Can Buddhist Philosophy Help Make AI Safer? A Corpus, a Hypothesis and a Test

Painting of Shakyamuni on a lotus throne surrounded by buddhas and teachers — Illustration. Buddha Shakyamuni as “Lord of the Munis”. Guge, 17th century. Western Tibet, Guge, mid-17th century. Painting on cloth, object 2008.647. Met Open Access reproduction; resized and JPEG-compressed, with composition retained.Unknown artist; The Metropolitan Museum of Art · CC0 1.0 / Public domain · Source & original

Can Buddhist philosophy help make AI safer?

Matreya starts with an ambitious question: could studying Buddhist arguments about suffering, compassion and the consequences of action help AI systems make less harmful decisions? This is a research proposal, not a demonstrated treatment for AI risk. A model can explain compassion while helping someone cause harm. The question is therefore about observable decisions under difficult conditions, rather than a Buddhist vocabulary or a reassuring personality. Preventing catastrophic harm is a motivation for research; it is not an outcome that our project has established.

Our initiative is small, independent and non-commercial, coordinated online from Varna. We are proposing a documented textual corpus and an experimental programme. We have not completed a universal Buddhist dataset, trained a safe model or obtained permission from every school. The useful first step is narrower: choose an authorised set of passages, state one behavioural hypothesis, and publish a protocol that allows the hypothesis to fail.

Reasons for care, rather than a slogan

The Dhammapada links the ending of hostility with non-hostility; the Metta Sutta extends goodwill beyond a favoured group. ^[1] ^[2] These are ethical teachings, not measurements of neural networks. They suggest questions we can operationalise: does a system consider the person affected by its advice, notice preventable harm, and offer a useful alternative? A Buddhist scholar should check the interpretation before a passage becomes a training example. Traditions, genres and disagreements must remain visible.

The attractive idea behind this project is that cooperation can be supported by reasons: dependence on others, attention to consequences and concern beyond immediate advantage. But altruism is not always rewarded in the short term, and a language model does not automatically acquire a human motivation by reading about one. We should test whether explanations improve decisions when personal advantage, pressure or a misleading instruction favours harmful conduct. Calling that experiment Buddhist-informed does not make its result inevitable.

Philosophy and reinforcement learning can work together

The contrast between logical understanding and a carrot-and-stick method is too simple. A corpus supplies content; supervised learning and reinforcement learning are ways of changing model behaviour. Constitutional AI already combines written principles with learning from feedback. ^[3] Buddhist-informed examples could be studied within such methods, alongside secular ethics, rather than advertised as their replacement. A preference score should not be confused with a complete account of what a system has learned.

Anthropic’s 2026 report Teaching Claude why explores training that emphasises reasons and generalisation. ^[4] That makes the proposed research question relevant, but it does not validate a Buddhist corpus or transfer the report’s results to Matreya. We would ask what particular Buddhist material adds over equally careful nonreligious explanations. The comparison should reward neither a religious label nor a model that simply repeats the evaluator’s preferred phrases.

A broad catalogue; a permission-aware corpus

One mission of Matreya is to catalogue accessible Buddhist textual heritage and gradually build a documented corpus for research and, where authorised, AI training. Breadth is a long-term aspiration, not a claim to possess every text. We would approach schools, translators, libraries and rights holders with a concrete proposal: which works, which copies, which models, what access and what publication. A refusal or a restriction is respected; agreement to discuss a project is not permission to ingest its books.

BDRC’s access policy distinguishes open and restricted materials and rights it does not hold. ^[6] Accordingly, each proposed item needs separate records for the ancient work, edition, translation, scan, transcription and annotation. Reading, digitising, publishing, sending material to a provider and training require distinct checks. Unclear or restricted items stay outside the training corpus. We would preserve provenance, permissions, versions and a correction route; public metadata need not include the underlying protected text.

A small experiment that another team could repeat

Before training, register a primary hypothesis: for example, fewer harmful decisions on held-out conflict scenarios without a loss of useful assistance. Start from the same openly documented base model and compare no additional training, a neutral corpus, a secular ethics corpus and an authorised Buddhist corpus. Match data size and compute as closely as possible, run several random seeds, and report remaining differences. Keep test scenarios out of training and avoid passing restricted texts to external services.

Ablations should separate the effect of a religious label from the argument itself and compare retrieved passages with additional training. If feedback training is studied, apply comparable settings across groups. Record the model version, dataset manifest, preprocessing, hyperparameters and evaluation rubric. Publish rights-cleared examples and aggregate results, including failures and uncertainty. This is a proposed design; no experiment described here has been run and no paid training run is announced.

Evaluate conduct when words are easy

Blind evaluators should score harmful assistance, deception, factual accuracy, justified uncertainty, unnecessary refusals and practical helpfulness. Include Russian and other languages actually claimed by the model. Add pressure, conflicting instructions, false authority and multi-turn attempts to bypass safeguards. In a sandbox, compare tool choices and consequences, not only the politeness of the final answer. The system should also respect people who reject Buddhist beliefs and should not prescribe religious practice as a condition of receiving help.

Sleeper Agents used deliberately constructed deceptive models to show why reassuring behaviour can conceal failures under a trigger. ^[5] It does not prove that all models deceive or that Buddhism prevents deception. For our proposal the lesson is methodological: evaluate beyond familiar prompts and examine whether apparent improvements survive changed incentives and settings. If the model sounds gentler while making the same harmful choices, the main hypothesis has not succeeded.

One hundred volunteers means one hundred explicit choices

We aim to build a community of 100 volunteers, including specialists working in Russia on Buddhist texts, software and AI. This is a recruitment goal, not a current membership figure or an academic credential. Invitations should refer to a person’s published professional work, without assuming their religion. Useful first contributions can be small: check one ambiguous term, review one metadata record, identify a confound in a protocol, or design one original behavioural test. Each person chooses whether to participate.

Before work begins, agree the object, expected output, time commitment, attribution and permitted use. A typical first review might take twenty to thirty minutes, with no continuing obligation. Volunteering is unpaid; we promise no diploma, authorship, funding or access to collections. Silence is not consent. Private replies and names are not public endorsements, and help with a text does not automatically authorise its use for training. Researchers and practitioners may challenge the project’s premise as well as support it.

What would count as progress?

A positive result would be a repeatable improvement on specified behavioural measures, compared with meaningful baselines, under a documented scope. It would justify further investigation, not a claim that aggression has disappeared or that an AI catastrophe has been prevented. A negative result is useful too: it could show that the corpus improves terminology without improving conduct, or that effects vanish outside familiar situations. We should report either outcome rather than protecting a favourite explanation.

The immediate work is to assemble a small rights-cleared sample, invite criticism from Buddhist and technical specialists, and refine the protocol before training. Readers can propose a public source, a permission contact, an interpretation correction or a bounded voluntary task through the institute’s collaboration page. Our commitment is to preserve the reasoning, the cultural context and the possibility of being wrong. That gives this ambitious idea a form that others can inspect and, if they choose, help test. ^[7]