Last updated: November 2026 · Reading time: about 12 minutes
 |
| Africa has more than 2,000 languages and less than 1% of AI training data. The gap is not in the model. It is in the recorder, the notebook and the hours of transcription. |
What this guide covers
What the conference actually said
ICAIR 2026 ran from 3 to 5 November 2026 at the University of Lagos, organised by the Machine Intelligence Research Group in the Faculty of Computing and Informatics. The theme was "Artificial Intelligence for Africa's Sustainable Economic Growth," and the conference drew roughly 300 research paper submissions from students and academics across multiple universities, with keynote speakers from Stellenbosch University in South Africa.
The call on language came in the closing communiqué, signed by the Chairman of the Local Organising Committee, Dr Sennaike Oladipupo:
"Participants observed that the limited representation of African languages in digital platforms, online content and AI training datasets continues to exclude significant sections of the population from the benefits of emerging technologies."
The framing is precise, and it matters. The conference did not argue that African languages are culturally valuable and should therefore be preserved. It argued that underrepresentation in training data is an exclusion mechanism. If a system cannot read your language, you cannot use it — and the people most affected are those whose primary means of communication is an indigenous language.
The communiqué went further than a language recommendation. It called for "a deliberate shift from technological dependence to indigenous innovation, from fragmented research efforts to meaningful collaboration, and from the consumption of imported solutions to the development of technologies that reflect Africa's priorities and aspirations."
The arithmetic of the gap
The numbers are stark, and they compound.
Africa has more than 2,000 spoken languages and more than a billion speakers. But less than 1% of widely used AI training data reflects African linguistic contexts, and less than 1% of internet content is in African languages — against a continent that holds roughly one-seventh of the world's population.
That data scarcity produces a measurable performance gap. Major language models score 10 to 30 percentage points lower on African-language benchmarks than on English. Around 98% of African languages are effectively unsupported. And the shortfall is not confined to minority languages: ChatGPT recognises only 10 to 20% of Hausa sentences, despite Hausa being spoken by roughly 94 million Nigerians.
| Measure | Figure | What it means |
|---|---|---|
| African languages | 2,000+ | Scale is not the problem |
| Share of AI training data | Under 1% | The corpus is the constraint |
| Languages with NLP-grade resources | Under 5% | Most languages cannot be trained on at all |
| Benchmark performance gap | 10–30 points | The gap is measurable, not anecdotal |
| Languages effectively unsupported | About 98% | A handful of languages carry the entire load |
The model already exists
This is the part the headline misses. The communiqué's call is not for a model that Nigeria has yet to build. The country already has one.
N-ATLaS — the Nigerian Atlas for Languages and AI at Scale — is an open-source multilingual large language model built by the National Centre for Artificial Intelligence and Robotics with Awarri Technologies, under the Federal Ministry of Communications, Innovation and Digital Economy. It is fine-tuned from Llama-3 8B and supports Yoruba, Hausa, Igbo and Nigerian-accented English, with automatic speech recognition for voice applications. It was announced in 2025, and the National AI Innovation Challenge 2026 was built around it, inviting Nigerian teams to ship working tools on the model.
It is not the only one. Vambo AI released MORENA, a 1.5-billion-parameter open-weight model trained from scratch for 12 African languages including Hausa, Yoruba, Igbo, Nigerian Pidgin, Kiswahili and ChiShona. A multilingual speech recognition model called SBPN covers Yoruba, Hausa, Igbo, Nigerian Pidgin and Nigerian English. Research groups including Masakhane have published benchmarks such as IrokoBench and Uhura across more than a dozen low-resource African languages.
