Development Impact Group

Select a EDS Sub navigation page selecting option, leaving this page

Image

Development Impact AI Lab

AI built for development. Tested where it matters.

The Development Impact AI Lab builds open AI tools designed to solve real development challenges.

Based in DECDI and launched in 2023, the Lab brings together AI researchers, economists and engineers to build tools with governments, World Bank teams and development partners, test their impact in the field, and integrate them into the systems people already use to make decisions.

Our work focuses on three areas where AI can make a meaningful difference:

  • Evidence for better decisions
  • Early warning for emerging crises
  • AI adapted to local languages and contexts

Our approach connects research, product development and impact evaluation in one continuous loop:

Understand the need → Build → Test → Embed → Measure → Improve

Everything we build is designed as a public good. Our tools, models and datasets are open wherever possible, global in ambition and grounded in local contexts.

What we do:

We build public-good AI systems for development: evidence tools, crisis early-warning dashboards, domain-specific models, and curated datasets. 

Better evidence into every policy decision

ImpactAI is a decision support tool built on the largest curated base of causal evidence: 15,000+ randomized trials across 150+ countries. Ask what works and it compares the effectiveness of programs, shows where the evidence is thin and suggests alternatives. Unlike other AI products, it puts every effect on one scale, traced back to its source, so programs can be compared rather than summarized. The AI extraction behind it is on par with human experts at about 85% accuracy, backed by our research at top tier computer science conferences and explained on our blog. A first cohort of adopters spans development banks, donors, foundations, governments and NGOs, including ADB, FCDO, the Gates Foundation, the Government of Odisha and Digital Green. ImpactAI was selected for Google.org’s first Generative AI Accelerator in 2024. See it in action in a one minute video.

Early warning for anticipatory action

Food crises rarely strike without warning. The signs are in the local news long before an assessment, yet only about 0.2% of humanitarian funding moves before the shock. ZeroHungerAI reads millions of news reports a day and turns them, with satellite data, into district level risk signals in 82 countries. Conflict, crops, displacement, prices and weather sit on one live map: 10 million events over 20 years, each traceable to its sources. To our knowledge, no other system offers high frequency risk data across sectors, at district level and on a global scale. It picks up food crises up to 12 months ahead and catches 46% more outbreaks than existing baselines (Science Advances, 2023). It is open access with an open API. Yet a signal only matters if it reaches the systems that act early. We are in discussions with the Bank’s Crisis Response Window and UN CERF to build it into their decisions.

With the AI Collaborative for Food Security, we combine our signals with WFP’s HungerMap and NASA Harvest to provide forecasts to governments and humanitarian partners. See it in action in a one minute video.

Small AI adapted to low resource contexts

AI underperforms where the social gains could be the highest. Models trained on Global North data do worse where the poorest live, and larger models will not fix it, because thousands of languages and domains each need local knowledge that no company has a strong incentive to collect. We showed it in Nigeria. Hate speech models that look strong on convenient test sets fall apart on representative Nigerian social media. A small language model we fine tuned on representative data we collected ourselves, NaijaXLM-T, closes the gap: 90% against 60% for ChatGPT (ACL 2024; blog part 1 and part 2). This is the small AI argument we made in the WDR 2026 spotlight. Few labs take a model all the way from training to a field trial of its impact. We did. In a randomized trial in Nigeria, advised by the Centre for Democracy and Development, the model sampled the users and measured the outcome of a preventive messaging campaign by Nigerian celebrities. The campaign led to a sustained reduction in hate speech, over 50% among the most exposed users (paper, under review at Nature Human Behaviour). 

Contact

For partnerships, demos, or general inquiries: dimeai@worldbank.org

 

Back to Programs
Meet our experts.
Get to know DIME AI

We are a multidisciplinary team of AI engineers, development economists, and software developers using impact evaluation and AI for social good.

Meet Our Team
ImpactAI logo and tagline (Generating Insights for Global Development)
Explore ImpactAI

The World Bank's Development Impact Group is using generative AI to help policymakers and practitioners access development economics research and make better policy decisions.

Learn More
Read our insights.
Impact Intelligence: DIME AI Blog

Explore the DIME AI Blog for updates on our events, in-depth looks at innovative tools like ImpactAI, and the latest in research and projects driving data-driven solutions for global development.

Read DIME AI Blog