Introduction: This article is translated from the War On The Rocks article “AI AT WAR” by Anthony King. April 27, 2023. The content of the article does not represent the translator's views.
There is a widespread belief that the world is on the brink of another military revolution. Artificial Intelligence (AI) is about to change the nature of warfare, just as gunpowder, tanks, airplanes, and atomic bombs did before. Today, many countries are proactively leveraging AI to gain military advantages. For example, China has announced plans to be a world leader in AI by 2030. China's “New Generation Artificial Intelligence Development Plan” claims: "Artificial intelligence is a strategic technology leading the future." Russian President Vladimir Putin has made similar statements: "Whoever becomes the leader in this area will become the ruler of the world." In response to the challenges posed by China and Russia, the United States has indicated it will implement the "Third Offset Strategy." The U.S. will invest heavily in AI, automation, and robotics to maintain its advantage in defense.
In light of these eye-catching developments, military commentators have developed a keen interest in the military applications of AI. For instance, in a recently published monograph, Ben Buchanan and Andrew Imrie claim that the emergence of AI is comparable to humanity's initial use of fire. AI-controlled automated weapons will become increasingly precise, rapid, and lethal. They represent the future of warfare. Many other scholars and experts agree with their views. For example, prominent computer scientist and AI pioneer Stuart Russell dedicated a special episode in his 2020 BBC Reith Lectures (BBC Reith Lectures) to the military potential of AI. He openly stated: slaughter machines and killer robots are on the rise. He described a scenario involving a deadly quadcopter the size of a jar, equipped with explosive devices: "Personnel-killing landmines can eliminate all males aged 16 to 60 in a city or all Jewish citizens within Israel, and unlike nuclear weapons, it won't affect urban infrastructure." Russell concluded: "There will be 8 million people wondering why you can't protect them from the robots' pursuit." Many other scholars also share Russell's views, including Christian Brose, Ken Payne, John Arquilla, David Hambling, and John Antal; they believe that with the development of second-generation AI, lethal autonomous weapons, such as killer drone swarms, may soon emerge.
The military revolution may not be as intense as its proponents envision at the outset. The military affairs revolution of the 1990s was indeed significant, bringing new possibilities to warfare, but it did not eliminate uncertainty. Similarly, the debate surrounding lethal automation and AI is often exaggerated. These discussions distort how AI currently operates and, in turn, misrepresent its potential impact on military operations in the foreseeable future. Despite the increasing importance of remote automated systems, the likelihood of replacing armies with automated drone swarms on the battlefield remains low, and it is also challenging to replace human commanders with supercomputers. AI became a major research topic in the 1950s. At that time, its operation was based on symbolic logic—programmers would encode inputs for AI to process. Such systems are also referred to as classical AI. AI made some progress, but its utility was very limited due to its operation based on manipulating specified symbols, especially in the real world. This led to an "AI winter" from the late 1970s until the end of the 1980s.
Since the late 1990s, second-generation AI, based on big data, massive computing power, and algorithms, has achieved some remarkable breakthroughs, including three significant events. On May 11, 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov. In 2011, IBM's Watson won on “Jeopardy!” Then came an even more remarkable achievement: in March 2016, AlphaGo defeated world Go champion Lee Sedol 4-1.
Deep Blue, Watson, and AlphaGo are all important milestones on this remarkable path. However, in just twenty years, AI transitioned from disappointing failures to unexpected victories. However, it is crucial to recognize the capabilities and limitations of second-generation AI. The development of second-generation AI revolves around a concept: neural networks. Machine learning programs can process vast amounts of data through their networks, continuously adjusting the weights assigned to specific data segments, ultimately generating logical answers. These systems are probabilistic and inductive. Programs and algorithms know nothing. From a human perspective, they do not understand the real world or the meaning of the data they process. Machine learning AI merely constructs statistical probability models through extensive trial and error. In this way, second-generation AI can identify various correlations within the data. As long as there is enough data, probabilistic induction can become a powerful predictive tool. Yes, AI cannot discern causal relationships or intent. Noted Silicon Valley entrepreneur Peter Thiel compellingly articulated the limitations of AI: "Forget those sci-fi fantasies; the real power of today's AI lies in its application to relatively simple, monotonous tasks, such as computer vision and data analysis." Thus, although machine learning performs far better than humans in some limited, constrained, and mathematically representable tasks, it remains weak. It entirely relies on the data used for its training, and even the slightest changes in the actual environment (i.e., data) can render it completely ineffective.
