Niclas Boehmer, Sara Fish and Ariel D. Procaccia, Generative Social Choice: The Next Generation
On many online platforms, users share their opinions, ranging from product reviews on Amazon to public policy discussions on Pol.is. But when hundreds or thousands of people weigh in, it becomes impossible for any one person to read everything.
We introduce the Proportional Slate Engine (PROSE), a system that creates a slate of statements from large, messy collections of opinions. Instead of repeating popular comments, PROSE uses a language model (like ChatGPT) to write and assess new statements that synthesize the beliefs of groups of users. Inspired by ideas from voting theory, PROSE aims for proportional fairness: each viewpoint receives space in the summary proportional to its popularity.
A key challenge is ensuring both clarity and fairness. To keep things simple, PROSE respects a word limit, so the summary doesn't become too long. To ensure fairness, we provide mathematical guarantees: even if the language model makes small mistakes, the summary still proportionally reflects what people believe.
We tested PROSE on real-world data like drug reviews and city improvement debates, and found it to be a powerful tool for summarizing public input in a readable and fair way.