Can an Algorithm Feed the World? AI in Agriculture

A model was trained on 54,306 images of healthy and diseased plant leaves, covering 14 crops and 26 diseases, and reached 99.35% accuracy on images held back for testing. It was then tried on 121 images from other sources taken in different conditions, and its accuracy fell to 31.4% (random guessing would give about 2.6%). This study, by researchers from Switzerland's EPFL and Pennsylvania State University (2016), is the most honest way to open a discussion of AI in agriculture: the tool is strong in its own conditions, and needs testing in the field.

A problem of production or of distribution?

An estimated 673 million people (8.2 per cent of the world's population) faced hunger in 2024, down from 8.5 per cent in 2023, according to the report The State of Food Security and Nutrition in the World 2025. The picture varies, however: in Africa the share rose above 20 per cent, and in Western Asia it reached about 12.7 per cent.

In its 2012 revision, the FAO estimates that the world's population will reach about 9.15 billion by 2050, and that agricultural production and consumption will be about 60 per cent higher than in 2005–2007. These are projections that depend on assumptions, not decreed fates.

Loss, for its part, is well documented: the FAO estimated that about 13.3 per cent of food was lost in 2023 between production and the stage before retail, which means that part of the solution is to reduce waste, not only to grow more. Here the proper place of AI appears: it is a tool for improvement, not a substitute for fairness in distribution.

What does the algorithm do in the field?

Satellites. In 2017 researchers at Stanford University published a method that estimates soybean yield at county level across the United States from publicly available remote-sensing data, and reported that it outperforms competing methods. This points to the possibility of estimating a harvest before it is gathered where field surveys are scarce; it does not necessarily mean the results transfer unchanged to every country and crop.

Early diagnosis. Rather than waiting for a pest to appear, models try to detect plant stress or its first signs from a photograph. As we saw at the start, laboratory accuracy does not guarantee field accuracy, so such applications need calibration on local images and the supervision of an agricultural adviser.

Irrigation. Water can be directed according to the needs of each plot, but the saving is measured farm by farm, so no single percentage should be generalised.

Efficiency for whom?

The legitimate fear is that the benefit will go first to large farms equipped with capital and digital infrastructure, and that the smallholder will turn from a person with knowledge of his land into a tenant of an algorithm whose logic he does not know, if soil and climate data become concentrated with a commercial intermediary. This is a risk assessment, not a description of documented reality, but it justifies asking before buying: who owns the data? And can the farmer understand why the system recommended what it did? (See the article "Algorithmic Bias" on the effect of a group being absent from the data.)

And even the most accurate algorithms do not repair a flaw in markets and policies: improving the supply chain may carry food from field to shelf efficiently, but it does not answer why someone who needs it remains unable to buy it.

The arid environment: our own case

Data from the Ministry of Environment, Water and Agriculture (Statistical Book 2019, as reported by Fanack) show that about 72% of water demand in the Kingdom went to agriculture, and a study cited by the site (2019) indicates that about 90% of agricultural water is groundwater, including aquifers that are non-renewable in some regions. In such a country precision irrigation and early warning take on special importance, and models trained in other climates are expected to need recalibration before adoption, by the same lesson with which we began. Universities and research centres have an opportunity here to build datasets and image collections from our hot, dry environments.

What to do

• Agricultural data the farmer has the right to access, not held by an intermediary.
• Testing the tool in the local field before relying on it, and asking for its accuracy outside the conditions of its training.
• Explainability, so that the farmer understands the reason for a recommendation.
• Training farmers and advisers to read the outputs and question them.
• Caution against a single profitable crop that impoverishes biodiversity.

It is not the algorithm that feeds people, for provision is from God, and ours is to take the means, to be grateful for the blessing and to guard it from waste. If we use these tools well, it will be to the extent that they conserve water, reduce waste and reach smallholders, not to the extent that they dazzle.