- Military
- History
- Security & Defense
- Artificial Intelligence
- Determining America's Role in the World
The wars in Ukraine and Gaza have shown the world that a new technology is now affecting the battlefield: artificial intelligence (AI). The Ukrainian and Russian armies are using AI to help locate and identify targets, pilot drones, and support tactical decision-making. Similarly, the Israel Defense Force has used AI in those roles to identify and locate Hamas leadership, command centers, and patterns. The introduction of AI to warfare has shown that the technology can and will increase the speed and scale of violence, resulting in effects that are both impressive and troubling. Integrating AI into warfare renders human oversight difficult in some cases, impossible in others, and in a sense, places the conduct of war into the hands of machines, intelligent or not. No longer confined to automating manual labor, machines are beginning to automate human reasoning. Understanding how this transition will, and should, fit into war is one of the defining problems of our time. However, introducing “thinking” machines into war is not that new. Militaries have been attempting it since World War II at least.
One of the most famous founding fathers of computer science, Alan Turing, became a household name by imagining how machines might think for humans and thus solve problems at greater scale and speed than humans can. The first pieces of this work were at Bletchley Park where Turing and his companions were working on breaking the German Enigma Code using new computers like the Bombe and the Colossus. Turing’s following work on computers as thinking machines made his name synonymous with the test for determining whether a machine had achieved consciousness or not. After Turing, both the British and the Americans continued work on integrating thinking machines into warfare. The U.S. Navy created a landmark system in 1958 called the Perceptron that was intended to learn tasks like a human. The New York Times reported that Perceptron was expected to learn to “walk, talk, see, write, reproduce itself, and be conscious of its existence,” in short order. As with most of the history of AI, the hype proved overly optimistic. Perceptron and other systems failed to deliver thinking machines in the 1950s and 1960s, sending the field of AI into its first “winter.”
However, the concurrent revolution provided by innovations in microchips still led to the introduction of microprocessors into military weaponry. This resulted in precision guided munitions that could hit targets on their own once fired but also in systems that could fire automatically. Many of those systems are still with us like the Navy’s Close-In Weapon System (CIWS) and smart anti-ship mines, and the Army’s Patriot Air Defense System. All can engage targets on their own without further human intervention once activated if allowed to do so, meeting the U.S. military definition of autonomy. None of these systems has any AI in them, in a modern sense, and no one would accuse CIWS or Patriot of “thinking.” Further, none of these systems has caused global concern over the last few decades even when they are responsible for catastrophic mistakes. Something in our perception of the technology has changed. Capabilities we now refer to as AI in military systems seem much more promising in their effects, and much more troubling.
Modern AI technology promises powerful improvements to any military fighting wars. For most of the history of war, technology has done some level of automating human labor. Whether throwing things, digging things, building things, or smashing things, machines were mostly used for labor. With the advent of computers, the idea of automating human cognitive labor seemed within reach. Whether PowerPoint is in fact a labor-saving device remains controversial but certainly transmission and storing of military plans and products with computers was helpful, and the value of precision munitions is without dispute. In the last ten years, the introduction of AI into software seemed to promise something new: the automation of reasoning.
Over the last several decades of war, the availability and sheer amount of warfighting information increased the cognitive load on military personnel beyond their ability to accomplish all that could be accomplished. As a result, the U.S. military’s first AI project in 2017, Project Maven, formed to help intelligence analysts sift through all the imagery and video gathered by thousands of sensors. This seemingly simple project, identify objects in digital files, was poised to transform the speed, precision, and resilience of military operations across the force by reducing cognitive load through the automation of low-level human reasoning. As Project Maven churned through this effort, new technologies began to appear that promised even more.
The generative AI revolution in 2022 supercharged discussions of what AI could do. Suddenly simple machine reasoning, or what Andrew Ng described as automating what would take a human about one second to process, became a low bar even though no one is able to jump over it consistently. New Large Language Models (LLMs) seem to perform very advanced, human-like reasoning and suddenly the idea of using machines to help humans make better military decisions, or even make military decisions on their own, appears within reach. These assumptions led to the development of several types of systems that are currently under development in most advanced militaries.
Most AI-enabled military systems use one of a few distinct but related technologies. First, object classification, sometimes colloquially referred to as computer vision, is used to identify things. This is the basis for the military capability called Automatic Target Recognition (ATR). ATR is the well-spring from which almost every AI-enabled system flows. Most military problems begin with being able to find and confirm a target faster and more reliably than your enemy. From drones like the Switchblade to targeting software like Gospel, an AI model is being used to find and identify the needle in the haystack of data.
