How will USAID use AI? The benefits and risks of new technology

The explosion of new artificial intelligence technologies, or AI, has some in the world of global development excited. Spending on AI is projected to hit $154 billion in 2023, a nearly 27% jump from the amount spent in 2022, and the world of aid sees emerging tools in the technology as potentially transformative for field work as wide-ranging as remote medical treatment, agriculture, the rebuilding of war-torn countries and make the industry smarter, faster, and more effective. USAID, the nearly $50 billion foreign aid agency, is also looking to deploy AI tools to advance its mission. In a new strategy unveiled in March, the agency said it could leverage AI, particularly in instances such as automation of tasks, to help in its localization efforts. This is the goal to spend at least a quarter of a $16 billion funding bucket with local organizations by 2025 and have at least half of its projects be led by local communities by 2030. The work has been slowed by a staffing crisis and USAID said that it was working with the U.S. Department of Homeland Security to test AI tools that can help execute the strategy. The wider development industry is optimistic that the technology offers opportunities to make transformative progress on key issues, ranging from local funding of aid to equitable hiring to disaster response, and could help democratize the world of foreign aid. “The advancement on the commercial side is really going to change what development can do,” said Alexis Bonnell, a former Google executive and former chief innovation officer at USAID. But others in the field warn that there needs to be better efforts made on issues such as privacy and technology biases to ensure that AI in development is ethical and inclusive. With that in mind, what are some of the opportunities and challenges facing USAID and others in development on its journey toward using AI? Leveling the localization playing field Experts say AI could become a tool to help more funds go to local organizations, by offering a way for some of the newer players to compete against traditional legacy contractors, who bring with them sprawling infrastructures and know-how that gives them an edge. AI can help with translation, for example, since not everyone who is interested in accessing funds or expertise from the likes of USAID will speak the language the majority of funding processes are communicated in. AI tools can help translate that material for prospective contractors who in the past may have been excluded due to language barriers. Similarly, foreign aid agencies will be able to offer feedback to bidders in their own language. Generative AI tools such as ChatGPT and Google’s Bard can also assist organizations to produce better, more readable proposals and make them competitive for contracts, Bonnell said. “To use AI as a way to kind of create a more level playing field at that procurement level is really interesting,” she said. Workforce management “Workforce development coming out of the pandemic are incredibly strained, not just the healthcare workforce and frontline workers,” said Sabs Quereshi, who teaches at the Elliott School of International Affairs of the George Washington University and has worked for United Nations agencies and USAID. “By utilizing AI to do some of this work that’s being done on a daily basis, we can really free up limited human resources to focus on more high-priority work.” USAID has said that it wants to deploy software to help its staff more quickly complete repetitive, rules-based tasks, such as the generation of negotiation memos to help cope with its self-described staffing crisis. “USAID is currently working with the Department of Homeland Security to test AI where appropriate within the universe of A&A processes,” the agency said, referring to the blueprint for how the agency gives grants and awards contracts and how it spends more than 85% of its funding. Disaster response Bonnell pointed to programs, such as the World Food Programme’s SKAI project, that leverage AI technologies to assess the impact of disasters such as floods or devastation caused by cyclones. While in the past this type of analysis could have taken weeks, thereby compounding a crisis, it now may take only hours. She also alluded to a Google supported program by the giving organization GiveDirectly that used satellite imagery and census information to identify who needed cash help after a hurricane. “[It] allows you to kind of go in and be very specific and very personalized,” Bonnell said. “What I loved about that was just that idea of really knowing and being able to meet someone kind of where they are and to understand where we can have the most impact.” Managing pitfalls USAID told Devex that it is aware of the potential of AI in development yet the agency is mindful of potential harms that the technology can facilitate. “Recognizing the immense power of AI, we believe we have a responsibility to help shape global deployment of AI in a responsive and ethical manner,” a spokesperson told Devex. Other experts also warn there need to be structures put in place that make the technology serve a more inclusive agenda. “There is a lack … of content, representation, of African narratives within the conversation of AI,” said Tulanana Bohela, co-founder of ONA Stories, a Tanzanian-based firm that works to create stories for augmented and virtual reality platforms. Bohela expressed a fear that African content may struggle to be included because of how fast the technology is moving. “Especially with generative AI, when we talk of people generating images or generating stories or texts or whatever the case [maybe], this generative