WHO Health AI Ethics Guidance

Overview

The "Ethics and governance of artificial intelligence for health: WHO guidance," published on June 28, 2021, represents the first global report of its kind, establishing a comprehensive ethical framework for the design, development, and deployment of artificial intelligence (AI) in the healthcare sector. Developed over an eighteen-month period of intensive deliberation, the guidance is the result of a multi-disciplinary effort involving a WHO Expert Group of 20 leading specialists in ethics, digital technology, law, and human rights, alongside contributions from various Ministries of Health. The document arrives at a critical juncture as digital transformation accelerates globally, particularly in the wake of the COVID-19 pandemic, which underscored both the immense potential of AI for public health surveillance and the significant risks associated with unregulated digital health interventions. The primary objective of the guidance is to ensure that AI technologies serve as a tool for health equity and universal health coverage rather than a driver of further disparity. The significance of this instrument lies in its transition from abstract ethical theory to actionable policy recommendations. While AI holds transformative promise for clinical diagnosis, drug discovery, and health system management, the WHO emphasizes that these benefits are not guaranteed. The document identifies a "socio-technical" approach, arguing that the ethics of AI cannot be divorced from the social, economic, and political contexts in which these systems are embedded. It specifically addresses the needs of low- and middle-income countries (LMICs), where the "digital divide" poses a risk that AI models trained on data from high-income populations may be ineffective or even harmful when applied to diverse global settings. By providing a set of consensus principles, the WHO aims to harmonize international efforts and provide a baseline for national regulatory bodies to build upon, ensuring that human rights and ethical considerations are "baked in" to the AI lifecycle from the very beginning.

Definitions

The guidance provides critical definitions to clarify the scope of AI in the health domain, moving beyond purely technical descriptions to encompass the ethical implications of these terms. It defines Artificial Intelligence not merely as a set of algorithms, but as a system capable of performing tasks that typically require human intelligence, such as pattern recognition, decision-making, and prediction. Within this, Machine Learning (ML) is identified as a subset of AI where systems "learn" from data to improve performance without explicit programming. A key distinction is made between "narrow AI," which is designed for specific medical tasks like interpreting radiological images, and the broader socio-technical systems that include the data pipelines, the human operators, and the institutional frameworks that govern their use. The document stresses that in health, AI must be understood as an assistive tool that operates within a "human-in-the-loop" or "human-on-the-loop" architecture, where human agency remains paramount. Furthermore, the document defines "Explainability" and "Intelligibility" as foundational requirements for medical AI. Explainability refers to the ability of a system to provide a human-understandable rationale for its outputs, which is essential for clinicians who must justify treatment decisions to patients. Intelligibility goes further, requiring that the logic of the AI system be understandable not just to developers, but to the end-users and regulators who oversee its application. The guidance also defines "Algorithmic Bias" as the systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others. This definition is particularly vital in health, where biased training data can lead to misdiagnosis or unequal access to care for marginalized populations. By standardizing these definitions, the WHO provides a common language for global stakeholders to discuss and mitigate the technical and moral risks of AI.

Governance and Institutional Framework

The institutional framework for this guidance is centered within the WHO Digital Health and Innovation Department, specifically the Health Ethics and Governance unit. This department is tasked with leading the organization's normative work on emerging technologies and supporting Member States in implementing the Global Strategy on Digital Health 2020–2025. The governance model proposed by the WHO is inherently multi-stakeholder, recognizing that no single entity can manage the complexities of health AI. It calls for the establishment of national and international oversight bodies that include representatives from government, academia, the private sector, and, crucially, civil society and patient advocacy groups. This inclusive approach ensures that the governance of AI is not captured by commercial interests and remains focused on the public good. The WHO itself acts as a central hub for knowledge sharing, providing technical support to countries as they develop their own national AI strategies. A significant component of the institutional framework is the role of regulatory agencies, which the WHO argues must evolve to meet the challenges of "black box" algorithms and autonomous systems. The guidance recommends that existing medical device regulators be empowered with new technical expertise and resources to conduct pre-market assessments and post-market surveillance of AI-based tools. Furthermore, the WHO promotes the creation of "AI Observatories" and "Collaborating Centres" (such as the one recently established at Delft University) to monitor the real-world impact of AI on health outcomes and ethical standards. This framework emphasizes that governance is not a one-time approval process but a continuous lifecycle of monitoring, evaluation, and adjustment. By fostering a global community of practice, the WHO ensures that lessons learned in one jurisdiction can inform the regulatory frameworks of others, promoting international alignment and trust.

