GiveDirectly’s AI experiment left vulnerable flood victims behind in Nigeria

This story was originally published by The New Humanitarian.

By Damilola Ayeni

When GiveDirectly launched an AI-powered flood relief programme in 2024 for Ibaji, a key Nigerian farming district, its algorithm was meant to help deliver anticipatory aid to one of the world’s most chronically flood-battered regions. It was meant to be an improvement on aid responses that arrive too late, after homes and fields have been submerged.

After the programme was completed, the New York-based cash aid NGO announced that 4,600 people received anticipatory aid, doubling their incomes, lowering food insecurity by 90%, and leaving 93% feeling more prepared for future floods.

But an investigation by The New Humanitarian – based on interviews with at least 10 flood-affected farmers, mostly in Ibaji’s Unale ward – has revealed gaps in the technology-driven pilot project that prevented some of the area’s neediest people from accessing aid before floods destroyed their farms. Would-be aid recipients said they were excluded for submitting personal data that did not perfectly match government or bank records or because they could not afford the cost of keeping their phones charged to receive verification calls.

Some of these farmers said they took on debt, expecting aid that never came.

Meanwhile, the programme’s verification and targeting systems allowed some aid payments to go to relatively wealthy people and people living outside the programme’s flood-affected target area – sometimes to multiple members of the same family.

This is not the first GiveDirectly programme to encounter pitfalls while experimenting with technological aid solutions. In 2023, the organisation reported that its anticipatory response to Cyclone Freddy in Mozambique ended up targeting villages that were relatively unaffected by floods while bypassing villages that were hit hardest.

GiveDirectly voluntarily reported to The New Humanitarian in 2023 that its own staff in the Democratic Republic of the Congo had defrauded beneficiaries of $1.2 million by stealing their SIM cards and intercepting mobile money transfers. GiveDirectly eventually paid the victims what they had been promised, but six told The New Humanitarian that the payments were not enough to cover debts they or their neighbours defaulted on because of the fraud.

After receiving questions from The New Humanitarian about the Ibaji programme, GiveDirectly updated an article on its website to acknowledge that the programme’s design inherently excluded some of the area’s most vulnerable residents, including people with limited phone access and “poor people who may have moved to higher grounds before the programme began [and] may have incurred flood-related losses but did not benefit”.

Stella Luk, GiveDirectly’s vice president of programmes, told The New Humanitarian that the design involved intentional tradeoffs meant to reach more people more quickly.

“We worked to mitigate the downside of those design choices and study the impact, so we can improve further,” she said. “We think we made the right tradeoffs in our Nigeria pilot and have been incorporating the lessons there into our future work.”

How the system worked

Every year between July and October, catastrophic floods hit the lowlands of Ibaji, displacing tens of thousands of Nigerians from farming villages along the Niger and Benue rivers.

These fertile floodplains are a powerhouse of cassava, maize, yam, and rice production. But the absence of basic infrastructure like paved roads and electricity – as well as a lack of protection from seasonal floods – leaves local Igala-speaking farmers impoverished and unable to harness their full agricultural potential.

In 2024, GiveDirectly and the International Rescue Committee jointly received a $4.6 million Google.org grant to roll out AI-assisted early-warning systems aimed at delivering anticipatory cash to help farmers prepare for floods before they hit. (The IRC implemented its programme in another part of Nigeria.)

Using historical flood data from Google, wealth indicators, and insights from the Nigeria Hydrological Services Agency (NIHSA), GiveDirectly narrowed down Ibaji’s 10 wards to the six most at risk of flooding. Within those wards, nearly 39,000 applicants pre-enrolled by submitting personal and bank information via short codes on their mobile phones, which the organisation said saved at least $80,000 in staff costs. Almost half of these applicants were automatically disqualified for self-reporting that they did not live in any of the six eligible wards.

Nearly 1,000 additional applicants failed to pass ID and name verification conducted via Smile ID, a biometric identity verification platform that cross-checks against telecommunications and government databases.

The remaining 18,000 applicants then underwent further verification, which involved calls from a third-party call centre and in-person visits. Over 13,000 people were disqualified because they could not be confirmed as living in an eligible area, including a “small minority” who could not be reached after five outreach attempts, according to GiveDirectly spokesperson Yonah Lieberman.

Cash payments followed a “trigger” based on information from Google’s AI-driven Flood Hub and hydrology analysis by the risk management firm JBA Global Resilience, according to GiveDirectly’s website. If 20% of a given community’s area was forecasted to flood, an automated payout command was issued to verified residents there: $105 up front to help them flee before floods struck, plus $210 later to help them rebuild.

“Satellite/AI flood modelling determined which communities [got paid] and when,” Federico Barreras, GiveDirectly’s senior manager for emergency cash, told The New Humanitarian.

He added that the organisation is now studying the anticipatory action model in Kenya and Mozambique, as well as in Bangladesh, where “we are running a large randomised controlled trial covering over 100,000 families that is planned for the flood season in the Jamuna River Basin”.

Exaggerated impact

GiveDirectly originally claimed on its website that “4,600 households” received cash transfers. But field reporting in Ibaji revealed that in several households, multiple members received separate payouts.

In response to questions about this discrepancy, GiveDirectly retracted the initial claim.

“The honest answer is that we were imprecise with our language: we used ‘household’ and ‘individual’ somewhat interchangeably in our public materials,” Luk said. “That is a mistake that we’ve since corrected. The programme enrolled and verified 4,600 individuals, not households.”

