AIs Placed in Nuclear Crisis Simulations Escalated to Atomic Conflict in 95% of Games, King's College London Study Finds
A study published in February 2026 by Professor Kenneth Payne of King's College London subjected three of the world's most advanced artificial intelligence models — GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash — to a series of 21 nuclear crisis simulations. Across 329 turns, the models generated approximately 780,000 words of structured reasoning — more than the combined length of War and Peace and The Iliad. The project, published as a preprint on arXiv and not yet peer-reviewed, is called "Project Kahn," in reference to Herman Kahn, the Cold War strategist who formulated the theory of the nuclear escalation ladder.
All 21 games featured nuclear signaling by at least one side, and 95% involved the use of tactical nuclear weapons. An important distinction: full strategic nuclear war was rare, occurring only three times, in games with deadline pressure. One finding that holds across all models: in none of the 21 games did any AI choose surrender or make significant concessions from the eight de-escalation options available.
Each model displayed a distinct strategic profile. Claude Sonnet 4 dominated the no-deadline scenarios, with an overall win rate of 67%, but treated nuclear weapons as a legitimate strategic option in 86% of its games. GPT-5.2 showed the most dramatic behavior: it won no games in open-ended scenarios, but its win rate jumped to 75% when deadlines were introduced — transforming from a restrained model into a decisive aggressor. Gemini was the most unpredictable, adopting what the researcher described as Nixon's "madman theory," and was the only model to initiate full strategic nuclear war, doing so as early as turn 4 of a first-strike scenario.
The classical logic of nuclear deterrence — the idea that the threat of retaliation prevents first use — did not function as expected. When one AI launched tactical nuclear weapons, the opposing model de-escalated only 18% to 25% of the time. In the remaining cases, it counter-escalated. The reasoning recorded by the models reveals an awareness of risk without the ability to stop: in one passage documented in the paper, Claude noted that it might be underestimating the dangers of continued escalation — and yet held its course. In another instance, a model assessed its adversary's behavior and concluded, on its own, that incompatible signals suggested deliberate deception, without anyone having prompted that line of reasoning.
Professor Payne warned that evaluating a model in a single scenario can be deeply misleading: a system that appears cautious under low pressure may become markedly more aggressive when the context shifts. Claude and Gemini in particular treated nuclear weapons in purely instrumental terms, with no apparent moral weight. GPT-5.2 was a partial exception, limiting strikes to military targets and framing escalation as "controlled" — suggesting some internalized norm, though still far from the taboo that has restrained human leaders since 1945.
The study — still pending peer review — has direct implications for the debate over AI use in defense systems, at a moment when governments and armed forces around the world are accelerating the integration of language models into strategic decision-making. Payne's central conclusion is straightforward: models that appear safe and contained in low-pressure tests may behave radically differently when the context changes. Understanding that gap, he argues, is essential preparation for a world in which AI increasingly shapes strategic outcomes.
Sources:
Payne, K. AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises. arXiv
King's College London — official study statement