In November 2022, Meta announced Cicero, which it described as the first AI to reach human-level performance in Diplomacy, and published the work in the journal Science. Diplomacy is unusual among games studied by AI. There is no dice and no hidden board, but winning depends on talking to the other players: forming alliances, negotiating joint moves, persuading, and at times betraying. It is a test of social reasoning, not just calculation, which is why it had resisted the techniques that conquered chess, Go, and poker.
Cicero’s approach was to join two strands of AI that had largely developed apart. One was natural language processing: a dialogue model that could hold free-form conversations with human players in plain English. The other was strategic reasoning: a planning engine, in the lineage of Meta’s earlier poker work, that worked out sound moves while anticipating what others would do. The crucial design choice was to ground the conversation in the plan. Rather than chatting freely and risking incoherent or dishonest promises, Cicero generated dialogue tied to the strategy it actually intended to pursue, so its words and its actions lined up.
The system performed well against people. In an online league on webDiplomacy.net, Cicero achieved more than double the average human score and finished in the top ten percent of players who played more than one game, often without other participants realizing they were negotiating with a machine.
Why business readers should care: most prior game-playing milestones lived in worlds of perfect competition with fixed rules. Cicero tackled a setting closer to real business life, one where progress depends on cooperation, persuasion, and managing relationships under uncertainty. It was an early, concrete demonstration that an AI could combine language and strategic planning to coordinate with people rather than simply outplay them.