Africa Needs AI That Serves Its Priorities

Across Africa, the most consequential uses of AI do not look like a chatbot. They look like a forecast that warns a nutrition team months before a crisis, a satellite map that shows a planner how land use is changing, a drought assessment that reaches a ministry in time to act. Most of this work runs on data that has nothing to do with language: imagery, health records, weather.

Yet whether any of it changes a person’s decision often comes down to language. One in six people worldwide has now used a generative AI product, according to Microsoft’s 2025 AI Diffusion Report, and Africa has about 1.5 billion people and more than 1,500 languages, yet most AI models were trained mainly on English and a handful of other global tongues. A farmer looking for planting advice in Dholuo, or a mother seeking health guidance in Amharic, may find that today’s systems cannot speak to them. Language is not everything in Africa’s AI story, but without it, everything else struggles to arrive.

Encouraging progress is being made, and much is led from within the continent. LINGUA Africa, an initiative of the Masakhane African Languages Hub with the Gates Foundation, the Microsoft AI for Good Lab and Google.org, funds open datasets, speech resources and practical language tools. Its recent call, designed to strengthen the language foundations needed for inclusive AI in Africa, drew more than 800 applications from 64 countries, 85 per cent of them African. The 26 selected projects span more than 50 African languages across 47 countries. Examples include Arusha Technical College’s work to build an open, community-led Tanzanian Sign Language resource, and Efficience Globale’s project in Guinea to make vaccination information accessible in Soussou, Pular and Maninka.

Voice matters as much as text on a continent with strong oral traditions. Microsoft Research Africa’s Paza project is improving speech recognition for low-resource languages, with a benchmark covering 39 African languages and new models for Swahili and five Kenyan languages, tested with farmers on ordinary mobile phones, amid patchy connectivity and background noise.

Yet fluency is not the same as usefulness. A system can answer a farmer in fluent Kikuyu and still provide a recommendation that makes no sense for the soil, the season or the family budget. Language opens the door, but trust depends on whether the advice reflects agriculture’s complex, local realities.

Data scarcity in Africa is not only a shortage of examples. It also means missing communities, outdated maps, and records that capture only the people who managed to reach a clinic. Train a model on data like that and it quietly inherits the same blind spots. Locally led data collection, documentation and long-term stewardship deserve as much investment as the models themselves. At the same time, scarcity is no reason to stand still: African innovation should be designed to work in today’s conditions, rather than waiting for perfect datasets (and compute capabilities) that may never arrive.

At Microsoft’s AI for Good Lab, where my team works on food security, public health, environmental monitoring and humanitarian response, we try to start from the decision that needs to improve rather than from a model we are eager to apply. In Kenya, together with Amref Health Africa and the Ministry of Health, we combined routine health records with satellite measurements of vegetation to forecast acute childhood malnutrition up to six months in advance. It is a promising proof of concept, not yet proof of healthier children; that takes deployment, evaluation and time.

Our geospatial work follows the same logic. With the Kenya Space Agency we built land-use maps tuned to local landscapes, and found that locally trained models can beat global, one-size-fits-all ones. Our work on building density and height from satellite imagery, including settlement growth around a refugee camp in Chad, provides critical information humanitarian planners need. Combined with population and cellular coverage data, this helps build connectivity maps that identify communities unlikely to receive early warnings disseminated through current digital infrastructure.

For AI solutions to deliver lasting impact, local institutions must own their development, deployment and long-term stewardship. To this end, we coordinated the launch of ADAPT-Kenya, a national initiative led by Kenya’s Ministry of Agriculture and Livestock Development and supported by the Gates Foundation and Microsoft’s AI for Good Lab. By convening nearly 50 partners, ADAPT-Kenya fosters the co-creation of AI-powered agricultural data products that integrate local knowledge with satellite observations to enhance crop monitoring, improve harvest forecasting, strengthen market access and trade, expand insurance services, and support disaster risk reduction.

The design has already been tested in an emergency. When drought hit the maize-harvesting counties of Kenya this season, ADAPT partners, including NASA Harvest, came together to assess conditions and give decision-makers a timely picture of the harvest. The team plans to do the same through the coming El Niño season, when decision-makers will again need timely insight for timely decisions.

As I argued recently in Nature Africa, the hardest part of AI is not the algorithm. Around 600 million people in sub-Saharan Africa still lack electricity, and connectivity, devices, skills and maintenance decide whether an impressive demonstration becomes a service people rely on every day. For Africa, these are not a distraction from the AI agenda; they are a large part of it.

Africa is not a single dataset, market or deployment environment, and its people should shape the problems, methods and standards by which progress is judged. The continent’s vibrant grassroots AI communities can lead that work, in collaboration with academia, small and medium enterprises, non-profits and UN organisations, and most importantly with governments, which play a dominant role across nearly every sector in Africa.

We should aim for more than AI that speaks Africa’s languages. We need AI that reflects its realities, strengthens its institutions and helps people make better decisions. Language belongs at the centre of that ambition, alongside good data, earned trust and the capacity to act.

Girmaw leads the Africa team of the Microsoft AI for Good Lab in Nairobi and is a member of the UN Secretary-General’s Independent International Scientific Panel on AI.