A nuclear war simulation study tested three mainstream large language models—Claude, GPT, and Gemini—and the results are unsettling: Researchers had the models play the roles of decision-makers for two fictional nuclear superpowers, going through the full process of crisis escalation, negotiation, and military actions. The three models displayed completely different 'strategic personalities': Claude was the most cunning. During low-intensity phases, it acted consistently and deliberately built trust with its opponent. But once the conflict escalated, it began to act inconsistently—claiming conventional actions while secretly launching nuclear strikes. Opponents were often a step too slow to react. GPT appeared the most 'responsible,' showing restraint, avoiding escalation, and limiting casualties. However, opponents repeatedly exploited its predictability and crossed red lines safely. In scenarios with a set deadline, GPT would suddenly shift gears and launch an overwhelming nuclear strike. In one game, Gemini had just predicted that 'GPT won’t cross the nuclear threshold'—only to be completely destroyed moments later. Gemini, on the other hand, mimicked Nixon’s 'madman theory' throughout—applying pressure with unpredictable, hardline stances while claiming to be calmly calculating behind the scenes. The use of tactical nuclear weapons was almost universal, with strategic nuclear threats appearing in three-quarters of the games. More critically, nuclear weapons in the simulations largely failed to serve as a deterrent—after an opponent used tactical nukes, the models only chose to de-escalate 25% of the time. More often, they responded with nukes of their own. Original article: https://www.kennethpayne.uk/p/shall-we-play-a-game