Back in 2020, we ran a story about a humanitarian project in Togo organized by the non-profit GiveDirectly, UC Berkeley, and Togo’s Ministry of Digital Economy. Its aim: use AI to deliver humanitarian aid in the form of money to the poorest populations in the country. Now some of the results have been published in Nature, and we wanted to give our readers an update on how it turned out.

How did the project work?

The team developed an algorithm that was able to pinpoint poor areas of the country. As people in different economic situations use their phones differently, data from phone calls, text messages, and phone plans help reliably discern the poor from the rich. The machine learning algorithm was programmed to recognize relevant differences and create a specific map of people who needed the aid the most.

The results are in

Why not just use governmental population data to decide on who gets delivered the cash? Governmental surveys were not fully accurate and could exclude members of the population who needed aid. The recent paper published in Nature revealed that the AI program resulted in a more representative and reliable map, decreasing exclusion error by four-to-21 percent.

Since November 2020, when the project started, the program has delivered 10 million dollars to around 137,000 of Togo’s poorest citizens. Previous studies’ data indicated that these cash transfers can help increase food security and psychological wellbeing. The team is still gathering its own data to see if this initiative has had the same impact, though they are hopeful it has.

Why is this important?

Like in every nation around the world, economic activity in Togo was slowed due to the pandemic, resulting in 54 percent of the Togolese population missing meals each week. COVID-19 pushed millions of people globally into extreme poverty, and governments have struggled to pinpoint who needs help the most.

This work shows how data research and machine learning algorithms can help humanitarian crises and deliver aid to those who need it most, especially in cases where traditional data surveys may not be available.

Although this project has been a great success so far, not all households in Tongo have access to a mobile phone. This being the case, data from around 15 percent of citizens are still missing from the project, and the team is looking at ways to expand their algorithm to accurately include these groups.

Source study: Nature – Machine learning and phone data can improve targeting of humanitarian aid