Session 7. 7.

"Over the course of three years, Large Language Models (LLMs) have become amazingly ubiquitous in our lives; ChatGPT helps us write our emails, set up our shopping lists for the grocery store, and even tell us a bit about Taylor Swift’s newest album. As these tools become more integrated into our lives, we are also facing a rapid growth in their power. Sam Altman, one of the most prominent creators in this space, projects that LLMs will be able to reason as well as, if not better than, humans in a “few thousand days”, so called Artificial General Intelligence (AGI). As we hurtle towards a profoundly new era of technological development, where machines can think as well as humans, this paper considers a particular set of institutions built into a prominent LLM, Anthropic’s Claude. What is the nesting of institutions in so-called Constitutional AI and what are some of the implications for how knowledge develops over the coming near term?

This paper uses the concept of nested institutions as found in the Institutional Analysis and Development (IAD) Framework: constitutional-, collective choice-, and operational-level rules and applies it to a specific process created to refine the development of Claude. Constitutional AI (Bai et al., 2022) uses a predetermined ‘constitution’ to guide Claude (a deontic) in how it grows its knowledge, a process that ultimately limits the role of humans after the initial creation of the ‘constitution’. The paper follows the process of constitution development and its implications for collective-choice and operational rules. The paper concludes with thoughts about future iterations of AGI-capable LLMs where humans are not part of the institution creation process.

Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., ... Kaplan, J. (2022). Constitutional AI: Harmlessness from AI feedback. arXiv. https://arxiv.org/abs/2212.08073"

Successful management of natural resources is crucial, especially when these resources are shared among multiple stakeholders. However, understanding which key factors favors the sustainable management of resources is difficult. While the literature abounds of individual cases, synthesis of these empirical studies is often hampered by cost and time. Here we leverage current advances in natural language processing and large language models in order to analyze the relationship between Ostrom Institutional Design principles with conflict, inequality and the state of natural resources in over 2000 published articles. By assessing the importance of Ostrom institutional design principles (DPs), we propose blueprints for success for different contexts. Our results contribute to the literature in two important ways: 1) furthering the synthesis related to common pool resources facilitating the understanding of which institutional arrangements increase the likelihood of success in different contexts; 2) methodologically, increase our ability to synthesize the literature in order to provide actionable information to policy makers and practitioners.

Collective decision-making in rural land management crucially demands equitable and inclusive participation, yet the stakeholders’ diversity often impedes true representation and active involvement. Advances in large language models (LLMs) have enabled more authentic simulations of stakeholder roles and behaviors, enhancing the transparency and accuracy of collective decision-making processes in rural settings. This study utilizes an agent-based modeling approach to investigate collective decision-making in the marketization of collectively owned rural land in China. By integrating insights from interviews with diverse local stakeholders—local farmers, corporate managers, and village officials—the model reflects the complexities of rural communities where stakeholder diversity often leads to conflicting interests and priorities. The simulation result, obtained through three voting rounds achieving a consensus with a two-thirds majority, demonstrates how large language model-driven simulations can effectively address these conflicts by ensuring diverse perspectives are adequately represented and considered. Additionally, the analysis highlights a notable shift in focus on keywords such as 'trust', 'transparency', and 'sustainability' across the voting rounds, illustrating how these evolving priorities play a pivotal role in achieving consensus by the final round. This approach illustrates the potential of agent-based models in facilitating more informed and equitable decision-making processes, thereby offering valuable perspectives for policy development and community planning in similar socio-economic settings.