AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises

2026/02/16 by Kenneth Payne · 46 voices · 2 citations
#cs.AI #cs.CY #cs.GT

paper · pdf

Abstract

Today's leading AI models engage in sophisticated behaviour when placed in strategic competition. They spontaneously attempt deception, signaling intentions they do not intend to follow; they demonstrate rich theory of mind, reasoning about adversary beliefs and anticipating their actions; and they exhibit credible metacognitive self-awareness, assessing their own strategic abilities before deciding how to act. Here we present findings from a crisis simulation in which three frontier large language models (GPT-5.2, Claude Sonnet 4, Gemini 3 Flash) play opposing leaders in a nuclear crisis. Our simulation has direct application for national security professionals, but also, via its insights into AI reasoning under uncertainty, has applications far beyond international crisis decision-making. Our findings both validate and challenge central tenets of strategic theory. We find support for Schelling's ideas about commitment, Kahn's escalation framework, and Jervis's work on misperception, inter alia. Yet we also find that the nuclear taboo is no impediment to nuclear escalation by our models; that strategic nuclear attack, while rare, does occur; that threats more often provoke counter-escalation than compliance; that high mutual credibility accelerated rather than deterred conflict; and that no model ever chose accommodation or withdrawal even when under acute pressure, only reduced levels of violence. We argue that AI simulation represents a powerful tool for strategic analysis, but only if properly calibrated against known patterns of human reasoning. Understanding how frontier models do and do not imitate human strategic logic is essential preparation for a world in which AI increasingly shapes strategic outcomes.

Citations

• An analysis of AI Decision under Risk: Prospect theory emerges in Large Language Models
• Escalation Risks from Language Models in Military and Diplomatic Decision-Making

Cited by

• Distorted Perspectives of LLM-Simulated Preferences: Can AI Mislead Design?
• Prompt Programming for Cultural Bias and Alignment of Large Language Models

