Stewards of Their Own Map: What Kakuma Teaches Us About AI Done Well

BevanSNZ — Mon, 24 Aug 2026 00:57:22 GMT

The most interesting AI projects are rarely the ones with the biggest models. They are the ones where somebody asked a good question about who holds the knowledge. A humanitarian mapping project in northwestern Kenya is a case in point, and it carries a lesson worth bringing home.

Kakuma refugee camp is the largest in Africa. Established in 1992 to shelter young people fleeing the war in Sudan, it has grown into a community of more than 300,000 people from over 20 countries, spread across roughly 15 square miles. For years its maps were severely outdated, hindering aid delivery, infrastructure planning and emergency response. The camp's irregular layout and diverse shelter types defeated traditional mapping methods.

Three organisations tackled it together. UNHCR's innovation lab, the Hive, defined the problem. The Humanitarian OpenStreetMap Team handled on-the-ground collection and community integration. Microsoft's AI for Good Lab built the models. “Collaboration was key because each person brought something unique to the table,” says Dr Simone Fobi Nsutezo, an applied research scientist at the Lab.

The residents did the work

This is the part that matters. Data collection was led entirely by locals, from introducing the project through to flying the drones. Refugees in the camp manually identified features, trained as mappers and interpreters, and produced the ground truth. HOT's team hand-tagged 10 square miles of imagery to build the training dataset. Only then did the Lab use Azure to develop models that could recognise buildings, sanitation blocks, solar panels, power poles and lines across the camp's unusual landscape.

“Once you've trained a model on a small amount of data, it's very fast to get predictions on new areas,” says Dr Amrita Gupta, another applied research scientist on the team. “We have open-source code for mapping solar panels, buildings, roof types, sanitation facilities, and anyone can independently use it.” Every model and dataset was released openly on GitHub, so other communities can adapt the work rather than repeat it.

Why it travels

The technical achievement is real, but the design choice is what public sector technologists should take from it. AI was used for pattern matching and time saving, applied on top of local knowledge rather than in place of it. As the project team put it: no one knows a community better than the people living there, and AI did not replace them but expanded their capabilities.

That principle should feel familiar here. New Zealand's Public Service AI Framework puts human-centred values and social licence at its core, and the Department of Conservation is already trialling AI-powered predator detection tools built for local conditions. Kakuma is a reminder that the strongest AI projects start by asking who already holds the expertise, then handing them better tools.

Sources: “Stewards of their environment”, Microsoft Unlocked, 17 December 2025; AI for Good Lab, microsoft.com

Whats new in GovTech articles

Stewards of Their Own Map: What Kakuma Teaches Us About AI Done Well