Collective Decision Making With Large Language Models
Abstract
Approved for public releaseThe typical social choice setting involves a small, predetermined set of alternatives (such as candidates in an election) and a set of participants who specify their preferences regarding these alternatives, often in the form of rankings. When democratic input is sought for more nuanced and open-ended decisions, however, the process may not fit into this neat template.To address this limitation, the PI has been developing a new paradigm for the design of democratic processes: generative social choice. It fuses the rigor of social choice theory with the flexibility and power of generative AI, in particular large language models (LLMs), to facilitate collective decisions on complex issues in a principled way. Generative social choice breaks the design of democratic processes into two interacting components: first, providing theoretical guarantees assuming the LLM is an oracle that can precisely answer certain types of queries, and second, implementing and empirically validating the required queries.The PI#s recent work introduced the generative social choice paradigm and used it to design and pilot a democratic process in one particular setting: the selection of representative opinion statements. The resulting slate of statements should satisfy axioms for proportional representation such as balanced justified representation (BJR). While previous systems can only select a slate among users# statements, the new process can generate new statements, which might find new common ground between participants and allow for more representative slates. The PI and his team piloted this process with hundreds of participants in order to investigate a question in the realm of AI governance.These results set the stage for a broader investigation of generative social choice. The proposed research includes three directions that aim to expand the effectiveness and reach of this paradigm:1.Additive utilities and fully justified representation: This direction focuses on relaxing assumptions on agent utilities made by prior work and achieving stronger representation guarantees through the design of practical LLM queries and novel algorithms.2.Increasing transparency: The main shortcoming of democraticprocesses based on generative social choice is potentially limited legitimacy due to their reliance on opaque LLMs. The proposed research aims to alleviate this shortcoming through the design of interactive democratic processes that are more transparent.3.Single-winner decision making: Prior work and the foregoing research directions focus on generating a representative slate of statements. The last proposed research direction extends the agenda to settings where only a single policy can be selected.Future Naval Relevance. Last month, the 2023 DoD Data, Analytics, and AI Adoption Strategy was announced. The document describes DoD#s #strategic goals and the AI hierarchy of needs,# and pays special attention to questions of responsible AI. An emerging approach to responsible AI, which has been explored in public and in private by major industry players, is to use democratic inputs for AI alignment. In fact, the PI#s pilot of the generative social choice paradigm was done as part of an OpenAI program, with the goal of answering ethical questions about the personalization of large language models. The proposed research aims to make the approach applicable to high- stakes scenarios, which means the Department of Defense would be able to apply it to responsible AI governance in its own applications.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 09, 2024
- Source ID
- N000142412704
Entities
People
- Ariel Procaccia
Organizations
- Office of Naval Research
- President and Fellows of Harvard College
- United States Navy