers. Governments would need to legislate or regulate to give any part of it legal force. Its practical weight comes from being the recognized international reference that regulators and health systems adopt.
What are large multimodal models in a health context?
They are AI systems that can take in more than one type of input, such as text and images, and generate a wide range of outputs. In practice these are the general-purpose, ChatGPT-style tools now appearing in clinical and administrative work, as distinct from a narrow imaging algorithm cleared as a specific medical device.
Which uses of AI does the guidance address?
Five broad applications: diagnosis and clinical care, patient-guided use such as symptom checkers, clerical and administrative tasks, medical and nursing education, and scientific research and drug development. The guidance attaches recommendations for governments, developers, and providers to each.
How should a US health system use this guidance?
As a ready-made structure for an AI governance policy. It supports procurement standards, clinician and patient involvement in design, user training, independent audit, and patient transparency, while keeping a human accountable for clinical decisions. Aligning to it now is far cheaper than retrofitting governance after an incident or a binding rule.
RELATED BRIEFINGS
Browse the full AI Regulation News tracker
Informational analysis for working professionals, not legal or medical advice. Confirm how any guidance applies to your practice with qualified counsel and your institution's governance in the relevant jurisdiction.