Harnessing AI for humanitarian action: Moving from response to prevention

Global humanitarian needs have surged due to escalating armed conflicts, climate-induced disasters, and political instability. Anticipatory action in the humanitarian sector is regarded as a key tool to help address this surge. This column discusses the potential for AI-based forecasting systems to inform anticipatory action in the humanitarian sector as well as the barriers to realising this potential. It then proposes areas of collaboration between academic researchers and humanitarians to overcome these barriers.

Quantitative prediction has made huge strides in the last decade thanks to a combination of increased data availability, advances in machine learning algorithms and a dramatic fall in computational costs. Prediction algorithms already complement human experts in a number of policy areas like medicine or conflict prevention.They can strengthen the knowledge base for policy makers by training on a large number of cases and with complex architectures that bring subtle patterns to light that would not have been visible to humans otherwise. In this way, quantitative forecasts are able to provide an extra ‘opinion’ that is data driven, and, hence, can form part of a broader trend towards quantitative policy evaluation. At its best, forecasting can be hugely valuable to humanitarian anticipatory action (AA) (Altay and Narayanan 2020, Wagner and Jaime 2020).

As shown in Figure 1, earlier provision of assistance can reduce the size and impact of a crisis, and forecast-based triggers should be able to play a central role in warning of a shock. However, although capabilities to anticipate crises have grown, only an estimated only 0.4% of humanitarian funding was allocated to AA. Most humanitarian aid is reactive, released only after disasters unfold. The low levels of funding given to AA may in part be a reflection of overall funding shortages in the humanitarian sector, but may also reflect a degree of scepticism towards predictive algorithms.

Figure 1 Anticipatory action (AA) enables earlier provision of assistance

Source: UNICEF

Using algorithms for humanitarian AA does indeed require careful consideration and faces multiple barriers:

1. Data availability: Reliable and frequently updated data are much more available for meteorological events than for targets such as armed conflicts or displacement. Machine learning algorithms also require a long history of data to learn from and the discontinuation of human-encoded data sets (such as the Polity Project by the Centre for Systemic Peace) creates a serious barrier to forecasting. Finally, human-made crises are a social phenomenon and so-called black swan events that do not have predecessors in the training data are more common.
2. Forecast literacy: Many applications of forecasting for AA will still require human(s)-in-the-loop who can contextualise and triangulate forecasts in order to decide on an appropriate response. However, a combination of algorithm aversion (e.g. Dietvorst et al. 2014) and insufficient confidence in how to interpret forecasts can hamper their use. Making forecasting results more explainable may increase their credibility. However, there are potential trade-offs between explainability and performance, as well as a risk of interpreting attempts at explaining feature contributions as causal drivers of predicted values (Yang et al. 2022).
3. Ethical risks: Forecasts can be misused by bad actors or trigger unintended responses. Therefore decision makers need to carefully consider the potential of forecasts to be self-fulling or self-defeating and what this means for the extent to which forecasts are made publicly available. There is also risk of bias as forecasting systems rely on historical data, which may be incomplete, inaccurate, or politically manipulated: if training data exclude specific events, forecasts may perpetuate biases, reinforcing unequal resource allocation.

Given these barriers, it is not surprising that forecasting systems have become more common in the context of AA for natural disasters like floods or earthquakes rather than in the context of armed conflict. Two examples of systems in place highlighted below are the Start Network’s Start Ready programme and the Anticipatory Action Framework piloted by the UN Office for the Coordination of Humanitarian Affairs’ (OCHA) in Bangladesh.

Start Ready, launched in 2022, pre-positions financing for recurring and predictable crises such as floods, droughts, and heatwaves. Start Network members collaborate at a national level to design and establish a disaster risk financing system, which involves analysing risks and developing contingency plans. Funds are rapidly disbursed when pre-agreed thresholds are reached in risk models and used by Start Network members to implement pre-agreed plans for anticipatory and early response activities. As the focus is on natural hazards for which there is highly reliable data, there is less need for additional human interpretation of monitoring and forecasting systems.

