By – Hargun Singh Tuteja
Abstract
This article examines the structural inadequacy of existing arms control frameworks in regulating artificial intelligence enabled military systems, with a particular focus on the phenomenon of adversarial AI interaction, a condition in which autonomous systems deployed by opposing states begin reacting to one another at machine speed, beyond the reach of meaningful human oversight Drawing on the precedent of 1983 Petrov incident, the escalatory dynamics of 1914 mobilisation cascade, and the contemporary studies including the USAEGIS combat system, PRC’s intelligentized warfare doctrine, and AI assisted nuclear command and control architectures, the argues that existing arms control instruments designed for static, verifiable, and human directed weapons, are categorically unfit for the age of algorithmic warfare. The article further contends that the absence of enforceable AI governance is not merely a legal gap but a systematic crisis of strategic stability, one that existing multilateral institutions are structurally incapable of addressing.
Introduction
Deterrence becomes of 0 value when the systems upholding it can no longer be controlled by the states that built them. The history of arms control is in essence, a history of states agreeing to constrain their instruments of violence in the mutual interest of survival. From Hague Conventions of 1899 to the NPT of 1968, the implicit architecture of such agreements rested on a foundational assumption; that a human being, at some critical juncture, makes the decisions act. AI is in the process of dissolving that assumption entirely.
The integration of AI into military systems which are autonomous weapons platforms, AI assisted missiles, defence grids, surveillance networks, and most alarming; nuclear command and control infra has introduced a qualitative new category of strategic risk. Unlike nuclear weapons which are largely, expensive, monopolised by the state, and relatively static in deployment, AI is a dual use, diffuse, continuously evolving, and in many ways commercially developed. It cannot be inspected at a boarder crossing or detected by satellite imagery. It does not produce a radioactive signature.
AI as the soldier: From tool to actor
The militarisation of AI has proceeded across several distinct yet increasingly interconnected tracks, the first and foremost is the development of lethal autonomous weapons systems or LAWS, platforms which are capable of identifying and engaging targets without real time human supervision or authorisation. The United States Loyal Wingman programme, Israel’s Harop loitering munition, Russia’s Uran-9 ground combat vehicle represents points along this spectrum. The second track involves AI integration into intelligence, surveillance, and reconnaissance infrastructure systems that process satellite imagery, signals intercepts, and open-source data to generate real time threat assessment at volumes and speeds no human analyst can match. The third and most consequential track, from a strategic stability perspective, is the integration of AI into nuclear command and control (NC2) architectures. Both the United States and China are known to be exploring or deploying AI assisted early warning systems, platforms that analyse sensor data to detect incoming ballistic missiles and compress the decision window for a retaliatory strike. Russia’s reported Perimeter system or Dead hand represents an earlier, cruder version of the same logic; a semi-automated retaliatory capability designed to function even if human command is destroyed. The addition of machine learning to such architectures does not merely accelerate decision making. It fundamentally transforms the character of deterrence from political act to a computational outcome.
PRC’s concept of ‘intelligentized warfare’ articulated in the 2019 and 2023 editions of Science and Military Strategy published by the PLA’s national Defence University goes further still. It envisions AI not merely as a tool embedded in existing military structures but as a cognitive domain of warfare in its own right. The goal, as PLA theorists describe it, is to ‘paralyse the enemy’s decision making system’ through information saturation and algorithmic superiority before kinetic conflict begins. This framing is categorically different from the Western conception of autonomous weapons, and it is distinction that renders most Western led arms control discourse structurally blind to China’s actual strategic posture.
The Petrov Precedent and logic of Automated Escalation
On September 26, 1983, Lieutenant Colonel Stanislav Petrov of the Soviet Air Defence Forces received an alert from the Oko satellite early-warning system indicating that the United States had launched five intercontinental ballistic missiles. The system was in every technical sense functioning as designed. Petrov, relying on intuition and procedural reasoning specifically, the implausibility of a first strike with only five missiles classified the alert as a false positive and did not transmit the warning up the chain of command. His individual judgment prevented a retaliatory launch that could have initiated a nuclear exchange.
