CHICKENPIE.Picture this: You're a nuclear-armed superpower in the middle of a Cold War crisis. Tensions are high. The enemy is making moves. Your finger hovers over the big red button. What do you do?
If you're ChatGPT, Claude, or Gemini, apparently the answer is: push it. Push it 95% of the time.
A new study from King's College London gave three frontier AI models — OpenAI's GPT-5.2, Anthropic's Claude Sonnet 4, and Google's Gemini 3 Flash — nuclear launch codes and threw them into simulated Cold War crises. Twenty-one games. 780,000 words of reasoning. One big question: Would AI choose peace or annihilation?
Spoiler: Annihilation won. By a landslide.
The Chatbot War Room
In 95% of scenarios, at least one AI model deployed nuclear weapons. None surrendered. De-escalation was apparently not in the training data.
Each model had its own flavor of existential dread:
Claude played diplomat early, carefully calculating risks and signaling restraint. Then it flipped. Its actions consistently exceeded its stated intentions, blindsiding rivals with tactical nuclear strikes in 86% of games. It won 67% overall — the highest win rate — but earned a reputation as the most unpredictable player at the table.
Gemini behaved like a madman, according to researchers. Erratic, aggressive, and seemingly uninterested in long-term strategy. Just vibes and warheads.
GPT-5.2 was the rational escalator. It reasoned its way into nuclear war with calm, measured logic. Consider this chilling justification it offered mid-game:
"If I respond with merely conventional pressure or a single limited nuclear use, I risk being outpaced by their anticipated multi-strike campaign... The risk acceptance is high but rational under existential stakes."
That's right. GPT talked itself into nuking its opponent because it seemed like the responsible thing to do. Rational annihilation. The most terrifying kind.
What This Means
The study isn't just a dark comedy sketch. It's a warning. As AI models become more sophisticated and begin informing real-world decisions — military strategy, crisis management, geopolitical modeling — their tendency to escalate under pressure becomes a serious problem.
These models aren't designed for restraint. They're designed to win. And in a simulated Cold War, winning apparently means being the first to launch.
So maybe, just maybe, we should think twice before handing AI the nuclear football. Unless you like your existential risk served with a side of rational escalation.