| Asset | What it does | Coverage |
|---|---|---|
| N-ATLaS | Open-source multilingual LLM, fine-tuned from Llama-3 8B, with speech recognition | Yoruba, Hausa, Igbo, Nigerian-accented English |
| Vambo AI MORENA | Open-weight 1.5B model trained from scratch | 12 African languages plus English and French |
| SBPN | Multilingual speech recognition | Yoruba, Hausa, Igbo, Nigerian Pidgin, Nigerian English |
| IrokoBench / Uhura | Human-translated evaluation benchmarks | 16 and 6 low-resource African languages respectively |
There is also a structural detail worth naming. N-ATLaS is fine-tuned from a Meta model. Fine-tuning an open-weight base is a legitimate and common approach, and it is far cheaper than training from scratch. But it means the base capability was developed elsewhere, and the African-language performance depends partly on how much of those languages the base model already absorbed. This is exactly the question Dr Ebun Fashina, head of UNILAG's Department of Artificial Intelligence and Robotics, put to the conference: "Are we going to build Africa's growth on other people's technologies? Is that how to have sustainable economic growth?"
The panel the communiqué treats separately
ICAIR 2026 ran a dedicated panel titled "Data Sovereignty in Africa: Economic Value, Quality of Service and the Future of Research." It is reported separately from the language recommendations. It should not be, because they are the same argument.
The panel's finding was that individuals, businesses and institutions across the continent generate valuable digital information, while the economic benefit from storing, processing and commercialising that data accrues largely to foreign technology companies. Olusola Teniola, Nigeria's national coordinator for the Alliance for Affordable Internet, highlighted the country's dependence on foreign telecommunications infrastructure, hardware, software and international internet routing, and called for investment in local data centres, domestic internet traffic exchange and indigenous technology development.
Sam Immanuel, chief executive of Semicolon, put it more directly: Africa must move beyond being a source of data and intellectual resources to becoming an owner and beneficiary of the value generated from them.
That is not a hypothetical concern. It is the mechanism behind the most common source of African-language training data, which the next section addresses.
Where the data actually comes from
One detail from the research literature explains more about the state of African-language AI than any policy document.
Much of the training data for smaller African languages is drawn from Bible translations. The reason is practical: parallel corpora — the same text in multiple languages, aligned verse by verse — are the most useful raw material for training translation models, and religious texts are among the few large, consistently translated, freely available parallel corpora that exist for many African languages.
The consequence is a domain mismatch that compounds the data gap.
| What the model learns from | What the model is then asked to do | The mismatch |
|---|---|---|
| Scriptural and formal register | Clinical instructions in a health app | Vocabulary for symptoms, dosage and consent is largely absent |
| Translated prose from an older era | Colloquial market and trading language | Everyday idiom and slang do not appear in the corpus |
| Consistent orthography in print | Inconsistent spelling in user-generated text | Code-switching and informal spelling break the model |
That is why the use cases the conference highlighted — AI tutors for children in remote areas, multilingual agricultural advisory systems, and health information for underserved communities — are precisely the ones the current data cannot support well. A model trained mostly on formal religious text will struggle with a farmer asking about cassava blight in a regional dialect.
The conference recognised this indirectly. Its communiqué called on research communities to "invest in locally relevant datasets, particularly those representing African languages, cultures and socioeconomic realities, and ensure that the communities contributing to these resources receive appropriate recognition and benefits."
What would actually close the gap
The communiqué lists recommendations across several fronts. Four of them are specific enough to act on.
- Fund datasets, not just models. The conference called for investment in locally relevant datasets covering African languages, cultures and socioeconomic realities. That means paying for collection, annotation, transcription and cleaning — the unglamorous work that no launch event celebrates and every model depends on.
- Pay the communities that produce the data. The communiqué explicitly says contributors should receive "appropriate recognition and benefits." That is a stronger position than most data governance frameworks take, and it is the difference between extraction and partnership.
- Build local compute and storage. The conference identified inadequate access to GPUs, CPUs, cloud facilities and reliable connectivity as a major impediment, and called for public and private investment in national and regional computing infrastructure, shared research facilities and accessible cloud resources. A model cannot be trained on infrastructure the country does not host.
- Strengthen data governance frameworks. The panel called for clearer rules on data ownership, local processing and commercial utilisation, so that African institutions can exercise control over data generated on the continent.
