Approximately 700 million people worldwide live in extreme poverty (living on less than US$2.15 per day). Eradicating poverty is one of the United Nations Sustainable Development Goals, but measuring poverty has been a challenge, mainly because data collection is expensive and time-consuming. Artificial intelligence (AI) can not only analyze data quickly, but also reach a wider range of people and identify patterns that experts may miss. The World Bank is also developing AI tools to predict food crises and violent conflicts and extract insights from aid intervention data.

However, AI models suffer from bias issues and may miss the poorest people for whom digital data are not recorded. Nonetheless, the current poverty assessment system is equally imperfect. Traditional methods such as household surveys are time-consuming and costly, but AI combined with satellite imagery and mobile phone data can more efficiently identify poor areas and individuals. For example, a research team from Stanford University in the United States used AI to analyze satellite images and successfully predicted poverty levels in African villages. The effect was equivalent to field surveys, but the cost was significantly reduced.

For example, the "NOVISSI" social security program in the West African country Togo used AI to analyze mobile phone usage patterns and satellite images to accurately allocate US$34 million in aid funds. Similar projects are underway in other African countries. Although AI predictions are not completely accurate, its ability to respond quickly is crucial in emergency situations.

However, although AI has shown potential in poverty alleviation, experts caution that it should be used with caution. AI cannot completely replace field surveys, especially in multidimensional poverty measurement. However, amid budget constraints and economic shocks, AI could be a critical tool in getting aid into the hands of those who need it most. In the future, the combination of AI and field surveys may become a new direction for poverty alleviation work.


Priority is given to the poorest villages and communities. We applied deep learning algorithms to high-resolution satellite imagery to produce microscopic estimates of wealth per 2.4 kilometer (km) grid cell (top left), combined these estimates with population density information for each grid cell (top middle), and used this information to identify the 100 poorest counties in Togo (top right).
Using ground-truth wealth and poverty data collected from a large telephone survey of active mobile phone users, the researchers trained a machine learning algorithm to estimate the wealth of each mobile user (above left). In the 100 poorest counties (red distribution on the right), those with estimated consumption of less than $1.25 per day are prioritized for inclusion in the Novissi program (dashed vertical line). These people are much poorer than the average resident of Togo (blue distribution). Source: Josh Blumenstock, University of California, Berkeley, January 11, 2021: Josh Blumenstock, University of California, Berkeley, January 11, 2021.