Africa's AI Gap Is a Language Gap

I've known for a long time that language would be one of the hardest problems for AI in Africa. I never expected it to be easy, or quick. Building good AI for languages with very little written data behind them was always going to take years of patient, expensive work.

What I didn't expect was how little of that work has actually been done. When I finally had a real reason to go looking, I hit a wall almost as soon as I started.

One note before I go on. I believe this is an Africa-wide problem. Almost every country on this continent works across several languages, and most of them face some version of what I'm about to describe. But I'm writing from Ghana, because Ghana is where I build and where I've seen this up close.

The moment it got personal

This year I started building a product for small Ghanaian businesses. I won't say much about it yet, except that it's the kind of tool you ask a question and it answers in plain terms. Early in the design, I did the exercise every builder should do. I pictured the people who would use it.

The answer was everyone. The accountant in Airport City with a laptop needs a clear answer just as much as the woman running a provisions shop in Kumasi, the fabricator in Suame or the aggregator moving maize out of the north. They all run real businesses, pay real taxes and deal with real paperwork. And whether they hold a degree or never finished school, most of them do their thinking, their bargaining and their worrying in Twi, Ga, Ewe or Dagbani. Given the choice, many would rather ask out loud, in the language they actually think in, the same way they'd send a voice note to a supplier. Literacy decides whether someone can type a question in English. It doesn't decide which language they'd rather ask a hard question in.

So the product needed voice, and it needed our languages. That's when I started digging into what was actually available. I expected the work to be early. I didn't expect it to be this thin.

What I found when I looked

If you're building in English, voice is close to a solved problem now. Platforms like ElevenLabs have made natural, human-sounding AI voices something a developer can plug into a product in an afternoon. The voices are good enough that you sometimes forget you're listening to a machine.

Look for that kind of capability built for Ghanaian languages and you come up almost empty. As far as I can find, there's no homegrown equivalent. Nothing a builder here can pick up and trust with real customers.

The closest thing we have is the work of Ghana NLP, a group that has done a lot of its work as volunteers, and their app Khaya, which works on speech and translation for Ghanaian languages. I have a lot of respect for what they've built with very little support. But I'll be honest. In my opinion it's still very early, and not yet ready to carry a production product that real businesses depend on. That isn't a criticism of the people. It's a measure of how little has been invested in the problem.

Text isn't much better. An independent benchmark called AfroBench tested large language models across 64 African languages. Even the strongest commercial models scored around 20 points lower in African languages than in English. The open models, the ones an African startup can actually afford to run and adapt, dropped by around 40 points. The reason is simple. These models learn from text on the internet, and there's very little of our languages on the internet.

None of this is secret. It just looks very different when it's the thing standing between your product and the people you built it for.

My mother and the voice note

My mother never had much formal education. She can't comfortably read or write a text message. But she's on WhatsApp every day, and she uses it entirely through voice notes. She records, she sends, she listens to the replies. Without that one feature, WhatsApp would be closed to her. With it, she's as connected as anyone in the family.

I think about that a lot now. One small feature brought her into the digital world on her own terms, in her own language, without asking her to become someone else first.

She isn't unusual. Ghana's 2021 census found that about three in ten people aged six and older can't read or write. Yet there are close to 42 million mobile connections in Ghana for a population of about 35 million, and roughly three in four people are now online. The phones are already in people's hands. The habit of speaking to them is already there. What's missing is AI that can listen back.

The bigger risk

This is where it stops being about my product, or anyone's product.

AI isn't just another app. It's quickly becoming the way people learn, get advice, find information, deal with institutions and do their work. Over the next decade it will sit inside education, healthcare, agriculture, banking and government services. Whoever can use it well will move faster than everyone else.

Right now, in Ghana and across Africa, the people who can use it well are mostly the ones who already had the advantages. Educated, comfortable in English, usually urban. Everyone else, which is most people, is being left on the other side.

We've seen a version of this before. The first digital divide was about access to phones and the internet, and mobile closed a lot of that gap faster than anyone expected. This one could be wider, and harder to close, because AI compounds. Every year the people who can use it pull further ahead of the people who can't.

There are plenty of other problems holding African AI back. Cost, connectivity, power, skills, data centres. I don't dismiss any of them. But I've come to believe language is the biggest risk of all, because it's the one that decides who gets to take part in the first place. You can fix the cost of a data bundle. You can't fix being spoken to in a language you don't think in.

If we don't solve it, Africa won't harness AI. AI will simply happen to Africa, shaped by other people's languages, built on other people's knowledge, and used fully by a small slice of our population.

In fairness to English

I want to be fair here, because this isn't a story about villains.

English is Ghana's official language. Most of the people building and buying software here work in it comfortably, and they're getting real value from AI today. Supporting dozens of languages properly is expensive, and the companies building global models have to start somewhere. The models are getting better every year.

All of that is true. But "start in English" has quietly turned into "stay in English," and the people left out by that choice are the majority of our population.

Who's already doing the work

The encouraging part is that Ghanaians haven't been waiting for permission. Groups like Ghana NLP have shown that the talent is here and that people care enough to work on this for free. That should embarrass the rest of us a little. It certainly shouldn't be the only plan.

Government has started saying the right things too. At the launch of the National AI Strategy in April, President Mahama called for our indigenous languages to be integrated into Ghana's AI work, alongside a $250 million AI computing centre.

I welcome that. But a computing centre with no Ghanaian language data in it will train very fast models that still can't understand a trader in Kejetia. The words were right. What matters now is whether the money follows them.

What I'd change

Government needs to start investing in language datasets, and take it far more seriously than it does today. Recording, transcribing and openly publishing speech and text for our major languages is national infrastructure, in the same way roads are. Everything anyone builds later depends on it.

Make it a real line in the AI strategy. Not a sentence in a speech, a budget with a target attached. How many hours of Twi, Ewe, Ga, Dagbani and Fante speech will exist as open data two years from now? Somebody should own that number.

Back the people already doing it. Ghana NLP and others like them have proven they can build. They need grants, contracts and compute, not just applause at events.

Use government's buying power. When public institutions procure AI tools, make voice and local language support a requirement. That alone would create a market for this work and force every serious vendor to take our languages seriously.

And to my fellow Ghanaian builders, this part is on us. Design for voice from the very first sketch. Test your product with someone who doesn't write English. Use whatever open language data exists, and give back to it when you can. Every AI product we ship in English only is a small vote for the divide. Every one that ships with voice in a local language, even a rough first version, makes the next one easier to build.

I keep coming back to my mother and her voice notes. She didn't need to change for WhatsApp. WhatsApp met her where she was.

Until AI can understand my mother, it hasn't really arrived in Africa.