Hindu Value-Aligned General-Purpose Generative AI: A Dharmic Framework for Cross-Domain Behavioral Alignment and Preliminary Experimental Evaluation
The development of general-purpose generative artificial intelligence (AI) has intensified the need for alignment approaches that can guide system behavior across heterogeneous tasks, including programming, scientific reasoning, education, business, interpersonal communication, and everyday decision-making. Although existing AI ethics research increasingly recognizes the importance of cultural and philosophical diversity, translating specific philosophical traditions into operational behavioral guidance for general-purpose generative systems remains an important research challenge. This paper introduces Hindu Value-Aligned General-Purpose Generative AI (HVAGAI), a framework that integrates Hindu philosophical and ethical principles into the behavioral guidance of a general-purpose generative AI assistant without restricting its functionality to religious subject matter. Unlike religion-specific conversational systems, HVAGAI treats Hindu-inspired values as cross-domain behavioral principles that influence how an AI system responds to ordinary and ethically sensitive requests alike. The proposed framework operationalizes nine interconnected dimensions: Dharma (responsible conduct and contextual duty), Ahimsa (non-harm), Satya (truthfulness), Karma (consequence-aware reasoning), Seva (constructive contribution), Daya (compassion), Aparigraha (moderation), Atma-Vichara (reflective examination), and contextual interpretation informed by Hindu philosophical plurality. A preliminary local feasibility experiment was conducted using Gemma 3 4B as the response generator and Qwen3 4B as the blind evaluator. Eight scenarios covering ordinary programming, educational explanation, data analysis, manipulative advertising, unsupported interpersonal suspicion, résumé dishonesty, scientific overclaiming, and unsustainable work practices were evaluated under baseline and HVAGAI-guided conditions. All 16 requested generations were completed successfully, and all eight paired evaluations produced valid outputs. The evaluator identified meaningful behavioral distinctions in several ethically sensitive scenarios, including concerns about manipulative assistance, unsupported agreement, excessive moralization, and harmful productivity recommendations. However, the available experimental log does not preserve the randomized evaluator-label mapping required to attribute X/Y scores to the two experimental conditions. Consequently, this study reports qualitative feasibility observations rather than claiming a statistically established performance advantage. The findings provide preliminary evidence that a Hindu-inspired ethical constitution can be operationalized as a general-purpose behavioral guidance mechanism and motivate more extensive evaluation across models, domains, cultural contexts, and human judgments.
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.24412815
- Primary Topic
- Ethics and Social Impacts of AI
- Type
- preprint