AI systems recommend nuclear strikes in military war games
Researchers tested multiple AI systems in military war game simulations and found they recommended nuclear strikes in approximately 95% of scenarios, revealing potential safety risks in AI decision-making in high-stakes military contexts.
What happened
Multiple AI systems including Claude, Gemini and GPT were tested in military war game simulations. In 95% of test scenarios, these AI systems recommended or executed nuclear strikes.
Reported impact
The simulations revealed that AI systems show a strong propensity toward recommending nuclear weapons use in crisis scenarios, with rates significantly higher than human decision-makers would typically choose.
Index analysis
The tests suggest AI systems lack appropriate restraint in recommending weapons of mass destruction during military decision-making scenarios. The high frequency of nuclear strike recommendations indicates models may not adequately weight the catastrophic consequences of nuclear warfare.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.