84000 and AI: Authority, Transmission, and Ethical Concerns

October 2025‍

Executive Summary

This concept paper examines the theoretical and practical implications of using artificial intelligence (AI)—particularly large-language models (LLMs), such as generative pre-trained transformers (GPTs) —to translate Buddhist scripture. The paper first considers “Translation and Transmission,” reflecting on the purpose of translating canonical Buddhist scriptures in historical context. 

Next, this paper reflects on “Reliability”: Since AI translations often produce linguistically fluent output, users—especially those without source language expertise—may struggle to judge the reliability of these translations. With this in mind the paper considers the importance of methodology in determining whether a translation is trustworthy and meaningful.

Thirdly, we consider “Productivity,” assessing possible gains afforded by the use of AI tools and potential disadvantages of prioritizing tool adoption over process.

Fourthly, we discuss issues of “Alignment” and how the incorporation of AI tools may affect the acquisition and maintenance of skills needed to translate and edit canonical Buddhist texts.  

In short, this paper examines both the potential benefits and risks of adopting—or rejecting—machine translation technologies. The resulting policy thus seeks to articulate a balanced and ethically grounded approach where AI assists human translators without replacing them. In this approach, AI tools are leveraged to ensure accuracy and maximize workflow efficiency without undermining the human goals and values that provide the raison d’être for the translation and transmission of Buddhist scriptures. 

In accord with this policy, we have concluded that, at this point, AI should not be used to produce canonical translations but can serve a valuable supporting role in source text comparison, quality assurance, glossary development, and other related tasks. To guide implementation, a policy on the use of AI for translators and editors at 84000 is included as an appendix. This policy describes approved and prohibited use cases, emphasizes scholarly oversight, and affirms 84000’s commitment to quality and transparency in all AI-related decisions.

Introduction

The advent of LLMs has the potential to profoundly reorder the toolbox used in translating canonical Buddhist materials and thus raises fundamental questions about 84000's role and purpose. Yet, as with any new technology, it is important to consider whether the use of AI is going to further our goals and whether AI technology aligns with our priorities as an organization devoted to the preservation, transmission, and sharing of Buddhist teachings. Our evolving policies around the use of AI have been guided by our reflections on the following questions:

• What is the purpose of translating Buddhist scriptures?
• How does integrating AI in the translation process advance or undermine the realization of this purpose?
• How does AI impact the acquisition of skills used in evaluating and understanding the Kangyur and Tengyur?
• How reliably do AI translations convey the meaning and affect of Buddhist scriptures?

In this concept paper, we consider these questions under the broad headers of (1) Translation and Transmission; (2) Reliability; (3) Productivity; and (4) Alignment.

Since the adoption of AI is likely to shape the future of Buddhist education, practice, and transmission itself, the issues considered under these headings are not only technological and economic, but also philosophical and social. We therefore feel these questions are of concern, not just to 84000, but to all Buddhists and, indeed, to everyone interested in the transmission of human cultures and values.

Part I Translation and Transmission: Our Mission and Our Vision

Transmission in Historical Context

Historically, the transmission of Buddhist teachings was, and indeed remains, primarily oral. According to accounts given in the Buddhist scriptures, the Dharma was preserved and passed down through recitation, memorization, and communal chanting for the first few centuries after the Buddha’s passing. Texts were recited by teachers to their students, who repeated the recitations communally and individually in order to commit these texts to memory. This transmission of texts was accompanied by oral explanations of the text that illuminated the text’s meaning and placed it in the wider context of Buddhist doctrine and practice. These orally transmitted lineages were upheld with great care and relied, for their authority, on the direct, personal transmission from teacher to student. 

A few centuries later, following the widespread adoption of writing in the Indian subcontinent in the 3^rd century BCE, Buddhist texts began to be written down. If the use of writing marked a significant shift, it may have been a gradual one, with written texts at first seen only as a useful accessory to the spoken word. But eventually, as Buddhist knowledge and beliefs were increasingly committed to writing, the written scriptures themselves came to be seen as the sacred basis for commentary and teaching, study and practice. Even so, the written text was never considered a sufficient end in itself. Rather, written texts were considered sacred vehicles or mediums through which their contents, the Buddha’s insights, might be transmitted and preserved. 

