Why shaping AI matters: Lessons from our work in Ethiopia about building fairer AI
By Darwish Thajudeen
Two years ago, we wrote about the great divide in AI: the uneven distribution of access to tools, infrastructure, and opportunity between the Global North and the Global South. That divide remains real. Yet our work since then has taught us something even more important: fairness in AI is not achieved simply by extending access to systems built elsewhere. It is achieved when more people from underserved regions have the opportunity to shape what these systems become.[1][2]
That lesson did not emerge from theory alone. It emerged from practice, from clinical workflows, from data questions, and from the daily realities of healthcare delivery in Ethiopia. Through our SUSTAIN-AI ECG project, developed together with the Armauer Hansen Research Institute (AHRI) in Addis Ababa and Technische Hochschule Mittelhessen (THM) in Germany, we have been working to improve cardiac diagnostics in a setting where specialist capacity is limited and early detection can make the difference between timely treatment and preventable harm.
Our initial research pointed out, that roughly 30 cardiologists serve a population of around 130 million people, which illustrates the scale of the diagnostic gap and the importance of practical support for frontline clinicians. In that context, AI can be helpful. But the deeper lesson is that usefulness alone is not enough. If AI is to deserve trust, it must also be open enough to inspect, representative enough to trust, and built with the communities it intends to serve.[3]
What our project in Ethiopia taught us?
When many people hear about AI for healthcare, they imagine a familiar story: a powerful model is developed in a research hub, deployed into a lower-resource setting, and expected to close service gaps. There is some truth in that story. AI can extend diagnostic support, improve efficiency, and help overstretched health systems make better use of limited expertise. But our experience in Ethiopia has shown that this narrative is incomplete.
The challenge was not only the shortage of cardiologists. It was also the fact that many ECG AI systems have historically been trained predominantly on Western populations. Recent research on AI in the Global South warns that applications developed and trained on datasets from high-income countries often need recalibration and contextual adaptation before they can perform reliably elsewhere. In other words, the issue is not simply access to AI models. It is representation in the data foundations from which expertise is being automated.
This matters because medical AI is never context-free. Disease prevalence differs. Clinical pathways differ. Health systems differ. Data practices differ. Even where an algorithm appears technically impressive, its performance may degrade when it encounters a population, workflow, or infrastructure context that was barely present in the data used to train it. That is one reason why the current policy debate has shifted away from abstract enthusiasm and toward questions of evidence, accountability, and local participation.[4][5]
Our work on SUSTAIN-AI ECG brought these questions into sharp focus. The project is designed not as an imported black box, but as a collaborative intervention that combines AI research, clinical implementation, and explainability methods with local integration into the Ethiopian healthcare system. That structure is not incidental. It reflects a broader conviction: if an AI system is going to operate in a real public-interest setting, the people affected by it must have a meaningful role in shaping how it is built, deployed, assessed, and improved.[6]
Lesson 1: Openness
For us, openness is not a branding exercise. It is part of what makes AI governable.
The World Health Organization has been explicit on this point. Its guidance on the ethics and governance of AI for health argues that transparency, explainability, and intelligibility are core principles for responsible use, and that sufficient information should be documented before deployment so that stakeholders can understand a system’s assumptions, data properties, limitations, and intended use. WHO also notes that transparency and participation can be increased through open-source approaches and through meaningful public consultation on how technologies are designed and used.[3][6]
For every organizations, this matters at two levels. First, openness is practical. If clinicians, researchers, regulators, and civil society actors cannot inspect the logic, limits, and provenance of a system, they cannot properly evaluate whether it is appropriate for the environments in which it is being used. Second, openness is democratic. It creates room for challenge, improvement, and accountability, especially in high-stakes settings such as health where decisions affect human lives.[7][8][9][3]
In our AI-ECG project in Ethiopia, explainability is not an optional extra. It is part of the infrastructure of trust. THM’s role in developing the underlying AI model includes work on explainability methods, and this is central to ensuring that clinicians can engage with the system critically rather than treating it as an unquestionable authority. That distinction is especially important in resource-constrained settings, where support tools must strengthen professional judgment rather than substitute for it.[1][2]
Lesson 2: Representation
If openness allows AI to be inspected, representation determines whether it can be trusted.
One of the central findings of the recent scoping review on AI in the Global South is that lack of data diversity remains a major impediment to fair and effective health AI. The review concludes that current AI health applications trained on high-income country data may perform poorly or produce inappropriate results in Global South settings, in part because local data are scarce, fragmented, under-digitized, or not available at all. This is not only a technical inconvenience. It is a structural source of inequity.
