AI in food systems: outcomes also depend on who controls the data

Predicting an agricultural yield and improving a diet are different outcomes. Treating them as interchangeable evidence for artificial intelligence makes it difficult to establish where adoption is producing value. That distinction is central to a contribution by Elisabetta Gotor and Ilaria Urbani, published on 20 October by the Alliance of Bioversity International and CIAT in the context of CFS 54 in Rome.

The authors describe stronger evidence for specific applications such as yield prediction, pest and disease identification, and early-warning systems. They consider evidence on food access, nutrition and diets more limited. The blog does not present a new quantitative synthesis: this assessment should be attributed to its authors rather than presented as a new experimental finding.

Different uses require different questions

The argument has five themes, summarised in the diagram. Alongside evidence, it considers local skills and infrastructure, data control, AI’s material costs and human judgement. A system may work on a defined task yet offer little value if connectivity, people able to assess it or institutions accountable for errors are missing. Context therefore becomes part of the outcome’s definition.

Data governance concerns who collects information, who can use it and who benefits from the conclusions. The contribution highlights under-representation and the concentration of infrastructure and algorithms, without quantifying their shares. Its discussion of energy, water and land is also qualitative. The five themes are therefore not a score, a ranking of tools or an environmental-impact calculation.

A reading for businesses

For a nutraceutical business, one possible application is to define the task before buying a system: screening literature, extracting document information and supporting decisions need different assessment criteria. This is GMA’s analysis, rather than a list of uses tested by the forum. Useful evaluation should connect sources, processing steps and corrections with a relevant outcome, such as extraction quality or review time.

The account highlights discussion priorities, including impact assessment before public procurement and local auditing capacity. It does not create new legal requirements. Its contribution this week is a practical adoption question: who can scrutinise the result, and with which data? Answering requires a verifiable process as well as an available tool.

Editorial note. This content is intended for industry professionals. It does not constitute medical advice, therapeutic guidance or regulatory advice. Read our editorial method.