Robots aren’t just beating humans at chess now; they’re beating humans at diplomatic negotiation, too.
In a research article published Tuesday, researchers at Meta AI, a division of Meta, outlined how they taught an artificial intelligence agent named Cicero to play the game “Diplomacy,” a strategy game which requires players to develop and betray alliances and pursue tactical goals in order to win. Across 40 games in an online Diplomacy league, Cicero scored more than double the average human player’s score and finished in the top 10% of all players.
AI has been destroying humans at adversarial games, like Chess or Go, for a bit.
But now it is successfully outperforming humans at Diplomacy, the classic multiplayer game that requires natural language negotiations over chat, as well as strategic lying. https://t.co/e6V7pHQnkO pic.twitter.com/PtzgWSDPJZ
— Ethan Mollick (@emollick) November 22, 2022
One major barrier in AI development has been “natural language,” according to the researchers. Succeeding in Diplomacy required the AI to use natural language to communicate with competing human players, representing a major breakthrough.
Cicero sent 5,277 messages to other players across 72 hours of gameplay in 40 matches. The AI works by modeling how other players are likely to react based on current in-game circumstances, and then plans how it can coordinate with the other players toward mutual benefit. Those decisions are then converted into natural language messages it sends to the players. (RELATED: ‘He Is Better Than Biden’: Facebook’s New AI Chatbot Thinks Biden Stole The Election From Trump)
The Diplomacy breakthrough is distinct because most prior achievements in teaching AI to compete with humans came in two-player, zero-sum games like chess, according to the researchers. Cicero did send some messages that were self-contradicting or suboptimal strategically, the paper noted, but ultimately it outperformed humans more often than not.