Proper transmission of the Buddha’s insights involved hearing the text explained by a human authority and subsequent internalization of the teachings through practice. Over time, some of the processes employed in transmission (such as reading transmissions or lung in Tibetan) were ritualized, but the essential basis was, and remains, the thorough, carefully verified communication of an entire way of thinking about, applying in practice, and realizing the Dharma. To the present day, this continues to be the basic model for transmitting the Buddha’s insights and its emphasis on the interpretation of authoritative masters has helped to ensure the sacred texts are not merely read but received as living teachings capable of effecting personal transformation.

This history of Buddhism illustrates that cultural transmission depends not solely on scriptures but also their enactment, in the act of reading or explaining. Transmission is not a product but a process, enacted by human actors with various motivations. Whether any one of us can claim access to the qualities of realization, we can in good conscience lay claim to participating in transmitting the qualities of scripture. In sum, the production of a translated text does not mark the culmination of translation, for the translated text is, at root, a vehicle and support for the human interaction it facilitates.

Translation in Historical Context

The propagation of Buddhism outside the Indian cultural sphere, along maritime and terrestrial trade routes to South-East Asia, China, Japan, Korea, Mongolia, and Tibet, eventually necessitated the translation of the Buddhist texts into languages other than the closely related Indic ones we now know as Sanskrit, Pali, etc. In Tibet, the translation of Buddhist texts was initially carried out under imperial sponsorship and direction, with immense care and attention to authenticity. It involved close collaboration between rigorously trained Tibetan translators and the Indian scholars. These Indian scholars explained not only the words of a given text but also the extensive background knowledge and practical experience necessary for a full understanding and application of its contents. 

In the extraordinary cultural transfer that continued for several centuries, the translation of Buddhist scriptures by teams of Tibetan translators and Indian experts involved the interpretation of meaning, a process informed by philological methodologies and also aesthetics—a feel for pace and pathos—in both the source and target languages. The methodologies adopted by these teams ensured that the Tibetan translations, later collected into the Kangyur and Tengyur, reliably represented the meaning and affect of the Indian texts from which they were translated. 

It is instructive to consider why the Tibetan translations, having crossed the divides of both language and culture, came to be treated as sacred scriptures themselves. The key factors seem to have been their careful transmission by respected scholars; the thorough training of translators; well-established methodologies for proofing, editing, and revision; and the gradual emergence of a sufficiently educated audience. Similar processes were followed in the translation of Buddhist texts into Chinese and other languages.

In short, the trustworthiness of the translation process was critical to the translated scriptures being received as reliable mediums for the Buddhist teachings with the capacity to meaningfully and positively impact the lives of its readers. This emphasis takes on signal importance and stands in distinct contrast to the obscure epistemology or “black box” of current machine translation models.

Translation and Transmission in the Twenty-first Century

Since 84000’s work is an extension of this historical process, we are actively investigating whether and how incorporating AI will affect this work. In this section, we focus on two areas in which AI tools have the potential to undermine, rather than support the translation and transmission of Buddhist ideas and values: skill acquisition and personalization of the Buddhist teachings. 

Cognitive Offloading and Skill Acquisition

Translation requires rigorous engagement with the source material. It fosters deep contemplation, scholarly dialogue, and cross-cultural exchange—elements that are central not just to the production of translations, but to their transmission and meaningful reception in the world. 

Furthermore, the skills and knowledge needed to produce a grounded, etymologically accurate, and idiomatic translation can only be acquired through a long process of active attention and must be maintained through frequent exercise or application of said skills. These skills include: 

• expertise in the grammar and lexicon of Tibetan, Sanskrit, and English; 
• familiarity with the subject matter; 
• appreciation for literary genres, devices, and styles; 
• reference to relevant commentaries; and
• reference to relevant secondary scholarship.

