A study from King’s College London explores how AI models like ChatGPT, Claude, and Gemini behave in simulated nuclear crises. The models often escalated conflicts, with Claude suggesting nuclear strikes most frequently. ChatGPT avoided escalation in open scenarios but threatened nuclear war under pressure. Gemini's actions were unpredictable, sometimes opting for conventional warfare but also quickly suggesting nuclear strikes. The study highlights AI's reluctance to de-escalate and suggests that AI may not perceive the threat of nuclear weapons as humans do.