Questions in this area
How can AI improve services in low-income countries?
What we've found
In our study with teachers in Sierra Leone, an AI chatbot through WhatsApp was 98% cheaper to access than a web page, including the cost of AI compute. Teachers rated its answers more relevant, helpful, and correct in blinded evaluations. This study focuses on teachers; it does not assess impacts on student learning.
What we're asking
Where can AI help deliver higher quality services at lower cost in low-income countries, in education, health, and other sectors? How can we quickly learn what applications work?
Can digital data expand access to credit?
What we've found
Mobile-phone behavior can help predict repayment for borrowers without formal credit histories. Similar methods have been employed in the emerging industry of digital credit. A randomized evaluation in Nigeria found that access to digital credit yields mostly small welfare effects, but found improvements in subjective well-being after roughly three months. A method to train machine learning models to be robust to manipulation improved predictive decisions in an experiment in Kenya.