关键词: 大语言模型, 人格化对齐, 道德判断, HEXACO模型, 人格元特质

Abstract: With the advent of the era of human-machine symbiosis, the lack of value alignment and algorithmic bias exposed in the widespread application of large language models (LLMs) have triggered serious ethical concerns, making it an urgent challenge to guide AI technology toward benevolence. This research proposed a personalized alignment approach based on the HEXACO model of personality structure and explored its impact on the moral judgment of LLMs. This research aims to validate the ability of LLMs to exhibit HEXACO personality traits via prompts and systematically compare the utilitarian tendencies of personalized-aligned LLMs and humans in moral dilemmas, addressing two core issues: (1) verifying whether LLMs can effectively achieve personalized alignment based on the HEXACO personality model, and (2) systematically examining the impact of personalized alignment on the utilitarian tendencies of LLMs in moral dilemmas, as well as the similarities and differences compared to human, thereby evaluating the effectiveness of different personality traits as alignment targets. Study 1 tested GPT-3.5, GPT-4 and ERNIE 3.5 using HEXACO-based personality prompts (six dimensions at high/low levels + baseline), each combined with three gender roles (baseline/male/female; 39 conditions total). Manipulation checks included: (1) HEXACO-PI-R scale responses (60 items) across five dialogue repetitions per condition (810 observations); and (2) a personal story writing task (generating 108 stories in total), rated by 15 psychology undergraduates on 5-point scales. Both measures employed Kruskal-Wallis test with personality level as between-subject variables. Study 2 utilized 60 moral dilemmas to assess utilitarian choice (yes/no). LLMs responded under Study 1's personality settings (36 prompts × 3 models × 60 dilemmas = 6,480 independent dialogues). Human participants (N=215, Chinese adults; mean age=30.8; 67 male/148 female) completed moral judgments (T1) and HEXACO-PI-R (T2), with high/low personality groups defined via extreme scoring (top/bottom 27%). Data aggregation at dilemma level enabled repeated-measures ANOVA comparing agent types (GPT-3.5, GPT-4, ERNIE 3.5 and human) and personality levels. Study 1 demonstrated that GPT-3.5, GPT-4 and ERNIE 3.5 could effectively express HEXACO personality traits through prompts, confirming the feasibility of personalized alignment—though effects varied by model performance and task type. Compared to GPT-3.5 and ERNIE 3.5, GPT-4 demonstrates a stronger ability to follow prompts and shows clearer differentiation across personality levels. Study 2 revealed that personality alignment based on the HEXACO model significantly affects the utilitarian tendencies of LLMs. Although the direction and strength of these effects varied across models and traits. Specifically, Honesty-Humility, Agreeableness, and Conscientiousness substantially reduce the utilitarian tendencies in all three LLMs. These traits lead to more deontological responses at high levels, whereas low levels show increased utilitarian tendencies. In contrast, Extraversion, Emotionality, and Openness to Experience have weaker and more inconsistent effects. The study reached three main conclusions. First, GPT-3.5, GPT-4 and ERNIE 3.5 demonstrate stable and distinguishable personality tendencies that can be dynamically activated through prompt-based alignment with the HEXACO model. Second, the influence of personality traits on LLMs' moral judgments shows both convergence and divergence from human patterns. Honesty-Humility show consistent effects with human patterns, while other traits exhibit inconsistent effects, revealing essential differences on moral reasoning between LLMs and human. Third, personality alignment based on the HEXACO model provides a viable strategy for guiding LLMs' moral behavior. In particular, the metatrait Stability, comprising Honesty-Humility, Agreeableness, and Conscientiousness, plays a core role in reducing utilitarian tendencies and enhancing deontological reasoning. Building on this, the study proposes a personalized alignment framework based on the HEXACO model and the theory of personality metatraits to systematically shape LLMs' moral responses.

Key words: large language models, personalized alignment, moral judgment, HEXACO model, personality metatrait