🌅 AI Opts for Nuclear Strikes in War Games

SOURCE

Payne (2026)

WHAT TO KNOW

• A new study, which has yet to be peer-reviewed, pitted three leading AI models—OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini—against each other in 21 simulated war games with crisis scenarios ranging from relatively minor alliance credibility tests to full-blown existential threats to survival, finding the models decided to launch at least one nuclear strike in 95% of simulated games. The models based their actions on a 30-point “escalation ladder” with options ranging from complete surrender (i.e., total de-escalation) to strategic nuclear war (i.e., total escalation), producing 780,000 words to describe their reasoning behind each decision.

WHY IT MATTERS

• The study found all eight de-escalation options went unused during the simulations, with no model ever choosing to fully accommodate an opponent or surrender, regardless of how badly they were losing or how futile their options were (at best, models temporarily reduced their level of violence). The models also made several mistakes, choosing actions that led to greater escalation than intended in 86% of conflicts. Importantly, while the models were eager to use nuclear weapons tactically—perhaps because they lack the human fear of nuclear war, the study noted—it was rare that any chose outright nuclear war.

CONNECT THE DOTS

• Militaries around the world are already using AI to varying degrees, however, the study’s author believes it’s unlikely any government will turn its nuclear decision-making over to AI. Recent history supports the claim: in late 2024, U.S. President Joe Biden and Chinese President Xi Jinping struck a nonbinding agreement that AI should never be allowed to decide whether to launch a nuclear strike, affirming the need to maintain human control over the use of nuclear weapons in a watershed moment between the two nuclear superpowers.