5 ways global development professionals can support better practices in AI

SAN FRANCISCO — Conversations on artificial intelligence and global development tend to focus on ways AI can transform the sector. Most of the time, global development professionals fail to see what they can bring to the evolution of machine learning and AI tools. For example, the AI for Good Global Summit in Geneva, Switzerland, this week asks how AI can accelerate progress toward the Sustainable Development Goals. But amid a debate on the ethics surrounding information collection in the wake of Facebook scandals over data misuse, some are arguing that it’s equally important for development professionals to weigh in on AI. “Whether you know how to code or not, I want to stress that it’s important for everyone from expert developers to development experts to play a role in ensuring that these tools are ultimately fair, effective, and inclusive,” said Amy Paul, a fellow at the United States Agency for International Development’s Global Development Lab, at a recent conference hosted by the Center for Effective Global Action in San Francisco. Through a Science & Technology Policy Fellowship from the American Association for the Advancement of Science, which allows scientists and engineers to work on federal policymaking, Paul works with USAID on ways emerging technologies can achieve the Sustainable Development Goals. “As development practitioners, it’s our job to look beyond the numbers at the real-world consequences, positive or negative, that they're having.” --— Amy Paul, USAID Global Development Lab fellow But she is also promoting the idea that experts in technology development and international development need to work together to shape the future of machine learning and AI technologies. She and her colleagues believe global development professionals have five key perspectives to offer the development of AI for good in order to maximize the benefits and minimize the risks. 1. Pick the right problem The more data that becomes available — and the easier it is to use machine learning-based tools — the more common it will be to use these approaches for any and all data, Paul explained. “But when resources are scarce, and human capacity to do this work is scarce, just because you can do it doesn’t mean that you should,” Paul said at her presentation in San Francisco. “And development practitioners will have a key role to play in identifying problems that are not only interesting, but relevant and critical problems for the populations they serve.” Machine learning is valuable in a few key areas, she told Devex in a follow-up interview, including understanding new insights about complex processes, processing lots of information quickly, and making tailored predictions about individuals, places, or instances. But Paul said a common problem is people get excited about the potential applications of this data without having data that is high quality enough to be productive. “Even after developing a model that works with training data, revisit that, and ask: Will the level of improvement be worth it?” she said. Key questions include: Who is this data about? What impact could your tool have? Could this data potentially be harmful? The reward has to outweigh the risk, Paul said. Those risks could include hard-coded biases in the data, particularly because of how challenging it can be to understand the layers of data that make up these deep neural networks, and technology failing unevenly across different elements. 2. Know the limitations of your data Machine learning and AI-based tools are only as good as the data used to train them, Paul explained, adding that data from developing countries can be particularly unreliable. “To really understand whether or not the data that you have are adequately representative of the context in which you want to use the tool, you have to understand where that data comes from,” she continued. “You have to understand who or what is represented, who or what might be overlooked, and what kinds of assumptions might be embedded in that data.” She said call detail records are one example of this problem. Drawing conclusions about individuals from call detail record patterns could pose a risk of mischaracterization, as a sim card could be used by more than one person. “While data scientists are going to do great work with the data they have, development practitioners can be an important resource when it comes to thinking about what data we should use as training data, how we label it, and what we should infer from it,” she said. 3. Leverage local expertise The Center for International Tropical Agriculture, a nonprofit research and development organization based in Palmira, Colombia, is leveraging AI to provide site-specific recommendations to smallholder farmers. Paul mentioned this organization as an example of best practices in leveraging local expertise in the development of these tools. As CIAT builds machine learning models, they are working with local growers associations to get access to data, asking extension agents if they are including the right variables, and having smallholder farmers review these models and ask for recommendations. “It’s really about being inclusive from start to finish, thinking about whose perspective matters, and asking what data will be appropriate for developing this tool, do I have relationships with actors who will share that with me, and how do I get feedback from them,” Paul told Devex. She and her colleagues at the Center for Digital Development at USAID are driving the message that the future of machine learning and AI depends a great deal on who makes up research, design, and development teams. “Building successful machine learning and AI-based tools require many different disciplines as well as local expertise, and development practitioners can be a key partner in understanding which voices are going to be important to bring around the table,” Paul said at her presentation in San Francisco, which was held at Google, where technologies are often developed in a setting far away from the populations that might ultimately use those tools. “How often have we seen tools that are developed in one context and deployed in another with less than fantastic results?” --— 4. Speak up for context When it comes to how effective computer vision algorithms can be, context is key — and global development professionals have on-the-ground insights developers might lack. “How often have we seen tools that are developed in one context and deployed in another with less than fantastic results?” said Paul. Local knowledge is crucial to understanding where to look for bias, such as in gender and credit scoring algorithms, and international development practitioners are key to making sure biases are mitigated to the extent possible. “We know the nuances, for example, about how a household survey is conducted and what the norms are for who responds to the questions and what perspectives you might be surfacing,” Paul told Devex. “These are knowledge points that can help inform someone who is building a machine learning model to know how to interpret that data.” 5. Keep people in the equation While machine learning and AI offer tremendous opportunity to advance the SDGs, these technologies can also pose risks to people. “As development practitioners, it’s our job to look beyond the numbers at the real-world consequences — positive or negative — that they're having,” Paul said. Regardless of whether people have technical skills, they can be advocates of the tools, as well as of the complementary policy and regulatory work that supports accountability, privacy protection, and responsible data sharing practices, she said. “One of the biggest things is refraining from this model where technology people are siloed from local context experts, domain experts, subject matter experts,” Paul told Devex. “It’s really about staying engaged with whoever is doing the building of the model, and seeing yourself, if not as an equal, then as an important part of that process.” Technology tends to advance faster than policy, she added, and the global development community will also play a key role in supporting privacy protection and mechanisms for redress and accountability. “These tools will fail sometimes. They’re not perfect. That doesn’t mean we shouldn’t use them. But we need to make sure we have mechanisms in place so that people aren’t unduly disadvantaged when they do,” she said.