Furthermore, the same skills required in the translation of Buddhist texts are required in both training a good machine translation model and editing a machine translation. One of our primary concerns is that offloading the cognitive work of translation and interpretation to an AI model will lead to a rapid atrophying of the skills in humans necessary to translate, edit, explain, and even understand Buddhist scripture. This concern reflects our understanding of cognitive development, where higher level cognitive skills such as evaluation and analysis are part of a holistic mental operation in which lower-level cognitive skills such as retention and comprehension provide the scaffolding that a person needs to effectively and independently carry out higher order tasks.  

An increasing number of studies suggest that cognitive offloading—decoupling lower and higher order cognitive tasks—leads to the erosion of the offloaded skills. In a paper on “The Impact of Generative AI on Critical Thinking” based on research carried out jointly by Microsoft and Carnegie Mellon, the authors observe, “[A] key irony of automation is that by mechanising routine tasks and leaving exception-handling to the human user, you deprive the user of the routine opportunities to practice their judgement and strengthen their cognitive musculature, leaving them atrophied and unprepared when the exceptions do arise.”

To put it another way, we are concerned that reliance on machine translation models will estrange us from the process of negotiating meaning to such an extent that we can no longer do it independently. As our own abilities to assess Buddhist scriptures decline, we—as translators, editors, and readers—will become increasingly reliant on technologies which cannot yet function as an independent authority. 

Thus, the use of AI tools to translate canonical Buddhist texts raises several questions, including: What do we gain and what do we lose in offloading the cognitive burden of assessing evidence and information? How will future generations acquire the skills necessary to verify a machine translation if reading and comprehension has been offloaded to the machine? 

Through our human-centric translation model, 84000 continues to emphasize the central role of highly trained scholar-practitioners—individuals who are deeply embedded in both the academic and spiritual contexts of the texts. This helps to ensure the integrity of the transmission and ensure that the necessary nuance and doctrinal accuracy are preserved in every translation.

Personalized Dharma: Transformation or Self-Affirmation?

Thus, while AI introduces unprecedented opportunities for engaging with Buddhist scripture in many new ways, it also presents serious challenges to the authority, integrity, and communal grounding of canonical texts. 

Already long before the advent of AI, Buddhism has had a multiplicity of canons and is well accustomed to scriptural plurality. Now, however, in a world where computers can potentially generate countless translations of Buddhist texts—with each tailored to individual preferences—new approaches will be required to reestablish communal standards of authority and ensure the continuity of the scriptural tradition. 

When multiple machine-generated translations begin to emerge, how can practitioners, and even scholars, discern which translations can be trusted? Perhaps more than ever before, questions of credibility, lineage, and scholarly integrity will then become paramount. Without clear validation criteria, it is possible that readers of Buddhist scriptures will begin gravitating toward those translations that are most accessible, rather than those that are most accurate or faithful to the tradition. 

Moreover, AI’s powerful ability to customize scripture to the individual—tailoring the language, message, and interpretation of the canonical source—could over time disrupt the traditional role of scripture as a fixed source of authority. As sacred texts become increasingly malleable and universally accessible through personalized AI interfaces, their authority may gradually diffuse as they will no longer be grounded in a stable, communal canon but rather in a proliferation of individually curated versions. 

As discussed below in “Part II Reliability: Methodology and Epistemology,” Buddhist traditions have long insisted on the necessity of personally verifying scripture. The algorithmic hyper-personalization, however, produces a text grounded in no clearly verifiable methodology or epistemology. A hyper-personalized AI-generated text would, by definition, defer to an individual’s interests, possibly overriding communally-accepted standards of verification in the process. Thus, AI-driven hyper-personalization is likely to produce texts that reinforce our pre-existing beliefs instead of describing a path to liberation while also eschewing the shared reference points that give Buddhist traditions their coherence. 

84000’s Role in Translation and Transmission 

Contemporary Buddhist communities mostly draw from later, indigenous texts, rather than from canonical Buddhist materials, for their monastic curriculums, public teachings, and everyday liturgies. Based on this observation, some scholars have concluded that the canons of Buddhist texts are simply objects of veneration that have little to do with the community’s study and practice of Buddhism. Yet such a conclusion, as well as devaluing veneration in general, overlooks the long-term importance of the canonical texts as the very foundation of how Buddhist knowledge and culture are actually transmitted. The Kangyur may stay untouched on temple shelves for most of the year but its contents are constantly discussed and debated through the medium of later commentaries produced in India and Tibet. No commentary or treatise would be seen as valid unless it was firmly anchored on statements cited from the canon. Thus, while canonical texts may not be the most commonly read titles, without them no other titles exist. It is because the canonical collections permeate and undergird later Buddhist thinking and texts that they are so venerated. 