AI is going off of what’s on the internet but also what’s been prompted,” she told Devex. “One way we are countering that is to bring awareness to this gap.” Other experts worry about whether AI will be able to tackle more entrenched issues, such as institutional racism. “AI can distract people from tackling problems that can only be addressed at a systemic level, and that can be so much harder than training or deploying a [machine learning] model,” said Aubra Anthony, who led a team at USAID that studied emerging technologies and is now a senior fellow at the Carnegie Endowment for International Peace. “But AI in general has a way of enthralling people, of overpromising what is possible, and the development sector isn’t immune to that.” Baking in inequity AI may help address inequity — but there are many instances of it also baking biases into systems. If the data it is trained on reflects human prejudices — and some warn AI data does — then AI will unthinkingly replicate it. USAID said that it is funding projects that help tackle some of these challenges. In Nigeria, in an initiative run by the University of Lagos and Nivi, a company that creates AI-powered medical chatbots, USAID is funding efforts to reduce gender biases that can emerge from the tool. The project won the agency’s Equitable AI Challenge and is now creating an auditing tool within its chatbot to be more accurately aware of the gender of its patients. “The tool will evaluate customer interactions with the chatbot — in English and Hausa — along with how the bot interprets each conversation and then responds,” an agency spokesperson said. Another effort by the agency is in Kenya, where it is helping to study the effectiveness of PROMPTS, a low-cost, AI-enabled digital platform developed by Jacaranda Health that answers questions via text message from expecting and new mothers. These are just some of the efforts the agency says are designed to ensure AI is guided towards responsible and locally relevant development. Data gathering Anthony, from the Carnegie Endowment for International Peace, pointed out that the data that powers AI technologies tends to come from parts of the world that are not the focus of global development, echoing comments by ONA Stories’ Bohela. These tools may struggle when in the field with models that “are markedly biased — because they’re trained on data about a demographic or group that is different from the one the AI is being rolled out for or with,” she said. Anthony pointed out that organizations such as the Mozilla Foundation’s Common Voice, the Lacuna Fund, Masakhane, and others are working to address these inequalities. But that process needs guardrails too, she said. “It’s especially important — and also especially tricky — to ensure that in the effort to build out stronger, representative data sets, that doesn’t turn into an extractive exercise,” she said. “Ensuring people are in charge of how their data are both collected and used for AI is not as straightforward as offering a ‘consent’ box.” In Africa, for example, countries need to think deeply about what can work in their environments and whether there are appropriate regulations to govern these technologies, said Vukosi Marivate, chair of data science at the University of Pretoria in South Africa and co-founder of the Masakhane initiative. “Things like data protections are still uneven across the continent. Some countries have data protections, some do not,” Marivate told Devex. He added that he and his colleagues at his university have been working with the likes of the African Union and other policymakers on the continent to develop such policies. The Masakhane initiative began in 2019 and is also working to build tools that Marivate says are by Africans and for Africans and that are available in different languages from the continent. Some of the data that informs global AI isn’t always inclusive of data from African languages, for example, something that needs to be fixed, said Marivate who is also a co-founder of Lelapa AI, a self-described Africa-centric AI research and product lab. “Some of our products and what we are developing currently we are taking the tough roads and making sure that we can build up the contexts and also be able to take some of the interesting problems that might be tackled with AI that might not be a big item in the West or Silicon Valley,” he said. Through Lelapa, they have built tools such as Vulavula, a language platform with possibilities of transcription and translation in multiple African languages. Bohela, from ONA Stories, told Devex that her outfit is creating a platform ONA Kesho — which means See Tomorrow — that will be a kind of “playground” for African creatives and technologists to experiment with these new tools. ONA has done hackathons to introduce these new tools to find ways to deploy platforms such as ChatGPT in a manner that can include African voices and narratives. Quereshi said now is the time to ensure the right regulations are in place so that these technologies are used to realize development goals and not undermine them. “Regulations with the right equity, with the right representation, with the right individuals, that are part of that collective leadership coming from local populations [who are] informed about what they're experiencing, we can find real solutions that are more tailored to them,” she told Devex. Bonnell compared the way to use AI to how you might work with a really good intern at your office. Very few people would take the work generated by an intern and post it as public policy without some kind of checks and balances. “AI can be a bit of a kickstart, but it doesn't absolve you from leveraging and using your expertise,” she said.