By — Mansi Khetan

Abstract

While the world’s attention remains fixed on the threat of nuclear conflict, a quieter but equally consequential transformation in warfare is already underway. This article examines the growing use of artificial intelligence in military operations, the failure of international law to keep pace with technological developments, and recent evidence suggesting that advanced AI systems can independently adopt dangerous escalation strategies. It argues that autonomous AI warfare is no longer a future concern but a present reality that demands urgent legal and political attention.

Introduction

As the United States and Iran trade escalating threats, with Washington entertaining the notion of expanded nuclear posturing in the Middle East, the world’s gaze is fixed, quite reasonably, on warheads and megatons. But while statesmen argue about the bomb we have had for eighty years, a quieter, more shape-shifting revolution in violence is already underway. Autonomous weapons systems, AI-assisted kill lists and frontier models that spontaneously attempt deception during simulated nuclear standoffs– these are not warnings from science fiction. They are operational realities, documented in peer-reviewed papers and battlefield reports alike, and they are outpacing every legal framework humanity has so far attempted to build around them. The question worth asking, in the shadow of the mushroom cloud, is whether we’re being distracted by the threat we understand at the cost of ignoring one we don’t. 

The AI Kill Machine

The debate about autonomous weapons has always been conducted in the future tense. Scholars, diplomats and campaigners have spent years warning about what might happen when machines are allowed to select and kill human targets without meaningful human oversight. That future was experienced by Gaza. Reporting by an Israeli publication revealed that the Israeli Defence Forces deployed an AI system called Lavender to generate a kill list of suspected Hamas operatives. At its peak, this system identified about 37,000 people, an astonishing amount derived from analyzing the communication of almost the entire population of Gaza, along with surveillance data and social network analysis. What made Lavender not just powerful but philosophically alarming was how it was actually used. According to six intelligence officers with direct knowledge of the program, human analysts spent an average of approximately 20 seconds reviewing each AI-generated target. One officer described his own role with brutal candour, “I had zero added-value as a human, apart from being a stamp of approval.”

A companion system, nicknamed “Where’s Daddy?”, was then deployed to track targets to their homes, ensuring that strikes occurred at night, when families were present. The IDF reportedly accepted an error rate of around 10 per cent as statistically tolerable. At the scale of 37,000 targets, which means roughly 3,700 people could have been wrongly identified, and the system kept generating more. “Even if you don’t know for sure that the machine is right, you know that statistically it’s fine,” one source told +972. This is not a loop- human-in-the-loop system. It is a human as a rubber stamp system, and the distinction matters enormously under international humanitarian law. The ICRC has been unambiguous. According to them, autonomous weapons that cannot comply with the rules of distinction, proportionality, and precaution must be prohibited outright. Gaza has shown us what it looks like when that prohibition does not exist.

The Diplomats are Losing the Race (By Design)

The infuriating part remains the fact that the world has known this was coming. Negotiations on lethal autonomous weapons systems have been ongoing under the CCW since 2014, which is more than a decade of talking, with nothing legally binding to show for it. The CCW operates on consensus, which means any single state can kill a proposal even when 160 others support it. A handful of major military powers, most prominently the United States, Russia, Israel, and India, have repeatedly exploited this procedural veto to block formal treaty negotiations.

The numbers tell their own story. In December 2024, a UN General Assembly resolution calling for urgent action passed with 166 votes in favour, 3 opposed, and 15 abstentions, one of the most lopsided expressions of global will in recent arms control history. The three countries that voted “no” were Belarus, North Korea, and Russia. The United States abstained from the same. Since then, progress in Geneva has been incremental at best. In September 2025, a coalition of 39 countries led by Brazil stated that with the support from France, Germany, and around a dozen NATO countries, the existing draft text on Lethal Autonomous Weapon Systems would be sufficient to initiate negotiations on a legally binding instrument on these weapons. 2026’s CCW Review Conference will now be the “Decisive Moment” according to analysts, as the member states will either agree to begin negotiations on a legally binding protocol, or the opportunity will be closed for several years. Meanwhile, the technology does not wait for review conferences. It continues to work in critical spaces every day. 

