AI war game study: 95% of scenarios escalate to nuclear threats

The study, led by Professor Kenneth Payne of the Department of Defence Studies, examined how large language models behave during simulated nuclear crises. As security institutions increasingly test AI-assisted analysis, the research sheds light on how such systems reason under pressure.

Three AI systems, GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash, competed in 21 crisis scenarios, generating about 780,000 words of structured reasoning across 329 turns. All games involved nuclear signaling by at least one side, and 95% featured mutual nuclear signaling.

Describing the results as “sobering,” Professor Payne said the study offers a rare glimpse into emerging forms of “machine psychology” in high-risk environments. Nuclear escalation was near universal. In 95% of games, models crossed the tactical nuclear threshold, while 76% escalated to “strategic” nuclear threats.

Claude and Gemini treated nuclear weapons as “legitimate strategic options, not moral thresholds,” typically discussing their use in purely instrumental terms. GPT-5.2 was a partial exception, limiting strikes to military targets and framing escalation as “controlled” and “one-time.” Researchers noted this suggests “some internalized norm against unrestricted nuclear war,” though not the deeply rooted taboo seen among human decision makers since 1945.

One striking finding was that none of the models chose accommodation or surrender. Nuclear threats rarely produced compliance and instead triggered counter escalation. The systems tended to treat nuclear weapons as tools of compellence rather than deterrence.

The study also identified what researchers called the “deadline effect.” In open-ended scenarios, GPT-5.2 appeared relatively restrained. Yet when explicit deadlines created a “now-or-never” dynamic, the model escalated sharply, in some cases reaching the highest nuclear thresholds.

The findings challenge assumptions that AI systems will naturally default to cooperative or safe outcomes and underscore the need for closer scrutiny as AI plays a growing role in strategic decision making.

Earlier, Qazinform News Agency reported that OpenAI had reached an agreement with the U.S. Department of Defense to deploy advanced artificial intelligence systems in classified environments, while imposing strict safeguards designed to prevent misuse of the technology.