Understanding human decision-making is a fundamental issue that has ramifications into policy making, management and even data science through microfoundations. State of the art in decision science hinges on elegant yet restrictive game theoretic models, behavioral models that often fail to replicate, or on agent-based simulations that are often criticized because of the large number of free parameters that they entail. In this dissertation, I propose that a third way forward is pursued with the usage of Large Language Models. First, I discuss the similarities between the decision making of LLMs and that of its human counterpart. I emphasize their shared reliance on anchors and cues that derive from the social framework in which strategic interaction takes place. In particular, I argue that more recent models are not necessarily the most humanlike in their judgement. Second, I propose that larger models can act as teachers for smaller models by transferring their sophisticated decision-making and theory of mind through fine tuning. Indeed, smaller models fine-tuned this way align their judgement to that of their larger pretrained "siblings" even on tasks upon which they have not been fine-tuned. Third, I offer a proof of concept of a “digital twin” with the goal of understanding how information modulation policies can affect the rate of cooperation in the repeated Prisoner’s Dilemma between LLM-empowered agents. Finally, I plan to investigate how LLMs interface with repeated social dilemmas using social learning and compare their behavior to that of human subjects. The results obtained could pave a way to the use of these algorithms in place of humans for lab and field experiments.