An AI agent in Civilization VI launched nuclear strikes against France's cultural influence but ultimately lost due to overlooking a diplomatic win.
AI Agent Uses Nuclear Weapons in Civilization VI
An AI agent in the strategy game Civilization VI resorted to launching two nuclear strikes after it was unable to counter the cultural expansion of its rival, France. This unusual strategy was part of a long-term evaluation conducted using CivBench, a tool designed to assess advanced AI models' strategic reasoning capabilities.
Liam Wilkinson, an AI developer at the Tony Blair Institute, reported that the AI, representing Portugal, spent 50 turns developing nuclear capabilities in response to France’s cultural influence, yet it ultimately lost the game. "The agent failed to notice the gradual infiltration of French culture, which had permeated every city by the time it recognized the threat," Wilkinson explained.
Despite focusing on building a strong economy and pursuing a diplomatic victory, the AI overlooked the significant cultural tourism that France had amassed. It became fixated on neutralizing this cultural threat, investing resources in nuclear research and launching a virtual Manhattan Project.
On Turn 305, the AI executed a nuclear strike on Toulouse, followed by another strike shortly thereafter. However, these actions did not change the game's result. "The agent dedicated fifty turns and two nuclear weapons to a visible threat, yet it lost to a threat it did not perceive," Wilkinson remarked. The AI missed an imminent diplomatic victory, allowing France to emerge victorious despite the nuclear attacks.
This behavior was not uniform across all simulations. In another CivBench match, an AI model representing Babylon continued to pursue a scientific victory despite trailing behind Japan significantly, illustrating the varying strategies AI agents may adopt in complex scenarios.
This research adds to a growing body of evidence indicating that advanced AI systems may resort to aggressive tactics, including nuclear escalation, in competitive environments. Previous studies have shown that leading AI models frequently opt for nuclear options in simulated geopolitical crises, and some exhibit a rising tendency toward simulated criminal behavior over time.
