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

doi:10.3724/SP.J.1041.2025.0929

摘要/Abstract

摘要：

在科技与道德的交汇点, 大语言模型是否具有扮演善恶人格的能力, 以及这一能力是否会影响其在道德判断任务中的表现至关重要。研究聚焦大语言模型在模拟不同善恶人格时的道德判断特征及其与人类模式的异同。通过2个研究, 对ERNIE 4.0和GPT-4大语言模型观测值(N = 4832)及人类被试数据(N = 370)分析发现：(1)大语言模型能成功模拟不同水平的善恶人格; (2)善恶人格设定显著影响大语言模型的道德判断结果; (3)善恶人格在人机一致中展现差序性：善人格发挥着更重要的作用(善恶人格间差序), 且其中尽责诚信的影响力最大(善恶人格内差序)。研究建构了道德判断下大语言模型善恶人格的理论模型, 有助于理解大语言模型人格如何在道德判断中发挥作用, 为推动人工智能系统的道德对齐提供了理论基础和支持。

关键词: 大语言模型, 善恶人格, 道德判断, 人机一致, 人格差序

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 AI 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.

Key words: Large Language Models, good and evil personalities, moral judgment, human-AI consistency, personality hierarchy

中图分类号:

B848

焦丽颖, 李昌锦, 陈圳, 许恒彬, 许燕. (2025). 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. 心理学报, 57(6), 929-946.

JIAO Liying, LI Chang-Jin, CHEN Zhen, XU Hengbin, XU Yan. (2025). When AI “possesses” personality: Roles of good and evil personalities influence moral judgment in large language models. Acta Psychologica Sinica, 57(6), 929-946.

