Conceptualising modes of control

Our dependent variable is modes of state control over military innovation. States ultimately seek reliable access to security-relevant technology, while denying it to rivals. In sectors such as AI where private corporate experts play a major role in innovation, states have control over military innovation when they prevent shirking (i.e., subpar efforts) and slippage (i.e., undesired development or use of security-relevant technology) among private innovators. There are various ways whereby states may pursue these goals, ranging from insourcing the expertise for military innovation to nurturing good rapport with private corporate experts so that the latter identify with state interests.

To distinguish between varying modes of control, research on indirect governance has highlighted that ‘governors’ may employ hierarchical or non-hierarchical modes of control for their ‘intermediaries’.Footnote ¹⁷ Hierarchical control involves hard, top-down oversight. It works with (threats of) coercion and sanctions. By contrast, non-hierarchical control relies on mutually beneficial collaboration, political, economic, or ideational inducements, soft (self-)regulation,Footnote ¹⁸ and/or nudging. It works with positive incentives and aims at intermediaries’ voluntary alignment with the governor’s goals.

Modes of state control may also vary with regard to the policy instrument employed to exercise control. Drawing on concepts from public policy research,Footnote ¹⁹ which have recently been introduced to the study of security policy-making,Footnote ²⁰ we distinguish between capacities- and rules-based modes. ‘Positive security states’ build and employ state capacities,Footnote ²¹ including administrative capacities and coercive capacities,Footnote ²² to ensure state control over military innovation. They spend state funding on building and maintaining internal capacities for military innovation within the state apparatus. Unlike positive security states, ‘regulatory security states’ draw on rules to steer and control military innovation performed by non-state actors.Footnote ²³ They spend state funding on buying security-relevant goods and services from private corporate experts. They seek to bring the latter’s behaviour in line with their own interests by means of both general (legislative or administrative) rules, which universally apply within their jurisdiction, and specific contractual rules governing the transactions between state buyers and private providers of goods and services. Contractual funding and the specific rules that come with it work as a mode of control by stipulating the obligations of private corporate experts and making (full and continued) payments contingent on their fulfilling the contractual obligations.

Combining the distinctions between hierarchical vs. non-hierarchical and capacity-based vs. rules-based approaches, we distinguish between four modes of control (see Table 1):

• • Hard regulation designates state reliance on the indirect provision of military innovation by private corporate experts and the promulgation of rules that stipulate, either in laws, regulations, or contracts, precise and binding obligations as well as sanctions in case of non-compliance. It also implies the build-up and use of hierarchical monitoring and enforcement procedures.

  Table 1.

  Conceptualising modes of state control.

  

• • Soft regulation encourages and endorses voluntary guidelines or codes of conduct without state enforcement. Specific rules in contractual provisions are limited in number and remain flexible with regard to their implementation.

• • Centralised capacity-building grants state policymakers hierarchical leverage over private corporate experts by making their contributions dispensable. In the extreme, it allows state institutions to autonomously make and implement their policies.Footnote ²⁴

• • Collaborative capacity-building refers to the pooling of state capacities with private corporate experts, complementing rather than substituting for private capacities. It aims to bring private corporate experts’ behaviour in line with state interests by creating incentives for sustained mutually beneficial cooperation.

These are ideal-typical distinctions, and modes may overlap in practice. Nonetheless, we can assess which mode is prevailing at a particular point in time, within a specific country, and in a particular field.

Explaining modes of control: The competence–control trade-off

Competence–control theory assumes that wherever governors rely on intermediaries, they face a trade-off.Footnote ²⁵ Relying on highly competent intermediaries, who are rich in expertise, operational capacities, or legitimacy, enhances the effectiveness of governance. But highly competent intermediaries are difficult to control. Enhancing control reduces the benefits from relying on intermediaries because it compromises intermediaries’ competence, their independent expertise, their efficient use of operational capacities, or their legitimacy. Incompetent intermediaries pose fewer control problems. But governors want their intermediaries to be competent because this enhances governance performance, and, at least under many circumstances, intermediaries also have incentives to increase their competence, which will inadvertently be hindered by hard, hierarchical state control. Governors have to navigate a competence–control (CC) trade-off. Maximising both competence and control remains elusive.Footnote ²⁶

We expect this trade-off to be pervasive in military innovation driven by private corporate experts.Footnote ²⁷ Where know-how and epistemic authority are concentrated within the private sector, reliance on their unconstrained expertise promises more effective military innovation. But this comes with a loss of state control. Uncontrolled private corporate experts may (ab)use their leeway to refrain from developing desired technologies; they may develop undesired technologies; or they may put technologies to undesired uses, including their diffusion to rivalling states. Lack of control also runs counter to standards of democratic accountability for decisions over the use of force. Thus, policymakers have incentives to foster private competence and impose state control.

