From the Archives | The Future of Military Applications of Artificial Intelligence

A Role For Confidence Building Measures?

Michael Horowitz, Laura Kahn, Carey Mahoney- Fall 2020 Vol.64. Issue 4

Today, militaries around the world, from those of great powers to smaller states,  are exploring how they can take advantage of artificial intelligence (AI)  applications. Novel military technologies, and the political-strategic opportunities that leaders imagine they create, have long shaped military strategy.  Drawing on the way that algorithms increasingly shape our interactions with the world,  from Netflix recommendations to Google search results, militaries looking for an edge are seeking to harness these advances in AI to further their own ends. There is still  significant uncertainty about what is possible concerning artificial intelligence, and  nearly all military applications still exist in the world of hypotheticals. The stakes still  seem high to policymakers, since militaries that can successfully combine  organizational change and new concepts of operation with appropriate technologies  often have had an edge in the history of warfare. The expectations that decision makers possess about technological use also drive the ways states seek to expand or  limit technological change itself. 

There is rising concern about how military uses of AI could create  international instability and conflict. Algorithmic bias, hacking, and data-poisoning  could create a race to the bottom, where militaries adopt uses of AI that promise  capabilities, but deliver accidents and greater risks of miscalculation. [2] The specific use  of lethal autonomous weapon systems on the battlefield could increase human  suffering and make war more likely.[3] Introducing AI in the nuclear domain could bring  the potential for automating early warning or for arming uninhabited platforms with  nuclear weapons, or it could increase the speed and escalation of conventional war,  engendering new pressures on strategic stability.[4] Each of these scenarios naturally  leads to the question of what individuals, countries, and the world can do to decrease  the risks surrounding military uses of AI. Thinking about risk reduction in the military  context reflexively raises proposals related to arms control. But how can states and  other stakeholders begin to think about controlling arms while their technologies and  doctrines have yet to fully emerge?  

To answer this question, we begin by taking stock of how militaries are  thinking about the potential applications of AI, showing that major, path-breaking  innovations likely lie beyond the horizon of what the world has seen to date.  Significant disagreement exists, even among those focused on how advances in AI  could shape the character of warfare, concerning what form these advances will take  and how they could influence military operations for different countries. We, then,  turn to the question of potential risk reduction through the lens of confidence-building  measures (CBM). 

During the Cold War, CBMs arguably helped decision makers manage similar  uncertainty amidst technological and doctrinal change in ways that facilitated great power cooperation. By comparing the present situation pertaining to AI to some of  the characteristics of the Cold War, we illustrate the extent to which applying CBMs  to military applications of AI may be both desirable and plausible. We argue that  learning among states plays a critical role in discovering the tacit rules by which they  engage in warfare enabled with new technologies. Therefore, measures that help  stimulate this learning pose the greatest chances of risk-reduction success at present.  The article concludes by highlighting three questions that any successful effort to  manage or otherwise control AI proliferation will need to consider.

The State of the Tech  

As AI continues to develop, it increasingly demonstrates the latent potential  to alter the global balance of power and strategic stability.[5] While many military  applications of AI are still more theory than reality, clearly this will not be the case for  long, as the promise of military AI has already led to significant changes in armament  dynamics. In fact, some of the world’s largest powers are prioritizing investment and  research and development (R&D) in new AI technologies. From AI researchers to  political scientists, many experts have raised their voices to warn of the unique risks  that accompany the proliferation of military applications of AI and the negative  implications it could have on the stability of international order.[6]

At present, national governments’ investments in AI are in their nascent  phases. The majority of advances in AI technologies and their applications are being  driven by the commercial and academic spheres.[7] As a result, these efforts have  generated a significant bedrock of publicly available, foundational research and open source tools, thereby decreasing the costs of these technologies. Although not  originally designed with military applications in mind, “Many of these capabilities are  dual- or multiuse across civilian and military contexts.”[8] With the direction of the public-private R&D pipeline thus effectively reversed, compared to when society more  frequently adopted military projects like the Defense Advanced Research Projects  Agency’s ARPANET (the predecessor of the internet), states and militaries must  grapple with the reality that the pace and direction of AI proliferation may be  impossible to control.  