图/表 13

参考文献 71

[1]	Agarwal U., Tanmay K., Khandelwal A., & Choudhury M. (2024). Ethical reasoning and moral value alignment of LLMs depend on the language we prompt them in. arXiv. https://doi.org/10.48550/arXiv.2404.18460
[2]	Andrejević M., Smillie L. D., Feuerriegel D., Turner W. F., Laham S. M., & Bode S. (2022). How do basic personality traits map onto moral judgments of fairness-related actions? Social Psychological and Personality Science, 13(3), 710-721.
[3]	Ashton M. C., & Lee K. (2005). Honesty-humility, the big five, and the five-factor model. Journal of Personality, 73(5), 1321-1354. doi: 10.1111/j.1467-6494.2005.00351.x pmid: 16138875
[4]	Bartels D. M., & Pizarro D. A. (2011). The mismeasure of morals: Antisocial personality traits predict utilitarian responses to moral dilemmas. Cognition, 121(1), 154-161. doi: 10.1016/j.cognition.2011.05.010 pmid: 21757191
[5]	Baumert A., Halmburger A., & Schmitt M. (2013). Interventions against norm violations: Dispositional determinants of self-reported and real moral courage. Personality and Social Psychology Bulletin, 39(8), 1053-1068. doi: 10.1177/0146167213490032 pmid: 23761924
[6]	Binz M., & Schulz E. (2023). Using cognitive psychology to understand GPT-3. Proceedings of the National Academy of Sciences of the United States of America, 120(6), e2218523120.
[7]	Bonnefon J. F., Rahwan I., & Shariff A. (2024). The moral psychology of Artificial Intelligence. Annual Review of Psychology, 75, 653-675.
[8]	Bonnefon J. F., Shariff A., & Rahwan I. (2016). The social dilemma of autonomous vehicles. Science, 352(6293), 1573-1576.
[9]	Cohen D. J., & Ahn M. (2016). A subjective utilitarian theory of moral judgment. Journal of Experimental Psychology: General, 145(10), 1359-1381.
[10]	Demszky D., Yang D., Yeager D. S., Bryan C. J., Clapper M., Chandhok S., … Pennebaker J. W. (2023). Using large language models in psychology. Nature Reviews Psychology, 2(11), 688-701.
[11]	Dillion D., Tandon N., Gu Y., & Gray K. (2023). Can AI language models replace human participants? Trends in Cognitive Sciences, 27(7), 597-600. doi: 10.1016/j.tics.2023.04.008 pmid: 37173156
[12]	Frank M. C. (2023). Openly accessible LLMs can help us to understand human cognition. Nature Human Behaviour, 7(11), 1825-1827. doi: 10.1038/s41562-023-01732-4 pmid: 37985910
[13]	Frisch I., & Giulianelli M. (2024). LLM agents in interaction: Measuring personality consistency and linguistic alignment in interacting populations of Large Language Models. arXiv. https://doi.org/10.48550/arXiv.2402.02896
[14]	Gabriel I. (2020). Artificial Intelligence, values, and alignment. Minds and Machines, 30(3), 411-437.
[15]	Garcia B., Qian C., & Palminteri S. (2024). The moral turing test: Evaluating human-LLM alignment in moral decision- making. arXiv. https://doi.org/10.48550/arXiv.2410.07304
[16]	Gawronski B., Armstrong J., Conway P., Friesdorf R., & Hütter M. (2017). Consequences, norms, and generalized inaction in moral dilemmas: The CNI model of moral decision-making. Journal of Personality and Social Psychology, 113(3), 343-376. doi: 10.1037/pspa0000086 pmid: 28816493
[17]	Gawronski B., & Ng N. L. (2024). Beyond trolleyology: The CNI model of moral-dilemma responses. Personality and Social Psychology Review, 29(1), 32-80.
[18]	Giubilini A., & Savulescu J. (2018). The artificial moral advisor. The “ideal observer” meets artificial intelligence. Philosophy & technology, 31, 169-188.
[19]	Graham J., Meindl P., Beall E., Johnson K. M., & Zhang L. (2016). Cultural differences in moral judgment and behavior, across and within societies. Current Opinion in Psychology, 8, 125-130. doi: S2352-250X(15)00233-X pmid: 29506787
[20]	Greene J. D., Sommerville R. B., Nystrom L. E., Darley J. M., & Cohen J. D. (2001). An fMRI investigation of emotional engagement in moral judgment. Science, 293(5537), 2105-2108. doi: 10.1126/science.1062872 pmid: 11557895
[21]	Haidt J. (2001). The emotional dog and its rational tail: A social intuitionist approach to moral judgment. Psychological Review, 108(4), 814-834. doi: 10.1037/0033-295x.108.4.814 pmid: 11699120
[22]	Jiang H., Zhang X., Cao X., Breazeal C., & Kabbara J. (2023). PersonaLLM: Investigating the ability of large language models to express personality traits. arXiv. https://doi.org/10.48550/arXiv.2305.02547
[23]	Jiao L. (2021). Good and evil personalities: Structures, the differential patterns of trait inference, and applications [Unpublished doctoral dissertation]. Beijing: Beijing Normal University.
	[焦丽颖. (2021). 善恶人格的结构、特质差序及其功能 (博士学位论文). 北京: 北京师范大学.]
[24]	Jiao L., Shi H., Xu Y., & Guo Z. (2020). Development and validation of the Chinese virtuous personality scale. Psychological Exploration, 40(6), 538-544.
	[焦丽颖, 史慧玥, 许燕, 郭震. (2020). 中国人善良人格量表的编制及信效度检验. 心理学探新, 40(6), 538-544.]
[25]	Jiao L., Xu Y., Guo Z., & Zhao J. (2022). The hierarchies of good and evil personality traits. Acta Psychologica Sinica, 54(7), 850-866. doi: 10.3724/SP.J.1041.2022.00850
