Large language models are developing fast and changing the way we work. We’ve seen exciting applications of generative AI in different areas – from transforming how we code to use cases in biology with generative molecular models. At Laterite, we’re constantly looking to integrate the latest tools into our workflows and offer better insights to our clients. Managing Partner Dimitri Stoelinga talks about some of the ways that LLMs could impact the future of development research.


Intelligent knowledge hubs

A persistent challenge in development lies not in the scarcity of data, but the fragmented use of evidence. At Laterite, we see this in the sectors we work in – education, public health, agriculture, livelihoods, and gender – where organizations are constrained by scattered knowledge. This limits problem-solving and stifles innovation. LLMs can transform the knowledge equation. By establishing databases with curated information from various sources enhanced by retrieval-augmented generation (RAG) and LLMs, we can create intelligent knowledge hubs. These would not only facilitate easier access to existing knowledge, but also enable the generation of entirely new insights. Laterite is already testing some of these solutions in the education space.

An entirely new class of data

In the social sciences, we classify data into two types: quantitative data (from surveys, administrative data or spatial data) and qualitative data (from focus group discussions, semi-structured interviews, etc). With their multimodal capabilities, LLMs blur this distinction by making it possible to extract insights from any type of information to translate data into a structured narrative. For example, pictures, powerpoints and publications can now be seamlessly parsed and analyzed with LLMs. This study by Lee Crawfurd, Christelle Saintis-Miller and Rory Todd from the Center for Global Development which studies gendered language in school books across 34 countries using LLMs is a great example of what is now possible.

Qualitative data at scale

Qualitative methods such as focus group discussions and key informant interviews offer incredibly rich insights and bring nuance to survey data. However, they aren’t scalable. A single interview might result in 20 pages of text that need to be transcribed, translated, coded, and analyzed. The number of pages for analysis can quickly run into the thousands beyond a sample size of 50. In addition to the time commitment for researchers, it can also lead to errors or missed insights – even with the help of software. LLMs make it possible to mimic the qualitative research process at scale and generate high quality insights effectively. We’ve been testing LLM agents for inductive and deductive qualitative analysis, with very promising results.

More inclusive development research

LLMs open the door to making research more inclusive. Discussions around inclusivity in research often pivot to ethics and the discourse on decolonizing research. However, the operational hurdles to achieving inclusive research receive less focus. Inclusive research should be answering questions that matter to the population being surveyed; integrating contributions from a broad spectrum of the population into the research instrument design; involving enumerators and participants in the analysis phase; and feeding results back to everyone involved. Project timelines and budgets make it very difficult to do this at scale.

By combining LLMs with mobile communication technologies, we can envision a future where research is more responsive. LLMs will make it possible to engage with more people throughout the research cycle. In the future, AI will help us share feedback with participants at scale, creating a more collaborative and engaging process.

What’s next?

These are just a few of the many possible applications of large languages models that could change the way we do development research. LLMs can’t replace in-person data collection and survey data, but they have the potential to make research more efficient, more inclusive, and more actionable for implementers and decision-makers alike.