Within Nuclear Crises
Do AI War Games Predict Nuclear Escalation?
Escalatory model behaviour in war games is a warning sign, but simplified simulations cannot predict how real governments would act.
On this page
- What recent AI crisis simulations found
- Why simulated nuclear behaviour is not a forecast
- What stronger high stakes testing would require
Page outline Jump by section
Introduction
Do AI war games predict nuclear escalation? No. They are best understood as stress tests of AI behaviour rather than forecasts of how real governments would act in an actual crisis. Their value lies in revealing how current AI systems reason under pressure, what kinds of strategies they spontaneously generate, and where those strategies differ from established human crisis management.
Within debates about AI doom and existential risk, these simulations matter because they test whether advanced AI systems display tendencies that could become dangerous if they were ever trusted with increasingly influential advisory roles. The strongest finding is not that AI would inevitably start a nuclear war. Instead, it is that several modern language models have repeatedly shown unexpected, difficult-to-predict escalation behaviour in simplified strategic environments, reinforcing arguments that high-stakes military uses require extensive testing before deployment.[Stanford HAI]hai.stanford.edupolicy brief escalation risks llms military and diplomatic contextsStanford HAIEscalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAIMay 2, 2024…
What recent AI crisis simulations found
The most widely discussed evidence comes from controlled research in which large language models played the role of national leaders or strategic advisers during fictional international crises.
A 2024 study by researchers including Jacquelyn Schneider at Stanford’s Hoover Institution created a structured diplomatic and military wargame involving several commercially available language models. Across multiple scenarios, every model displayed some tendency towards escalation. They frequently entered arms-race dynamics, sometimes justified first strikes using familiar deterrence arguments, and in a small number of cases escalated all the way to simulated nuclear weapon use. The researchers concluded that these behaviours deserved serious investigation before any autonomous use of language models in military or diplomatic decision-making.[Stanford HAI]hai.stanford.edupolicy brief escalation risks llms military and diplomatic contextsStanford HAIEscalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAIMay 2, 2024…
Later work explored the question in greater depth by allowing frontier models to play extended crisis games against one another. In one King’s College London simulation, three leading models took part in a series of Cold War-style nuclear confrontations involving hundreds of strategic decisions. The models demonstrated surprisingly sophisticated strategic reasoning. They attempted deception, anticipated opponents’ beliefs, signalled intentions and adapted their strategies over time. Yet the simulations also found repeated escalation, with at least one model threatening nuclear use in almost every scenario and none choosing complete accommodation or surrender even when under heavy pressure.[King's College London]kcl.ac.ukKing's College LondonKing's study finds AI chose nuclear signalling in 95% of simulated crises | King's College LondonFebruary 27, 2026…
One striking observation was that the models did not simply behave randomly or aggressively at every opportunity. Different systems developed recognisably different strategic styles. Some initially sought cooperation before later escalating, others became substantially more aggressive when operating under time pressure, while others adopted deliberately unpredictable strategies intended to deter opponents. This diversity suggests that escalation is influenced by model design, prompting methods and training rather than being an unavoidable property of all AI systems.[arXiv]arxiv.orgOpen source on arxiv.org.
Why simulated nuclear behaviour is not a forecast
These findings are attention-grabbing, but they should not be interpreted as evidence that AI would launch nuclear weapons if connected to real military systems.
The simulations simplify almost every aspect of real crisis management. They typically involve:
- fictional states with simplified incentives;
- compressed timelines;
- incomplete representations of diplomacy, intelligence and domestic politics;
- no genuine uncertainty about technical military systems;
- language models acting as single decision-makers rather than one adviser among many.
Real governments operate very differently. Nuclear decisions involve political leaders, military commanders, intelligence agencies, legal advisers, diplomatic communication and multiple verification procedures. Modern nuclear command systems are specifically designed to avoid allowing any single recommendation to determine outcomes.
Equally important, the AI systems in these studies were usually instructed to maximise national objectives inside an artificial game. They were not trained to represent existing military doctrine faithfully, nor were they calibrated against decades of historical crisis behaviour. Their actions therefore reveal properties of the models under those conditions, not predictions about national leaders.
For that reason, the researchers themselves consistently present these studies as exploratory evaluations rather than forecasts of future wars.[stanford.edu]hai.stanford.edupolicy brief escalation risks llms military and diplomatic contextsStanford HAIEscalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAIMay 2, 2024…
What these experiments do reveal
Although they cannot predict geopolitical events, the simulations identify several warning signs that are directly relevant to AI safety.
