• Long before today's generative AI tools, FEWS NET was using machine learning, remote sensing, and advanced statistical models to support early warning efforts.

• This blog explains why AI can improve famine early warning systems but cannot replace the local expertise, accountability, and contextual understanding that underpin FEWS NET’s work.

• The blog also explores how FEWS NET is building the foundations for smarter and more accountable forecasting through causal graphs, enhanced market analysis, structured assumptions databases, and AI-enabled reasoning.

Who on the planet is at greatest risk of life-threatening hunger? Where should humanitarian efforts be focused, and how should limited aid resources be allocated to avert loss of life?

For decades, governments and humanitarian organizations have relied on forecasts and monthly analyses produced by the Famine Early Warning Systems Network (FEWS NET) to understand where hunger is expected to spike and make timely, cost-efficient decisions.

Each month, FEWS NET estimates the population in need of humanitarian assistance around the world eight months into the future. These estimates, along with detailed reporting on the drivers of hunger emergencies, help donors plan ahead and strategically deliver aid before the worst outcomes occur.

FEWS NET has been doing this work for 40 years. From its earliest days, the network has adopted and advanced what would today be described as artificial intelligence (AI), even before the term became ubiquitous.

As new programmatic and analytical methods have emerged – like advanced statistical models, remote sensing, automation, and machine learning (ML) – FEWS NET has integrated them into its workflows to improve forecasting, scale analysis, and reduce uncertainty.

Today, much of the discussion around famine early warning centers on how AI can be used to automate systems like FEWS NET.

AI has a consistent technical definition, but what people refer to as AI tends to change over time. Methods that were once considered cutting-edge – like optical character recognition or early chess engines – are still technically AI, even though they are now mostly described as standard software. Much of today’s public conversation focuses on a specific subset of AI: generative AI powered by large language models (LLMs).

Now many are asking: Can today’s AI completely replace FEWS NET?

“It’s not about replacing FEWS NET with AI,” University of Maryland Assistant Research Professor Dr. Weston Anderson said. “It’s about using new tools responsibly and effectively. FEWS NET has been around for 40 years because it’s been exceptionally effective at taking up new methods and technologies and applying them in an accountable way.”

That accountability and responsibility are where FEWS NET experts are focusing their efforts as they work to understand what’s coming next with AI, and how it can be integrated to make existing systems more nimble and efficient.

How FEWS NET already leverages different types of AI

Examining how FEWS NET uses different types of AI today begins with understanding the complexity of the questions that must be answered to assess acute food insecurity.

Food security depends on far more than rainfall and harvests. Staple food and commodity prices, seasonal labor demand, livestock health, pest outbreaks, government and international assistance, and conflict all affect people’s ability to grow, buy, gather, or otherwise obtain food.

Each month, FEWS NET analysts follow a structured process to assess and forecast food security outcomes. They start by setting the parameters for the scenario and describing the current situation on the ground. From there, they make key assumptions about what might happen to households’ income and food sources, and use these to project both household-level food security and broader community-level outcomes. Finally, analysts consider events that could change the scenario, like a failed rainy season or a sudden spike in conflict, so that forecasts remain flexible and responsive to emerging risks.

“One of FEWS NET’s greatest values is the collection, aggregation, and standardization of many disparate data sources,” FEWS NET Program Manager Kevin Coffey explained. “Acute food insecurity is driven by a complex web of factors like crop production, economic stability, local and global conflicts, and more. Compiling comparable, cleaned, and verifiable data is something FEWS NET is already using AI to do better every day.”

While some of the workflows involved in collecting, cleaning, and standardizing these datasets already benefit from AI, there is still immense value in maintaining the domain, regional, and local expertise needed to interpret what the indicators actually mean.

Local context can vary even within a single country, and applying trained AI models across different regions of the world is even more challenging.

For example, imagine a model trained in the Horn of Africa, where diverse economies, widespread pastoral livelihoods, and years of consecutive drought shape current food security outcomes. If that same model were applied to countries in Latin America and the Caribbean – where livelihoods such as coffee farming or mining labor dominate, and drought is not currently a primary shock – how could we expect its outputs to remain accurate?

“Let’s say you manage to build a global model,” FEWS NET Principal Data Engineer Dave English added, “and suddenly you have something like the Ukraine war, where major producers of wheat and fertilizer stop exporting and prices skyrocket. How does that model account for a shock it has never seen before?”

