The World Health Organization published a report arguing that ethics review has not kept pace with the speed at which AI tools are entering health research. The document calls for stronger oversight of how AI is developed and used in studies involving patients and their data.
The concern is structural. Models are often trained on clinical data collected for other purposes, sometimes across borders, and their performance can drift after deployment. Conventional research-ethics frameworks were written for protocols with fixed endpoints, not for systems that update, generalise unpredictably, or infer sensitive attributes from routine records.
WHO’s recommendations centre on governance rather than specific tools: clearer accountability for who answers when a model causes harm, transparency about training data and validation populations, ongoing monitoring after a study ends, and meaningful patient involvement in design and review.
The report lands in a crowded policy week. Regulators in the United Kingdom have issued their own recommendations on clinical AI, India’s drug regulator has formed an expert committee, and several United States bills are moving on health care AI governance. Each framework defines risk differently, which creates friction for developers running multi-country studies.
For research groups, the practical effect is likely more documentation: provenance for training data, subgroup performance reporting, and a monitoring plan that outlives the trial. For regulators, WHO’s push adds pressure to align definitions before national rules harden in incompatible directions.