But the intensity of the CC trade-off and the prevalence of competence or control imperatives vary across instances of military innovation, suggesting different control modes.

When the CC trade-off is mild, as the state’s dependence on private competence is limited, the private sector does not have a clear edge over state institutions, and state concerns about control are moderate, policymakers will engage in collaborative capacity-building. In such settings, collaborative capacity-building promises to enhance overall capacities. Moreover, it endows the state with sufficient non-hierarchical avenues of influence on private actors by giving the latter a stake in continued collaboration, inducing behavioural alignment with state interests.

In dynamic contexts that call for prioritising private competence, state policymakers will seek to nurture the unrestrained competence of private corporate experts and choose soft regulation. To be sure, in some instances, collaborative capacity-building rather than soft regulation may seem most conducive to the development of private competence. This is especially so when the challenge for the state is to get an industry and its competence for innovation going, while market incentives for private innovation are lacking. But, first, for dual-use technologies, including AI, there are usually strong business incentives for market-based private innovation. Second, soft regulation that leaves business unconstrained generally enhances the speed of competence growth as it does not require lengthy public–private coordination. Finally, the more dynamic the environment for military innovation, the greater the comparative advantage of unleashed business innovation and the fewer assets the state would have to offer, through collaborative capacity-building, to industry. Accordingly, in such dynamic settings that call for unrestrained private competence, state policymakers will opt for soft regulation, contenting themselves with broad guidelines or encouraging private self-regulation. Public monitoring of compliance will be weak, and enforcement will be absent, compromising state control.

In settings that lead state policymakers to prioritise state control, they will focus on centralised state capacity-building to substitute for private corporate experts’ capacities. Instead of indirect governance, the state focuses on in-house, direct, and command-and-control-based governance that renders private contributions dispensable. This maximises the state’s hierarchical leverage. Yet, it implies foregoing the benefits of relying on the competence of specialised private corporate experts.

If there is an intense CC trade-off, as the demand for both competence and control is strong, state policymakers will opt for hard regulation. They will promulgate a large number of strict rules for private corporate experts. There will be strong oversight bodies and sanctions imposed for rule violations. While falling short of the control achieved by centralised capacity-building, hard regulation can provide policymakers with significant hierarchical control. The flip side is that (access to) private corporate experts’ competence also suffers – though less than with centralised state capacity-building.

Technological change and geopolitical pressures as drivers of the CC trade-off

CC theory helps us grasp oscillation between different modes of state control because it suggests pervasive dissatisfaction with any chosen mode and, consequently, recurring institutional reforms. But the intensity of the CC trade-off varies, and CC theory does not specify under what conditions policymakers will prioritise either competence or control. To capture the outcomes of recurring dissatisfaction-driven reforms, we draw on research on transformations in security policy-making. We highlight two drivers – technological change and geopolitical competition – that condition the intensity of the CC trade-off and state policymakers’ propensity to prioritise competence or control.

Technological change has historically challenged existing patterns of how to provide security.Footnote ²⁸ We distinguish between strong and moderate technological change in terms of the speed and nature of technological developments. When technological innovation occurs rapidly and in a disruptive way, that is, through major breakthroughs, technological change is strong. In contrast, when technological development occurs slowly and incrementally, that is, through the gradual adaptation of existing technologies, technological change is moderate. We assume general-purpose technologies, such as AI, to be associated with stronger technological change,Footnote ²⁹ making the stalling of technological innovation unlikely. Yet, we still expect variation in the more or less rapid speed and the more or less disruptive nature of innovation in such technologies.

Ceteris paribus, strong technological change contributes to a blurring of the civilian–military divideFootnote ³⁰ and increases the relevance of technical experts for security policy-making.Footnote ³¹ When technological change is strong and its main sources come from the private sector, generalist bureaucracies will lack the technical knowledge to meet the challenges posed by technological change. While under certain conditions the state may be important to jump-start private competence for innovation, the stronger the technological change, the greater becomes the comparative advantage of unconstrained private innovation. Policymakers’ desire to control private corporate experts may well intensify as the amount of technological innovation produced by the private sector increases. Yet, their ability to control private experts through hierarchical and capacities-based means without compromising their much-needed competence declines. Therefore, state actors will be inclined to opt for rules-based modes of control, and more precisely, the non-hierarchical approach of soft regulation.

