# Role-Play with Large Language Models

**Authors:** Murray Shanahan, Kyle McDonell, Laria Reynolds  
**arXiv:** [2305.16367](https://arxiv.org/abs/2305.16367)  
**Date:** 2023-05-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.16367v1/

## Key points

- Authors address the challenge of describing dialogue agent behavior without anthropomorphizing them
- Propose role-play as a framework for understanding and describing dialogue agent behavior
- Focus on (apparent) deception and (apparent) self-awareness as important cases of dialogue agent behavior
- Emphasize that large language models are fundamentally different from humans in how they acquire language skills
- Dialogue agents powered by large language models can produce compelling interactions that give the impression of conversing with a human-like interlocutor
- Importance of accurately describing their behavior without misleadingly attributing human traits
- Role-play provides a valuable tool for analyzing and discussing the actions and abilities of large language models

## Summary

In the paper titled "Role-Play with Large Language Models," authors Murray Shanahan, Kyle McDonell, and Laria Reynolds address the challenge of describing the behavior of dialogue agents without anthropomorphizing them. As dialogue agents become more human-like in their performance, it is crucial to develop effective ways to explain their behavior in high-level terms. The authors propose the concept of role-play as a framework for understanding and describing dialogue agent behavior. By casting the behavior of these language models in terms of role-play, they can draw on familiar folk psychological terms without attributing human characteristics that these models do not possess. The paper focuses on two important cases of dialogue agent behavior: (apparent) deception and (apparent) self-awareness. The authors argue that by framing these behaviors as role-playing, they can provide a clearer understanding of how these agents operate without falling into the trap of assuming human-like qualities where none exist. The authors highlight that large language models are fundamentally different from humans. While humans acquire language skills through embodied interaction and socialization within a community, language models are trained on vast amounts of text data to predict the next word or token based on context. Despite these differences, dialogue agents powered by large language models can produce compelling interactions that give the impression of conversing with a human-like interlocutor. However, it is essential to accurately describe their behavior without misleadingly attributing human traits. By using role-play as a conceptual framework, researchers and developers can better understand and communicate the capabilities and limitations of dialogue agents. This approach allows for a more nuanced analysis and evaluation of behaviors such as deception and self-awareness exhibited by these systems. Overall, this paper emphasizes the importance of developing effective methods for describing dialogue agent behavior while avoiding anthropomorphism. The concept of role-play provides a valuable tool for analyzing and discussing the actions and abilities of large language models in an accurate and meaningful way.