This background note informs the CFS High-Level Forum on Harnessing Artificial Intelligence (AI), Digitalization and Data Governance for Food Security and Nutrition. Building on prior work by the Food and Agriculture Organization of the United Nations (FAO), the World Bank, the United Nations Development Programme (UNDP), CGIAR and CFS itself, it focuses on what has been observed in deployment rather than future potential not yet supported by evidence and aims to provide a shared technical basis for the discussions at the Forum.

The term "AI" covers a spectrum of technologies, from narrow, task-specific machine learning (ML) to general-purpose generative systems. The note simplifies this into specific-purpose AI and generalist AI, since the two carry different infrastructure requirements, application risks and governance implications. Most proven advances are made by specific-purpose tools embedded in existing physical and institutional systems. Generalist AI offers complementary value, particularly as a human-languageinterface to data and expertise, but its deployments in high-stakes FSN functions remain limited and its benefits less consistently demonstrated.

The FSN dimensions whose impacts are the most documented are availability, through increased productivity, and stability thanks to better early warning systems. The effects on accessibility are very context specific. The effects on other dimensions like utilization, especially for nutrition are not well documented yet. Agency is the dimension most under threat because of unequal access to AI and because of the risks posed by potentially redelegating decisions to opaque algorithmic systems weakening accountability and contestability in food governance.

AI and digital tools amplify existing institutional and social capacity rather than substituting for it. Where that capacity is present, the right tools can accelerate progress; where it is absent, no digital tool alone can supply it. The World Bank's "4Cs" (Connectivity, Compute, Context and Competency) describe the cumulative preconditions for adoption, adaptation or innovation and many countries do not yet meet baseline conditions.

Three categories of risk are attached to deployment regardless of context: Operational risks, including hallucinations and a documented gap between vendor claims and observed benefits. Social and political risks, including algorithmic exclusion in areas of low contextual data availability (“data deserts”) that often coincide with areas of threatened food security (e.g. Sahel, Horn of Africa), reduced institutional agency and power asymmetries between multi-billion-dollar private firms and public sector institutions. Competition for resources, with projected AI data-center investment far exceeding the UN estimate to end world hunger and the large environmental footprint of AI. These risks compound on those most vulnerable and least likely to capture the benefits of adoption: smallholders, women and Indigenous Peoples.

Across these findings, a strategic lever is data. Technical infrastructure, however well-engineered, serves whoever controls the underlying data. Regulatory frameworks for FSN data exist in some countries but often lack international coordination and technical implementation in the form of dataspaces that operationalize sovereignty, fair benefit-sharing and contestability.

The note's way forward starts from the recognition that adoption is already underway, often ahead of the conditions that make it serve the public. This requires public and institutional capacity with digital public infrastructure that is publicly owned, democratically governed and built with those it serves and sustained investment in local competency and in the hybrid technical-policy expertise needed to specify, procure, audit and contest digital systems.