A Public Good That Cannot Speak Your Language - Cognitive Sovereignty: Beyond the Fine Print

Dr Kenechukwu Ikebuaku | 22 September 2026

Abuja, July 2026

In late July, telecommunications ministers from across Africa met in Abuja for the African Telecommunications Union's Conference of Plenipotentiaries. They adopted the Abuja Ministerial Declaration on Meaningful Connectivity for Africa, and launched ATLAS Umoja AI, a pan-African initiative to build artificial intelligence models that understand African languages. It brings together the governments of Benin, Kenya, Namibia, Nigeria and Togo with the GSMA and African technology partners.

Nigeria's Minister of Communications, Innovation and Digital Economy, Dr Bosun Tijani, opened with a figure that deserves to travel further than the conference hall. Africa is home to nearly one third of the world's languages. Yet fewer than 2% of them are meaningfully supported by modern AI systems (CIO Africa, 2026). He framed the initiative explicitly as a matter of digital sovereignty.

I write as one of the partners in that initiative, and as someone whose organisation works on a less visible but equally important part of this question: who, within an African country, is actually able to build, evaluate and answer for an AI system.

Gabriela Ramos and Emilija Stojmenova Duh make an important advance by shifting the debate from data centres and procurement contracts to the citizen. Their provocation rightly argues that sovereignty is about preserving democratic agency, not merely controlling infrastructure. I agree with the diagnosis, and the remedies they propose are necessary. Yet there is a deeper layer, largely invisible in current debates, that lies beneath the fine print and beyond the ownership of infrastructure. I extend the concept of cognitive sovereignty to AI governance, defining it as the capacity of a society to ensure that the intelligence governing its public life can understand its languages, reflect its knowledge, and remain subject to its own human judgment.

Beyond the Fine Print

Digital sovereignty has evolved through three increasingly demanding conceptions.

Infrastructure sovereignty asks where the servers sit and who owns the compute. This is the debate of the last decade, and the one the provocation rightly describes as insufficient.

Citizenship sovereignty, the provocation's own contribution, asks whether the citizen's relationship to the state remains legible, contestable and accountable once mediated by proprietary systems. This is a genuine advance on the first.

Cognitive sovereignty asks a third question. Whose language, whose categories, whose memory and whose reasoning patterns does the system encode? A model is not a neutral instrument. It is a compression of a corpus, and every corpus is a selection of the world.

This is not a metaphysical concern. Joshi et al. (2020) showed that the world's languages sit in a steep resource hierarchy within natural language processing, with a small group absorbing most data and research attention while the majority remain unserved. Birhane (2020) named the consequence: systems conceived elsewhere and deployed here do not merely fail to fit, they impose. Couldry and Mejias (2019) describe the wider arrangement as data colonialism, the appropriation of human life as raw material for capital.

This creates a real problem for governments. A state can achieve full data residency, negotiate every safeguard the provocation recommends, insert portable exit and meaningful penalties, and still find that its citizens meet the state through a system that cannot hold their language, their idiom, or the categories by which they understand their lives. No clause has been breached. Something has been surrendered. No clause can remedy a system that is fundamentally blind to the society it governs.

Sovereignty as capability, not possession

The provocation's working definition of sovereignty is possessive: what a state owns, controls or can contractually compel. I want to propose an alternative, drawing on the capability tradition in development economics.

Sen (1999) argued that development is best understood not as the accumulation of resources but as the expansion of what people are actually able to do and to be. Resources matter instrumentally. Capability matters intrinsically. The distinction is not just academic. Two countries holding identical resources can have radically different freedoms, because converting resource into capability depends on institutions, skills and social arrangement.

This distinction extends well beyond digital governance.  Examining Nigeria's entrepreneurship education programme, we found that formal training expanded while entrepreneurial capability did not, because the institutional apparatus needed to convert instruction into real opportunity was absent (Ikebuaku and Dinbabo, 2018).

Applied to digital sovereignty, the picture changes. Sovereignty is not what a state possesses. It is what a state, and its citizens, are able to do. A country that owns a data centre it cannot staff has a building. A country holding audit rights it cannot exercise has a document. In both cases the resource is present and the capability is absent.

This is the sovereignty paradox as I would restate it. Smaller and developing nations are told they cannot be sovereign because they lack the capital to build alternatives to global infrastructure. But sovereignty was never only about the infrastructure. It was about the capability to govern what runs on it. That capability is far cheaper to build than a data centre, and it is the binding constraint far more often than compute.

The paradox also obscures what a country holds. Africa's position is usually described in deficit terms: less compute, less capital, less infrastructure. Yet the continent holds an outsized share of the world's linguistic diversity and the youngest population on earth, precisely the two inputs the next phase of AI development is short of. Scarcity runs in both directions.

The audit gap

Here the argument becomes empirical, and I can speak from measurement rather than assertion.

Through the GSMA's African AI Languages Model Project, I have led talent landscape mapping across 48 African countries. The exercise synthesised 9,937 AI job postings, 25,489 research publications and 27,349 open-source code repositories into country-level capability matrices, establishing where the continent's AI capability actually sits rather than where it is assumed to sit.

The sharpest shortages are not in model building. They cluster in the functions on which enforcement depends: model evaluation, MLOps and LLMOps, AI safety and AI governance. These are precisely the capabilities needed to exercise the rights the provocation recommends writing into contracts.

This gives the fine print argument a harder edge. A clause requiring algorithmic transparency creates an entitlement. Exercising it requires a domestic team able to specify what an audit must show, interrogate the system, evaluate the vendor's evidence, and credibly threaten to walk away. Where that team does not exist, the clause is a right on paper held by a party with no means of enforcement. The vendor knows this, and it shapes the negotiation before it begins. Ramos and Stojmenova Duh are right that governments have leverage. Leverage, however, is not a property of the contract. It is a property of the counterparty.

Sovereignty is ultimately exercised through human judgment. Every AI system, however capable, reaches a point where someone must decide whether its recommendation should be accepted, challenged or overridden. Algorithms classify, predict and optimise. They cannot bear responsibility. AI systems may automate decisions, but they cannot automate accountability. Responsibility remains irreducibly human. Sovereignty resides not only in infrastructure, code or contracts, but in the capability of people and institutions to exercise judgment over them. Training AI evaluators, auditors and governance specialists is therefore not a workforce investment. It is an investment in sovereign capacity.

One practical illustration comes from our work at Mozisha, where we train African professionals in AI evaluation, governance and human oversight and place them into organisations where those capabilities are exercised. In a pilot AI Employability Lab run with the University of the Western Cape between February and May 2026, involving approximately 100 students from law, medical bioscience, commerce and mathematical statistics, participants rated the career value at 8.5 out of 10, and every one recommended it onward. The sample is small and illustrative. It shows that sovereign capability can be built quickly and affordably inside ordinary public universities. The gap is a policy choice, not a natural constraint.

The unit of sovereignty is not the nation

The provocation closes with a suggestion deserving more than the single sentence it receives: that governments could form alliances with other countries consuming the same goods.

This is the most practical answer available to the sovereignty paradox, and it is already being tested. No single African state has the market weight to set terms with a frontier laboratory, or the corpus depth to build a competitive multilingual model alone. A bloc has both. ATLAS Umoja AI pools datasets, technical expertise and research across five founding governments, building on N-ATLAS, the open-source multilingual model Nigeria launched in 2025 as Africa's first government-backed open-source large language model.

Three features of that arrangement are worth generalising beyond Africa. First, the coalition pools the full stack of what AI actually requires: language data, expertise, talent, research and, in time, infrastructure. No single African state could assemble this alone; a bloc can. Second, the coalition includes both states and domestic firms, so capability accumulates in the local economy, not in a ministry alone. Third, the artefact is open source, which makes it a digital public good in the provocation's own sense.

The immediate test is whether the Abuja positions translate into a coordinated bloc at the ITU Plenipotentiary Conference in Doha later this year. If they do, the model transfers beyond this continent.

What follows for policy

Five recommendations.

