The Six Ethics Principles the WHO Set for AI in Health Care

In June 2021 the World Health Organization published its first global report on artificial intelligence for health and named six ethics principles to guide the design and use of these tools: protect human autonomy; promote human well-being, safety, and the public interest; ensure transparency and explainability; foster responsibility and accountability; ensure inclusiveness and equity; and promote AI that is responsive and sustainable. The value of these six is not that they settle any single deployment. Read as a set of questions rather than slogans, they give clinicians, buyers, and patients a shared way to interrogate a product before trusting it with a decision.

This article treats the principles as an appraisal lens. None of them, on its own, tells you whether a specific algorithm is good enough for a specific job. Used together, they expose where the marketing stops and the evidence begins.

Where the principles come from

The 2021 guidance was assembled over roughly eighteen months by a WHO expert group spanning ethics, law, human rights, and digital technology, and it was published on 28 June 2021 under the title Ethics and governance of artificial intelligence for health. The document is deliberately framed for a global audience, which is why several principles reach beyond the exam room into questions of access, oversight, and environmental cost. In January 2024 the WHO extended this work with separate guidance on large multi-modal models, the generative systems that accept text and images and produce open-ended outputs, adding more than forty recommendations for governments, developers, and health providers. The 2024 document does not replace the six principles; it applies them to a newer and messier class of tool.

The six, read as questions

Protect autonomy

The first principle keeps humans in control of medical decisions and protects informed consent, privacy, and confidentiality. As an appraisal question it becomes concrete: does the tool preserve a clinician's ability to override it, and does it make the patient's role in the decision clearer or dimmer? A system that quietly narrows what a clinician or patient can choose fails this test even if its outputs are accurate.

Promote safety and well-being

Here the WHO asks that tools not harm people and that designers meet regulatory requirements for safety, accuracy, and efficacy for well-defined uses. The load-bearing phrase is well-defined. A model validated to flag one finding on one type of scan in one population has not been shown safe for anything else. Much of the risk in health AI comes from tools used outside the narrow setting where they were actually tested.

Ensure transparency and explainability

The principle asks that enough information be published or documented before a tool is deployed, so that it is intelligible to the people who build and use it. This is not a demand that every model be fully interpretable, which is often impossible. It is a demand for disclosure: what data trained it, on whom, with what known failure modes. When a vendor cannot answer those questions, that silence is itself an appraisal finding.

Foster responsibility and accountability

Someone must remain answerable when a tool contributes to harm, and there must be mechanisms for questioning a result and seeking redress. Automation can blur this. If no human or institution can be identified as responsible for a bad output, the accountability principle has not been met, regardless of how the liability language reads. The useful question is simple: when this tool is wrong, who is answerable, and how does a patient contest it?

Ensure inclusiveness and equity

The WHO asks that AI be designed to encourage the widest possible equitable use, regardless of age, sex, income, or geography. Equity failures in health AI are usually quiet. A model can perform well on average while performing poorly for a subgroup that was thin in its training data. The appraisal question is whether performance was measured across the groups who will actually be exposed to the tool, not only in aggregate.

Promote sustainability

The final principle is the one most often skipped. It asks that systems be responsive, monitored after deployment, and mindful of environmental cost and energy use. Responsiveness matters because a model can drift as clinical practice and populations change; a tool accurate at launch is not guaranteed accurate two years later without monitoring. Sustainability here means both keeping the tool honest over time and accounting for its resource footprint.

Using the lens without overclaiming

Two cautions keep this framework useful. First, the principles are guidance, not law. They do not certify any product, and passing them in spirit is not the same as clearing a regulator such as the FDA or a national medicines agency. Second, they can pull against each other. A more transparent model may be less accurate; a tool that serves a broad population may be tuned less tightly for any single subgroup. The WHO does not pretend these tensions dissolve. The honest use of the six principles is to surface the trade-offs and force them into the open, so that whoever adopts a tool chooses those trade-offs deliberately rather than inheriting them by default.

That is why the framework is worth knowing even if you never write policy. When a health AI product is presented to you, the six principles convert a glossy claim into six specific questions, and the questions a vendor cannot answer tell you more than the ones it can. This article is educational and not medical advice.