DOES THE AI RACE PRODUCE SECURITY

Artificial intelligence (AI) has become central to contemporary security discourse. Across United States policy documents, strategic reports, and congressional testimony, AI is framed not merely as a driver of economic growth or administrative efficiency, but as a decisive factor in national survival. Falling behind in the development and deployment of advanced AI systems is repeatedly presented as a threat to military superiority, economic competitiveness, and even democratic order itself (National Security Commission on Artificial Intelligence 2021; White House 2025). This framing has produced a dominant conception of security structured around the metaphor of an “AI arms race,” in which speed, scale, and technological leadership are treated as prerequisites for security.

This essay evaluates the plausibility of that conception. Rather than focusing on how AI has been securitised as a discursive move, it asks whether the security practices authorised by the arms race framing actually produce security in practice. Drawing on Critical Security Studies, the analysis assesses security not in terms of intent or rhetoric, but through its material effects, political consequences, and distributional outcomes (Bigo 2002; de Goede 2017). While competitive acceleration is presented as necessary to avoid existential vulnerability, it remains unclear whether such practices reduce insecurity or instead generate new and uneven forms of risk.

The central argument advanced here is that the AI arms race constitutes an implausible conception of security. Although framed as essential to national protection, it misidentifies the primary object of security by privileging state capacity and corporate competitiveness over societal safety and democratic accountability. Moreover, the practices it legitimates (accelerated deployment, deregulation, and state–industry fusion) tend to amplify systemic risks while redistributing insecurity across societies and globally (AI Now Institute 2025).

The essay proceeds by outlining the dominant conception of AI security, examining how it is operationalised in practice, and evaluating its effects. It then analyses the distributional consequences of AI securitisation before considering how AI security might be rethought beyond arms-race logic.

1The Dominant Conception of Security: AI as a Strategic Race

Contemporary AI policy is governed by a competitive, state-centric conception of security in which technological leadership is treated as a prerequisite for national survival. Within this framework, AI is not merely a domain of economic innovation or governance, but a strategic capability whose distribution is understood to determine geopolitical power. Security is therefore defined relationally: to be secure is to maintain an advantage over rivals, while insecurity is equated with the risk of “falling behind” in the development and deployment of advanced AI systems.

This conception is articulated most clearly in United States policy discourse. The National Security Commission on Artificial Intelligence frames AI as a decisive factor in great-power competition, warning that the United States is “not prepared to defend or compete in the AI era” and that failure to act decisively risks “strategic surprise” (NSCAI 2021, 13). AI leadership is presented as inseparable from national security, economic prosperity, and military effectiveness, collapsing these domains into a single strategic imperative. Technological development itself is thereby elevated to the status of a security objective.

This logic is reinforced across executive, defence, and industrial policy. America’s AI Action Plan identifies AI as “a foundational technology for national security” and emphasises the need to “outcompete strategic rivals” across the AI value chain (White House 2025), while the Department of Defense links AI adoption to “decision advantage” achieved through speed, automation, and information dominance (DoD 2023). Similarly, the CHIPS and Science Act frames advanced computing capacity as a national security concern, justifying economic intervention through the language of strategic competition (CRS R47523). Across these domains, security is consistently tied to acceleration and relative performance rather than to the management of social or systemic risk.

Taken together, these policies construct a coherent conception of AI security defined by three core assumptions: security is zero-sum and relational; technological superiority (particularly speed of innovation and deployment) is the primary means of achieving it; and the principal threats addressed are external and strategic rather than societal. Institutionalised across authoritative sites of policy production, this arms-race framing becomes normalised as common sense, marginalising alternative understandings of AI as a governance challenge or source of distributed social risk.

2 Objects, Actors, and Practices of AI Security

Although the AI arms race framing presents itself as a response to broad societal vulnerability, it rests on a narrower (often implicit) definition of what and who is being secured. Official discourse invokes expansive objects of security such as democracy, privacy, and public welfare (The White House 2025). In practice, however, AI securitisation privileges the protection of state capacity, military advantage, and corporate capability, rather than societal security, a prioritisation that becomes visible through the actors empowered and the practices authorised.

