WHO's Six Ethical Principles for LMMs in Practice

WHO's six consensus principles from its 2021 guidance remain the ethical core for LMM governance. Here is how Hong Kong healthcare professionals can apply them when encountering generative AI tools.

1. Protect autonomy

Humans must remain in control of medical decisions. Patients should understand when AI assists their care; data privacy and valid informed consent must be protected. LMMs must not replace the right to decline AI-assisted care when alternatives exist.

2. Promote human well-being, safety, and the public interest

Tools should meet defined safety and efficacy standards for their intended use. AI should not be used when it causes avoidable harm that alternatives could prevent. Quality monitoring must continue after deployment — not only at launch.

3. Ensure transparency, explainability, and intelligibility

Developers, clinicians, patients, and regulators need sufficient information about how an LMM was built and its limitations. Disclosure should enable public debate — including labelling AI-generated content so users know a machine, not a clinician, produced it.

4. Foster responsibility and accountability

Appropriately trained people must use AI under suitable conditions. Mechanisms for questioning decisions and redress when people are harmed are essential. Human supervision points should exist upstream and downstream of algorithms.

5. Ensure inclusiveness and equity

AI must not encode biases against identifiable groups. Access should extend beyond high-income settings to low- and middle-income contexts — relevant for Hong Kong's diverse population and cross-border care. Monitor for disproportionate harms.

6. Promote responsive and sustainable AI

AI should support sustainable health systems, workplaces, and environments — linking clinical adoption to workforce planning and the carbon/water costs of large models.

Practical checklist

Before adopting an LMM tool, map each principle to a concrete question: Who is accountable if the output harms a patient? Can we explain its use to patients in plain language? Was it tested on populations similar to ours?

Source: WHO — Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models (2024)