Generative Social Choice

TL;DR

Generative social choice marries social choice theory with large language models to address open-ended democratic questions by generating and evaluating new alternatives. The paper introduces a two-component framework: (i) guarantees under ideal queries and (ii) empirical validation of LLM-based queries, and formalizes a strong representation axiom, balanced justified representation (BJR). It proves polynomial-time BJR guarantees with perfect queries, analyzes limits under bounded-query regimes, and demonstrates practical viability through a case study on summarizing abortion opinions into a representative slate of statements, achieving high perceived representativeness in a validation survey. The work highlights both theoretical and empirical pathways for scalable, AI-assisted democratic processes, while acknowledging trust and robustness challenges and suggesting a path toward responsible deployment. Overall, it provides a principled framework and encouraging empirical results for open-ended, AI-augmented social choice.

Abstract

The mathematical study of voting, social choice theory, has traditionally only been applicable to choices among a few predetermined alternatives, but not to open-ended decisions such as collectively selecting a textual statement. We introduce generative social choice, a design methodology for open-ended democratic processes that combines the rigor of social choice theory with the capability of large language models to generate text and extrapolate preferences. Our framework divides the design of AI-augmented democratic processes into two components: first, proving that the process satisfies representation guarantees when given access to oracle queries; second, empirically validating that these queries can be approximately implemented using a large language model. We apply this framework to the problem of summarizing free-form opinions into a proportionally representative slate of opinion statements; specifically, we develop a democratic process with representation guarantees and use this process to portray the opinions of participants in a survey about abortion policy. In a trial with 100 representative US residents, we find that 84 out of 100 participants feel "excellently" or "exceptionally" represented by the slate of five statements we extracted.

Figures (8)

• Figure 1: Distribution of discriminative query predictions on the five held-out statements for each of the 100 participants. The x-axis shows the rating level selected by the participant, the y-axis the distribution of predictions. Means are represented by triangles.
• Figure 2: Predicted preference ordering for pairs of participant and held-out statement. Cell $(x, y)$ aggregates all pairs where the first participant rates the first statement at level $x$ and the second participant rates the second statement at level $y$. Colorful areas indicate in what fraction the discriminative query for the first pair gives a higher, equal, or lower prediction than the discriminative query for the second pair. The number within each cell indicates the number of relevant pairs. The lower half of the grid is symmetric to the upper half and therefore omitted.
• Figure 3: Win rate of LLM-generated statements over all four participant-written statements, in terms of the minimum predicted utility across the four participants. The dashed line indicates a win rate of 20%, which would be obtained if the generated statement followed the same distribution as the human-written statements. Win rates are computed for 1000 random groups of four participants, and ranges indicate 95% confidence intervals.
• Figure 4: Distribution of the top-20 rating of LLM-generated statements on 50 randomly subsampled sets of 80 agents, across different numbers of generation attempts. Statements were generated by applying the generation prompt to one, two, three, or four nearest-neighbor clusters of size 10, and taking the statement with highest top-20 rating.
• Figure 6: Demographic composition of both samples, compared to the US population as of the 2020 Census USCensus2020 and the 2020 U.S. Presidential Election results ElectionResults2020.

Theorems & Definitions (21)

• Definition 1
• Theorem 2
• Proposition 2
• Proposition 2
• Theorem 3
• Theorem 4
• Proposition 5
• proof
• Definition 6
• Proposition 7