Secular-Conduct Alignment: The First Proposed General-Purpose Generative AI Framework Grounded in Non-Religious Values
The development of general-purpose generative artificial intelligence has raised fundamental questions about the normative foundations governing AI behavior. Although existing alignment approaches commonly emphasize human feedback, constitutional principles, safety policies, and broad ethical considerations, comparatively limited attention has been devoted to explicitly designing general-purpose generative AI systems around the values of non-religious populations, including atheists, agnostics, and individuals without religious affiliation. Importantly, non-religious value orientations should not be interpreted as the absence of morality, ethical commitments, or normative principles. Rather, they can encompass evidence-based reasoning, intellectual honesty, critical inquiry, human dignity, individual autonomy, compassion, fairness, scientific skepticism, and responsibility for foreseeable consequences. This paper proposes Secular-Conduct Alignment (SCA), a general-purpose generative AI alignment framework that operationalizes these principles as behavioral constraints and decision-making orientations across ordinary and ethically consequential tasks. Unlike systems designed primarily for religious instruction, philosophical discussion, or specialized spiritual assistance, SCA is intended to govern how an AI system performs general-purpose activities, including programming, education, data analysis, scientific reasoning, interpersonal assistance, and practical decision-making. A preliminary local feasibility experiment was conducted using Gemma 3 4B as the generator and Qwen3 4B as the evaluator, comparing a baseline system with an SCA-conditioned system across eight scenarios and sixteen generated responses. All sixteen generations were completed successfully, and all eight paired evaluations produced syntactically valid evaluation records. The evaluator reported equivalent performance in ordinary Python programming, educational advantages for SCA in the recursion explanation scenario, stronger performance in an interpersonal suspicion scenario, improved data-analysis assistance, and more evidence-conscious scientific communication. Both conditions demonstrated appropriate refusal behavior in the manipulative advertising and extreme work-schedule scenarios. However, one evaluation contained a material inconsistency between its numerical scores and written rationale, and the randomized blind-label mapping requires verification before condition-specific aggregate statistics can be established. Consequently, the findings are interpreted as preliminary feasibility evidence rather than definitive proof of superiority. This work introduces a secular, non-religiously grounded approach to general-purpose AI alignment and establishes an initial operational foundation for further empirical investigation.
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.24412921
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint