One of AI’s clearest public-interest applications did not generate an answer for a user. It helped decide who would receive emergency money. In a March 2022 account, researchers Emily Aiken and Joshua Blumenstock reported that their partnership with Togo’s government and GiveDirectly had delivered nearly $10 million to roughly 137,000 people. Those figures describe the programme at that time, not a new 2026 disbursement.[2]
The operation paired machine-learning-based selection with mobile enrolment and payments. Its first beneficiaries were enrolled and paid in November 2020. The authors describe collecting around 15,000 phone surveys to relate living conditions to patterns in mobile-network data, giving the targeting model a basis in household evidence rather than an unsupported guess.[2]
The useful comparison was what government could actually do
The Nature paper, published on 16 March 2022, examined the rural expansion of Novissi. Satellite-based estimates helped identify poorer areas, and phone metadata helped prioritise subscribers within them. In the programme’s evaluated rural setting, funding could reach only 29% of people in eligible areas, making selection consequential.[1]
Against a survey-based proxy for poverty, the phone-based method had a 53% exclusion-error rate, compared with 59–78% for the feasible alternatives examined. Lower is better: exclusion means a person classified as poor by the reference measure is left out. The reference was a proxy means test, not a complete direct measurement of everyone’s consumption.[1]
The same paper also found that approaches requiring a comprehensive social registry could outperform phone-based selection in a hypothetical comparison. Togo did not have such a registry available for the emergency. The result supports using AI to complement missing administrative information, not declaring it superior to every alternative.[1]
That is an important form of benefit. A government need not have a perfect measurement system for an improved, usable one to matter. But the cash delivered and the targeting comparison remain different pieces of evidence: neither establishes that the algorithm itself caused a measured reduction in poverty, hunger or distress.
What the newer evidence adds
A 2026 preprint by M. Merritt Smith, Aiken, Blumenstock and Sveta Milusheva compares phone-data prediction across Afghanistan, Côte d’Ivoire, Malawi and Togo. It finds that longer-term measures such as wealth are generally more predictable than transient vulnerabilities such as food insecurity or mental health. The study evaluates prediction, not a new cash-transfer impact trial.[3]
Its findings suggest a practical limit. A household can suffer a sudden loss before its historical phone-use pattern changes enough to reveal it. That is a reason to preserve ways of reporting new need, rather than assuming a statistical portrait is a complete account of someone’s present circumstances.
The four-country paper also explicitly limits its evidence to mobile subscribers. It does not establish performance for the whole population, including people without phones. The original programme’s authors likewise identified phone access as an unresolved route to exclusion.[3][2]
Togo’s experience is compelling because the delivery was real and the research did not hide the alternatives. AI helped turn imperfect existing data into an actionable route for emergency support. The lasting lesson is to judge the entire service: who is reached, who remains invisible, and whether a mistaken decision can be corrected.
Sources & further reading
- Machine learning and phone data can improve targeting of humanitarian aid Nature · 16 Mar 2022 · Accessed 27 Oct 2026
- How AI helped deliver cash aid to many of the poorest people in Togo Emily Aiken and Joshua Blumenstock, via Fast Company / The Conversation · 19 Mar 2022 · Accessed 27 Oct 2026
- Predicting Well-Being with Mobile Phone Data: Evidence from Four Countries arXiv · 2 Feb 2026 · Accessed 27 Oct 2026