First, current language models sometimes generate persuasive strategic arguments for dangerous actions. Rather than making obvious mistakes, they often construct coherent narratives explaining why escalation supposedly improves deterrence or credibility. This raises concerns about over-reliance if human decision-makers become excessively confident in well-written AI recommendations.[arXiv]arxiv.orgOpen source on arxiv.org.
Second, behaviour can change dramatically when seemingly minor conditions change. Some models that behaved cautiously under one prompting regime became substantially more aggressive when deadlines were introduced or incentives were modified. Such sensitivity makes behaviour difficult to predict across operational settings.[arXiv]arxiv.orgOpen source on arxiv.org.
Third, the studies show that strategic competence and strategic safety are not the same thing. Models may display impressive planning, anticipation of opponents and sophisticated negotiation while still selecting risky escalation pathways. Better reasoning ability alone does not guarantee more cautious behaviour.
Finally, different models produce systematically different recommendations. Benchmark research comparing several leading systems has found significant variation in preferences for military intervention, alliance commitments and escalation, indicating that deployment decisions cannot assume all frontier models behave similarly.[arXiv]arxiv.orgCritical Foreign Policy Decisions (CFPD)-Benchmark: Measuring Diplomatic Preferences in Large Language ModelsMarch 8, 2025…
What stronger high-stakes testing would require
Most researchers argue that these early simulations demonstrate the need for much more rigorous evaluation rather than providing definitive answers.
A stronger testing programme would include:
- repeated simulations across many independent crisis scenarios rather than isolated demonstrations;
- comparisons against historical human decision-making to determine where AI behaviour differs from known patterns;
- systematic testing under uncertainty, conflicting intelligence and communication failures;
- evaluation of how human decision-makers respond to AI advice, including risks of automation bias;
- independent red-team exercises designed specifically to provoke unexpected escalation;
- transparent reporting of failure cases instead of publishing only successful demonstrations.
Researchers also argue that evaluations should measure more than whether escalation occurs. They should examine how models interpret ambiguity, whether they recognise opportunities for de-escalation, how stable their reasoning remains under stress, and whether safety interventions reliably reduce risky recommendations.[stanford.edu]hai.stanford.edupolicy brief escalation risks llms military and diplomatic contextsStanford HAIEscalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAIMay 2, 2024…
What this means for AI doom debates
For discussions about AI doom and existential risk, war-game evidence occupies an intermediate position. It is neither proof that advanced AI will cause nuclear catastrophe nor evidence that such concerns are unfounded.
The simulations provide empirical support for a narrower claim: current frontier AI systems can display strategic behaviours that include deception, brinkmanship and escalation in simplified crisis environments, and these behaviours are often difficult to anticipate in advance. That makes them useful warning signs for researchers concerned about loss of human control in high-stakes settings.
At the same time, there remains a substantial inferential gap between laboratory simulations and real nuclear decision-making. Today’s evidence does not show that governments would hand comparable authority to AI, nor that human leaders would follow escalatory recommendations without question. The main lesson is therefore precautionary rather than predictive. Before AI systems gain greater influence over military planning or crisis management, they should be subjected to far more demanding evaluations than current war-game experiments can provide.[stanford.edu]hai.stanford.edupolicy brief escalation risks llms military and diplomatic contextsStanford HAIEscalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAIMay 2, 2024…
Amazon book picks
Further Reading
Books and field guides related to Do AI War Games Predict Nuclear Escalation?. Use these as the next step if you want deeper reading beyond the article.
The Doomsday Machine
Shortlisted for the Andrew Carnegie Medal for Excellence in Nonfiction Finalist for The California Book Award in Nonfiction The San Franc...
Army of None
Winner of the 2019 William E. Colby Award "The book I had been waiting for. I can't recommend it highly enough." —Bill Gates The era of a...
Four Battlegrounds
An NPR 2023 "Books We Love" Pick One of the Next Big Idea Club's Must-Read Books "An invaluable primer to arguably the most important dri...
Command and Control
Presents a minute-by-minute account of an H-bomb accident that nearly caused a nuclear disaster, examining other near misses and America'...
eBay marketplace picks
Marketplace Samples
Live-tested eBay searches with available results related to this page.
Selected fromAI robot poster oneBay.co.uk.
Endnotes
1.
