Student Blog: Understanding the “Anthropic Moment”:An Impossible Trinity of Military AI and the Opportunities Behind
By Roger (Tse) Lo*
Introduction
A recent incident involving tensions between Anthropic and the U.S. Department of Defense (now the Department of War), an event the author calls the “Anthropic moment”, has drawn interdisciplinary attention across academic communities, mainly ethics-wise. As widely recognised, the deployment of military AI necessitates a more rigorous examination of the so-called “black box” problem, to which this article does not seek to resolve the most intricate issue. Instead, it offers an introductory framework that brings international security and international law into conversation.
The author will first explain how key stakeholders interact and then introduce the perspective that military AI is being integrated through procurement and operational pipelines faster than oversight and legal compliance mechanisms can keep pace, allowing the market to exert its negotiating power over the “security urgency” to outstrip democratic control.
In response, this article argues for systematic legal review and proposes the formulation of a special act grounded in the existing legal instruments, e.g. Article 36 of the Additional Protocol I, to be fostered transnationally by legislators worldwide to safeguard the Democratic Principle and uphold respect for humanity under an international consensus.
Starts from the “impossible trinity”
Traditionally, many IHL scholars have supported banning fully autonomous weapons on the ground that humans would lose meaningful control. Yet growing concern about cognitive distortions, such as automation bias, suggests that even “human-in-the-loop” designs may not, ensure adequate due diligence in practice. Worse, from a broader, macro-level perspective on the evolving models of military AI, it is increasingly difficult to deny the increasing influence of the private sector in national security decision-making. Building on Putnam’s two-level games, this article suggests that under the context of techno-politics, the three parties, namely government, society, and the market, are unlikely to reach a stable equilibrium, which this author calls an “impossible trinity.”
Briefing the “impossible trinity” problem: the state (government), society (civil society), and the market (big tech and its analogues), each claims a different source of legitimacy, including inter alia, security imperatives, democratic accountability, or innovation capacity. Still, any two tend to crowd out the third, usually the society. When military AI is procured and fielded at speed, this tension becomes structural rather than episodic, shaping what is built, what is used, and who may meaningfully interrupt the ordinary course of events.
Take the U.S. for example, the governmental decision-making is often shaped by lobbying networks and other political intermediaries behind the scenes. As a result, while public participation in these processes remains extremely limited, the private sector may “dictate how states use AI in war,” warned a group of Stanford scholars about the spiralling risk.
This dynamic is not driven by partisan sentiment; rather, it persists across different presidential administrations, whether Democratic or Republican. In the Biden era, policy orientation converged with part of public opinion, yet the result was that the market’s demand for exporting high-tech products could not be satisfied; stepping forward to Trump’s second term, government decision-making was visibly pulled by market lobbying, and the policy constraints that had originally been formed on the basis of public opinion were forced to give way (for the public opinion, almost 60% of Americans would refuse to such a result, see the survey from Chicago Council on Global Affairs and Carnegie Endowment for International Peace).
This also indicates a notable deviation from the classic two-level game setting: the domestic “second level” no longer generates effective policy pressure, and social forces cease to function as meaningful constraints; instead, they are increasingly instrumentalised by big tech as a source of legitimacy for its preferred policy outcomes.
It must be the Congress that regulates
What is missing? Genuine democratic representation. Legislatures, whether Congress, Parliament, or their functional equivalents, are empowered by the public to set binding rules; regulation of autonomous weapons systems, and more broadly the use of military AI, should therefore be led by elected lawmakers (in fact, these are within their mandates commissioned by the people) rather than substituted by executive agencies or even private firms, as echoed by scholars such as Michael Depp, Alan Z. Rozenshtein, and a group of Stanford experts from various disciplinary backgrounds.
To reach a new balancing point within the framework of two-level games, we can only reasonably expect the Legislature to form an internal check-and-balance within the “state” (in the sense of the trinity), since none of the other actors can reliably do this job — government, the market, or think tanks.
First, experiences from state practice suggest that armed forces, despite increasingly sophisticated risk-assessment mechanisms (e.g., collateral damage estimation), remain prone to deviating from legal norms when operational imperatives dominate. Aside from history, one recent illustration is the tension in U.S. policy on nuclear command and control: While official positions emphasise sustained human control, institutional signals and contestation suggest pressure to incorporate increasingly autonomous capabilities, such as demonstrations directed at the Pentagon and debates over expanding autonomy in decision-support systems. If these systems begin to function as a substantive basis for operational decision-making, the risk grows that practice will drift from stated commitments.
Second, it is reported that 40% of enterprise AI leaders have expressed concern about the militarisation of AI. However, even when these concerns are widely shared, private-sector actors cannot resolve the problem on their own, because they do not possess the institutional authority to make binding law. Paramountly, the market is not a single, unified actor, but highly differentiated and power-asymmetric, structured by the fact that a small number of firms (emphasis added on Nvidia, see news from Bloomberg or Reuters; see also a recent interview with its CEO Jensen Huang) with critical technology and market position can monopolise policy discourse, even when there is no consensus in the market or society.
Third, ostensibly independent research institutes may also become “securitised”, particularly when they depend on government funding or are embedded in national-security policy networks. For example, reporting on the Alan Turing Institute has raised a question about the degree to which public sponsors can shape research priorities and their leadership changes. Such influence would barely equate to removing independence assurance, if becoming decisive, independent institutions may struggle to function as “credible”, autonomous checks within the broader ecosystem of military AI governance.