Geopolitical competition also shapes how states organise the provision of security policies.Footnote ³² Following a neoclassical realist view, which stresses material power but also includes domestic actors’ perceptions,Footnote ³³ we conceive of geopolitical competition as a function of the distribution of military capabilities among leading powers and/or between the leading power and its main challenger(s), while also taking into account rivalling states’ perceptions of threat, shaped by the perceived revisionist or status-quo oriented intentions of their counterparts. Ceteris paribus, strong geopolitical competition suggests a reassertion of state control over military innovation. Under strong geopolitical competition, military innovation becomes a strategic priority. Policymakers will be ready to make vast investments in capacity-building for military innovation, while their tolerance for private corporate experts’ autonomy will be reduced. Private corporate experts good at technological advancement are less proficient in geopolitical strategising. Geopolitical priorities may even be at odds with their interests in maximising business opportunities, including with rivals of ‘their’ home country. Strong geopolitical competition thus incentivises policymakers to pursue a capacities-based and hierarchical mode of control.

There is an intricate relationship between technological change and geopolitical competition. Technological innovation helps states gain an edge over geopolitical rivals. Accordingly, geopolitical competition induces states to promote technological change in an effort to succeed in the international power race.Footnote ³⁴ Yet, with dual-use technologies, the link between technological change and geopolitical competition is quite diffuse. Far from any change in dual-use, technology eventually impacts on geopolitical competition, and it is often hard for states to foresee the military implications of change in dual-use technologies. Moreover, for private commercial firms, their efforts at technological innovation are primarily contingent not on geopolitical pressures but rather on business opportunities. Most importantly, geopolitical competition and technological change suggest different modes of state control: Technological change calls for expertocratic governance relying on unconstrained non-state experts and tech companies, whereas geopolitical competition suggests hierarchical grand strategising by policymakers, militaries, and generalist bureaucrats.

We expect technological change and geopolitical competition to vary across states, fields, and over time. State actors may perceive technological change and geopolitical competition as – absolutely or relatively – more or less salient. We claim that the CC trade-off and the chosen control modes vary, depending on the relative strength of both drivers. If both technological change and geopolitical competition are moderate, the CC trade-off is mild, and state policymakers will engage in collaborative capacity-building. If technological change is strong but geopolitical competition is moderate, state policymakers will opt for soft regulation. If technological change is moderate, while geopolitical competition is strong, state policymakers will pursue centralised capacity-building. If both technological change and geopolitical competition are strong, state policymakers will pursue hard regulation (see Table 2).

Table 2.

Explaining modes of control: varying drivers of the CC trade-off.

Yet, neither of these modes represents a stable institutional equilibrium. As maximising both private competence and state control is appealing but elusive, there are pervasive incentives for institutional re-design. This does not rule out the persistence of some institutions; yet, it implies recurring reforms of control modes, pointing in contradictory directions.

Research design and operationalisation

We apply this theory to explain modes of state control over military AI innovation in the United States. Its democratic procedures for security policy-making, capable military–bureaucratic apparatus, extensive experience in regulating big tech companies, vast investments in the spin-on of private AI innovations, and early AI institution-building efforts render the United States a security state that faces incentives to promote both (private) competence and (state) control while possessing the prerequisites for either of the two strategies. In other words, the United States should be subject to a pronounced CC trade-off; it is a typical case, suitable for empirically demonstrating a new theoretical argument and mechanism.Footnote ³⁵ This case selection limits the generalisability of our findings. Yet, we expect the CC trade-off and our theory to apply to other advanced democratic military powers too, as (long as) they have incentives for both ensuring state control and enlisting private competence while possessing enough potential for capacity-building and regulatory capability so that their choice of control modes is not a foregone conclusion.

We focus on autonomous weapon systems (AWS), which, ‘once activated, can select and engage targets without further intervention by a human operator’.Footnote ³⁶ Again, AWS are a typical case for the CC trade-off, lending itself to in-depth process-tracing. There should be strong incentives for both (fostering) private competence and (ensuring) state control: Private AWS experts have a competence advantage vis-à-vis the state. But there are also strong incentives to seek state control over AWS innovation due to their relevance and risks.

