The World Health Organization (WHO) has released guidance on the ethics and governance of large multimodal models (LMMs), setting out how governments, technology companies and healthcare providers should approach AI systems that can work across text, images and video.
The document includes more than 40 recommendations intended to support appropriate use of LMMs in ways that promote and protect public health. The central message is practical: these tools may be useful in medicine, but they need clear rules, defined responsibilities and careful oversight before they are trusted in health-related decisions.
Why large multimodal models matter in health
LMMs are different from narrower AI tools because they can process several kinds of information and produce several kinds of output. In a medical setting, that matters because healthcare work often involves mixed information: written notes, patient questions, images, video and structured records can all be part of the same workflow.
That flexibility is the reason the WHO is focusing on ethics and governance now. A model that can respond to patient questions, support documentation or assist research may be useful, but the same flexibility also creates more points where errors, bias or misuse can enter the system.
Dr. Jeremy Farrar, Chief Scientist at WHO, stresses the need for transparent information and guidance around how LMMs are designed, developed and used. The goal, as described by WHO, is not only better health outcomes, but also the reduction of existing health inequalities.
Five areas where WHO sees medical use
The WHO guideline identifies five broad health applications for LMMs. Together, they show how widely these systems could touch medicine if they are adopted across healthcare settings.
- Diagnosis and clinical care, including responding to patients’ written queries.
- Patient-guided use, including investigating symptoms and treatment.
- Clerical and administrative tasks, including documenting and summarizing patient visits within electronic health records.
- Medical and nursing education, including simulated patient encounters for trainees.
- Scientific research and drug development, including identifying new compounds.
These categories range from direct patient interaction to behind-the-scenes administrative work. That range is important because the risk profile changes depending on the task. A model used to summarize a visit inside an electronic health record is not being used in the same way as one helping a person investigate symptoms and treatment.
The WHO guidance does not treat all use cases as equal. Instead, it points toward a more controlled approach: LMMs should be developed for well-defined tasks, and those tasks should meet the required levels of accuracy and reliability.
The risks are not only technical
The WHO highlights documented risks that LMMs may provide false, inaccurate, biased, or incomplete information. In medicine, that kind of output can be harmful when people use it to make health-related decisions.
One concern is the quality of the data used to train these systems. If an LMM is trained on low-quality or biased data, those weaknesses can influence what the model produces. That matters in healthcare because biased or incomplete information can affect how symptoms, care options or patient needs are interpreted.
The guidance also points to broader risks for health systems. These include the accessibility and affordability of the most capable LMMs, automation bias and cybersecurity risks. In other words, the issue is not only whether a model can produce a useful answer. It is also whether health systems can access the best tools fairly, protect themselves from digital threats and keep human judgment in the loop.
Automation bias is a particular concern in clinical settings. The WHO describes it as a situation in which healthcare professionals miss errors they would otherwise have caught, or inappropriately delegate difficult decisions to an LMM. That risk cuts to the core of medical responsibility: AI can assist, but the guidance warns against letting the tool become an unexamined authority.
Governance has to include more than developers
WHO says safe and effective LMMs require involvement from many stakeholders throughout model development and implementation. That includes governments, technology companies, healthcare providers, patients and civil society.
This is a significant governance point. The source of expertise cannot be limited to scientists and engineers. Developers of LMMs should involve potential users as well as direct and indirect stakeholders, because the effects of medical AI extend beyond the team building the model.
For governments, the recommendations include investing in not-for-profit or public infrastructure. The guideline also calls for laws and regulations that uphold ethical obligations and human rights standards, along with regulatory bodies and mandatory post-publication review and impact assessment.
That approach frames medical LMMs as part of public health infrastructure, not just software products. If a model is used in healthcare, WHO’s guidance suggests that its responsibilities continue after release. Review and impact assessment are presented as necessary parts of governance, not optional extras.
Built on earlier WHO AI guidance
The new LMM guidelines build on the WHO Guidelines on Ethics and Governance of AI for Health, published in June 2021. The connection matters because it places multimodal AI inside a broader health AI framework rather than treating it as an isolated technology trend.
AI is already being used in several areas of medicine, including diagnostics in various disciplines, psychotherapy and drug development, and it is showing initial success. The WHO guidance does not reject that progress. Instead, it sets boundaries for how more capable multimodal systems should be designed, governed and used.
The practical lesson is straightforward: LMMs may become useful across clinical care, administration, training and research, but healthcare adoption needs more than technical performance. It requires transparency, defined tasks, stakeholder involvement, attention to bias and affordability, and governance that can respond after systems are deployed.