The United Nations General Assembly (UNGA) is built to generate buzz. World leaders descend upon New York, shut down streets, and make grand proclamations. Imagine throwing the topic of generative AI, with all its hype, into the mix—this year’s UNGA did just that. Ahead of the High Level Week, UN Secretary-General António Guterres emphasized the urgency of the moment, stating in an interview that most of the Sustainable Development Goals are not being met, and that “We need something to allow them to have a quantum leap. And that quantum leap can come exactly from artificial intelligence if it is properly democratized.”
Quantum leaps sound exciting, but what do implementations on the ground look like, and are they effective? On September 23, CGD and OpenAI co-hosted an event that featured nonprofits from Rio de Janeiro to New Delhi. These organizations presented lightning demos of AI-powered products they are deploying at scale to coach teachers, provide advice to farmers, and speed up care in intensive care units. Senior leaders in the development and humanitarian sectors—David Miliband, Isobel Coleman, Cina Lawson, and Thomas Davin—then took to the stage to discuss emerging trends and challenges for the use of AI in development.
As former implementers of development programs—having both worked at organizations like GiveDirectly and Innovations for Poverty Action—we (Han from CGD and Alex from OpenAI) believe that AI offers practical applications that could, in the future, hold up under rigorous impact evaluation. While we are a few years away from a comprehensive body of evidence, promising trends are emerging. Here are four key takeaways and three open questions from the event:
Takeaways
- Generative AI is being applied to diverse sectors and can help scale interventions. One nonprofit featured at the event, Digital Green, is using generative AI to deliver targeted advice on pest management and weather conditions to smallholder farmers. Kytabu, a social enterprise, leverages AI to tutor students in Kenya at their own level and pace. While neither organization’s AI-powered intervention has undergone an impact evaluation, both draw from an evidence base that suggests AI tools can help scale effective programs. For example, providing digital agricultural advice has been shown to be cost-effective at improving farmer practices and income prior to the introduction of AI. Phone-based tutoring prior to the advent of AI has also shown cost-effectiveness, and early results from a WhatsApp-based AI tutor in Ghana has shown promise.
-
Part of the value proposition of generative AI is its ability to dramatically increase program scale at a fraction of the cost. Digital Green estimates that its Farmer.Chat product can reduce the cost of traditional extension services to $0.35 cents per farmer through a digital service—a 100x cost reduction. Sam Altman, CEO of OpenAI, echoed this point during a fireside chat, explaining that the cost of using OpenAI’s models has decreased significantly—from $15 per million tokens with the GPT-3 model to just $0.15 with GPT-4o. As USAID Deputy Administrator Coleman noted during the panel, these cost reductions can help reach more people in an era when global needs are outpacing aid budgets. While costs are falling rapidly and many AI-powered digital interventions are more inexpensive than their analog versions, it remains to be seen if falling costs can keep pace with the potentially massive increases in reach.
-
Generative AI could free up more time for human interaction, by saving workers time on laborious tasks. Kytabu, one of the organizations featured, shared that teachers in Kenya spend 40 percent of their time creating lesson plans. Kytabu’s AI tools have helped reduce this burden, increasing the time teachers spend with students by 42 percent. In India, 10BedICU uses AI to transcribe doctor-patient interactions and convert them into Electronic medical record entries, reducing the time health care providers spend on data entry by over 50 percent. GPT-4 Vision, OpenAI’s model that processes images, analyzes photos of old hospital monitors to capture data and integrate it into records, allowing healthcare workers to focus more on patient care.
-
AI can bring expertise to the masses, but it also enables grassroot communities to provide information to authorities. “Democratizing expertise” is a popular term that connotes how ordinary users can access virtual “experts” with generative AI. A number of nonprofits are however trying to flow information the other way—enabling local knowledge to be more readily accessible by governments and traditional experts with more resources. VisaoCoop, one of the featured nonprofits, enables remote communities in the Amazon to share environmental data—such as photos or voice memos—through familiar platforms like WhatsApp. They can send a photo, voice memo, or text message, and generative AI then analyzes the data, feeding into a crisis response map. Crisis responders and policymakers then use this map to target resources, intervening in time to protect communities who used to be ignored. Similarly, farmers using Digital Green’s Farmer.Chat can upload photos of their crops and receive a real time diagnosis of any issues, while contributing to a database of locally-derived agricultural best practices.
Open Questions
-
Will the hypotheses for impact hold up? The organizations featured are seeing scaled adoption and retention, but impact requires a step further. For example, a popular generative AI-based job training program may enhance skills but may not uplift livelihoods if labor demand is low. Similarly, a popular children's literacy program might keep students engaged, but lack impact without a proven phonetics curriculum. The evidence base is growing, with a new report by AI for Education finding at least six experimental or quasi-experimental studies on AI-driven education interventions.
-
To ensure equitable access, how will we build capacity for AI? The event’s panelists discussed how AI brings about new challenges such as AI-generated disinformation and touched on how helping populations develop both digital skills and “critical thinking” can mitigate its effects. Beyond these household-level capacity gaps, Minister Lawson also highlighted large scale infrastructure gaps facing sub-Saharan African countries such as lack of internet access, electricity, and data centers. She argued that these need to be invested in concurrent to investments in microeconomic interventions.
-
Looking beyond impacts on individual households, how will generative AI affect labor markets and the macro economy? The event focused largely on individual use cases of AI, but broader questions remain: how will AI affect labor markets? Will AI widen or close the gap between rich and poor countries? In which sectors will it be labor substituting, versus complementary? And as these AI-driven solutions scale, will we begin to see their effects at the macroeconomic level?
As with any good discussion on transformative technology, AI raises more questions than it answers. But discussions like these demonstrate the rapid maturation of the AI-for-development ecosystem and the possibilities ahead. As AI continues to evolve, so too must our applied research and commitment to ensuring its benefits reach the people who need them most.
Han Sheng Chia is an Advancing Evidence for Policy Fellow at the Center for Global Development.
Alex Nawar works for the Global Impact team at OpenAI.
Disclaimer
CGD blog posts reflect the views of the authors, drawing on prior research and experience in their areas of expertise. CGD is a nonpartisan, independent organization and does not take institutional positions.
Image credit for social media/web: OpenAI