This divergence is evident when moving from rhetoric to institutional practice. While policy documents portray citizens as the ultimate beneficiaries of AI leadership (NSCAI 2021; White House 2025), defence strategy defines AI primarily as a means of achieving “decision advantage” and enhancing warfighting effectiveness (DoD 2023). Security is thus operationalised through organisational performance, speed, and technological integration rather than through the mitigation of social harms associated with AI deployment.

Bigo’s account of security as a field of professional practices helps explain this shift. Security is produced less through abstract commitments to public welfare than through the routine activities of specialised actors (military planners, intelligence agencies, and technocratic experts) whose authority rests on claims to expertise and urgency (Bigo 2002). From this perspective, AI securitisation reorganises power by privileging institutions best positioned to define threats and propose technical solutions, rendering state and military capacity the effective objects of security.

Private industry occupies a parallel position within this security assemblage. Congressional testimony portrays major AI firms as indispensable security actors, framing corporate access to data, talent, and capital as a national necessity (Jensen 2025). This alignment is reinforced through industrial policy, including the CHIPS and Science Act, which secures technological infrastructure and supply chains in the name of national security (CRS R47523). While justified through societal protection, these arrangements primarily secure state and corporate capacity, with citizens positioned largely as indirect beneficiaries.

These priorities are enacted through a set of interlocking practices, most notably acceleration. Policy documents emphasise the need to “accelerate the development and adoption of AI” to avoid strategic disadvantage (NSCAI 2021, 17), while the Department of Defense links AI adoption to faster data processing and operational tempo (DoD 2023). Within this logic, delay, oversight, and democratic deliberation are reframed as security liabilities.

Taken together, the objects of AI security are revealed through the practices by which security is pursued. Rather than protecting society in a broad sense, the AI arms race secures institutional capacity and technological dominance, shaping how security success is defined.

3 Securitisation as Process: Chains, Anticipation, and Political Economy

The AI arms race is sustained not through singular declarations of threat, but through processes that translate speculative futures into present governance decisions. De Goede’s concept of “chains of securitisation” captures this dynamic, showing how anticipatory logics mobilise uncertainty to justify pre-emptive intervention (de Goede 2017). In the AI domain, imagined scenarios of rival technological dominance legitimise immediate acceleration, investment, and institutional restructuring.

These anticipatory practices are stabilised through professional routines. As Bigo argues, securitisation is reproduced through the everyday activities of security experts whose authority derives from claims to specialised knowledge and urgency (Bigo 2002). Defence planners and policy experts translate abstract concerns about competition into funding priorities, organisational reforms, and regulatory choices, embedding the arms-race logic in bureaucratic common sense.

Crucially, these chains intersect with political economy. Security discourse directs public resources toward particular technological trajectories while insulating these choices from democratic contestation (AI Now Institute 2023). Once institutions and markets are aligned around competitive acceleration, alternative approaches to AI governance become increasingly difficult to pursue. In this way, the arms race becomes self-reinforcing, producing the conditions under which its own practices appear necessary.

4 Plausibility Test: Does the AI Arms Race Produce Security?

Whether the AI arms race constitutes a plausible conception of security ultimately depends on its effects. If the practices it authorises reduce vulnerability, enhance safety, and mitigate meaningful risks, then its competitive logic may be justified. If, however, they generate new forms of insecurity or fail to address the most significant threats associated with AI, the conception must be judged implausible. Evaluated against this standard, the AI arms race performs poorly.

A first problem lies in the misdiagnosis of insecurity. The dominant framing treats insecurity primarily as an external, strategic condition: the risk that rival states will achieve superior AI capabilities. As a result, security practices prioritise speed, scale, and deregulation. Yet many of the most pressing risks associated with AI (such as system opacity, misalignment, accidental harm, misuse by non-state actors, and social destabilisation) do not arise from falling behind an adversary, but from the rapid and weakly governed deployment of powerful technologies. By privileging external rivalry over internal and systemic risk, the arms race framing diverts attention and resources away from the very sources of insecurity it claims to address.

