WHO Releases First Global Guidelines on AI in Health

Published 2025-02-18 · Policy

The World Health Organization released its first comprehensive global guidelines on the use of artificial intelligence in healthcare on February 16, 2025, establishing a framework that the organisation hopes will serve as a baseline for national regulators worldwide. The 94-page document, titled "Ethics and Governance of Artificial Intelligence for Health: WHO Guidance," is the product of a two-year consultation process involving 84 experts from 31 countries, 14 regulatory agencies, and 22 academic institutions.

The guidelines arrive at a moment when AI in health is proliferating faster than regulatory frameworks can adapt. The WHO estimates that over 500 AI-based health tools have entered the market globally since 2020, ranging from diagnostic imaging algorithms to drug discovery platforms to patient-facing chatbots. Most of these tools are regulated — if at all — under medical device frameworks that were designed for conventional software and hardware, not for adaptive algorithms that learn and change over time.

Core Recommendations

The guidelines are organised around six core principles, each accompanied by specific recommendations addressed to governments, developers, healthcare providers, and international organisations.

1. Mandatory Bias Auditing

The WHO recommends that all AI health tools undergo prospective bias audits before deployment and at regular intervals thereafter. The audits should assess whether the tool's performance varies across demographic groups defined by age, sex, gender identity, ethnicity, socioeconomic status, and geographic location. The guidelines specify that audits should use validated statistical methods and should be conducted by independent third parties where feasible.

The recommendation responds to a growing body of evidence that AI health tools can perpetuate and amplify existing healthcare disparities. A 2023 study in JAMA Internal Medicine found that a widely deployed commercial algorithm for predicting healthcare needs systematically underestimated the illness severity of Black patients, resulting in fewer referrals for specialist care. The WHO guidelines cite this study as a cautionary example and recommend that training datasets be audited for representativeness before model development begins.

2. Patient Consent for AI-Assisted Diagnoses

Patients should be informed when an AI system contributes to their clinical assessment and should have the right to request a human-only evaluation, the guidelines state. This recommendation goes further than the consent requirements in most existing regulatory frameworks, which typically mandate disclosure only when an AI system is the primary decision-maker rather than a supplementary tool.

Dr. Ren Minghui, the WHO's Assistant Director-General for Universal Health Coverage, described the consent requirement as fundamentally about trust: "Patients will not accept AI in their care if they feel it is being imposed without their knowledge. Transparency is not just an ethical obligation — it is a practical prerequisite for adoption."

3. Training Data Transparency

The guidelines call on AI developers to publish detailed documentation of the datasets used to train health AI models, including the size, source, demographic composition, and known limitations of each dataset. Where proprietary considerations prevent full public disclosure, the WHO recommends that a minimum dataset description be made available to regulators and that independent auditors be granted access to full documentation under confidentiality agreements.

This recommendation aligns with emerging regulatory practice. The EU AI Act requires high-risk AI systems to be accompanied by detailed data governance documentation, and the US FDA's proposed framework for regulating AI-based software as a medical device includes similar data transparency provisions. The WHO guidelines aim to establish a global minimum standard that applies regardless of jurisdiction.

4. Post-Market Surveillance

Once deployed, AI health tools should be subject to continuous monitoring for performance degradation, adverse events, and emerging biases. The WHO recommends that developers implement automated performance monitoring dashboards and that national regulators establish adverse event reporting systems specific to AI tools — analogous to the pharmacovigilance systems that exist for pharmaceuticals.

The recommendation addresses a specific vulnerability of adaptive AI systems: model drift. Unlike conventional medical devices whose performance is fixed at the time of manufacture, machine learning models can degrade as the data they encounter in deployment diverges from their training data. The guidelines recommend that developers specify acceptable performance thresholds and that regulators have the authority to require remedial action when thresholds are breached.

5. Human Oversight

The guidelines affirm that AI should augment, not replace, human clinical judgement. No AI system should be deployed in a context where it makes autonomous clinical decisions without the ability for a qualified healthcare professional to review, override, or discontinue its recommendations. The WHO acknowledges that this principle may evolve as evidence accumulates, but maintains that current evidence does not support fully autonomous AI diagnosis or treatment in any clinical domain.

6. Equity and Access

The final principle addresses the risk that AI health tools could widen rather than narrow global health disparities. The guidelines recommend that developers prioritise applications addressing the health needs of low- and middle-income countries, that pricing models be designed to avoid creating two tiers of care, and that open-source licensing be considered for tools developed with public funding.

Reception and Criticism

The guidelines have been broadly welcomed by public health advocates and regulatory bodies, though some industry voices have expressed concern about the practical implications. The Advanced Medical Technology Association (AdvaMed), which represents over 400 medical device companies, issued a statement supporting the principles while cautioning that "overly prescriptive requirements could slow the pace of innovation and delay patient access to beneficial technologies."

Dr. Sridhar Venkatapuram, a bioethicist at King's College London who served on the WHO expert panel, pushed back on this framing: "The question is not whether regulation slows innovation. The question is whether unregulated innovation harms patients. The history of medical technology is replete with examples of tools that were marketed prematurely and caused significant harm — from the Dalkon Shield to metal-on-metal hip implants. AI is not exempt from this pattern."

Some researchers have noted that the guidelines lack enforcement mechanisms. The WHO does not have regulatory authority over national governments; its guidelines are advisory, not binding. Whether the recommendations translate into actual regulatory change will depend on individual countries' willingness to adopt them. Early signals are mixed: the EU's AI Act incorporates many of the WHO's principles, as do proposed regulations in Brazil and South Korea. The United States has not announced plans to adopt the guidelines, though the FDA's existing approach to AI-based medical devices overlaps with several WHO recommendations.

Implications for Specific Technologies

The guidelines have particular relevance for several categories of AI health tools that are currently in active development or early deployment. Large language model-based diagnostic assistants, such as those studied by Stanford researchers for depression screening, would need to undergo bias auditing and patient consent procedures under the recommended framework. AI-driven drug discovery platforms like AlphaFold 3 and NVIDIA BioNeMo would fall under the training data transparency provisions. And patient-facing tools like the Woebot anxiety therapeutic would need to satisfy the human oversight and post-market surveillance requirements.

The WHO has committed to updating the guidelines every two years as evidence accumulates and technology evolves. The next edition, planned for 2027, is expected to address the use of generative AI in clinical settings — a category that has exploded since the release of ChatGPT in late 2022 and that the current guidelines address only at a high level.

For broader coverage of AI regulation and policy, visit our AI Ethics and Policy research repository. Related articles include EU AI Act enforcement for health companies and our analysis of Australia's AI mental health triage programme.

← Back to Publications