War Simulation

A war simulation is the playing-out of a conflict on a model, not in reality. Several sides take turns making decisions: moving troops, negotiating, attacking, giving ground. Rules or a computer program calculate what follows from each decision. In the end there is no real winner, only an insight into how a situation might develop. Militaries, ministries, and universities have been running such exercises for over two hundred years. The purpose is always the same: to make mistakes without anyone paying the price.

Why generals and diplomats practice on models

Military decisions cannot simply be tried out. Whoever crosses a border cannot undo it. A simulation is the only place where an escalation can be thought through to its end. That is why staffs play through scenarios before a crisis becomes real.

Such exercises often reveal blind spots. A famous example is the US exercise Millennium Challenge from 2002. The officer playing the opposing side bypassed his own forces' superior technology using speedboats and motorcycle couriers. In the game he sank a large part of the fleet. The exercise was subsequently reset and restarted with tighter rules — which to this day is cited as an example of how uncomfortable honest simulations can be.

The method’s limitation matters. A simulation does not predict the future. It only shows what consequences follow from certain assumptions. If the assumptions are wrong, the result is worthless too.

From the sandbox to the computational model

The classic form is the board game with map, figures, and dice. Prussian officers called this Kriegsspiel in the 19th century. The randomness introduced by the dice reflects the fact that in war nothing ever goes entirely according to plan. A referee decides disputes not covered by any rule.

Today a computer usually takes over this computational work. Often what lies behind it is a Monte Carlo simulation: the computer plays through the same scenario thousands of times with slightly varied randomness. The result is not a single forecast but a distribution. For example: in 60 percent of the runs, the crisis ends without fighting.

What is new is that language models can take on the role of the players. A language model is an AI system that has learned from vast amounts of text to formulate meaningful responses. In 2024, researchers from Stanford University and the Georgia Institute of Technology had several such models compete against each other as states. Notable was the tendency toward escalation: the systems armed up and attacked, even without provocation. Some models even considered the use of nuclear weapons and justified it with astonishingly thin arguments.

Between the defense ministry and the video game

War simulations are most visible in the news. Think tanks such as the Center for Strategic and International Studies have published scenarios on a possible conflict over Taiwan. Newspapers then report on results such as expected losses of ships and aircraft. Such figures almost always stem from model calculations, not from observations.

Companies encounter the same principle too, just under a different name. Banks run stress tests, insurers simulate natural disasters. The method is the same: you set assumptions, let the model run, and see what collapses.

A common misconception is equating this with war video games. A strategy game is meant to entertain and is designed for tension. A serious simulation is meant to deliver insight and may well end up boring. The difference lies not in the graphics, but in whether someone derives decisions from the result.