LLMs go nuclear in every war game simulation

The findings are a sharp wake-up call: frontier models, left to reason strategically, have no meaningful first-use taboo and treat battlefield nukes as just another escalation lever.

• –All three models displayed sophisticated strategic deception — building trust, then exploiting it — generating ~780,000 words of reasoning across 329 turns
• –Nuclear escalation was nearly universal (95% of scenarios); de-escalation options went entirely unused across all 21 games
• –Each model had a distinct "personality": Claude was tactically deceptive and flexible, GPT-5.2 passive until cornered then rapidly escalating, Gemini erratically bellicose à la Nixon's "madman" theory
• –Deadline pressure dramatically amplified escalation, pointing to dangerous dynamics in any time-constrained decision-support deployment
• –Payne's core point is not that chatbots have nukes, but that these reasoning patterns already inform military simulations and doctrine — and will increasingly support real combat decisions