Key Provisions

The core of the guidance consists of six consensus ethical principles designed to guide all stakeholders in the health AI ecosystem. The first principle is "Protecting Human Autonomy," which asserts that humans should remain in control of medical decisions. This includes the right to informed consent and the protection of privacy and confidentiality, ensuring that AI does not become a tool for coercion or unauthorized surveillance. The second principle, "Promoting Human Well-being and Safety," mandates that AI developers satisfy rigorous standards for safety, accuracy, and efficacy. It emphasizes that AI should only be deployed if it provides a clear benefit over existing practices and does not cause physical or mental harm. The third principle is "Ensuring Transparency, Explainability, and Intelligibility," requiring that the internal logic of AI systems be accessible to those who use and are affected by them, thereby fostering trust and enabling informed clinical judgment. The fourth principle, "Fostering Responsibility and Accountability," addresses the legal and moral challenges of AI-driven errors. It clarifies that while AI can assist in decision-making, the ultimate responsibility for clinical outcomes remains with human practitioners and the institutions that deploy the technology. It calls for clear liability regimes to ensure that individuals harmed by AI have access to redress. The fifth principle, "Ensuring Inclusiveness and Equity," is perhaps the most critical for the WHO's mission. It requires that AI be designed to be accessible to all, regardless of age, sex, gender, income, or ethnicity, and that it actively works to close the digital divide. Finally, the sixth principle, "Promoting AI that is Responsive and Sustainable," encourages the development of AI that is environmentally friendly and aligned with the broader goals of health system sustainability. Together, these provisions form a "moral compass" for the development of responsible AI in health.

Scope and Application

The scope of the WHO guidance is global and comprehensive, covering the entire lifecycle of AI technologies from initial research and data collection to deployment, commercialization, and eventual retirement. It applies to a wide range of AI applications in health, including clinical care (diagnosis and treatment), health research and drug development, public health surveillance, and health systems management (such as optimizing hospital workflows). Geographically, the guidance is intended for all 194 WHO Member States, but it places a special emphasis on the unique challenges faced by low- and middle-income countries. The document recognizes that these regions often lack the robust regulatory infrastructure found in high-income countries, making them more vulnerable to "ethics dumping" or the deployment of unproven technologies. Consequently, the guidance serves as a foundational template for LMICs to develop their own protective legal frameworks. In terms of application, the guidance identifies four primary target audiences, each with specific roles and responsibilities. First, for "Governments and Ministries of Health," it provides a roadmap for national policy and regulation. Second, for "AI Technology Developers," it offers an "ethics-by-design" framework to ensure that ethical considerations are integrated into the technical architecture of the software. Third, for "Health Care Providers and Workers," it provides guidance on how to use AI tools safely and ethically in clinical settings, emphasizing the importance of maintaining the patient-provider relationship. Fourth, for "Industry and Commercial Entities," it sets expectations for transparency, data sharing, and corporate responsibility. By addressing this diverse set of stakeholders, the WHO ensures that the guidance is not just a theoretical document but a practical tool that can be applied across the entire health technology sector.

Implementation Framework

The implementation framework of the WHO guidance is built around a set of practical tools and recommendations designed to translate ethical principles into daily practice. A central recommendation is the use of "Ethical Impact Assessments" (EIAs) and "Human Rights Impact Assessments" (HRIAs). These assessments should be conducted before an AI system is deployed and at regular intervals thereafter to identify potential risks such as bias, privacy violations, or negative impacts on the health workforce. The guidance provides checklists for developers and health ministers to ensure that these assessments are thorough and transparent. Furthermore, the WHO advocates for "Public Engagement" as a core implementation strategy, requiring that the communities most affected by AI—particularly patients and marginalized groups—be involved in the design and governance process to ensure the technology meets their actual needs. Another key pillar of the implementation framework is the promotion of "Technical Standards" and "Interoperability." The WHO encourages the adoption of global standards for health data to ensure that AI systems can work together across different platforms and jurisdictions. This is coupled with a call for "Capacity Building," recognizing that many health systems lack the digital literacy and technical expertise needed to manage AI effectively. The WHO commits to providing training and resources to help countries build this expertise. Additionally, the guidance suggests that implementation should be supported by "Regulatory Sandboxes," where new AI tools can be tested in a controlled environment under close supervision before being scaled up. This phased approach allows for innovation while maintaining strict safeguards for patient safety and ethical integrity.

