Beginning in late 2020, with the COVID-19 pandemic in full swing, the mobile phones of tens of thousands of impoverished villagers in Togo pinged with good news. Their money was ready. With the help of artificial intelligence (AI), these villagers in the narrow strip of land in West Africa had been approved to receive around US$10 every two weeks, delivered directly to their mobile money accounts. Although that might not sound like much, it could keep people from going hungry.
Conventional poverty-relief projects rely on data gathered through in-person surveys — which were not possible during the pandemic. But Togo’s effort, dubbed Novissi, which means ‘solidarity’ in the local Éwé language, incorporated AI to determine who should receive aid. The project, led by Cina Lawson, the Togo Minister of Digital Economy and Transformation, along with scientists from the University of California, Berkeley, and the non-governmental organization (NGO) GiveDirectly, analysed satellite images and data from mobile-phone networks to estimate the wealth of particular regions and individuals1. “We needed a surgical approach,” says Lawson. It was an important moment for the use of AI in anti-poverty work, she says.
Some 700 million people around the globe live in extreme poverty, defined by the World Bank as living on less than $2.15 per day. Ending that poverty, one of the United Nations’ Sustainable Development Goals, requires an understanding of who is in need and what their needs are. But measuring poverty has long been a challenge, in large part because of the time and cost involved in trying to collect data from the poorest and most vulnerable populations.
Combine AI with citizen science to fight poverty
AI allowed Lawson to leapfrog the conventional hurdles of using old and incomplete data to quickly make the most of her limited budget. It’s an approach that is garnering both interest and controversy, says Joshua Blumenstock, a computer scientist at University of California, Berkeley, who collaborated on Novissi.
AI tools can not only be fast, says Ariel BenYishay, a development economist at the AidData Research Lab at William & Mary, a university in Williamsburg, Virginia, but they can also include a larger, more representative portion of the population than household surveys do, and identify patterns in data that even specialists could miss. AI might also help researchers to evaluate how well programmes meet their objectives and demonstrate how investments in areas such as health, agriculture, education and infrastructure pay off — or not. The World Bank recognizes this value and has been developing advanced AI tools to try to forecast food crises and violent conflicts, and to pull insights from large swathes of data gathered after an aid intervention. It concluded its Poverty, Prosperity, and Planet report2 in October 2024 by noting that anti-poverty “efforts should focus on leveraging machine learning and artificial intelligence models to close data gaps and enable more timely monitoring”.
But there are reasons to be cautious, says human geographer Ola Hall at Lund University in Sweden, who researches the intersection of AI and poverty. AI models have been criticized for being racist, sexist and otherwise biased. Just as household surveys often miss the poorest families because they do not have permanent housing, AI-driven programmes might not help individuals who do not have digital data trails, Hall says. They are nowhere near accurate enough to determine who qualifies for aid or cash subsidies and who doesn’t, he says.
However flawed AI might be, though, current systems of evaluating poverty are just as abysmal, says BenYishay. “The baseline isn’t perfect data. It’s actually very crappy data,” he says.
Measuring poverty
Table of Contents
British social reformer Charles Booth undertook an early effort to quantify poverty from 1886 to 1903 when he criss-crossed London’s cobblestones collecting data on people’s incomes and social class. He created a colour-coded map of the city and reported his findings in a treatise titled Life and Labour of the People in London. English sociologist Seebohm Rowntree and his team interviewed 11,560 families in York, UK, and published the findings in a 1901 book called Poverty: A Study of Town Life. The team calculated poverty on the basis of the ability to meet a person’s “physical efficiency”, or their minimal nutritional requirements. A sample minimal diet might include bread, porridge, boiled bacon, potatoes, skimmed milk and little else.
After US President Lyndon Johnson declared a ‘War on Poverty’ in 1964, the Office of Economic Opportunity adopted a poverty threshold devised by economist Mollie Orshansky that took a similar approach. It defined poverty as the bare minimum income required to cover food, shelter and other basic costs. Around the same time, India performed similar calculations for its populace. Although each expert tinkered with their formulae to account for local variations in rent and food costs, they all defined poverty on the basis of how much money a person lives on per day.
The dollar-per-day approach is blunt and easy to communicate, says Dean Jolliffe, an economist at the World Bank in Washington DC. Yet, how much money a person spends to get by is just one aspect of poverty. Economist and Anglican priest Sabina Alkire advocates for a more nuanced way to define poverty. “I want to know how many poor people lack a house, how many poor people have a kid out of school, so I can actually respond in very tangible, direct ways,” says Alkire, who is the director of the Oxford Poverty and Human Development Initiative at the University of Oxford, UK.