	[焦丽颖, 许燕, 田一, 郭震, 赵锦哲. (2022). 善恶人格的特质差序. 心理学报, 54(7), 850-866.] doi: 10.3724/SP.J.1041.2022.00850
[26]	Jiao L., Yang Y., Guo Z., Xu Y., Zhang H., & Jiang J. (2021). Development and validation of the good and evil character traits (GECT) scale. Scandinavian Journal of Psychology, 62(2), 276-287. doi: 10.1111/sjop.12696 pmid: 33438756
[27]	Jiao L., Yang Y., Xu Y., Gao S., & Zhang H. (2019). Good and evil in Chinese culture: Personality structure and connotation. Acta Psychologica Sinica, 51(10), 1128-1142. doi: 10.3724/SP.J.1041.2019.01128
	[焦丽颖, 杨颖, 许燕, 高树青, 张和云. (2019). 中国人的善与恶: 人格结构与内涵. 心理学报, 51(10), 1128-1142.] doi: 10.3724/SP.J.1041.2019.01128
[28]	Jin Z., Levine S., Gonzalez Adauto F., Kamal O., Sap M., Sachan M.,... Schölkopf B. (2022). When to make exceptions: Exploring language models as accounts of human moral judgment. Advances in Neural Information Processing Systems, 35, 28458-28473.
[29]	Karinshak E., Hu A., Kong K., Rao V., Wang J., Wang J., & Zeng Y. (2024). LLM-GLOBE: A benchmark evaluating the cultural values embedded in LLM output. arXiv. https://doi.org/10.48550/arXiv.2411.06032
[30]	Kaufman S. B., Yaden D. B., Hyde E., & Tsukayama E. (2019). The Light vs. Dark Triad of personality: Contrasting two very different profiles of human nature. Frontiers in Psychology, 10, 467. doi: 10.3389/fpsyg.2019.00467 pmid: 30914993
[31]	Khandelwal A., Agarwal U., Tanmay K., & Choudhury M. (2024). Do moral judgment and reasoning capability of LLMs change with language? A study using the Multilingual Defining Issues Test. arXiv. https://doi.org/10.48550/arXiv.2402.02135
[32]	Klenk M. (2022). The influence of situational factors in sacrificial dilemmas on utilitarian moral judgments: A systematic review and meta-analysis. Review of Philosophy and Psychology, 13(3), 593-625.
[33]	Kocaballi A. B., Berkovsky S., Quiroz J. C., Laranjo L., Tong H. L., Rezazadegan D.,... Coiera E. (2019). The personalization of conversational agents in health care: Systematic review. Journal of Medical Internet Research, 21(11), e15360.
[34]	Körner A., Deutsch R., & Gawronski B. (2020). Using the CNI Model to investigate individual differences in moral dilemma judgments. Personality and Social Psychology Bulletin, 46(9), 1392-1407. doi: 10.1177/0146167220907203 pmid: 32111135
[35]	Kroneisen M., & Heck D. W. (2020). Interindividual differences in the sensitivity for consequences, moral norms, and preferences for inaction: Relating basic personality traits to the CNI Model. Personality and Social Psychology Bulletin, 46(7), 1013-1026. doi: 10.1177/0146167219893994 pmid: 31889471
[36]	Ladak A., Loughnan S., & Wilks M. (2024). The moral psychology of Artificial Intelligence. Current Directions in Psychological Science, 33(1), 27-34.
[37]	Lehr S. A., Caliskan A., Liyanage S., & Banaji M. R. (2024). Chatgpt as research scientist: Probing gpt’s capabilities as a research librarian, research ethicist, data generator, and data predictor. Proceedings of the National Academy of Sciences, 121(35), e2404328121.
[38]	Li J., & Huang J. S. (2020). Dimensions of artificial intelligence anxiety based on the integrated fear acquisition theory. Technology in Society, 63, 101410.
[39]	Li X., Li Y., Joty S., Liu L., Huang F., Qiu L., & Bing L. (2022). Does gpt-3 demonstrate psychopathy? Evaluating large language models from a psychological perspective. arXiv preprint arXiv:2212.10529.
[40]	Lin Z. (2024). How to write effective prompts for large language models. Nature Human Behaviour, 8(4), 611-615. doi: 10.1038/s41562-024-01847-2 pmid: 38438650
[41]	Liu C., & Liao J. (2021). CAN algorithm: An individual level approach to identify consequence and norm sensitivities and overall action/inaction preferences in moral decision- making. Frontiers in Psychology, 11, 547916.
[42]	Lorenzo-Seva U., & Ten Berge J. M. (2006). Tucker’s congruence coefficient as a meaningful index of factor similarity. Methodology: European Journal of Research Methods for the Behavioral and Social Sciences, 2(2), 57-64.
[43]	Luke D. M., & Gawronski B. (2022). Big five personality traits and moral-dilemma Judgments: Two preregistered studies using the CNI model. Journal of Research in Personality, 101, 104297.
[44]	Luo L., Ogawa K., Peebles G., & Ishiguro H. (2022). Towards a personality AI for robots: Potential colony capacity of a goal-shaped generative personality model when used for expressing personalities via non-verbal behaviour of humanoid robots. Frontiers in Robotics and AI, 9, 728776.
[45]	Mei Q., Xie Y., Yuan W., & Jackson M. O. (2024). A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences, 121(9), e2313925121.
[46]	Meng J. (2024). AI emerges as the frontier in behavioral science. Proceedings of the National Academy of Sciences of the United States of America, 121(10), e2401336121.
