Generative Social Choice

Introduces a design methodology for open-ended democratic processes combining social choice theory with LLM capabilities to generate text and extrapolate preferences, enabling collective selection of textual statements unlike traditional voting limited to predetermined alternatives. The approach divides AI-augmented democratic process design into proving representation guarantees via oracle queries and empirically validating their implementation via LLMs. Applied to summarizing free-form opinions into proportionally representative slates—in a trial on abortion policy, 84 of 100 participants felt “excellently” or “exceptionally” represented.