Negotiation in the game of Diplomacy · Contested (since 2022)

Meta's Cicero, combining a language model with strategic planning, scores more than double the human average over 40 online games and ranks in the top 10 percent of players with more than one game.

The games were anonymous five-minute blitz rounds on webDiplomacy; opponents did not know they were talking to a machine, and 40 games is a small sample against a mixed field rather than a tournament of the best. Cicero was designed to be honest and still produced inconsistent messages that human players sometimes noticed. Persuading and coordinating with people through language in a competitive game was shown to work; whether it holds against top players who know what they face, or in negotiations where words bind, has not been shown.