Curation

Thus, a Buddhist community’s engagement with its canon is shaped by an ever-changing process of curation by the community and its leaders. In response to the perceived needs of the community, certain scriptures are emphasized which, in turn, inform the writing of new texts for new audiences. Throughout history, Buddhism has frequently adapted to new technologies and cultural changes, but the adoption of new interpretations and practices has always been done on the basis of established Buddhist norms, and with a solid understanding of them. New emphases and ways of practice evolve in a dynamic that seeks to meet a community’s immediate needs using the resources of accepted Buddhist teaching. In short, traditionally, Buddhist authorities have acted as curators of the teachings, to borrow a phrase from Donald Lopez.

Partnering 

Our mission at 84000 is to unlock the Tibetan Buddhist canon for all by producing reliable and readable translations of the Kangyur and Tengyur. Our vision, however, of awakening humanity is much larger and requires the cooperation of and partnership with many other stakeholders, including both traditional authorities and contemporary organizations that share our goals and values. Thus, just like the great Indian masters and Tibetan translators before us, our mission and our vision cannot be decoupled. Through co-sponsored events, joint endeavors, institutional collaborations, and expertise exchanges, we hope to use our individual and collective expertise in interpreting, contextualizing, and explaining the canonical materials to facilitate the transmission of Buddhist ideas and values, especially with a modern readership in mind.

Part II Reliability

Methodology and Epistemology

Historically, Buddhists have consistently maintained that the Dharma must be rooted in, and shaped by, conscious intent. In other words, for Buddhist scripture to be trustworthy, it must also be produced by someone who crafted it knowingly based on reliable sources and is willing to assume responsibility for the final product. The great importance placed on reliable provenance is, ironically, seen especially in many medieval apocryphal texts, where the authors have chosen to present their work as the received words of the Buddha.  

From a Buddhist perspective, reliability is contingent upon epistemological validation: the principle that knowledge must arise through valid means such as direct perception, reasoned analysis, or the testimony of a reliable person. A translation, therefore, must not only read fluently but must also be justifiable through recourse to one of these three means. Similarly, in philological traditions, meaning is determined by following established methodologies: assessment of meaning in light of evidence within and external to the immediate text, careful comparison of editions and recensions, recourse to relevant lexical and commentarial authorities, the consistent use of terminology, and so on. 

Unlike human translators, AI systems lack such an intentional engagement with texts and there is no one who accepts professional accountability for their outputs. Since machine translation occurs in an epistemological “black box,” by design AI systems cannot justify their own translations. The use of AI in the translation process thus runs counter to the traditional Buddhist emphasis on the validation of sources of knowledge.  

LLMs do not translate by understanding the meaning of the texts and they follow no discernible methodology apart from Bayesian probability. LLMs operate by predicting the most probable token based on statistical patterns contained in their training data combined with subsequent reinforcement learning to make their output align with human preferences. The accuracy of their algorithmic output thus depends on the degree to which the text in question conforms to the materials of its training. The more unique a word or passage, the less likely its algorithm is to generate an appropriate translation. In other words, their application is greater with generically conforming texts like operation manuals for household appliances than it is with texts marked by high levels of specificity, e.g. texts that make subtle distinctions between like concepts and whose material is largely unprecedented.  Consequently, their ability to render doctrinal nuance with a sensitivity to cultural context is limited. 

Furthermore, LLMs fabricate, hallucinate and vacillate with a frequency belied by the smooth surface and superficial plausibility of their output. Superficially apt but in fact incorrect translations can be very difficult to spot because AI generated translations “seem” to produce linguistically fluent and plausible output. But the fluency often comes at the expense of accuracy as LLMs gloss over the challenging aspects of a text. In this way users will struggle to judge the reliability of an AI translation, and they may not even be aware of the underlying problems with the translation at all. Thus, without rigorous human oversight, AI translations will often omit or distort subtle but essential dimensions of meaning. 

