Using AI in agriculture could boost global food security – but we want to anticipate the risks

As the worldwide population has expanded over time, agricultural modernisation has been humanity’s prevailing approach to staving off famine.

Quite a lot of mechanical and chemical innovations delivered through the Fifties and Nineteen Sixties represented the third agricultural revolution. The adoption of pesticides, fertilisers and high-yield crop breeds, amongst other measures, transformed agriculture and ensured a secure food supply for a lot of tens of millions of individuals over several a long time.

Concurrently, modern agriculture has emerged as a wrongdoer of worldwide warming, accountable for one-third of greenhouse gas emissions, namely carbon dioxide and methane.

Meanwhile, inflation on the worth of food is reaching an all-time high, while malnutrition is rising dramatically. Today, an estimated two billion people are afflicted by food insecurity (where accessing secure, sufficient and nutrient-rich food isn’t guaranteed). Some 690 million persons are undernourished.

The third agricultural revolution can have run its course. And as we seek for innovation to usher in a fourth agricultural revolution with urgency, all eyes are on artificial intelligence (AI).

AI, which has advanced rapidly over the past 20 years, encompasses a broad range of technologies able to performing human-like cognitive processes, reminiscent of reasoning. It’s trained to make these decisions based on information from vast amounts of knowledge.

Using AI in agriculture

In assisting humans in fields and factories, AI may process, synthesise and analyse large amounts of knowledge steadily and ceaselessly. It can outperform humans in detecting and diagnosing anomalies, reminiscent of plant diseases, and making predictions including about yield and weather.

Across several agricultural tasks, AI may relieve growers from labour entirely, automating tilling (preparing the soil), planting, fertilising, monitoring and harvesting.

Algorithms already regulate drip-irrigation grids, command fleets of topsoil-monitoring robots, and supervise weed-detecting rovers, self-driving tractors and mix harvesters. A fascination with the prospects of AI creates incentives to delegate it with further agency and autonomy.

This technology is hailed because the solution to revolutionise agriculture. The World Economic Forum, a global nonprofit promoting public-private partnerships, has set AI and AI-powered agricultural robots (called “agbots”) on the forefront of the fourth agricultural revolution.

But in deploying AI swiftly and widely, we may increase agricultural productivity on the expense of safety. In our recent paper published in Nature Machine Intelligence, we have now considered the risks that would include rolling out these advanced and autonomous technologies in agriculture.

From hackers to accidents

First, given these technologies are connected to the web, criminals may attempt to hack them.

Disrupting certain forms of agbots would cause hefty damages. In the US alone, soil erosion costs US$44 billion (£33.6 billion) annually. This has been a growing driver of the demand for precision agriculture, including swarm robotics, that may also help farms to administer and lessen its effects. But these swarms of topsoil-monitoring robots depend on interconnected computer networks and hence are vulnerable to cyber-sabotage and shutdown.

Similarly, tampering with weed-detecting rovers would let weeds loose at a considerable cost. We may also see interference with sprayers, autonomous drones or robotic harvesters, any of which could cripple cropping operations.

Beyond the farm gate, with increasing digitisation and automation, entire agrifood supply chains are prone to malicious cyber-attacks. At least 40 malware and ransomware attacks targeting food manufacturers, processors and packagers were registered within the US in 2021. The most notable was the US$11 million ransomware attack against the world’s largest meatpacker, JBS.

Then there are accidental risks. Before a rover is distributed into the sphere, it’s instructed by its human operator to sense certain parameters and detect particular anomalies, reminiscent of plant pests. It disregards, whether by its own mechanical limitations or by command, all other aspects.

The same applies to wireless sensor networks deployed in farms, designed to note and act on particular parameters, for instance, soil nitrogen content. By imprudent design, these autonomous systems might prioritise short-term crop productivity over long-term ecological integrity. To increase yields, they may apply excessive herbicides, pesticides and fertilisers to fields, which could have harmful effects on soil and waterways.

Rovers and sensor networks might also malfunction, as machines occasionally do, sending commands based on erroneous data to sprayers and agrochemical dispensers. And there’s the chance we could see human error in programming the machines.

Safety over speed

Agriculture is simply too vital a site for us to permit hasty deployment of potent but insufficiently supervised and sometimes experimental technologies. If we do, the result could also be that they intensify harvests but undermine ecosystems. As we emphasise in our paper, probably the most effective method to treat risks is prediction and prevention.

We must be careful in how we design AI for agricultural use and will involve experts from different fields in the method. For example, applied ecologists could advise on possible unintended environmental consequences of agricultural AI, reminiscent of nutrient exhaustion of topsoil, or excessive use of nitrogen and phosphorus fertilisers.

Also, hardware and software prototypes must be fastidiously tested in supervised environments (called “digital sandboxes”) before they’re deployed more widely. In these spaces, ethical hackers, also referred to as white hackers, could search for vulnerabilities in safety and security.

This precautionary approach may barely decelerate the diffusion of AI. Yet it should be sure that those machines that graduate the sandbox are sufficiently sensitive, secure and secure. Half a billion farms, global food security and a fourth agricultural revolution hang within the balance.