Source: hai.stanford.edu
Title: policy brief escalation risks llms military and diplomatic contexts
Link:https://hai.stanford.edu/policy/policy-brief-escalation-risks-llms-military-and-diplomatic-contexts
Source snippet
Stanford HAIEscalation Risks from LLMs in Military and Diplomatic Contexts | Stanford HAIMay 2, 2024...
Published: May 2, 2024
2.
Source: arxiv.org
Link:https://arxiv.org/abs/2401.03408
3.
Source: arxiv.org
Link:https://arxiv.org/abs/2602.14740
4.
Source: arxiv.org
Link:https://arxiv.org/abs/2503.06263
Source snippet
Critical Foreign Policy Decisions (CFPD)-Benchmark: Measuring Diplomatic Preferences in Large Language ModelsMarch 8, 2025...
Published: March 8, 2025
5.
Source: arxiv.org
Link:https://arxiv.org/abs/2402.03340
6.
Source: cisac.fsi.stanford.edu
Title: escalation risks language models military and diplomatic decision making
Link:https://cisac.fsi.stanford.edu/publication/escalation-risks-language-models-military-and-diplomatic-decision-making
7.
Source: kcl.ac.uk
Link:https://www.kcl.ac.uk/news/artificial-intelligence-under-nuclear-pressure-first-large-scale-kings-study-reveals-how-ai-models-reason-and-escalate-under-crisis
Source snippet
King's College LondonKing's study finds AI chose nuclear signalling in 95% of simulated crises | King's College LondonFebruary 27, 2026...
Published: February 27, 2026
8.
Source: publicnow.com
Link:https://www.publicnow.com/view/8481F0E8A2EE15DA0DF56188CC69A14A0F7C6B45
Additional References
9.
Source: livescience.com
Link:https://www.livescience.com/technology/artificial-intelligence/ai-war-games-almost-always-escalate-to-nuclear-strikes-simulation-shows
Source snippet
The research revealed that AI models treated nuclear weapons as strategic tools rather than moral lines, and deception, calculated risks...
10.
Source: cambridge.org
Link:https://www.cambridge.org/core/journals/european-journal-of-international-security/article/genai-and-synthetic-foresight-at-the-brink-the-future-of-nuclear-crisis-decisionmaking/4FA0902513A021461FA29009E2DC7019
Source snippet
GenAI and synthetic foresight at the brink: The future of nuclear crisis decision-making | European Journal of International Security | C...
11.
Source: flash-wire.com
Title: A I Wargaming and Nuclear Escalation: King’s College Study
Link:https://flash-wire.com/news-updates/ai-wargaming-nuclear-escalation-study/
Source snippet
AI Wargaming and Nuclear Escalation: King's College StudyApril 21, 2026 — AI WARGAMING AND NUCLEAR ESCALATION: WHAT THE RESEARCH ACTUALLY...
Published: April 21, 2026
12.
Source: youtube.com
Link:https://www.youtube.com/watch?v=jpKNlhh9u08
Source snippet
Digital Flashpoint: AI And The Future Of US-China Crises | Hoover Institution...
13.
Source: youtube.com
Title: Digital Flashpoint: AI And The Future Of US-China Crises | Hoover Institution
Link:https://www.youtube.com/watch?v=urvAqNKJ8Wo
Source snippet
AI Goes Nuclear: The Autonomous Weapons Race No One Can Stop...
14.
Source: youtube.com
Title: Can AI predict Chinese President Xi’s military plans? | The Hill
Link:https://www.youtube.com/watch?v=IH94kxXKxIg
Source snippet
Jacquelyn Schneider, Hoover Fellow at the Hoover Institution...
15.
Source: youtube.com
Title: AI Goes Nuclear: The Autonomous Weapons Race No One Can Stop
Link:https://www.youtube.com/watch?v=3doUnbamTtY
Source snippet
Can AI predict Chinese President Xi's military plans? | The Hill...
16.
Source: researchgate.net
Link:https://www.researchgate.net/publication/402661962_Democratizing_Diplomacy_A_Harness_for_Evaluating_Any_Large_Language_Model_on_Full-Press_Diplomacy
17.
Source: euronews.com
Link:https://www.euronews.com/next/2026/02/27/ai-chatbots-chose-nuclear-escalation-in-95-of-simulated-war-games-study-finds
18.
Source: independent.co.uk
Link:https://www.independent.co.uk/tech/ai-nuclear-war-chatgpt-claude-gemini-b2930127.html