Second, insight generation from large data sets is where AI models identify patterns that humans cannot see because the size of the data set means it would take too long for a human to make the same connection, if ever. Insight generation is where LLMs get their real mojo from as well, but what makes them so powerful is that the interface is in natural language. You don’t need a spreadsheet, you just ask a question and the model generates the insight. There is some overlap here but at its simplest any LLM is about making it easier for humans to access complex models without expertise to gain those insights and pattern recognition. This feels like interacting with an intelligence, but it is not.
This type of technology is very prominent in decision support systems (DSS) and, coupled with ATR, can make human decision-making faster and possibly higher quality. The ultimate prize in an AI-enabled DSS is to achieve the Move 37 effect which refers to the AI Model AlphaGo that defeated the Go player Lee Sodol in 2016 by using a game strategy never before seen by humans and therefore unanticipated and difficult to counter. This is catnip for military leaders and the ultimate decision advantage in war.
Third, learning behavior is where an AI model learns how to behave in ways that help complete a task. Object avoidance and target tracking in drones are good examples. The average human cannot hope to fly a drone through a forest at a target but with object avoidance and target tracking, the AI helps smooth out your mistakes so you can focus on your target and not become distracted by avoiding trees.
Fourth is a combination of many technologies resulting in some type of mission autonomy. An autonomous system would likely never make use of one AI technology. Instead, it would be a few AI technologies like object avoidance and ATR in a drone that allow the system to act without human supervision. While difficult to define, autonomy in some forms is making its appearance on the battlefield.
AI-enabled autonomy can be described in three forms. Closed loop systems are those that are a single weapons system that performs its own functions. A lethal autonomous drone is an example. It navigates itself, finds the target, computes a way to engage the target, decides to engage, shoots, then assesses its actions. This type of system is desirable because it allows for speed and resilience of mission, but it is very hard to get right.
A more likely type of AI-enabled autonomy is the distributed system. These systems are a patchwork of many systems. Perhaps a remotely piloted drone has ATR on it that can send target info to an AI-enabled DSS in a command center which determines the best allocation of resources for engaging thousands of targets and after deciding which weapon is the best shooter, sends the firing order to a missile battery that uses AI to decide which is the safest position to fire from, providing its human operators with direction and guidance on where to move and when to fire.
A distributed system has humans involved, and humans can be the chokepoint preventing decisive speed to defeat the enemy. One way to remove the slow humans is through agentic warfare. An AI agent is a DSS that can stack complex series of tasks to perform several missions in support of a higher human intent. In the above scenarios the AI-enabled DSS could be an AI agent that has been told to establish air superiority and then constructs a task-stack to achieve that goal, controlling many systems along the way.
If this all sounds unsettling, good. It is. There are some problems with AI technology that prevent these systems from being the flawless partners in war that we might want. AI technology, of any currently fielded variety, is a probabilistic tool. Meaning everything it outputs is based on a probability of being the desired output. Traditional software is deterministic which means that you enter the same input 100 times, you will get the same output 100 times. Not so with probabilistic software like AI.
AI learns from a set of training data how to give you the output you want. This is a strength in that it can learn complex things like how to pick out a dog from a cat. But it has a problem called an asymptote. The model will never be 100 percent certain that its output is correct. And the more the input varies from the training data, the more likely the output will be incorrect.
War is incredibly complex and there are many instances where real world input is far from the AI’s training data. This renders the model unreliable or makes it reliable in very narrow circumstances. Having to employ an AI in narrow contexts or face severe reliability issues means that humans must retain control to an extent that renders the machine more a hindrance than an asset. This is why drones in Ukraine remain piloted by humans rather than AI.
The promise of AI-enabled systems in war is so promising though that no military will forego developing these systems. So, what is to be done? The U.S. military has initiated a policy that lethal autonomous systems, powered by AI, must reflect appropriate levels of human judgment. The international community insists that any AI-enabled weapon must be under meaningful human control. But what is meaningful control, and when does meaningful control sacrifice meaningful autonomy to the point where having the system is no longer useful? When must judgment or control be applied? Continuously? Only when conducting a strike? When the system is being manufactured? There are no clear answers to these questions and it is likely that militaries will be forced to make hard choices based on the severity of their security situation. In the meantime, the machines are learning, and war is growing faster and ever more frightening.