That is why one of the most important achievements of the Ethiopian pilot has been the collection of approximately 7,200 ECG recordings from Ethiopian patients within just a few months. These recordings are being anonymized and prepared for open-source release, forming the basis of what should become, to the best of our knowledge, the first ECG dataset specifically focused on Ethiopian populations. This is a significant milestone not only because it improves the relevance of future ECG models for Ethiopian patients, but because it contributes to a more representative evidence base for medical AI more broadly.
There is a tendency in AI discourse to speak as though better models simply emerge from larger compute budgets or more complex architectures. That view is incomplete. Better models also depend on broader data, better validation, and better representation of the populations for whom the systems are intended. In this sense, data diversity is not a secondary fairness add-on. It is part of the scientific basis of model reliability.
For funders and policy actors, this has a clear implication. Investments in local data infrastructure, governance, and stewardship are not peripheral to innovation. They are foundational to building AI systems that can travel responsibly across contexts, or better yet, that can be co-developed within them.[6]
Lesson 3: Participation
Representation alone is not enough if local actors are only treated as data providers while key design decisions remain elsewhere.
WHO’s guidance emphasizes that end-users and direct and indirect stakeholders should be engaged from the early stages of AI development in structured, inclusive, and transparent design processes. It also calls for the public to be involved in understanding data use, assessing social acceptability, and expressing concerns and expectations about how AI is applied in health. Recent scholarship echoes this view, warning that AI in healthcare can reproduce or exacerbate existing inequities when community perspectives are not incorporated into technology deployment.[5][10]
This is why we increasingly think of participation not as a communications principle but as a research principle. Participation improves validity because local clinicians and institutions can identify workflow realities, data gaps, and contextual constraints that are often invisible from a distance. Participation also improves legitimacy because people are more likely to trust systems they have helped shape, scrutinize, and adapt to their needs.[9][3]
SUSTAIN-AI ECG works because it is grounded in partnership. AHRI leads clinical implementation and ensures alignment with Ethiopian health structures. THM contributes the technical and scientific basis of the AI model. Local hospitals and healthcare professionals shape daily use, validation, and integration into existing processes. During the pilot, the system was deployed in nine hospitals, more than 30 healthcare professionals were trained to use it in practice and the feedback of practitioners was thoroughly collected and documented. It creates a basis for a future more localized version of our ECG4Africa platform for which we are currently looking for funding. This kind of collaboration is not simply operational convenience. It is what makes the system responsive to the setting it serves.
This collaborative model also challenges a common assumption in global AI debates: that innovation flows in one direction, from a few technical centers outward. Our experience suggests the opposite. Sites such as Ethiopia are not merely places where technology is applied. They are places where better questions are asked, better evidence is generated, and the limitations of supposedly universal systems become visible.
Our different vision of global AI
The language of the Global North and Global South can sometimes flatten important differences. There are inequities within high-income countries and important innovations across lower-resource settings. Yet the division still captures something real about who usually sets the terms of technological development and whose contexts are assumed to be universal.
Our work in Ethiopia has made one point especially clear: the future of AI will not become fair simply because it spreads more widely. It will become fair when more of the world has the power to shape how AI is designed, trained, validated, governed, and improved. That is a higher standard than access alone. It requires authorship, participation, and shared ownership.
This is the movement we believe in. An AI ecosystem that is open enough to inspect. AI systems that are representative enough to trust. Technologies that are built with the communities they intend to serve, not merely delivered to them after the key decisions have already been made.
The SUSTAIN-AI ECG pilot has shown what this can look like in practice: real clinical deployment, local training, data generation, cross-continental partnership, and a pathway toward a more representative evidence base for cardiovascular AI. The significance of that work extends beyond one project and one country. It points toward a more mature model of AI for good, one grounded not in technological heroism, but in collaboration, humility, and institutional trustworthiness.
Building it together
For supporters, mentors, and funding organizations, the invitation is clear. Help back AI that can be inspected rather than merely marketed. Support data ecosystems that improve representation rather than deepen bias. Encourage partnerships in which local institutions are co-authors of innovation, not just sites of deployment.
For researchers, that may mean asking whose data is missing and whose expertise is absent from validation. For policy actors, it may mean turning high-level principles into real governance expectations. For philanthropic and public funders, it means recognizing that long-term fairness depends on investing in capacity, infrastructure, and open evidence, not just in tools.
We take optimism from the fact that this work is already underway. In Ethiopia, clinicians, researchers, public institutions, and technical partners are demonstrating that fairer AI can be built through participation rather than assumption. The next step is to widen that circle.
If the future of AI is to serve humanity fairly, then humanity must be present in how it is built. That is not only a moral aspiration. It is a scientific and institutional necessity, we urge you to participate in!
Best regards
MI4People-Team