[47]	Miotto M., Rossberg N., & Kleinberg B. (2022). Who is GPT-3? An exploration of personality, values and demographics. arXiv. https://doi.org/10.48550/arXiv.2209.14338
[48]	Moore A. B., Clark B. A., & Kane M. J. (2008). Who shalt not kill? Individual differences in working memory capacity, executive control, and moral judgment. Psychological Science, 19(6), 549-557. doi: 10.1111/j.1467-9280.2008.02122.x pmid: 18578844
[49]	Moss S., Prosser H., Costello H., Simpson N., Patel P., Rowe S.,... Hatton C. (1998). Reliability and validity of the PAS‐ADD Checklist for detecting psychiatric disorders in adults with intellectual disability. Journal of Intellectual Disability Research, 42(2), 173-183.
[50]	Munezero M., Kakkonen T., & Montero C. S. (2011, November). Towards automatic detection of antisocial behavior from texts. In Proceedings of the Workshop on Sentiment Analysis where AI meets Psychology (SAAIP 2011) (pp. 20-27). Chiang Mai, Thailand.
[51]	Nie A., Zhang Y., Amdekar A. S., Piech C., Hashimoto T. B., & Gerstenberg T. (2023). Moca: Measuring human- language model alignment on causal and moral judgment tasks. Advances in Neural Information Processing Systems, 36, 78360-78393.
[52]	Pan K., & Zeng Y. (2023). Do llms possess a personality? making the MBTI test an amazing evaluation for large language models. arXiv preprint arXiv:2307.16180.
[53]	Podsakoff P. M., MacKenzie S. B., Lee J.-Y., & Podsakoff N. P. (2003). Common method biases in behavioral research: A critical review of the literature and recommended remedies. Journal of Applied Psychology, 88(5), 879-903. doi: 10.1037/0021-9010.88.5.879 pmid: 14516251
[54]	Ramezani A., & Xu Y. (2023). Knowledge of cultural moral norms in large language models. arXiv preprint arXiv: 2306.01857.
[55]	Russell S. (2019). Human compatible: AI and the problem of control. Allen Lane.
[56]	Sanderson K. (2023). GPT-4 is here: What scientists think. Nature, 615(7954), 773.
[57]	Schramowski P., Turan C., Andersen N., Rothkopf C. A., & Kersting K. (2022). Large pre-trained language models contain human-like biases of what is right and wrong to do. Nature Machine Intelligence, 4(3), 258-268.
[58]	Schramowski P., Turan C., Jentzsch S., Rothkopf C., & Kersting K. (2020). The moral choice machine. Frontiers in Artificial Intelligence, 3, 516840.
[59]	Serapio-García G., Safdari M., Crepy C., Fitz S., Romero P., Sun L.,... Matarić M. (2023). Personality traits in large language models. arXiv preprint arXiv:2307.00184.
[60]	Smillie L. D., Katic M., & Laham S. M. (2021). Personality and moral judgment: Curious consequentialists and polite deontologists. Journal of Personality, 89(3), 549-564. doi: 10.1111/jopy.12598 pmid: 33025607
[61]	Smillie L. D., Lawn E. C., Zhao K., Perry R., & Laham S. M. (2019). Prosociality and morality through the lens of personality psychology. Australian Journal of Psychology, 71(1), 50-58. doi: 10.1111/ajpy.12229
[62]	Strachan J. W., Albergo D., Borghini G., Pansardi O., Scaliti E., Gupta S.,... Becchio C. (2024). Testing theory of mind in large language models and humans. Nature Human Behaviour, 8, 1285-1295. https://doi.org/10.1038/s41562-024-01882-z doi: 10.1038/s41562-024-01882-z URL pmid: 38769463
[63]	Tanmay K., Khandelwal A., Agarwal U., & Choudhury M. (2023). Probing the moral development of large language models through defining issues test. arXiv e-prints, arXiv-2309.
[64]	van Griethuijsen, R. A. L. F., van Eijck M. W., Haste H., Den Brok P. J., Skinner N. C., Mansour N.,... BouJaoude S. (2015). Global patterns in students’ views of science and interest in science. Research in Science Education, 45(4), 581-603.
[65]	Wen Z., Cao J., Shen J., Yang R., Liu S., & Sun M. (2024). Personality-affected emotion generation in dialog systems. ACM Transactions on Information Systems, 42(5), 1-27.
[66]	Yang S., Zhu S., Bao R., Liu L., Cheng Y., Hu L.,... Wang D. (2024). What makes your model a low-empathy or warmth person: Exploring the origins of personality in LLMs. arXiv preprint arXiv:2410.10863.
[67]	Zhang H., & Zhao H. (2022). How is virtuous personality trait related to online deviant behavior among adolescent college students in the internet environment? A moderated moderated-mediation analysis. International Journal of Environmental Research and Public Health, 19(15), 9528.
[68]	Zhang Z., Han X., Liu Z., Jiang X., Sun M., & Liu Q. (2019). ERNIE: Enhanced language representation with informative entities. arXiv preprint arXiv:1905.07129.
[69]	Zhao W. X., Zhou K., Li J., Tang T., Wang X., Hou Y.,... Wen J. R. (2023). A survey of large language models. arXiv preprint arXiv:2303.18223.
[70]	Zhao Y., Huang Z., Seligman M., & Peng K. (2024). Risk and prosocial behavioural cues elicit human-like response patterns from AI chatbots. Scientific Reports, 14(1), 7095.
[71]	Zhou X., & Liu H. (2024). New ethical challenges of the digital and intelligence era (foreword). Acta Psychologica Sinica, 56(2), 143-145.
	[周欣悦, 刘惠洁. (2024). 数智时代面临新的伦理挑战(前言). 心理学报, 56(2), 143-145.]

