AI is a real threat, but empirics prove that norms solve---legal apparatuses are key
Maas 19, Matthijs---Faculty of Law, Centre for International Law, Conflict and Crisis, University of Copenhagen, Copenhagen, Denmark; Center for the Governance of AI, Future of Humanity Institute, University of Oxford, Oxford, UK (“How viable is international arms control for military artificial intelligence? Three lessons from nuclear weapons,” Contemporary Security Policy, vol 40, no 3, 2019)
More often, norms—implicitly the “nuclear taboo” against the (first) use of nuclear weapons (Carranza, 2018), and explicitly the norms encoded by international legal instruments, including the Nuclear Non-Proliferation Treaty (NPT) and the Comprehensive Nuclear Test-Ban Treaty (CTBT)—appear to have served as a factor in constraining nuclear proliferation, though such norm enforcement is not without challenge (Knopf, 2018). Nonetheless, these global legal instruments can function, in part because they can provide shared normative frameworks that disseminate and promote non-proliferation norms, -interests and -identities at the domestic-political level, tipping the balance of domestic contestation towards coalitions seeking non-proliferation. While one might expect such effects of “norms capture” to be stronger in liberal societies than in non-liberal ones, some scholars have suggested that even in the latter case, state elites who are not accountable to their own publics simply come to internalize the normative characterization of a successful state as one that abides by its treaty commitments (Rublee, Bertsch, & Wiarda, 2009, p. 222), or at the very least, have incentives for compliance with international non-proliferation norms, to foster a reputation for reliability in the eyes of other states (Williamson, 2003, p. 81). Moreover, global international regimes—defined by Krasner (1982) as “sets of implicit or explicit principles, norms, rules and decision-making procedures around which actors’ expectations converge in a given area of international relations” (p. 2)—can serve as Schelling Points around which global society can coordinate multilateral collective sanctions or rewards (Müller & Schmidt, 2010).
To what extent have non-proliferation norms driven nuclear restraint? Of course, it can be hard to disentangle causal connections between membership in normative (legal) instruments such as the NPT and nuclear restraint, since such behavior could reflect existing policy preferences by the states. Some reviews of the nuclear (non-)proliferation records have suggested that “NPT membership and the NPT regime's norms have modest or marginal impacts on nuclear proliferation” (Jo & Gartzke, 2007, p. 185), though others found that, when accounting for states’ ex ante treaty commitment preferences, state ratification of the NPT treaty regime was “robustly associated with a lower likelihood of pursuing nuclear weapons” (Fuhrmann & Lupu, 2016, p. 530).
Intriguingly, while public norms seem to be able to strengthen the hands of (non-)proliferation coalitions, they do not seem to reliably shift state policymaking where these coalitions do not already exist in sufficient strength. For instance, in 1994 Ukraine chose to join the NPT and renounce its nuclear arsenal in spite of respectable Ukrainian public support for retaining the weapons (Sagan, 1996, p. 80). Conversely, in 1999 the U.S. Senate rejected the CTBT in the face of widespread U.S. public support.
As with all history, it can be hard to distill unambiguous causal chains; yet surveys of the distinct state rationales for nuclear proliferation and of –non-proliferation or the abandonment of ongoing nuclear programs (Garnett, 2012; Sagan, 1996, 2011; Solingen, 1994; van der Meer, 2011, 2014), suggests that, far from proliferation cascades fueled by security and strategic concerns, a far wider array of motives shaped these decision-making processes, with elements of all three models—security, domestic politics, and norms—playing different roles amongst different states, and often contributing to a decision to forego or abandon nuclear proliferation. For instance, Solingen has charted the role of ascendant liberalizing coalitions, in countries such as Taiwan, South Korea, and Argentina, in shifting towards nuclear restraint, because of the favorable impact of this decision on international trade, aid, technology and investment opportunities, as well as to reduce perceived wasteful budgets for such military programs (Solingen, 1994). The broad history of nuclear restraint suggests that, far from a foregone conclusion, arms races involving strategically appealing technologies can be slowed, channeled, or stopped. This suggests that halting, managing or containing military AI arms races is viable—and hints at a range of considerations for doing so.
In the first place, it suggests that security concerns are conducive but not decisive to arms races, and that a limited number of “first-mover” major powers may share an interest in supporting global legal regimes aimed at the non-proliferation of certain forms of military AI, such as cyber warfare systems, which might otherwise empower conventionally weaker (or non-state) rivals. Given that this group of “leading” states is initially small, bilateral agreements may suffice; however, the unevenly shared stakes in the technology may render eventual multilateral negotiations more difficult (cf. Picker, 2001).
