AI & The Future of Food Security: Using AI to Build Food System Resilience and Opportunity in a Changing Climate

In collaboration with Google DeepMind, Gro Intelligence hosted an in-person roundtable in New York City in September 2023 to explore the ability of artificial intelligence systems to help increase global food security. The event was attended by policymakers, academics, business executives, and AI practitioners. Below we walk through the roundtable discussion insights that emerged around how AI can help businesses, governments and NGOs develop and maintain more sustainable, resilient, and crisis-proof food systems in the face of an increasingly changed climate.

Challenges and Opportunities

The world is experiencing huge growth in the capabilities and applications of AI systems, particularly machine learning (ML) and large language models (LLMs). At the same time, our global food system appears increasingly susceptible to political and environmental impacts, including those related to anthropogenic climate change. The past 18 months have been tragically illustrative in this respect, as Russia’s invasion of Ukraine, pre-existing global price pressures and simultaneous, ten-year high drought conditions in several of the world’s major grain-producing regions combined with devastating effect. According to the UN Food and Agriculture Organization, between 691 million and 783 million people were hungry globally in 2022, with some 180 million more people facing severe food insecurity than the year before.

What is food security? Food security describes the ability of people, households, or communities to reliably access a sufficient quantity of affordable and nutritious food that meets their dietary and cultural needs for a healthy and active life. It is typically described and measured (as is its conceptual opposite, food insecurity) across four intertwined dimensions:

• physical availability of food supply; 
• economic access to available food; 
• utilization, a measure of the safety and quality of available food such that can be adequately metabolized and used by the body; and 
• the stability over time of such availability, access, and utilization. 

According to the UN (FAO, 2022) the ‘major drivers’ of food insecurity are: armed conflict; climate extremes; economic shocks; and socio-economic inequality. Within a population, low levels of education, weak social networks, limited social capital, low household income, and being unemployed are the factors most strongly associated with the likelihood of experiencing food insecurity (USDA ERS, 2010).

Driven globally by a web of intertwined complexities – armed conflict, climate extremes, economic shocks, and socio-economic inequality – and locally by low levels of education, weak social networks, limited social capital, low household income and unemployment, food insecurity is considered by many to be a quintessentially ‘wicked’ problem.  

Defined in 1973 by UC Berkeley design professors Horst Rittel and Melvin Webber, the term refers to real-world, societal problems that are difficult, if not impossible, to definitely solve because of incomplete, contradictory or changing requirements that are often difficult to recognize. Complicating matters, wicked problems are all essentially unique, can be described as the symptom of other problems and are susceptible to multiple approaches and possible solutions. Wicked problems, according to Rittel and Webber, must be solved not once but many times as real-world conditions change and evolve.

Viewed in this context, our crisis of rising, global food insecurity is a critical humanitarian and national security challenge that AI is ideally, if not uniquely among information technologies, poised to help solve.

With its ability to analyze large and complex datasets, to identify patterns and trends that would be difficult or impossible for humans to identify on their own, and to simulate complex, networked cause and effect scenarios at scale, AI systems are already being used in the food security space for data analysis, system monitoring and predictive modeling. Typically centering on the physical availability of food as well as the market economics of national and global food supplies, examples of current food security-related AI use cases include: predicting crop yields and supply; forecasting demand in the short, medium and long-term; assessing potential supply chain disruptions; analyzing the impact of climate change on agricultural production; and optimizing crop yields at the farm level.

But there is real potential for AI – particularly predictive machine learning and large language models – to do much more in the next five to ten years to help us materially increase the economic and environmental resilience of the global food system.  

