Development Impact Group

The challenge: Research insights can guide investments that reduce poverty and promote shared prosperity, but the vastness of research literature makes it challenging for decision-makers to find and synthesize this information to maximize development impact. Traditional methods of accessing research are slow and cumbersome, with critical insights buried in extensive literature. With a growing development finance gap, estimated at $4 trillion in 2024 – now more than ever – it is critical to allocate the limited resources for maximum impact. 

DIME AI’s innovation: ImpactAI, our flagship project, harnesses the power of large language models (LLMs) to sift through swathes of impact evaluation research and accurately extract insights that provide effortless access to causal evidence for improved decision-making. With a generative AI layer, users can ask the tool questions in natural language and receive quick, reliable, actionable responses. Unlike existing AI-for-research tools, ImpactAI specializes in causal evidence and delivers aggregated, quantitative research summaries and interactive visualizations – delivering months of analysis in seconds. Primary users include development practitioners, policymakers, academics, think tanks, and impact investors. As a one-stop source that lowers the barriers to accessing high-quality research, ImpactAI offers easy access to cutting-edge ideas, best practices, and empirical evidence, thereby enhancing the effectiveness of development projects worldwide.

Scaling our impact: ImpactAI provides a powerful foundation to advance development in various ways. In the future, it will also include predictive models to forecast the impact of policy interventions. Future iterations of the tool will also incorporate more qualitative insights to help users understand the context and mechanisms behind successful interventions. Thereby, we may also be able to assess and guide the design of development project – directly informed by the latest research evidence. To provide diverse and representative research insights, future updates are also expected to include studies in other languages as the tool evolves. 

Get to know the ImpactAI team and learn more about our tool here.

View Publications

EconBERTa: Towards Robust Extraction of Named Entities in Economics | GitHub Repository | Hugging Face Repository

Exploring generative AI for development impact: Learnings from a high-level panel hosted by DIME, supported by Google.org

From research to action: Introducing ImpactAI at the Google.org Impact Summit

View Presentations

ImpactAI: Impact Evaluation Using Generative AI (Measuring Development 2024)

ImpactAI: Product Demo

The challenge: Food Security threatens hundreds of millions globally, with undernourishment rising from 580 million in 2012 to 730 million in 2022 (FAO, 2024). Climate change, population growth, political instability, and economic disparities exacerbate the issue, disproportionately affecting vulnerable communities. Thus, predicting these crises is crucial but hindered due to incomplete and outdated data. 

DIME AI’s innovation: Our research pilot (research paper, blog article) uses natural language processing on non-traditional data sources, such as news articles, to identify and track risk drivers to detect early signs of food crises. Real-time news integration improves forecasts by up to 50% a year in advance and aids decision-making in data scarce environments. By allowing more time for anticipatory actions to be taken, one can improve humanitarian responses and save lives.  

Scaling our impact: The team is now turning the successful research pilot from 37 countries into a live public good with weekly updates at the district level across all 75 IDA countries. Additionally, the project is continuously advancing the modeling approach, including working with local languages to improve predictive accuracy and advancing geospatial signals to provide higher-resolution predictions. Likewise, our model outputs will ultimately feed into anticipatory cash transfer programs to illustrate the value-add of better, more timely food insecurity predictions.

View Publications

Predicting food crisis outbreaks using natural language processing of news streams

Predicting food crises using news streams | GitHub Repository

View Presentations

Food and Humanitarian Crises: Science and Policies for Prevention and Mitigation

The challenge: Large Language Models (LLMs) are vastly underexplored beyond Western, English-speaking contexts from where most training data is sourced (Longpre et al., 2024), rendering their generalizability in the Global South insufficient (Naous et al., 2024). This problem is compounded in low- and middle-income countries (LMICs) where digitized data on local languages is scarcer. As a result, the transformative benefits of LLM-based AI products – for instance to combat online harms and to increase social cohesion – are limited in these ‘lower resource languages.’  

DIME AI’s innovation: Our team builds context-specific LLMs for application in LMICs, capable of working in lower resource languages. We achieve this by undertaking tailored data collection and constructing realistic benchmarks for LLM development and evaluation. In a recent example on Nigerian X (Twitter), we developed a culturally-aware hate speech detection (HSD) model which achieved a 30-point (10x) improvement on HSD compared to the most advanced models at the time. We are progressively building on our context-specific models to explore novel applications. On Nigerian X (Twitter), we used our tailored HSD model to identify users posting hate speech and target them with prosocial advertisements – we found that this can reduce online/offline violence by X% for treated users and Y% for their connections.  

Scaling our impact: Motivated by these strong results, our team is actively exploring opportunities to scale this line of work to (1) adjacent themes like misinformation, (2) further geographies, (3) additional social media platforms, and (4) other modes of data like videos. We see great need and value in context-specific LLMs, wherefore we are also pioneering the augmentation of societal surveys with silicon samples in low-resource contexts. 

View Publications

NaijaHate: Evaluating Hate Speech Detection on Nigerian Twitter Using Representative Data | GitHub Repository | Hugging Face Repository

Large-Scale Demographic Inference of Social Media Users in a Low-Resource Scenario

View Presentations

Detecting Hate Content (Measuring Development 2024)