When the novel coronavirus reached Togo in March, its leaders, like those of many countries, responded with stay-at-home orders to suppress contagion and an economic assistance program to replace lost income. But the way Togo targeted and delivered that aid was in some ways more tech-centric than many larger and richer countries. No one got a paper check in the mail.
Instead, Togo’s government quickly assembled a system to support its poorest people with mobile cash payments—a technology more established in Africa than in the rich nations supposedly at the forefront of mobile technology. The most recent payments, funded by nonprofit GiveDirectly, were targeted with help from machine learning algorithms, which seek signs of poverty in satellite photos and cellphone data.
Togo’s project is an example of the pandemic forcing urgent experimentation that may lead to lasting change. The turn to satellite and cell phone data was driven, in part, by a shortage of reliable data on citizens and their needs. Shegun Bakari, an adviser to Togo’s president, says it worked so well that the data-centric approach will likely be used more widely. “This project is foundational for us in terms of how we can set up our social protection system in Togo in the future,” he says.
The new aid system is called Novissi, meaning “solidarity” in the local Ewe language, and took shape during 10 intense days of work starting in late March. Cina Lawson, Togo’s minister of digital economy, was motivated by fear of the side effects of pandemic shutdowns. Half of Togo’s 8 million people live on less than $1.90 a day. Most Togolese work in the so-called informal sector, for example as manual laborers or as seamstresses, and COVID-19 restrictions abruptly cut off their income. “We were thinking we’ve got to support these people because if they don’t die of COVID, they’ll die of starvation,” Lawson says.
Novissi launched on April 8 and sent aid that same day to informal workers in and around Togo’s capital, Lomé. Radio ads asked people to text message a special number that walked them through a brief questionnaire over SMS. Payments were sent more or less instantly, if a check against Togo’s voter ID database, which covers 93 percent of the population, confirmed a person had previously declared an informal occupation and lived in an eligible area. The program was quickly expanded to the area around Togo’s second largest city, Sokodé.
Men received CFA$10,500 (francs) each month, roughly $20, in biweekly installments, and women CFA$12,250 (francs), roughly $23; the difference was by design to better support families. The amounts were aimed at replacing roughly one-third of Togo’s minimum wage. So far the government has sent roughly $22 million through Novissi to nearly 600,000 people.
Lawson was proud to see government aid sent so fast, but as COVID-19 spread she also worried her program wasn’t able to target the people most in need of help, in part because she didn’t know where to find them. Government officials contacted Joshua Blumenstock, codirector of University of UC Berkeley’s Center for Effective Global Action, who’d been researching how big data can fill information gaps facing countries like Togo. His lab had shown that phone records could predict individual wealth in Rwanda about as well as in-person surveys and that satellite images could track areas of poverty in sub-Saharan Africa.
Blumenstock offered to adapt his technology to help and enlisted a team that came to include Berkeley grad students, two faculty members from Northwestern, and the nonprofit Innovations for Poverty Action. He also connected Lawson with GiveDirectly, which distributes cash payments in poor countries. GiveDirectly had talked with Blumenstock before about using his work to prioritize aid and now saw a chance to put the idea into action.
GiveDirectly’s payments usually reflect information gathered by staffers who visit poor communities and perform household surveys. But that posed risks during a pandemic. Han Sheng Chia, the organization’s special projects director, was curious whether satellite and similar data could help the group distribute aid faster and more widely. “The scale of need we’re facing this year is so huge,” he says. The World Bank estimated in October that the number of people in extreme poverty will rise by about 100 million this year, the first global increase in 20 years.
Blumenstock and his team trained image analysis algorithms to create a fine-grained map of Togo from satellite images, calibrated using a 2018 household survey that had reached only part of the country. The algorithms picked up indicators of wealth and poverty such as different roofing materials and road surfaces. The researchers built a second system that estimates the wealth of users of Togo’s two primary cell networks, using calling patterns and other account details, like credit top-ups. That part of the system was based on a phone survey in September of about 10,000 people in the poorest regions flagged by the satellite analysis. GiveDirectly also sent a small team to Togo to gather additional information on communities in need.
A new, more automated system launched in November, using GiveDirectly’s money. In the areas identified as least wealthy, people the algorithms flagged as likely to live on less than $1.25 a day received text messages inviting them to apply for help, a process that takes less than 3 minutes. Men receive five monthly payments of roughly $13 each, and women roughly $15 each. Applicants are verified against Togo’s voter ID database and GiveDirectly’s requirements.
Within two weeks, Chia says, the program had paid 30,000 of Togo’s poorest people, many in rural areas. “To cover that geographical span would have taken huge field teams upwards of 200 people months,” he says, adding that the approach may be applicable elsewhere.
Blumenstock says this is the first time he has seen proxies for poverty used to directly route cash, not just to inform aid decisions. “This entire aid mechanism is contactless,” he says—although his team is using phone surveys to retrospectively audit the program and plans an in-person survey in Togo next year. GiveDirectly has so far distributed nearly $800,000 out of a planned $10 million budget intended to reach about 115,000 people.
Togo’s project is not the first experiment in using algorithms to direct aid to some of the world’s poorest. Population density maps created by Facebook machine learning experts helped guide a targeted cholera vaccination campaign in Mozambique last year after a cyclone caused widespread damage and flooding. Also last year, the Rockefeller Foundation helped launch a startup called Atlas AI to commercialize Stanford University research on measuring poverty and crop yields using satellite imagery and machine learning.
Zia Khan, senior vice president of innovation at the foundation, says that technology should help programs like its work on agricultural development, or deciding where to support construction of rural solar “mini-grids” to improve access to electricity. Measuring electrical infrastructure from space photos can be less time-consuming and can sidestep terrestrial sensitivities that prevent a clear picture of a community’s needs. “Sometimes there are political issues around how accurately government ministries want to depict the poverty in rural areas,” Khan says.
Tapping satellites and algorithms doesn’t guarantee accuracy or empirical truth, though. To be reliable, machine learning models must be trained on data representative of the situation where they will be used. “If you put biased data in you’ll get biased decisions out,” Khan says.
Rockefeller backs a project called the Lacuna Fund launched earlier this year to help create data sets to support use of machine learning in low-income countries. It is initially focusing on sub-Saharan Africa, including ways to better identify crops and pests found in that region that are unfamiliar to most people in Western AI labs.
How machine learning can help—or fail—humanitarian projects will become more apparent as governments and donors use it more. Togo may be among the leading experimenters. Bakari, the adviser to the country’s president, says Novissi has inspired interest in using the technology for other assistance programs and to help government finances. “If you can use big data to target the poorest, you can use the same technology to know who you should be asking to pay more tax that will support the poorest parts of the country,” he says.
This story originally appeared on wired.com.