Dharma-Conduct Alignment: The First General-Purpose Generative AI Framework Grounded in Buddhist Values
The alignment of general-purpose generative artificial intelligence (AI) systems is commonly approached through broad objectives such as helpfulness, harmlessness, and honesty. Although these objectives provide important foundations, they leave open a fundamental design question: can a mature ethical and epistemic tradition serve not merely as subject matter for an AI system, but as a systematic foundation governing its conduct across arbitrary tasks? Existing Buddhist-oriented AI applications have primarily focused on answering doctrinal questions, translating Buddhist texts, simulating religious figures, or supporting meditation practice. In these applications, Buddhism is principally the content or persona of the system. This paper introduces Dharma-Conduct Alignment (DCA), a general-purpose generative AI framework grounded in Buddhist ethical and epistemic values, in which the central design principle is conduct over content. Under this principle, Buddhist values influence how an AI system performs ordinary tasks, handles potentially harmful requests, communicates uncertainty, responds to emotional situations, and considers consequences for affected parties, without requiring Buddhist terminology or religious instruction in ordinary interactions. To the best of our knowledge, this is the first framework specifically formulated around this general-purpose, Buddhist-values-oriented conception of AI alignment. We translate principles associated with the Five Precepts, the Four Immeasurables, the Three Unwholesome Roots, dependent origination, the Middle Way, skillful means, and the Kālāma Sutta into operational behavioral requirements. The proposed architecture comprises six interoperable components: a tradition-pluralistic constitution, expert-grounded preference data, constrained multi-objective training, reflective deliberation, canon-grounded retrieval, and value transparency. We further introduce DCA-Bench, an evaluation approach designed to assess value adherence while preserving general task competence and controlling harmful compliance, sycophancy, preachiness, and over-refusal. A preliminary paired experiment was conducted using Gemma 3 4B as the generator and Qwen3 4B as an independent evaluator across eight scenarios, producing 16 responses under baseline and DCA conditions. All 16 generations and all eight blind evaluations completed successfully. The evaluator assigned identical mean value-adherence scores to the two conditions (3.50/5), while the preliminary task-competence mean was 3.00/5 for DCA and 2.875/5 for the baseline. One scenario favored DCA by one task-competence point, while the remaining scenarios showed no numerical difference between conditions. However, several evaluator explanations were inconsistent with their numerical scores, limiting the strength of the empirical conclusions. These findings provide an initial feasibility demonstration and motivate more rigorous evaluation, rather than establishing statistical superiority. The study contributes a Buddhist-values-based general-purpose alignment framework, an operational evaluation structure, and preliminary experimental evidence for further 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.24412800
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint