- GiveDirectly utilizes AI/ML algorithms to quickly target and pay people in extreme poverty after a crisis. While effective, this complex system is challenging to explain to recipients in extreme poverty.
- To develop a clearer explanation, we worked with researchers and local field teams and conducted a recipient consultation process.
- The most understandable explanation includes scenario-based examples instead of the usual fact-based explanation.
When a crisis hits a low-income country, aid groups often struggle to quickly find and help those most in need because governments have incomplete or outdated information on their citizens (‘social registries’). This is why GiveDirectly has used mobile phone metadata and AI to predict people’s level of need and send them emergency cash quickly—often 4-8x faster than conventional aid.
Using large algorithms to identify those in need makes it harder for these communities to understand why some people get help and others don’t (‘targeting’), and explaining how these systems work is especially challenging with an audience unfamiliar with many tech or AI concepts. This blog explains our efforts to bridge this gap.
Our first explanation of how we used AI was hard for our recipients to understand
We’d been using ‘fact-based explanations’ to explain this mobile data and AI approach to communities, which you can read an excerpt of below:
📝 Click to read our initial fact-based explanation
The program wants to make sure that money goes to the people who are the poorest. There are two ways we do this, one for households with phones and one for households without phones. For households with phones, we look at their level of phone usage. Telecom companies like Airtel already record data anytime you use your mobile phone –– like the phone numbers that you call, the duration of the call, what time the call took place. The content of the calls that you place or the messages you send are not recorded.
It turns out that how someone uses their mobile phone can indicate how much money they live on per day. Recently, we worked with Airtel and TNM to look at how poorer and wealthier people in Mzuzu use their phones. There were some patterns, for example people who tend to place longer phone calls and place international calls tend to spend more money each day.
It is important to explain these systems so people can better understand and consent to how their data is used, so we interviewed 67 GiveDirectly recipients across Malawi and Kenya in February and October 2024 to get their thoughts. We learned our ‘fact-based explanation’ did not work well enough, as most recipients needed multiple, long explanations to understand the idea, even the more tech–savvy young adults in urban areas. We returned to the drawing board.
So we developed a clearer explanation in collaboration with recipients
In October 2024, we tested and developed a better way to explain how we use mobile phone data and AI. Working with Zoe Kahn, a PhD Candidate at UC Berkeley, and our local team in Malawi, we iterated over a three-day period:
- Day 1: A local staffer explained the process using visuals and fact-based explanations to recipients. Visuals helped, but the fact-based explanation alone wasn’t clear. These visuals were adapted from visuals developed by Kahn for previous research in Togo (Kahn et al. 2025).
- Day 2: We added local, scenario-based examples alongside visuals. This improved understanding significantly, and the scenarios alone worked better than just facts.
- Day 3: Since we cannot always use visuals when responding remotely to a crisis, we tested a mix of verbal-only scenario-based and fact-based explanations with more recipients. This approach proved concise and more understandable, resulting in the following:
👩🏿🤝👨🏾 Click to read our new real life scenario-based explanation
GiveDirectly is piloting a program to deliver emergency cash aid to households facing poverty due to rising food prices. We plan to give money to the poorest households to help them during the lean season, but due to limited funds, we will identify the poorest households using two ways.
First, we will visit households and ask questions about living conditions, such as whether they own or rent their home, and whether they own a refrigerator. For example, if John rents a 1-room home and lacks a refrigerator, while Mary owns a 3-room house and has two refrigerators. We’ll gather more such data to assess their circumstances.
Second, with permission, we will use mobile phone usage data from Airtel or TNM to see how the household uses mobile phones, such as their airtime purchases, the duration of their calls, and where they make calls from and to. For example, if John buys small amounts of airtime, and rarely uses his phone, while Mary uses her phone a lot and makes frequent long-distance calls to people outside the city, we might determine John is poorer and eligible for assistance.
Your information will remain confidential and used solely for eligibility purposes in compliance with Malawian laws. You can inquire and make requests about your data, including to delete your data, at any time by calling our call center. The phone data will only share how you have used your phone, not what you talked about. If you don’t have a phone or choose not to provide permission, we will only use the first method.
Do you understand and wish to sign up for the program, consenting to GiveDirectly accessing your phone usage data from the past six months?