States on every continent are forging their own paths toward corralling their  national capabilities to bring about battlefield-ready AI. A foundational example is the  U.S. Defense Department’s Project Maven. The Project attempts to provide better,  more actionable intelligence by creating “computer vision” algorithms that recognize  and identify object classes from drone video footage to support targeting decision making.[9] Another U.S. program, Sea Hunter, aims to field an autonomous ship  designed to assist with anti-submarine and electronic warfare, offering an example of  a U.S. AI system more intricately integrated into a battlefield platform.[10]

At the same time, the People’s Republic of China has been working to fulfil a  three-phase plan to become the world leader in AI by 2030. As early as 2011, the  Chinese military defined an AI weapon as “one that utilizes AI to pursue, distinguish,  and destroy enemy targets automatically; often composed of information collection  and management systems, knowledge base systems, decision assistance systems,  mission implementation systems, etc.” [11] Steady progress in the People’s Liberation  Army’s broader modernization program across a range of capabilities has included  efforts to create AI systems that support Chinese interests both at home and abroad.  Such initiatives range from an underwater drone program[12] to extensively integrating  advanced cloud computing, surveillance technology, and facial recognition software as  part of the Chinese Social Credit System.[13] Echoing China, since 2017 when Vladimir Putin commented that whoever  leads the world in AI will rule it, the Russian Federation has also been investing in  AI—albeit at a much lower level than the United States or China.[14] Still, the Russian military already has deployed the Nerekhta small autonomous combat vehicle and the  Uran-9 unmanned light tank.[15] Russia continues to invest in developing additional  technologies, including an early-stage lethal autonomous armored turret, “capable of  independently acquiring, identifying and engaging targets.”[16]

Other countries also are beginning to test AI-enabled or increasingly  autonomous military technologies. Israel’s Rafael Advanced Defense Systems unveiled  a new Spice-250 precision-guided munition, which purports to use AI and deep  learning to enable “automatic target recognition.”[17] A collaborative project between  Israel’s Rafael and France’s Atos Information Technology, contracted by the German  military, has been working to generate a “glass battlefield,” in which a combination of  AI, unmanned aerial vehicles (UAVs), combat vehicles, computer vision, and more,  provide real-time 3D situational awareness for mobile operations. [18] Singapore is using  AI to “conduct predictive maintenance on navy ships.” [19] The Republic of Korea has  two ongoing AI information-processing and response projects named Exobrain and  ADAMs, as well as a Gyun-Ma, an unmanned combat vehicle project.[20] Other militaries, though further from testing and deploying systems, are also  increasing AI investments. Argentina’s 2019 military revitalization plans identified AI  and autonomy as a necessary component for future command and control (C2) systems  and “sensitive” operations. [21] Among African and Middle Eastern states, where  spending on AI was set to total $290 million in 2019, some countries such as Egypt  have focused on increasing human capital rather than making immediate technical  progress. [22] In sum, by July 2018, 24 countries, along with the European Union  Commission and the Nordic-Baltic region, had published national strategies to  promote the use and development of AI.[23] But to what strategic end? 

Shifting Technology, Shifting Expectations  

A 2015 “Open Letter from AI & Robotics Researchers” warned, “If any major  military power pushes ahead with AI weapon development, a global arms race is  virtually inevitable, and the endpoint of this technological trajectory is obvious:  autonomous weapons will become the Kalashnikovs of tomorrow”—that is, as  ubiquitous as the Soviet-era machine guns that fill small-arms markets around the  globe.[24] Whether or not one agrees that recent developments in military AI will  precipitate a global arms race, many believe that although militaries constantly seek to  advance their capabilities, there is a unique sense of urgency and competition that  accompanies governments’ quests to integrate AI into their national security  strategies.[25] 

One way to explain this sense of urgency is to attribute it to the uncertainty  that states face about where exactly the military AI revolution will lead. New technologies not only create real, material opportunities for states to push and pull on  one another in new ways in international politics, but they also change states’ broader  strategic expectations about how they can best achieve their goals. How technologies  shift these expectations, though, is determined by more than who invests the most or  does so most quickly. In other words, no amount of spending on R&D can fully secure  a state against the unknown. 

