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

Often, the most captivating AI initiatives aren’t necessarily those that involve the largest models. Instead, they’re the ones based on insightful questions about who possesses the knowledge needed. A notable example is a humanitarian mapping project in northwestern Kenya, which presents a valuable lesson for us all.

Kakuma refugee camp, Africa’s largest, was established in 1992 to provide refuge to young people fleeing the Sudanese war. Today, it has evolved into a community of over 300,000 individuals from more than 20 nations, spread across an area of around 15 square miles. For years, the camp’s maps were outdated, which significantly hampered aid delivery, infrastructure development, and emergency response efforts. The camp’s irregular structure and variety of shelter types posed challenges for traditional mapping techniques.

Three organisations collaborated to tackle this challenge. UNHCR’s innovation hub, the Hive, identified the core problem. The Humanitarian OpenStreetMap Team managed the on-the-ground data collection and integration with the community. Meanwhile, Microsoft’s AI for Good Lab focused on developing the necessary models. Dr. Simone Fobi Nsutezo, an applied research scientist at the Lab, shared, “Collaboration was essential because each contributor brought something distinct to the project.”

Empowering the Residents

Here’s the vital part: the locals led the data collection. From the project’s introduction to operating drones, refugees within the camp identified various features, trained as mappers and interpreters, and established the ground truth. The Humanitarian OpenStreetMap Team carefully tagged 10 square miles of imagery to create a comprehensive training dataset. It was only after this that the Lab utilised Azure to create models capable of recognising buildings, sanitation blocks, solar panels, and power infrastructure within the camp’s complex layout.

“After training a model with a small dataset, it becomes quick to predict new areas,” explains Dr. Amrita Gupta, also an applied research scientist in the team. “We’ve made open-source code available for mapping solar panels, buildings, roof types, and sanitation facilities, which anyone can use independently.” All models and datasets were made accessible on GitHub, allowing other communities to adapt this work rather than start from scratch.

Key Takeaways

The technical achievement is impressive, but it’s the design choices that public sector technologists should prioritise. AI is being utilised for pattern recognition and efficiency, enhancing local knowledge rather than replacing it. As the project team aptly put it, nobody understands a community better than those who live there, and AI has simply supplemented their existing capabilities.

This principle resonates widely. New Zealand’s Public Service AI Framework emphasises human-centred values and social consent. The Department of Conservation, for instance, is already testing AI-equipped tools for predator detection tailored to local conditions. The Kakuma project reminds us that the most effective AI initiatives begin by identifying local expertise and providing those individuals with enhanced tools.

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