# PersonaLLM: Investigating the Ability of GPT-3.5 to Express Personality Traits and Gender Differences

**Authors:** Hang Jiang, Xiajie Zhang, Xubo Cao, Jad Kabbara, Deb Roy  
**arXiv:** [2305.02547](https://arxiv.org/abs/2305.02547)  
**Date:** 2023-05-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.02547v1/

## Key points

- Study aimed to investigate if personalized Large Language Models (LLMs) can accurately reflect personality traits consistently
- Used GPT-3.5 to create LLM personas assigned with Big Five personality types and gender roles
- Created 320 LLM personas for each of the 32 Big Five personality types, with 5 females and 5 males in each group
- LLM personas' self-reported Big Five Inventory (BFI) scores aligned consistently with their assigned personality types
- Significant correlations found between assigned personality types and Linguistic Inquiry and Word Count (LIWC) psycholinguistic features in writings
- Differences observed in use of technological and cultural words between female and male LLM-generated personas
- Research serves as foundational step towards exploring personalized LLMs in Human-AI conversations

## Summary

The study "PersonaLLM: Investigating the Ability of GPT-3.5 to Express Personality Traits and Gender Differences" conducted by Hang Jiang, Xiajie Zhang, Xubo Cao, Jad Kabbara, and Deb Roy from MIT Media Lab and Stanford Graduate School aimed to address the gap in research regarding whether personalized Large Language Models (LLMs) can accurately reflect certain personality traits consistently. The researchers focused on studying LLM-based simulated agents known as LLM personas and conducted a case study using GPT-3.5 (text-davinci-003) to explore if LLMs could generate content with consistent personalized traits when assigned Big Five personality types and gender roles. To conduct the study, 320 LLM personas were created for each of the 32 Big Five personality types, with 5 females and 5 males represented in each group. These personas were prompted to complete the classic 44-item Big Five Inventory (BFI) and then write an 800-word story about their childhood. The results showed that the self-reported BFI scores of the LLM personas aligned consistently with their assigned personality types, demonstrating large effect sizes across all five traits. Furthermore, significant correlations were found between the assigned personality types of the LLM personas and some Linguistic Inquiry and Word Count (LIWC) psycholinguistic features present in their writings. For example, extroversion was associated with pro-social and active words while neuroticism correlated with words related to negative emotions and mental health. Additionally, notable differences were observed in the use of technological and cultural words between female and male LLM-generated personas. This research serves as a foundational step towards further exploration of personalized LLMs and their potential applications in Human-AI conversations. The findings contribute valuable insights into how LLMs can effectively express personality traits while also shedding light on gender differences in generated content. This study was accepted for presentation at the 9th International Conference on Computational Social Science (IC2S2), held from July 17-20, 2023 in Copenhagen, Denmark.