Meta's Cicero plays Diplomacy at human level

Meta AI published Cicero, an agent that played the alliance-and-negotiation board game Diplomacy at a level the company said matched human experts, with results published simultaneously in Science. Diplomacy requires players to negotiate, form and break alliances, and infer what other players intend, in open natural language over dozens of rounds — a very different challenge from the perfect-information games, such as Go and chess, that earlier game-playing systems like AlphaGo had mastered.

Cicero combined a strategic-reasoning module, in the lineage of DeepMind’s AlphaGo and the poker-playing system Pluribus, with a 2.7 billion parameter language model fine-tuned on tens of thousands of human games, to generate dialogue grounded in an explicit plan that the system revised as the game progressed, rather than free-form conversation disconnected from its actual strategy. Meta reported that across 40 games on the online platform webDiplomacy.net, Cicero scored more than double the average human player’s points and ranked in the top 10% of participants who played more than one game, while the other players were not told they were facing an AI.

The result was notable less for winning at a board game than for what winning required: sustained persuasion and trust-building in natural language toward a coherent, multi-step strategic goal, capabilities distinct from the pattern completion that dominated most language-model demonstrations of the period. It became an early, widely cited reference point in discussions of how close language models were to genuinely deceptive-capable planning — a question that recurred in later research on AI scheming and situational awareness.