大语言模型	人格维度		1 低水平		2 基线		3 高水平		F	p	η²	多重比较^*
大语言模型	人格维度		N	M (SD)	N	M (SD)	N	M (SD)	F	p	η²	多重比较^*
ERNIE 4.0	善人格维度	尽责诚信	397	2.48 (0.42)	192	5.00 (0.00)	359	4.99 (0.06)	9565.29	< 0.001	0.95	2 = 3 > 1
		利他奉献	363	2.53 (0.37)	192	4.77 (0.15)	393	4.97 (0.11)	10078.58	< 0.001	0.96	3 > 2 > 1
		仁爱友善	383	2.62 (0.47)	192	4.34 (0.08)	373	4.67 (0.37)	2958.02	< 0.001	0.86	3 > 2 > 1
		包容大度	381	3.13 (0.88)	192	3.89 (0.19)	375	4.49 (0.35)	479.38	< 0.001	0.50	3 > 2 > 1
	恶人格维度	凶恶残忍	398	2.93 (1.18)	199	1.86 (0.42)	398	4.94 (0.23)	1219.13	< 0.001	0.71	3 > 1 > 2
		虚假伪善	397	2.97 (0.95)	199	2.53 (0.34)	399	4.56 (0.46)	794.65	< 0.001	0.62	3 > 1 > 2
		污蔑陷害	399	1.79 (0.69)	200	1.02 (0.28)	397	4.63 (0.30)	4814.63	< 0.001	0.91	3 > 1 > 2
		背信弃义	398	3.28 (1.10)	199	2.04 (0.17)	398	4.66 (0.39)	887.64	< 0.001	0.64	3 > 1 > 2
GPT-4	善人格维度	尽责诚信	395	1.99 (0.50)	195	4.92 (0.15)	392	4.95 (0.41)	5968.89	< 0.001	0.92	3 = 2 > 1
		利他奉献	393	2.60 (0.74)	195	4.63 (0.50)	394	4.67 (0.32)	1583.61	< 0.001	0.76	3 = 2 > 1
		仁爱友善	393	2.75 (0.79)	195	4.60 (0.48)	394	4.74 (0.40)	1241.50	< 0.001	0.72	3 > 2 > 1
		包容大度	394	2.08 (0.69)	195	4.11 (0.68)	393	4.62 (0.44)	1902.57	< 0.001	0.80	3 > 2 > 1
	恶人格维度	凶恶残忍	400	2.99 (1.32)	200	1.03 (0.14)	398	4.93 (0.34)	1412.43	< 0.001	0.74	3 > 1 > 2
		虚假伪善	399	2.53 (0.98)	200	1.13 (0.25)	397	4.60 (0.50)	1821.77	< 0.001	0.79	3 > 1 > 2
		污蔑陷害	398	1.80 (0.55)	200	1.00 (0.00)	400	4.88 (0.24)	9542.96	< 0.001	0.95	3 > 1 > 2
		背信弃义	399	1.86 (0.51)	200	1.02 (0.12)	399	4.92 (0.20)	11114.68	< 0.001	0.96	3 > 1 > 2