Another example of why expert interpretation remains essential draws on lessons learned from FEWS NET’s experimental use of satellite imagery and trained data models to track livestock herd dynamics in East Africa. Across the region, livestock like cattle, sheep, goats remain key sources of both food and income for millions of people.

Motivated by the need to better understand herd size, movement patterns, and water access, FEWS NET analyzed images taken by satellites at the same time every day.  

“The problem here is, these satellite images are taken at noon when the sun is in the center of the sky,” English explained. “At that time of day, livestock like cattle are sheltering beneath trees for shade.”

Analysts reviewing the imagery would find no cattle simply because the animals were hidden under the tree canopy.

“This example illustrates that machine learning can be valuable, but there are cases where it just doesn’t work, and moments when it’s essential to rely on experts in specific domains and contexts to ensure accurate interpretation,” English noted.

When it comes to applying its scenario development process across diverse contexts to analyze food security conditions and forecast populations in need of assistance, FEWS NET depends on a wide range of workflows: local teams gathering and verifying information on the ground, regional scientists using advanced computing to track weather patterns and emerging shocks, data engineers leveraging AI to efficiently clean, standardize, and store disparate datasets, and much more.

AI plays an important and growing role, but it cannot replace the full breadth of expertise, judgment, and grounded understanding that these efforts require.

Trust and accountability in the use of AI

Global conversations about AI must center on accountability and responsibility, especially when the insights inform life-or-death decisions by governments and humanitarian agencies responding to emerging food crises. 

As FEWS NET reflects on its use of AI to date and looks toward the next frontier, understanding the challenges of responsible AI adoption remains a critical part of the discussion.

To frame this conversation about accountability and responsibility, it is first important to recognize the value of relationships and experience in gathering information for early warning systems.

“The experience we’ve accumulated over the years is a major advantage,” FEWS NET Deputy Chief of Party Tim Hoffine said. “FEWS NET must remain agile as the types of information available evolve. For example, we cannot rely too heavily on a single source, like price data from a government source that may eventually become unavailable. Systems that depend on a specific type of evidence will break if that source disappears, so we need both flexible systems and people who can interpret the data and build a complete picture.”

Forty years of experience has not only helped FEWS NET build that agility but also establish trust among a diverse array of stakeholders, all of whom must cooperate to provide the wide-ranging data needed for food security analysis.

“If a new model comes together, it can’t be immediately accountable or have trusted relationships with governments and humanitarian organizations,” Becker-Reshef added.

In the context of food security and humanitarian planning, trust and accountability are paramount. Decisions informed by early warning systems can be a matter of life or death – determining where aid is sent, when it arrives, and who receives it first. Stakeholders need confidence that the information guiding these decisions is accurate, validated, and interpreted by experts who understand the local and regional context.

Accountability also means that FEWS NET can explain how its analyses were produced, why certain assumptions were made, and what uncertainties remain. This transparency builds confidence among partners and allows for more effective collaboration when responding to crises.

Trust and accountability cannot be built overnight. They are the result of decades of consistent engagement with governments, local organizations, and humanitarian agencies, as well as the rigorous application of scientific methods. 

While AI can enhance efficiency and scale data processing, it cannot replace the relationships, credibility, and institutional knowledge that underpin responsible decision-making in this sector.

Charting the frontier of AI for early warning systems 

So, what’s next for the use of AI in famine early warning systems?

FEWS NET is already adopting LLMs to analyze thousands of historic reports and identify logical connections between the factors that contribute to food insecurity.

But the conversation is expanding far beyond text analysis, as the network explores how AI can meaningfully enhance forecasting while staying accountable and grounded in real-world expertise.

As noted above, AI is a moving target. The frontier of AI in the case of famine early warning combines advanced reasoning, causal-chain extraction, and knowledge graphs to better understand complex relationships between events, markets, and food security outcomes.

By structuring assumptions and historic data, FEWS NET can test models that detect shifts in local conditions and improve predictions of populations in need. 

To move in this direction, early work will include building causal graphs from reports, improving price and market modeling, and refining assumptions databases. These foundational efforts are designed to enable smarter, faster, and more accountable forecasting over time. 

The long-term vision is ambitious: a guided system for field analysts, where structured assumptions, historic data, and AI-driven reasoning combine to produce accurate, actionable forecasts. 

By continuing to invest in frontier AI, FEWS NET can maintain its leadership in early warning systems, staying ahead of emerging challenges while enhancing the speed, precision, and reliability of its analyses.

Sign up to receive FEWS NET updates directly in your inbox.