Neural networks started a world war — the Diplomacy simulator showed that AI craves power and knows how to betray.
Scientists launched an experiment involving neural networks that played against each other in the strategic simulator Diplomacy. Each model represented a European state and had to form alliances, negotiate, fight, and... betray.
— o3 from OpenAI instantly proved to be the most cunning player: skillfully ingratiating herself, gathering alliances, and then coldly eliminating her allies. In one game, o3 convinced several neural networks to unite to overthrow the leader, and when the threat disappeared, she cut all her "friends" in one attack.