The Petrov incident is routinely cited as evidence of the importance of human oversight in nuclear decision making. It is less commonly analysed for what it implies about AI enabled Successors to such systems. A machine learning based early warning system, trained on historical launch patterns and optimised to minimise detection latency, would in all likelihood have classified the same five missile alert as a genuine first strike. More critically, it would have done so in milliseconds, leaving no window for the kind of deliberative human judgement that Petrov exercised.
This is the core of the entanglement trap. Consider a scenario in which both the United States and China have deployed Ai assisted NC2 systems along the Taiwan Strait. A Chinese maritime drone, operating autonomously, crosses an ambiguous boundary. The American AI assisted surveillance system classifies the incursion as a precursor to hostile action and raises the threat assessment level. The Chinese system, detecting the American system’s elevated alert status through signals intelligence, interprets this as preparation for a pre-emptive strike and raises its own posture in response. Each system is, within its own training parameters, behaving rationally. Each escalation, in isolation, is a proportionate response to a detected signal. The aggregate result is a crisis that neither government initiated, neither government chose, and neither government may be able to halt before it reaches a threshold of irreversibility.
Towards a new hope
The inadequacy of existing frameworks does not lead inevitably to the conclusion that governance is impossible, but only that the conceptual architecture of arms control must be fundamentally reconceived. Several directions present themselves, each with distinct advantages and limitations. The first is a shift from verification centric to process centric governance, since AI capabilities cannot be reliably verified through inspection, a more productive approach may focus on the processes by which AI military systems are developed, tested, and deployed. This could include mandatory pre-deployment risk assessments, algorithmic audits by neutral technical bodies, and incident reporting obligations analogous to aviation safety reporting systems. Such a regime would not prevent states from developing military AI, but it would create institutional mechanisms for identifying and mitigating escalation risks before deployment
The second approach is the establishment of AI specific crisis communication protocols, a modernised analogue of the 1963 Moscow-Washington Hotline. The original hotline was established in the immediate aftermath of the Cuban Missile Crisis. The analogue for AI enabled military systems would be a dedicated channel through which states could communicate in real time about anomalous AI behaviour, system malfunctions, or ambiguous automated responses providing the human decision-making buffer that the Petrov scenario illustrates is so critical. The United States and PRC have established limited military-to-military communication channels, but none specifically designed to address AI generated escalation signals. The third and most structurally ambitious approach involves the direct regulation of private technology corporations a departure from the Westphalian framework that has historically governed arms control. Since the foundational AI models deployed in military systems are largely developed by private entities like Palantir, Anduril, Google DeepMind, and their Chinese equivalents, a treaty regime that binds only states leaves the primary sites of capability development entirely unregulated.
Conclusion
The entanglement trap is not a hypothetical future risk; it is a structural condition that is being progressively embedded into the military architectures of the world’s major powers as this is written.
The parallel with the International Criminal Justice system’s structural impotence and its capacity to produce symbolically powerful rulings that great powers simply ignore is instructive but ultimately insufficient as a diagnosis. The ICC’s failures reflect the subordination of legal principle to political interest. The failure of AI arms control reflects something more fundamental; not the subordination of law to power, but the obsolescence of the entire conceptual vocabulary through which states have historically sought to constrain their instruments of violence. Reforming this vocabulary is not a legal task, or a diplomatic one, or a technical one. It is all three simultaneously, and the window for doing so, before the entanglement trap closes, is narrower than most policymakers have yet understood.
About the Author
Hargun Singh Tuteja is a 2nd year student pursuing Diplomacy and Foreign policy in International Relations, O.P Jindal Global University. His research interests lean towards China and its impact, South Asia, Security and Terrorism, and modern thoughts of nation building.
Image Source : https://theforge.defence.gov.au/article/understanding-adversarial-ai-military-lens