One of the clearest peer-reviewed cases of machine learning delivering a social benefit is Togo's 2020 emergency cash program, Novissi. A study in Nature (March 16, 2022) found that models trained on mobile phone data cut wrongful exclusion of poor households by 4 to 21 percent compared with the geographic targeting the government had available. The same study found the method did worse than a full social registry would have, so the benefit depends on what data a country lacks.
Key takeaways
- Novissi began in April 2020, when Togo needed to reach poor households quickly during COVID-19 lockdowns.
- Compared with geographic targeting, the machine learning approach reduced exclusion errors by 4 to 21 percent.
- Compared with a hypothetical comprehensive social registry, it increased exclusion errors by 9 to 35 percent.
- The authors describe it as a fast option when current data are missing, not a replacement for traditional targeting.
- A benefit claim is only as strong as its comparison point.
The problem Novissi had to solve
Cash transfers only help if they reach the right people. In April 2020, Togo launched Novissi, a digital cash program, and needed to find poor recipients without a current registry: according to the Abdul Latif Jameel Poverty Action Lab (J-PAL), the country's last census was from 2011 and held nothing useful for targeting, and building a new registry during the pandemic was impractical. Registration worked on basic phones, and recipients received mobile money payments of about USD 20 per month for three months.
The first phase aimed at informal workers in the regions hit hardest by lockdowns, identified through a voter database with self-declared location and occupation. The rural expansion, which is the part the Nature paper evaluates, needed another way to find the poorest people.
What the models did
Two layers of machine learning were used. According to J-PAL, models applied to high-resolution satellite imagery estimated relative wealth in grid cells of 2.4 square kilometers, and the estimates were averaged by population to rank the 100 poorest cantons. Within those cantons, researchers surveyed mobile subscribers about wealth or consumption, matched the answers to call detail records, and trained a model to predict each subscriber's consumption so the poorest could be prioritized.
The paper, Machine learning and phone data can improve targeting of humanitarian aid, was written by Emily Aiken, Suzanne Bellue, Dean Karlan, Christopher Udry, and Joshua Blumenstock and published in Nature 603 in March 2022.
Reading the numbers against the right baseline
| Comparison | Result reported | What it means |
|---|---|---|
| Versus geographic targeting (the option Togo had) | Exclusion errors lower by 4 to 21 percent | Fewer poor people were left out than under area-wide blanketing |
| Versus a hypothetical comprehensive social registry | Exclusion errors higher by 9 to 35 percent | A good registry would have done better, but Togo had none |
| Fairness checks | No unfair exclusion of women or of any ethnic, religious, or age group found, per J-PAL | Group-level patterns were examined, not only overall accuracy |
The pairing is the instructive part. A headline of "AI improves aid targeting" is true against one baseline and false against another. The honest reading is that the method helped when the alternative was crude and fast action mattered.
Limits the authors state
J-PAL's summary quotes the researchers on the boundary of the claim: the approach is a rapid, cost-effective option when up-to-date data are missing, but it should not replace traditional targeting methods when the data and timeline permit. Several questions fall outside what the study could settle, including whether the same models would hold up in other countries, how phone-data access should be governed, and what happens to people who do not own phones. This article does not cover those questions beyond flagging them.
A checklist for any claimed social benefit of AI
- What was the comparison: no tool, a simpler tool, or the ideal tool?
- Was the result published with methods that outsiders can check?
- Were outcomes measured for different groups, not only on average?
- Who holds the data, and under what rules?
- Does the claim describe a pilot in one place or a general result?
- Do the authors themselves state limits?
The questions do not prove or disprove a benefit. They show where a headline is leaning on an unstated assumption.
FAQ
Did AI decide who received money in Togo?
Models ranked cantons and individuals by predicted poverty, and that ranking guided who was prioritized for the rural expansion of the program. The program design, payment amounts, and registration process were set by the government and its partners.
Is 4 to 21 percent a large improvement?
It is a reduction in the share of poor people wrongly excluded, measured against geographic targeting. Whether it is large depends on how many people are affected and what the alternative costs, which the percentage alone does not show.
Does the study show AI is better than a social registry?
No. Against a hypothetical comprehensive registry, the machine learning approach increased exclusion errors by 9 to 35 percent. Its value came from filling a data gap.
Can the result be applied in Canada or the United States?
Not directly. Both countries have extensive administrative data, so the data gap that made the Togo approach useful is much smaller, and the study makes no claim about wealthy countries.
Related reading
- AI and modern society
- Ethical questions around AI
- How AI programs analyze large datasets
- AI, sustainability, and society
More in AI Basics.
Last verified: October 2026
Sources
- Aiken E, Bellue S, Karlan D, Udry C, Blumenstock JE, Machine learning and phone data can improve targeting of humanitarian aid, Nature 603, 2022
- Using Mobile Phone and Satellite Data to Target Togo's Emergency Cash Transfer Program, Abdul Latif Jameel Poverty Action Lab (J-PAL), 2026