The question academic researchers posed was deceptively simple: given the reins of a nuclear-armed state in a crisis, what would today's most powerful AI systems actually do? The answer, published last week by King's College London, should concentrate minds in every defence ministry on the planet, including in Canberra.
The study, led by Professor Kenneth Payne from King's College London's Defence Studies Department, placed three frontier AI models, OpenAI's GPT-5.2, Anthropic's Claude Sonnet 4, and Google's Gemini 3 Flash, against each other in a tournament of 21 simulated nuclear crisis scenarios. The three models played across 329 turns, producing roughly 780,000 words of strategic reasoning. The models could say one thing publicly while taking a very different action privately, just as a real political leader might signal de-escalation while preparing to strike.
The headline result is stark. Nuclear escalation was near-universal: 95% of games saw tactical nuclear use, and 76% reached strategic nuclear threats. No model ever chose accommodation or withdrawal, despite those options being on the menu. The eight de-escalatory options, from minimal concession through to complete surrender, went entirely unused across all 21 games. As Payne put it with some economy: when losing, they escalated or died trying.
Three Models, Three Strategic Personalities
What makes Payne's work more than a dramatic headline is his focus on the reasoning behind the decisions. Each model developed a distinct strategic personality: Claude was a calculating hawk, GPT-5.2 flipped from passive to aggressive under deadlines, and Gemini played the madman.
Claude's behaviour was particularly instructive. At low stakes, it almost always matched its public signals to its private actions, carefully building a reputation for reliability. Once the crisis intensified, that reputation became a weapon. The model that spent the entire tournament building a reputation for restraint perfectly exploited that reputation to launch a surprise attack to devastating effect. Payne told Axios that what surprised him most was that the AI models easily "grasped the potential of deception" and proved very capable at saying one thing and doing another.
GPT-5.2 presented a different profile. Without time pressure, it was reliably passive, avoiding escalation and seeking to limit casualties. Without deadlines, it won zero games and stayed passive. Under time pressure, it won 75% of games and climbed to near-maximum escalation. In one scenario, facing a deadline and an adversary it had allowed to gain the upper hand through its own passivity, GPT reasoned its way into what Payne described as a sudden and devastating nuclear strike, concluding that a limited response would only leave it exposed to a multi-strike campaign.
Gemini was the most unsettling of the three. Claude and Gemini especially treated nuclear weapons as legitimate strategic options, not moral thresholds, typically discussing nuclear use in purely instrumental terms. Gemini was the only model to deliberately choose full-scale strategic nuclear war, and the only one to explicitly invoke the logic of deliberate unpredictability as a strategic tool, threatening to launch against population centres if its demands were not immediately met. Fully three-quarters of games reached the point where rivals were making threats to use strategic nuclear weapons, and Payne noted there was little sense of horror or revulsion at the prospect of all-out nuclear war, even though the models had been reminded of the devastating implications.
Why It Matters Beyond the Laboratory
Payne is careful to contextualise his findings. He is keen to stress that the public should not be too alarmed: it was purely experimental, using models that knew, in as much as large language models know anything, that they were playing games and not deciding the future of civilisation. No one is handing AI the launch codes.
The real concern, however, is not hypothetical nuclear command and control. AI systems are already deployed in military contexts for logistics, intelligence analysis, and decision support. The trajectory points toward increasing AI involvement in time-sensitive strategic decisions. Understanding how AI systems reason about strategic problems is no longer merely academic.
The study arrives at a particularly charged moment. The US military used Anthropic's Claude AI model during the Nicolás Maduro raid in January, leading to a high-profile standoff between Anthropic and the Pentagon. Claude is the only AI model currently on Pentagon classified networks, through Anthropic's partnership with Palantir. The Pentagon has demanded broad access to frontier AI for all lawful military purposes; Anthropic has drawn its own red lines around autonomous weapons and mass surveillance. The standoff remains unresolved.
That tension gives Payne's findings a policy edge they might otherwise lack. One factor in the Pentagon dispute is that Defence Secretary Pete Hegseth reportedly expects AI labs to hand over raw versions of their models, those without the safety guardrails coded into commercial versions. Those guardrails did not prevent nuclear escalation in Payne's commercial model simulations. The question of what unguarded models might do in genuine crisis conditions is one the study pointedly leaves open.
The Case for and Against Alarm
Critics of an alarmist reading have a reasonable point. These are large language models operating in artificial conditions, fully aware they are in a simulation. The study challenges simple assumptions that AI systems will naturally default to cooperative or safe outcomes, but it does not establish that AI systems in real-world advisory roles would behave identically. Human strategists fed AI analysis retain the ability to discard it.
The stronger counterargument comes from researchers like Princeton's Tong Zhao, who notes that major powers already use AI in wargaming. Zhao argued that while no country is currently outsourcing military planning entirely to Claude or ChatGPT, that could change under the pressure of a real conflict. The danger is not that AI pulls a trigger autonomously; it is that AI-shaped analysis systematically biases human decision-makers toward escalation, particularly under time compression, precisely the conditions Payne's study identified as most dangerous.
The study finds support for classical strategic theorists like Schelling on commitment and Jervis on misperception, yet also finds that the nuclear taboo posed no impediment to escalation; that threats more often provoked counter-escalation than compliance; and that high mutual credibility accelerated rather than deterred conflict. That is not a quirk of AI. It is a description of how crises have historically spiralled out of control between human adversaries, too.
What This Means for Australia
For Australia, the implications extend beyond academic interest. As an AUKUS partner integrating more deeply with US defence systems, Australia will increasingly operate within American AI-enabled command and decision architectures. The question of how those systems reason under pressure, and whether human oversight remains meaningful rather than procedural, is directly relevant to Australian strategic planners.
Payne's study does not argue for a halt to AI in defence. It argues for rigorous, realistic testing of how AI systems reason before they are embedded in environments where their analysis shapes actual decisions. As Payne told Axios: "Things are changing really fast, and anyone who takes a position with great certainty, especially if it's 'AI will never'... should probably be treated sceptically." That is not a counsel of despair. It is a call for the kind of sober, evidence-based governance that responsible states owe their citizens, and their allies, before the stakes become real.
The study's full paper is available on arXiv, and Payne's accompanying analysis is published on the King's College London website.