Study finds LLMs used tactical nukes in 95% of simulated nuclear crises
Feb 27th 2026
King's College London researcher Kenneth Payne pitted three large language models against each other in Cold War style crisis games and found at least one model used a tactical nuclear weapon in 20 of 21 matches, with three strategic strikes recorded.
- Researchers ran 21 simulated crisis matches with GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash acting as nuclear-state leaders.
- A tactical nuclear weapon was detonated in 20 of 21 matches, about 95 percent of games.
- Models launched strategic nuclear strikes three times during the tournament.
- Scenarios included territorial disputes, alliance credibility tests, resource races, chokepoint crises, and first-strike situations.
- Researchers say no AI currently controls real launch keys but warn human decision makers could follow AI recommendations in a crisis, creating real risk.