Who Controls AI Adoption? The WHO on Generative AI and Health System Divides

Sam: The World Health Organization just released a striking warning that health systems lacking AI expertise risk being exposed to fragmented, vendor-driven adoption, and today we are breaking down their strategy to fix that.

Maya: We are looking at a document titled ARTIFICIAL INTELLIGENCE, which is a draft guidance for responsible AI in health published by the World Health Organization Regional Office for the Western Pacific on 17 July 2026.

Sam: And before we dive in, a quick reminder that our voices are AI-generated, and this episode is a summary of a public document.

Maya: This document is fascinating because it is positioned as a provisional agenda item 9 for the 77th session of the Regional Committee taking place in Manila, Philippines from 19 to 22 October 2026. It is a direct response to artificial intelligence becoming a national political commitment for many Member States in the Western Pacific Region.

Sam: And they are looking at this through a very specific lens. They point out that ministries of health across the Western Pacific are confronting growing and interlinked system pressures. We are talking about demographic changes reshaping the demands for care, specifically ageing populations and the rising prevalence of noncommunicable diseases.

Maya: Exactly. Those demographic shifts are straining health services. But on top of that, many countries are facing workforce shortages and maldistribution. They have constrained fiscal space, fragmented data systems, and limited interoperability.

Sam: They also add climate change and emergencies to that list, noting that these have increased the frequency and complexity of public health threats. And they highlight that misinformation, along with broader social and commercial determinants of health, are further complicating decision-making and service delivery.

Maya: Which brings us to why this document exists. The WHO notes that these challenges weigh heavier on the Region's lower-middle income countries. Because of limited resources, health ministries have recognized they need innovation to expand access in remote areas, strengthen the workforce, and improve system effectiveness. And in that context, AI has emerged as a potential enabler of broader health system transformation.

Sam: But they make a really interesting distinction about the timeline of AI here. They mention that over the past decade, before the proliferation of generative AI, machine learning and deep learning were already being used.

Maya: Right, they call those task-specific applications. They give examples like augmenting epidemiological modelling for climate-sensitive diseases, supporting screening and diagnosis for high-burden conditions, digitizing paper-based health records, and improving specific aspects of service delivery.

Sam: But then generative AI arrives, and the WHO says it plays a much broader role. Because it is trained on vast archives of recorded information and knowledge, they say generative AI has moved beyond solving specific problems to providing more general-purpose support for health system functions.

Maya: And that shift to general-purpose tools that can influence clinical documentation and decision-making at scale is what is driving this new guidance. Because right now, well-resourced settings are already operationalizing AI in health systems, while many resource-constrained settings in the Region have yet to start.

Sam: They are very blunt about why that gap exists. They point to limitations in infrastructure, availability and quality of data, financing, governance, and regulatory oversight. But they also highlight the concerns coming out of those well-resourced settings that are already using AI.

Maya: Yes, early evidence emphasizes concerns related to performance and evidence gaps, safety, and system effects. The document specifically calls out risks of hallucination, bias, over-reliance on AI, and the potential deskilling of health workers.

Sam: That term, deskilling of health workers, is something we hear whispered about a lot, but seeing it documented by the WHO as a formal concern is striking. Yet, despite all these risks, their ultimate conclusion is not to hit the brakes.

Maya: Far from it. The document states explicitly that as AI technologies evolve and transform societies, they will continue to transform health systems. They write, the cost of not engaging AI will be high.

Sam: They frame it as an inevitability. They say health systems architects who understand the abilities and limitations of AI will be uniquely positioned to unlock the full potential of AI for system transformation. The question they pose is not whether ministries of health should adopt AI, but what actions should be taken today so that AI helps to deliver better health and well-being for people tomorrow.

Maya: To answer that question, the WHO developed 3 interconnected operational frameworks to shape the role of AI in public health. These were developed through expert and Member State consultations and validated through country workshops. The goal is to help countries start with use cases that are feasible today, while building the capacity needed for more transformative AI use over time.

Sam: Let us break those 3 frameworks down. The first one prioritizes use cases. The second helps determine the minimum necessary ecosystem. And the third helps build institutional and workforce capacity.

Maya: Starting with the first framework, prioritization. The WHO insists that the starting point for responsible adoption of AI is not the selection of a technology. It is the clear definition of the issue to be solved.

Sam: They give some great examples of those issues, like improving access to screening for high-burden diseases in underserved areas, or supporting frontline and village health workers in care delivery, or enabling more timely decisions informed by higher-quality data. The logic is that clearly defining the challenge determines the appropriate AI use case, which then determines the choice of technology.

Maya: Because as they point out, not every problem warrants an AI solution, and not every AI use case warrants public investment. They want ministries to conduct a problem-driven assessment of public value before even looking at specific technologies.

