The games are on github (https://github.com/kennethpayne01/project_kahn_public/blob/m...) which might give better context as to how the simulation was run. Based on the code the LLMs only have a rough idea of the rules of the game. For example you can use 'Strategic Nuclear War' in order to force a draw as long as the opponent cannot win on the same turn. So as long as on your first turn you do 'Limited Nuclear Use' then presumably its impossible to actually lose a game unless you are so handicapped that your opponent can force a win with the same strategy. I suspect with knowledge of the internal mechanics of the game you can play in a risk free way where you try to make progress towards a win but if your opponent threatens to move into a winning position then you can just execute the 'Strategic Nuclear War' action.

From the article:

> They also made mistakes in the fog of war: accidents happened in 86 per cent of the conflicts, with an action escalating higher than the AI intended to, based on its reasoning.

Which I guess is technically true but also seems a bit misleading because it seems to imply the AI made these mistakes but these mistakes are just part of the simulation. The AI chooses an action then there is some chance that a different action will actually be selected instead.