I. I was sitting with a pamphlet in front of me, an AI safety awareness flyer we were preparing for a project called Safer Schools, Safer Communities, designed to be printed and displayed as posters in schools and community spaces across northern Nigeria. My job was to translate it into Hausa so that the students, teachers, and community members who walked past it every day could actually read it and understand it.
I got stuck almost immediately.
Not because the Hausa language is limited. Hausa is one of the most widely spoken languages in Africa, with over 60 million native speakers and a literary and oral tradition that stretches back centuries. It is a language capable of precision, nuance, and abstraction. The problem was not Hausa. The problem was that the concepts on that flyer had been built entirely inside a world that had never once considered Hausa speakers when it decided what AI safety should mean.
Take the word deepfakes. In English, one word carries a precise, widely understood meaning. In Hausa, I had to write an entire sentence: an image or recording that has been convincingly altered and manipulated to misrepresent someone as doing or saying something that was not actually done or said. The concept existed in every language and culture, but the vocabulary did not. Then I reached the word prompt. Not the everyday meaning of the word, but its technical AI meaning is the instruction a human gives to an AI system. There is no Hausa equivalent. Not an approximate one. Not a borrowed one. Nothing. And prompt is not a peripheral term in AI safety. It is foundational. Every subsequent concept that builds on it, prompt injection, adversarial prompts, the relationship between human instruction and machine behaviour became harder to convey with each word I turned the page to.
This was not a policy brief or an academic paper. It was a flyer. Something designed to be understood at a glance by a fourteen-year-old walking past it in a school corridor in Gombe. When a complex AI safety concept fails to translate in an academic paper, it is a problem. When it fails on a poster designed for a child in a northern Nigerian classroom, it is a different order of problem entirely.
So I did what translators in this situation always do. I approximated. I borrowed it. I used the closest available phrase and hoped the meaning carried across. Sometimes it did. Often it didn't. And every time it didn't, I thought: if the language of AI safety does not exist in Hausa, then the communities I work with are not just excluded from the conversation, they are excluded from the protection the conversation is supposed to provide.
II. In March 2026, Anthonio Oladimeji wrote on the EA forum that the AI safety conversation is missing 1.4 billion people. He is right. And in 2023, Ashura Batungwanayo and Hayley Martin raised a harder tension that AI safety messaging does not inherently speak to the immediate challenges Africans face daily, and that focusing on existential risks can inadvertently diminish the urgency of poverty, education, and health. That tension is real and I want to engage with it honestly.
My argument is that the Hausa language exclusion sits on both sides of that divide at once. It is an immediate problem happening now, in communities where AI systems are already being deployed without Hausa-language safeguards and it is a long-term safety problem, because governance frameworks built without Hausa speakers will have systematic blind spots that matter most when the stakes are highest. This is not a choice between urgent and existential. It is an example of where they are the same thing.
I am writing from the Hausa context because that is where my direct experience lies. But I want to be clear: this is not a northern Nigeria problem. Yoruba and Igbo Nigeria's other two major languages, each spoken by tens of millions of people, are equally absent from AI safety discourse, safety evaluations, red-teaming exercises, and governance frameworks. The three languages together represent the vast majority of Nigeria's 200 million people. None of them have a meaningful presence in the rooms where AI safety is being designed. The exclusion I am describing is not regional. It is national. And it is a failure that Nigeria's own emerging AI governance frameworks currently being drafted without reference to any of these languages are on track to institutionalise permanently if no one intervenes.
This is what I mean by epistemic exclusion.
It is not simply that Hausa, Yoruba, and Igbo speakers are underrepresented in AI safety research though they are, almost entirely. It is that the knowledge structures of AI safety have been built in ways that make them invisible as subjects of concern, as sources of expertise, and as participants in governance. When safety evaluations are conducted only in English, they cannot catch failure modes that emerge specifically in these linguistic and cultural contexts different metaphors, different social norms, different ways of expressing refusal, consent, or trust. When red-teaming exercises happen only in English, the edge cases that matter in Lagos, Enugu, or Kano never get tested. When governance frameworks are designed without these communities at the table, the blind spots that result are not random. They are systematic. And they will matter most when AI systems are deployed in exactly the communities that were excluded from designing their safeguards.
This is not a diversity argument. It is a safety argument. I am not asking for inclusion as a courtesy. I am pointing out that the field of AI safety has a specific technical and governance failure the failure to build safety knowledge in the languages and contexts of the majority of the world's population and that this failure has consequences the field should care about on its own terms.
III. Let me be concrete about what those consequences look like from where I sit.
I am currently the Stakeholder Relations Officer at New Incentives, a GiveWell top-recommended organisation running a conditional cash transfer programme for childhood immunisation across nine states in northern Nigeria. I work daily with state primary health care agencies, government ministries, and community health workers. I have spent years building the kind of institutional trust in Gombe State across 13 government ministries and departments that makes evidence-based interventions actually work on the ground.
Nigeria is moving faster on AI governance than most African nations. It has signed the Paris Charter on AI, released a National AI Strategy calling for a dedicated governance regulatory body, introduced a National AI Bill proposing a National AI Regulatory Authority, and, as of May 2026, has an Internet Code of Practice in force that includes AI deployment notification requirements. President Tinubu has already deployed AI-enabled surveillance cameras across Plateau State. The Central Bank has mandated AI for real-time financial monitoring. AI is not arriving in Nigeria; it is here.