Part III Productivity 

Improved productivity—and, by extension, significant cost savings—is one of the most commonly cited arguments in favor of adopting AI tools. Given that 30% of the Kangyur, and nearly all of the even larger Tengyur have yet to be translated, a significant amount of work remains to be done. And since producing high-quality translations through human effort is very resource intensive, integrating AI in the workflow—especially for generating initial drafts or managing repetitive, time-consuming tasks—could reduce expenses. If AI tools can improve the speed while not compromising the quality of our work, it is important to consider incorporating it into our workflow. Potential gains, however, depend on how well-suited such tools are to the tasks we ask them to do. It is therefore important to consider whether AI tools actually speed up the work of translating ancient Buddhist scriptures and whether automating tasks improves the final product.

Recent studies have indicated that the adoption of generative AI into workflows does not invariably lead to gains in productivity. A study by the Upwork Research Institute, for instance, found that the incorporation of AI tools into workflows actually hampered productivity, increased employee workload, and contributed to burnout.

In an article published in the Harvard Business Review, Ben Waber and Nathanael J. Fast identify two core problems posed by the adoption of LLMs in workflows:

1. their persistent ability to produce convincing falsities and
2. the likely long-term negative effects of using LLMs on employees and internal processes.

84000 translators and editors have independently identified the first of these problems in their own assessments of machine translation tools, noting that the superficial plausibility of AI outputs can create a false sense of security or trust in the results, which makes the detection of errors more difficult. Translators and editors consistently report that an emphasis on speed can shift one’s priorities away from scholarly depth and spiritual integrity. Unlike accuracy or clarity, speed alone is not a compelling reason to adopt a technology when working with sacred texts. 

In fact, if speed becomes the primary goal, it might subtly undermine the perceived value of the work itself. There is a real danger that Dharma scriptures—if translated rapidly and in high volume—may begin to be treated as just another form of cheap commodity: consumed quickly, taken for granted, and no longer regarded with the reverence they deserve. If this shift were to occur, the Dharma would lose not only its sacredness in the eyes of its readers, but also its transformative power. 

For this reason, any use of AI must be justified by its contribution to accuracy, consistency, and understanding—not by how quickly it can deliver results. All decisions must remain anchored in our long-term commitment to preserving the authenticity and living transmission of the Buddha’s words for future generations.

The second problem identified by Waber and Fast relates to job satisfaction. Our translators have often described translation as a profound spiritual practice—an intimate relationship with the text that deepens their own understanding and practice. If translators who have dedicated years to mastering classical Tibetan, Sanskrit, Chinese and Buddhist philosophy find themselves relegated to “prompt engineering” and correcting machine output, they are likely to experience reduced job satisfaction and a diminished sense of meaningful connection to the texts, potentially leading to attrition of qualified scholars from the field. If this dimension is lost, we risk not only the quality of our translations but also the preservation of a living tradition of engaged Buddhist scholarship that has been cultivated over generations.

Furthermore, Waber and Fast report several specific findings that bear closely upon 84000’s work, including the following two:

1. There were fewer productivity gains when “LLMs had poor data coverage or required reasoning that was unlikely to be represented in online text.”
2. “Early research also indicates that when it becomes known that generative AI tools are being used for content generation in interpersonal communication, trust can be significantly reduced.”

Waber and Fast conclude that the incorporation of generative AI is risky for “projects and workflows where the truth matters” while productivity gains are most promising “where generating lots of non-factual ideas quickly is useful.”

Canonical Buddhist materials have relatively “poor data coverage” in comparison to more conventional texts written in contemporary languages. And, given the prevalence of intertextuality and assumed knowledge, the translation of Buddhist scriptures also requires a high degree of reasoning, e.g. inference, of material that is often not represented in the text at hand. For these reasons, at present, machine translations of Buddhist texts require careful editing to bring the translation up to a standard where it is reliably useful to scholars, practitioners, or even general readers. Thus, in assessing possible productivity gains of AI-assisted translation, one must factor the need for careful editing into our project timelines.  

