How AI can put 20 years of development evidence to work

A farmer in rural Ghana can use her phone to take a picture of a blighted leaf, upload it, and within seconds she gets a diagnosis that takes account of the soil, the weather, and the other local conditions. Score one for AI. We don't have enough human experts because they are hard to train, and harder still to retain, so apps like this have huge potential. But it doesn't mean they will succeed.

Iqbal Dhaliwal, global executive director at J-PAL, has watched ambitious tech dreams for development fall apart before. Remember one laptop per child, the plan that failed because schools didn't have technical support, the right lesson plans -- or even electricity? He tells Tim Phillips that AI can succeed, as long as whoever is using it has thought through the theory of change, designs for impact rather than downloads, and can scale up successful experiments. J-PAL has launched the AI Evidence Playbook to help policymakers solve these knotty problems.

The research behind this episode

Dhaliwal, Iqbal, Sam Carter, Attaullah Abbasi, and Audrey Lorvo. 2026. AI Evidence Playbook: A Practical Guide. Cambridge, MA: Abdul Latif Jameel Poverty Action Lab (J-PAL).

To cite this episode

Phillips, Tim, and Iqbal Dhaliwal. 2026. "How AI can put 20 years of development evidence to work." VoxDev Talks (podcast).

About the guest

Iqbal Dhaliwal is global executive director of the Abdul Latif Jameel Poverty Action Lab (J-PAL), based at MIT, and co-chairs Project AI Evidence. He co-directs J-PAL's South Asia office with Esther Duflo. Before joining J-PAL in 2009 he served in the Indian Administrative Service, where he ran a statewide welfare department and led a public company, delivering large-scale programmes in the field. His work spans the design, evaluation, and scale-up of anti-poverty programmes, and the question of when a technology genuinely changes lives rather than dashboards.

Research and concepts cited in this episode

Project AI Evidence (PAIE) is J-PAL's initiative to identify, evaluate, and scale applications of AI for social good, and to scale down those that may cause harm. The AI Evidence Playbook is its practical output, a reference for policymakers, practitioners, and donors weighing whether and how to adopt AI-enabled programmes.

The six pathways. The playbook groups two decades of development evidence into six areas where AI could raise impact or cut cost: improving needs prediction and targeting; increasing access to personalised, timely support; maximising the effectiveness of frontline service providers; improving organisational and programmatic efficiency; reducing bias and ensuring fairness; and boosting government resource mobilisation, including more progressive taxation.

Machine learning targeting in Togo. During the Covid-19 pandemic, Togo's government wanted to reach its poorest households but lacked the administrative data to find them. Researchers used satellite imagery to identify likely-poor neighbourhoods from features such as house size and roof quality, then paired it with mobile phone records to narrow down poorer individuals within them; a practical, rapid way to fill a data gap without a social registry.

Theory of change. The sequence of steps connecting an input to a final outcome. Dhaliwal's point is that most people hold an optimistic hypothesis rather than a theory of change: the inputs and the hoped-for outcome are clear, but the intermediary links, adoption, trust, workflows, repair, are missing. Writing it on paper is where the gaps show.

Design for impact, design for scale. Engagement numbers and downloads are a first step, not proof of impact. Designing for scale means designing for a farmer who may lack a device, may lack reliable connectivity, and may not trust the technology without the extension worker they have known for fifteen years.

One laptop per child, and smokeless stoves. Dhaliwal raises both as technologies whose theory of change was sound but whose delivery was not thought through: the school system could not absorb the laptop, the electricity and connectivity were not there, and no one planned for breakage. The relevant cautionary parallel for AI, whose problems he argues are not just inherited but exacerbated by the pace of change.

Global knowledge, local context. Dhaliwal separates rules that can be changed from genuine context differences. A programme that works in India but "won't work in Ghana" because teachers there work six hours rather than six hours fifteen minutes is a rule problem, and solvable. A classroom of 1:20 versus 1:60 is a real contextual difference worth worrying about.

More VoxDev Talks episodes

The development economics of AI: Lessons and questions. Oliver Hanney and Deena Mousa take stock of what an entire series of conversations revealed about where AI helps in development, and where the evidence runs thin.

Related reading on VoxDev.org

AI and development economics: Early evidence and how to keep up, a running VoxDev reading list on the impacts of AI in low- and middle-income countries that includes J-PAL's AI Evidence Playbook among its resources.