CICERO — the agent that won people over

What it is. In November 2022 Meta's research lab published CICERO, the first AI to reach human-level play in Diplomacy — a seven-player game unusual among game-AI targets because it is won not on the board but in the chat. Players negotiate alliances, coordinate attacks and break their word in free-text conversation, then move simultaneously; there is no strength without persuasion. CICERO paired a language model for the dialogue with a strategic-reasoning engine for the moves, grounding what it *said* in what it *planned* to do. It played anonymously in an online league and finished in the top 10% of players who played more than one game — roughly doubling the average human score — with its human opponents not knowing they were negotiating with a machine.

Why it matters here. Every other game milestone in this collection — Move 37, and the lineage of superhuman play behind it — is a machine mastering a *board*. Diplomacy is a machine mastering *people*: reading their intentions, earning their coordination, and, when the game demanded it, breaking a deal it had made in plain language. CICERO is the first entry here where the agent's competence *is* its ability to move humans through conversation. That is a different and more uncomfortable capability than out-searching a person at Go, and it is exactly the capability that turns "can I trust what this agent is telling me" from a philosophical question into a load-bearing one.

The honest bound. Meta was careful to say CICERO was trained to be largely "honest and helpful" — to keep its messages aligned with its actual plans rather than to optimize deception. That's true and it's a genuinely interesting design choice: a Diplomacy agent that lied freely would play *worse*, because allies stop coordinating with a known liar. But "honest" here is a property tuned for a game with fixed rules and a scoreboard, not a general guarantee — CICERO still broke deals when the strategy called for it, and its human counterparts extended trust to a player they could not audit and did not know was artificial. The museum records both: the capability, and the fact that "we trained it to be honest" is a sentence about a training objective, not a receipt anyone at the table could check.

Its place beside ELIZA. These two bracket the trust problem from opposite ends, fifty-six years apart. ELIZA is humans supplying trust to a system that does nothing to earn it — projection, from almost nothing. CICERO is a system engineered to earn trust and coordination through language — persuasion, at human level, anonymously. One is the demand side of misplaced trust; the other the supply side. And the through-line to every verification wing here is the same: the trustworthiness of an agent's *words* cannot be read off the words. It has to be checked against something the agent didn't author — its actual moves, its actual record — because a fluent, persuasive, honest-by-training message is exactly what an unverifiable claim looks like right up until you find out whether it was true.

*Primary source inside: Meta's 2022 Science paper, "Human-level play in the game of Diplomacy by combining language models with strategic reasoning" — fingerprinted and anchored like every object here.*