When the Algorithm Knows Where the Need Is: AI in Social Work

Aid is often sent to where the latest news arrived, not to where the need is greatest. That is the complaint of some charitable and development work: it is not money alone that is lacking, but precise knowledge of who needs what, where and when. Here, precisely, AI is being tried. Let us see what the experiments have shown and what they have not.

Mapping poverty from the sky and from the phone

Where precise household surveys are absent, researchers have turned to indirect sources. In a study published in Science in 2016, a Stanford University team combined daytime satellite images, night-time light images and village survey data, and trained a model to recognise roads, towns and farmland in order to estimate the wealth of villages in several African countries. In Rwanda, Joshua Blumenstock and colleagues, in a study also published in Science in November 2015, linked call and text records held by a telecoms company to the answers of a thousand subscribers about their living conditions, and extracted patterns associated with wealth and poverty. Blumenstock described his work at the time as a "proof of concept", not a ready tool.

Who deserves aid? The Togo experiment

In Togo, when the government wished to pay out emergency cash transfers during the pandemic without a reliable registry of the poor, targeting based on phone data and machine learning beat the feasible alternatives in the rural scenario (an exclusion error of about 53% against 59% to 78%), yet nearly half of those eligible still stayed off the list, models age quickly, and anyone without a phone may not appear in the data at all. The experiment and its limits are set out in the article "AI and the Gateways to Good: How Can Giving Reach Its People?".

The authors see the method as a fast supplement in crises or where registries are lacking, not a replacement for traditional enquiry.

Warning before hunger

In 2023, Balashankar, Subramanian and Fraiberger published in Science Advances a model that predicts the degree of food insecurity up to 12 months ahead. It combines traditional indicators (conflict, prices, rainfall) with what an algorithm extracts from about 11.2 million English-language news articles between 1980 and 2020 tied to 37 countries, of which 21 were used because their data were complete. Adding the news cut the error by about 45% relative to the traditional model for three-month-ahead predictions. Among its limits are that it relied on English-language news, and that its accuracy is higher where media coverage is greater, which means those overlooked by the media may be less well served by the warning.

A patient teacher for those who have none

In a pilot by the World Bank in Edo State, Nigeria, over six weeks from June to July 2024, a generative model was used as an after-school tutor. The researchers estimated the effect at about 0.3 standard deviations, which they described as nearly two years of typical learning. These are preliminary findings published in a blog before the full report; attendance suffered from flooding and a teachers' strike, and the long-term effects and possible negative effects are unknown. Large policies should not yet be built on them.

Limits not to be forgotten

These experiments yield five lessons for any body working with those in need, whether a charity or an institution:

• Data is a trust: a poor person's data is not spoil; collect only what is needed and guard it.
• Those absent from the data are absent from the algorithm: channels are needed to reach people who have no phone or account.
• Models age: they must be updated and reviewed regularly.
• The decision is human: the tool shortlists and warns, but withholding and granting are not left to an algorithm whose logic is unknown. For more on accountability, see the article "Responsible Artificial Intelligence: Four Principles for Governance and Trust".
• Listen to the community: those who know the need at first hand see what the figures miss.

Perhaps the most telling thing about these tools is that they do not excuse us: once we see the need clearly, we are bound to act, and God knows the intentions.