Innovation multiplies the number of actors who have militarized AI available --- that shreds deterrence

The strategic implications of artificial intelligence run deep as they have the potential to upend the fundamental calculus behind deterrence by putting the premium on offence and pre-emption (Rickli 2017). Lynn-Jones notes that “conflict and war will be more likely when offence has the advantage, while peace and cooperation are more probable when defence has the advantage” (Lynn-Jones 1995, 691). Therefore, as argued by Rickli, the impact that AWS have on the offence-defence balance is an essential element in assessing its implications for strategic stability (Rickli, 2018). The use of AI in command-and-control structures, such as automated decision-making and weapons in the form of AWS, risks tilting the balance towards the offence.

First, at the tactical level, how AI-powered weapon systems are likely to be used largely favours offensive use (Rickli 2018). One such example is swarming tactics, which aim at overwhelming enemy defence systems and neutralising any advantage gained by a defensive posture. Following Lynn-Jones’s argument, an environment in which defensive postures are likely to be inefficient enhances the possibility of conflict and strategic instability by incentivising pre-emptive strikes. Second, the opacity inherent to artificial intelligence algorithms similarly affects the offence-defence balance, as it leads to an incapacity to appropriately judge an adversary’s capabilities and their impact (Horowitz 2018, 55). Unlike traditional military capabilities, which can be more easily observed, a state’s armament in AI-related capabilities would be almost impossible for another state to measure accurately. Assessing the depth of automation of an adversary’s arsenal, the quality of the code, the efficiency of the autonomous weapons, and their capabilities would prove difficult. This could create an environment where states are pushed to assume the worst of their rivals and over-estimate their capabilities (Horowitz 2018, 55).

The lack of first-hand battlefield experience of these capabilities could similarly push states to over-estimate the destructive potential of adversary’s weapon systems, leading to an increased probability of arms races resulting from security dilemmas (Horowitz 2019, 779). In this case, states will not only have difficulty in discerning a state’s intention but would also have trouble assessing capabilities. This adds a dimension to Jervis’s traditional understanding of arms races resulting from security dilemmas, where states can measure each other’s capabilities but not intentions (Horowitz 2019, 779). As posited by Price, Walker, and Wiley, understanding leadership, their intentions, and how they might react to a certain event is part of deterrence. In the case of automated decision-making and AWS, it would be impossible for a state to determine how AWS would respond to one state’s action, increasing mutual unintelligibility between adversaries and strategic uncertainty (Price et al. 2018, 101).

Furthermore, in a “humans out of the loop” scenario, opportunities for de-escalation and non-use of force would be scarce. Lastly, as Horowitz argues, “the relationship between speed and crisis stability in a world of deployed LAWS [lethal autonomous weapon systems] represents one of the clearest risk factors associated with autonomous weapons” (Horowitz 2019, 781). In fact, with the expectation that hostilities will be conducted at machine speed, states might suffer from a fear of losing quicker, hereby creating escalation pressure and incentives to act first with a pre-emptive strike (Horowitz 2019, 782).

It is, however, worth noting that the argument of the offensive advantage provided by artificial intelligence is a disputed one. For instance, Hammes (2021) makes the opposite point by arguing that emerging technologies and AI usher in an era of defensive dominance on the battlefield because the attackers always leave multiple signatures that can be more easily identified by the defenders. While this argument holds some truth, it is limited as it assumes that strategic stability only relies on one material factor. The growing development of stealth technology also undermines the point. However, strategic stability depends on several tangible factors and relies on perceptions. Thus, we agree with Paul Scharre’s doctoral thesis that “widespread deployment of fully autonomous weapons would likely undermine crisis stability, escalation control and war termination because of the risk of unintended lethal engagements” (2020, 289).

Vertical and Horizontal Proliferation: Strategic Impact of Underdogs

Artificial intelligence is inherently subject to quick proliferation due to its nature as a dual-use technology, whereby developments in the commercial sector subsequently trickle up to the military realm (Rickli and Ienca 2021). Therefore, AI is a commercial endeavour developed in the private sector to satisfy business demands (Morgan et al. 2020, 13). Private corporations are some of the most aggressive investors in the technology, including tech giants such as Microsoft investing US$1 billion in OpenAI, an AI research Lab dedicated to researching artificial general intelligence (AGI), or Google having acquired Deepmind for more than US$500 million to conduct research in the same field (Shu 2014; Vincent, 2019). In the case of AI, proliferation is not only a function of state funding and development – as is traditionally the case for nuclear armament – but also a function of private sector demands, R&D, and production. Horowitz argues that military capabilities proliferate faster when there is “an underlying commercial demand and the unit costs are low” (Horowitz 2019, 777). Therefore, the horizontal proliferation of military AI is only likely to accelerate. As stated by Morgan et al. in a RAND corporation study, “we can reasonably expect most future military applications to be adaptations of technologies developed in the commercial sector” (Morgan et al. 2020, 13).

The US benefits from a first-mover advantage in AI, with China following close behind and closing the technology gap quickly (Johnson 2021, 370). However, military AI is unlikely to remain a two-horse race due to the dual-use nature of AI and because AI research and development is taking place globally. This entails that multiple states and non-state actors will have access to at least some degree of military AI. Indeed, while China and the US are far ahead in the field, worldwide economic interdependencies and linkages mean that AI innovations are quickly diffused around the globe. Michael Raska argues that the sixth (and current) revolution in military affairs (RMAs), the AI-RMA, differs from previous RMAs in the diffusion pattern of its technologies (Raska 2021, 469). In this latest AI-driven RMA, the benefits of technological innovation will not be reaped only by major powers (Raska 2021, 473). Middle powers, such as Singapore or Israel, are developing technologies to “advance their economic competitiveness, political influence, and status in the international arena” (Raska 2021, 473). These medium powers are developing niche capabilities and will become “critical influencers in shaping future security, economics and global norms in dual-use AI” (Johnson 2021, 370). Proliferation is thus unlikely to only take place horizontally between states, but vertically too, among non-state actors and even individuals. The fact that innovation is business-driven multiplies the number of actors who have access to disruptive technologies by lowering the bar for their adoption. Furthermore, unlike traditional military capabilities such as fighter jets, algorithms, the backbone of AI, can be replicated and proliferated at little cost.