An artificial intelligence tool may accelerate medical research without simplifying the responsibilities around it. On 21 September 2026, the World Health Organization is holding the public launch of its report on ethics review and oversight of AI-related health research. Published in July, the document examines challenges for researchers, ethics committees, regulators, funders and scientific journals. The webinar does not create global law or approve any product. It brings practical recommendations into public discussion so that speed of analysis or model power does not become a shortcut around consent, equity and accountability.
Three forms of research that should not be confused
WHO distinguishes three categories. The first is health-related data science using AI, such as an algorithm searching large information sets for associations. The second covers research conducted with AI tools and technologies, where the technology forms part of the research method. The third is health-related research on AI tools and technologies themselves. These situations may overlap, but they do not always raise the same questions. Evaluating an experimental model, using an assistant to analyse documents and re-examining health records require different evidence and different levels of data access.
This classification helps ethics committees ask more precise questions. Who counts as a participant when research reuses previously collected data? What consent was obtained, and does it truly cover the new use? Can a model trained in one country or hospital be studied elsewhere without reproducing inequality? AI does not remove the principles of health research. It can make their application harder because data, code, technology providers and decisions are distributed among several actors. The report therefore extends scrutiny beyond the team whose names appear on the protocol.
Ethics committees confronting opaque systems
A committee cannot readily assess what it cannot understand or document. AI systems create challenges involving traceability, bias, variable performance and explanation. High average accuracy can conceal poorer results for particular groups. A model update can also change behaviour after the initial review. WHO examines gaps in existing standards and oversight and ways to strengthen committee capacity. This does not require every member to become an engineer. It means assembling the expertise needed and requiring researchers to provide information clear enough for risks to be judged.
Ethical oversight is not a one-time formality before a project begins. When a system learns, changes version or gains access to new data, monitoring should test whether its assumptions still hold. Responsibilities must be defined: who watches for incidents, who authorizes a modification and who informs affected people or institutions? The report also considers possible uses of AI in the work of ethics committees themselves. Automating a literature search may help; handing over judgment on a study’s acceptability without oversight would add a new layer of risk. A tool can support deliberation, not carry moral responsibility for it.
Funders, journals and regulators enter the frame
Responsibility does not stop at the laboratory door. WHO considers complementary roles for funders, data-access mechanisms, scientific journals, publishers and regulatory agencies. A funder can make support conditional on risk assessments and benefit-sharing plans. A journal can require a description of data, limitations and validation methods. A regulator can demand evidence suited to the intended use. None replaces the ethics committee, but each can prevent a known weakness from simply being passed to the next stage.
This chain matters because publication can give a result lasting authority. If data are biased, exclusions poorly explained or a proprietary model impossible to examine, caution should be visible in the conclusions. Transparency does not necessarily mean releasing all sensitive data or commercial code. At minimum, it requires that choices affecting validity, rights and the possibility of reproducing or contesting a result be understandable. Trust cannot be announced into existence. It depends on showing how a decision was made and how it can be corrected.
The risk of extractive research
The report pays particular attention to low- and middle-income countries. It addresses fairness, benefit sharing, data colonialism, the movement of ethically questionable research into less protected settings and power imbalances. A population may provide valuable data without gaining access to the resulting tools, care or revenue. A local committee may be asked to evaluate a complex technology without sufficient funding or expertise. Describing these risks does not mean every international collaboration is abusive. It requires asking who sets priorities, who controls data and who benefits from the result.
For B-EMPIRE, the significance of the 21 September launch lies in this shift: ethics in medical AI concerns not only model quality but the entire organization of research. Confirmed facts include a 69-page report, the three categories it examines and a virtual launch focused on consent, bias and committee capacity. Its real influence will depend on adoption of the recommendations, resources for oversight and institutions’ willingness to slow a project when safeguards are inadequate. In health, moving faster is progress only when no one becomes invisible in the process.