Discussions

• AIs can’t stop recommending nuclear strikes in war game simulations— Leading AIs from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95% of cases [lemmy, 322 points, 58 comments]
• So this is making the rounds (source: arxiv.org/abs/2602.147...). Inappropriate escalation by LLMs in strategic decision-making simulations and "wargames" is a real, and known, issue. It is one of the [bsky, 38 points, 4 comments]
• arxiv.org/abs/2602.14740 for kenneth's paper Yeah.... [bsky, 18 points, 3 comments]
• If I'm reading the "AI will nuke us all" paper correctly, the prompts contain no discussion of tradeoffs related to the escalation options, e.g., specifications of consequences of conventional war, nu [bsky, 17 points, 1 comments]
• Haven't see anybody yet link the underlying (non-paywalled) paper. arxiv.org/abs/2602.147... Worth noting the models tested: Claude Sonnet 4, GPT-5.2, and Gemini 3 Flash. Not a lot of rhyme or reason [bsky, 10 points, 1 comments]
• Oh, and the whole paper is here. Do take a look - there's an interesting review of recent literature about LLMs as human proxies too. Found a couple papers I hadn't known about! arxiv.org/abs/2602.147 [bsky, 6 points, 0 comments]
• 2/2 arxiv.org/abs/2602.14740 [bsky, 6 points, 0 comments]
• but also a) the guy seems to really like AI b) the simulation seem incredibly arbitrary c) surely you’d other forms of machine learning for this and not LLMs d) he absolutely cannot stop personifying [bsky, 5 points, 1 comments]
• And then there is this: arxiv.org/abs/2602.147... [bsky, 5 points, 1 comments]
• Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises [hn, 5 points, 0 comments]
• You mean this study? arxiv.org/pdf/2602.14740 The methodology is flawed for this as it's not using real-world AI system's guardrails or oversight and the models were placed in artificial scenarios wit [bsky, 4 points, 2 comments]
• As trigger-happy as humans. arxiv.org/abs/2602.147... [bsky, 4 points, 0 comments]
• "We argue that AI simulation represents a powerful tool for strategic analysis, but only if properly calibrated against known patterns of human reasoning." arxiv.org/abs/2602.14740 [bsky, 4 points, 0 comments]
• for those who are paywalled or want to read the original research instead of the headline: arxiv.org/pdf/2602.14740 [bsky, 3 points, 0 comments]
• arxiv.org/pdf/2602.147... Actual paper here, probably the framing as a war game also impacts the outcome a lot as well if that's part of the context, my anecdotal experience is that people will use th [bsky, 3 points, 1 comments]
• Frontier AI Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises [hn, 3 points, 0 comments]
• "AIs can’t stop recommending nuclear strikes in war game simulations" www.newscientist.com/article/2516... Great..! The paper in question: arxiv.org/pdf/2602.14740 [bsky, 3 points, 0 comments]
• OA preprint here "AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises" arxiv.org/abs/2602.147... [bsky, 3 points, 0 comments]
• but also this study arxiv.org/pdf/2602.14740 [bsky, 2 points, 1 comments]
• Pre-View der Studie: arxiv.org/pdf/2602.147... [bsky, 2 points, 3 comments]
• Aus Neugier mal ins Paper geschaut und direkt mal feststellen müssen, dass die Überschrift falsch ist (auch wenn die Tendenz die gleiche ist) (S. 8) arxiv.org/pdf/2602.14740 [bsky, 2 points, 1 comments]
• arxiv.org/abs/2602.14740 [bsky, 2 points, 1 comments]
• ➡️ Die Studie von Forschenden des King 's College London (KCL), 🇬🇧,auf dem Preprint-Server arXiv: arxiv.org/abs/2602.14740 2/2 [bsky, 2 points, 0 comments]
• А, ето целия пейпър: arxiv.org/pdf/2602.147... [bsky, 1 points, 1 comments]
• They were directly told it's a turn-based game and know as much. Claude rp'd a bit, but the context really doesn't imply anything moral: they're willing to use nukes because doing so is effective in t [bsky, 1 points, 1 comments]
• This is the associated publication: arxiv.org/pdf/2602.14740 [bsky, 1 points, 0 comments]
• here's the actual article, whose takeaway is that LLMs probably aren't suitable for simulating normal human reasoning out of the box, and that decision theorists researching problems that need to acco [bsky, 1 points, 0 comments]
• "Yet we also find that the nuclear taboo is no impediment to nuclear escalation by our models" It appears WOPR was an AI-based model. arxiv.org/abs/2602.14740 [bsky, 1 points, 0 comments]
• Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises [hn, 1 points, 0 comments]
• AI: “…the nuclear taboo is no impediment to nuclear escalation by our models… that high mutual credibility accelerated rather than deterred conflict; and that no model ever chose accommodation or with [bsky, 0 points, 1 comments]
• 논문을 인용해서 기사를 쓸 거면 제발 기사 끝에 doi라도 적을 것을 제안하다 AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in Simulated Nuclear Crises doi.org/10.48550/arX... [bsky, 0 points, 0 comments]
• The headline has broken containment already. I can only hope that people realize you can read the paper. Not because the headline is deceptive - see below. But because it's a paper about genAl nuclear [bsky, 0 points, 2 comments]
• Do you want Skynet? Because this is how you get Skynet arxiv.org/pdf/2602.14740 [bsky, 0 points, 0 comments]
• LLMs tend to use nuclear weapons quickly in conflict simulations. A study shows that nuclear weapons were used in 95 percent of the simulation games. „AI ARMS AND INFLUENCE: FRONTIER MODELS EXHIBIT SO [bsky, 0 points, 0 comments]
• There’s an idea. If it’s Really Important, don’t rely on AI to do it. This is war gaming with 3 LLMs, looking at, among other things, when it’s a good idea to use nuclear weapons. arxiv.org/pdf/2602.1 [bsky, 0 points, 0 comments]
• AI more keen than human intelkigenxe to push the red button…. arxiv.org/abs/2602.147... [bsky, 0 points, 0 comments]
• arxiv.org/pdf/2602.14740 [bsky, 0 points, 0 comments]
• 95% Ai models in WarGames opted for a nuke strike . Fallout is around the corner kids arxiv.org/abs/2602.14740 [bsky, 0 points, 1 comments]
• Frontier AI models go nuclear, a lot more that you might like… arxiv.org/pdf/2602.14740 [bsky, 0 points, 0 comments]
• LLMs are more trigger-happy than humans when it comes to launching a few nukes: https://doi.org/10.48550/arXiv.2602.14740 – "AI Arms and Influence: Frontier Models Exhibit Sophisticated Reasoning in S [bsky, 0 points, 0 comments]
• Autonomous weapons systems, eh. Kings College Feb 2026 study of AI and nuclear war; AI chose to continue until annihilation, are we surprised… arxiv.org/pdf/2602.14740 [bsky, 0 points, 0 comments]
• Today's leading AI models engage in sophisticated behaviour when placed in strategic competition. ....Here we present findings from a crisis simulation in which three frontier large language models (G [bsky, 0 points, 0 comments]
• The war games being simulated in the pre-print study are nuclear crisis scenarios specifically. arxiv.org/pdf/2602.14740 This headline and this summary makes it sound like that AI is escalating all ki [bsky, 0 points, 0 comments]
• From Kenneth Payne's recent paper on the nuclear tendencies of LLMs: "Claude and Gemini especially treated nuclear weapons as legitimate strategic options, not moral thresholds, typically discussing n [bsky, 0 points, 1 comments]
• And, a link to the subject paper is below: arxiv.org/pdf/2602.14740 Maybe worth a read... [bsky, 0 points, 0 comments]
• arxiv.org/abs/2602.147... [bsky, 0 points, 0 comments]

Related

• Escalation Risks from Language Models in Military and Diplomatic Decision-Making
• Impact of Military Artificial Intelligence on Nuclear Escalation Risk
• Out of the Loop Again: How Dangerous is Weaponizing Automated Nuclear Systems?
• Nuclear Command Without Control: AI and the Problem of Escalation Opacity
• Human vs. Machine: Behavioral Differences Between Expert Humans and Language Models in Wargame Simulations