We test a set of theoretical expectations: First, we expect to observe a correlation over time between constellations of drivers, the prevailing imperative in the CC trade-off, and chosen control modes (see Table 2). If we do not find such a correlation, we would consider our theory falsified. We measure geopolitical competition and technological change through objectivist and perception-based indicators. To assess geopolitical competition, we first focus on the relative distribution of military capacities in the US–China power rivalry: When military spending in general as well as for AI specifically gets more even (in favour of China), we take this as an indicator of intensifying geopolitical competition.Footnote ³⁷ Moreover, we take an increasing perception of Chinese revisionist ambitions in the US government discourse as an indication of heightened competition. Similarly, we assess the strength of technological change through AI performance, as measured by computing power used for training,Footnote ³⁸ and perceived technological breakthroughs, captured by global patenting activityFootnote ³⁹ as well as the identification of ‘revolutionary’ technology advancements in policy reports such as the Stanford AI Index Report.Footnote ⁴⁰ While on an individual firm level, filing for patents may also reflect a desire to protect innovation in highly competitive contexts, on the macro level, global patenting activity is widely regarded as a useful indicator of technological change in a certain field.Footnote ⁴¹ Therefore, an acceleration of patenting activity and prominent references to technological game changers together indicate strong technological change.

Beyond correlational evidence, we first expect a pattern of recurrent institutional reforms, which point in different, if not contradictory, directions. Relatedly, we expect to see that any reform produces some dissatisfaction, which prepares the ground for another round of reform. Second, we check for process evidence by identifying ‘empirical fingerprints’ in official statements and documents drawing connections between a) the contextual pressure(s) policymakers are facing and the prioritisation of private competence or state control, and b) the expressed prioritisation of competence or control and the design of control modes. If we find such evidence, our confidence in our theory will be considerably strengthened.

From collaborative capacity-building (1990s) to soft regulation (2000s–mid 2010s)

From the 1990s to the mid-2010s, innovation in military AI – and specifically AWS – evolved from a nascent field of R&D to a field of high military and commercial relevance. This also entailed a change in the state’s governance of innovation, which can be described as a shift from collaborative capacity-building to soft regulation.

In the 1990s, both geopolitical competition and technological change were moderate, creating a relatively mild CC trade-off. Geopolitical competition was limited compared to later (and earlier) decades, and US defence spending gradually decreased. Nevertheless, the country still considerably outspent, by about thirteen times, its (future) main competitor, China.Footnote ⁴² In the 1990s, there was strong hope among many US policymakers that a rising China could be integrated into a US-led global order.Footnote ⁴³ Technological change was also moderate. The 1990s started off as the last ‘AI winter’,Footnote ⁴⁴ characterised by technological stagnation and funding cutbacks, neared its end. Despite some important advances,Footnote ⁴⁵ these developments only laid the groundwork for more dynamic progress in the 2000s. As OECD data indicate, global AI-related patenting would pick up only in the late 1990s and, in particular, the 2000s.Footnote ⁴⁶

As expected by our theory, the resulting mild CC trade-off translated into a control mode characterised by collaborative capacity-building, where public and private capacity(-building) coexisted and was, frequently, pooled. In the 1990s, policymakers, in particular via the Defense Advanced Research Projects Agency (DARPA), were the drivers of AI-related innovation, in collaboration with academic and private experts. DARPA was critical in stimulating the establishment of AI as a field of academic research, whose beginnings are usually dated back to the Dartmouth Conference of 1959.Footnote ⁴⁷ Its Strategic Computing Initiative (SCI, 1983–93) entailed big investments in public and private R&D aimed at collaborative capacity-building. However, the SCI was discontinued as it failed to achieve its ambitious goal of creating artificial general intelligence.Footnote ⁴⁸ Similar to these general developments in AI innovation, policymakers sought to advance AWS through collaboration with the private sector. But again, the track record of these attempts was mixed. One of the first attempts at developing fully autonomous loitering weapons,Footnote ⁴⁹ the Low-Cost Autonomous Attack System (LOCAAS) project was abandoned.Footnote ⁵⁰ Doubts about the effectiveness of collaborative capacity-building increased.

In addition, technological change accelerated in the 2000s. The performance of AI models increased, with the computing power used in training notable AI models doubling approximately every eighteen to twenty-four months.Footnote ⁵¹ Moreover, from 2000 to 2010, the average yearly number of global AI patents almost tripled.Footnote ⁵² The 2000s saw a variety of breakthroughs, including the rise of deep learning,Footnote ⁵³ which increased the commercial and military value of AI. The leadership role of private experts in driving cutting-edge technology development became entrenched. These included traditional defence contractors such as Boeing, Lockheed Martin, Northrop Grumman, and Raytheon (now RTX Corporation), and big tech companies, such as Google, Microsoft, and Amazon.Footnote ⁵⁴ Specialised AI firms, such as Databricks, C3 AI, and Palantir, proliferated, and the competence advantage of private (commercial) actors vis-à-vis the public sector became apparent.Footnote ⁵⁵

