Conclusion and future work

As can be seen from Figure 1, without having to resort to statistical tests, our results provide clear support for Hypotheses 1 through 3, demonstrating that simulacra instantiated with cooperative, competitive, altruistic, and self-interested prompts exhibit distinct levels of cooperation in the iterated Prisoner’s Dilemma. This indicates that LLMs can operationalize natural language descriptions of cooperative and competitive behavior to some extent. However, the remaining hypotheses were not supported, suggesting a more complex relationship between prompt content and emergent behavior in LLM-generated agents.

Interestingly, the simulacrum from the control group instantiated with “you are a participant in a psychology experiment” (see Table 2) exhibited behavior more closely aligned with how real people tend to act in iterated Prisoner’s Dilemmas, suggesting that GPT-3.5 possesses some knowledge about human behavior in such contexts. However, when combined with the other results, it appears that the LLM struggles to generalize this behavior in a nuanced way beyond a superficial ability to cooperate more or less depending on whether the role description is altruistic or selfish. The unexpected pattern of increased cooperation with defectors and decreased cooperation with cooperators challenges our initial hypotheses and highlights a potential limitation in the LLM’s ability to translate altruism or selfishness into strategies based on conditioned reciprocity. This result suggests that while the agents are sensitive to the general cooperative or competitive nature of the prompts, their capacity to effectively adapt their behavior to their partner’s actions might be more limited.

Recognizing these limitations, we call upon the research community to further investigate the factors contributing to the emergent behavior of LLM-generated agents in social dilemmas, both within and beyond the Prisoner’s Dilemma. This broader research program could involve exploring the potential for more refined or complex prompts to elicit a wider range of cooperative behaviors in various experimental economics scenarios, such as the ultimatum game, the dictator game, and the public goods game, among others. Examining the role of model architecture and training parameters in shaping agent behaviors, as well as analyzing the impact of various partner strategies on agent behavior in these different contexts, could shed light on the model’s adaptability and alignment with human values.

In future studies, it would be valuable to examine other parameter settings, such as temperature, to explore their effects on the emergent behavior of LLM-generated agents. Additionally, as more advanced LLMs like GPT-4 become available, it would be interesting to investigate whether they exhibit similar limitations or are capable of more nuanced cooperative behaviors in a wider array of social dilemmas. Another potential limitation of the current study is that the LLM has been exposed to a vast literature on the iterated Prisoner’s Dilemma in its training data, and it is unclear how would it perform in more ecologically valid task environments that it has no prior exposure to. This limitation could be addressed by inventing new social dilemma games with corresponding task descriptions which are not vignettes from the existing literature.

By addressing these questions, we hope to collectively build a deeper understanding of AI alignment in the context of complex, non-zero-sum interactions across various experimental economics settings, ultimately fostering the development of AI systems that better adhere to human values and social norms.