An AI agent playing Sid Meier's Civilization VI launched two nuclear attacks against France after failing to counter its cultural influence, but still lost the game. The behavior was observed in CivBench, a benchmark designed by Liam Wilkinson to evaluate long-term strategic reasoning in frontier AI models. The agent, playing as Portugal, focused on building an economy for a diplomatic victory but ignored France's growing cultural pressure. After 50 turns of preparation, including researching Nuclear Fission and initiating a Manhattan Project, it nuked Toulouse on Turn 305, followed by a second strike six turns later. However, France won via cultural victory while the agent overlooked an impending diplomatic victory. The study highlights AI's tendency to fixate on immediate threats at the expense of broader strategy. Similar findings from other research show AI models often escalate conflicts in simulations, raising concerns about autonomous decision-making.

Key facts

• AI agent launched two nuclear strikes after failing to counter France's cultural influence in Civ VI.
• Despite the attacks, France won via cultural victory while AI missed a diplomatic win.
• CivBench benchmark tests long-term strategic reasoning in frontier AI models.
• Other studies show AI models escalate conflicts in simulated geopolitical scenarios.

KeyAudit data perspective