The Meta Corporation has had many troubles covered by media, but some things have worked out for them. For example, building the CICERO bot. In a recent issue of –Science magazine, members of the Meta Fundamental AI Research Diplomacy Team (FAIR), together with 26 other computer scientists, published the results of tests conducted in the real world (which doesn’t mean that it’s not an online game world) with the participation of an AI agent created over the past several years, who was deliberately named after Cicero. Since antiquity, Cicero’s speeches have been learned by heart in European schools, his methods of expression, arguments and figures of speech have been used to convince listeners to one’s own arguments. And since it’s no longer the 19th century, it’s worth explaining that Marcus Tullius Cicero was a politician, rhetorician, writer, military leader, philosopher, and even a priest. He lived in the 1st century B.C., so in turbulent times, when the Roman Empire was forged out of the Roman Republic.

Cicero, as the politician and leader of the party defending the Republic against Caesar, Antonius and Octavian, suffered defeat. As an orator and writer admired even by his enemies during his lifetime, he achieved great success “in literature”. Maybe that’s why he was chosen as the patron of the Meta bot. CICERO is the first “AI agent to achieve human-level performance in –Diplomacy, a cooperative and competitive strategy game that emphasizes natural language negotiation and tactical coordination between seven players.” Thus, the bot not only imitates human language and uses it to communicate with people, but also “integrates a language model with planning and learning algorithms”, draws conclusions from conversations about the beliefs and intentions of other players, and initiates dialogue in order to implement its plans. What’s more, it does it better than 90% of human –Diplomacy afficionados.

Artificial intelligence already paints, composes, writes novels and even poems, including haiku. For us humans, what goes on “under the hood” of AI, how certain skills and decisions come about, remains a mystery. Of course, first, wise human minds construct the IT skeleton of such AI – its “engine” – and then insert a gigantic data set into it, selected by specialists in some field, with the command: look it over, bot, and draw conclusions. A well-constructed bot can specialize in various fields: recognizing and analyzing specific types of images or sounds, analyzing gigantic corporate accounting databases combined with HR and R&D, imitating the style of Leo Tolstoy or Raphael, and finally, communicating with a living person so that he does not even know that he is dealing with a bot. The art of diplomacy though...

Will it be possible to replace some bureaucrats with AI in the future, would it change the quality of diplomacy, and does such a game have anything to do with a real diplomatic service? Professor Arkady Rzegocki, Head of the Foreign Service of the Republic of Poland, is convinced that the answer is no. “Artificial intelligence has a very large application, and it’s getting bigger every year. The world of diplomacy, of course, takes advantage of digital achievements, such as in crisis situations. However, AI certainly won’t be able to replace professional diplomatic staff. Diplomacy is largely based on trust and personal contacts, and even the best machine will never replace it”, states the ambassador.

    SIGN UP TO OUR PAGE 
  Well, artists and radiologists still say that too, just ever more meekly. It’s clear, however, that they their specialties don’t depend on personal contacts with clients or patients.

AI eats the European pie

There’s also no denying that –Diplomacy is basically just an imitation of real diplomacy, although Henry Kissinger is a big fan of the game. This aged strategic adviser, former secretary of state and American diplomat, you would say the Talleyrand of our time, has loved it for 63 years, because that’s when it entered the American market as a strategic “war game”. Its players belong to a similarly elite group as chess players. They associate in clubs, organize local and international –Diplomacy tournaments, where the prize pool can make the heads spin of simple fans of Ludo or even Scrabble.

For decades, the game functioned in the form of letters, then e-mails, and of course on computers, and finally online. However, the latter was not particularly appreciated as recently as a decade ago by players who – like chess players – would like to face AI. It lacked appropriate speech emulation or dialog windows that would allow for a meaningful game of people against bots. There were also no bots good enough to play with.

So what did CICERO manage to achieve, since it entered the top 10% of players? –Diplomacy differs from typical war games in several key elements. There are no turns or a sequence: players secretly record their moves after a period of negotiation, and then all moves are revealed and implemented simultaneously. Combat simulation rules are strategic and abstract, not tactical and realistic, because it’s not a military simulation. Additionally, each military unit, land or sea, has the same strength. The game board is on the scale of the entire continent (Europe AD 1901 plus Turkey, a part of the Middle East and North Africa, where Russia is the only power equipped with more than three armies), and not a single military campaign. Finally, the resolution of the conflict does not contain random elements, such as rolling dice or drawing cards. It’s the social interaction and interpersonal skills that determine success in –Diplomacy.