While innovation in AWS was largely driven by advances in AI, the boom of the field in the 2000s was also fuelled by the proliferation of drone technology. In the context of US interventions in the global ‘War on Terror’,Footnote ⁵⁶ US drone arsenals saw a forty-fold increase from 2002 to 2010.Footnote ⁵⁷ The availability of (still remotely controlled) drone technology raised interest in ‘taking the human out of the loop’.Footnote ⁵⁸

This incentivised policymakers to foster private competence. Geopolitical competition was still moderate, as military spendingFootnote ⁵⁹ and the perception of China as a primarily economic competitor rather than a security rivalFootnote ⁶⁰ suggest. Therefore, the urge to exploit new technological possibilities by spurring growing private ingenuity ‘beyond the traditional performer base’ prevailed.Footnote ⁶¹ Competence was temporarily prioritised over control. A mode of governance described as soft regulation ensued.

In the broader military AI field, the state initially acted as an important stimulator of R&D. DARPA-funded private research in different AI subfields facilitated the breakthroughs in deep learning in the 2010s.Footnote ⁶² Several DARPA programs, such as the SyNAPSE Program on brain-inspired ‘neuromorphic’ computingFootnote ⁶³ and the Autonomous Robotic Manipulation Program,Footnote ⁶⁴ made large investments in the private sector to help build up AI competence. This stimulation and enlistment of private competence was supported by institutional reforms, such as the establishment of the Defense Innovation Unit Experimental (DIUx, later Defense Innovation Unit, DIU). As ‘the only DoD organisation focused on accelerating the adoption of commercial and dual-use technology to solve operational challenges at speed and scale’,Footnote ⁶⁵ it served to improve the relations between the Pentagon and non-defence tech companies, reflecting the recognition that AI competence had become concentrated within the private sector.

In line with a soft (self-)regulation approach, state oversight was limited. Setting rules for AI innovation was mostly left to the private experts performing the R&D. Early industry attempts at self-regulation included the establishment of an internal AI ethics board when Google acquired DeepMind in 2014,Footnote ⁶⁶ and the establishment of OpenAI as a non-profit company, which stimulated a discourse about principles of AI R&D such as openness, transparency, and respect for intellectual property rights.Footnote ⁶⁷ Overall, the public laissez-faire approach incrementally gave way to initiatives endorsing industry self-governance in the mid-2010s.

Harnessing private competence for the DoD and promoting industry self-regulation were also inscribed in the mandates of new bodies such as the Defense Innovation Board (DIB, est. in 2016), which gave representatives of commercial AI companies a leadership role and was first chaired by Eric Schmidt, former Executive Chairman of Alphabet.Footnote ⁶⁸ Industry initiatives included the Partnership on AI (2016), cofounded by big US tech companies,Footnote ⁶⁹ and company-specific guidelines such as Google’s AI Principles.Footnote ⁷⁰ The DIB’s own Recommendations on the Ethical Use of Artificial Intelligence by the Department of Defense were drafted in collaboration with experts from OpenAI, Microsoft, Facebook, and Google, among others.Footnote ⁷¹

Efforts to incentivise increasingly competent private corporate experts also intensified in the subfield of AWS, as the benefits of unconstrained private innovation capacity became apparent. State regulation and interference in AWS innovation were limited and soft. Three DARPA GRAND Challenges, organised from 2004 to 2007 and offering between 10 and 20 million USD in prize money, were particularly influential. They demonstrated the feasibility of autonomous movement in difficult environments and gave a boost to the development and commercialisation of autonomous technologies.Footnote ⁷² The growth in largely unconstrained private competence for AWS innovation was accompanied by a loss of state control. Despite DARPA’s critical role in stimulating early research, the US security state lost its role as the global driver of innovation in the field to the private sector, and, within a matter of years, the field of autonomous vehicle development would become dominated by private actors from the automotive and IT industries.Footnote ⁷³

Several R&D projects aimed at enlisting private sector competence fell short of their ambitious goals and were given up. Prominent examples include the Joint Unmanned Combat Air Systems (J-UCAS) programFootnote ⁷⁴ and the Army’s modernisation program on Future Combat Systems (FCS), which aimed at developing teams of networked unmanned and manned vehicles by heavily relying on industry subcontractors. The FCS program was restructured multiple times and eventually cancelled. A RAND evaluation of the program identified significant management and planning mistakes and found that it was predicated on unrealistic requirements and ‘significant leaps in technology’. This caused serious doubts about the military’s ability to make effective use of private innovation capacities.Footnote ⁷⁵ A key point of criticism concerned the non-traditional partner-like relationship of the Army with its main contractor, Boeing, which produced significant oversight problems for DoD leadership, ‘especially when the government is disadvantaged in terms of workforce and skills’.Footnote ⁷⁶

Thus, towards the end of the second – soft regulation – phase, the opportunity costs of this approach had become obvious. In particular, concerns about a lack of state control over increasingly powerful private experts and the state’s inability to translate private competence into military capacities of the state emerged. This created incentives for another round of reforms in the second half of the 2010s.

