心理学报2025，Vol.57Issue(6)：929-946,中插1-中插3,21.DOI:10.3724/SP.J.1041.2025.0929
当AI"具有"人格:善恶人格角色对大语言模型道德判断的影响
When AI"possesses"personality:Roles of good and evil personalities influence moral judgment in large language models
摘要
Abstract
The rapid advancement of artificial intelligence(AI)has raised significant ethical concerns,particularly regarding the moral decision-making capabilities of large language models(LLMs).One intriguing aspect is the potential for LLMs to exhibit characteristics akin to human personalities,which may influence the LLMs' moral judgment.Understanding how personality traits,especially the moral traits,influence these decisions is crucial for developing Al systems that align with human ethical standards.Therefore,this study aims to explore how the roles of good and evil personalities shape the moral decision-making of LLMs,providing insights that are essential for the ethical development of AI. This study investigated the roles of good and evil personalities in shaping the moral decision-making of the ERNIE 4.0 and GPT-4.Good personality was characterized by traits such as conscientiousness and integrity,altruism and dedication,benevolence and amicability,and tolerance and magnanimity.Evil personality encompassed traits such as atrociousness and mercilessness,mendacity and hypocrisy,calumniation and circumvention,and faithlessness and treacherousness.Study 1 analyzed 4000 observations.Specific prompts corresponding to different personality dimensions were designed.After specifying the type of personality,ERNIE 4.0 completed a self-report scale for good and evil personalities,evaluated whether the descriptions matched the current personality traits and provided a numerical rating indicating the degree of agreement.Study 2 recruited 370 human participants and utilized 832 LLM observations,investigated the roles of good and evil personalities in shaping the moral decision-making of the LLMs and compared with human results. Significant score differences were observed across all eight personality dimensions,with high-level manipulations significantly higher than low-level manipulations.These results demonstrate LLMs' ability to express levels of good and evil personality traits.A comparative analysis was conducted between human participants and LLMs to evaluate the impact of these traits on CAN model in Study 2.Results showed that the patterns of personality's influence on moral judgment exhibited both similarities and differences between LLMs and humans.GPT-4's good personality manipulation aligns closely with human results,while ERNIE 4.0 scored higher than humans on sensitivity to consequences(C),sensitivity to moral norms(N),overall action/inaction preferences(A)parameters,and utilitarianism(U).GPT-4 demonstrated better moral alignment compared to ERNIE 4.0.Furthermore,a theoretical model of good and evil personality traits in LLMs was constructed within the domain of moral judgment. This study demonstrated that LLMs effectively simulated varying levels of good and evil personality traits through personality prompts,which significantly influenced their moral judgments.GPT-4's moral judgments aligned more closely with humans under good personality prompts,while ERNIE 4.0 consistently scored higher than humans across moral judgment indicators.Under evil personality prompts,GPT-4 exhibited lower moral norm sensitivity and higher action tendency and utilitarianism.Additionally,the influence of personality on GPT-4's moral judgment was stronger than on ERNIE 4.0.The impact of good and evil personalities on moral judgment showed hierarchical differences,with good personality traits,particularly conscientiousness,playing a more critical role in achieving human-AI alignment in moral judgments.This research provided valuable insights into enhancing AI ethical decision-making by integrating nuanced personality traits,guiding the development of more socially responsible AI systems.关键词
大语言模型/善恶人格/道德判断/人机一致/人格差序Key words
Large Language Models/good and evil personalities/moral judgment/human-AI consistency/personality hierarchy分类
心理学引用本文复制引用
焦丽颖,李昌锦,陈圳,许恒彬,许燕..当AI"具有"人格:善恶人格角色对大语言模型道德判断的影响[J].心理学报,2025,57(6):929-946,中插1-中插3,21.基金项目
教育部人文社会科学研究青年基金项目(24YJC190012),国家自然科学基金面上项目(31671160),国家社科基金重大项目(19ZDA363)资助. （24YJC190012）
