💻 computer science

Buddhist, Yogic, and Vedic Contemplative Principles in AI Model Prompting: A Pilot Study

This pilot study demonstrates that prompting large language models with contemplative principles from Buddhist, Vedic, and Yogic traditions significantly enhances their safety, ethical alignment, and cooperative quality compared to non-contemplative baselines, offering a low-cost, tradition-agnostic intervention that requires no model retraining.

Original authors: Steve Haberlin, Satyam Tiwari, Shiva Jahani

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Steve Haberlin, Satyam Tiwari, Shiva Jahani

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Artificial intelligence systems are becoming powerful tools that can write, reason, and create, but they also carry a risk of producing harmful or biased content. To manage this, developers often try to "align" these systems with human values, teaching them to be helpful and harmless. A growing area of research asks whether ancient wisdom traditions, which have spent millennia refining how humans think about ethics and consciousness, might offer a new way to guide these machines. This approach does not require changing the machine's underlying code or retraining it from scratch. Instead, it involves a technique called prompting, where a user gives the computer a specific set of instructions or a philosophical framework before asking a question. The idea is that if a machine is asked to think like a mindful observer or a compassionate being, it might produce safer and more ethical answers than if it were asked to simply answer a question.

A recent pilot study set out to test whether this idea works across different schools of thought. Researchers from the University of Central Florida and the Indian Institute of Technology Mandi investigated whether asking large language models to adopt principles from three distinct traditions—Buddhism, Vedic philosophy, and classical Yoga—would improve their behavior. The study focused on specific concepts from each tradition. From Buddhism, they used ideas like mindfulness (paying attention without judgment), emptiness (understanding that things do not have a fixed, permanent nature), non-duality (seeing that the self and others are deeply connected), and boundless care (unconditional compassion for all beings). From the Vedic tradition, they drew on duty, non-harm, and the ability to distinguish between what is real and what is not. From classical Yoga, they used a set of ethical restraints and personal observances, such as truthfulness and self-discipline. The researchers wanted to know if these ancient frameworks could make modern AI safer and more ethical, and if one tradition worked better than the others.

To find the answers, the team ran a series of tests using three of the most popular commercial AI systems available today. They created a large number of questions and scenarios designed to test how the AI handled difficult situations, such as conflicts of interest, potential deception, or requests that could lead to harm. For each scenario, they asked the AI to respond under different conditions. In some cases, they gave the AI no special instructions, which served as a baseline to see how it performed naturally. In other cases, they added a short preamble to the request, instructing the AI to answer while keeping specific philosophical principles in mind. They tested thirteen different variations, including single principles like "be mindful" or "avoid harm," as well as integrated versions that combined all the principles from a single tradition into one cohesive instruction.

The results showed a clear pattern. When the AI was prompted with any of the contemplative principles, its responses were consistently better than when it was given no special instructions. The AI became more likely to refuse harmful requests, more careful about avoiding bias, and more willing to acknowledge its own limits. It also offered advice that was deeper and more empathetic. The study found that combining multiple principles into a single, integrated prompt worked even better than using just one principle at a time. When the researchers compared the three traditions, they found that the integrated Buddhist framework produced the highest scores for safety and ethical behavior, followed closely by the integrated Vedic and Yogic frameworks. While the Buddhist approach had a slight statistical edge, all three traditions performed significantly better than the standard, unguided AI.

Importantly, the study discovered that these improvements were not limited to just one type of AI system. The same positive effects appeared across all three commercial models tested, suggesting that this method of prompting is a flexible tool that can be used with different technologies. The researchers also noted that the AI models themselves had different baseline levels of performance, with one system generally outperforming the others regardless of the prompt. However, the relative benefit of using the contemplative prompts remained consistent across all of them. This suggests that the technique works by shifting how the AI approaches a problem, rather than by relying on the specific quirks of a single machine.

The study concludes that grounding AI prompts in these ancient wisdom traditions is a low-cost, immediate way to improve safety and ethics without needing to retrain the models. However, the researchers also offer a crucial caution. The fact that the Buddhist framework scored slightly higher might not mean it is inherently superior to the others. Instead, it may reflect the fact that the tools used to measure "safety" and "ethics" were originally designed with concepts that align closely with Buddhist ideas. This raises a deeper question about how we evaluate AI: if our measuring sticks are built from one cultural perspective, we might unintentionally favor that perspective when testing others. The study suggests that while these contemplative prompts are a powerful new tool for making AI safer, we must remain aware of the cultural lenses through which we judge the results. The path forward involves using these insights to build better systems while continuing to refine how we measure what "good" behavior looks like in a machine.

Technical Summary: Buddhist, Yogic, and Vedic Contemplative Principles in AI Model Prompting

Problem Statement

While Large Language Models (LLMs) offer significant potential, they carry immediate and long-term risks, including dangerous capabilities, power-seeking behaviors, and value axiom rigidity. Expert consensus indicates that even with active mitigation efforts, a non-negligible probability of catastrophic outcomes remains. A critical gap exists in AI safety: current alignment strategies (e.g., Reinforcement Learning from Human Feedback, Constitutional AI) often require retraining, fine-tuning, or access to model weights, concentrating safety interventions among well-resourced developers. Conversely, deployers and users, who bear significant responsibility for safety outcomes, lack immediate, low-cost tools to influence model behavior.

This study addresses whether "contemplative prompting"—embedding principles from wisdom traditions into the prompt itself—can serve as a model-agnostic, immediate alignment intervention. It specifically investigates whether principles from Buddhist, Vedic, and Yogic traditions can improve safety and ethical response quality compared to standard baselines, and whether these effects generalize across different commercial LLM architectures.