Notably absent from the communiqué is any binding requirement. The recommendations are addressed to governments, universities, the private sector and development partners, and they are framed as urgings rather than obligations. That is the structural weakness of a conference communiqué: it can name the gap precisely and still have no mechanism for closing it.
What to watch
| Indicator | Why it matters |
|---|---|
| Whether N-ATLaS adds languages beyond the current four | The gap between four languages and 500+ is the entire problem. Expansion is the test of whether this is a platform or a proof of concept. |
| Whether the National AI Innovation Challenge produces language tools | Applications closed in October 2026. What shipped on N-ATLaS will show whether the model is being used or merely cited. |
| Any published data-licensing terms | If African-language corpora are released under terms that let foreign firms commercialise them without local benefit, the sovereignty concern is unaddressed. |
| Local compute capacity | Data centres and domestic traffic exchange are the precondition for training anything at scale. Watch whether the investment materialises or stays an announcement. |
| Benchmark transparency | The 10–30 point gap is an aggregate. Published per-language results for Nigerian languages would make the gap trackable rather than rhetorical. |
| Whether ICAIR 2027 revisits the same call | If the 2027 communiqué asks for the same thing in the same words, the recommendations did not convert into action. |
Frequently asked questions
What was ICAIR 2026?
The 6th International Conference on Artificial Intelligence and Robotics, held from 3 to 5 November 2026 at the University of Lagos. It was organised by the Machine Intelligence Research Group in the Faculty of Computing and Informatics, under the theme "Artificial Intelligence for Africa's Sustainable Economic Growth," and drew roughly 300 research paper submissions.
What did the conference call for on African languages?
The closing communiqué called for AI systems capable of understanding African indigenous languages and cultural contexts, warning that limited representation in digital platforms and AI training datasets is excluding significant sections of the population from the benefits of the technology. It urged investment in locally relevant datasets and asked that the communities contributing to those resources receive appropriate recognition and benefits.
Does Nigeria already have an AI model that understands local languages?
Yes. N-ATLaS — the Nigerian Atlas for Languages and AI at Scale — is an open-source multilingual model built by the National Centre for Artificial Intelligence and Robotics with Awarri Technologies under the Federal Ministry of Communications, Innovation and Digital Economy. It supports Yoruba, Hausa, Igbo and Nigerian-accented English, with speech recognition, and was fine-tuned from Llama-3 8B.
How big is the African language AI gap?
Africa has more than 2,000 languages, but less than 1% of widely used AI training data reflects African linguistic contexts. Fewer than 5% of African languages have the resources needed for natural language processing, roughly 98% are effectively unsupported by major models, and leading models score 10 to 30 percentage points lower on African-language benchmarks than on English.
Why does the data gap matter more than the model gap?
Because a model can only learn from what it is given. Nigeria already has a multilingual model; what it lacks is the corpus to train it on at scale. Building another model without fixing the data shortage produces the same four-language coverage with a different name on the launch slide.
Where does most African-language training data come from?
Much of the training data for smaller African languages is drawn from Bible translations, because parallel corpora — aligned text in multiple languages — are scarce and religious texts are among the few large, consistently translated sources available. The result is a domain mismatch: models learn formal scriptural register and are then asked to handle clinical, agricultural or colloquial language.
What did the conference say about data sovereignty?
A dedicated panel on "Data Sovereignty in Africa" found that African individuals, businesses and institutions generate valuable digital information while the economic benefit from storing, processing and commercialising it accrues largely to foreign technology companies. It called for stronger data governance frameworks, local storage and processing capacity, domestic internet traffic exchange and indigenous software and hardware.
Is this the same call experts have been making for years?
Largely yes, which is the point worth noting. The 2026 communiqué is more precise about the data layer than earlier versions, and it lands alongside a functioning Nigerian model and a funded innovation challenge. Whether that converts the call into action is the thing to watch at ICAIR 2027.
Related reading on Dedon Tech: DevsInGovernment November 2026: what Nigeria's KOICA partnership has already built · CBN open banking framework: what the rails do, and what they don't yet cover · right of way: why Nigerian fibre costs more than it should · full site map
0 Comments: