The 1978 Alma-Ata Declaration placed primary health care at the base of the health system – an architecture conventionally represented as a pyramid, with tertiary care at its apex. What is less often emphasised is that the quality of care can be graded along the same structure, with a declining gradient towards the base. Although often masked by the referral system, the inequity cost of this gradient is vast.
Artificial intelligence has been heralded as the technology to flatten this gradient, given its potential to leapfrog the challenges health systems face. The claim is no longer aspirational: AI applications are already deployed in low-resource settings, from tuberculosis screening to diabetic retinopathy. Against a projected global shortage of 10 million health workers by 2030, particularly along the last mile, AI becomes a credible tool for bringing health systems to the most vulnerable populations.
The obstacle lies in what the models have learnt from. Health data concentrate in a few high-resource settings, and it is well established that performance degrades when models cross institutional and demographic boundaries. With this realisation has come a shift towards consolidating data from the Global South, home to more than 80% of the world’s population. Consequently, data sovereignty has moved to the centre of health policy, with localisation requirements asserting data as a national asset. This is necessary but insufficient. Localisation governs where data reside, not which patients the data represent.
The problem of ‘apex capture’
Nor is this phenomenon confined to poorer countries. Even in high-resource countries with mature digital health systems, most datasets used to train clinical-grade deep learning algorithms originate from a few accessible urban hubs. Globally, most clinical AI evaluation studies have been conducted in secondary or tertiary care. Even without governance failure, data accumulate where capital and infrastructure exist. Where internal disparities are steep, this ‘apex capture’ cascades: most models are built using data from the resource-rich apex – and are thus built for the apex. Such models may come to represent the country without ever being evaluated on the populations carrying most of its disease burden. This is sovereignty by proxy – digital colonialism within borders rather than across them.
Globalisation compounds this. Across many low- and middle-income countries, foreign investment is growing in precisely the tertiary hospitals that generate structured, coded, imaging-rich data. This offers a plausible route to greater representation in global models, although one still marred by selection bias. Regulators may miss it too: many AI devices cleared for market do not disclose whether the model was tested beyond a single site.
Moreover, harm from such models is not easily seen, particularly in the last mile, where they will be used to meet unmet health needs. There, a slightly miscalibrated model may be overly trusted by a front-line worker – automation bias leads the worker to treat its output as authoritative. The result is an echo chamber whose pull strengthens as clinical experience falls. Sustained exposure to such models may also erode unassisted performance, deskilling the very professionals AI is meant to support. Post-deployment monitoring should catch such failures, yet reviews of deployments find real-world performance goes largely unassessed.
Building AI sovereignty
The solution is not restraint, which would deny the last mile the access that AI is best placed to extend. Rather, AI sovereignty must be strengthened across the technology’s lifecycle. Sovereignty here is not protectionism but the institutional capacity to evaluate and govern AI using one’s own evidence. Its operational core is verification: no model should reach patients without demonstrating performance in the population it will serve. Institutional and infrastructural readiness and evaluation are thus prerequisites, and monitoring becomes the embedded layer. Evidence generation must be financed as public infrastructure, and pooled regional evaluation can give smaller states this capacity. Such capacity determines whether governance itself is sovereign – built on the scaffold of enforceable national standards, clinical pathways and procurement conditions, informed by global guidance. It is what allows jurisdictions to avoid the isomorphic mimicry already documented in national digital health strategies.
Forty years ago, the Ottawa Charter defined health promotion as enabling people to increase control over the determinants of their health. Today we confront a new digital determinant – one that can conceal its own provenance. A country that owns its evidence, its objectives and its last mile can make AI move expertise outward rather than patients inward, easing the burden on the apex of the health system.
