Abstract
Trust is a fundamental component of human social and economic interaction. For AI agents to effectively integrate into human society, they must be capable of understanding and navigating these dynamics. This paper presents a mathematical framework and reinforcement learning simulations to study the emergence of trust in AI agents using the "Trust Game." By employing Thompson Sampling, we demonstrate that trust is not merely a moral sentiment but can emerge as a computationally optimal strategy. Our findings show that agents learn to provide either zero or full trust based on whether the expected return from a partner exceeds a specific threshold, regardless of the partner's specific behavioral response function.
Introduction
Trust is often described as the bedrock of human morality and a prerequisite for economic growth. In the context of artificial intelligence, the ability to trust and be trustworthy is essential for collaborative agents. This research addresses the challenge of quantifying trust and determining how it can be learned by a machine.
The study focuses on the Trust Game (TG), a canonical task in psychology and economics. In this game, a "Trustor" decides how much of an endowment to send to a "Trustee." The amount sent is multiplied by a factor $K$. The Trustee then decides what fraction of the received amount to return to the Trustor. The risk for the Trustor lies in the potential for the Trustee to be selfish, while the reward lies in the potential for mutual gain through cooperation.
Methods
The researchers utilized a combination of formal mathematical proofs and Reinforcement Learning (RL) simulations to investigate trust dynamics.
Theoretical Framework
The authors derived a theorem to identify the optimal fraction of an endowment a Trustor should send to maximize expected rewards. This depends on the Trustee’s probability of returning funds ($p$) and the fraction they return ($\alpha$). A key proposition (Proposition 1) suggests that under power-function behaviors, the optimal strategy is binary: complete trust (100% offer) if the product of the return fraction, return probability, and multiplier ($\alpha_0 p_0 K$) is greater than 1, and zero trust otherwise.
Reinforcement Learning Setup
The Trustor was modeled as an RL agent using Thompson Sampling, an algorithm that balances exploration and exploitation by maintaining a probability distribution over the expected rewards of different actions. The agent chose from 11 discrete trust levels (0 to 1 in 0.1 increments).
The environment consisted of a modeled Trustee whose behavior followed different social response patterns:
- Constant: Returns a fixed percentage.
- Linear: Returns scale linearly with the amount received.
- Quadratic: Returns scale exponentially with the amount received.
Simulations consisted of 20,000 trials per condition to observe the learning trajectory of the agents.
Results
The RL agents consistently converged to the mathematically predicted optimal strategies. The learning process demonstrated that trust emerges as a rational response to an environment that rewards cooperation.
As illustrated in Figure 1, the agents' behavior was highly sensitive to the "market conditions" of the game. When the conditions for trust were poor (top row), the frequency of the "0% offer" (red dashed line) rose to nearly 100%. Conversely, when the environment rewarded trust (bottom row), the "100% offer" (green solid line) became the dominant strategy. This convergence occurred regardless of whether the Trustee's return logic was constant, linear, or quadratic, showcasing the robustness of the Thompson Sampling approach.
Discussion and Limitations
This work demonstrates that "Machines that Trust" can be developed using structured cognitive models and reinforcement learning. By framing trust as a rational decision-making process, we can create agents that collaborate more effectively in social and economic settings.
However, several limitations remain:
- Static Trustees: The Trustee's behavior parameters were fixed, whereas real-world interactions involve co-adaptation where both parties learn simultaneously.
- Action Granularity: The discrete action space (11 levels) simplifies the continuous nature of real-world trust.
- Task Complexity: While the Trust Game is a gold standard, it is a simplified model. Future research should explore multi-stage social dilemmas and more complex interaction environments.
Ultimately, this research provides a path toward AI that can better understand and participate in the social fabric of human society through the lens of computational rationality.
Summary updated