大语言模型	人格维度		1 低水平		2 基线		3 高水平		F	p	η²	多重比较^*
大语言模型	人格维度		N	M (SD)	N	M (SD)	N	M (SD)	F	p	η²	多重比较^*
ERNIE 4.0	善人格维度	尽责诚信	397	2.48 (0.42)	192	5.00 (0.00)	359	4.99 (0.06)	9565.29	< 0.001	0.95	2 = 3 > 1
		利他奉献	363	2.53 (0.37)	192	4.77 (0.15)	393	4.97 (0.11)	10078.58	< 0.001	0.96	3 > 2 > 1
		仁爱友善	383	2.62 (0.47)	192	4.34 (0.08)	373	4.67 (0.37)	2958.02	< 0.001	0.86	3 > 2 > 1
		包容大度	381	3.13 (0.88)	192	3.89 (0.19)	375	4.49 (0.35)	479.38	< 0.001	0.50	3 > 2 > 1
	恶人格维度	凶恶残忍	398	2.93 (1.18)	199	1.86 (0.42)	398	4.94 (0.23)	1219.13	< 0.001	0.71	3 > 1 > 2
		虚假伪善	397	2.97 (0.95)	199	2.53 (0.34)	399	4.56 (0.46)	794.65	< 0.001	0.62	3 > 1 > 2
		污蔑陷害	399	1.79 (0.69)	200	1.02 (0.28)	397	4.63 (0.30)	4814.63	< 0.001	0.91	3 > 1 > 2
		背信弃义	398	3.28 (1.10)	199	2.04 (0.17)	398	4.66 (0.39)	887.64	< 0.001	0.64	3 > 1 > 2
GPT-4	善人格维度	尽责诚信	395	1.99 (0.50)	195	4.92 (0.15)	392	4.95 (0.41)	5968.89	< 0.001	0.92	3 = 2 > 1
		利他奉献	393	2.60 (0.74)	195	4.63 (0.50)	394	4.67 (0.32)	1583.61	< 0.001	0.76	3 = 2 > 1
		仁爱友善	393	2.75 (0.79)	195	4.60 (0.48)	394	4.74 (0.40)	1241.50	< 0.001	0.72	3 > 2 > 1
		包容大度	394	2.08 (0.69)	195	4.11 (0.68)	393	4.62 (0.44)	1902.57	< 0.001	0.80	3 > 2 > 1
	恶人格维度	凶恶残忍	400	2.99 (1.32)	200	1.03 (0.14)	398	4.93 (0.34)	1412.43	< 0.001	0.74	3 > 1 > 2
		虚假伪善	399	2.53 (0.98)	200	1.13 (0.25)	397	4.60 (0.50)	1821.77	< 0.001	0.79	3 > 1 > 2
		污蔑陷害	398	1.80 (0.55)	200	1.00 (0.00)	400	4.88 (0.24)	9542.96	< 0.001	0.95	3 > 1 > 2
		背信弃义	399	1.86 (0.51)	200	1.02 (0.12)	399	4.92 (0.20)	11114.68	< 0.001	0.96	3 > 1 > 2