What is a ‘digital twin’? A digital twin is “an integrated multiphysics, multiscale, probabilistic simulation of an as-built system . . . that uses the best available models, sensor information, and input data to mirror and predict activities/performance over the life of its corresponding physical twin.” (US DoD Digital Engineering Strategy)

Opportunities for such advancement, several of which could be enabled by AI-powered food systems and weather “digital twins,” include the development of better in-season or ‘tactical’ decision-making tools such as:

• Adaptive systems for monitoring and managing logistics and supply chain networks;
• Predictive midterm to seasonal weather forecast models;
• Automated early warning systems for extreme weather;
• LLM-enabled technical and non-technical end user engagement with, and understanding of, complex food system data sets;

as well as more capable tools supporting strategic policy action and local, regional and national resilience planning such as:

• Supply chain impact assessment (climate change, price shocks, conflict) and scenarios planning for resilience;
• Food supply system optimization to maximize useable customer delivery and minimize waste;
• Development and maintenance of emissions and impact mitigation.

With this collective understanding, participants were asked to consider structural issues they understand to be important, if not critical, for enabling the full potential of AI to help solve global food security challenges.

In anticipation of the roundtable discussion, AI and food security experts at Gro Intelligence identified three key questions to help guide the conversation: 

Data: In order to unlock the potential of AI for Food Security, we need data - and lots of it. What is the most critical missing data (type; category; source), and what actions are needed to ensure its equitable availability and accuracy?

AI Governance: Access to safe and adequate food is a basic human right. How can we ensure that AI deployed to improve food systems is responsibly and transparently managed?

Equity: The viability of powerful AI applications may ultimately rest on popular and political support for the ethical use of what are effectively massive aggregations of personal or private data. In the context of food system applications, what is the most effective way to build and maintain social license for such data use?

Roundtable Insights

Conducted under the Chatham House Rule with Gro Intelligence CEO Sara Menker moderating, the roundtable converged on three main themes:

• AI is Uniquely Poised to Help Increase Global Food Security
• Data and Collaboration are Paramount
• Focus on Access is Essential for Equitable AI Outcomes

AI is Uniquely Poised to Help Increase Global Food Security

Consistent with the idea that food security can be seen as one of modern society’s ‘wicked’ problems, several participants spoke to both the nature of the food security challenge and the potential for AI to help human decision makers to meet that challenge:  “Food is as complex as it gets from the biological processes of how it's grown to the way in which it makes its way through the markets.” 

Accepting that, as one participant said “AI is only a tool, not a magic bullet,” the roundtable recognized and agreed that AI is uniquely suited to help humans monitor and manage complex, dynamic systems and that food security is just such a system - a “multilayered dynamic involving weather patterns, individual farming decisions, incentives, global logistics and transport.”

Two potential applications of AI in the food security space were discussed at some length: the ability of generative models to increase and democratize access to food security resources and expertise, and the ability of AI to support the development of early warning detection systems for environmental, market, and systemic food security challenges.

Capabilities enabling the first application are moving forward with remarkable speed, as the power and functionality of generative language models rapidly advance. One roundtable participant described the potential for cost-effectively improving the subject-matter accuracy and responsiveness of large ‘horizontal’ generative AI models by pairing them with ‘vertical’ expert food system knowledge analytical and data hierarchies. Doing so could revolutionize access to complex and dynamic food security data, as LLM-powered user interfaces would allow nonexperts to conversationally pose questions, the answers to which require complex statistical data analysis, with the ability to ask in a chat thread for context-driven clarification, explanation, and recommendations. Another participant with knowledge of agricultural policies in southern Asia highlighted the potential for AI to help small scale farmers apply for subsidies. Currently such farmers often face language or literacy and knowledge barriers when faced with the need to read and complete long and varied legal documents which ultimately prevent them from successfully applying for available government funding. Natural language and LLM AI could help address this challenge by enabling small-hold farmers to understand and complete such applications.

Participants similarly recognized the potential for AI to power new and valuable early warning and predictive food system models, including those powered by digital twins of agrifood systems and subsystems. Rapid advances in AI-specific computing systems and analytic tools will allow decision makers to iteratively model system dynamics at various geographical and systemic resolutions to explore, test, and prepare for a range of scenarios and potential futures. There was general agreement that our ability to create such AI-powered monitoring and management systems would enable the development and implementation of more effective policies and processes for mitigation and adaptation to risks. Given the acknowledged complexity of the global food system and the socio-political and environmental dynamics of food security, however, one participant cautioned that in developing and deploying AI tools in the food security space, we must continually ask ourselves “what problem are we trying to solve for?”

