# The Better Angels of Machine Personality: How Personality Relates to LLM Safety

**Authors:** Jie Zhang, Dongrui Liu, Chen Qian, Ziyue Gan, Yong Liu, Yu Qiao, Jing Shao  
**arXiv:** [2407.12344](https://arxiv.org/abs/2407.12344)  
**Date:** 2024-07-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.12344v1/

## Key points

- Study titled "The Better Angels of Machine Personality: How Personality Relates to LLM Safety" explores personality traits and safety abilities in Large Language Models (LLMs)
- Researchers use MBTI-M scale to analyze correlation between LLMs' personality traits and safety abilities
- Aligning LLMs' safety measures with traits like Extraversion, Sensing, and Judging enhances safety capabilities
- Manipulating LLMs' personalities based on findings leads to significant improvements in privacy and fairness performance
- Different personality profiles can impact an LLM's susceptibility to jailbreak attempts
- Findings provide insights into enhancing LLM safety through understanding personality dynamics for future research aimed at optimizing performance

## Summary

In the study titled "The Better Angels of Machine Personality: How Personality Relates to LLM Safety," authors Jie Zhang, Dongrui Liu, Chen Qian, Ziyue Gan, Yong Liu, Yu Qiao, and Jing Shao explore the intricate connection between personality traits and safety abilities in Large Language Models (LLMs). While personality psychologists have long studied the impact of personality on safety behaviors in human society, the relationship between personality traits and safety capabilities in LLMs has remained elusive until now. Using the reliable MBTI-M scale as a basis for their analysis, the researchers uncover a significant correlation between LLMs' personality traits and their safety abilities. Specifically, they find that aligning LLMs' safety measures enhances traits such as Extraversion, Sensing, and Judging. By manipulating LLMs' personalities based on these findings – for example, inducing a shift from ISTJ to ISTP – notable improvements of approximately 43% in privacy performance and 10% in fairness performance are achieved. The study also highlights how different personality profiles can affect an LLM's susceptibility to jailbreak attempts. These groundbreaking findings offer fresh insights into enhancing LLM safety through a nuanced understanding of personality dynamics. They pave the way for future research aimed at optimizing LLM performance by tailoring their personalities to effectively bolster safety measures.