The discourse around artificial intelligence has undergone a dramatic shift from the initial excitement about its potential benefits and concerns about model safety. Throughout 2024, AI slowly moved to the domain of security. Drawing from my graduate studies at LSE and my dissertation on securitisation theory, I believe this framework offers valuable insights into understanding this transition and its implications for AI governance.
While this post serves as a foundation for a more comprehensive paper, it aims to introduce how securitisation theory can illuminate current developments in AI policy and governance. Please reach out if you would like to explore this topic further with me!
Disclaimer: This piece does not aim to pass value judgment on securitisation nor reflect personal opinions on AI governance. Rather, it seeks to examine how securitisation manifests in this particular field and provide a few representative examples.
Short Introduction to Securitisation
Securitisation theory encompasses two main theoretical frameworks: the Copenhagen School and the Paris School. The Copenhagen School first coined the term securitisation and the Paris School then broadened the scope of what is meant by it. The following paragraphs summarise the key arguments of these schools and serve as a basic introduction to this theory.
Reusing a part of my dissertation, here I outline the traditional perception of securitisation as defined by the Copenhagen School:
“The term securitisation was first defined by Ole Wæver as a ‘speech act’ (Wæver, 1995) claiming that ‘by saying the words, something is done’ which implies an assumption that words construct a reality (Buzan et al., 1998, p. 26). Building on this perspective, the Copenhagen School of Security claims that the securitizing actor ‘moves a particular development into a specific area, and thereby claims a special right to use whatever means are necessary to block it’ (Wæver, 1995, p. 55). Constructing issues in a specific socio-political context as threats and the acceptance of them as such by the audience then results in a ‘politics of exception’ which allows the deployment of extraordinary, unconventional, and disproportionate measures and mechanisms, exerting power over the labelled domain (Buzan et al., 1998; Donnelly, 2017).”
This perception is expanded by another theoretical approach to securitisation – the Paris School – which argues that securitisation is not performed solely through discourse but also through actions. Therefore, they emphasise the politics of insecurity rather than just the actual level of insecurity faced by society, highlighting the context as well as the use of securitisation to control the population through “a continuum of threats” that enables the securitising actor to accumulate power (Bigo, 2002, p. 63; Huysmans, 2004). The “securitising actor” can be the government, various agencies, or even private sector representatives.
The Evolution of AI Securitisation
The trajectory of AI securitisation can be traced through several key developments, demonstrating both discursive and practical manifestations of this process.
The roots of this approach to AI can be traced to 2019 when President Trump signed the “Maintaining American Leadership in Artificial Intelligence” order, saying, "Continued American leadership in Artificial Intelligence is of paramount importance to maintaining the economic and national security of the United States.” Highlighting the competitive dynamics and the security aspects of AI, this speech was an exemplary securitising act as defined by the Copenhagen School and an early instance of securitisation.
Narratives about AI security have been on the rise since November 2022 (as Google Trends nicely show). This trend is captured by Leopold Aschenbrenner’s essay Situational Awareness which represents a pivotal moment in AI securitisation discourse. Defining AI as a crucial technology for preserving national security, this essay introduces “AGI realism” and predicts a necessary government takeover of leading AI companies – justifying extraordinary measures through the security framing. Its subsequent sharing by President Trump’s daughter Ivanka Trump ahead of the US elections suggests its potential influence on future policy approaches.
Most recently, the renaming of the UK’s AI Safety Institute to the AI Security Institute demonstrates institutional securitisation. The government website claims that this change was made to strengthen “protections against the risks AI poses to national security and crime.” As Politico writes: “On its website the institute has now dropped talk of ‘societal impacts’ as a reason for evaluating models, changing it to ‘societal resilience.’ References to the risk of AI creating ‘unequal outcomes’ and ‘harming individual welfare’ have also gone. The institute has also dropped ‘public accountability’ as a reason for evaluating models, changing it to keeping the ‘public safe and secure.’” The new language indicates the more “hard security” focus of the Institute, reshaping institutional priorities and approaches.
