Most conversations about AI and language work assume a starting point: that the language in question already has enough data behind it. Enough text, enough transcripts, enough training material for a model to learn from.
Kinyarwanda doesn’t have that. Neither do most of the languages spoken across East Africa.
I’ve spent twenty years working across Kinyarwanda, English, French, and Swahili — legal translation for international tribunals, humanitarian communication, terminology work where getting it wrong has real consequences. That work taught me something the AI conversation often misses: language isn’t just vocabulary and grammar rules a model can absorb from a large enough dataset. It’s context, register, the difference between how you address a judge and how you address a neighbor — and for low-resource languages, there often isn’t enough data for a model to learn that distinction at all.
That gap is exactly where I think native speakers with deep linguistic and domain expertise have the most to offer right now — not just translating content, but helping build and evaluate the training data, terminology, and quality standards that AI systems need to actually work for languages like Kinyarwanda, rather than treating them as an afterthought.
The next chapter of AI language work won’t be won by whoever has the biggest model. It’ll be won by whoever takes the time to get the languages most models still get wrong.
If you’re working on multilingual AI, low-resource language data, or localization for African languages, I’d love to compare notes.