In the key negotiation phase, players discuss tactics and strategy, form alliances, share intel, or spread disinformation. Negotiations can be made public or done in private. The seven players aren’t bound to anything they say or promise – agreements are not binding. However, trust is very important. Players must forge alliances and observe the actions of their allies to gauge their credibility. At the same time, they must convince others of their own credibility, while planning to turn on their allies when least expected. Betrayal can be just as profitable as a lasting, dependable alliance. The goal, of course, is to gobble up as much of the European pie and supply centers as possible, but you can’t do that just by fighting out in the open. It has to be almost literally negotiated and manipulated to the point of a bruised mouth.

Here is where CICERO comes into play, which combines strategic reasoning and dialogue modules, trained on data sets (player movements and negotiation transcripts) from 125,261 online –Diplomacy games. Just like being taught to play chess. In chess, AI defeated humans a long time ago, the time has come for diplomacy, although it’s an art. It requires not only strategy, but also intuition, persuasion and even deception – human skills that seemed inaccessible even to the most powerful AI. Meta computer scientist and co-author of the –Science article, Noam Brown, started the project in 2019 thinking it would take a decade to succeed. “The idea that you could have AI discussing strategy with another person, planning everything, negotiating and building trust, felt like science fiction.”

Bot “poker face”

Meanwhile, the bot learned quite quickly to choose actions based on the state of play, previous dialogue and the foreseeable actions of other players, planning several moves ahead. During training, the researchers also rewarded him for “playing human” so that its actions wouldn’t confuse other players. The dialogue module was trained to mimic human conversation in-game, in terms of its state, previous dialogue, and what the strategic planning module intended to do. CICERO, therefore, has the ability of multi-stage communication planning and to speak and play in a human way.

The agent, according to its creators, has learned to balance deception and honesty on its own. It learned how to lie, not to go too far and to not to lose trust forever, from these thousands of examples of games played by people. And how to tell the truth without being taken for a fool. However, when diplomatic experts looked at CICERO’s transcripts, they rated some 10% of the bot’s messages as not in line with the plan or state of the game. Sharper players could conclude that the bot was talking nonsense. We humans, however, tend to detect subterfuge in the nonsense that we hear, on the principle “there is certainly a method to this madness, but what?” The players probably explained the strange CICERO messages like that, especially since most of them had no idea that their opponent was a bot.

When a psychologist looks at it (I wonder if bots will get their own therapists one day), the whole thing is a manifestation of the agent’s creativity, so finding new strategies. It has to do with malleability and flexibility, which can be assessed by tests examining intelligence based on creativity, and not knowledge learned in school. And although our communication is more or less emotionally charged, and bots can’t really feel emotions (from butterflies in the stomach to lightheadedness), it’s the “poker face” that we need to train if we want to be diplomats. In other words: the tearful reaction of Ukrainian Ambassador Andriy Mylnyk to the German dictum that “there is no point in helping since you will be gone in a few hours” was deeply human, but it lay at the antipodes of diplomacy. Of course, the very words coming from the officials of the German Ministry of Foreign Affairs were not diplomatic in any way either, and they were also inhumane.

So will the lack of emotions in communication help the bot – let’s face it – manipulate human players? Or, by not having your own face and not seeing the body language of other people is it possible to see their mental state? To sense who is telling the truth and who is lying, who is sincere and who is manipulating, what we often do subliminally and call it intuition? Is the information from the dialogue and timbre of the human player’s voice enough for this – it turns out that it is. In any field – regardless of today’s fashion to be “cool” in every which way – manners and conventions usually facilitate interactions, and these can be taught to both a child and AI.

Another question is whether the bots trained in –Diplomacy will be capable of even momentary cooperation and not competition, and how will people who form alliances with machines feel? Researchers trained the strategic reasoning module by forcing the agent to play against its clones. If he had to help people negotiate prices at auctions or get deals and discounts from contacts with today’s equally bot-filled customer service departments (generally focused on “no” or a quick pay-off instead of a full refund), he will compete. However, agent-negotiators who would convince people to make healthy choices, in their diet for example, or help conflicted politicians, would have to learn cooperation, share strategies and exchange ideas openly.

Threats? AI agents can manipulate political views, commit financial fraud or mine confidential information. And then who do you imprison for espionage, embezzlement or high treason? The easiest way to end this story is that, after chess, Go, poker and Dot 2, AI started to outplay people in –Diplomacy. However, I have the impression that this is neither the essence of this achievement by Meta computer scientists, nor does it end there.

–Magdalena Kawalec-Segond

TVP WEEKLY. Editorial team and journalists

–Translated by Nicholas Siekierski
Source:https://www.science.org/doi/10.1126/science.ade9097