The Blueprint

Doing AI Right in Africa

This essay was first published at HAKI

My first essay in this series argued that Africa’s position in the AI revolution is one that offers strategic choice. I posed that Africa’s lateness allows it to choose better, and that its diversity and polycentric structure can be deployed as leverage. All of this remains true. However, optionality is not strategy. Africa still needs to make her choices.

This article concerns the making of those choices, and outlines what doing AI right in Africa requires.

1. Fit for purpose

The dominant dogma today is the development of ever-larger AI models on ever-larger datasets. This produces systems optimised for the languages, contexts, and concerns of the high-income and homogenised populations of the global north. African institutions cannot win this race, and they need not attempt to. Instead, they should pursue a more important and more rewarding question, which is how they can build the most useful models possible. Inevitably, the usefulness of technologies in African contexts will demand design choices that deliberately diverge from the frontier. This shall be our path, and we should follow it with fervour.

Many high-impact problems across African agriculture, education, economics, health and governance do not need frontier models with billions of parameters. They need specialised models that are trained on locally relevant data, built for extant hardware and connectivity conditions, and maintained by practitioners who understand both the technical architecture and the area it serves. A diagnostic model running on a tablet in a clinic with intermittent connectivity, and calibrated to the clinical patterns of a specific population, is more valuable in that clinic than a frontier model whose operational requirements cannot be met there and whose outputs cannot be useful there. The real driver of innovation is fitness for purpose. And fitness for African purposes is something African institutions are better positioned to build than institutions in Silicon Valley or Shenzhen whose grasp of those purposes is indifferent.

Understanding this also transforms the resource calculus. Small, specialised models require substantially less compute to train and less infrastructure to deploy. Our bottleneck therefore shifts from capital and hardware to domain expertise and problem formulation: the capacity to identify tractable problems, gather sufficient data, and build systems that practitioners can use, maintain, and trust. This means our hurdle is not financial scale, it is actually intellectual and institutional work.

Open-source solutions provide the technical foundation that makes this approach viable across the continent. As foundational models, training frameworks, and deployment tools are becoming openly available, African developers should build on such infrastructure. Fine-tuning open-source models on local data produces models that local institutions can own outright. Contributing to open-source projects also builds capability that cannot be extracted: the spirit of the open-source software movement lives in people and institutions. Every developer who contributes to open foundations and every institution that funds them strengthens a technical commons whose value compounds across the rest of the world, forever.

2. Research as Problem-Solving

Universities and research institutions are the anchors of technical capability. However, their current incentive structures pull research capacity in the wrong direction. Funding, tenure, and institutional prestige still track international publication metrics whose research priorities are set by a few high-income countries and for their own meritocratic purposes. A researcher building a diagnostic system for a neglected tropical disease, or developing a small agricultural advisory model for smallholder farmers, faces a structurally harder path to career advancement than a colleague whose work aligns with the funding priorities of international donors and publication interests of “high-impact” scientific journals.

This is not a failure of individual will, it is the predictable outcome of the structures that African researchers currently inhabit. These structures need to be rebuilt. The continental research community should define high-impact African problems by prevalence, tractability, and the absence of existing solutions, rather than by the availability of international funding or the proximity to global research trends. By channelling existing funding toward researchers pursuing such problems, we would redirect intellectual capacity without asking anyone to sacrifice sustainability. Additionally, African publication and recognition infrastructure should be incentivised to treat demonstrated local impact as a legitimate scholarly credential, just as strongly or even more than they currently regard international standing.

Curriculum reform is also crucial. AI training should be grounded in African problem contexts from the outset. This will produce graduates whose skills are calibrated to the problems they will encounter. Embedding intellectual property management, data governance, and translational research alongside core technical training will also equip researchers and engineers for the institutional and commercial dimensions of AI development, not only the computational ones. The European Institute of Innovation and Technology’s Deep Tech Talent Initiative trained over one million people in three years. Comparable ambition is not unrealistic at our continental scale: programmes like DELTAS Africa in the Science for Africa Foundation already provide examples to emulate.

3. Coordinate, then Negotiate

No single African country has the market, scale, legal capacity, or technical resources to negotiate bilateral AI agreements from a position of comparable strength with global economic powers or the largest global corporations. This is a structural condition which cannot be remedied by preparation alone. The appropriate strategy is coordination, and coordination is most effective when built on the domestic capabilities my preceding sections describe.

The logic is simple. A nation or corporation negotiating for access to African data or markets with a single country faces an inferior counterpart that they can outmanoeuvre. The same institution negotiating with a coordinated continental bloc faces a different calculation entirely. They are compelled to contend with a billion consumers, conditions of access across all member states, and collective legal and institutional intelligence. This leverage is structural, and it only requires that African countries align on the specific and limited set of issues where fragmentation currently works against our collective interests.

