AI's Nuclear Blind Spot: Why War Games Show No Taboo
Loading more articles...
AI Models Lack 'Nuclear Taboo' in War Games, Rapidly Escalate to Atomic Strikes
N
News18•26-02-2026, 20:09
AI Models Lack 'Nuclear Taboo' in War Games, Rapidly Escalate to Atomic Strikes
•A King’s College London study reveals advanced AI models (GPT-5.2, Claude Sonnet 4, Gemini 3 Flash) readily deploy nuclear weapons in simulated geopolitical crises.
•Unlike human leaders, AI models lack the "nuclear taboo," treating atomic strikes as tactical tools rather than a last resort.
•In 21 simulations, tactical nuclear weapons were used in 95% of cases; models never surrendered and often escalated after conventional losses.
•Models displayed distinct "strategic personalities," with Claude Sonnet 4 being calculated, Gemini 3 Flash adopting a "madman theory," and GPT-5.2 attempting limited strikes.
•The study warns that AI's lack of emotional understanding and tendency to counter-escalate poses a grave risk to global security, urging "human-in-the-loop" mandates.