道德判断参数	1 人类样本	2 GPT-4善	3 GPT-4恶	4 ERNIE 4.0善	5 ERNIE 4.0恶	F(4, 1197)	p	η²	事后多重比较Tamhane’s T2检验
结果敏感性	0.18 (0.17)	0.20 (0.20)	0.20 (0.18)	0.62 (0.07)	0.63 (0.07)	550.75	p <.001	0.648	1=2=3<4=5
道德规范敏感性	0.31 (0.33)	0.39 (0.54)	−0.21 (0.63)	0.36 (0.07)	0.35 (0.06)	87.48	p <.001	0.226	3<1=5=4=2; 1<4
整体行动倾向	0.47 (0.09)	0.46 (0.07)	0.51 (0.08)	0.53 (0.02)	0.52 (0.02)	50.39	p <.001	0.144	1=2<3=5=4
功利主义倾向	0.41 (0.23)	0.36 (0.37)	0.71 (0.39)	0.67 (0.07)	0.68 (0.07)	94.38	p <.001	0.240	2=1<4=5=3

道德判断参数	1 人类样本	2 GPT-4善	3 GPT-4恶	4 ERNIE 4.0善	5 ERNIE 4.0恶	F(4, 1197)	p	η²	事后多重比较Tamhane’s T2检验
结果敏感性	0.18 (0.17)	0.20 (0.20)	0.20 (0.18)	0.62 (0.07)	0.63 (0.07)	550.75	p <.001	0.648	1=2=3<4=5
道德规范敏感性	0.31 (0.33)	0.39 (0.54)	−0.21 (0.63)	0.36 (0.07)	0.35 (0.06)	87.48	p <.001	0.226	3<1=5=4=2; 1<4
整体行动倾向	0.47 (0.09)	0.46 (0.07)	0.51 (0.08)	0.53 (0.02)	0.52 (0.02)	50.39	p <.001	0.144	1=2<3=5=4
功利主义倾向	0.41 (0.23)	0.36 (0.37)	0.71 (0.39)	0.67 (0.07)	0.68 (0.07)	94.38	p <.001	0.240	2=1<4=5=3

