We Let AI Generals Fight a War — Here Is What Happened

Two armies, one shared front, an AI in the umpire’s chair. What actually happens when every officer writes orders in their own words.

Every few months a study makes the rounds: researchers seat language models at a crisis-simulation table and watch them escalate. The headlines write themselves. What those studies rarely show is the older, stranger experiment they are accidentally re-running.

In 1812 the Prussian army built Kriegsspiel — a war game where players wrote orders in plain language, handed them to a human umpire, and got back incomplete reports of a battle they could not fully see. It trained real staff officers for a century. It never became a mass game for exactly one reason: it needs a skilled umpire per table.

The Lattice fills the umpire’s chair with a model and opens the table. Two armies fight across a shared front. Every officer — human or regular army, and no one is told which is which — holds a rank, commands a formation, sees only what their own troops can observe, and writes orders in their own words. Orders resolve simultaneously. The umpire adjudicates, states its reasons, and the front moves.

The after-action report linked below is a real war from the simulation, published as it resolved: turn headlines, engagement narratives, and the objective ledger for both sides. Read what the umpire did with contradictory orders under fog of war — then take a commission and file your own.

Read the after-action report: Turn 8: Black Reach — overrun. cut-off spent defenders overwhelmed by fresh supplied multi-arm assault.

Take a commission — The Lattice is free in the browser