AI Can Now Spot the Poorest Villages Using Phone and Satellite Data

Governments and charities face a simple but brutal problem: figuring out who is poor. Household surveys and national censuses are expensive and happen only every few years, so the data is often stale by the time it's used. In Sri Lanka, researchers tested a cheaper shortcut — feeding mobile phone call records and satellite imagery into a machine learning model to predict poverty across 13,985 local divisions.

But here's the twist that makes this paper interesting. The researchers argue that average prediction accuracy is the wrong thing to measure. What actually matters is whether a map helps you find the poorest places when your budget only stretches to helping a few. So they graded the models on a practical question: of the 25 poorest districts, how many did the map correctly flag? The combined phone-plus-satellite model found 86% of them. Satellite images alone managed 69%, and phone data alone 67%.

One detail stood out. Isolated communities — villages far from others — were about 11% harder for the model to read correctly. That's a warning sign, because remote areas are often exactly where the poorest people live. Getting them wrong means aid misses the people it was meant for.

The team also cautions that agreeing with a wealth index built from census assets is not the same as actually measuring how much people consume or earn. A map can look accurate and still misjudge real hardship. Still, the broader message is a practical one: AI poverty maps should be judged by whether they help deliver aid to the right doorsteps, not by how clever the algorithm sounds.