Centralised capacity-building (mid–late 2010s)

When slowed-down technological change was increasingly overshadowed, throughout the 2010s, by strong competition between the United States and China,Footnote ⁷⁷ a pronounced geo-politicisation of AI drove a more hierarchical approach to controlling military AI innovation in the second half of the decade. The mode of control shifted towards attempts at centralised capacity-building to enhance state control over private corporate experts and military AI innovation.

While technological change certainly did not come to a halt, it became more moderate and its dynamic temporarily slowed down compared to previous and subsequent phases: The number of global AI patents, which had notably increased in the previous period, remained at a plateau during most of this period, and even dropped around the year 2016, before a drastic increase would usher in the beginning of the fourth phase of booming AI R&D in the 2020s.Footnote ⁷⁸ Regarding computing power usage, the trend observable in previous decades – a constant growth according to Moore’s law – continued. This stands in stark contrast to the subsequent phase, which would be perceived as a period of historically unprecedented speed and magnitude in AI progress.Footnote ⁷⁹ These broader technological trends were mirrored in the political discourse, where a strong increase in geopolitical competition moved to the forefront while the excitement about technological change subsided.

In the 2010s, the gap between US and Chinese defence spending diminished due to a significant growth in Chinese spending: In 2010, the US spent approximately seven times more than China, whereas merely five years later, it only outspent its competitor by a factor of 3.7.Footnote ⁸⁰ While reliable data is scarce, estimates suggest that United States and Chinese spending on military AI reached similar levels in the mid to late 2010s.Footnote ⁸¹

In the United States, where, previously, enthusiasm about the potential of AI had dominated, these shifts in the international distribution of material capabilities were accompanied by a growing fear of falling behind an increasingly revisionist China and a new arms and technology race.Footnote ⁸² This was exacerbated by the publication of China’s ambition to become the leader in AI by 2030, as stated in its New Generation AI Development Plan.Footnote ⁸³ US policymakers and military officials likened the increasing US–China competition to the Cold War context, criticised the lack of a coordinated approach to military AI innovation, and called for more – and more centralised – capacity-building to avoid a new ‘Sputnik moment’.Footnote ⁸⁴

Hence, due to heightened geopolitical competition, combined with mounting dissatisfaction with the soft regulation mode of the previous phase, state control over military innovation was given precedence over (fostering) private competence. Centralised capacity-building became the dominant mode of governance. Through several reforms, new strategies, bodies, and policies were created to nurture AI capacities within the security state. Most importantly, the DoD’s Defense Innovation Initiative bluntly cited geopolitical competition as a rationale for increasing state capacities in military AI to ‘sustain and advance America’s military dominance for the 21st century’.Footnote ⁸⁵ This Third Offset Strategy stood in the tradition of previous offset strategies aimed at maintaining superiority during the Cold War by investing in promising technologies. While it also responded to the diminishing role of the DoD in driving military innovation by suggesting new ways of collaborating with the private sector,Footnote ⁸⁶ its focus was on strengthening the broader innovation capacity across the DoD and the military services. Hence, internal research agencies were the primary recipients of increased R&D funding.Footnote ⁸⁷ This was accompanied by the development of new operational concepts and other measures to improve innovation management.Footnote ⁸⁸ In sum, the initiative represents a top-down effort not only to acquire particular technologies but also to ‘change the way we innovate, operate, and do business’.Footnote ⁸⁹

At the level of military AI, the expansion of in-house R&D was accompanied by institutional reforms, most importantly the establishment of the Joint Artificial Intelligence Center (JAIC, 2018) to coordinate AI efforts across the DoD, accelerate AI adoption, and educate DoD staff on AI.Footnote ⁹⁰ The military services also established governance bodies to advance capacity-building, such as the Air Force AI Accelerator or the Army AI Task Force.Footnote ⁹¹

Concerning AWS, various strategy papers discussed options for enhancing AWS capabilities ‘to help ensure overmatch against increasingly capable enemies’.Footnote ⁹² DoD Directive 3000.09, titled ‘Autonomy in Weapons Systems’, laid out policies and assigned responsibilities for their development and use.Footnote ⁹³ Additionally, with the Third Offset Strategy, autonomy-dedicated research budgets of the military services significantly increased in the second half of the 2010s.Footnote ⁹⁴