This critique is not confined to critical security scholarship. Within strategic studies itself, the analytical coherence of the AI arms race metaphor has been questioned. Scharre argues that treating AI as an arms race obscures fundamental differences between AI and traditional military technologies, including the difficulty of measuring capabilities, the diffuse nature of innovation, and the prevalence of dual-use applications (Scharre 2021). Unlike nuclear weapons or conventional arms, AI does not lend itself to clear benchmarks of superiority, rendering race-based assessments inherently speculative.

A second problem concerns risk amplification through acceleration. As shown earlier, the arms race logic treats speed as a security good, incentivising rapid development and deployment. However, accelerated innovation can itself become a source of insecurity. AI systems deployed without adequate testing, oversight, or governance are more likely to produce discriminatory outcomes, safety failures, and vulnerabilities to misuse. For example, the deployment of predictive policing systems in several U.S. cities has reproduced racial bias and intensified surveillance of marginalised communities, illustrating how accelerated AI adoption can generate social harm rather than security (Shapiro 2019). In this sense, practices intended to secure society may instead increase exposure to technological risk.

The arms race framing also generates self-reinforcing escalation dynamics. Heeks and He show how narratives of an AI “Cold War” exaggerate rival capabilities and intentions, encouraging mutual securitisation and foreclosing opportunities for cooperation (Heeks & He 2024). As states respond to perceived threats by accelerating their own AI programmes, insecurity is reproduced regardless of whether initial threat assessments are accurate. Security thus becomes a moving target, continually deferred through further escalation.

Most fundamentally, the arms race conception fails to account for the collective-action character of AI risk. Many dangers associated with advanced AI (such as loss of control, systemic bias, or large-scale misuse) are transnational and non-zero-sum Addressing them effectively requires coordination, shared standards, and collective restraint (Dempsey et al 2022). Yet the competitive logic of the arms race frames openness and cooperation as strategic liabilities, actively undermining the conditions necessary for meaningful risk mitigation.

Contrasting this approach with alternative frameworks further underscores its implausibility. Mazzucato and Perez argue that framing AI primarily through competition obscures the state’s capacity to shape markets toward public purpose, safety, and long-term societal goals (Mazzucato et al. 2022). From this perspective, security is achieved not through dominance, but through the deliberate direction of innovation toward socially beneficial and risk-reducing ends.

Taken together, these considerations demonstrate that the AI arms race does not plausibly deliver the security it promises. By misidentifying insecurity, privileging acceleration over safety, and reproducing escalation dynamics, it generates new vulnerabilities while leaving fundamental risks unaddressed. The following section examines how these failures are distributed and how inequality itself emerges as a security outcome of AI securitisation.

5 Distributional Consequences: Inequality as a Security Outcome

If the plausibility of a security conception depends on its effects, then the distribution of those effects becomes the central question. The AI arms race framing does not merely fail to eliminate insecurity; it redistributes it. While the benefits of competitive AI development accrue to a relatively small set of state institutions and corporate actors, the risks associated with accelerated deployment, weak governance, and technological concentration are displaced onto less powerful groups. Inequality, in this sense, is not an external policy problem but a security outcome produced by the dominant conception itself.

Empirical analyses of AI governance illustrate this asymmetry clearly. The AI Now Institute documents how race-based rhetoric has facilitated the concentration of key resources (compute, data, capital, and talent) within a small number of firms deemed strategically indispensable (AI Now Institute 2025). Public investment and regulatory flexibility are justified on security grounds, yet the resulting gains are largely privatised, while costs linked to labour displacement, surveillance expansion, and social disruption are socialised. Security policy thus operates less to reduce risk than to reallocate it across society.

From a critical security perspective, this pattern is neither accidental nor anomalous. Cooper’s analysis of the “inequality of security” demonstrates how security practices systematically generate differentiated protections, producing “winners and losers” within risk societies (Cooper 2008). Security, in this account, is not a universal condition but a stratified one, unevenly distributed according to political and economic power. Applied to AI, this insight highlights how securitisation privileges actors best positioned to define threats and propose technical solutions, while exposing others to heightened vulnerability.