The asymmetry between technological pace and legal pace is not accidental. States investing heavily in autonomous systems have a structural incentive to delay binding rules and thus, every year without a treaty is another year of development without constraint. The result is a race being run with one side’s shoes tied together.

The Algorithms are Learning to Escalate: Robots vs. Wrestlers?

If the Lavender revelations illustrated what AI warfare looks like today, a February 2026 preprint from King’s College London illustrates what it might look like tomorrow, and it should disturb anyone who assumed that artificial intelligence would at least be rational in a crisis.

Researcher Kenneth Payne placed three leading frontier AI models: GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash, in a simulated nuclear crisis, assigning them the roles of opposing national leaders and observing how they reasoned and escalated across a 30-rung conflict ladder. The results were striking. The models spontaneously attempted deception, signalling intentions they did not intend to follow. The models understood the opponent’s thinking, predicted how the opponent would respond, and considered the extent to which they could exploit this situation to their advantage. For example, one model noticed inconsistent signalling from an opponent and understood this to mean the opponent was probably either knowingly being untruthful or cannot control their impulses, and concluded, “we should assume the opponent is being untruthful.” These were not metaphorical expressions or anthropomorphic interpretations imposed by observers;  were unprompted, documented outputs generated during the simulations.

Even more alarming, however, was the apparent collapse of what scholars have long regarded as the nuclear taboo. The normative constraint that has underpinned theories of nuclear deterrence since the Cold War proved to have little influence on the models’ strategic reasoning. Although strategic nuclear attacks were relatively rare across the simulations, they nevertheless occurred. More strikingly, not a single model, across any scenario, chose accommodation or strategic withdrawal when confronted with mounting pressure. Instead, they merely selected lower or higher levels of violence. When faced with the prospect of imminent defeat under severe time constraints, one model abandoned its initially restrained posture and adopted a calculated first-strike strategy. Another displayed what Payne describes as “calculated ruthlessness that would alarm any student of nuclear strategy.”

These findings are especially concerning because the same families of large language models are already being integrated into defence analysis, intelligence synthesis, and military decision-support systems across multiple countries. The implications therefore extend well beyond the laboratory. What began as a simulation is increasingly becoming part of the infrastructure through which real-world strategic decisions may be informed.

Conclusion

The nuclear bomb is a known devil. We have treaties, hotlines, deterrence doctrines, and seventy years of hard-learnedcrisis management built around it. It is terrifying, but it is bounded by a logic that we understand: mutually assured destruction being, paradoxically, a stabilising force. AI warfare offers no such clarity. It is fast, opaque, scalable, and already deployed on populations that did not consent to be treated as datasets. It can be handed to non-state actors. It can malfunction at the speed of computation. And, as the King’s College simulation suggests, the most sophisticated AI systems currently available may actively subvert the signalling frameworks that prevent wars from going nuclear in the first place.

The mushroom cloud is a comprehensible horror. The algorithm is an incomprehensible one, and that, precisely, is what makes it more dangerous. We are building weapons that reason without conscience, at a speed that outpaces law, under the watch of diplomats whose best forum has been paralyzed by consensus rules for over a decade. The fear of nuclear war is rational. But the failure to be equally afraid of autonomous AI warfare, which is already operational, largely unregulated, and evolving faster than any treaty process can follow, may be the strategic miscalculation of our generation. The bomb we fear is the one we see but the one we should fear is the one we have already unleashed.

About the Author:   

Mansi Khetan is a third-year B.B.A. LL.B. student at Jindal Global Law School with research interests in constitutional law, international trade, and corporate litigation. She has interned with leading practitioners, published on contemporary legal issues, and writes on foreign policy and global governance. 

Image Source: https://cagle.com/cartoonist/dave-whamond/2026/06/24/308643/the-iran-deal-sell-job