Sam: And once a use case emerges as a priority, they say it should be subject to tool-specific assessment. That means looking at evidence of validity and regulatory status before any implementation decisions are made. They ask ministries to balance benefits against risks.

Maya: Exactly. They want ministries to ask: Does addressing this problem with AI generate meaningful benefit relative to the actual standard of care available in the intended setting? They specifically emphasize looking at this for underserved populations.

Sam: They also talk about inherent risk. They define this based on the severity of harm if the AI-influenced decision is wrong, the degree of human oversight between the AI output and the impact, and the reversibility of the decision.

Maya: This is where the document gets highly practical for health-system leaders. They introduce a risk-proportionate approach. They acknowledge that across global AI governance frameworks, a low-, medium- and high-risk typology is widely used. But they argue that from a health system perspective, risk is not solely determined by the use case.

Sam: Right, they introduce this concept of implementation risk. They define this as the risk that arises from how, where, and by whom an AI system is deployed within a specific health system. They state that the exact same AI technology can present very different levels of risk depending on the deployment context.

Maya: The example they use is perfect for clinicians to understand. They describe an AI system for diabetic retinopathy screening. If you deploy that system as a productivity tool in a hospital setting where a specialist is providing oversight, the implementation risk is relatively low.

Sam: But if you take that exact same system and deploy it in a rural clinic that has no access to an ophthalmologist, the AI becomes the primary decision-maker. The WHO states that the implementation risk in that latter setting is higher, even though the underlying technology hasn't changed at all.

Maya: And here is the crucial caveat they add: the potential benefits in that higher-risk setting may also be greater, because the alternative might be no screening at all. That is why they say high risk does not necessarily mean do not deploy. It just means that stronger governance, validation, monitoring and accountability mechanisms are required.

Sam: That transitions perfectly into their second framework, which is the minimum ecosystem for implementation. They argue that technology is only as effective as the systems and governance within which it is embedded.

Maya: Quick note before we carry on. This spot is open for a sponsor. If your company builds or sells AI for healthcare and wants to reach the clinicians, health-system leaders and industry teams who listen to this show, the link to our sponsorship page is in the show notes.

Sam: And now, back to the document.

Maya: Yes, they say the end goal is not deployment of an AI solution; it is resolving the issue. And between the starting point and the ending point, an enabling ecosystem governed by the public sector is the critical success factor. This ecosystem requires aligned changes to care pathways, workforce roles, incentives, accountability arrangements, information systems, and service delivery models.

Sam: And they break this minimum ecosystem down into 5 distinct areas that need to be assessed against the specific AI solution and deployment context. The first area is governance.

Maya: For governance, they state there should be a clear public authority or function responsible for approving, overseeing, and when necessary, suspending the use of AI in health. And this oversight should be proportionate to risk. Lighter oversight for low-risk implementation, but higher-risk implementation requires clearer accountability, stronger review, and ethics or regulatory engagement.

Sam: The second area in the ecosystem is data and digital infrastructure. They emphasize that health data used by AI should be protected, secure, and fit for the intended purpose.

Maya: They are very strong on data sovereignty here. They state that countries should be able to control where data are stored, who can access them, how they are shared, and crucially, whether they are used to train or improve external systems. The arrangements need to support interoperability, privacy, and data sovereignty.

Sam: The third area is validation. There has to be evidence that the AI performs as intended in the setting where it will be used, or in comparable contexts. And for higher-risk implementations, they require stronger prospective evidence and monitoring of performance.

Maya: Because they want systems in place to detect errors, bias, unintended effects, or loss of effectiveness over time. AI models can drift, and performance can degrade, so continuous validation is part of the minimum ecosystem.

Sam: The fourth area is a massive one for the business side: financing. The WHO states clearly that funding should cover more than the initial procurement. Ministries are told to consider the full life cycle of implementation.

Maya: This is a vital point for companies selling into these systems. The WHO says life cycle costs include system integration, training, monitoring, cybersecurity, and any additional service demands generated by the AI application. If an AI screening tool suddenly uncovers a massive backlog of patients needing care, the system needs the financing to handle that service demand.

Sam: They state each implementation should have a clear pathway to sustainability, with decision points for continuation, scale-up, or discontinuation based on performance and value. They really want to avoid short-term pilots that lack long-term system integration plans.

Maya: And the fifth and final area of the ecosystem is capacity. They note that the institutions and individuals involved in governing, procuring, using, and overseeing AI must have the skills and authority to fulfil their roles. Higher-risk applications naturally require stronger technical, clinical, and managerial capacity.

Sam: Which leads us directly to the third framework: capacity-building. They want to ensure that all the actors involved can do what they need to do, in line with national priorities and ecosystem readiness. And they define core capacities for 4 different stakeholder groups.