For the prospects of an AI military revolution, the vulnerability of data-driven inductive machine learning in real environments is a significant issue. Both AI supporters and opponents suggest that in the near future, autonomous drones will be able to fly, identify, and attack targets quite easily in urban environments. After all, autonomous drone swarms have been demonstrated—but that was in recognized, controlled environments. However, training drones to operate autonomously in ground combat is extremely challenging. The battlefield environment is highly dynamic and complex, especially in urban settings where civilians and military personnel are mixed. Currently, there seems to be no reliable data available for training drone swarms—battlefield conditions are simply too variable. Similarly, it is also difficult to understand how algorithms make command decisions. Command decisions require parsing different structured information and balancing political and military factors; all of this requires judgment. Avi Goldfarb and Jon R. Lindsay argue in a recent article that if there were perfect data, then data and AI would be the best choice for making simple decisions. It can be said that military command decisions inherently encompass complexity and uncertainty. It is worth noting that although Google and Amazon are both outstanding data companies, their managers have never envisioned using algorithms to replace them in making strategic and operational decisions. Data, after being rapidly processed by algorithms, can help company executives better understand the market—the depth and granularity of which are beyond what their competitors can achieve. Information advantage has helped them achieve dominance. However, machine learning has not replaced management functions.
Therefore, in the near future, it is highly unlikely that AI-controlled lethal autonomous drones or killer robots will take over the battlefield. Computers or supercomputers are also unlikely to replace commanders. However, this does not mean that AI, data, and machine learning are unimportant for contemporary and future military operations. They are very important. However, the primary function of AI and data is not lethal—they do not resemble how humans learned to use fire. Data is digital information stored in cyberspace, and its importance lies in enabling national governments to understand themselves and their adversaries more broadly, deeply, and credibly. When effectively processed using AI, military commanders can perceive the battlefield space with unprecedented depth, speed, and resolution. Data and AI are also crucial for cyber military operations and information warfare. They have become indispensable parts of defense and offense. While AI and data may not yet represent humanity's new fire in terms of utilizing cyberspace for digital military intelligence, they can serve as a significant new source of information. AI can help us “see the other side of the hill”; it is revolutionary in this regard. For modern warfare, data and AI are a critical intelligence function, and it can even be said to be the only critical intelligence function.
Noted military commentator Paul Scharre has stated that AI will inevitably lead to lethal automation. In his bestselling book “Army of None” (Army of None), published in 2019, he depicted the rise of remote automated weapon systems. Scharre proposed in the book that AI will revolutionize warfare: “In future wars, machines may decide life and death.” Even though AI's potential still fascinates him, his thoughts have undergone significant changes. In his new book “Four Battlegrounds” (Four Battlegrounds published in February 2023), he made substantial revisions to his initial views. In the new book, he abandoned the catastrophic scenarios depicted in “Army of None.” If “Army of None” is a sci-fi essay, then “Four Battlegrounds” is a work of political economy. It interprets the specific issues of great power competition and the industrial strategies and regulatory systems underlying that competition. This book describes the impact of digital intelligence on military competition. Scharre analyzes the regulatory environment needed to harness the power of data. He compellingly argues that in the superpower competition between the U.S. and China, the race for data and the advantage of processing data with AI will become a decisive factor in military affairs. Data will provide significant intelligence advantages. In Scharre's view, there are four key resources that will determine the winner of this intelligence race: “The country that leads in these four battlegrounds—data, computing, talent, and institutions (tech companies)—will gain a significant advantage in AI.” He believes that the U.S. and China will engage in a life-and-death struggle over these four resources. Both China and the U.S. have fully recognized that whichever country gains an advantage in AI will significantly lead in political, economic, and critical military domains. That country will know more than its adversary. That country will be able to use military force more efficiently. That country will dominate information and cyberspace. That country will be more lethal.