Taken together, absent any foreseeable external intervention, internal balancing will be the primary pathway to improve deliberation and democratic representation. Otherwise, the win-set will likely narrow, increasing the risk that the rivalry devolves into a prisoner’s dilemma.
International alignment Re-centred
At present, the U.S. approach to international alignment is to issue a declaration and invite others to sign a Call for Action that largely reiterates those U.S.-formulated principles. Concurring with Michael Depp’s view, this article favours an internal balancing, which means the Call for Action must be complemented by a process that draws on a wider range of national perspectives, particularly from states that are not leading AI powers. Broadening participation helps ensure that emerging practices are tested against established debates in international relations and international law, rather than being normalised by a small number of leading AI states, most notably the U.S. and the UK, who set the trajectory of frontier AI and continue to export technologies that help to operate key segments of the military data supply chain. This concentration heightens proliferation risks around military data and can amplify escalation dynamics across multiple theatres.
To illustrate, the author observes that multiple international initiatives now address “responsible” AI development, including the Asilomar AI Principles (2017), OECD AI Principles (2019), IEEE’s Ethically Aligned Design (2019), the UN-GGE reports on LAWS (2019; 2023), the G7 Digital and Tech Ministers’ Joint Declaration (2023), the GC REAIM (2023), and the Chair Summary of the Vienna Conference on Autonomous Weapons Systems (2024)…etc. Yet these instruments are largely soft law: they do not, by themselves, create binding obligations for states. Their non-binding character leaves room for manoeuvre in interpretation and implementation, enabling governments to selectively invoke high-level commitments while preserving strategic discretion—thereby sustaining uncertainty over the concrete legal status of frontier military AI deployments.
Aggregating the above, this author suggests that the challenge of ensuring armed forces’ compliance with international law cannot be understood solely at the level of abstract legal principles; the Legislature must operationalise the compliance, moving beyond formal reiterations of rules of engagement and toward governance frameworks that translate international consensus into operational practice. If failed, they will thereby be responsible for reinforcing path dependencies within the military hierarchy.
Operationalising Compliance with the Trinity
While the “impossible trinity” reflects structural limitations within and beyond the legal aspect, the core issue lies in how compliance is operationalised within institutional and technological systems. Effective compliance depends on sustained interaction between domestic institutions and multilateral processes. Against this background, this article closes with three suggestions for legislators operating under different constraints:
Legislatures should set clear statutory baselines, executives should internalise them through system design, and international forums should converge on implementable standards that can be adopted across jurisdictions.
Concerning its interconnected nature, lawmakers should deepen cross-national parliamentary engagement, rather than leaving coordination primarily to executive branches. In particular, the Vienna Conference process offers a promising model for iterative, inclusive alignment.
Transnational conferences should be used to identify concrete regulatory gaps, develop new instruments where needed, and clarify how existing treaty frameworks, including the Geneva Conventions, the Hague Regulations, and the Convention on Certain Conventional Weapons, continue applying to frontier military AI deployments.
These initiatives have been consistently supported and advanced by the Responsible by Design Institute, who convening an international group of experts to develop practical legal guidance, with a focus on weapons reviews under Article 36 of Additional Protocol I to the Geneva Conventions. Without such alignment, legal norms risk remaining aspirational; with it, international consensus can serve as a practical constraint on the development and deployment of military AI, reducing the likelihood of escalation and systemic harm.
Conclusion: Parliament-led legislation is the opportunity behind
Through the lens of the “impossible trinity”, we are allowed to observe how the techno-politics transformed the essence of the classic international theory (the “two-level games”) and how civil society is gradually losing its voice.
To uphold the spirit of democracy —— if people still believe it —— the Congress, Parliament, or any other institution that may equally function shall intervene now. Without internal balancing, it is almost certain that human-beings will witness our self-destruction by creating things we can no longer control, and starting to rediscover why "perpetual peace" matters, as Kant already put in front of our eyes in 1795. We know it is an ideal illusion, yet for a brief moment in the 1990s, globalisation made it feel almost real.
If the international community, in particular the legislators around the world, cannot assemble together, in the foreseeable future, not far away, we may grow accustomed to AI-enabled warfare while continuing to place our faith in notions of "power" shaped by the ideal political regime that too often empowers ambitious individuals seeking to rule the world rather than advance the common welfare of humankind.
*Roger (Tse) Lo is a junior Diplomacy student at National Chengchi University (NCCU), Taiwan (R.O.C), pursuing a double major in International Law and Digital Governance, with minors in Legal Studies and Philosophy. A two-time champion of the IHL moot court competition in Taiwan, he is also the youngest recipient of NCCU’s AI Interdisciplinary Research Scholarship. Currently, he serves as a research assistant at the NCCU Centre for International Legal Studies and the Asia-Pacific Regional Coordinator of the Student Ambassador Scheme at the Responsible by Design Institute.
The views expressed in this post are those of the author, and do not necessarily reflect the position of the Responsible by Design Institute.
The Student Blog Series is a forum for students to share opinions and cultivate ideas on the responsible design and development of new military technologies. RBDI does not screen posts to fit a particular editorial agenda, nor endorse or advocate material that is published.
