Two vets, two thousand images and a hoax to look out for. The future of the supply chain

What can artificial intelligence learn from a mozzarella? The answer is simple yet far from trivial: before learning how to do something, one must first learn to understand what matters. And to grasp that, almost always, one needs help from those who know the job inside out. The proof comes from a place that, in the collective imagination, would seem a world away from Silicon Valley: a buffalo farm in Campania.

A group of researchers from the University of Naples Federico II has developed a system that uses a thermal camera, connected to a milking robot, to monitor the condition of buffalo udders and identify signs that can help with the early detection of potential health problems. To train the system, 2,148 thermal images were collected. The most interesting detail, however, concerns the work carried out before the algorithm became capable of recognising anything: two vets examined the images and indicated, one by one, the areas to be observed. They taught the machine where to look.

It may seem like a technical detail, but it is a small lesson in politics, economics and culture. We have become accustomed to discussing AI as if its destiny were to replace human expertise: the doctor, the journalist, the professor, the craftsman. The history of fake news reminds us that expertise, even before it might be automated, must first be understood. Without the expertise of vets, images remain just images. Without a clear definition of the problem to be solved, computing power risks being nothing more than wasted power.

This does not mean that artificial intelligence is destined to remain a mere obedient assistant. It means something more concrete: its value increases when it comes into contact with the knowledge accumulated within the professions. It can regularly measure what a human being cannot monitor continuously, identify anomalies and provide information to enable earlier intervention. However, the system developed in Naples does not, for the time being, formulate a diagnosis on its own: it provides information to help the vet make a decision. This distinction is not a limitation to be concealed; it is the starting point.

The same principle applies to the cheesemaker who can interpret a production process, to the farmer who recognises a change in an animal’s behaviour, and to anyone who can distinguish a correct figure from a sound decision. It applies to a newspaper, too. It is not enough simply to ask AI to write: you must know how to teach it what constitutes newsworthy content, which details matter, and where to look for errors.

From 28 to 30 October, in Sorrento, the World Buffalo Congress will bring together scientists and businesses to discuss the future of the buffalo industry. Amidst sensors, genetics and predictive systems, the most interesting message might well be this: true innovation does not begin when the machine replaces the expert, but when the expert can teach the machine something useful. Even to produce better mozzarella, one must first learn how to observe a buffalo.