The Navy’s UCLASS program illustrates both the aspirations for centralised capacity-building and the challenges of implementing them: Initial plans foresaw the development of a semi-autonomous combat drone that could carry large payloads. However, Navy requirements underwent several significant alterations, particularly a reduction of strike capabilities. This caused strong criticism from Congress, the Government Accountability Office, and suppliers criticising the repeated changes of course, after they had already made significant investments.Footnote ⁹⁵ Eventually, efforts were redirected to the development of a drone used for refuelling to expand the range of manned fighters.Footnote ⁹⁶

As predicted by our theory, the results of this centralised capacity-building approach were widely perceived as unsatisfactory, and a debate about innovation obstacles in the security state unfolded. Commonly cited problems included resistance to change and lack of expertise among the staff, as the private sector was a ‘more effective competitor for talent’.Footnote ⁹⁷ The mandates of newly created AI bodies were perceived as unclear or limited. The state appeared to have failed at substituting private competence with state capacities, while at the same time struggling to enlist private expertise (beyond that of traditional defence contractors). This was ascribed to a hardware-centred innovation system, complex acquisition processes, limited data availability, and intellectual property concerns.Footnote ⁹⁸ This dissatisfaction contributed to the momentum for yet another round of reforms.

Hard regulation (early 2020s)

In the early 2020s, geopolitical competition remained strong, while technological change accelerated significantly. The co-presence of these two strong drivers exacerbated the CC trade-off. Coupled with the dissatisfaction about centralised capacity-building efforts, it led to a shift towards hard regulation of heavily enlisted private corporate experts.

On the one hand, the perception of US–China rivalry reached a new urgency. While the gap between US and Chinese military expenditures had barely changed since 2015,Footnote ⁹⁹ US fears of Chinese revisionism had grown. China was seen as ‘the only competitor with both the intent to reshape the international order and, increasingly, the economic, diplomatic, military, and technological power to do it’.Footnote ¹⁰⁰ While, in the previous phase, China had primarily been depicted as a technological rival, it was now perceived as an immediate threat to US domestic security,Footnote ¹⁰¹ for example, through attacks on critical infrastructure. State control over military AI innovation to limit safety and security risks associated with the diffusion of AI technology, therefore, became a critical concern.

On the other hand, from 2021 to 2022 alone, the number of AI patents increased by 62.7 per cent.Footnote ¹⁰² By the early 2020s, AI models across all domains had become much more sophisticated, and computing power usage in training was doubling every 3.4 months.Footnote ¹⁰³ As this development further strengthened the advantage of private industry with more resources at its disposal, the perceived technological breakthroughs primarily originated from the commercial sector. For instance, in 2023, over 70 per cent of all ‘foundation models’ (trained on large datasets and adaptable to a large variety of tasks), originated from industry, with US big tech companies at the forefront.Footnote ¹⁰⁴ In the subfield of AWS, increased sophistication and availability of commercially developed drone technology created additional incentives to enhance collaboration with commercial manufacturers.Footnote ¹⁰⁵

The National Security Commission on AIFootnote ¹⁰⁶ explicitly highlighted this dual challenge of improving collaboration with the private sector, on the one hand, and, on the other hand, avoiding excessive reliance on uncontrolled private actors. This resulted in the hard regulation of private corporate experts.

AI-related government spending to contractors strongly increased in the early 2020s,Footnote ¹⁰⁷ and strategy papers re-emphasised the importance of collaborating with private corporate experts. At the same time, they stressed the necessity of state oversight to ensure that private innovation served the goals of technological modernisation and the maintenance of military superiority.Footnote ¹⁰⁸ Justified with a strong national security rationale, efforts to increase state control included improved oversight over US private corporate experts as well as regulatory restrictions imposed on domestic and international technology firms. For that purpose, in ‘one of the most extensive technological blockades ever attempted’, the government restricted Chinese access to critical AI technology and banned the export of AI chips to China, causing considerable criticism from US producers, who feared competitive disadvantages.Footnote ¹⁰⁹ This was accompanied by a new focus in the strategic orientation of existing institutions: The DIU’s importance as a ‘gateway’ to commercial innovation significantly increased due to a considerable budget increase,Footnote ¹¹⁰ and a newly granted authority to award other transactions (OT) agreements,Footnote ¹¹¹ making it a critical contracting authority. Similarly, DARPA sought partnerships with the private sector in its two major AI investment campaigns, the AI NEXT CampaignFootnote ¹¹² and the AI ForwardFootnote ¹¹³ initiative, while shoring up its oversight activities.

