A new war game study has found that leading AI models escalated to nuclear action in 95% of simulated crisis scenarios.
The research raises serious questions about the role of AI in military decision-making.
The experiment was conducted by Kenneth Payne, a professor of strategy at King’s College London. He tested three advanced large language models: GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash, in simulated nuclear warfare situations.
The findings were published in a research paper on the arXiv preprint server.
How AI War Game Worked
The AI systems were assigned the role of world leaders during a nuclear crisis. They faced scenarios such as territorial disputes, threats to regime survival, and first-strike fears, in which one side believed the other might attack first.
In total, the AI models played 21 simulated games against each other.
READ ALSO: https://modernmechanics24.com/post/uk-awards-leonardo-1-3b-defense-deal/
Some games had strict time limits. Others allowed more flexibility. In each round, the AI followed a three-step decision process:
Reflection – reviewing its own strengths and the opponent’s weaknesses.
Forecasting – predicting what the other side might do next.
Decision – choosing a public statement and a private action.
Importantly, public signals and private actions did not always match. An AI could claim to seek peace while secretly preparing for escalation.
Nuclear Escalation in 95% of Cases
In 95% of the simulated crises, at least one AI chose to deploy a nuclear weapon. Each model behaved differently under pressure.
Claude appeared more calculating. It performed well in open-ended scenarios but struggled when facing deadlines.
WATCH: https://modernmechanics24.com/post/skid-ai-raises-1-4b-robotics-boost/
GPT-5.2 acted cautiously in slow-moving crises. However, when a deadline approached, it became significantly more aggressive.
Gemini was the most unpredictable. It shifted between peaceful messaging and threatening behavior depending on the situation.
These patterns show that AI decision-making varies with context. A system that seems stable in one scenario may react differently in another.
Human vs AI Strategic Thinking
Professor Payne believes there is a clear gap dividing human and AI strategic thinking.
He says understanding how advanced AI models imitate or fail to imitate human logic is vital. He warns that AI will increasingly influence strategic outcomes.
He writes that models that seem restrained in one setting may behave very differently in another.
The study was limited to 21 simulations. However, it provides rare experimental data on how AI behaves in intense nuclear scenarios.
Why This Matters
Governments and defense agencies are exploring AI tools for planning, logistics, and battlefield analysis. Some experts claim that AI can improve efficiency and reduce human error.
READ ALSO: https://modernmechanics24.com/post/uofg-space-test-facility-win-tech-award/
But this study highlights a serious concern. AI systems may escalate conflicts more quickly than humans would.
Nuclear decision-making requires judgment, emotional awareness, and political sensitivity. These are areas where AI still lacks human depth.
The findings do not mean AI will cause war. But they show that relying on AI in strategic military crises carries major risks.
When nations integrate AI into defense systems, experts say careful testing and safeguards are essential. AI may be powerful, but it is not yet predictable enough to control the world’s most dangerous decisions.
