Responsibly developing data-driven models for humanitarian decision-making: our research on AI for Food Security Monitoring and what we can learn from it

Despite the increasing availability of data from various sources, it remains difficult for humanitarians and governments to respond adequately and quickly to unfolding humanitarian crises.  One of the problems that causes this, is the challenge that decision-makers face in assessing the impact of a given shock or hazard on the local population. Therefore, a major issue for the development of early warning systems for humanitarian action lies in contextualizing the data to go from a specific hazard or shock event to its impact on the local population. Given these problems and the developments in AI and data-driven modelling in the past decade, there are many hopes that AI can close this information gap. 

However, many scholars and practitioners are apprehensive about using (often complex) data-driven models for actual humanitarian decision-making in practice, and rightfully so. Different documented cases from the public sector such as the Dutch child-benefit scandal or the American COMPAS case have shown what harm the irresponsible use of AI-informed decision-making can do to already vulnerable and marginalized populations. Thus, the question remains how to responsibly develop data-driven models that are useful to the humanitarian community.

In our trans-disciplinary research in collaboration with the Integrated Food Security Phase Classification (IPC), we spent a year exploring this question in a case study on how data-driven models impact the IPC's decision-making process on Acute Food Insecurity analysis updates. Using a human-centered design approach, we systematically analysed the current IPC decision-making process and their information needs and evaluated existing food insecurity models with respect to their suitability, while simultaneously conducting literature research on how to develop AI solutions in a value-driven way. These explorations indicated that the common approaches to developing data-driven models, as well as existing theoretical frameworks with regard to responsible AI implementation, have clear mismatches and shortcomings compared to what may be needed in practice. From this, we draw several lessons on how to improve, so that models for humanitarian decision-making can bring actionable insights, are understandable by its end-users, and embody the humanitarian values.