Scaling AI in Health Responsibly: Barriers, Risks, and Policy Pathways
The rapid growth of AI investment and the spread of certified tools across healthcare systems have yet to translate into meaningful, equitably distributed benefits for patients and providers alike. Addressing the structural barriers that prevent responsible scaling of Artificial Intelligence in health was the central focus of a workshop co-organised by the Organisation for Economic Co-operation and Development (OECD) and the Digital Transformations for Health Lab, where experts were invited to review and enrich a draft policy recommendation.
Professor Roda put forward several additions she considered essential to the recommendation:
- Technical infrastructure as part of the trust framework. Health AI applications typically run on infrastructure that are not owned or governed by the deployers, raising serious questions of accountability, data governance, and transparency. Infrastructure should therefore be explicitly included in any trust framework the recommendation seeks to build.
- Selective, evidence-based scaling. Before committing to scale, a structured evaluation should determine whether AI genuinely delivers advantages, and at what cost. Decision-makers should ask whether equivalent or better outcomes could be achieved through other technologies or non-technological approaches, with added benefits in employment, training, and the quality of human interaction. Economic, societal, and environmental consequences must be mapped, and it must be clear who bears the costs and who captures the gains. Vulnerable populations, including children, deserve particular attention in this assessment.
- A more rigorous understanding of human oversight. Requiring a person to validate AI output is often treated as sufficient, but this interpretation is too narrow. When AI handles analytical tasks over extended periods, operators gradually lose the ability to detect its errors, a phenomenon known as cognitive offloading. This dilutes human responsibility, with potentially devastating consequences. The recommendation should affirm that humans must remain the primary decision-makers, maintain a clear understanding of AI’s limitations, and continue to develop mastery of medical knowledge.
- Education of medical professionals and the public. Continuous training for healthcare personnel and accessible information for the general public should be treated as prerequisites for responsible deployment.
- Clear accountability and liability before deployment. Whenever AI in health applications, responsibility must be assigned prior to deployment. All actors should be required to conduct due diligence, submit to independent auditing, maintain incident registers, and communicate clearly with the individuals and communities affected.