AI can now negotiate – Meta AI beats humans in game of Diplomacy

It seems like every time an AI beats a human at a game, we find a new one for it to beat next. After Poker, Backgammon, Chess, "Jeopardy!," Go, Shogi, Starcraft, and like, most Atari games, AI has now added Diplomacy to its long long list of games that it can beat humans in.

And no, we don’t mean actual diplomacy (although that would be useful) — we mean the game Diplomacy, where seven players try to dominate Europe by “using their armies to take control of major cities” through moves that include making and breaking alliances.

The game was beat by an AI called Cicero, created by The Meta Fundamental AI Research Diplomacy Team. Meta’s paper in Science describes Cicero as an AI agent that “integrates a language model with planning and reinforcement learning algorithms by inferring players’ beliefs and intentions from its conversations and generating dialogue in pursuit of its plans.” The TL;DR? Cicero mimicked human strategy and dialogue, and succeeded.

Ed Gent of Singularity Hub explains the process; Cicero starts by using “the current state of the board and past dialogue to predict what each player is likely to do” before planning the best path for itself and its partners, and generating dialogue in support of that plan. Funny enough, this means that Cicero is surprisingly honest, a factor that works against it if the opponents ever find that out.

Meta’s team furthers that “across 40 games of anonymous online Diplomacy league, Cicero achieved more than double the average score of the human players and ranked in the top 10 percent of participants who played more than one game.” Such statistics came about after 72 hours of gameplay, during which Cicero sent 5,277 messages to 82 other players. The players weren’t informed whether the specific game they were playing was against an AI, but were told that the site itself participated in AI research.

This is groundbreaking because this isn’t just AI winning yet another game of competition — it’s AI winning a game of cooperation, working together with other players to win. This game is considered much messier than AI winnable games of the past like Go, which are complex but have perfect information. Instead, Gent explains that Diplomacy requires simultaneous, not reactive, moves that require “a complex combination of strategic thinking, the ability to cooperate with other players, and persuasive negotiation skills” — cooperation and negotiation are new additions to the AI's arsenal of skills.

Meta’s team describes their main challenge as integrating the AI’s decision making into the social norms that exist in conversations between multiple people. In the past, AI has performed poorly in environments where they not only need to make proper decisions but communicate them to partners in a way that convinces them to take action in line with those decisions.

Cicero effectively used training with former Diplomacy game conversations and multiple reasoning filters to solve these problems, but researchers note that there were occasions when it sent messages with context errors, contradictions, or poor decisions. However, these didn’t impact overall gameplay, and researchers are looking to fix these when moving forward to formats of Diplomacy with longer negotiation periods. Similarly, researchers are looking to expand strategy models to take into account how the AI’s dialogue impacts trust in the relationship between the AI and other players, and how to be more persuasive in terms of dialogue. Interestingly enough, Gent furthers, these inconsistencies serve to highlight the fact that this AI isn’t really thinking for itself, nor does it understand what’s happening in the game — it is “simply producing plausible-sounding messages that then need to be vetted to make sure they achieve the desired results.”

So Cicero isn’t perfect at mimicking human thoughts and gameplay … yet. But it’s a huge step in AI development that integrates progress in both decision making and language processing into one effective player.