States come to understand the potential of their innovations with greater  certainty over time through a process of learning about the ways military applications  of new technologies can exert maximum strategic influence. Naval carriers entered  several major powers’ fleets in the interwar period, for instance, and there was intense  experimentation. However, it was not until the improvisation of the early part of  World War II that states adopted doctrines for their more effective (and unforeseen)  use as platforms for extending the stretch of air power. [26]  

Regarding battlefield applications of AI, warfare among autonomous systems  that decrease the presence of humans in combat potentially draws the biggest  disjuncture with previous forms of war and, by extension, politics. The processing  power of AI that enables split-second decision-making may, in interaction with other  AI systems, drastically increase the pace of battle. Response loops between combatants  that unintentionally escalate or otherwise spin out of control would, under such  circumstances, challenge the ability of human decision-makers to manage battlefield  interactions. Conscious of these potential unintended consequences, armed forces are  using wargames to experiment with how better-trained commanders and new AI  technologies might limit such risks and to identify requirements for future systems.[27]

Such efforts can indeed help states learn how AI can enhance their national  capabilities. Yet, experimenting and wargaming short of combat can only go so far in  “buying down” the unforeseeable risks of a new technology. Critically, what such  efforts cannot do is assess fully how AI-driven solutions built to fill gaps in existing  doctrinal concepts—let alone major AI breakthroughs—will aggregate into novel  strategic settings where all parties have adopted AI to their own perceived needs and  purposes.[28] Given this scenario, only experience over time can tell what net effect,  however large or small, AI will have on the character of war and on how states form  expectations about the ways AI might best help leverage their power. 

A part of the strategic dynamics of any interaction is the metaphorical “rules  of the road,” both implicit and explicit, that guide how actors on one side of a confrontation interpret and form expectations about the actions of those on the  other.[29] As novel technologies and the doctrines that guide their use in armed conflict  emerge—and experiences of militaries’ use and observation of new technologies in  battle accumulate—decision-makers come to form clearer ideas and expectations  about the new rules of the game. 

An earlier general-purpose technology applied to military use like AI—electric  telegraphy and the global network of transoceanic cables it spawned—provides a case  in point. As international telegraphy grew, it gave states, particularly those with  overseas colonies, an unprecedented capability to send nearly instantaneous orders to,  and receive reports from, military commanders across the globe. Yet, long-distance  cables, including those undersea, were vulnerable to being severed physically by  adversary navies. International law prohibited interference in official state  communications, but such rules explicitly did not apply when state security was at  stake.[30]

Despite owning the plurality of the world’s long-distance connections at the  turn of the nineteenth century, hegemonic Great Britain remained vulnerable to the  possibility that an enemy could sever telegraph cables to prevent London from  communicating with its forces in a future theater of war. By 1897, attuned to this risk,  British leaders were calling for a British-owned, globe-spanning network of redundant  connections to defend against this contingency.[31] Indeed, in 1898, the imagined use  of cable-cutting tactics came to pass when the U.S. Navy ploy to isolate Spanish  commanders in the Philippines and Cuba led the Americans to cut Spanish and British owned cables. [32]

Strategists drew a key lesson from this debut episode of cable-cutting warfare.  Although it neither delivered Spanish forces nor doomed the Americans, the U.S.  Navy’s failure to locate the final cable off the coast of Cuba in the employ of Spanish  forces did show that, to succeed, cable-cutting relied on a significant degree of chance.  The Spanish-American War in the Caribbean drove home the reality that finding cables  in a seabed is no easy task. This realization would have a significant impact on the  strategic use of cable-cutting going forward. [33] The British, who owned most of the  world’s cables and the plurality of the world’s naval forces by 1900, made two choices  in response to the Americans’ 1898 experience that made good strategic sense in light  of this lesson-learned. First, accepting the cutting of undersea cables, even of those  owned by non-combatants, as a legitimate tactic of warfare helped establish a tacit precedent that Great Britain could leverage in its own future statecraft.[34] Second, the  choice to construct redundant, though costly, cables to secure British mastery of global  military communications, in interaction with states with smaller navies and less expansive cable networks, represented a trump card against other states’ potential  cable-cutting threats. For lesser powers, challenging British telegraphic dominance in  wartime with a cable-cutting strategy not guaranteed to succeed simply did not make  sense. Indeed, it was not until the German fleet’s blitz to match British naval power  in the years preceding World War I that the British monopoly on cable-cutting as a  strategic wartime ploy met serious operational, if not strategic, challenges.[35] 

