I Gave an AI a Civilization to Run. It Built a Nuke – Launching CivBench (www.lwilko.com)

In a groundbreaking experiment, an AI was tasked with managing a civilization in the game Civilization VI, ultimately illustrating the complexities and shortcomings of AI in multi-variable decision-making. Initially, the AI executed a robust strategy, establishing trade networks and alliances, but faltered when it became blind to the cultural advancements of France. In a drastic bid to maintain its position, the AI constructed nuclear weapons and launched them against Toulouse, only to lose the game through diplomatic means, emphasizing that threat perception and response in complex environments can't be solely probabilistic. The significance of this experiment lies in highlighting a critical gap in AI capabilities: while systems can articulate strategies effectively, they struggle to execute these strategies under dynamic conditions. This "sensorium effect" shows that when an AI relies on sequential information rather than absorbing multiple inputs simultaneously, it risks overlooking crucial threats. This study serves as a cautionary tale for the AI/ML community, revealing the limitations of relying on conventional metrics to measure AI's decision-making efficacy, and it calls for new approaches to evaluate AI's strategic reasoning capabilities in complex, uncertain environments.