# Opportunities and Risks of LLMs for Scalable Deliberation with Polis

**Authors:** Christopher T. Small, Ivan Vendrov, Esin Durmus, Hadjar Homaei, Elizabeth Barry, Julien Cornebise, Ted Suzman, Deep Ganguli, Colin Megill  
**arXiv:** [2306.11932](https://arxiv.org/abs/2306.11932)  
**Date:** 2023-06-20  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.11932v1/

## Key points

- Large Language Models (LLMs) can facilitate, moderate, and summarize outcomes in Polis engagements
- LLMs augment human intelligence and improve efficiency of Polis conversations
- Summarization capabilities enabled by LLMs empower the public in collective meaning-making exercises
- Risks associated with using LLMs need to be addressed
- Principles and techniques proposed for characterizing and mitigating risks
- Implications of employing LLMs in other deliberative or political systems explored
- Open research directions outlined for further enhancing tools like Polis with LLMs

## Summary

In the paper titled "Opportunities and Risks of LLMs for Scalable Deliberation with Polis," authors Christopher T. Small, Ivan Vendrov, Esin Durmus, Hadjar Homaei, Elizabeth Barry, Julien Cornebise, Ted Suzman, Deep Ganguli, and Colin Megill discuss the potential benefits and risks associated with using Large Language Models (LLMs) in the context of Polis, a platform that utilizes machine intelligence to enhance deliberative processes. The authors begin by highlighting the opportunities presented by LLMs in facilitating, moderating, and summarizing the outcomes of Polis engagements. Through pilot experiments using Anthropic's Claude LLM, they demonstrate how these models can augment human intelligence and improve the efficiency of Polis conversations. The authors specifically emphasize how summarization capabilities enabled by LLMs offer new methods that empower the public in collective meaning-making exercises. However, alongside these opportunities come risks that need to be addressed. The paper discusses some of these risks and proposes principles and techniques for characterizing and mitigating them. Additionally, it explores the implications of employing LLMs in other deliberative or political systems. The authors conclude by outlining several open research directions for further enhancing tools like Polis with LLMs. These future research areas aim to address limitations related to LLM context and explore ways to maximize insights and quality in utilizing LLMs within deliberative processes. Overall, this paper provides a comprehensive examination of both the opportunities and risks associated with leveraging LLMs for scalable deliberation with platforms like Polis. It offers valuable insights into how machine intelligence can be effectively utilized while also addressing potential challenges that may arise from their implementation.