Private industry exemplifies this dynamic. Congressional testimony portraying major AI firms as guardians of national security positions corporate success as a proxy for societal safety (Jensen 2025). This alignment legitimises policies that shield firms from regulation and subsidise their expansion, while marginalising concerns about workplace precarity, algorithmic bias, and democratic oversight. In effect, security is achieved for some by transferring insecurity onto others.

These distributional effects also extend beyond national borders. The AI arms race concentrates technological capacity within a small number of advanced economies, framing access, cooperation, and technology transfer as strategic risks. States outside this core become dependent on foreign platforms and infrastructures while remaining excluded from meaningful participation in governance decisions (Salami 2024). From a global security perspective, such asymmetries generate instability rather than resilience.

Crucially, these inequalities undermine the very security objectives the arms race seeks to achieve. Social insecurity (manifested in job displacement, surveillance, and weakened democratic accountability) can itself become a source of political instability and mistrust. When AI is experienced as an imposed risk rather than a protective technology, the legitimacy of security institutions erodes. Recognising inequality as a security outcome therefore exposes a core contradiction within the dominant conception and reinforces the case for rethinking AI security beyond competitive securitisation.

6 Rethinking Security Beyond Simple Desecuritisation

The critique developed thus far might appear to invite a straightforward conclusion: if the AI arms race produces insecurity and inequality, then AI should be desecuritised and returned to the realm of normal politics. While intuitive, such a move risks oversimplifying both the problem and the stakes involved. Critical security scholarship cautions that desecuritisation is not inherently emancipatory and that appeals to “normal politics” can themselves reproduce exclusion, hierarchy, and harm (Balzacq et al. 2015). Rethinking AI security therefore requires a more careful move, one that neither uncritically accepts competitive securitisation nor assumes that security discourse can simply be abandoned.

Balzacq and others argue that securitisation is best understood as a contextual and practice-based process rather than a binary shift between politicisation and depoliticisation (Balzacq et al. 2015). From this perspective, the problem with the AI arms race is not that AI has been treated as a security issue per se, but that it has been securitised in a particular way: through competitive urgency, exceptionalism, and technocratic closure. The question is therefore not whether AI should be a matter of security, but what kind of security politics is appropriate given the nature of AI-related risks.

This distinction matters because AI does pose genuine security challenges. Risks associated with large-scale surveillance, autonomous decision-making, systemic bias, and loss of control over complex systems cannot be adequately addressed through market mechanisms alone. Treating AI solely as an economic or innovation policy issue would likely obscure these dangers rather than resolve them, leaving significant sources of insecurity intact.

What is required instead is a shift from competitive securitisation toward reflexive and accountable security governance. This move aligns with Bigo’s emphasis on interrogating the everyday practices and institutional arrangements through which security is produced (Bigo 2002). Rather than privileging speed and dominance, a rethought conception of AI security would foreground risk reduction, democratic oversight, and the mitigation of socially distributed harms. Security, in this formulation, would be measured not by relative advantage over rivals, but by the capacity to stabilise social relations and prevent avoidable harm.

Alternative governance frameworks point in this direction. Mazzucato and colleagues propose a market-shaping approach in which the state actively directs technological development toward public purpose, embedding safety, accountability, and social value into innovation trajectories from the outset (Mazzucato et al. 2022). While not framed explicitly as a security theory, this approach offers a more plausible foundation for AI security than the arms race logic, precisely because it treats risk and inequality as central governance concerns rather than externalities.

In this sense, the alternative to the AI arms race is not desecuritisation, but a different mode of securitisation: one oriented toward collective resilience, precaution, and democratic accountability rather than competitive acceleration. Such a redefinition does not deny the geopolitical dimensions of AI, but it rejects the assumption that security can only be achieved through rivalry and dominance.

Conclusion

This essay has evaluated the plausibility of the AI arms race as a conception of security by examining whether the competitive practices it authorises actually produce security in practice. On this basis, it has argued that the AI arms race constitutes an implausible conception of security. While presented as a necessary response to geopolitical uncertainty, it undermines the very conditions of security it claims to protect.