Concerning AWS, extensive enlistment of private expertise was accompanied by a growing quantity and quality of regulation. For instance, the DIU’s Blue Unmanned Aerial Systems program developed a vetted roster of drones and relevant components to allow for more efficient acquisition while ensuring that manufacturers met military needs and safety standards. This way, policymakers sought to improve not only the adoption of private expertise but also the control over it, in an attempt to counter risks such as data breaches or unauthorised access.Footnote ¹¹⁴ Additionally, new regulation aimed at increasing the reliability of technology by enhancing validation and verification (V&V) processes for drones operated by the DoD.Footnote ¹¹⁵

Efforts to increase control were not restricted to the domestic industry. While drones produced by Da-Jiang Innovations (DJI) had before been widely used by government agencies and commercial actors,Footnote ¹¹⁶ their access to the US airspace now raised fears of transmission of sensitive information to China.Footnote ¹¹⁷ Chinese companies, therefore, were the main target of new legislation prohibiting government agencies from buying drone technology manufactured in ‘adversary’ countries.Footnote ¹¹⁸

Thus, driven by strong technological change and strong geopolitical competition, reform efforts served the dual goal of improving the use of private competence, while, at the same time, increasing state control over the (domestic and foreign) actors driving innovation. It is too early to tell whether hard regulation will succeed in balancing competence and control and whether this mode of control will prove more stable than previous ones. Our theory raises significant doubts about this. Figure 1 summarises the predominant drivers and resulting modes of governance in the field of AWS within the four phases discussed above.

Figure 1.

Drivers of CC trade-off and modes of control over AWS innovation.

Alternative explanations

While our theory captures shifting modes of control, an alternative institutional field explanation might hold that the frequent institutional change is hardly puzzling. As path-dependence,Footnote ¹¹⁹ vested interests,Footnote ¹²⁰ and veto pointsFootnote ¹²¹ are less pronounced in a young institutional field such as military AI, institutions can change with relative ease. We agree that the unsettled character of the field has facilitated reforms. Yet, we emphasise that there has not only been a lot of institutional change in a relatively short timeframe, which might indeed be expected for an unsettled field. Rather, the field is characterised by recurring dissatisfaction-driven and partly contradictory institutional reforms. Our theory uniquely captures this particular pattern of reforms.

An explanation centred on party politics and leaders’ belief systems might claimFootnote ¹²² that the chosen modes of control reflect the preferences and ideological inclinations of parties and individuals in government. We agree that party-based and individual preconceptions of the appropriate relationships between the state and the market might shape the politics of military innovation. Yet, empirical patterns of varying control modes, which reflect more or less business-friendly and more or less statist approaches, largely contradict the expectations suggested by these theories. While in the 2000s, stimulating the growth of the private industry, coupled with regulatory laissez-faire, was in line with the market-friendly orientation of the Republican G. W. Bush administration, this approach continued under the first administration of Democrat Barack Obama (2009–13). Only during Obama’s second term (2013–17) did a reorientation towards centralised capacity-building emerge. Against theoretical expectations, it reached its heyday under the Republican Trump administration, before heavy reliance on private industry innovation, coupled with hard regulation, was pursued under Democrat Joe Biden in the early 2020s. Shifts in control modes correspond much better to changes in geopolitical competition and technological change than to electoral cycles and varying political leadership.

Yet another explanation might highlight the impact of business interests. According to this perspective,Footnote ¹²³ private industry actors will lobby for policies that give them key privileges vis-à-vis international competitors. Proponents of this view might hold that governments will seek to support domestic industry by introducing more restrictive control modes in periods when domestic firms are falling behind international competitors. While we agree that some control measures, including import restrictions for Chinese drones, directly benefited US producers, there are several problems with this explanation: First, the competitive position of US firms in the broader AI/AWS field has not been correlated with the variation in control modes. While Chinese competitors have caught up over time, commercial US technology companies have remained at the forefront of R&D. Hence, the introduction of harder control modes barely seems to be a state response to save a faltering industry. Second, as patenting data indicates, Chinese competitors in the AI/AWS field have, to some extent, caught up in the last decade. Yet, it has been in this very period that the United States has shifted from an approach prioritising control (phase III) to one that sought to re-balance competence and control considerations (phase IV). Third, the introduction of hard regulation has aimed at strengthening state control over international and domestic private corporate experts. The tightening of export control measures to the detriment of US AI firms is a case in point.