Abdul Latif Jameel Poverty Action Lab (J-PAL), at Massachusetts Institute of Technology (MIT)

A successful first Global Dialogue would move the conversation from broad principles to practical guidance on how AI can improve real-world outcomes while managing risks. From the perspective of the AI Evidence Playbook, success would mean creating stronger consensus that AI governance should focus not only on model performance, but on whether AI-enabled programs improve outcomes such as learning, health, income, access to services, fairness, and well-being. The Dialogue should also help countries and institutions identify where AI can add the most value, what conditions are needed for AI tools to work in practice, and how to evaluate their impact before scaling. This is especially important for low- and middle-income countries, where AI tools may fail if they are not adapted to local languages, infrastructure, institutional capacity, and user needs. A strong outcome would include clear pathways for capacity-building, evidence generation, and shared learning across governments, researchers, civil society, and implementers. It would also be valuable if the Dialogue encouraged funders and policymakers to support rigorous pilots, impact evaluations, and responsible experimentation. This would help ensure that AI investments are guided by evidence, rather than hype, and that AI governance remains grounded in public benefit, accountability, and equity.

• Safe, secure and trustworthy AI
• Social, economic, ethical, cultural, linguistic and technical implications of AI
• Transparency, accountability, and human oversight
• AI capacity-building

These priorities are central to ensuring that AI is used responsibly and effectively for social impact. The AI Evidence Playbook emphasizes that AI should be treated as a means to improve real-world outcomes, not as an end in itself. Capacity-building is therefore essential, especially for governments, implementers, and civil society organizations that need practical tools to decide where AI can add value, how to design programs around it, and how to evaluate whether it works. The social, economic, ethical, cultural, linguistic, and technical implications of AI are also urgent because AI systems often perform differently across contexts. Tools that work in one setting may not translate into impact elsewhere if they do not account for local languages, frontline workflows, institutional constraints, user trust, and unequal access. Transparency, accountability, and human oversight are important because AI-enabled programs can affect public services, livelihoods, and access to opportunities. Strong oversight helps ensure that decisions remain explainable, contestable, and aligned with public interest. Finally, safe, secure, and trustworthy AI is a core condition for adoption and impact. Without trust, safeguards, and evidence on real-world performance, AI systems may either fail to improve outcomes or create new harms. Together, these priorities support an approach to AI governance that is practical, evidence-based, and focused on measurable public benefit.

One cross-cutting issue is the need to build stronger evidence on when AI actually improves outcomes in real-world settings. Much of the current AI discussion focuses on technical performance, such as accuracy, engagement, or benchmark scores. However, early evidence shows that strong model performance does not automatically translate into better outcomes for people. AI tools may be accurate in isolation but still fail if they are poorly integrated into workflows, not trusted by users, or unsupported by institutions. Another emerging issue is the gap between AI experimentation and responsible scaling. Governments and implementers need practical guidance on when to pilot, when to adapt, and when there is enough evidence to scale. This requires investment in rigorous evaluation, including randomized evaluations where appropriate, as well as monitoring for unintended consequences. A further issue is ensuring that AI governance reflects the needs of low-resource settings. Many AI systems are designed around high-income country infrastructure, languages, and institutional capacity. Governance discussions should therefore pay closer attention to locally appropriate AI, including tools that are affordable, multilingual, usable by frontline workers, and resilient in settings with limited connectivity or data systems. Overall, evidence generation should be treated as a core part of AI governance, not a separate technical exercise.

Governance gaps in these areas are affecting the global AI-for-development sector in several ways. The most significant challenge is that AI adoption is moving faster than the evidence base. Governments, donors, and implementers are increasingly interested in using AI to improve education, health, labor markets, agriculture, and public service delivery, but there is still limited evidence on when AI-enabled programs actually improve real-world outcomes. This creates a risk that resources are directed toward tools that perform well technically but do not translate into better learning, income, health, access, or well-being. A second challenge is uneven capacity. Many low- and middle-income countries, civil society organizations, and frontline institutions face constraints related to data quality, infrastructure, procurement, staff training, and local technical expertise. These gaps make it harder to assess AI tools, adapt them to local needs, and govern their use responsibly. There are also important risks around accountability, human oversight, and equity. AI tools used in public systems can affect access to services, jobs, credit, or benefits. Without clear oversight, transparency, and mechanisms for contesting decisions, AI may reinforce exclusion or reduce trust. At the same time, the opportunity is significant. If designed and evaluated well, AI can help improve targeting, provide personalized and timely support, strengthen frontline service delivery, improve organizational efficiency, reduce bias, and support government resource mobilization. The key opportunity is to make evidence generation a central part of AI governance. Pilots, impact evaluations, and shared learning can help the global community identify where AI adds value, where it does not, and what safeguards are needed before scaling.

