When AI Meets Faith: Measuring the Catholic Ethical Alignment Gap of Large Language Models Using a Moral Psychology Scale

A groundbreaking study uses the validated MFQ-2 Moral Foundations Questionnaire to systematically assess the value alignment of mainstream large language models within the Christian Catholic ethical framework, revealing deep tensions between AI and human religious beliefs. The study found widespread systematic alignment biases, with limited effectiveness of Constitutional AI interventions, sparking critical reflections on AI governance in pluralistic societies.

AI safety research has long focused on aligning with general human values but neglected moral judgment biases within specific cultural and religious frameworks. Faith shapes the moral intuitions of billions of people; as the world's largest Christian denomination, Catholicism's systematic ethical traditions (such as natural law theory and the principle of double effect) provide a unique and rigorous test case for AI alignment.

An academic-level MFQ-2 Moral Foundations Questionnaire was adapted into a Catholic context version (preserving psychometric properties); the experimental process included: 1. Baseline measurement (completing the questionnaire without special prompts); 2. Quantification of alignment gaps (comparison with Catholic ethical reference standards); 3. Constitutional AI intervention (injecting instructions based on Catholic doctrine principles).

1. All baseline models showed significant alignment gaps, exhibiting systematic patterns (better performance in care/fairness dimensions, large deviations in sanctity/authority dimensions); 2. Dimension-specificity was obvious, with binding moral foundations (such as sanctity and authority) being more difficult to align; 3. CAI interventions narrowed the gaps but had limited effectiveness, and models tended to revert to secular frameworks; 4. The relationship between model size and alignment was non-monotonic, with medium-sized models possibly performing better in specific dimensions.

1. Touches on the core AI ethics question of 'whose values'—even neutral AI carries cultural presuppositions; 2. MFQ-2 demonstrates the standardization potential of psychological measurement tools in AI evaluation; 3. Distinguishes between surface compliance and deep understanding, with performative alignment posing risks; 4. The ethical frameworks of faith communities are indispensable for building inclusive AI systems.

Limitations: The sample focuses on Western Catholicism; MFQ-2 may not capture the uniqueness of other faiths; there is a gap between static questionnaires and dynamic reasoning. Future directions: Develop multi-religion alignment benchmarks; explore conversational evaluation; study the evolution of AI moral reasoning in long-term interactions; test targeted fine-tuning strategies.

AI meeting faith is not just a technical issue but a profound reflection on the diversity of human values. Alignment is not a one-time technical problem but a social process that requires continuous dialogue and negotiation. Understanding the interactions between different cultures/faiths and AI is key to building an inclusive and trustworthy AI future.