If it can't speak Twi, it isn't general
Why fluency in African languages is a prerequisite for artificial general intelligence, not a localization afterthought.
Ask one of today’s leading AI models to explain a loan agreement in English, and you will usually get a clear, careful answer. Ask for the same explanation in Twi, Ewe, Hausa or Tigrinya, and the answer often gets worse: stiffer phrasing, thinner vocabulary, sometimes a quiet slide back into English. Many speakers of these languages know that gap well. We think it says something important about how far the field still is from general intelligence.
Picture a cloth trader in Kumasi who wants help pricing a new shipment, answering a supplier in Ho and planning for a slow month. She thinks and bargains in Twi; her supplier writes back in Ewe. A system that can reason brilliantly in English but cannot work with her in both languages is not general. It is specialized in the internet’s majority languages.
Fluency in African languages, in other words, belongs inside the definition of artificial general intelligence. Here is why, and what we are doing about it at SISU AI.
General for whom?
The “general” in artificial general intelligence carries a lot of weight. It promises a system that can learn and reason across the full range of human activity, and much of that activity runs through language. People teach, trade, pray, argue and remember in words.
Humans live in thousands of languages. More than 2,000 are spoken in Africa alone, and more than 75 of them have over a million speakers. For many people on the continent, one of these is the language of home, of the market and of first thoughts.
So any claim of generality invites a follow-up question: measured against what? Measure a model against the text that happens to be online, and English counts for more than any other language. Measure it against the people who might use it, and the bar rises sharply. Only the second measure deserves the word.
Generality should be measured across humanity, not across the web’s majority.
How the gap gets built
Start with the data. Large language models learn from enormous amounts of text, much of it crawled from the web. Close to half of what gets crawled is in English, and most African languages barely register. A model can only become fluent in what it has read.
Then there is the token tax. Models read text in small pieces called tokens, and every token costs computing power. Research has shown that tokenizers split many non-English languages into far more tokens than English. So the same conversation can cost more in those languages, and less of it fits into the model’s working memory.
Next come the marks on the page. Yorùbá, for example, is written with tone marks that help tell words apart, and text pipelines often strip diacritics like these as if they were noise. What survives is a flatter, more ambiguous version of the language, and that is the version a model learns.
Finally, a great deal of knowledge is spoken rather than written. It lives in stories, songs, sermons and conversations between grandparents and grandchildren. Very little of that reaches a web crawl, and none of it should get there without the consent of the people who hold it.
Each of these causes comes down to choices about what to collect, how to clean it and what to measure. Different choices are possible.
Why AGI needs Twi
It is tempting to treat all of this as a localization problem: build the intelligence first, translate it later. We think that gets the order wrong, for five reasons.
- Coverage. Generalizing across languages is part of what general intelligence means. A system that shines only where its training data is thick may simply be good at its training data. Carrying skills into languages with little text is one of the clearest tests of whether a model has learned something that travels.
- Knowledge. African languages carry knowledge about farming, medicine, ecology and history, and some of it may exist nowhere in English. Imagine a farmer in northern Ghana explaining, in her own language, how she reads the soil before planting. If no model can understand her, what she knows stays out of reach of the systems we are calling general.
- How people actually talk. For many speakers, moving between English, Pidgin or Sheng and a local language is the norm, sometimes inside a single sentence. A model that handles each language only in isolation will stumble over ordinary conversation. Real-world generality means keeping up with the mix.
- Safety. Research has shown that safety guardrails tuned in English can be bypassed in low-resource languages. A guardrail that holds in English but gives way in another language protects only some of the people using the model. Responsible AI has to be multilingual.
- The next billion users. Africa is home to about 1.5 billion people, with a median age of about 19, and its population is projected to reach roughly 2.5 billion by 2050, about a quarter of humanity. The next billion users will not type in English, and many of them would rather not type at all.
What fluency sounds like
Fluency is a higher bar than translation. A word-for-word rendering of a Twi proverb into English can be accurate and still miss the point entirely, and the same is true in the other direction.
A fluent system knows the idiom and the proverb. It hears tone, both the kind that changes what a word means and the kind that turns a remark into a joke or a rebuke. It knows which honorifics show respect to an elder, when humor is a polite way of saying no, and when a question is not really a question.
No one should have to translate themselves to be understood by a machine.
Fluency also has to be spoken. Many people prefer speaking to typing: it is faster, it does not depend on a keyboard or on reading a particular script, and it is how a great deal of knowledge has long been passed on. That is why we think voice should come first, with text close behind.
Building Alkebulan
Our tagline at SISU AI is “Intelligence without borders,” and language is often the first border. Alkebulan, our Pan-African large language model, is multilingual by design and built to preserve what generic corpora miss.
The v0.0.17 release, “Accra,” expands Alkebulan’s cultural knowledge packs, adds foundational support for edge-friendly deployment and includes responsible-AI guardrails tuned with community reviewers to flag cultural misinterpretations. We think that last choice is the right one: the people best placed to spot a cultural misreading are the people being misread.
More is in development. The roadmap includes a Pan-African storytelling mode co-created with historians and cultural custodians, a regulatory toolkit with data-residency controls, a datasets pipeline and hub, and integration with Quadrantica+, efficient multilingual NLP built for cost and quality at scale. On storytelling, we want the people who keep the stories to shape how a model tells them. On efficiency, the reasoning is the same as for the token tax: when a language is expensive to serve, its speakers end up paying.
The same commitment runs through our companies. Vaxus launched VAXUS GPT in 2023 for climate awareness in local languages, and Vibelane aims to bring mental health support to people in their native language. Altera fine-tunes language models for personalized learning, with speech processing, in pursuit of “AI that speaks your language.”
Our partner OpenAI4Africa, founded in 2023 under the banner “Borderless AI: for Africans, for all,” lists a set of words that makes this argument better than any essay could. Every one of them means “language.”
Asụsụ. Èdè. Yare. Rurimi. aneɛ. Gbegbɔgblɔ. Kasa. Wiemɔ. Lugha. لغة. Taal. Fiteny.
That is 12 words for one idea, and Africa’s more than 2,000 languages hold many more. Several carry accents, dots or special letters such as ɛ and ɔ, and one is written in Arabic script: exactly the kind of text a careless pipeline would mangle. If general intelligence is the goal, every language behind those words is part of the test.
What we ask of the field
No single company can close this gap, and it should not be closed carelessly. We would ask everyone building large models, ourselves included, to commit to four things.
- Collect data with consent. Voices, stories and conversations belong to the people who share them. Contributors should know how their words will be used and be free to say no.
- Pay contributors fairly. Recording, transcribing, translating and reviewing African-language data is skilled work, and it should be paid as skilled work.
- Build in the open. Where contributors agree, publish open datasets and benchmarks for African languages, so progress can be checked in public and no single company owns the yardstick.
- Report results per language. A single multilingual average can hide a model that fails badly in the languages with the least text online. Publish the score for every language, especially the weak ones.
Back to Kumasi
Return to the trader in Kumasi. What would actually help her is a model that understands her when she speaks Twi, answers in Twi that sounds like Twi, keeps up when she slips into English mid-sentence, and costs little enough to use every day. Building it is slow, careful work, and it is the same work general intelligence demands.
That is a high bar, and the right one. The day an AI can bargain with her in Twi and hold its own, we will be glad to call it general.