Similarly, UN OCHA’s Anticipatory Action Framework in Bangladesh targets severe monsoon floods along the Jamuna River. By using dual triggers from the Global Flood Awareness System (GloFAS) and Bangladesh’s Flood Forecasting & Warning Centre (FFWC), OCHA activates funding when flood forecasts surpass pre-defined thresholds. In 2020, this system preemptively allocated $5.2 million, providing aid to 200,000 people ahead of expected floods. Historical trigger data from GloFAS and FFWC (Figure 3) showcase how water discharge and flood risk indicators inform funding decisions, enabling multi-sectoral pre-disaster interventions that reduce the impact of floods on vulnerable communities.

Figure 3 Historical analysis of triggers in the GloFAS and FFWC models for the anticipatory humanitarian action pilot for the 2020 monsoon floods in Bangladesh

In the case of armed conflict, using forecasts to aid AA remains limited due to data constraints and political sensitivities. The data landscape of armed conflict events has hugely benefited from the working of the Uppsala Conflict Data Program (UCDP) and Armed Conflict Location and Event Data (ACLED). However, human-encoded data is also subject to unintentional omissions as well as changes in encoding over time. Forecasting armed conflict is inherently complex and potential errors in the data can be exacerbated when used for prediction.

Crucially, unlike floods, conflicts are driven by human decisions, making them endogenous to reporting and predictions. Compared to setting up flood defences or prepositioning goods to aid people affected by drought, potential AA in the context of conflict is therefore extremely sensitive as it can signal a forecast. Even though approximations suggest huge potential benefits of intervening early (Mueller et al. 2023), the reality of international politics often makes this difficult in practice.

The Humanitarian Forecast Working Group consists of the participants of a workshop held at the Institute for Economic Analysis (CSIC) in Barcelona in May 2024. ⁵ A key insight from the workshop was that the adoption of AI-systems in the humanitarian sector will require collaboration across academia, foreign offices, development as well as humanitarian organisations. A first product of this collaboration is a joint Policy Insight which provides design considerations for adopters in the sector and highlights examples of best practices (Humanitarian Forecast Working Group 2024).

However, the paper is also a call for action. In order for forecasts to be effectively used for AA, we suggests three key pillars for further development:

1. Improving the data landscape: Without good data there is no good forecast. Continued maintenance of data sets such as UCDP and ACLED is vital. Developing new data sets (e.g. for displacement) are also critical, with the Complex Risk Analytics Fund (CRAF’d) playing a crucial role in enabling such efforts. Especially great value lies in data that is highly granular both in its spatial and temporal resolution and provides long histories for training machine learning models.
2. Expanding forecasting methods: One dimension for improvement is the better integration between econometric panel data methods and machine learning. Another is to improve efforts to make predictions of distributions rather than point predictions. ⁶ This could be further expanded to predicting joint distributions in time and space so that these forecasts could be used flexibly to answer an infinite number of conditional questions. Further efforts could be made to address endogeneity issues of self-fulfilling and self-defeating forecasts by trying to account for the potential impacts of interventions in a “dual” forecasting system.
3. Continued exchange: Forecasting conflict and humanitarian need is highly political. To make academic research most useful for the policy community, regular collaborations and exchanges are indispensable. One key problem identified by the group was, for example, the communication between ‘analytical’ and ‘intuitive’ staff within humanitarian organisations, development agencies, and foreign offices. These interfaces come with a whole bandwidth of problems such as the different requirements for update frequency and forecasting horizons, as well as the need for contextualisation and the explainability of machine learning outputs. Most of these aspects can only be addressed through collaborations. Researchers, data providers and data scientists need to understand what is feasible (ethically, politically and organisationally) in practice, but can also help shape these constraints through capacity building and the development of new methods.

While the path forward involves navigating technical, political, and ethical challenges, the potential rewards are immense: saving lives, reducing suffering, and improving aid efficiency.

Source : VOXeu