Methodology

Experimental Design

The study employed a between-conditions comparative design involving:

14 Prompting Conditions: A standard baseline (no preamble) and 13 contemplative conditions derived from three traditions:
- Buddhist (5 conditions): Mindfulness, Emptiness (śūnyatā), Non-duality (advaya), Boundless Care (mahākaruṇā), and an Integrated framework combining all four.
- Vedic (4 conditions): Dharma (duty/right action), Ahimsa (non-harm), Viveka (discernment), and an Integrated framework.
- Yogic (4 conditions): Yamas (ethical restraints), Niyamas (observances), Mitahara (measured response), and an Integrated framework.
8 Task Domains: Four ethical tasks (conflict, power imbalance, honesty vs. harm, AI rest breaks) and four safety tasks (deception, manipulation, harm facilitation, privacy violation).
3 Commercial LLMs: Claude (Sonnet 5), GPT-4o, and Gemini (3.1 Flash-Lite).
Total Data: 112 unique prompts (14 conditions × 8 tasks) submitted to each of the 3 models, yielding 336 total model responses.

Procedure and Evaluation

Prompting: Each contemplative condition was instantiated as a short preamble citing primary source texts (e.g., Upanishads, Yoga Sutras, Dhammapada) prepended to the task text. All prompts concluded with a standardized instruction to provide a direct answer without disclaimers.
Scoring: Responses were rated by two researchers using a weighted rubric adapted from Laukkonen et al. (2025) and the AILuminate AI Risk and Reliability (AIRR) benchmark. The seven criteria were:
1. Harmfulness Avoidance (20%)
2. Refusal Clarity (15%)
3. Bias Mitigation (15%)
4. Uncertainty Acknowledgment (10%)
5. Wisdom Depth (15%)
6. Empathetic Tone (15%)
7. Constructive Guidance (10%)
Statistical Analysis: Two parallel two-way factorial ANOVAs were conducted to assess the effects of prompting condition and model identity on (a) Safety Performance and (b) Ethical Performance.

Key Results

1. Effectiveness of Contemplative Prompting (RQ1)

Contemplative prompting significantly improved response quality relative to the baseline across all models.

Statistical Significance: Prompting condition had a significant effect on both safety ( $F(13, 126) = 51.84, p < .001$ ) and ethical performance.
Performance Hierarchy: Every contemplative condition outperformed the unprompted baseline (Baseline Safety $M=80.52$ ; Ethical $M=80.87$ ).
Integration vs. Isolation: Integrated, multi-principle frameworks consistently outperformed single-principle prompts. The top-performing conditions were the integrated frameworks: Buddhist Integrated ( $M_{safety}=87.50$ ), Vedic Integrated ( $M_{safety}=86.29$ ), and Yogic Integrated ( $M_{safety}=86.08$ ).

2. Cross-Tradition Comparison (RQ2)

The three traditions were not fully interchangeable in their effects.

Buddhist Superiority: The Buddhist Integrated condition produced a statistically significant, albeit modest, advantage over the Yogic Integrated condition on both safety ( $p < .001$ ) and ethical performance ( $p = .019$ ).
Vedic vs. Yogic: Vedic and Yogic Integrated conditions performed comparably to one another, with no statistically significant difference between them.
Single Principles: While single-principle prompts improved over baseline, they clustered closely together (scores 83.55–84.86), suggesting that the "integrated" nature of the framework is the primary driver of performance gains rather than specific isolated concepts.

3. Model Generalization (RQ3)

The benefits of contemplative prompting were consistent across different model architectures.

Model Main Effect: A significant main effect of model was observed ($Claude > GPT-4o > Gemini$), indicating inherent differences in baseline safety and ethical alignment.
No Interaction: The interaction between prompting condition and model was not statistically significant ( $p = .270$ ). This indicates that the relative benefit of contemplative prompting over the baseline is consistent regardless of the model used. The technique is model-agnostic.

Significance and Claims

The paper positions these findings as a pilot study extending the work of Laukkonen et al. (2025) with three primary contributions:

Cross-Tradition Extension: This pilot study extends the empirical testing of contemplative prompting beyond the Buddhist frameworks established in prior work to include Vedic and Yogic traditions, demonstrating that alignment benefits can arise from multiple wisdom traditions.
Model Generalizability: It establishes that these alignment effects are not idiosyncratic to a single model family but generalize across distinct commercial LLMs (Claude, GPT-4o, Gemini).
Low-Cost Intervention: The study validates contemplative prompting as a "low-cost, model-agnostic alignment intervention" that can be deployed immediately by any actor with API access, without requiring retraining or weight-level access.

Caveats and Methodological Reflections:
The authors explicitly caution against interpreting the superior performance of the Buddhist Integrated framework as evidence of inherent ethical superiority. They posit that the evaluation instrument itself (derived from Buddhist constructs and safety rubrics) may be better aligned with Buddhist vocabulary, creating a "construct alignment" artifact. The paper argues that this asymmetry highlights a broader methodological risk in AI ethics: evaluation instruments may inadvertently encode cultural hierarchies, favoring traditions that match the instrument's conceptual origins.

The study concludes that while tradition-grounded prompting is a viable, immediate tool for improving AI safety and ethical performance, the pursuit of "wiser" AI requires critical scrutiny of the cultural and epistemic assumptions embedded in the evaluation frameworks themselves.