Mobile phone data can be used alongside satellite imagery and other data to estimate poverty across regions and countries.Credit: Kelvin Juma
In the early 2000s, Alkire wanted a way to capture poverty’s various affects on people. Just because someone has enough money to buy food doesn’t mean they have enough for medical care or school fees, says Alkire. In 2008, Alkire worked with James Foster, an economist at the George Washington University in Washington DC, to develop what’s called the Multidimensional Poverty Index (MPI)3. The approach estimates a unified measure of poverty by tallying up the deprivations and their intensity, with a total of ten indicators, including nutrition, school attendance, access to drinking water and what a household uses for cooking fuel.
For the field of poverty, this was a sea change. It allowed policymakers and others to measure, dissect and target the interacting variables that contribute to poverty at the household level. The United Nations Development Programme replaced its Human Poverty Index, which focused on survival, literacy and standard of living, with Alkire and Foster’s MPI in 2010, although certain United Nations agencies along with the World Bank continue to rely on the dollar-per-day definition.
‘It is chaos’: US funding freezes are endangering global health
Researchers and aid agencies have developed myriad ways other than the MPI to define poverty. These methods vary in the factors they include, depending on what they want to measure and the data at hand, says Jennifer Davis, who heads the Program on Water, Health & Development at Stanford University in California. In a 2024 paper, a team led by Davis and her graduate student Christine Pu evaluated four definitions of poverty that are used in the field, including daily per capita expenditure but not MPI, and found massive differences in how those definitions rank households in Ethiopia, Ghana and Uganda4. “When we did our analysis, not only did we not find much agreement at all for the full sample, we didn’t find it for urban households, or for the bottom 20%, or for the bottom 1% where we might expect the greatest need,” says Pu.
Along with the lack of agreement on definitions, there’s the problem of time. Even a well-oiled field team needs several hours to survey a single family, says Jolliffe. Although poverty researchers have refined their metrics and are incorporating the latest computational methods for analysing data, they often continue to rely on on-the-ground surveys to collect those data. A lot of people are surprised that we still do household surveys, says Jolliffe. But, “this notion that we have data on everything about everybody is very much a rich-world perspective”.
Turning to AI
As a PhD student in agricultural and resource economics, Marshall Burke was familiar with laborious data collection. To learn about farming and agriculture practices in East Africa, Burke travelled to Kenya and Uganda, where he spent months talking to farmers and walking their fields. But when Burke started the Environmental Change and Human Outcomes Lab at Stanford University in 2015, he wondered whether the computer revolution might offer better approaches.
These experiments could lift millions out of dire poverty
David Lobell, who had vast experience in remote sensing, occupied the office next to him. Around the same time, a specialist in AI and image recognition, Stefano Ermon, also joined Stanford University. The trio’s discussions turned to how the ever-increasing data from satellite images could be used to help identify people living in poverty around the world. Knowing that night-time lighting can be a rough proxy for wealth, the researchers used night-time satellite imagery of areas across Africa alongside daytime imagery to teach computer models to identify features associated with wealth.
Asking a computer to compare images of areas that are already known to be extremely rich or extremely poor is an electronic version of the game “spot the difference”, Burke says. The algorithms compare the distribution and condition of roads, the amount of green space, the size and spacing of buildings and a multitude of other variables. “All the sorts of things you and I would think to look for in an image are a little bit predictive,” says Burke. “A machine can sort through all that data,” and determine what aspects are most relevant.
In 2016, the team reported that the AI analyses of the satellite images correlated strongly with on-the-ground measurements of poverty5. As machine learning advanced, Lobell, Burke and Ermon refined their models by incorporating the latest techniques.
Using a pan-African data set of publicly available satellite images, the trio tested an updated approach in May 2020. When the team compared its machine-learning predictions with wealth-related survey data from 20,000 villages, the algorithm performed just as well as the laborious surveys, but at a fraction of the effort and cost6 (see ‘Poverty predictions’).
Source: Ref. 6
“This was a pretty seminal improvement concept for the development community,” says BenYishay. Other teams are joining the experiment, throwing out a lot of different ideas, he says. Scientists are applying machine learning to look for patterns in satellite images and mobile-phone data, and to analyse the impacts of drought, agricultural productivity, infrastructure investments and more, he says.