AI and the Gateways to Good: How Can Giving Reach Its People?

During the pandemic the government of Togo launched an emergency cash-transfer programme called Novissi. The problem it faced is an old one in every act of giving: no comprehensive social register showed who deserved help. So how does money reach those it is meant for?

The idea: what phones say about their owners

Researchers trained a machine-learning model on call records stripped of names from two telecom operators, together with a national household survey (2018–2019) and a phone survey (2020), to estimate which households were poorer from their pattern of phone use. They compared this method with geographic targeting, occupation-based targeting and others, in a study published in Nature in 2022 (Aiken and colleagues).

What the numbers said, for and against

The result was useful but far from perfect:

• In the programme's rural expansion, the wrongful exclusion of eligible people reached 53% with this method, against 59% to 78% for the alternatives feasible at the time.
• The method reduced exclusion errors by 4% to 21% relative to geographic targeting.
• But it increased them by 9% to 35% relative to methods that assume a comprehensive social register, a hypothetical comparison since Togo had no such register.

That is, more than half of the eligible people in the rural expansion remained outside the assistance. This does not discredit the idea; it is the best available when there is no register. But it forbids us to imagine that the algorithm reaches perfection, and shows why field verification and channels for objection and appeal must run alongside it.

Fairness and privacy

No systematic differences in exclusion appeared by gender, ethnicity, religion, age or household type, but small differences in accuracy between groups were found, and no method achieved perfect parity among them. The researchers noted that conceptions of fairness can themselves conflict, and that phone data with names removed can in some cases be re-identified. There is a double trust in this kind of giving, because the owners of this data are among the weakest and most in need of discretion. These concerns meet what the Kingdom's AI ethics principles (SDAIA, second version, 2023) arrive at: privacy and security, fairness, social and environmental benefits, and accountability. The general principles are treated in our article "Responsible Artificial Intelligence: Four Principles for Governance and Trust".

Other gateways to benefit

Good is not only cash transfers. In the Nature study on floods (2024), researchers reported that their model forecasts river flow up to seven days ahead and operates in more than 80 countries, free and without registration, and that it raised forecast quality in Africa to a level approaching what the European comparison system achieves in Europe. A warning that arrives days ahead may give families and the authorities time to prepare.

In the field of charity

It is not right to make the algorithm an authority in deciding who deserves charity under Sharia. The categories of those entitled to zakat are rulings of Sharia, governed by the competent bodies and scholars. The machine may help reach people who do not appear in records, or organise applications and track delivery, after which the responsible official verifies and decides. A charitable institution that uses such tools should ask three questions: what proportion of those eligible did we miss? How do we reach them? And what data of theirs must we protect?

So the "gateway to good" stays open with insight: AI shortens the distance between donor and recipient, while the decision, the trust and the prayer for acceptance remain with their owners, and God is the one whose help is sought.