AI safety, in the languages
the models forget.
We study how frontier AI models behave in African languages, starting with Kiswahili and Sheng. Most safety testing happens in English. The models are then shipped to hundreds of millions of Kiswahili speakers, largely untested in the language they speak.
A dedicated lab for African-language AI safety.
Hekima AI Research operates as a dedicated lab (working name Sauti, Kiswahili for “voice”). We build the things that let a model’s safety be measured where it has not been: red-team datasets, rigorous evaluations and testing by fluent native speakers rather than machine translation.
The findings are meant for the organizations accountable for model safety: frontier AI labs, safety institutes and the teams building guardrails. The scarce ingredient is not compute. It is people who can attack a model fluently in Kiswahili, Sheng and beyond, and document what they find with scientific care. That is what we are here to do.
The work is held to one rule above all others: honesty. We publish nothing we have not measured, and we do not overclaim what we find.
Publications are on the way.
This is where our papers, datasets and evaluations will live as each one is ready. None is posted before it clears independent review.
The Swahili Jailbreak Report
A review of how little frontier-model safety has been measured in Kiswahili, and a pre-registered method for measuring it properly with native-speaker red-teamers. In preparation. Nothing is published until it carries a named author and independent expert review.
African-language safety leaderboard
An open, reproducible comparison of how major AI models refuse or comply with harmful requests across African languages and their code-switched variants. Planned as the free, public face of the work.
Work in AI safety?
If you are at a lab, a safety institute or a research group and care about how models behave in under-tested languages, we would like to talk about evaluations, datasets or native-speaker red-teaming.
Get in touch →