WHO Ethics and Governance of Artificial Intelligence for Health
The World Health Organization (WHO) published its guidance on Ethics and Governance of Artificial Intelligence for Health in 2021, establishing a framework for the ethical development and deployment of AI technologies in healthcare. The document identifies six core principles that should guide AI in health contexts: protect autonomy, promote human well-being and safety, ensure transparency and explainability, foster responsibility and accountability, ensure inclusiveness and equity, and promote AI that is responsive and sustainable.
The Six Ethical Principles
Protecting autonomy means that AI systems should not override human decision-making—health professionals must retain the final say in diagnosis and treatment decisions, and patients must have the right to informed consent about AI involvement in their care. Promoting well-being and safety requires rigorous testing before deployment, continuous monitoring for harm, and regulatory oversight similar to medical devices. Transparency and explainability demand that AI systems be understandable to the clinicians and patients who use them—black-box systems that provide predictions without explanations are ethically problematic in health contexts where decisions affect life and death.
# Explainable AI example: SHAP values for medical diagnosis
import shap
import xgboost as xgb
model = xgb.XGBClassifier()
model.fit(X_train, y_train)
# Explain a single prediction
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_patient)
feature_importance = list(zip(feature_names, shap_values[0]))
feature_importance.sort(key=lambda x: abs(x[1]), reverse=True)
# Show top 3 factors influencing the diagnosis
for feature, impact in feature_importance[:3]:
direction = "increases" if impact > 0 else "decreases"
print(f"{feature}: {direction} risk by {abs(impact):.4f}")Key Challenges Identified by the WHO
Bias and fairness is a major concern: AI models trained on data from wealthy, predominantly white populations may perform poorly on marginalized groups. The WHO cites examples where dermatology AI trained primarily on light skin tones misdiagnoses skin cancer in darker skin. Data privacy is another critical issue—health data is highly sensitive, and AI systems that share data across institutions must implement robust de-identification, consent management, and security measures. Intellectual property rights for AI-generated discoveries (e.g., a novel drug molecule designed by an AI system) create legal gray areas that existing patent law does not fully address.
# Detecting dataset bias in health AI
def check_demographic_balance(dataset):
groups = dataset.groupby(["race", "age_group", "gender"]).size()
total = len(dataset)
underrepresented = []
for group, count in groups.items():
proportion = count / total
if proportion < 0.01: # Less than 1% representation
underrepresented.append((group, proportion))
return underrepresented
bias_report = check_demographic_balance(health_dataset)
for group, prop in bias_report:
print(f"WARNING: Underrepresented group {group} ({prop:.1%})")Governance Recommendations
The WHO recommends that governments establish regulatory frameworks for AI in health, requiring pre-market validation, post-market surveillance, and mandatory adverse event reporting. AI systems should be regulated as medical devices—the EU AI Act and FDA's evolving framework for AI/ML-based SaMD (Software as a Medical Device) provide emerging regulatory models. The guidance emphasizes that AI should complement rather than replace health workers, particularly in low-resource settings where AI could help address workforce shortages by assisting with triage, screening, and diagnostic support. Human oversight mechanisms must be built into every AI health system, with clear escalation paths when the AI encounters cases beyond its training distribution or confidence thresholds.
Global Implementation and Country Examples
Several countries have begun implementing AI ethics frameworks aligned with WHO guidance. The European Union's AI Act (2024) classifies health AI as "high-risk," requiring conformity assessments, human oversight, and transparency documentation before market approval. The US FDA has approved over 1000 AI-enabled medical devices through its De Novo and 510(k) pathways, with a growing emphasis on real-world performance monitoring after approval. China's Ministry of Health issued guidelines requiring AI diagnostic systems to undergo clinical validation in Chinese populations before deployment. India's NITI Aayog published a national AI strategy that prioritizes health applications while acknowledging the need for regulatory frameworks that protect privacy in a context where digital health ID systems are expanding rapidly. These national approaches vary in stringency but converge on the core WHO principles: AI in health must be safe, effective, equitable, and subject to human oversight. The WHO's global guidance provides a common language for international collaboration, enabling mutual recognition of AI system approvals and shared best practices for post-market surveillance across jurisdictions.
AI and Health Equity
The WHO guidance strongly emphasizes that AI should not exacerbate existing health inequities. In practice, this means ensuring training data represents diverse populations (not just data from wealthy urban hospitals), that AI tools are accessible in low-resource settings (offline-capable, low-bandwidth, affordable), and that deployment does not divert resources from proven public health interventions toward unproven AI solutions. Community engagement throughout the AI lifecycle ensures that AI addresses actual community needs rather than researcher interests. The WHO recommends that AI investments be accompanied by investments in digital infrastructure and health worker training to ensure that AI benefits reach all populations equitably.