Much as the telegraph network’s material characteristics as a hard-to-find cable  at the bottom of the sea ultimately dictated the constraints within which early  twentieth-century powers would strategize, so, too, is military AI bound to shape the  rules by which twenty-first century powers will play. A key lesson is that these rules  tend to reveal themselves as states use their new technologies in interaction with one  another. Only as they learn about what AI can do in broader-scale military applications  will states gain a clear picture of the strategic dynamics and power shifts prompted by  military AI’s emergence.  

The bad news is that, in the meantime, states’ plans to use AI in battlefield  contexts will remain largely hypothetical: informed by expertise, to be sure, but subject  to significant learning and change. The opacity, for instance, of Chinese plans to apply  AI to the People’s Liberation Army’s “information-centric” operational concepts  makes it difficult for U.S. strategists to gauge whether U.S. firepower-enhancing and  “maneuver-centric” approaches to AI are indeed a best response.[36]  The good news is that once decision-makers understand these dynamics and  particularly if they have written them into international law, states’ behaviors that  follow or break rules are more easily interpretable by other states. Rules thus help  states communicate (to allies and adversaries alike) their intentions in ways that can  reduce the chances of unintentional escalation.[37] As Caitlin Talmadge shows in the  context of Cold War crises, it is rare that emergent technologies themselves caused  escalation risks to increase or decrease.[38] Rather, as with the telegraph, technological  shifts provide one of many vehicles through which decision-makers shifted their  expectations about how best to apply coercive power to pursue national interests. 

Managing Uncertainty About the Rules of the Road  

In the meantime, while the rules relating to new military technology have yet  to find definition, how are states to cope? Would attempts to establish such rules not  merely be crude guesswork about a form of warfare whose contours, once discovered,  would quickly antiquate early attempts to contain it?  

The Cold War between the United States and Soviet Union provides one  setting in which to analyze the conditions that make CBMs both plausible and  desirable. The nuclear attacks on Japan and subsequent tests made clear the material  effects of atomic weapons: quick, assured, and causing massive destruction. Like the  state of AI operational concepts today, though, optimal doctrine in nuclear warfighting  remained subject to impassioned debate on each side of the Iron Curtain for much of  the Cold War. This lack of consensus both within and between states created strategic  uncertainty. Leaders in neither the East nor West were confident that they could fight  and win a nuclear war at an acceptable cost. In fact, chances were that the opposite  was true. Given that basic, macabre shared understanding, it was common interest  that brought the parties together to develop specific measures to build confidence.  Both sides would detect a surprise first strike by the other, thereby rendering a “bolt  from the blue” a useless ploy in the nuclear strategy rulebook.  

Confidence-building measures, per se, were formally introduced in U.S.  President Dwight Eisenhower’s “Open Skies” Proposal in 1955. Eisenhower  suggested swapping U.S. and Soviet blueprints of military establishments and  providing facilities for aerial photography as immediate, “practical” steps toward  “lessening danger and relaxing tension” by providing each superpower an opportunity  to demonstrate a serious commitment to arms control.[39] Later, others would consider  the communications channel, or “hotline,” the Soviet Union and the United States put  in place after the Cuban Missile Crisis to be a CBM, as well.[40] However, the  Commission on Security and Cooperation in Europe Helsinki Conference in 1975 is  perhaps most widely considered the birthplace of modern CBMs. [41] In a seminal monograph on the subject, Political Scientist Marie-France  Desjardins argues that the CBMs that emerged from the 1975 Helsinki Final Act arose  from a mutual bargaining interest in preventing nuclear war. The CBM’s were intended  to reduce the likelihood of a surprise attack via the facilitation of the “exchange of  information, notification, and observation of major military activities.” [42] According to  Desjardins, three conditions make CBMs an appropriate tool for mitigating the risks  of potentially destabilizing technologies. First, the chief concern must be that lack of information could escalate the likelihood of an attack. Second, communications or  verification technology exists that could remedy the lack of information. [43]Third, the  parties to a CBM must be states with high degrees of command and control over their  armed forces. In 1975, the U.S.-Soviet confrontation met all three of these conditions.  Both parties feared that a lack of information and uncertainty about the political intent  of the other would engender the necessary circumstances for a surprise nuclear attack.  

