Responsible AI Adoption in Health Research

The constraint on responsible AI adoption in population-scale health research is not the capability of the models. It is the capability of the institution to establish what a workflow costs today, to compare it against a version assisted by a model, to detect a quality regression before it reaches a publication or a participant, and to repeat that exercise when the model changes underneath it. This paper argues that the portfolio should follow that capability. Where it leads instead, the organisation is buying something it cannot yet assess. An organisation that can run an instrumented baseline may pilot a reversible internal workflow, and one that can run a pre-registered controlled comparison may consider a use case whose output leaves the building. It argues equally firmly against several deployments that are currently easy to buy.