Coordination does not require uniformity. Individual countries will and should make different choices about domestic AI applications, regulatory architecture, and the balance between innovation and caution. These differences are productive: they generate comparative evidence that allows better approaches to displace weaker ones across the continent. What coordination demands is alignment on the minimum standards for bilateral data agreements. It ensures shared legal intelligence so that no government signs terms a neighbour has already identified as harmful, and guarantees joint negotiating positions when foreign actors approach the continent with proposals whose acceptance by any single country weakens the position of all. That African countries have repeatedly negotiated separately, and on inferior terms, is evidence of the cost of this fragmentation

Shared legal capacity is the most immediately achievable and most consistently undervalued form of coordination. A continental AI legal intelligence function that is jointly funded, accessible to all member states, and maintains a living database of precedents and recommended minimum terms would directly improve the quality of bilateral negotiations everywhere. The East African Community, ECOWAS, and the African Union carry existing institutional mandates that can accommodate this function without new structures. What is required is political prioritisation: treating AI governance as a continental coordination problem rather than fifty-four separate national ones.

4. Enforceable Terms

African governments are negotiating bilateral agreements that exchange national data for infrastructure, technical capacity, or funding at increasing frequency and scale. The terms of exchange will determine whether these agreements build genuine capability or entrench dependencies that outlast the partnerships that created them. When domestic capability is established and regional coordination in place, African governments will be far better positioned to insist on terms that protect long-term interests. For every bilateral agreement, these four provisions are minimum conditions that must be met. Recent bilateral agreements concluded by African governments demonstrate that each of these protections has been omitted in practice.

First: a specified and binding benefit-sharing formula, defined before any deal is signed. Commercial products derived from any data generate value over timescales that extend well beyond the lifecycles of the agreements that spurred them. Deferring benefit-sharing to future subsidiary negotiations that neither party is obligated to conclude leaves the originating country with legal ownership of data whose commercial value has already been extracted, without any guaranteed benefits in recompense. Minimum royalties, tied to gross revenues from products demonstrably derived from national data, triggered automatically rather than contingent on further negotiation, are the appropriate standard.

Second: explicit licensing provisions for AI models trained on national data. A model trained on national agricultural, health, economic or genomic data embeds the informational value of that data in a form that persists independently of the original dataset. Agreements that protect the data whilst leaving model ownership unaddressed protect the less valuable asset whilst surrendering the more valuable one. Perpetual, royalty-free licensing of resulting models to the country of origin is the appropriate standard provision.

Third: mandatory data deletion with certified third-party verification upon agreement expiration. Data retention requirements vary across jurisdictions, and what a partner country’s law demands may conflict directly with what an African government’s data protection framework requires. A nation’s legal ownership of data provides no practical protection once a copy of that data resides in a foreign jurisdiction, especially when enforceable deletion obligations were not secured. Such provisions must specify not merely that data “may be disposed of” but that it shall be permanently deleted within a defined window, with cryptographically verified confirmation returned to the originating government.

Fourth: binding arbitration with defined jurisdiction, monetary penalties, and enforceable timelines. Dispute resolution limited to diplomatic consultations carry no binding outcome, no timeline, and no financial consequence for non-compliance. They provide the form of protection, not its substance. Arbitration must be demanded under internationally recognised rules, with jurisdiction explicitly granted to the data’s country of origin, and penalties applied proportionately to commercial value at stake. This converts aspirational commitments into enforceable obligations.

None of these terms are unusual in well-negotiated commercial data agreements. Their absence from sovereign agreements reflects a capacity gap, a gap that the capability-building and coordination described in the preceding sections directly addresses.

5. What is Success?

Success in African AI deserves a precise definition, because imprecision here carries a strategic cost. Success is not a continent that has replicated any existing model of AI development, nor one that has built some number of data centres or reached some abstract condition of technological self-sufficiency. It is more specific and more demanding: it is AI systems accountable to African governance; data generated by African populations contributing to African analytical capability; problems identified by African communities addressed by systems built by Africans; and economic value created by African talent retained within African institutions.

This standard requires no special pleading. Every major AI-developing region applies it to itself as settled policy. What distinguishes Africa is not that the standard is unreasonable but that the institutional conditions for meeting it, which are research investment calibrated to local problems, technical capacity anchored locally, coordinated regional standards, and enforceable data governance, are still under construction.

This work does not need to wait for continental consensus. It advances in the Kenyan team building the next legal AI system on open-source infrastructure. It deepens in the Nigerian university redirecting a research programme toward a problem its communities in Abuja face, and in building the recognition infrastructure in such nations to make this redirection attractive and sustainable. It consolidates in the regional body establishing shared legal intelligence before the next bilateral agreement reaches the negotiating table. Furthermore, it matures in the governments that insist on enforceable deletion clauses as a condition for signature.

The architecture of African AI sovereignty is built from many such decisions, which will be made by many institutions. “Any institution can initiate change that propagates across a polycentric network, and the window in which that propagation can still determine the outcome remains open, but it will not remain open indefinitely.