Beyond the Limits: Generative AI and the Dawn of Post-Scarcity Democracy Research

The legacy models of democratic governance—notably including liberal democracy—increasingly face challenges and criticism. While alternative governance models have been theorized and broadly discussed (e.g., Landemore 2020; Bächtiger et al. 2018), there remains a significant gap between theory and reform. One important cause of the gap is the traditional social science methodologies that make empirical exploration of democracy's design space slow, expensive, and ethically fraught. I will explore if the increasingly powerful and accessible tools of in silico experimentation bear promise to ease the social sciences' predicament. In particular, could the continued capability-progress of generative AI help the social scientists escape the habitual constraints? My central hypothesis is that generative AI—especially the large language models (LLMs)—enables a potential breakthrough into a post-scarcity paradigm for democracy research. LLM-driven simulations of individuals and communities offer unprecedented efficiency and scalability (Grossmann et al. 2023), enabling previously infeasible experimentation with innovative sociopolitical configurations. LLMs can be shaped into "generative agents" (Park et al. 2023)—or digital homunculi, if you like—and have proven capable of providing believable simulations of human behavior (Argyle et al. 2023). They respond to their environment, interact, share information, and coordinate, resulting in emergent patterns that resemble those of real societies. The opportunity to study democracies in silico appears just a few steps ahead. Especially so with AI's exponentially growing capabilities that enable ever more robust and fine-grained simulations of human social worlds. Of course, the post-scarcity research paradigm inevitably carries limitations and risks. An LLM should not be overly anthropomorphized. Rather than a person, it is an alien substrate capable of acting out a human role if so instructed. Therefore, the resulting simulation of human society may not remain true to "what real humans would do," especially in novel situations that we want to test for, which are not present in LLMs' training data. Further complexties arise due to the proprietary nature of the cutting edge models and their constant fine-tuning by developers, potentially hindering research reproducibility (Grossmann et al. 2023). These and other concerns will require us to develop creative methodologies for the post-scarcity democracy research. However, despite the risks, the opportunity to easily explore "what-if" scenarios not readily observable in real world and even prototype reforms appears too attractive to pass. LLM-based simulations could transform democracy research and help advance democratic practice for the future. It is our job to figure out how to make them serve this purpose. The purpose of my contribution is to lay down the groundwork.