“Four Battlegrounds” depicts the intense competition between China and the U.S. over data and AI. It describes the recent progress of these two countries and assesses their relative strengths. China still lags behind the U.S. in some areas. The U.S. has a talent advantage and leads in research and technology. “China is stagnant in chip production.” However, Scharre warns the U.S. not to be complacent. In fact, this book vividly showcases the author's concerns that the U.S. will fall behind in the data competition. Therefore, Scharre emphasizes China's advantages and its rapid development. China has 900 million internet users, and its data volume far exceeds that of the U.S. In economic sectors like ride-hailing, China's level of digitization far surpasses that of the U.S. For example, the U.S. lacks a product comparable to WeChat. Many of China's applications outperform similar applications in the U.S. Additionally, the Chinese government is not constrained by legal or citizen privacy concerns. The Chinese Communist Party actively monitors its citizens' digital profiles—collecting their data and recording their activities.
Government control also benefits China's tech companies: “The Chinese Communist Party's massive investments in intelligence monitoring and social control have greatly propelled the development of Chinese AI companies and tied them closely to the government.” The Chinese government collaborates closely with the tech industry. China's regulatory advantages also far exceed those of the U.S. The Chinese Communist Party provides guarantees for tech giants like Baidu and Alibaba: “China's investment in the tech industry is paying off.” Scharre concludes: “China is not only creating a new model of digital authoritarianism but is also actively exporting it.”
How will the U.S. government respond to China's quest for dominance in data and AI? Scharre writes interestingly on this aspect. The U.S. government needs to make significant regulatory changes to harness the military potential of data effectively. The armed forces need to collaborate closely with the tech industry. They “cannot only work with traditional defense contractors but must also engage with startups.” This is not an easy task. Scharre documents the challenging regulatory environment in the U.S.: “In the U.S., large tech companies like Amazon, Apple, Meta, and Google are independent power centers that often disagree with the government on specific issues.” Scharre discusses Google's notorious protest in 2017—when employees refused to participate in the Department of Defense's Project Maven. Some in the U.S. tech industry are skeptical about the military applications of AI.
U.S. tech companies may have always been reluctant to collaborate with the armed forces, and the Department of Defense has not provided assistance. This has inadvertently hindered collaboration between the military and tech companies. The relationship between the Department of Defense and the defense industry has always been close. For example, in 1961, President Dwight D. Eisenhower warned about the threat of the “military-industrial complex” to democracy. The Department of Defense has developed a procurement and contracting process primarily designed for government procurement of combat platforms, such as tanks, ships, and aircraft. To meet the diverse requirements of the Department of Defense, Lockheed Martin and Northrop Grumman have become very adept at delivering corresponding weapon systems. However, tech companies do not operate this way. As one of Scharre's interviewees noted: “Buying AI is not like buying bullets.” Tech companies do not sell specific functionalities like firearms. They sell data, software, and computing power—in summary, they sell expertise. The best way to develop algorithms and programs for a very specific problem is through iterative development. For a military task, the full potential of certain software and algorithms may not be immediately apparent, even to tech companies. The operating environment of tech companies is highly competitive, so they prefer to establish more flexible, open contracting systems with the Department of Defense—tech companies need secure and rapid returns on investment. What tech companies seek is collaborative participation, not just a contract to build a specific platform.
