Societal Alignment Frameworks Can Improve LLM Alignment

The paper presents a framework for improving large language model (LLM) alignment by integrating insights from societal alignment frameworks, which encompass social, economic, and contractual dimensions. It critiques current alignment methods for their tendency to create misspecified objectives and explores how uncertainty in societal frameworks can inform LLM alignment strategies. This approach suggests a shift in perspective, viewing the underspecified nature of LLM objectives as a potential avenue for enhancing alignment, while also advocating for participatory design in alignment interfaces to better meet human values.