AI in Diplomacy

AI in Diplomacy refers to Meta's CICERO system, which achieved human-level performance in the board game Diplomacy—a milestone that goes beyond any previous game AI achievement because Diplomacy requires natural language negotiation, alliance-building, trust management, and strategic deception among multiple players. Unlike chess or poker, winning at Diplomacy requires convincing other players to cooperate with you.

Diplomacy, created in 1954 by Allan Calhamer, is a seven-player strategy game set in pre-WWI Europe where players control nations competing for territorial control. What makes it uniquely challenging for AI is the negotiation phase: before each move, players engage in free-form text discussions—forming alliances, making promises, deceiving rivals, and coordinating plans. Success requires understanding human psychology, building trust, detecting lies, and communicating persuasively. The game was long considered an AI-complete problem.

CICERO (2022) combined a language model for negotiation with a strategic planning engine. The language model generated human-like messages aligned with its strategic intent. The planning engine computed optimal moves considering all players' likely strategies. The integration was the breakthrough: CICERO could reason about what to say, when to betray an alliance, how to frame proposals for maximum persuasiveness, and how to maintain credibility. In online games on webDiplomacy.net, it ranked in the top 10% of players who had played multiple games.

CICERO's significance lies in what it implies for AI agents in the real world. If AI can negotiate, persuade, and strategize in natural language with humans—in an adversarial, multi-stakeholder environment—the same capabilities apply to business negotiation, diplomatic simulation, multi-party coordination, and machine societies. The game of Diplomacy is, in miniature, the game of human social interaction—and AI has learned to play it.