Despite the perception that this might be unfair to other households with only one recipient or those who received nothing, Luk added that GiveDirectly was “comfortable” with multiple people in the same household getting paid separately because more funds “will only help them more”.

Mismatches and disconnections

When The New Humanitarian visited Unale, one of Ibaji’s eligible wards, in June 2026, local farmers waded through thigh-deep road puddles and struggled to push cargo tricycles through the muck.

Some of the area’s 200,000 residents bore painful memories of the 2024 season and GiveDirectly’s digital experiment.

Joshua Ojoyemenenyo, an Unale resident whose fields were wiped out by floods that year, said he had been disqualified because the name he submitted to GiveDirectly did not perfectly match official records.

Friday Fedegwu, a local community leader, said mismatches in names, phone numbers, and bank verification numbers resulted in applicants being “cut out of the process entirely”.

Data mismatches are a well-documented barrier to financial inclusion and are common in communities where relatives register SIM cards on each other’s behalf.

GiveDirectly’s Lieberman said 939 applicants were excluded for failing identity verification via the Smile ID platform.

Another common verification barrier was Ibaji’s lack of electricity and mobile connectivity. To use their phones, many residents rely on solar or generator-powered charging hubs, which come with exorbitant fees.

Applicants who could not afford to charge their phones missed verification calls, and some were disqualified. GiveDirectly said on its website that it was unable to reach 3,000 applicants by phone, adding that the inability to be reached after “five phone calls and two in-person visits” was a common reason for exclusion.

In the update to their website following questions from The New Humanitarian, GiveDirectly explained further that “people weren’t near their phones when we called (having left their phone to charge, for example), and in some poor-signal areas people had to climb trees or get onto rooftops to make a call”.

Luk said human reviews of database anomalies and in-person outreach attempts were meant to resolve these issues.

But for applicants like Ojoyemenenyo, the organisation’s field staff could not reconcile the inconsistencies. He said he was forced to borrow three bags of rice seeds at a 100% interest rate from a local farmer who had received aid from GiveDirectly. To pay it back, he surrendered six bags of rice from his harvest the following year.

“Excluded without an explanation”

The same system that screened out some of Ibaji’s poorest residents benefited those who were relatively well-off.

One Unale household secured four separate payouts and used the funds to build a new building, according to Fedegwu. In another household, two local politicians received separate payouts and – with capital and agricultural training they already had – reaped a bountiful harvest despite the floods.

Unale residents also told The New Humanitarian that people living in cities, outside GiveDirectly’s rural target area, received cash after convincing the call centre that they were flood-affected farmers. One Lagos-based recipient said he was in Idah, the urban cultural capital of the Igala people, when he received the verification call. Familiar with Ibaji, he answered the agent’s questions over the phone, passed the screening, and secured a cash transfer without an in-person visit.

GiveDirectly defended its programme’s methodology, insisting almost every beneficiary had been rightly targeted with aid.

“We believe the overwhelming majority of recipients were living in extreme poverty,” Luk said, citing recipient surveys. She added that GiveDirectly “screened out the overwhelming majority of enrollees who do not actually live in the targeted rural wards, ensuring that only people actually living in the poorest, most flood-prone areas would get cash”.

The New Humanitarian spoke with several farmers within the programme’s coverage area who said they were excluded without explanation.

Samuel Adejoh said he and several family members gave GiveDirectly’s call centre their National Identification Numbers and account numbers but received no payment or any reason for their disqualification.

“They sent all of us a message that we were not qualified – my mum, my wife, my father, all of us,” he said.

Floods destroyed his rice field, and he was forced to leave Unale permanently the following year, migrating hundreds of kilometres to Ogun State in search of a new means of survival.

Aladi Samuel said her family were disqualified for unspecified reasons, leaving them unable to buy a canoe they needed to salvage their produce when floods destroyed their rice field. When a GiveDirectly team visited Unale later to evaluate the aid programme, she told them network failures prevented her from reaching the organisation’s call centre. She said the team gave her a vague promise of a future resolution.

Ladi Ashoba, the headmistress of Unale’s only public primary school, said some disqualified residents took out loans after GiveDirectly evaluators assured them the system would work “next time”. Although GiveDirectly sent out text messages in December 2024 announcing that the aid programme was “closed”, residents said they believed the verbal promise, hoping to pay off their debt with the next year’s payout.

“Being excluded without an explanation costs trust,” GiveDirectly said on its website. Leaving thousands of pre-enrolled people empty-handed without a clear explanation is “an area we’re actively working to do better”, they said.

Luk told The New Humanitarian similar misunderstandings were “also a risk in programmes that use human field infrastructure”. She added that the organisation conducted “re-sensitisation visits ward by ward” to clarify that the aid programme was a one-time event.

On their website, GiveDirectly said their geographic trigger – designed to pay by area rather than individual need – was “a deliberate tradeoff: speed and coverage over precision targeting”.

But the benefits of that tradeoff were inconsistent. While most beneficiaries received payments before the floods hit, nearly a third only received them in late October 2024, after flooding had peaked, because of identity verification issues, according to the website.

Joel, another Unale resident, recalled seeing local widows unable to access emergency funds as their homes collapsed amid rising waters.

“That is our pain,” he said.

This article was produced in collaboration with Egab. Edited by Jacob Goldberg and Lilian Wagdy.

–––––

The New Humanitarian puts quality, independent journalism at the service of the millions of people affected by humanitarian crises around the world. Find out more at www.thenewhumanitarian.org.