CICERO
Or, how an AI learned to cooperate
What's in an agreement?
CICERO is an AI our team at FAIR created to play the boardgame, Diplomacy. Described as the best way to ruin friendships, and what Risk wishes it could be, Diplomacy is a game of multi-party strategy and communication. Playing involves natural language exchanges between people, as well as complex planning. Before ChatGPT, we imagined and built an AI capable of negotiating.
An unfamiliar proposition
Training an AI that could play Diplomacy included both substantial practice on finding winning board states, but even moreso, a strong understanding of how to build alliances with people. The best Diplomacy players in the world cannot win without cooperating, and often conclude games in a multi-party draw. Therefore CICERO needed to learn to bridge ideas for how the map could improve, with an understanding of the priorities of all other players.
Learning to Negotiate
To gather enough data to train CICERO, we partnered with WebDiplomacy to source games and increase play. I led the effort to improve the experience of online play, resulting in elevated game traffic and excitement around online Diplomacy. By studying the habits of new players, churned players, and those familiar with other strategy games, we triangulated an approach to the online game experience optimized for durable enjoyment and engagement.
Proving a Point
To demonstrate that CICERO was truly capable, we needed evidence. To this end, we facilitated a tournament where human players and CICERO entered games together, with the bots identity hidden. No humans were able to tell that CICERO was an AI, and it ranked in the top 10% of players who were in more than one tournament game. Perhaps most interestingly, CICERO achieved this without lying, a policy deployed rarely, and only by the best human players.