Data and Collaboration are Paramount

Together with the discussion on the potential for using AI tools to improve agrifood system performance and food security decision making, the roundtable focused on the criticality of, and challenges surrounding, access to consistent and reliable global food security data.

Acknowledging the fundamental need for comprehensive and reliable data to train AI models, participants grappled with the question of what needs to happen to make such data the rule, rather than the exception for food security. 

While LLMs can be successfully trained on a seemingly infinite and openly accessible ‘internet of text and images,’ no such data resource exists for training food systems and food security models. In addition to challenges related to language, format and frequency, comprehensive food system data, when available, is often commercially sensitive, competitive or considered a matter of national interest, if not national security. “At the subnational level, the quality of data is variable, the kind of data that we have is variable, it's not shared.” said one participant, “Countries are not comfortable sharing food security data because there is a lot of sensitivity around it.” Such barriers reflect the socio-economic complexity of food systems and, as another participant highlighted, make it challenging to get to the “ground truth” about food system conditions.

Understanding the role that the decades-old Protein Data Bank collaborative had in  supporting and enabling the recent development of the AlphaFold deep-learning protein structure prediction software, several participants identified the potential for an analogous food security ‘data commons’ to enable AI-powered advancements for food systems.  The power of generative AI to help fill data gaps (e.g., from primary earth observation sensors) to create valuable new global training data sets was also discussed.

There was consensus that there is no easy fix to such data-centric issues, but that there is “a big need for data investment.” Not only is food security data often sensitive, it is frequently expensive to create, collect and harmonize. Unfortunately, it was not immediately clear to participants who – beyond governments or large market incumbents – has the economic means and incentive to pay for such data generation in this context.  But there was no question regarding the need for such data generation, and that it is unlikely to occur without sustained and genuine cooperation across geographies and across the spectrum of food system participants.

Focus on Access is Essential for Equitable AI Outcomes

As a natural corollary to the roundtable’s discussion regarding data access and availability, the group also focused on the overall economics of AI and its implications for equity of access to the benefits that current and future AI food system tools offer.  

Participants recognized the need to address information asymmetries in the global food security and food system marketplaces so that food system power does not become further centralized in wealthier developed jurisdictions and the benefits of new and existing AI capabilities are distributed equitably. According to several participants, a focus on equity is critical as impacts to the global food system already fall most heavily on net food importing countries in the Global South and with a changed and changing climate, “this is a situation that's only going to get more difficult - the crisis of trust is going to get amplified.”

The discussion prompted one participant to ask how we might “create the right market to scale solutions to lower costs for participating in the use of AI?”

While no panacea, AI was posed as one potentially important solution. Looping back to its earlier discussion on the potential value AI can offer, the roundtable discussed the potential for generative, LLM-based applications to foster a new and transformational democratization of food system information access and use. AI offers the potential for real-time, natural language sharing at scale – across geographies, languages, and education levels – including complex, system-level food system information and analytics that previously have only been meaningfully accessible to a relatively small group of subject-matter experts.

Unlocking AI’s Food Security Potential

While the ‘wicked’ problem of food insecurity, and the questions posed to the roundtable participants, are far beyond solving or answering within the scope of a single roundtable discussion, there was consensus among the participants on several important points:

Advances in AI have the potential to improve the productivity, resiliency, and equity of food systems, including by improved monitoring and adaptive management and by simplifying and delivering decision support in natural human modalities like language. Such AI systems will require sustained collaboration and investment in terms of data generation and sharing, and shared governance of those data and systems will be necessary to ensure the informational advantages provided by these AI systems also benefit those most at risk of food insecurity.