The AI Dialogue can play an important role by creating a shared space where governments, researchers, civil society, implementers, and funders can move from broad AI principles to practical cooperation. In the AI-for-development sector, there is strong interest in using AI to improve outcomes in education, health, labor markets, agriculture, governance, and public service delivery, but the evidence base remains limited. The Dialogue can help align international cooperation around a core question: not only whether AI systems are technically capable, but whether they improve real-world outcomes safely, equitably, and cost-effectively. From J-PAL's perspective, including through Project AI Evidence (PAIE) and the AI Evidence Playbook, the Dialogue could help make rigorous evidence generation a central part of AI governance. This includes encouraging responsible pilots, randomized evaluations where appropriate, implementation research, and transparent learning from both successful and unsuccessful AI applications. The Dialogue can also help ensure that low- and middle-income countries are not only recipients of AI tools, but active contributors to global governance debates. Their priorities include capacity-building, local language access, infrastructure constraints, frontline delivery challenges, and institutional readiness. Overall, the Dialogue should support a global evidence and governance agenda that keeps AI focused on measurable public benefit, responsible experimentation, and informed scaling.

The AI Dialogue should connect with initiatives that are already building evidence, capacity, and practical guidance for responsible AI use. One relevant example is J-PAL's Project AI Evidence (PAIE), which supports rigorous evaluations of AI applications for social impact and helps connect researchers, implementers, and funders around promising AI-enabled programs. PAIE's work, including the AI Evidence Playbook, focuses on identifying where AI can add value, what conditions are needed for successful implementation, and how AI programs should be evaluated before scaling. The Dialogue could also build on related efforts across the AI-for-development ecosystem, including initiatives focused on AI capacity-building, responsible data governance, digital public infrastructure, human rights, public sector innovation, and AI safety. These efforts often address complementary parts of the same challenge. Some focus on technical safeguards, others on governance norms, and others on implementation and evidence. The added value of the AI Dialogue would be to bring these strands together and elevate evidence as a core component of AI governance. It could encourage countries and funders to ask practical questions before scaling AI: What problem is being solved? Why is AI needed? What safeguards are in place? How will impact be measured? And what conditions are needed for the tool to work in practice? By connecting initiatives like PAIE with global governance discussions, the Dialogue can help ensure that AI policy is informed by real-world evidence, not only technical promise.

The AI Dialogue should create structured channels for governments, researchers, civil society, implementers, affected communities, and funders to contribute based on their respective roles. Governments can identify priority public service challenges and policy needs. Researchers can contribute evidence on whether AI-enabled programs improve real-world outcomes. Civil society and affected communities can identify risks related to exclusion, rights, trust, and accountability. Implementers can share practical lessons on delivery, adoption, and institutional readiness. Funders can support responsible pilots, capacity-building, and rigorous evaluation. The Dialogue would be most useful if organized around concrete use cases and evidence needs, rather than only broad principles. Sector-focused sessions on education, health, labor markets, agriculture, and governance could examine what is already known, what remains uncertain, and what should be tested before scaling.

Global AI governance discussions often underrepresent low- and middle-income country governments, frontline service providers, local civil society organizations, and communities directly affected by AI-enabled public services. Researchers and implementers working on real-world delivery challenges are also often less visible, especially those operating in settings with limited infrastructure, local language needs, and weak data systems. These perspectives could be included through regional consultations, funded participation, multilingual engagement, and structured opportunities to present evidence from practice. The Dialogue should also include organizations supporting rigorous evaluation of AI applications for social impact, including initiatives such as J-PAL's Project AI Evidence, so that governance discussions are informed by implementation realities, evidence needs, and measurable public benefit.

The Dialogue could use practical, evidence-focused formats that connect governance debates with real-world implementation. Use-case clinics would allow governments or implementers to present a concrete AI application and receive input from researchers, civil society, technical experts, and funders. Sectoral evidence roundtables could review existing evidence, identify knowledge gaps, and define priority questions for future evaluation. The Dialogue could also include learning sessions on cases where technically promising AI tools did not translate into real-world impact. This would help create a more balanced discussion of both opportunities and risks. Matchmaking sessions could connect implementers, researchers, and funders to support responsible pilots, impact evaluations, and capacity-building. These formats would help ensure that the Dialogue produces practical cooperation, not only general statements of principle.

Effective AI governance should combine clear safeguards with practical evidence on whether AI improves real-world outcomes. Several approaches are especially useful. First, evidence-based evaluation mechanisms can help governments and funders decide when AI should be piloted, adapted, or scaled. J-PAL's Project AI Evidence (PAIE) is one example. It supports rigorous evaluations of AI applications for social impact, including pilots and randomized evaluations, and connects researchers, implementers, and funders around concrete use cases. Second, practical guidance tools can help decision-makers assess AI beyond technical performance. The AI Evidence Playbook offers one such approach by asking where AI can add value, what conditions are needed for implementation, and how impact should be measured. This helps shift governance from abstract principles to decisions about real programs, safeguards, and outcomes. Third, governments can strengthen AI governance through responsible public procurement standards, including requirements for transparency, human oversight, data protection, bias assessment, independent evaluation, and mechanisms for contesting decisions. These standards are especially important when AI is used in public services. Fourth, regulatory sandboxes and policy innovation labs can allow governments to test AI tools in controlled settings before wider deployment. These should include affected communities, frontline workers, technical experts, and independent evaluators. Finally, public learning platforms can help share results from both successful and unsuccessful AI applications. This is important because technically promising tools do not always improve real-world outcomes. Governance systems should therefore reward transparency, learning, and adaptation, not only adoption. Together, these practices can make AI governance more accountable, practical, and focused on measurable public benefit.