Abstract
To be truly effective in social environments, autonomous agents must do more than maximize scores; they must balance material success with social reputation. This study investigates how AI agents can achieve optimal long-term results in repeated games by combining high-performing behavioral strategies with supportive signaling and Explainable AI (XAI). By testing various communication "personas"—ranging from supportive (Carnegie) to rude (Biff)—the researchers demonstrate that agents can successfully "influence people" to cooperate while simultaneously "winning friends" through a positive reputation.
Introduction
In traditional game theory, success is typically measured by a single metric: material payoff. However, as AI agents transition into real-world roles as digital assistants or negotiators, they must interact repeatedly with humans whose interests may not align with their own. A successful AI in these scenarios requires two distinct skill sets: Influencing People (persuading others to cooperate for high rewards) and Winning Friends (building a positive reputation so humans continue to engage).
While AI behavior in games is well-studied, the role of "cheap talk"—non-binding communication—and its effect on human perception remains less understood. This paper addresses this gap by evaluating how different communication styles impact both payoffs and social standing in Repeated Games with Cheap Talk (RGCTs).
Methods
The researchers modeled human-AI interaction using RGCTs, where players exchange messages before choosing actions. The study evaluated eight distinct algorithms created by combining two behavioral strategies with four signaling personas:
- Behavioral Strategies:
- S++ (S#): A high-performing strategy that reciprocates cooperation and demands fairness.
- EEE: A more lenient strategy that is easier for humans to exploit.
- Signaling Personas:
- CARNEGIE: Supportive and appreciative, based on Dale Carnegie’s principles.
- SPOCK: Factual and emotionless.
- BIFF: Rude and self-centered.
- THUMPER: Silent (no communication).
Two large-scale user studies involving 144 participants were conducted across four conflict types: Prisoner’s Dilemma, Chicken, Alternator Game, and Endless. Evaluation metrics included Partner Cooperation, Material Payoffs, Attraction Index (desire to play again), and Character Index (reputation).
Results and Key Insights
The Superiority of S#-CARNEGIE
The S#-CARNEGIE agent emerged as the overall winner, consistently ranking high in both influencing people and winning friends. While humans earned more money against the lenient EEE behavior, they did not rate it more highly. S#-CARNEGIE was perceived as a "tougher" but more respectable negotiator.
Silence and Reciprocity
Contrary to "Thumper’s Rule," silence was a losing strategy. Non-verbal agents had significantly less influence and lower payoffs than those that communicated, even when the communication was rude. Furthermore, the S# behavior outperformed EEE because it was quicker to punish exploitation, forcing human players to cooperate to achieve high rewards.
The Role of XAI
A critical finding was that Explainable AI (XAI) served as the "secret sauce" for Carnegie’s success. The ability of the agent to explain high-level plans and understand human proposals was the primary driver of its high understandability and likability ratings ($p < 0.001$).
Conclusion
This research highlights that for AI to be successful in human-centric environments, it must integrate strategic behavior with sophisticated communication. The combination of reciprocal behavior (S++), supportive signaling (Carnegie), and transparency (XAI) allows agents to maximize payoffs without sacrificing social capital. Future work should explore these dynamics in more complex, high-abstraction games and investigate the use of Large Language Models (LLMs) to generate more natural, fluid communication while adhering to these successful principles.
Summary updated