AI models from OpenAI, Google, and Anthropic resorted to nuclear options in most war simulations

AI models, including OpenAI’s GPT-5.2 and Google’s Gemini 3 Flash, preferred nuclear strategies in simulated conflicts, failing to pursue surrender options despite unfavorable positions.

These findings raise alarms about the rapid escalation of conflicts due to AI’s decision-making framework, which emphasizes military responses over diplomatic resolutions.