The humanitarian and development sectors, including WASH (Water, Sanitation, and Hygiene), are filled with numerous success stories. These narratives of success are undoubtedly crucial, offering insights into what works and helping organizations avoid reinventing the wheel. However, the prevalent culture of primarily publishing positive outcomes is limiting the scope of our understanding and potential learning. Negative data, i.e., information on projects or initiatives that didn't go as planned or outright failed, is largely overlooked. This lack of visibility significantly narrows the scope of knowledge that feeds into AI systems like Retrieval Augmented Generation (RAG).
AI, particularly systems built on RAG, depend heavily on the quality and diversity of the information they're trained on or store for later retrieval. They are designed to provide the most accurate advice based on the available content. If the available data mainly consists of successful cases and lacks the critical learnings from unsuccessful attempts, the AI system can be biased, ultimately providing inadequate advice.
To illustrate, let's consider a hypothetical situation. A development organization planning a sanitation program in a rural area wants to leverage an AI system for the best course of action. The AI, trained only on published data, primarily successful projects, may suggest a community-led total sanitation approach.
However, a similar program previously failed in a comparable context, due to factors such as community resistance, lack of ongoing support, or ineffective triggering techniques. This crucial information was, unfortunately, not documented and shared publicly due to the perceived 'failure' of the project.
The AI system, being unaware of this important negative data, might recommend a similar strategy, thereby leading the organization into potentially avoidable pitfalls.
To leverage AI's full potential in humanitarian and development sectors, it is paramount to balance the scales of knowledge. Organizations must challenge the prevailing culture of primarily publishing success stories and make a conscious effort to document, learn from, and share unsuccessful attempts. This practice is not only a more ethical and transparent way of working but also critical to improving AI systems' ability to give robust advice.
A comprehensive database, inclusive of both positive and negative data, will lead to more well-rounded advice from AI systems like RAG. This approach will not only save resources but also help avoid the repetitive cycle of the same mistakes, ultimately leading to more effective and sustainable projects.