By defining insecurity primarily in terms of relative technological decline, the arms race misidentifies the sources of risk associated with advanced AI systems. Many of the most consequential threats (such as systemic bias, governance failures, accidental harm, and social destabilisation) are not addressed through competitive acceleration. On the contrary, practices such as deregulation, rapid deployment, and state–industry fusion tend to amplify these risks, rendering security synonymous with speed and scale rather than with safety, resilience, or democratic accountability.

The analysis further demonstrated how AI securitisation operates through anticipatory chains that translate speculative futures into present governance choices. Once embedded into institutions’ agendas, the arms race logic becomes self-reinforcing, redistributing insecurity rather than eliminating it. The benefits of securitisation accrue disproportionately to state institutions and powerful firms, while its costs are borne by workers, marginalised communities, and less technologically advantaged regions. Inequality thus emerges not as a side effect, but as a constitutive outcome of the dominant conception of AI security.

Recognising these failures does not imply that AI should simply be desecuritised. Instead, it calls for a rethinking of how AI is secured. A more plausible conception of AI security would shift the focus from rivalry to risk reduction, from market capture to public purpose, and from strategic advantage to collective resilience. More broadly, this analysis highlights a central challenge for contemporary security politics: the tendency to conflate security with competitiveness in domains shaped by transnational risks. Addressing this challenge requires not the abandonment of security discourse, but its repoliticisation; anchoring security in democratic accountability and in the lived conditions of safety and vulnerability it purports to address.

Bibliography

AI Now Institute (2023) Confronting Tech Power: 2023 Landscape. New York: AI Now Institute.

AI Now Institute (2025) Artificial Power Series: AI Arms Race 2.0. New York: AI Now Institute.

Balzacq, T., Léonard, S., & Ruzicka, J. (2015). ‘Securitization’ revisited: theory and cases. International Relations, 30(4), 494-531.

Bigo, D. (2002) ‘Security and immigration: toward a critique of the governmentality of unease.’ Alternatives, 27 (Suppl.), 63–92.

Cooper, M. (2008). The inequality of security: Winners and losers in the risk society. Human Relations, 61(9), 1229-1258.

de Goede, M. (2017) ‘Chains of securitization.’ Finance and Society, 3 (2), 197–207.

Dempsey, M., McBride, K., Haataja, M., & Bryson, J. (2022). Transnational digital governance and its impact on artificial intelligence. In The Oxford Handbook of AI Governance (p. 20). Oxford University Press.

Department of Defense (2023) Data, Analytics and AI Adoption Strategy. Washington, DC: U.S. DoD.

Heeks, R. and He, Y. (2024) Analysing the US–China “AI Cold War” Narrative. Manchester: University of Manchester.

Jensen, B. (2025) ‘Protecting Our Edge: Trade Secrets and the Global AI Arms Race.’ Testimony to the House Judiciary Subcommittee on Courts, Intellectual Property, and the Internet, 7 May 2025.

Mazzucato, M. and Perez, C. (2022). Redirecting Growth: inclusive, sustainable and innovation-led. UCL Institute for Innovation and Public Purpose, Working Paper Series (IIPP WP 2022-16).

National Security Commission on Artificial Intelligence (2021) Final Report. Washington, DC: U.S. Government Printing Office.

Salami, A. O. (2024). Artificial intelligence, digital colonialism, and the implications for Africa’s future development. Data & Policy, 6, e67. doi:10.1017/dap.2024.75

State Department (2025) Enterprise Data and AI Strategy. Washington, DC: U.S. Department of State.

Scharre, P. (2021) ‘Debunking the AI arms race theory.’ Texas National Security Review, 4 (4), 14–29.

Shapiro, A. (2019) ‘Predictive Policing for Reform? Indeterminacy and Intervention in Big Data Policing’ Surveillance and Society. 17(¾): 456-472. University of Pennsylvania

The White House (2022) National Security Strategy. Washington, DC.

The White House (2025) America’s AI Action Plan. Washington, DC.

United States Congress. (2022) CHIPS and Science Act of 2022. Public Law No. 117-167. Signed 9 August 2022.

Zeng, J. (2021) ‘Securitization of Artificial Intelligence in China.’ Journal of Contemporary China, 30 (129), 1–17.