Francis Elendu, a secondary school teacher in Enugu, Nigeria, goes out into the world with two tools: a guide cane and his phone. The cane is, in his words, "like the best tool for mobility." The phone runs apps like Envision AI and Seeing AI that show him poles and overhead objects the cane can't detect. Most days, that works. Some days, it doesn't — because the navigation and vision models that power those apps were trained on streets that look like streets in America or Europe, not on Nigerian roads, not on the textures of his actual environment. "I don't know if when they were developing the apps, they considered some places that are not America or Europe," he says. "Sometimes, these apps will be seeing things that are completely different."
Elendu's experience is one user's data point. Scale it up and it becomes a civilizational one: more than 100 million Africans live with a disability, and according to WHO's 2022 Global Report on Health Equity, over 1 in 8 people across the 47 countries of the WHO African Region have a significant disability. Yet the local sign languages, speech patterns, and lived realities of those 100 million people are, for the most part, absent from the datasets that train global AI. A continent is about to build its AI future on top of models that were never taught how to see, hear, or understand its disabled citizens. And in Nairobi, that gap just got a coalition aimed at closing it.
The Disability Data Desert
The evidence for the gap comes from a 2025 scoping study commissioned by Artificial Intelligence for Development (AI4D), which examined Ghana, Kenya, and Rwanda with more than 585 participants. The findings are blunt: persons with disabilities are underrepresented in the datasets used to train AI systems, and the disability data that does exist is fragmented and poorly mapped. Kenyan Sign Language data, for example, is extremely limited compared with American Sign Language. Kinyarwanda voice data involving speech impairments is, for all practical purposes, non-existent in global datasets.
The consortium calling this the "disability data desert" is deliberately understating it. A desert doesn't mean "a bit dry." It means: the water table is so far down that nobody has bothered digging, so the crops never grow, so nobody builds an irrigation system, so the water table stays far down. That's the trap, and it's why the response from researchers and disability advocates isn't another app — it's the plumbing underneath all the apps.
What HAIDI and ADDN Actually Are
Unveiled at the Global Data Festival in Nairobi, the Hub for AI and Disability Inclusion (HAIDI) and the African Disability Data Network (ADDN) are two complementary initiatives backed by the International Development Research Centre (IDRC), the UK's Foreign, Commonwealth and Development Office (FCDO), and the AI4D programme, in a partnership involving the Responsible AI Lab (RAIL) at Ghana's Kwame Nkrumah University of Science and Technology (KNUST), the Next Step Foundation, the Assistive Technologies for Disability Trust (AT4D), and Maseno University's Centre for Applied AI (MCAAI).
The important design decision is what they're not. HAIDI is not another application layer, and ADDN is not a central data vault. ADDN is built as a coordination and discovery network: it helps researchers, startups, and policymakers find out what disability-related datasets exist, who manages them, and how to request access under that manager's own governance and consent framework. Instead of transferring ownership of sensitive disability data to a central platform — which is exactly how the last generation of development data went wrong — it acts as a bridge between those who need data and those who already hold it.
Christopher Harrison, Head of AI at Next Step Foundation, puts the logic plainly: "Every AI system begins with data. The question is whether that data reflects the diversity of the people who will ultimately be affected by the technology. When persons with disabilities are missing from datasets, their experiences, needs and perspectives risk being overlooked in the systems built from them. Closing that gap is not simply a technical challenge; it is essential to building a more equitable digital future."
The First-Year Agenda
HAIDI's first year is built around answering a question the consortium admits it can't yet answer: what disability data even exists across Africa? The scoping study surfaced the gap; it was never a full inventory. So the plan is a continent-wide mapping exercise that will locate disability-related datasets across sign language, speech, and health, document where they're held, and identify the biggest gaps across languages, disability types, geographies, and data modalities.
Alongside the mapping, HAIDI is developing three Africa-focused flagship datasets:
- Sign languages — to build what no one else has: a continental resource across multiple African sign languages, not just one national sign language.
- Mental health and sexual and reproductive health — initially with participants aged 18 to 30, with 13-17 only after parental consent arrangements are worked out.
- Inclusive employment — to give AI tools a real signal about how disabled people actually work and get hired.
The collaboration is already past the slide-deck stage. The consortium is engaging more than 80 founders building sign language technologies through a Pan-African webinar series, deliberately treating accessibility as a continental problem rather than 54 national ones. And to make the data translate into products, HAIDI launched an 18-month programme for three startups building AI for persons with disabilities, with grants of up to KES 5 million (~$38,600) each, plus technical support, community testing, and access to funding networks — open to education, employment, mental health, caregiving, and assistive technology.
The Consent Problem
Here's the part that separates this from a ordinary data-collection drive, and the part the consortium handles with unusual honesty. Disability data is intimate data. Every activity involving persons with disabilities goes through ethics review and institutional review boards; participants give informed consent; personal data is anonymised wherever possible and stored securely; participants can withdraw.