When these conditions are met, CBMs are an attractive first step for two  reasons: first, they work to reduce risk through building trust; second, CBMs can be  “stacked,” or incrementally layered, to increase this trust and can therefore serve as a  building block to further reduce tensions. CBMs do not tend to deal with root causes  of conflicts. Rather, they serve as low-risk, “path clearers” on the road to better  relations between states by reducing the likelihood that a worst case scenario comes to  fruition or, alternatively, by mitigating a specific risk, such as an accidental or surprise  attack as was feared in 1975.[44] 

Still, CBMs are not risk-free themselves. “Negotiating, and sometimes even  executing, these measures can actually damage international security,” Desjardins  argues, by increasing rather than decreasing mutual suspicion. “CBMs are only as  strong as the fundamental political will for compromise and successful negotiations.” [45] One party’s defection from an agreement can reinforce perceptions of  untrustworthiness. The defection exacerbates conflict if this leads an actor to capitalize  on such noncompliance by using it to justify antagonistic responses. [46] Such outcomes  are likely when parties lack the political will to prioritize overcoming challenges in CBM  negotiations and when CBM provisions are too broad, vague, or unverifiable.  Furthermore, parties can also use CBMs as a stalling tactic or as leverage in other  negotiations.  

The 1975 Helsinki CBMs process overcame these hurdles, though, and were  considered an early confidence-building success. They served as a template for  measures in other international settings over the following decades. Their immediate  effect in the U.S.-Soviet context was to reduce uncertainty of the proverbial “rules of  the road” of nuclear war—most helpfully, without either side needing to experience it.  By generating credible information that constituted a reciprocal early-detection system,  CBMs narrowed the range of plausibly effective strategies for nuclear use that either  side might imagine adopting. The shared interest in reducing the probability of surprise  nuclear attack enabled the success of CBMs. 

From Confidence to Control?  

Researchers and policy practitioners are considering whether confidence building measures might help avoid worst case scenarios related to military AI. These  scenarios range from battlefield accidents to unconstrained arms racing. Some experts  are calling for more stringent controls on the most worrisome AI systems through  blanket bans of lethal autonomous weapons systems.[47] In light of their historical  application, how much might CBMs—or even formal arms control—plausibly help  manage the risks of AI?  

Skeptics might point to how little the present military uses of AI meets  Desjardin’s three conditions for when the CBMs are likely to succeed as the Helsinki  measures did. First, the threat of mutual annihilation is not the most salient fear that  the states investing most heavily in AI currently experience. As a result, the  overlapping interests that make bargains possible are less likely to drive parties to a  CBM agreement in the same way that the fear of surprise attack catalyzed cooperation  in Helsinki. Second, technologies to verify use of AI in conflict are not mature enough  to both respect an offending state’s desire to keep algorithms secret and to provide an  investigator credible evidence that, say, an AI accident was indeed unintentional. Third, states and military chains-of-command may not have the level of command and control  over AI-infused forces that they think they do. Delegating analytic tasks and decision making authority to computer systems will likely require changes to C2 systems that  may remain untested in battle for the foreseeable future. This situation would make it  harder for CBM negotiators to make reliable commitments to ensure the restraint of  such systems. 

These challenges are complicated further by the lack of opportunities for  “near-battlefield” learning about military AI applications. Despite the presence of  factors that limit the chances of developing traditional CBMs, we contend that  governments could benefit from pursuing CBMs to facilitate such learning to reduce  shared strategic risks. Although states appear to lack as strong a common, motivating  political interest as the ones that led to Helsinki agreements, they do share a common  lack of understanding of how AI will change the implicit rules of strategic engagement  in international politics. Borrowing from the CBM toolkit would be one way of  pursuing a shared interest in reducing the uncertainty associated with the “rules of the  road” that only collective international action can create.  

A CBM-like focus on international transparency and information-sharing  could substantially alleviate risks associated with AI and its nature as a technology  prone to oversimplification and miscomprehension. As an enabling technology with  many discrete applications, the amorphous quality of AI exacerbates uncertainty over  how its integration into existing platforms and doctrine will change the character of  warfare. So far, no international consensus has emerged on how the explicit rules of  warfare, that is, international law, should change. 

Toward the goal of developing understanding of the tacit rules for AI, it looks  like there is much that could be done. Indeed, AI “does not lend itself well to  verification, much less controls, unlike the delivery platforms and warheads generally  associated with nuclear deterrence,” says Lt. Gen. Jack Shanahan.[48] But, as Director  of the Pentagon’s Joint Artificial Intelligence Center (JAIC), he has argued that early  efforts by the U.S. Department of Defense to incorporate ethics into discussions of  adopting AI “almost” represent confidence building measures. [49] Focusing on ethics— the rules that guide humane decision-making in any domain—in international  discussions may well help reveal common interests or points of divergence. Dialogue  among various parties could provide significant insight into how warfighters,  engineers, and policymakers are considering, for instance, the delegation of life-and death decisions to computers.  

For analysts, identifying such common interests among states pursuing  military applications of AI is a critical starting point from which to develop proposals  for international confidence-building or to move more ambitiously from a focus on  confidence to one on arms control. During the Cold War, the desire to avoid nuclear  annihilation was obvious and a credible enough shared interest. Today, given the non obvious technological, doctrinal, and strategic impacts of AI, and given the chances  AI’s most useful military applications could be in its generalized application to  warfighting, the search for such overlaps in states’ political interests is bound to be a  far more subtle task. 

Toward a New Template  

The central problem among states who may wish to identify the implicit rules  of the road for using military AI resembles less of a prisoner’s dilemma and more of a  coordination problem. That is, states are less worried about whether they will get  “suckered” by another party’s “defecting” by blitzing to some full-scale AI force or all empowering, critical discovery. Indeed, research budgets are substantial on all sides,  and the problems for AI innovation are less technical than they are organizational or  doctrinal. Instead, this essay argues, an under-acknowledged part of the shared  insecurity states feel from the AI militaries on the horizon comes from the difficulty  of coordinating a consensus on what the rules of the road of AI warfare will be. Whether such rules will emerge, we have argued history suggests, is a moot discussion. 

The CBMs of old may not offer an exact template for how best to manage the  transition from conventional to AI-laden warfare, and we do not seek to offer specific  prescriptions for the shape confidence-building for AI should take. But, framing  today’s situation of international uncertainty as a coordination problem about the  future tacit rules for military AI use leads us to pose three questions that efforts to  manage AI proliferation will need to answer in order to resolve strategic ambiguity and  build viable risk-reduction solutions.  

First: What can military AI technology do? Knowledge about the material facts,  that is, the technical and organizational features of military-relevant AI applications, is  the foundation on which to build understandings about the risks themselves and the  means to reduce them. In- and near-combat experiences with AI, like those of the  telegraphic cable-cutters or naval carriers in prior centuries, would provide military  planners ways to optimize doctrine and strategy. As argued above, AI system tests,  operational experiments, wargames, and early introduction into strategic settings  obviously help states learn about their national capabilities. Only when information is  available to all parties in a strategic interaction, though, can the tacit boundaries that  help limit escalation and unintended consequences of military force emerge.  

There were limits to the availability of information about rarely used  technology in the Cold War, too, but that did not prevent CBMs and arms control.  One reason for this dynamic was the risk of global destruction due to a U.S.-Soviet  nuclear war. After the Cuban Missile Crisis, new perceptions of risks, generated by the  ambiguities of nuclear brinkmanship, created an impetus for CBMs. In-depth CBMs  and more intrusive arms control agreements only emerged a generation after the  creation of nuclear weapons, once arms buildups made the destructive potential of  nuclear weapons even clearer and states’ thinking about their use had evolved. 

