# Towards Human-AI Deliberation: Design and Evaluation of LLM-Empowered Deliberative AI for AI-Assisted Decision-Making

**Authors:** Shuai Ma, Qiaoyi Chen, Xinru Wang, Chengbo Zheng, Zhenhui Peng, Ming Yin, Xiaojuan Ma  
**arXiv:** [2403.16812](https://arxiv.org/abs/2403.16812)  
**Date:** 2024-03-25  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2403.16812v1/

## Key points

- Traditional approach in decision-making involves passive review of AI suggestions
- Passive paradigm hinders analytical thinking and communication of conflicting opinions
- Novel framework called [framework name] proposed to address this challenge
- Framework aims to foster human reflection and discussion on conflicting human-AI opinions
- [Framework name] engages humans and AI in dimension-level opinion elicitation, deliberative discussion, and decision updates
- Model utilizes large language models as a bridge between humans and domain-specific models for flexible conversational interactions
- [Framework name] outperformed conventional explainable AI assistants in enhancing reliance and task performance
- Participants engaged in doubts, reflections, self-corrections, and discussions with the AI during conversations
- Discussing with the AI provided new knowledge, insights, fresh perspectives, identified biases, promoted balance between objective knowledge and subjective opinions, facilitated informed decision-making
- Some limitations identified by participants: reduced agency due to active thinking prompted by the AI, perceived mental demands of deliberative discussions
- [Framework name] can enhance decision-making processes by encouraging balanced perspectives between human subjectivity and AI objectivity while highlighting potential biases within the AI model

## Summary

In the realm of , a common approach involves humans passively reviewing AI suggestions and making decisions based on whether to accept or reject them as a whole. However, this passive paradigm often hinders analytical thinking and communication of conflicting opinions when disagreements arise. To address this challenge, a novel framework called  has been proposed. This framework aims to foster human reflection and discussion on conflicting human-AI opinions in decision-making processes. By leveraging theories in human deliberation,  engages humans and AI in dimension-level opinion elicitation, deliberative discussion, and decision updates. To enable AI with deliberative capabilities,  has been designed. This model utilizes large language models (LLMs) as a bridge between humans and domain-specific models to facilitate flexible conversational interactions and provide accurate information. An exploratory evaluation conducted on a graduate admissions task demonstrated that  outperformed conventional explainable AI (XAI) assistants in enhancing appropriate reliance and task performance among users. The thematic analysis of open-ended questions revealed that participants engaged in doubts, reflections, self-corrections, and discussions with the AI during conversations. Many participants acknowledged that discussing with the AI provided new knowledge, insights, fresh perspectives, helped identify biases in decision-making processes, promoted balance between objective knowledge and subjective opinions, and facilitated informed decision-making. However, some limitations were identified by participants regarding reduced agency due to active thinking prompted by the AI and perceived mental demands of deliberative discussions. Overall, the findings suggest that  can enhance decision-making processes by encouraging , promoting balanced perspectives between human subjectivity and AI objectivity while also highlighting potential biases within the AI model. The study provides valuable insights for future design considerations of AI-assisted decision tools to optimize user experience and decision outcomes.