Monitoring and Evaluation

Monitoring and evaluation (M&E) are treated as essential, ongoing processes within the WHO guidance to ensure that AI technologies remain aligned with ethical standards over time. The document emphasizes that because AI systems, particularly those using machine learning, can evolve as they process new data, a one-time "static" approval is insufficient. Instead, it calls for "Post-Market Surveillance" regimes that track the real-world performance of AI in clinical settings. This includes monitoring for "algorithmic drift," where a model's accuracy degrades over time, and "emergent bias," where a system begins to produce unfair outcomes as it encounters new populations. The WHO recommends that health systems establish clear reporting mechanisms for AI-related "adverse events," similar to those used for pharmaceuticals and traditional medical devices. The guidance also outlines the need for "Independent Audits" of AI systems. These audits should be conducted by third parties to verify that developers are adhering to their stated ethical principles and that the systems are performing as intended. The results of these audits should be made public to foster transparency and accountability. At the organizational level, the WHO commits to periodically reviewing and updating the guidance itself to reflect technological advancements and emerging ethical challenges, such as the rise of Large Multimodal Models (LMMs). This commitment to "Responsive Governance" ensures that the WHO remains at the forefront of the field. By establishing these M&E frameworks, the WHO provides a mechanism for continuous improvement, ensuring that AI for health remains a safe, effective, and ethical tool for all.

Relationship to Other Instruments

The WHO AI Ethics Guidance does not exist in a vacuum; it is deeply integrated with existing international human rights law and other global governance instruments. It explicitly builds upon the "Universal Declaration of Human Rights" and the "International Covenant on Economic, Social and Cultural Rights," asserting that the right to health and the right to enjoy the benefits of scientific progress are the legal foundations for ethical AI. The guidance also aligns with the "UNESCO Recommendation on the Ethics of Artificial Intelligence," which provides a broader cross-sectoral framework for AI governance. By grounding its health-specific recommendations in these universal principles, the WHO ensures that its guidance has the weight of international law behind it and contributes to a coherent global approach to AI ethics. Furthermore, the document is a key component of the "WHO Global Strategy on Digital Health 2020–2025," which provides the overarching strategic vision for the organization's digital initiatives. It also complements the "WHO Regulatory Considerations on Artificial Intelligence for Health" (2023), which provides more technical detail on the legal and regulatory requirements for AI medical devices. The guidance also references the "OECD Principles on Artificial Intelligence," particularly regarding transparency and accountability. By referencing and aligning with these various instruments, the WHO avoids "fragmentation" in AI governance and promotes a unified global standard. This interconnectedness allows Member States to leverage existing legal and policy frameworks when implementing the WHO's health-specific AI recommendations.

International Alignment

International alignment is a central theme of the WHO guidance, reflecting the reality that AI development and data flows are inherently cross-border. The WHO actively collaborates with other United Nations agencies through the "UN Inter-Agency Working Group on AI" to ensure that health ethics are represented in broader UN-wide AI initiatives. The guidance also seeks to align with the efforts of the G20 and G7, which have increasingly focused on the digital economy and the responsible use of data. By participating in these high-level forums, the WHO advocates for the "Global South," ensuring that the interests of low-income countries are not sidelined in international negotiations on AI standards. This alignment is crucial for preventing a "race to the bottom" in regulatory standards and for promoting the mutual recognition of AI certifications between countries. A key aspect of this alignment is the promotion of "Data Sovereignty" and "Equitable Data Sharing." The WHO guidance encourages international cooperation to create "Global Data Commons" for health, where high-quality, diverse datasets can be shared ethically to train AI models that work for everyone. This is aligned with the "Sustainable Development Goals" (SDGs), particularly SDG 3 (Good Health and Well-being) and SDG 10 (Reduced Inequalities). The WHO argues that without international alignment on AI ethics, the technology could exacerbate global health inequities. Therefore, the guidance serves as a call to action for international organizations, governments, and the private sector to work together to create a "Global Health AI Ecosystem" that is governed by shared values of solidarity, equity, and human rights.

Implementation Timeline

Adoption and Endorsement

Sources and References