Introduction and Purpose: Large language models (LLMs) are increasingly integrated into human-facing domains, yet a thorough understanding of their underlying functionality, particularly during autonomous interactions with humans, remains limited. Consequently, questions regarding their psychological characteristics, especially personality traits, have become more prominent. The purpose of this paper is to synthesize existing findings on the personality profiles of LLMs using validated personality inventories.


Methodology: A systematic review was conducted in accordance with PRISMA guidelines, focusing on studies from 2019 to 2025 to capture recent developments in information and communication technology. A search of PubMed identified 36 results, with 4 meeting the criteria for quantitative analysis using validated instruments. An additional search of Xiv.org yielded 53 results, of which 3 met the inclusion criteria.


Results: Most instruction-tuned models produce stable, prosocial, and emotionally resilient personality profiles. However, results vary considerably across software versions, prompting strategies, and architectural configurations. Temporal stability is limited, and human-like biases such as social desirability are evident. Certain LLMs also display elevated levels of dark traits, such as Machiavellianism, which raises ethical concerns.


Conclusion: This analysis underscores both the diagnostic potential and the limitations of psychometric methods in evaluating artificial intelligence. Methodological heterogeneity, construct validity, and the effects of fine-tuning warrant further investigation. Implications for AI alignment, psychological safety, and the responsible design of personality-driven AI agents are discussed, including an overview of an emerging legal application for end-users. Future research should employ experimental designs to further examine the dynamics of change in LLMs and the subjective experiences of users.