Why Africa should build trust in AI differently
Trust in AI depends on more than technology. Africa's opportunity is not simply to deploy AI diagnostics, but to show how trustworthy AI is built: through strong institutions, local evidence, and continuous evaluation.
The clinical usefulness of health technology algorithms in diabetic retinopathy screening was established using retinal images from Kenya's Nakuru eye disease cohort. Photo: Aisha Bappah
In eight states across Nigeria, mobile teams have spent the past several years bringing ultra-portable digital X-ray units to hard-to-reach communities, screening tens of thousands of people for TB using CAD4TB, a computer-aided detection algorithm that flags likely disease on a chest radiograph in seconds. It is, on paper, exactly the kind of tool global health has long hoped for: a way to put specialist-level triage in front of patients who may never see a radiologist. Independent validation studies in Tanzania and South Africa have shown that CAD4TB can accurately distinguish TB from healthy lungs, and its use aligns with the World Health Organization (WHO) 2020 recommendation that computer-aided detection can serve as an alternative to human interpretation for TB screening.
Yet a 2024 comparison of commercially available chest X-ray algorithms on the same South African population found CAD4TB performing meaningfully below several newer competing products, including Lunit and Nexus. CAD4TB was not fraudulent; nor was it unsafe. It remained a clinically useful tool. What changed was the availability of better comparative evidence.
This is the real state of AI diagnostics in Africa: not a story of broken algorithms, but of incomplete information. A health ministry procuring a TB screening tool, or a clinician relying on one, typically sees a single vendor's validation study, not an independent, head-to-head comparison showing how multiple products perform on local populations as technologies evolve. This is an information and governance gap, not simply an algorithmic one. And it reframes the trust question.
Much of the global conversation about AI diagnostics still asks whether we can trust AI. Africa should ask a different question: what kind of health system produces trustworthy AI? Trust is not something that can be built into an algorithm. It is earned through institutions that continuously evaluate, govern, and improve how AI performs in real-world care.
Clinical usefulness is a systems outcome
Health technologies rarely transform health systems on their own; systems determine the extent to which technology's promise becomes realised care. An algorithm can perform exceptionally well in a peer-reviewed validation study and still add little value in practice because the workflow around it assumes referral capacity, connectivity, or digital records that the deploying facility simply does not have.
Landry Dongmo Tsague, director of Africa CDC’s Center for Primary Health Care, has repeatedly argued that AI cannot compensate for weak primary healthcare systems. Where facilities lack reliable electricity, connectivity, or functioning referral systems, AI does not solve the underlying problem.
A similar pattern is visible in diabetic retinopathy screening, a second area where African evidence has accumulated steadily rather than all at once. An early study using retinal images from Kenya's Nakuru eye disease cohort helped establish that automated detection was feasible in an African population. A more rigorous prospective validation followed in Zambia's Copperbelt province, where a deep-learning model (originally trained on a Singaporean cohort) was tested against retinal specialists' grading of images from over 1,500 Zambians with diabetes. It performed with clinically acceptable accuracy despite having been trained on a different population entirely.
Building on those results, a randomised controlled trial is now underway in Tanzania, testing not just diagnostic accuracy but whether AI screening actually gets more patients to timely ophthalmology follow-up compared with standard care. That progression, from feasibility, to cross-population validation, to a trial measuring real referral outcomes, is closer to the ideal sequencing.
Yet even here, no African study has run the kind of independent, head-to-head comparison between competing retinopathy algorithms that exposed CAD4TB's relative standing. The gap that matters most—knowing not only whether a tool works, but how it compares with its alternatives on local populations—remains in even these better-evidenced cases.
Trust is earned through institutions
Popular commentary treats trust as something to be engineered into software or asserted through ethical principles. Ethical frameworks matter, but Africa is no longer starting from a blank page.
The African Union's (AU) Continental Artificial Intelligence Strategy identifies health as a priority sector while emphasising public value, local capacity, and responsible governance. Africa CDC has gone further through its Continental Health Data Governance Framework, building on the Malabo Convention and the AU Data Policy Framework to establish African approaches to collecting, sharing, and governing health data across borders.
The challenge today is less of drafting new principles and increasingly that of implementing, coordinating, and resourcing the institutions that already exist. Trust grows when ministries can independently evaluate products before procurement. It grows when regulators can monitor algorithms after deployment rather than only before approval. It grows when clinicians know that software updates, performance changes, and safety concerns are subject to continuous oversight. Trust is not an abstract ethical aspiration; it is an institutional capability.
Local representation is clinical evidence
The CAD4TB example illustrates a further lesson. Diagnostic performance is not a fixed property of an algorithm. It varies by population, imaging equipment, disease prevalence, and clinical setting—a phenomenon researchers call "dataset shift", in which algorithms lose performance when deployed on populations that differ from those on which they were developed and validated. That makes local representation more than a question of fairness; it makes it a question of clinical evidence. Africa CDC has consistently argued that AI must be trained and continuously evaluated using data that is accurate, timely, and representative of the populations it serves, and that such data should remain governed within Africa, for Africa's benefit. This is often discussed as data sovereignty. It is, as the TB and retinopathy examples show, a question of whether a health ministry can see how a given tool actually performs on its own population before relying on it, regardless of how sophisticated the underlying model appears on paper.
What success should look like
Africa should resist measuring progress by the number of AI tools deployed, or by impressive accuracy scores reported in validation studies conducted elsewhere.
We should be asking: can a TB programme in Kaduna or Kampala compare competing AI products on local data before procurement, rather than relying primarily on manufacturers' own studies? Can regulators continue evaluating those products as evidence evolves, just as the Tanzania retinopathy trial is now testing real referral outcomes? Will Africa CDC's Continental Health Data Governance Framework, once endorsed, be implemented and resourced at national level rather than joining the list of frameworks that simply remain on the shelf? Will the infrastructure ambitions set out in Africa CDC's Primary Health Care Digitalization Framework be realised so that AI tools are introduced into health systems capable of sustaining them? These are not questions about algorithms; they are questions about institutions.
None of this argues against the use of AI diagnostics in African health systems—quite the opposite. It argues that Africa's greatest contribution to the global AI conversation may not be developing the next breakthrough algorithm. It may be demonstrating how trustworthy AI is built: through resilient primary healthcare systems, strong public institutions, locally generated evidence, and continuous evaluation after deployment.
If Africa gets that sequencing right—infrastructure, governance, and locally validated evidence, in that order—trust will not be the obstacle to AI-enabled diagnostics. It will be its most durable outcome.
The opinions expressed are those of the author and do not necessarily reflect the position of Re:solve Global Health.
Caroline S Mbindyo is director of impact and innovation at Amref Health Africa. She is a futures researcher operating at the intersection of digital transformation, global health, and international development. With a strong foundation in innovation strategy, health R&D, and evidence-based practices, Caroline is dedicated to honouring local communities and is currently investigating how AI could shape health across Africa.
