In Diplomacy, a strategy game set in Europe at the turn of the 20th century, success relies not on chance but on negotiation skills. Players, each representing a superpower, invest significant time in building trust and forming alliances, only to ultimately betray opponents (or allies) to gain territory. How is it that an artificial intelligence can achieve victory over human players?
The game’s distributor, Avalon Hill, may assert that “the most skillful negotiator will achieve victory,” but researchers have sought to understand how, in 2022, an AI model participated in an online Diplomacy tournament and triumphed over humans in 40 games.
Is this software capable of mastering the intricacies of human communication, with its subtleties and unspoken cues?
Appearances can be misleading, according to researchers from the University of Southern California (USC), the University of Maryland, Princeton, and the University of Sydney.
The authors of the study aimed to uncover how CICERO, an AI model developed by Meta, the parent company of Facebook, managed to succeed.
They found that these victories stemmed more from the AI’s strategic capabilities rather than its communication skills, which the researchers believe still lag behind those of humans.
The study authors suggest that the findings could enhance AI’s ability to communicate with humans in addition to planning strategic moves; not only for board games but also for addressing everyday challenges.
“We’re exploring this because we want to model communications between AI and humans,” stated Jonathan May, a USC associate professor and co-author of the study. “At the heart of this study is a question that’s both significant and complex: How adept is AI at bluffing?”
Decoding communications
To ascertain this, the researchers organized a series of Diplomacy games with CICERO pitted against human players. Over 24 games and 200 hours of competition, they collected more than 27,000 exchanged messages. Unlike previous studies, the authors shifted their focus from CICERO’s impressive win rate to its ability to utilize bluffing and communication skills that are central to the board game.
To evaluate the AI’s bluffing abilities, the team developed a conversation analysis system utilizing a technique called abstract semantic representation (ASR), which translates complex natural language messages into data that can be processed by computers.
This ASR enabled the researchers to compare what players claimed they intended to achieve in their messages with what they actually accomplished during the game. For instance, if Germany stated to England, “I will support your invasion of Sweden next turn,” the researchers would verify whether that player indeed offered support or instead made a contradictory move.
This approach allowed the researchers to identify instances of bluffing and persuasion while also comparing CICERO’s communication skills to those of humans.
Strategy goes beyond communication
Despite winning 20 of the 24 games included in the study, research revealed that CICERO’s messages were sometimes inconsistent and did not accurately reflect the computer’s true strategic intentions.
“If you pay attention to what the machine says about the game, it’s nonsensical,” remarked Professor May. “What the AI expresses is simply stuff that players in Diplomacy have already stated. It doesn’t reflect its actions.”
The researchers also conducted experiments where CICERO’s communication abilities were restricted in various ways. In some games, the AI was unable to send messages at all, while in others, its exchanges were limited to basic strategic information. Adjusting these parameters did not significantly alter the results, suggesting that negotiation skills are largely unrelated to the computer’s proficiency at Diplomacy.
Humans, conversely, proved to be adept at deception. The study indicated that CICERO was less treacherous and persuasive than its human counterparts and was also less likely to be convinced to take specific actions.
Humans not only lied more but were also more effective at persuading one another compared to CICERO. Furthermore, when human players discovered CICERO was an AI, they began to deceive it more frequently.
“What makes CICERO excel at Diplomacy is that the AI has recorded numerous games, enabling it to play effectively,” May noted. “However, the computer struggles to be convincingly deceptive and does not genuinely respond to what other players are saying.”
Although it’s just a game, understanding the nature of AI deception in Diplomacy could pave the way for future studies on more significant forms of adversarial communication.
May mentioned that this new insight could assist in developing applications to counter AI-generated threats in realistic scenarios, such as a digital assistant helping humans identify misinformation and advising on whom to trust online.
Subscribe to our sprawling newsletter
Support us for the price of a coffee
In Diplomacy, a strategy game set in Europe at the turn of the 20th century, victory does not depend on luck, but on the ability to negotiate. Players, each representing a superpower, spend a good part of their time building trust, forming alliances, before betraying their opponents (or their allies) to seize territory. How is it that an artificial intelligence is able to triumph over flesh-and-blood players?
In 2022, an AI model participated in a tournament of Diplomacy online and triumphed over humans in 40 games. This prompted researchers from esteemed institutions like the University of Southern California (USC), the University of Maryland, Princeton, and the University of Sydney to delve deep into the mechanics of this victory.
Understanding AI in Diplomacy
The game’s distributor, Avalon Hill, claims that “the most skillful negotiator will achieve victory.” Yet, the capacity of AI models like CICERO, developed by Meta (Facebook’s parent company), has left many aghast. How does an AI manage to outsmart talented human negotiators?
Decoding Communications in Diplomacy
Researchers organized a series of games to analyze how CICERO interacted and communicated during gameplay. They engaged the AI in 24 games and examined over 27,000 messages exchanged between players. The study aimed to understand the bluffing and persuasion aspects that are critical in Diplomacy.
Bluffing Techniques in AI
The advanced communication analysis system employed in the study is based on a technique called abstract semantic representation (ASR). This system allowed the researchers to convert complex natural language messages into structured data.
Some key points about CICERO’s bluffing and communication capabilities include:
- CICERO’s messages were often inconsistent and did not truly reflect its strategic intentions.
- Humans displayed a greater ability to deceive and persuade compared to the AI.
- When humans learned CICERO was an AI, they adjusted their strategies, attempting to outwit it.
Game Dynamics: AI vs. Humans
Though CICERO triumphed in 20 out of 24 games, its ability to communicate effectively fell short. Professor Jonathan May pointed out that much of the AI’s communication was based on established patterns in previous games rather than innovative persuasion techniques.
Communication Constraints
To truly challenge CICERO’s capabilities, researchers imposed restrictions on its ability to communicate. For some games, it was unable to send any messages, while in others, only basic strategies could be conveyed. The results revealed:
- Negotiation effectiveness was not heavily impacted by communication capabilities.
- Human players demonstrated deceit more competently and adjusted their tactics significantly when interacting with the AI.
Takeaways from the Study
What makes AI like CICERO stand out? The study highlights that while it recorded an impressive win rate due to deep strategic knowledge, its understanding and execution of human-like communication were lacking. The findings imply:
| Aspect | CICERO | Humans |
|---|---|---|
| Win Rate | 20 out of 24 | N/A |
| Communication Skills | Lacking precision | Advanced and deceptive |
| Adapting Strategy | Highly adaptive | |
| Deception Level | Minimal | High |
Implications for Future AI Developments
The research sheds light on potential advancements in AI communication, particularly in crafting systems that can navigate not just the strategic but also the nuanced social dynamics of human interaction. This could extend beyond board games to real-world applications, prompting more research into designing digital assistants capable of discerning and navigating misinformation and fostering more human-like conversations.
Real-World Applications
As AI continues to evolve, the implications of this study may lead to significant benefits, including:
- Enhancing AI’s ability to understand and interact with humans in nuanced scenarios.
- Creating algorithms designed to detect misinformation online.
- Improving strategic negotiation AI for various applications, from business negotiations to conflict resolutions.
Final Thoughts on AI and Human Interaction
In examining CICERO’s approach to Diplomacy, researchers can glean crucial insights into the complexities of trust, deception, and negotiation in human behavior. The advancements in understanding AI’s capabilities could lay the foundation for future developments that enhance not just entertainment experiences but also practical, real-world applications. As we navigate these exciting developments in AI and communication, the journey to understanding and modelling communication complexities will continue to unfold.
