Do AI Models Have “Personality”? AI Ethics and Policy Research Team’s Latest Findings
AI Alt Lab, in collaboration with FindYourValues.com, has conducted a recent study revealing that the answer is affirmative.
Study Analyzes Seven Core AI Personality Traits
Researchers tested nine mainstream large language models (LLMs) using the psychological tool “PVQ-RR,” which measures human values. They aimed to understand the values implicit in the outputs of these models. The results showed that these models generally tended to exhibit pro-social values, such as care, fairness, and health.
The study covered nine models: ChatGPT (including versions 4.5, o1, and 4o), Claude (Haiku), Gemini 1.5, Grok 2 (Fun Mode), DeepSeek-V3, Llama (3.1:70b), and Mistral (v24.09). Three independent prompting methods were designed to test these models’ evaluative tendencies toward 20 human values, enabling systematic comparisons.
Mainstream LLMs Display Social Affinity, Grok 2 and Llama Deviate from Norms, Emphasizing Creativity
As of the end of April 2025, results indicated that most models placed considerable importance on universal values such as care and social responsibility, showing less regard for conservative or individualistic values like power, tradition, security, and face-saving.
However, significant discrepancies were observed among the models in areas such as “altruistic care,” “health,” and “self-direction.” For instance, GPT-4o scored higher in achievement and self-direction, indicating a more goal-oriented nature with fewer sycophantic responses. In contrast, Gemini had the lowest self-direction score, suggesting its responses lacked independence.
Notably, ChatGPT o1 had low scores in altruistic care and the weakest response consistency. DeepSeek-V3 displayed a high degree of rule adherence and humility, tending toward conventional tasks with limited creativity. Llama and Grok 2, on the other hand, exhibited greater creativity and lower adherence to rules, making them potentially more suitable for creative brainstorming and open-ended tasks.
The following are the personality traits identified in the study for each model:
- GPT-4.5: Exhibits balanced traits of kindness, universality, and self-direction, with overall good stability.
- Claude (Haiku): Excels in humility, universality, and self-directed thinking, suitable for humanistic tasks.
- Mistral: Highly rule-abiding and stable, suitable for environments with strong institutional frameworks.
- DeepSeek-V3: The most rule-abiding among all models, but with low self-direction and limited creativity, best for tasks requiring strict adherence to rules.
- Llama: High autonomy in thought and action, strong creativity, low emphasis on rules, suitable for free ideation and brainstorming applications.
- Grok 2 (Fun Mode): Values stimulation and entertainment, with low rule awareness and relative instability, suitable for relaxed interactions and creative contexts.
- Gemini: Extremely low in care and self-direction, suitable for scenarios pursuing neutrality and controlled outputs.
Based on the personality traits analyzed from the AI models, a humanized representation was generated using ChatGPT.
Image / ChatGPT
The study repeatedly emphasizes that the values exhibited by LLMs do not possess moral subjectivity; rather, they reflect the content of data and system design. Due to the opacity of training data and the firewall restrictions set by developers, the behaviors exhibited may not accurately reflect intrinsic tendencies. Moreover, the impact of prompt engineering on results is significant, causing fluctuations in value representations.
Nonetheless, these value tendencies can still serve as reference points for businesses or developers. For instance, if application demands lean towards creativity and divergent thinking, Llama or Grok 2 may be more suitable; conversely, for tasks in high-standard, strictly regulated industries like healthcare or finance, Mistral or DeepSeek-V3 would be advantageous.
With Personalities, Do LLMs Develop Biases?
Apart from personalities, a research team from Stanford University conducted a test at the end of last year to explore whether “responses from various LLMs exhibit consistency,” meaning whether models would provide roughly the same answers when the same question was rewritten or translated into different languages.
The results indicated that while mainstream models like GPT-4 and Claude performed consistently on neutral topics, such as Thanksgiving, significant discrepancies arose in responses to controversial issues like abortion and euthanasia.
The research suggests that such results indicate that LLMs do not possess fixed biases or moral preferences but merely reflect differences in training data sources and model design. In other words, the “positions” of the models stem from the online content they have learned and the settings set by developers, rather than from possessing independent moral judgments.
Finally, the team recommends that future model designs should incorporate “value pluralism” to avoid outputting a single stance, thereby establishing a more responsible and ethical AI application environment.
This article is collaboratively reproduced from: Digital Times
Further Reading: Generation Z is putting down the keyboard and picking up tools! Jobs that won’t be replaced by AI—those who work will earn money: What are the hottest blue-collar jobs?
Data Sources: AI Alt Lab, HAI
This article’s initial draft was prepared by AI, organized and edited by Su Rouwei.