"Wherever possible" is doing a lot of work in that sentence. Fola Adeleke, an AI governance and data protection expert, notes that de-identification doesn't reliably prevent re-identification, depending on the type and scale of the data. And the withdrawal right has a real edge to it: you can remove someone's data from a published dataset, but once that data has been downloaded and used to train a model, getting it out of the model is, practically speaking, impossible without a full retraining. "Once the data is used to train the AI, it may be technically difficult or impossible to withdraw that data without retraining the model," Adeleke says. "So informed consent is crucial before the data is used."
That's a governance position that global AI labs have not been famous for applying. It's also, quietly, the most advanced disability-data framework being built anywhere, and it's being built by African institutions. That's not an accident. It's a data-sovereignty play: Africa is deciding the terms on which its own people's most intimate data is collected, governed, and reused — and it's doing it before the data gets colonised by foreign model builders.
Why 100 Million Is a Market, Not a Rounding Error
For the optimists, the framing matters. One in eight is not a charity statistic. It's a user base. It's the number of people who will use chatbots, navigation apps, hiring tools, and health assistants for the rest of their lives, and it's the number of people those products will either include or keep excluding. Abdulazeez Hamdallah, a Lagos-based disability advocate with a hearing impairment, runs his working life on ChatGPT, Google Meet, Microsoft Teams, and Otter.ai — and tells TechCabal that live captioning routinely fails on accents and background noise, and that sign language support is poor across the board. "I have been in meetings and places where I did not understand what was being said," he says. "So, it can be very frustrating at times."
Every one of those frustrated users is a customer the global AI industry has not served. A sign-language model that works across African sign languages, a voice model that understands Kinyarwanda with speech impairments, an employment tool trained on how disabled people in the region actually work — none of that exists yet, and the coalition in Nairobi is literally mapping where it needs to be built. Yinka Olaito, Executive Director of the Centre for Disability and Inclusion Africa, is direct about the stakes: "It is important for us to think about AI through the lens of disability inclusion because otherwise we will continue to have discrimination around speech and other areas of disability, leaving people out of technologies that should benefit them."
What It Means
1. Inclusion is now a data-infrastructure question, not a features question. The industry keeps asking "which checkbox do we add to the accessibility settings?" The harder and more correct question is upstream: whose data was the model trained on, who governs it, and can the people in it withdraw consent in a meaningful way? HAIDI and ADDN are betting the future of African AI is decided in that upstream layer, not in the product UI.
2. The withdrawal limitation is the single most important sentence in this story. "You can remove it from the dataset, but not from the model that already learned it" is the exact constraint that every AI governance framework on Earth should be engineered around, and it's being articulated first in a pan-African disability-data context. That's a genuine first-mover position on AI governance that the continent will own, whether the West notices it or not.
3. This is the quietest and most consequential data-sovereignty move of the year. No chips, no export controls, no compute wars. A network of universities, trusts, and disability organisations deciding, before foreign labs build the default, that the data representing 100 million Africans will be discovered, governed, and reused on terms set by the people it represents. The EV playbook had an industrial version. This is the civic version, and it's being written on the ground.
🔥 Hot Takes
1. The $38,600 grants are the least important number in this story, and the smartest one. A handful of six-figure grants to three startups is nothing on a global scale. But it's the pipeline that turns a data coalition into a commercial ecosystem. Whoever builds the first truly good African sign-language model gets a distribution channel (the ADDN network, the 80+ founders, the disability organisations) that no VC fund in London or San Francisco has ever assembled. That's an unfair advantage, and it's being built for free, right now, in Nairobi.
2. "1 in 8" should be on the cover of every African national AI strategy, and most of them don't have one yet. HAIDI's one-year agenda includes assessing how disability inclusion shows up in national AI strategies across the continent. That framing quietly tells you the answer you're going to get: it mostly won't. The gap between "we are building an AI economy" and "our AI economy knows what 125 million of its citizens are" is the single biggest unpriced risk in African AI, and this is the first public document that puts the two halves of that sentence on the same page.
3. The West should be paying attention to the consent architecture, not the datasets. The datasets will be useful. The consent-and-withdrawal framework is the export. The moment an EU or US lab wants to train on African data — sign language, mental health, employment — they're going to have to do it through ADDN's governance, on ADDN's terms, with ADDN's withdrawal rules. That's how you build a data exchange without building a data colony, and it's being designed by the parties most historically harmed by the other way of doing things.
The Bottom Line
Africa's AI boom is real, and it has a blind spot that is measured in millions of people: 100 million disabled citizens whose languages, voices, and working lives have never been properly represented in the systems that will shape their next two decades. HAIDI and ADDN are not going to fix that with a model. They're going to fix it with a map, three datasets, a discovery network, and a consent framework that takes withdrawal seriously — and by doing so, they're building the data-sovereignty floor that any future African AI economy will stand on. Whether the continent's next trillion-dollar industry includes its disabled citizens, or quietly prices them out again, is now a logistics question. And for the first time, the answer is being written by the people who live inside the problem.