But none of these governance documents mentions Hausa. Not one framework, standard, or regulatory proposal has been designed with explicit attention to what it means to govern AI in a context where the primary language of community trust, local government, and daily life is not English. Nigeria has no comprehensive AI law in force yet, and the frameworks being drafted to fill that gap are being written entirely without Hausa speakers at the table. The gap I am describing is not theoretical. It is the gap between what is being built and who it is being built for.
The biosecurity dimension makes this more urgent still. Northern Nigeria has lived through insurgency, repeated disease outbreaks, and the constant challenge of building health-seeking behaviour in communities where trust is hard-won and easily broken. AI-enabled misinformation campaigns, synthetic biology threats, or the failure of AI-assisted surveillance tools to account for Hausa-language communication patterns, these are not distant risks. They are risks with a specific geography, a specific language, and a specific community of people who have almost no representation in the rooms where the safeguards are being designed.
When I ask myself what an AI-enabled bioattack would look like in northern Nigeria, how it would spread, how communities would respond, how institutions would coordinate, I realise that the entire response infrastructure depends on trust networks and communication channels built in Hausa. An AI safety community that has never engaged with Hausa has also never seriously engaged with that scenario.
IV. I want to acknowledge what exists because the ecosystem is more alive than it appears from the outside, and these efforts deserve to be in the same conversations as the Anthropic safety teams and the GovAI fellows, not operating in parallel universes.
Masakhane has done foundational work on African language NLP, including Hausa datasets and models. The African Institute for Artificial Intelligence, where I serve as Strategic Advisor, is working to bring African voices into AI governance across the continent. Cecil Abungu's team at ILINA is doing some of the most rigorous thinking on African-context AI development. Sumaya Nur Adan's CASA centre and her work with the Oxford AIGI are building the institutional connections between African researchers and the global governance architecture. The UCT African Hub on AI Safety, Peace and Security and the Global Centre on AI Governance's white paper Toward an African Agenda for AI Safety are providing the policy frameworks this work needs.
Anthonio Oladimeji, in the comments on his post, put it well: the problem is not that the work does not exist, it is that it is not visible enough to the broader EA and AI safety community. I agree. And I want to add one more layer: even within this growing ecosystem, AI safety, specifically alignment research, safety evaluations, red-teaming, and biosecurity governance has almost no engagement with Hausa as a language. You can find Hausa NLP datasets. You cannot find Hausa red-teaming exercises. You can find African AI ethics guidelines. You cannot find biosecurity governance frameworks designed for contexts where Hausa is the primary language of community trust.
That is the specific gap this piece is pointing at.
V. Here is what I think should happen.
First, AI safety organisations should fund translation and localisation work as a safety priority, not a communications add-on. Translating safety concepts into Hausa properly, with the linguistic and cultural work that real translation requires is not outreach. It is safety infrastructure. It should be funded accordingly.
Second, red-teaming and safety evaluation programmes should actively recruit Hausa, Yoruba, and Igbo speakers and build evaluation datasets in these languages. The failure modes that matter in Lagos, Enugu, and Kano will not be discovered by English-speaking evaluators working with English-language prompts. This is a technical gap with a straightforward solution if the field decides to prioritise it.
Third, the Nigerian governance frameworks currently being drafted, the National AI Bill, the National AI Strategy, the Internet Code of Practice should include explicit language representation requirements. A governance framework for a country where the three largest languages are Hausa, Yoruba, and Igbo, written entirely in English and without reference to any of them, is not a complete governance framework. Civil society advocates, including those of us working at the intersection of public health and AI safety, should be pushing for this in every consultation process we can access.
I am trying to do my part. Through my work with the African Institute for Artificial Intelligence, I am advising on AI governance from a northern Nigeria perspective. I am working to establish an AI safety and biosecurity community in northern Nigeria, bringing government health officials, civil servants, and civil society advocates into structured conversations about AI safety using materials that I am working to make available in Hausa. The Safer Schools, Safer Communities project, where I first encountered the translation gap that opened this piece, is one concrete attempt to build that infrastructure from the ground up.
And I am writing this piece because the first step to closing an epistemic gap is naming it precisely enough that people can act on it.
VI. When I was sitting with that flyer and running out of Hausa words, I was not experiencing a translation problem. I was experiencing the consequence of a field that built its knowledge without asking who else needed to be in the conversation.
Sixty million Hausa speakers are not a footnote to AI safety. Neither are the tens of millions of Yoruba and Igbo speakers alongside them. They are not edge cases to be addressed after the hard problems are solved. They are people who will live with the consequences of decisions being made right now in English, in rooms they were not invited into, using concepts that do not yet exist in their languages. And as Nigeria's AI governance frameworks take shape without a single reference to the languages spoken by the majority of its population, that exclusion is becoming structural.
The communities most vulnerable to AI-enabled harm are the least represented in the rooms designing the safeguards. That is not an accident of geography. It is a choice that the AI safety community and Nigeria's own policymakers still have time to make differently.
Hadiya Usman is the Stakeholder Relations Officer at New Incentives and a Strategic Advisor at the African Institute for Artificial Intelligence. She is completing an MPA in Economic and Sustainable Development at Iconic Open University and working to establish an AI safety and biosecurity community in northern Nigeria. She speaks English, Hausa, and Fulfulde.
I used claude to help draft/edit this post; all arguments were reviewed and modified by me.