The U.S. military, especially the Department of Defense, finds that this new approach is not always an easy contracting model. In the past, the bureaucratic response to demands was sluggish—the procurement process could take 7 to 10 years. However, despite many tensions and the system being far from perfect, Scharre also notes that the regulatory environment is changing. He describes the emergence of a new military-tech complex in the U.S. Of course, Project Maven is an example of this process. In 2017, Bob Work released a now-famous memo announcing the “Algorithmic Warfare Cross-Functional Team”—Project Maven. Since the emergence of surveillance drones and military satellites during the global war on terror, the U.S. military has been immersed in full-motion video streams. The value of those recordings is unparalleled. For example, in 2019, the U.S. Air Force used the 24/7 aerial surveillance system Gorgon Stare to trace back a car bomb explosion in Kabul (which killed 126 civilians) and ultimately located the safe house used to carry out the attack. However, using humans to do this is simply too slow. Therefore, the Air Force began experimenting with computer vision algorithms to screen full-motion video. The goal of Project Maven is to assist the Air Force. But it requires a new contracting environment. Work did not adopt the lengthy procurement process but instead introduced a 90-day short-term process. Companies had three months to demonstrate their utility. If they made progress, contracts would be executed with them—if not, they would be out. Meanwhile, Work declassified drone footage so that Project Maven could train its algorithms. By July 2017, Project Maven had an initial operating system capable of detecting 38 different target categories. By the end of that year, the system was deployed in operations against ISIS: “This tool is quite simple, capable of identifying and tracking personnel, vehicles, and other targets from video footage captured by the ScanEagle drone used by special operators.”
Since Project Maven, the Department of Defense has also launched several other initiatives to promote collaboration between the military and the tech industry. The Defense Innovation Unit (The Defense Innovation Unit accelerates the relationship between the Department of Defense and Silicon Valley companies, offering contracts with a 26-day deadline instead of months or years. In the first five years of the Defense Innovation Unit, it signed contracts with 120 “non-traditional” companies. Under the leadership of General Jack Shanahan, the Joint Artificial Intelligence Center has played a crucial role in advancing cooperation between the armed forces and tech companies, involving tasks such as personnel rescue and disaster relief operations, developing software for mapping wildfires and post-disaster assessments—we do not know whether these examples in Scharre's book imply more military uses. After experiencing early difficulties, General James Mattis created the Joint Enterprise Defense Infrastructure during his tenure as Secretary of Defense, revolutionizing the procurement system for tech companies. For example, the Department of Defense invested nearly $100 million in 2021 to help Anduril develop an AI-based anti-drone system.
“Four Battlegrounds” is an excellent complement to the literature on AI and warfare, rich in content. The message this book conveys is clear. For the military, data and AI are now important and will continue to be crucial in the future. However, data and AI will not fundamentally change the nature of combat—the operation of lethal weapon systems will still primarily be in human hands, including remote weapon systems capable of killing, as seen in the brutal war in Ukraine. The conditions in combat are complex and bewildering. To maximize the effectiveness of weapons, human judgment, skill, and cunning are essential. However, any military hoping to win on future battlefields must harness the potential of big data—the military must master the surging digital information in the battlefield space. Humans alone are incapable of achieving this. Therefore, headquarters need algorithms and software to process that data. The military needs to establish close partnerships with tech companies to create these systems, and operational command centers need data scientists, engineers, and programmers to ensure these systems function properly. If the armed forces can achieve this, data will allow them to see the battlefield space with greater depth and breadth. This cannot completely resolve the issues of military operations—the fog and friction will still persist. However, commanders empowered by data may be able to deploy forces more effectively and efficiently. Data can enhance the lethality of the armed forces and human combat units. The Russia-Ukraine war has already shown us that data-centric military operations can outperform adversaries that still operate in a simulated manner. Scharre's book is a clarion call to ensure that the fate of the Russian military in Ukraine does not fall upon the U.S. when the next war arrives.