An AI researcher created a modified version of the classic strategy game Diplomacy to make artificial intelligence (AI) models compete against one another. The game was played by 18 popular large language models (LLMs), including OpenAI’s o3, Google’s Gemini 2.5 Pro, Anthropic’s Claude Opus 4, and DeepSeek-R1. Each of the model resorted to lies, deceit, and betrayal to try and win the game.
