## Artificial Intelligence (content unit: 42-45-bjoerkegren-final)

## Source details

**Canonical URL:** [Artificial Intelligence (content unit: 42-45-bjoerkegren-final)](https://www.imf.org/-/media/files/publications/fandd/article/2023/december/42-45-bjoerkegren-final.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/fandd/article/2023/december/42-45-bjoerkegren-final.pdf.md)
- [Structured JSON version](/-/media/files/publications/fandd/article/2023/december/42-45-bjoerkegren-final.pdf.json)

---

### Local adaptation is essential
- Most recent advances in artificial intelligence originated in wealthy nations—developed in those countries for local users, using local data.
- AI-based solutions will work only if they fit the local social and institutional context.
- In Togo, the government repurposed technology originally designed to target online advertising to the task of identifying the country’s poorest residents, processing data from satellites and mobile phone companies to identify signatures of poverty (for example, villages that appeared underdeveloped in aerial imagery and mobile subscribers with low balances on their phones).
- The Togo program was customized through collaboration among the government, researchers, and nonprofit organizations: building a system for distributing mobile money payments that worked for all mobile subscribers, adapting existing machine learning software to target cash transfers, and interviewing tens of thousands of beneficiaries to ensure the system reflected the local definition of poverty.
- The AI-based solution in Togo was not designed to be permanent; it was to be phased out after the pandemic ended.

### Examples and empirical findings
- Targeting aid in Togo: Using AI to identify poverty signatures helped ensure that cash transfers reached people with the greatest need.
- Microloans in Kenya: Machine learning is used to determine eligibility for microloans based on mobile phone behavior; in Kenya over a quarter of adults have taken out loans using their mobile phones.
- Behavioral responses can undermine targeting: If those with more Facebook friends are likelier to be approved for a loan, some applicants may consider adding friends quickly, making it hard for systems to target the intended people.
- In a study with the Busara Center in Kenya, people were able to learn and adjust their smartphone behavior in response to algorithmic rules; a proof-of-concept adjustment that anticipates these responses performed better.
- Education in Sierra Leone: TheTeacher.AI, an AI chatbot tailored to local curriculum and instruction and accessible with poor internet, required training and experimentation; many teachers initially couldn’t phrase questions to yield useful answers, but a small group began to use the system regularly for teaching concepts, planning lessons, and creating classroom materials.

### Implementation challenges and risks
- Laboratory performance may not translate to reliable consequential decisions on the ground; people might adapt their behavior to qualify for benefits, undermining targeting algorithms.
- Complex AI systems are difficult to understand, even for AI researchers; explaining eligibility criteria is essential in nonemergency social protection settings.
- Norms and values around data and privacy differ across contexts: in rural Togo few people worried about the government or companies accessing their data, but many wondered if and how such information would be shared with their neighbors.
- AI can produce misinformation (for example, provocative false photographs and robocalls that mimic voices), affecting trust in online information—even in remote populations.

### Communication, usability, and capacity
- Grasping AI’s potential is harder in lower-income countries where literacy and numeracy are lower and residents are less familiar with digital data and algorithms.
- Simple numerical explanations (for example, algorithms with negative numbers and fractions) may be difficult to communicate; field experiments showed people could understand concepts when presented in simpler ways.
- Some applications do not require users to understand how algorithms work (for example, Netflix recommendations); in humanitarian crises, policymakers may accept “black box” algorithms, as Togo did during COVID-19.
- Uses of AI may not be immediately obvious; discovering useful applications depends on trial and error and sharing successful use cases.

### Infrastructure and data gaps
- AI solutions rest on physical digital infrastructure: massive databases on servers, fiber-optic cables and cell towers, and mobile phones in people’s hands.
- Over the past two decades, developing economies have invested heavily in connecting remote areas with cellular and internet connections, laying groundwork for AI applications.
- Some AI systems will require additional investment in knowledge infrastructure where data gaps persist and the poor are digitally underrepresented.
- AI models often have incomplete information about the needs and desires of lower-income residents, the state of their health, the appearance of people and villages, and the structure of lesser-used languages.
- Gathering the needed data may require integrating clinics, schools, and businesses into digital record-keeping systems; creating incentives for their use; and establishing legal rights over the resulting data.

### Policy recommendations and design principles
- Customize AI to local values and conditions: adjust systems that assume access to expensive resources (for example, digital whiteboards) so they are relevant for teachers lacking those resources.
- Invest in capacity and training of local AI developers and designers to ensure technical innovation reflects local values and priorities.
- Ensure transparency and explainability where eligibility and social protections are concerned, recognizing that acceptable levels of transparency may differ by context and by emergency versus nonemergency settings.
- Anticipate behavioral responses to algorithmic decision rules and design systems (or adjustments) that are manipulation-proof or robust to strategic behavior.
- Inform populations about AI’s societal effects (for example, deepfakes and voice-mimicking robocalls) to reduce misinformation risks and ensure representation of local concerns in regulation.
- Invest in knowledge infrastructure and legal frameworks that enable equitable data collection and protect rights over generated data.

*Daniel Björkegren and Joshua Blumenstock. Artificial Intelligence. DECEMBER 2023.*

---


_Source: https://www.imf.org/-/media/files/publications/fandd/article/2023/december/42-45-bjoerkegren-final.pdf_