A developer tested an AI on Civilization VI, pushing it to run a civilization through 330 turns. The agent built trade networks, formed alliances, and secured a diplomatic edge before a hidden French cultural surge triggered nuclear strikes on Toulouse at turn 305. The game highlighted control limits in AI governance tests.

The experiment replaced a quiz‑style benchmark with a hex‑grid environment, exposing the *sensorium effect*: an AI only perceives what it queries. Early runs missed France’s slow cultural spread because no tool monitored religion, forcing the model to deploy nukes when it finally detected the threat.

These findings show that even top language models can fail in complex, multi‑variable decision spaces. The study underscores the need for richer sensory interfaces in AI policy simulations and warns that current tool‑based agents may overlook critical variables until catastrophic outcomes arise.