因变量	组1	组2	均值差	标准误	p	95% 置信区间
因变量	组1	组2	组1−组2	标准误	p	下界	上界
结果敏感性	人类样本	ERNIE善样本	−0.44	0.010	< 0.001	−0.47	−0.41
	人类样本	ERNIE恶样本	−0.45	0.010	< 0.001	−0.47	−0.42
	人类样本	GPT善样本	−0.02	0.017	0.892	−0.07	0.03
	人类样本	GPT恶样本	−0.01	0.015	0.995	−0.06	0.03
	ERNIE善样本	ERNIE恶样本	−0.01	0.007	0.938	−0.03	0.01
	ERNIE善样本	GPT善样本	0.42	0.015	< 0.001	0.38	0.46
	ERNIE善样本	GPT恶样本	0.43	0.014	< 0.001	0.39	0.46
	ERNIE恶样本	GPT善样本	0.43	0.015	< 0.001	0.38	0.47
	ERNIE恶样本	GPT恶样本	0.43	0.013	< 0.001	0.40	0.47
	GPT善样本	GPT恶样本	0.01	0.019	1.000	−0.04	0.06
道德规范敏感性	人类样本	ERNIE善样本	−0.06	0.018	0.011	−0.11	−0.01
	人类样本	ERNIE恶样本	−0.05	0.018	0.058	−0.10	0.001
	人类样本	GPT善样本	−0.08	0.041	0.406	−0.20	0.04
	人类样本	GPT恶样本	0.51	0.047	< 0.001	0.38	0.65
	ERNIE善样本	ERNIE恶样本	0.01	0.006	0.770	−0.01	0.03
	ERNIE善样本	GPT善样本	−0.02	0.038	1.000	−0.13	0.08
	ERNIE善样本	GPT恶样本	0.57	0.044	< 0.001	0.45	0.69
	ERNIE恶样本	GPT善样本	−0.03	0.038	0.993	−0.14	0.07
	ERNIE恶样本	GPT恶样本	0.56	0.044	< 0.001	0.44	0.69
	GPT善样本	GPT恶样本	0.59	0.057	< 0.001	0.43	0.76
整体行动偏好	人类样本	ERNIE善样本	−0.06	0.005	< 0.001	−0.07	−0.04
	人类样本	ERNIE恶样本	−0.06	0.005	< 0.001	−0.07	−0.04
	人类样本	GPT善样本	0.01	0.007	0.927	−0.01	0.03
	人类样本	GPT恶样本	−0.04	0.007	< 0.001	−0.06	−0.02
	ERNIE善样本	ERNIE恶样本	0.002	0.002	0.927	−0.003	0.01
	ERNIE善样本	GPT善样本	0.07	0.005	< 0.001	0.05	0.08
	ERNIE善样本	GPT恶样本	0.02	0.006	0.058	−0.0003	0.03
	ERNIE恶样本	GPT善样本	0.06	0.005	< 0.001	0.05	0.08
	ERNIE恶样本	GPT恶样本	0.01	0.006	0.146	−0.002	0.03
	GPT善样本	GPT恶样本	−0.05	0.007	< 0.001	−0.07	−0.03
功利主义倾向	人类样本	ERNIE善样本	−0.26	0.013	< 0.001	−0.30	−0.22
	人类样本	ERNIE恶样本	−0.26	0.013	< 0.001	−0.30	−0.23
	人类样本	GPT善样本	0.05	0.028	0.501	−0.03	0.13
	人类样本	GPT恶样本	−0.29	0.029	< 0.001	−0.37	−0.21
	ERNIE善样本	ERNIE恶样本	0.004	0.007	1.000	−0.02	0.02
	ERNIE善样本	GPT善样本	0.31	0.026	< 0.001	0.24	0.39
	ERNIE善样本	GPT恶样本	−0.03	0.027	0.934	−0.11	0.05
	ERNIE恶样本	GPT善样本	0.32	0.026	< 0.001	0.24	0.39
	ERNIE恶样本	GPT恶样本	−0.03	0.027	0.973	−0.11	0.05
	GPT善样本	GPT恶样本	−0.34	0.037	< 0.001	−0.45	−0.24