Improving comprehension even further would come with tradeoffs
More people in poverty will understand this clearer explanation, but still not everyone will. There are trade-offs involved with making it more widely comprehensible, as we need to:
- provide transparency and comprehension without increasing the risk of fraud from explaining our systems in too much detail
- keep explanations short enough to fit into other program communications, especially during a crisis
Next, we’ll test how to best deliver this explanation through different channels including SMS text and IVR automated phone calls. For more on our approach to responsible AI/ML, read here →
Acknowledgements: This work would not have been possible without Zoe Kahn, Emily Aiken, Christine Lukhele, John Pendame, Faith Tauka, Swathi Ramprasad and Andrew Chang.
Appendix 1 – Our AI/ML explanations before and after testing
📝 Click to read our initial fact-based explanation
“The program wants to make sure that money goes to the people who are the poorest. There are two ways we do this, one for households with phones and one for households without phones. For households with phones, we look at their level of phone usage. Telecom companies like Airtel already record data anytime you use your mobile phone –– like the phone numbers that you call, the duration of the call, what time the call took place. The content of the calls that you place or the messages you send are not recorded.
It turns out that how someone uses their mobile phone can indicate how much money they live on per day. Recently, we worked with Airtel and TNM to look at how poorer and wealthier people in Mzuzu use their phones. There were some patterns, for example people who tend to place longer phone calls and place international calls tend to spend more money each day.”
👩🏿🤝👨🏾 Click to read our new real life scenario-based explanation
GiveDirectly is piloting a program to deliver emergency cash aid to households facing poverty due to rising food prices. We plan to give money to the poorest households to help them during the lean season, but due to limited funds, we will identify the poorest households using two ways.
First, we will visit households and ask questions about living conditions, such as whether they own or rent their home, and whether they own a refrigerator. For example, if John rents a 1-room home and lacks a refrigerator, while Mary owns a 3-room house and has two refrigerators. We’ll gather more such data to assess their circumstances.
Second, with permission, we will use mobile phone usage data from Airtel or TNM to see how the household uses mobile phones, such as their airtime purchases, the duration of their calls, and where they make calls from and to. For example, if John buys small amounts of airtime, and rarely uses his phone, while Mary uses her phone a lot and makes frequent long-distance calls to people outside the city, we might determine John is poorer and eligible for assistance.
Your information will remain confidential and used solely for eligibility purposes in compliance with Malawian laws. You can inquire and make requests about your data, including to delete your data, at any time by calling our call center. The phone data will only share how you have used your phone, not what you talked about. If you don’t have a phone or choose not to provide permission, we will only use the first method.
Do you understand and wish to sign up for the program, consenting to GiveDirectly accessing your phone usage data from the past six months?
Appendix 2 – Additional responses from recipient interviewing on AI/ML
During our February 2024 interviews with recipients, we also learned about their preferences related to the use of mobile phone data and AI for targeting. Click below to learn more:
📲 They wanted speed and convenience
The majority of recipients with mobile phones cited two advantages of using AI and mobile phones for targeting and enrollment: speed and convenience. A majority of phone users preferred more convenient and faster ways of accessing aid, and some shared that needing to participate in numerous in-person surveys with GiveDirectly staff reduced time for employment or other monetary activities.
🎯 They were concerned about exclusions
Many recipients expressed concern that this approach would leave behind people without phones, or the digitally illiterate. Some shared that it would be better for GiveDirectly to use both AI and in-person surveys to make sure people were not excluded from programs.
🏛️ They trust GiveDirectly, but don’t necessarily trust others
Almost all recipients we spoke with were not concerned with GiveDirectly accessing their mobile phone data, though this may be confounded by a desire to participate in GiveDirectly programs. Some did not have similar trust in other institutions, such as certain government organizations. Some recipients also shared that they preferred the use of mobile phone data instead of relying on village leaders to select recipients for aid (which is common in the aid sector, but not an approach that GiveDirectly uses) as they felt the latter was more likely to be unfair.
🛖 They worried about information in data leaks getting to their neighbors
Most recipients we spoke with were less concerned about the risk of their mobile phone data being leaked or shared with other organizations (like the government). They were more concerned about the risk of their data being shared with family and community members, having neighbors learn about their poverty level or other details.