(Illustration by Caroline Gamon)
Artificial intelligence (AI) has the potential to transform our lives. Like the internet, it’s a general-purpose technology that spans sectors, is widely accessible, has a low marginal cost of adding users, and is constantly improving. Tech companies are rapidly deploying more capable AI models that are seeping into our personal lives and work.
AI is also swiftly penetrating the social sector. Governments, social enterprises, and NGOs are infusing AI into programs, while public treasuries and donors are working hard to understand where to invest. For example, AI is being deployed to improve health diagnostics, map flood-prone areas for better relief targeting, grade students’ essays to free up teachers’ time for student interaction, assist governments in detecting tax fraud, and enable agricultural extension workers to customize advice.
But the social sector is also rife with examples over the past two decades of technologies touted as silver bullets that fell short of expectations, including One Laptop Per Child, SMS reminders to take medication, and smokeless stoves to reduce indoor air pollution. To avoid a similar fate, AI-infused programs must incorporate insights from years of evidence generated by rigorous impact evaluations and be scaled in an informed way through concurrent evaluations.
Specifically, implementers of such programs must pay attention to three elements. First, they must use research insights on where AI is likely to have the greatest social impact. Decades of research using randomized controlled trials and other exacting empirical work provide us with insights across sectors on where and how AI can play the most effective role in social programs.
Second, they must incorporate research lessons on how to effectively infuse AI into existing social programs. We have decades of research on when and why technologies succeed or fail in the social sector that can help guide AI adopters (governments, social enterprises, NGOs), tech companies, and donors to avoid pitfalls and design effective programs that work in the field.
Third, we must promote the rigorous evaluation of AI in the social sector so that we disseminate trustworthy information about what works and what does not. We must motivate adopters, tech companies, and donors to conduct independent, rigorous, concurrent impact evaluations of promising AI applications across social sectors (including impact on workers themselves); draw insights emerging across multiple studies; and disseminate those insights widely so that the benefits of AI can be maximized and its harms understood and minimized. Taking these steps can also help build trust in AI among social sector players and program participants more broadly.
Five Paths
The good news is that researchers, including in our network at the Abdul Latif Jameel Poverty Action Lab (J-PAL) at MIT, have already been evaluating social programs around the world that use machine learning and AI tools to increase impact. From this growing body of evidence, we believe that AI has massive potential to fight poverty and climate change. Specifically, we think AI can greatly boost our efforts in at least the following five ways.
We need to know what works and what doesn’t when we try to infuse our existing social programs with a new general-purpose technology like AI or design new programs based on it.
Improved targeting and needs prediction | Social programs use a variety of tools to identify who the participants should be, whether they be recipients of targeted resources or potential students in an entrepreneurship course. The problem is that these tools, including surveys and poverty thresholds, can be inexact, expensive, and ineffective. But AI can help mitigate these problems by enabling us to leverage tremendous amounts of data to identify who is most in need of support or most likely to succeed by participating in that program.
For example, in India, low-income households in remote areas are particularly vulnerable to floods, which can disrupt or destroy lives and inflict costly damages. More than nine million people in the Lower Ganges are affected by each flood, but it may not be clear which communities require the most support, and the resources and capacity to provide help are often in short supply. In partnership with Google, J-PAL-affiliated researchers are evaluating an AI model in Bihar that forecasts floods. The model predicts flood locations and water levels, sending alerts to smartphones in affected areas. These warnings help at-risk households prepare and evacuate and could enable the government to preemptively direct resources to improve resilience in rural areas. Of course, such programs have to be careful about using the wrong training data or incorrectly programmed models that could result in denial of service to those most in need or most likely to use it.
Increasing access to services | Most social programs and organizations do not have enough capacity to reach everybody who needs them. AI-powered tools can bridge the gap between existing supply of service providers and underserved populations, ensuring that essential services are available to more people, especially in remote or resource-constrained areas.
For example, not enough agricultural extension workers are available to visit every smallholder farmer to offer recommendations. But J-PAL is funding research that is currently evaluating ongoing advances in machine learning that enable any farmer with access to a smartphone to take a picture of a crop and receive instant advice on how to diagnose and treat pests and diseases to improve yields. AI could significantly eliminate knowledge gaps for those who would most benefit from these kinds of services. But we must also ensure that people will trust this new technology and its recommendations.
Maximizing the effectiveness of frontline service providers | Frontline workers typically have the skills necessary to carry out basic duties, but they may lack the advanced knowledge or training to handle more complex challenges. AI can enhance the decision-making capabilities of frontline workers and enable them to perform more high-value work.
For example, community health workers (CHWs) are a crucial source of basic care in the Global South, which suffers a shortage of doctors. But a CHW may not have a formal degree in medicine or a group of experienced peers with whom they can consult on a medical case. A properly trained model could provide diagnostic assistance, evidence-based treatment protocols, and smart triaging, thus leading to better health outcomes. J-PAL-affiliated researchers are proposing the development of an automated screening tool that uses machine learning to predict the results of expensive, “gold standard” diagnostic tests (e.g., an echocardiogram, or ECG) from data provided by less expensive, mobile tests (e.g., an ECG mobile tool, or a blood pressure and heart rate check) that can be performed at a patient’s home.
Reducing biases and ensuring fairness | True, algorithms can carry biases, but they are not inherently biased, and we can eliminate these issues through good engineering. When programmed correctly, AI can even help identify and mitigate existing biases in government and social programs, ensuring fairer distribution of services.
For example, evidence from researchers affiliated with J-PAL has shown that human-made systems, such as hiring processes, can be rife with bias. On the other hand, algorithms designed to screen candidates based on their potential rather than simply favoring those with conforming résumés can be used to address these biases and drive more equitable outcomes. But assuring good results requires rigorous evaluations to make sure that the algorithms are not in fact magnifying existing biases.
Boosting resources and efficiency to tackle the biggest challenges | The fight against climate change and poverty needs a huge increase in resources, but governments worldwide are facing unsustainable fiscal deficits, thanks in part to pandemic-era growth shocks and high spending on support programs. This fiscal problem requires that countries increase tax compliance. Here AI can help with audits of tax records—something that was previously tedious, labor-intensive, and expensive—and turn them into quick and affordable processes, freeing up tax collectors to do tasks that humans are better suited for and reducing the need to expand staffing.
For example, J-PAL-affiliated researchers working in India found that machine-learning techniques can identify likely tax evaders, potentially widening the tax base massively. Human workers can then verify these cases with in-person follow-ups. However, it’s crucial to evaluate these techniques to avoid unintended consequences, such as overreliance on AI at the expense of human auditors.
Evidence-Informed AI
While we are all justifiably excited about the potential impact that AI can have in these areas, the actual impact remains to be seen in the fight against poverty and climate change. In addition, questions remain about AI’s effects on the labor market and welfare. Will it displace jobs or help the less skilled move up the productivity and income chain? Will it displace jobs from high-income countries to low- and middle-income countries, or bring offshored jobs back to the affluent ones? Will the current tax structures that favor capital investment and penalize employment shift AI from being a skill enhancer to a job displacer?
To inform our future decisions, we need to know what works and what doesn’t when we try to infuse our existing social programs with a new general-purpose technology like AI or design new programs based on it. Thanks to more than two decades of building the field of rigorous impact evaluations in partnership with implementers, donors, and local communities, we can shape the direction of AI by learning from past evaluations and designing evidence-enhanced AI programs from the start.
Read more stories by Iqbal Dhaliwal & Michael Hou.