Maya: First, for policy-makers and institutional leaders, the core capacity is stewardship. This encompasses understanding opportunities and challenges, setting priorities, and establishing those proportionate governance and regulatory pathways. They also have to define boundaries for data use, clarify accountability, and make informed decisions about the lifecycle of AI applications.

Sam: Then for planners, evaluators, and implementers, the priority is implementation readiness and evidence generation. This means having the ability to assess technologies, estimate those full life cycle costs we just talked about, and design and implement evaluations.

Maya: Third, for practitioners and providers, the focus is on safe and appropriate use in routine practice. Health workers need to understand both the intended use and the limitations of AI tools. Crucially, they need to know when human judgment should override AI outputs.

Sam: That override capability is key. They also need to be able to communicate AI-supported decisions to patients, and identify and report any errors or unintended effects they see on the ground.

Maya: Finally, the fourth group is patients, communities, and civil society. For them, the core capacity is meaningful participation. They need the ability to engage in decisions about whether and where AI is introduced, and what safeguards are required. They should also be involved in how benefits and risks are communicated, and how effects on access, equity, trust, and rights are monitored over time.

Sam: And across all these groups, the WHO lists foundational competencies that are required: people-centredness, ethical responsibility, critical appraisal, evidence-informed decision-making, communication, collaboration, and accountability.

Maya: Taking all of this together, the document translates these frameworks into 5 specific recommendations for Member States. The first recommendation is to lead with ministry-owned prioritization of AI use cases. They want ministries to create an AI road map that addresses national health priorities and advances equity, rather than allowing adoption to be driven primarily by vendor offerings, donor projects, or ad hoc institutional initiatives.

Sam: Their second recommendation is to establish a risk-proportionate governance pathway. Ministries should designate or strengthen mechanisms with the mandate to approve, oversee, and suspend the use of AI. And they note that where national regulatory capacity is limited, Member States may draw on reliance mechanisms and trusted external assessments, as long as they retain national decision-making authority.

Maya: Third, for each priority use case, they recommend ministries define and assess the minimum ecosystem required. This assessment should dictate whether implementation can proceed now, whether it should be staged with parallel system strengthening, or whether it should be deferred until key gaps are addressed.

Sam: Fourth is the financing recommendation we touched on: fund the ecosystem, not only the solution. Public investment decisions must cover the full implementation life cycle and any downstream service demands generated by AI use. Sustainability plans are required.

Maya: And the fifth recommendation is to build institutional and workforce capacity for AI, embedded into existing education and professional development systems, starting with the minimum capacities needed for priority use cases.

Sam: The document also outlines the role of the WHO Secretariat in supporting all of this. They state the WHO will assist Member States by turning global principles into practical national actions and facilitating shared learning to reduce duplication.

Maya: They list several specific actions for the Secretariat. First, developing operational tools for implementation, which may include prioritization tools, ecosystem assessment tools, and monitoring approaches aligned with health technology assessment principles.

Sam: Second, convening policy dialogue and intergovernmental exchange. They mention establishing a regional technical advisory mechanism to provide expert input on priority issues. Third, they will catalyze country implementation, providing technical assistance and supporting AI use cases where the public health value is high but market incentives or local capacity are insufficient.

Maya: Fourth, providing a regional learning platform. They want to support a regional catalogue documenting AI use cases, implementation risk classifications, evidence status, and lessons from implementation. They mention using existing platforms like the Pacific Open Learning Health Net and the Global Health Learning Centre for structured training pathways.

Sam: And fifth, building partnerships for public goods. The WHO wants to facilitate collaboration among ministries, academic institutions, implementers, civil society, technology developers, and open-source communities to develop regional public goods, like common AI models adapted to settings where they are needed in the Region.

Maya: The conclusion of this draft guidance is powerful. They state that AI is no longer an optional consideration for health systems in the Western Pacific Region. Used strategically, it can expand access and improve performance. But used poorly or not at all, countries run the risk of deepening inequities.

Sam: They reiterate that the central policy task is how to shape AI use so that it serves the public, creates value, and promotes equity and safety. And they leave Member States with this thought: The time to begin is not when every condition is perfect, but when countries have a clear public-interest purpose, adequate safeguards, and a commitment to monitor, learn, and adapt over time.

Maya: To recap, this WHO draft guidance provides a strategic framework for resource-constrained health systems to manage generative AI. It emphasizes that risk depends heavily on the implementation context, and it urges ministries of health to govern the adoption process strictly, fund the entire ecosystem rather than just the software, and build the capacity of both the workforce and the public.

Sam: A must-read for health-system leaders and developers looking at global regulatory baselines. The link to the full document is in the show notes. And as always, a reminder that this podcast is for informational purposes only and is not medical advice.