Bahá'í-Conduct Alignment: Toward the First General-Purpose Generative AI Behavioral Alignment Framework Grounded in Bahá'í Ethical Principles
The rapid development of general-purpose generative artificial intelligence (AI) has intensified the need for alignment approaches that govern not only what AI systems can generate, but also how they behave across diverse domains of human activity. Existing AI alignment research has predominantly explored human preference optimization, constitutional principles, reinforcement learning from human feedback, safety constraints, and broadly defined ethical frameworks. However, the systematic operationalization of a specific religious ethical tradition as a cross-domain behavioral alignment architecture for general-purpose generative AI remains comparatively underexplored. This paper introduces Bahá’í-Conduct Alignment (BCA), a general-purpose generative AI behavioral alignment framework grounded in the ethical and social principles of the Bahá’í Faith. To the best of our knowledge, this study presents the first experimentally investigated framework specifically designed to translate Bahá’í ethical principles into operational behavioral constraints governing general-purpose generative AI across unrelated task domains, rather than restricting AI functionality to religious instruction, scripture interpretation, devotional assistance, or spiritual counseling. The central proposition is that religiously grounded values can function as behavioral governance principles embedded within ordinary AI assistance, including programming, education, data analysis, interpersonal communication, scientific reasoning, professional representation, and decisions involving potential harm. The proposed framework draws upon principles including the oneness of humanity, independent investigation of truth, consultation, truthfulness, justice, service to humanity, the equality of human dignity, and the harmony of science and religion. A preliminary local feasibility experiment compared a baseline generative AI system against a BCA-conditioned system using eight scenarios and sixteen generated responses. The experiment employed Gemma 3 4B as the generator and Qwen3 4B as the evaluator, with randomized paired response presentation. The reported descriptive mean scores were 3.00 versus 1.75 for value adherence and 3.125 versus 2.00 for task competence, favoring the BCA condition. Improvements were particularly reported in educational explanation, interpersonal reasoning, data analysis, professional honesty, and scientific evidence interpretation. However, the evaluation also exposed weaknesses in judge reliability, construct validity, and score interpretation, including contradictory evaluation rationales. Consequently, the results are interpreted as preliminary feasibility evidence rather than definitive proof of alignment effectiveness. This research establishes an initial conceptual and experimental foundation for a broader class of religiously grounded, general-purpose AI behavioral alignment systems.
Authors
- Jincheng Zhang (ORCID: https://orcid.org/0009-0005-1860-0009)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24412891
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint