AI for social good: An evidence-informed agenda

Artificial intelligence is reshaping how we learn, work, and govern, much like earlier general-purpose technologies that rewired entire economies. The social sector is no exception: governments, NGOs, and social enterprises are rapidly weaving AI into programs, from education to financial inclusion to social protection. 

In addition to questions about AI’s big-picture impacts, pressing questions remain about the effects specific use cases will have on outcomes in the real world.

History gives us reason to be careful. Over the past two decades, new technologies launched with great promise often fell short when context, delivery, and incentives were overlooked—from laptops in classrooms to clean cookstoves. To avoid repeating those mistakes, policymakers and funders should invest in evidence-informed programming and real-world evaluations that measure what changes for people and surface unintended effects. That’s how we leverage both evidence and innovation to learn what actually works, and for whom.

This post, the first in a series, outlines how we can bring this discipline to AI in the social sector. Drawing on a growing base of evidence and J-PAL’s earlier article in the Stanford Social Innovation Review, we outline six promising opportunities for AI to contribute to social good, highlighting emerging and ongoing research: 
 

1. Improving program targeting and needs prediction

Governments and NGOs often struggle to identify who needs support the most. Traditional methods for collecting this information, such as surveys and administrative data, can be slow, expensive, and inaccurate (see Chapter 4 of the Social Protection Initiative Evidence Review for a synthesis of existing evidence on targeting).

AI can improve targeting by analyzing vast data to predict needs more accurately, allowing for earlier and/or more targeted action. For example, in Togo, using algorithms to analyze satellite and phone data enabled rapid, cost-effective, targeted cash transfers during COVID-19. Researchers are working with Google in India to test strategies for disseminating AI-powered flood forecasts that alert households of likely flooding in their area. In another example, researchers are building a predictive model to identify families at risk of eviction in Missouri so that timely assistance reaches them before displacement. 

2. Increasing access to relevant, personalized information

AI tools can expand access to context-specific information, benefits, and programs through personalized content. In Kenya, providing pregnant women with informational messages and an AI-powered helpdesk led to improved knowledge, birth preparedness, and newborn and postpartum care. In the United States, an AI-powered chatbot to support admitted students in the summer before their first year of university increased on-time enrollment in one context—but increased enrollment only for a subset of students in another, emphasizing the need for continued evaluation as interventions expand to new contexts

3. Maximizing the effectiveness of frontline service providers

Through triaging and personalized training, AI can help frontline workers, including health workers, teachers, and law enforcement officers, focus their limited time on the highest-value work. For example, automated essay grading and feedback enabled teachers in Brazil to spend more time engaging with students, improving student performance on a standardized exam. Through J-PAL funding, researchers are also assessing the impact of an AI-based low-cost referral tool that predicts the risk of silent heart attacks in India. 

4. Improving organizational efficiency

AI-enabled solutions can improve the efficiency of organizations by augmenting capabilities, optimizing service delivery, and automating labor-intensive processes. In Mexico, informing workers about conciliation services and machine learning predictions of case outcomes increased immediate settlements when workers personally received the information, resulting in reduced caseloads for courts. In Chile, providing procurement officers with AI-assisted monthly performance reports improved efficiency, but only when the reports were also visible to managers.

Researchers have also identified opportunities for AI to improve efficiency in hiring at both a nonprofit in Ghana and a recruitment firm in the Philippines. Researchers are now exploring related questions with a recruitment software platform in Mexico, and also with the Public Employment Service in France. 
 

5. Reducing bias and ensuring fairness

AI can amplify bias if trained on skewed data or deployed without safeguards such as independent audits and human oversight. However, when carefully designed, it has the potential to reduce human bias in decision-making.

J-PAL affiliated researchers have developed new, diversity-enhancing algorithms for resume screening during hiring processes. Ongoing research is evaluating whether algorithms can help teachers in Italy reduce gender discrimination in educational opportunities.  

6. Boosting government resource mobilization

Many low- and middle-income countries face severe fiscal constraints, making it harder to address pressing challenges like poverty and climate change. AI can help governments mobilize resources more effectively and support a shift toward fairer taxation by improving compliance, detecting fraud and leakages, and strengthening revenue collection and management. For instance, a “digital property tax census” in Senegal led to more accurate and progressive property value assessments than relying on bureaucrats’ in-person assessments.

It’s essential to build these tools to be useful within real-world systems. In India, a machine learning tool flagged tax fraud more accurately than existing methods, but tax collection did not improve, in part due to institutional processes that made it difficult to act on the improved recommendations. Examples like this highlight the importance of context-appropriate interventions and continued evaluation.   

Responsible AI adoption must include evaluation

AI offers an opportunity to help governments, donors, and civil society organizations do more with less while staying evidence-informed. But responsible adoption demands concurrent impact evaluation and transparency about results.

In the coming months, we will feature sector deep dives, from education and health to labor and firms, highlighting where AI is showing promise, where evidence is mixed, and what it takes to deliver impact in practice. Over time, we will also write about using AI for social science research and host guest blogs from researchers within our network. Stay tuned for the launch of J-PAL’s AI Evidence Playbook in coming months, a guide that distills what we know and still need to learn about AI’s role across the social sector.

Are you a funder interested in being part of this effort? Contact us—and to stay up to date on the latest AI research and policy efforts, subscribe to our eNews and select “artificial intelligence” as an interest area in the subscription form.