Psychological Profiling of LLM Moral Reasoning: Linking Value Classifications to Human Traits

Drawing inspiration from Zangari et al. (2024) and Gloor et al. (2022), this research would systematically analyze the moral value classifications made by LLMs across diverse contexts, treating the model as a “subject” whose moral outputs can be profiled using tools from psychology (e.g., Schwartz’s value survey, Big Five traits, or MFT). The goal is to uncover whether LLMs cluster into recognizable “moral personalities”—such as being more “conscientious,” “liberal,” or “authoritarian”—and how this varies across architectures, training data, or languages (see Agarwal et al., 2024 for multilingual effects). This could reveal hidden biases, but also suggest new ways to tune models for specific applications (e.g., matching the moral “personality” of a model to the context of use). It synthesizes methods from psychology, ethics, and NLP, and could yield both practical tools for model selection and deeper theoretical insights into AI alignment.

References:

1. A survey on moral foundation theory and pre-trained language models: current advances and challenges. Lorenzo Zangari, C. M. Greco, Davide Picca, A. Tagarelli (2024). Ai & Society.
2. Measuring ethical behavior with AI and natural language processing to assess business success. P. Gloor, Andrea Fronzetti Colladon, F. Grippa (2022). Scientific Reports.
3. Ethical Reasoning and Moral Value Alignment of LLMs Depend on the Language We Prompt Them in. Utkarsh Agarwal, K. Tanmay, Aditi Khandelwal, Monojit Choudhury (2024). International Conference on Language Resources and Evaluation.