Is persona enough for personality? Using ChatGPT to reconstruct an agent's latent personality from simple descriptions

Abstract: Personality, a fundamental aspect of human cognition, contains a range of traits that influence behaviors, thoughts, and emotions. This paper explores the capabilities of LLMs in reconstructing these complex cognitive attributes based only on simple descriptions containing socio-demographic and personality type information. Utilizing the HEXACO personality framework, our study examines the consistency of LLMs in recovering and predicting underlying (latent) personality dimensions from simple descriptions. Our experiments reveal a significant degree of consistency in personality reconstruction, although some inconsistencies and biases, such as a tendency to default to positive traits in the absence of explicit information, are also observed. Additionally, socio-demographic factors like age and number of children were found to influence the reconstructed personality dimensions. These findings have implications for building sophisticated agent-based simulacra using LLMs and highlight the need for further research on robust personality generation in LLMs.

• The paper demonstrates that GPT-3.5 and GPT-4 can reconstruct HEXACO personality traits with approximately 72% consistency from basic persona descriptions.
• The methodology integrates socio-demographic factors with personality dimensions to generate diverse persona narratives for simulation testing.
• The study reveals that LLMs tend to overestimate positive traits, highlighting biases linked to missing or demographic data.

"Is Persona Enough for Personality? Using ChatGPT to Reconstruct an Agent's Latent Personality from Simple Descriptions" (2406.12216)

Introduction

This paper investigates the ability of LLMs, specifically GPT-3.5 and GPT-4, to reconstruct complex human personality traits from basic persona descriptions. Utilizing the HEXACO personality framework, the study examines the extent to which LLMs can replicate underlying personality dimensions using socio-demographic and personality type information. This line of research broadens our understanding of the emergent cognitive capabilities of LLMs, particularly in the context of agent-based simulations and personality reconstruction.

Persona Construction and Experimental Setup

The study constructs persona descriptions by integrating socio-demographic factors (age, gender, marital status, annual household income, number of children) with personality types using the HEXACO framework. For each personality dimension, descriptive sentences are generated, and personas are randomly composed by selecting and combining these dimensions. By applying a consistent methodology involving GPT-3.5-Turbo and GPT-4-Turbo through the OpenAI API, the study establishes a reproducible experimental setup to evaluate personality reconstruction.

Results and Analysis

Out of 1000 personas tested, GPT-3.5-Turbo showed a significant degree of consistency (71.88%) in reproducing high or low expected scores across HEXACO dimensions. However, discrepancies were common where LLMs assigned unintended high scores, indicating a bias towards positive personality traits when faced with missing information. Additionally, omitted personality dimensions were also more likely to be attributed high scores by LLMs.

Further analysis through ANOVA revealed that socio-demographic aspects such as age and number of children significantly influenced reconstructed personality dimensions, although not uniformly across all dimensions. The observed pattern underscores the tendency of LLMs to overestimate the influence of such factors, deviating from empirically-established correlations in personality psychology.

Implications and Future Work

The findings highlight LLMs' potential in reconstructing human-like personas, which has promising applications in crafting realistic simulated agents. Nevertheless, the inclination of these models towards biased output underscores the need for more robust calibration methods. Future research should aim to address such biases and enhance the fidelity of personality reconstruction, contributing to agent simulations that accurately mirror human cognitive complexity.

Conclusion

This study advances the exploration of LLMs' capabilities in understanding and replicating human personalities using simple persona descriptions. It emphasizes the critical role of comprehensive data representation in guiding LLM predictions, while acknowledging the models' existing biases and limitations. Understanding these dynamics is vital for developing sophisticated agent-based simulations employing LLM technology.