1. Treat national language and cultural corpora as sovereign public assets. They should be licensed under enforceable terms, not scraped as free training input. One mechanism is the corpus trust: publicly governed repositories that steward national and regional language resources, set licensing conditions, ensure communities retain oversight of their knowledge, and reinvest revenues into domestic AI capability. Conceptually, the closest analogue is the sovereign wealth fund, except that the strategic asset is linguistic and cultural capital rather than mineral wealth.

2. Tie every transparency clause to a named enforcement capability. Contracts should not merely grant audit rights. They should specify who will exercise them. Where domestic capacity does not exist, the contract should fund its creation as a condition of award, with the training obligation running to public institutions rather than the vendor's staff. Transparency without an auditor is not a right. It is a courtesy.

3. Require a national AI capability audit before major procurement. States routinely assess fiscal and legal readiness before signing. Capability readiness belongs on the same footing, using the kind of mapping the GSMA exercise has shown feasible at continental scale.

4. Draw the red line at unexplainable adjudication. No proprietary, non-interpretable model should determine an outcome against which a citizen holds a right of appeal, including welfare eligibility, immigration status, publicly guaranteed credit and any criminal justice function, unless the state can explain that determination in a language the citizen actually speaks. This is where cognitive sovereignty stops being abstract. An explanation the citizen cannot understand is not an explanation, and a right of appeal that cannot be exercised is not a right.

5. Negotiate as blocs, and license as blocs. Regional bodies should hold pooled corpus licensing and pooled procurement mandates. The asymmetry between one small state and a frontier laboratory is unbridgeable. The asymmetry between a continental bloc holding nearly one third of the world's languages and that same laboratory is a negotiation.

Return to Abuja

Ramos and Stojmenova Duh write that if a government cannot explain, audit or alter the systems that govern its people, then it is no longer fully sovereign. I would extend the sentence. A government cannot explain what it cannot read. It cannot audit what it has no one trained to interrogate. And it cannot alter what was never built in a language its people speak.

The ministers who met in Abuja understood this. They did not begin with infrastructure. They began with language, recognising that a continent holding nearly one third of the world's linguistic diversity is served by AI systems that meaningfully support fewer than 2% of its languages. That is the fine print no lawyer drafted, and no procurement clause alone can repair.

The future of digital sovereignty will not be decided by those who own the largest data centres, but by those capable of questioning, governing and ultimately reshaping the intelligence that runs upon them. Sovereignty that lives only in contracts, and not in the people capable of exercising judgment over the systems those contracts govern, is sovereignty on loan.

References

Birhane, A. (2020) 'Algorithmic Colonization of Africa'. SCRIPTed, 17(2), pp. 389–409.

CIO Africa (2026) 'African Ministers Adopt Connectivity Pact, Launch Pan-African AI Language Initiative'. CIO Africa, 27 July. Available at: https://cioafrica.co/african-ministers-adopt-connectivity-pact-launch-pan-african-ai-language-initiative/ (Accessed: 27 July 2026).

Couldry, N. and Mejias, U.A. (2019) The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford, CA: Stanford University Press.

GSMA (2026) GSMA African AI Languages Model Project: Talent Mapping Dashboard. Available at: https://www.gsma.com/talent-mapping/dashboard/ (Accessed: 27 July 2026).

Ikebuaku, K. and Dinbabo, M.F. (2018) 'Beyond entrepreneurship education: Business incubation and entrepreneurial capabilities'. Journal of Entrepreneurship in Emerging Economies. doi:10.1108/JEEE-03-2017-0022.

Joshi, P., Santy, S., Budhiraja, A., Bali, K. and Choudhury, M. (2020) 'The State and Fate of Linguistic Diversity and Inclusion in the NLP World'. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pp. 6282–6293.

Ramos, G. and Stojmenova Duh, E. (2026) Digital Sovereignty Is in the Fine Print. Draft provocation for The New Alignment, June 2026. London: UCL Institute for Global Prosperity.

Sen, A. (1999) Development as Freedom. Oxford: Oxford University Press.