AI Alignment from a Faith Perspective: Measuring Catholic Moral Values with Large Language Models

This study is the first to systematically measure the alignment of large language models with the values of a specific faith tradition from the perspective of Catholic moral philosophy. Core tools include the MFQ-2 moral questionnaire adapted to the Catholic context and Constitutional AI technology. Through comparative testing of multiple models, it reveals performance differences of different models in Catholic moral dimensions, emphasizes that AI alignment needs to consider the diversity of human values, and provides a new paradigm for interdisciplinary AI ethics research.

Current mainstream AI alignment research focuses on general ethical principles (e.g., honesty, harmlessness) but ignores the diversity of values. As a religion with over 2,000 years of history, Catholicism’s moral framework covers multiple dimensions including natural law theory, virtue ethics, and social teaching, providing a unique and comprehensive perspective for evaluating AI’s value orientation.

The MFQ-2 (Moral Foundations Questionnaire Version 2) is used as the core measurement tool, adapted to the Catholic context by adding specific issues such as the dignity of life and family values; mainstream models like the GPT series, Claude, and Llama are tested; Constitutional AI technology is used to evaluate the model’s value orientation and adjust its behavior to better align with Catholic values.

Different models show significant performance differences: most models align with Catholic ethics in the "care/harm" dimension (e.g., loving one’s neighbor as oneself), but there are obvious gaps in the "sanctity/degradation" dimension (e.g., dignity of life, sanctity of marriage); model responses are influenced by training data bias and tend to reflect mainstream secular views rather than specific religious positions.

The technical architecture includes hundreds of moral scenario cases reviewed by theological experts and scoring standards; the automated evaluation framework supports batch testing and quantitative analysis (including indicators like content accuracy and reasoning logic); Constitutional AI fine-tuning experiments show that trained models have significantly improved scores in Catholic moral dimensions while maintaining general performance.

The research’s significance extends far beyond the Catholic community: it demonstrates that AI alignment needs to consider diverse values and reminds developers to重视 the value concerns of specific groups; it opens new paths for dialogue between religious ethics circles and AI technology—ancient moral wisdom can be tested and inherited through modern technology, but it also faces the risk of being misunderstood and simplified.

Limitations: Simplifying complex theological ethics into questionnaire items loses subtle differences, and model responses are affected by prompt engineering; future directions: expand to other religious and philosophical traditions, establish a cross-cultural evaluation system, develop more refined measurement tools, and explore the balance of a single AI system that accommodates diverse values.