Ethics of AI in health
About 25 minutes · plain language
Ethics asks not only “can we use AI?” but “should we, and how?” In this lesson you will learn the six ethical principles WHO set out for AI in health, and how to apply them to real situations.
- Apply the six WHO principles to a real case.
- Spot bias and explain whom it harms.
Key terms
- Autonomy
- The right of people to make informed decisions about their own care.Example: A patient is told when an AI tool has influenced a recommendation.
- Beneficence and non-maleficence
- Doing good and avoiding harm.Example: Testing an AI tool before relying on it.
- Transparency
- Being open about what an AI tool does, what data it uses and how well it works.Example: A plain-language sheet explaining a ward’s AI tool.
- Explainability
- Being able to give an understandable reason for an AI output.Example: Showing which vital signs raised a risk score.
- Accountability
- Being answerable for decisions and their results. People and organisations are accountable; tools are not.Example: The nurse remains accountable for care even if an AI suggested it.
- Algorithmic bias
- When an AI tool works worse, or treats people unfairly, for some groups.Example: A tool that underestimates risk in women or in a minority group.
- Equity
- Fairness in access to care and in outcomes.Example: Checking that a tool works equally well for patients who do not speak English.
- Proxy
- A measure used in place of the thing you really care about.Example: Using past health-care spending as a stand-in for health need.
- Sustainability
- Meeting today’s needs without harming people or the planet in the future. WHO’s sixth principle includes the environmental cost of AI.Example: Choosing a smaller AI tool when it does the job as well as a large one.
The six WHO principles
In 2021 WHO published six principles for AI in health, shown in the figure: protect autonomy; promote human well-being, safety and the public interest; ensure transparency, explainability and intelligibility; foster responsibility and accountability; ensure inclusiveness and equity; and promote AI that is responsive and sustainable. They give a shared language for discussing any AI tool.
Autonomy and consent
People should stay in control of decisions about their care. Patients should be told when AI plays a meaningful part in their care, and clinicians must be free to disagree with it and override it.
Bias: how it happens
Bias can enter at every stage. Data may under-represent some groups. Design may use a misleading proxy. Use may differ between groups. A well-known US example used past health-care costs as a proxy for health need. Because less money had historically been spent on Black patients with the same level of illness, the algorithm judged them healthier than equally sick White patients, and fewer were referred to extra care. Fixing the proxy greatly reduced the bias.
The lesson: bias is found by checking results separately for each group, not by assuming numbers are neutral.
Accountability
A tool cannot be held responsible; people and organisations can. Nursing and medical codes keep the professional accountable for care, even when AI contributed. The American Nurses Association, for example, states that AI must support, not replace, nursing judgement.
Transparency and explainability
People affected by an AI tool should be able to learn what it does, what data it uses and how well it works. Full technical detail is not always possible, but honest information about purpose, limits and performance always is.
The environmental cost of AI
WHO’s sixth principle asks for AI that is responsive and sustainable, including its effect on the environment. AI runs in data centres that use electricity and water for cooling. The International Energy Agency estimated that data centres used about 1.5% of the world’s electricity in 2024, and that this could roughly double by 2030, driven largely by AI. Generating text or images uses much more energy per task than a simple search or a small, task-specific model.
This is not a reason to avoid AI where it helps patients. It is a reason to use it with purpose: prefer a smaller or task-specific tool when it is good enough, avoid generating content nobody needs, and ask vendors and institutions to report the energy use of the tools they buy. Health care already produces a large share of national emissions, and many health systems have set targets to reduce them.
Case: the no-show predictor
A clinic wants to use an AI tool that predicts which patients will miss appointments, so it can double-book their slots. Checking results by group, the team finds the tool flags patients from poorer areas far more often, mainly because of transport problems. Double-booking would make their access even worse. The clinic decides to use the prediction differently: flagged patients get a reminder call and help with transport.
Common mistakes
- Assuming an algorithm is neutral because it uses numbers.
- Checking accuracy overall but not by group.
- Letting “the system decided” replace a person’s responsibility.
Check your understanding
1. A tool is 90% accurate overall. Why might it still be unfair?
It could be much less accurate for some groups. Overall accuracy can hide differences between groups.
2. Who is accountable if a nurse follows a wrong AI suggestion?
People and organisations, including the nurse and the organisation that deployed the tool. The tool itself cannot be accountable.
3. Give one example of a misleading proxy.
Using past health-care spending as a measure of health need.
Try it
- Choose an AI tool you know and test it against each of the six WHO principles. Which is weakest?
Remember
- Use the six WHO principles as a checklist.
- Check performance in each patient group to find bias.
- People stay accountable; tools do not.
- Sustainability is part of AI ethics: use AI with purpose, and prefer smaller tools when they are good enough.
- World Health Organization. Ethics and governance of artificial intelligence for health: WHO guidance. Geneva: WHO; 2021. www.who.int
- UNESCO. Recommendation on the Ethics of Artificial Intelligence. Paris: UNESCO; 2022 (adopted November 2021). unesdoc.unesco.org
- Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ. 2025;388:e081554. doi.org
- American Nurses Association. The ethical use of artificial intelligence in nursing practice: position statement. ANA; 2022. Republished in: Online J Issues Nurs. 2025;30(2). ojin.nursingworld.org
- Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-53. pubmed.ncbi.nlm.nih.gov
- International Energy Agency. Energy and AI. Paris: IEA; 2025. www.iea.org
- Luccioni AS, Jernite Y, Strubell E. Power hungry processing: watts driving the cost of AI deployment? In: Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency (FAccT). ACM; 2024. p. 85-99. arxiv.org
Lesson text is our own plain-language summary of these sources; it is not endorsed by their authors. Check the original before relying on any detail.