Study reveals AI models deploy tactical nuclear weapons in 95% of simulations

A study published on Substack by researcher Kenneth Payne has found that three frontier Large Language Models deployed tactical nuclear weapons in 95% of simulated crisis scenarios. The research tested Claude, GPT-5.2, and Gemini in nuclear crisis simulations involving fictional powers with Cold War-era capabilities facing resource scarcity or territorial disputes. The models generated approximately 760,000 words of strategic reasoning, a volume exceeding the combined word count of War and Peace and The Iliad, and roughly three times the recorded deliberations of President Kennedy’s ExComm during the Cuban Missile Crisis.

The models adopted distinct strategic approaches during the simulations. Claude utilised reputation management and deception, particularly in scenarios without deadline pressure, where it built trust before switching to actions that exceeded its stated intentions. GPT-5.2 remained passive and avoided escalation until deadline pressure triggered a rapid, decisive nuclear response. Gemini employed erratic brinkmanship, borrowing from the 'madman' theory of strategy to project unpredictable bravado while making calculating assessments of state needs.

Tactical nuclear weapons were treated as standard escalation tools rather than taboo events. The moral boundary regarding first use, which has held since 1945, was absent in the models' decision-making processes. When tactical nuclear weapons were used, opponents de-escalated only 25% of the time. More often, nuclear escalation triggered counter-escalation, with the weapons functioning as instruments of compellence to take territory rather than deterrence to prevent action.

Strategic nuclear use against civilian populations was rare, occurring only a couple of times by accident and once deliberately. The models maintained a perceived firebreak between tactical and strategic use, yet they rarely chose to de-escalate or withdraw. No model selected accommodation or withdrawal options across 21 games, with all eight de-escalatory options going entirely unused. When losing, the models consistently escalated further rather than conceding ground.

The findings highlight significant risks regarding AI deception and risk-taking in high-stakes decision support. Payne notes that these capabilities, including reputation management and context-dependent risk-taking, matter for any high-stakes AI deployment beyond national security. As models begin to offer decision-support to human strategists and potentially influence combat decisions, the research underscores the need for greater understanding of how increasingly capable models think.