The securitisation of AI extends beyond government actors. In the introduction of the Stargate Project, OpenAI claims that “this project will not only support the re-industrialization of the United States but also provide a strategic capability to protect the national security of America and its allies,” which is probably one of the reasons why this private initiative was presented at the White House in the first place. While using a somewhat softer language, Anthropic’s Dario Amodei also seems to favour this perception, most recently advocating for export controls to prevent China from gaining a technological advantage. The focus on the US-China competition and the vocabulary around it persuasively align AI with national security narratives.
Why Does this Matter?
The securitisation of AI has three primary implications:
Legitimising extraordinary measures and power concentration
Shaping public and political discourse
Fuelling global AI competition
1. Legitimising extraordinary measures and power concentration
Securitisation enables and justifies the concentration of power in AI development among a small number of actors. Framing AI companies as national champions in a race to achieve dominance in international competition facilitates market concentration. At the same time, it gives the government legitimacy to make regulatory exceptions for the AI industry and disregard safety measures to boost rapid progress and preserve competitive advantage. This approach might result in increased market oligopolisation, relaxed safety standards, and potentially even government intervention or nationalisation as suggested by Leopold Aschenbrenner.
2. Shaping public and political discourse
Words matter and they shape reality to a non-trivial degree – this is the foundation of securitisation theory. As Foucault claims, ‘discourse transmits and produces power’ (Foucault, 1998, p. 100-1) and the public discussions about the importance of AI for national security very likely shape public perception of this technology as a whole. Construing AI as a security risk leads to prioritisation of control over democratisation. This can result in the public favouring government takeover of AI labs without considering other governance models and blocking broader benefit-sharing.
3. Fuelling global competition
Connecting AI to national security and framing it as a strategic asset casts AI as a zero-sum game. This perception reduces the potential for productive international collaboration and can lead to mistrust, nationalism, protectionism, and possibly greater militarisation, resulting in Cold War-like dynamics. Apart from potentially increasing the likelihood of an international conflict, the securitisation of AI can hinder possible benefit-sharing schemes, fostering greater isolationism and global technological fragmentation.
Moving Forward
Understanding the securitisation of AI through this theoretical framework reveals important patterns in how the technology is being framed and governed. Various groups are (possibly unwittingly) actively trying to securitise this technology both through speech and actions, and this influences the direction that AI development at large follows. While securitisation might serve certain strategic purposes, it also risks constraining our ability to address AI's challenges collaboratively and democratically. Considering the possible consequences, it is important to think about the words we use and their impact on constructing the social reality we share.
The goal of this post was to outline how the securitisation theory applies to AI and why it can provide a very relevant theoretical lens for analysing the various recent developments in this field. I am planning to expand this piece into an academic article and I would love to collaborate on this with other AI governance researchers. I am also happy to discuss this topic and listen to any feedback you might have. Thanks for reading!
Securitisation Literature:
Buzan, B., Wæver, O., & Wilde, J. de. (1997). Security: A New Framework for Analysis. Lynne Rienner Publishers.
Bigo, D. (2002). Security and Immigration: Toward a Critique of the Governmentality of Unease. Alternatives: Global, Local, Political, 27(1_suppl), 63–92. https://doi.org/10.1177/03043754020270S105.
Donnelly, F. (2017). In the name of (de)securitization: Speaking security to protect migrants, refugees and internally displaced persons? International Review of the Red Cross, 99(904), 241-261. doi:10.1017/S1816383117000650.
Foucault, M. (1998). The History of Sexuality: The Will to Knowledge. London, Penguin.
Huysmans, J. (2004). A Foucaultian view on spill-over: freedom and security in the EU. Journal of International Relations and Development, 7(3), 294–318. https://doi.org/10.1057/palgrave.jird.1800018.
Waever, O. (1995). Securitization and Desecuritization. In R. D. Lipschutz (Ed.), O Security. essay, Columbia University Press.