Technology for Development

Source details

AI holds promise to target aid and credit better and improve access to tailored teaching and medical advice in developing economies.
Successful AI applications in low-income settings require local innovation and adaptation; plug-and-play deployment of systems developed for wealthy nations is unlikely to deliver intended benefits.
Major implementation challenges include behavioral responses to algorithms, communication barriers, data gaps, and infrastructure and institutional constraints.

The government repurposed technology originally designed to target online advertising to identify the country’s poorest residents.
Data sources used: satellites and mobile phone company data.
Poverty “signatures” included villages that appeared underdeveloped in aerial imagery and mobile subscribers with low balances on their phones.
Outcomes and implementation practices:
- Targeting based on these signatures helped ensure that cash transfers reached people with the greatest need (Aiken and others 2022).
- Success depended on government collaboration with researchers and nonprofit organizations, customizing the technology to local needs.
- Operational adaptations included building a mobile money distribution system that worked for all mobile subscribers, adapting existing machine learning software to target cash transfers, and interviewing tens of thousands of beneficiaries to align the system with local definitions of poverty.
- The AI-based solution was designed to be phased out after the pandemic ended.
Risks identified:
- Algorithms that perform well in laboratory settings may be unreliable in consequential real-world deployments.
- Behavioral responses (people adapting behavior to qualify for benefits) can undermine targeting effectiveness.

Context: Over a quarter of adults in Kenya have taken out loans using their mobile phones.
Machine learning used to determine microloan eligibility based on mobile phone behavior (Björkegren and Grissen 2020).
Risk: Observable features that predict creditworthiness (for example, having more Facebook friends) can induce strategic behavior (applicants adding friends), degrading targeting accuracy.
Evidence from field research:
- In a study with the Busara Center in Kenya, people learned and adjusted smartphone behavior in response to algorithmic rules (Björkegren, Blumenstock, and Knight, forthcoming).
- A proof-of-concept algorithm adjustment that anticipates such responses performed better.
Implication: Algorithmic design must anticipate manipulative or adaptive behaviors; technology alone cannot solve real-world implementation problems.

Local partner piloted an AI chatbot for teachers called TheTeacher.AI, similar to ChatGPT but:
- Tailored to local curriculum and instruction.
- Accessible even when internet connections are poor.
Implementation observations:
- Many teachers initially could not phrase questions to get useful answers.
- A subset of teachers adopted the system regularly to assist with teaching concepts, planning lessons, and creating classroom materials (Choi and others 2023).
- Adoption required training, experimentation, and sharing of effective use-cases.
Lesson: Some AI uses are not immediately obvious; discovery depends on trial and error and peer learning.

Explaining algorithmic concepts can be difficult where literacy, numeracy, and familiarity with digital data are low:
- Field experiment in Nairobi found challenges explaining simple algorithms with negative numbers and fractions to low-income residents; simpler communication approaches were developed.
Not all applications require users to understand internal algorithmic workings (black box models may be acceptable in crises).
Transparency is critical in nonemergency social protection settings:
- Explaining eligibility criteria to potential beneficiaries is essential.
- Norms and values around data and privacy differ across contexts; in rural Togo few interviewees worried about government or companies accessing their data, but many were concerned about how information would be shared with neighbors.
Broader societal impacts:
- AI can generate false images and voice-mimicking robocalls, reducing trust in online information.
- Even remote populations need awareness of these harms so they are not misled and so their concerns inform regulation.

AI solutions depend on physical digital infrastructure: servers, fiber-optic cables, cell towers, and mobile phones.
Over the past two decades developing economies have invested heavily in cellular and internet connections, creating a foundation for AI applications.
Persistent data and representation gaps:
- AI models in developing economies may lack information about the needs and desires of lower-income residents, the state of their health, visual appearance of people and villages, and the structure of lesser-used languages.
Required actions to improve knowledge infrastructure:
- Integrate clinics, schools, and businesses into digital recordkeeping systems.
- Create incentives for the use of such systems.
- Establish legal rights over resulting data.
Local tailoring of models:
- Western AI systems may propose expensive resources (digital whiteboards, slide presentations) that are not feasible; systems must be adjusted to local resource constraints.
- Invest in capacity and training of local AI developers and designers so technical innovation better reflects local values and priorities.

Adaptation and localization:
- AI must be customized to local social and institutional contexts to deliver benefits to the poor.
Anticipate behavioral responses:
- Design algorithms that account for strategic behavioral adjustments that can undermine targeting.
Emphasize transparency where appropriate:
- Explain eligibility criteria in nonemergency social protection programs to align with beneficiaries’ expectations and norms.
Build knowledge infrastructure and legal frameworks:
- Integrate public service institutions into digital recordkeeping, incentivize their use, and establish legal rights over data collected.
Invest in local capacity:
- Train and fund local AI developers and designers to produce solutions attuned to local resources, languages, and values.
Promote public awareness:
- Inform populations, including remote communities, about AI-driven misinformation risks (deepfakes, voice-mimicking robocalls) so trust and regulatory responses can be calibrated appropriately.

Source: Technology for Development, F&D Magazine, Daniel Björkegren and Joshua Blumenstock.

F&D: Technology for Development
- F&D: Technology for Development (Markdown version)
- F&D: Technology for Development (PDF)