In addition to reducing uncertainty about rules, establishing collaborative  efforts to exchange technical information might help form the basis for verification or,  in the case of misuse, attribution systems more robust nonproliferation measures  would need in the future. Decades of global, scientific exchange directed at supporting  nuclear test explosion monitoring beginning in the 1970s were what laid a similar basis  for negotiations of the Comprehensive Nuclear-Test-Ban Treaty in the 1990s.  

Second: What military-strategic trade-offs will different military AI rules imply? The  asymmetrical and pocketed development of military AI systems around the globe  suggests that politically viable bargains may need to be tailored to the interests of each  of their parties.⁵⁰ Coordination problems theoretically lack solutions that benefit some  parties more than others, but this is rarely the case in reality. Understanding how  agreeing to one set of standards, rules, or information-sharing and transparency  practices will disproportionately affect those at a comparative disadvantage in one area  is likely key to identifying stable, self-enforcing conceptions of shared rules and ethics.  

Third: What political trade-offs will different military AI rules imply? International  cooperation often encounters political obstacles at both the international and domestic levels. Rather than cataloguing the many roadblocks that AI risk-reduction efforts may  face, considering instead those that are novel to military AI today foregrounds two  relevant features. First, new actors in national security technology politics, particularly  those in Silicon Valley in the U.S. case, have unique business and organizational  interests that, as Google’s withdrawal from Project Maven laid bare, can disorient  governments. Assumptions that a state holds about its ability to access national  capabilities in the private sector (or the university systems of a foreign nation) for the  sake of national security, if proven false, might have a cascading effect on their AI  strategies on the global stage. Second, states that preside over populations  disproportionately targeted by domestic police may have unique constituencies to  which they or complicit technology developers must answer, like Chinese citizens in  Hong Kong or Black Americans, in response to domestic applications of militarized  AI. ⁵¹ Subsequent political consequences of retaliatory upheaval to perceived  violations, or the threat thereof, might push states’ interests to value some rules for  international AI engagement over others. 

Looking Ahead

There is growing desire in the international community to figure out ways to  reduce the risk that military uses of AI undermine international stability. An expanding  chorus of experts suggests that enhancing dialogue among countries, including the  United States and China, would mitigate such risks.⁵² Considering the conditions in  which CBMs have met with historic success, we recognize that current trends act as  factors that may decrease the overall chances that CBMs will govern military uses of  AI in the near future.  

Yet, the alternative framework we present for thinking about the strategic  effects of emerging technology helps show that being clear-eyed about the prospects  for AI risk reduction need not imply unqualified pessimism. If more formal CBMs,  resembling their Cold War predecessors, are potentially less feasible at present, then  other tools from the broader CBM toolkit might still have a positive impact in reducing  the risks of uncharted strategic territory. A Stockholm International Peace Research  Institute report has suggested that even the simplest of measures, like states publishing  lists of their ongoing R&D related to military AI systems, could have a positive effect  on international stability. Exchanging such information might not identify every  possible way that introducing AI to combat would change the rules of the road. But, it would at least help strategists know which lanes are safest for them and their  competitors to drive in. Despite challenges states face in pursuing “traditional” CBMs,  future policy research and innovation can seek to resolve the uncertainty about how  AI might change the tacit, and even formal, rules of international politics and conflict. 

Arms control skeptics marshal strong arguments, which we take seriously,  against the chances that efforts to encourage international cooperation on military  applications of AI will succeed in any meaningful sense in the short term. Moreover,  we have not explicitly considered the possibility that AI might, in fact, have little to no  discernable impact on how decision-makers evaluate the use of force in international  politics. The premise of the argument here, though, is that finding out such facts to  be true or not by leveraging the CBM toolkit would offer useful knowledge for states  rushing to militarize AI without a clear and high-confidence picture of its strategic  effects in mind. Still, should negotiations toward this end fail, even the most realpolitik minded Machiavellian should recognize that keeping one’s enemies close, particularly on the precipice of a shift in the character of warfare driven by  AI, is sage advice.