In the second place, the domestic-politics model suggests that strengthening the hand of domestic coalitions pursuing the non-proliferation of AI weapons is one pathway towards shifting state decision-making away from pursuing more problematic categories of military AI, even in the face of clear national security interests. Of course, one caveat here concerns the fact that military AI may have far broader appeal than nuclear weapons did, such that it is harder to find domestic coalitions that are clearly opposed to its development in all cases. For instance, the strategic benefits of developing nuclear weapons are almost solely military and relatively discrete—a half-finished nuclear weapon is not even half as useful as a finished one, and the road from starting a nuclear program towards developing, not just a working weapon or small arsenal, but a credibly survivable and deliverable deterrent is long and potentially less useful (because more provocative to well-armed adversaries) than not initiating a nuclear breakout to begin with. In contrast, the benefits of pursuing military AI might be more linear and gradual, with intermediate advances in subfields (e.g., image recognition or drone swarming command and control) enabling not just immediate application to battlefield roles, but also economically productive spin-offs to civilian applications. These features, combined with the comparatively lower reputational costs, may make some forms of military AI more palatable. But not all. Exceptions may be found in high-performance adversarial contexts (such as cyber warfare or aerial warfare) where AI systems or platforms end up directly engaging with each other. In such cases, as Payne (2018a) has argued, “marginal quality might prove totally decisive” because “other things being equal, we can expect higher-quality AI to comprehensively defeat inferior rivals” (p. 24). In such domains, the incentives for parties to independently develop “second-rate” military AI capabilities might be lower. Conversely, where AI systems do not have to “fight their like” directly (e.g., logistics; facial recognition), second-best AI systems still offer military advantage, and could proliferate widely (Horowitz, 2018a, 2018b). This suggests that the precise appeal of military AI systems to different parties may be more complex—which offers openings for tailored engagement with domestic coalitions.
Thirdly, the “norms”-model suggests that, while policy-makers may pursue the development of AI in general because of its “symbolic” value as a marker of global scientific leadership, this may not transfer to the development of AI for military purposes. Instead, the degree to which military AI confers status may be mixed: Pursuing openly autonomous “killer robots” may indeed remain unappealing for states. For instance, over the past years, global surveys of public opinion show that in most countries (excepting India), majorities oppose the deployment of autonomous weapons (Open Roboethics Initiative, 2015; Roff, 2017). An opposition appears to be on the rise: an even more recent survey by Ipsos showed an increase in global opposition (from 56% to 61%) since 2017 (IPSOS, 2019). At the same time, other surveys have shown that public opposition to these weapons can be very context-dependent, and drops off if their usage is framed as being defensive and aimed at reducing casualties amongst friendly troops (Horowitz, 2016; West, 2018). In another U.S. survey, Americans generally expressed mixed support for the United States investing more in AI military capabilities, but also for the United States to cooperate with China to avoid the dangers of an AI arms race (Zhang & Dafoe, 2019, pp. 26–30).
Moreover, advocacy efforts might well be able to shift these public norms on military AI further—and thereby alter the reputational penalties and rewards of deploying new systems or for complying with restrictive global regulation, respectively. Indeed, it is important to recognize the considerable efforts that have been put into making “killer robots” normatively unpalatable, notably by movements such as the Campaign to Stop Killer Robots, a coalition of 89 NGOs from 50 countries (cf. Joshi, 2019). In fall 2018, both the European Parliament as well as United Nations Secretary General António Guterres called for bans on autonomous weapons (European Parliament, 2018; Guterres, 2018); and at present at least two dozen states are pursuing such a legally binding ban—although states such as the United States, United Kingdom, and Russia still explicitly oppose such an initiative (Joshi, 2019). Even if such public advocacy efforts have not (yet) produced a ban on autonomous weapons, this does not mean they have not already influenced the normative space around military AI.
However, to what extent will a specific opprobrium on the—important, but narrow—category of autonomous weapons transfer to other types of military AI? Indeed, beyond “killer robots,” it is unclear to what extent states will face a meaningful or strong “military AI taboo” with the same strength as the “nuclear taboo.” After all, the latter norm was possible and potent, because nuclear weapons are a “single” technology with a single discrete, publicly visible and viscerally horrifying use mode. This created a natural and unambiguous “red line” in usage, not to be crossed. Conversely, the deployment and use of AI in many (non-kinetic) military applications is already a fact, such that this Rubicon has been crossed. Moreover, the technology is moreover very heterogeneous, such that whereas visceral applications (e.g., “killer robots”) may generate public opprobrium and restrictive activism, more diffuse or less kinetic ones (e.g., logistics systems; capabilities to track missile submarines) may not. It would therefore be advisable that organizations pursuing bans of the technology, consider the degree to which framings of “killer robots” continue to correspond to developments in military AI, including other usages which are potentially unethical, unsafe, or destabilizing.
Finally, while public norms or activism against military AI may strengthen domestic political coalitions already opposed to these weapons, they alone are not always able to sway policymakers in the first place. A key route lies therefore in shaping policymakers’ norms (and indirectly the domestic political landscapes). This relies on the (top-down) norm-shaping influence exerted by global legal instruments and regimes, but also on the (bottom-up) institutionalization of norms by “elite entrepreneurship in norm change” (Lantis, 2018), and specifically through “epistemic communities” of expert groups.