Social and biological factors influence health outcomes through conventional pathways and, increasingly, through digital pathways of data and algorithms.

• Biomedical data disadvantage is a health risk factor for most of the world's population; representation bias and distribution shifts in health data are key mechanisms through which it acts.

• AI greatly empowers precision medicine but simultaneously opens up a major pathway through which this risk factor can exert its effects.

• AI-powered precision medicine can be systematically less precise for data-disadvantaged populations, producing new or amplified health disparities. These disparities may vary continuously with genetic ancestry rather than conforming to discrete population categories.

• These health disparities can affect any disease for which biomedical data inequality exists, making their potential impacts broad.

• The digital pathways also provide targets for algorithmic intervention. Building on transfer learning, we are advancing toward in-context learning (ICL) with tabular foundation models and moving from discrete ancestry categories to the genetic ancestry continuum to develop a new framework for robust omics-based prediction of cancer risk and outcomes, aiming for a Pareto improvement: improving prediction for data-disadvantaged populations without reducing performance for populations already well served by existing models.