Part IV Alignment

Environmental Concerns

Beyond the above-mentioned concerns regarding the potential loss of necessary knowledge and skill due to cognitive offloading, the use of AI also raises ethical issues from an environmental perspective—an area of deep importance for an organization like 84000. Training and operating large language models consumes vast amounts of energy, which contributes significantly to carbon emissions and environmental degradation. This runs counter to core Buddhist values such as non-harm and care for the environment that we share with other beings.

As we consider our role in the world of Dharma transmission, we should therefore also weigh the environmental consequences of our technological choices. Choosing whether to engage with AI requires not only scholarly discernment, but also a sincere reflection on whether we should strive to minimize harm to sentient beings and the planet itself.

Misaligned AI systems

Besides the immediate harm caused to the environment, there are also an increasing number of AI researchers who warn that AI systems might pose existential risks to humanity.  Several AI industry insiders have raised such concerns, in particular since the launch of ChatGPT in 2022, and recently the founding AI researcher and Nobel Prize laureate Geoffrey Hinton estimated the risk of such planet-wide human extinction caused by AI at 10-20% over the next three decades.  With such potential dangers tied to the technology, a responsible Buddhist organization must not only consider whether to make any use of AI, but also develop strategies for actively mitigating these potential dangers.

Conclusion

Different Applications of AI

Even if AI is not used to produce an initial draft translation, it may be useful at many points during the translation and editing process, from source text comparison, drafting glossary definitions, and checking translation drafts. 

Particularly useful are AI tools that aid the editing process by assisting in the alignment of Tibetan and English texts and identifying parallel passages in other parts of the Tibetan, Sanskrit, Pali, and Chinese canons. It is also possible that AI tools may be developed to highlight omissions, flag inconsistencies, suggest alternative terminology, check glossaries and highlight passages that require a careful editorial check. This would help the editors in their review of the texts by allowing them to focus mainly on the most challenging passages in the translations. However, such tools are currently not available and will need to be developed. Currently, commercial LLMs can perform these functions to some degree, but the lack of quality and instability in their performance makes them highly unsuitable at present.

Using AI as a tool in this way would still allow human scholars to focus on the source material, interpret its meaning, and convey that meaning in translation. Rather than being used as a method to speed up a translation or bypass the subsequent care and attention of an editor, the use of AI should be deployed to allow greater care and attention.

A Principled but Flexible and Evolving Stance

As 84000 considers how to integrate machine translation into its workflow, it is essential to assess both the opportunities and risks this technology presents—not only from a practical or technical perspective, but also in terms of our core mission: the faithful and authentic transmission of the Buddha's words.

As AI-generated translations will likely proliferate—some of them quite polished in appearance—there is also a risk that our more careful but slower work could be overshadowed in public view. Without adequate explanation, readers may not understand the qualitative differences, and could gravitate toward faster, more accessible (but less reliable) alternatives.

In an era where AI technologies are being adopted rapidly, and often uncritically, choosing not to use machine translation to produce initial draft translations sends a strong message about 84000’s commitment to authenticity, care, and responsibility. It distinguishes our work from the growing volume of AI-generated content, much of which lacks transparency, context, and quality assurance. Thus, a human-centric approach to translation could strengthen the trust and confidence that many Buddhist communities, particularly those rooted in traditional lineages, place in our work, and reinforce our identity as caretakers of the authentic Dharma.

Translation of sacred texts is not simply a technical task but a deeply human endeavor that requires lived understanding, doctrinal sensitivity, and ethical responsibility. Therefore, rather than fully embracing or rejecting AI at this point, 84000 may consider a middle way: To avoid using AI to draft translations directly, while embracing its use as a supportive tool. 

This approach preserves the integrity of the translation process while still taking advantage of technological innovations that can improve our workflow and the quality of the published translations. It allows us to remain rooted in tradition while adapting responsibly to new tools. By integrating AI carefully and ethically—as an assistant, not a substitute—we can honor our mission and ensure the Buddha’s words continue to be transmitted with clarity, authenticity, and care for generations to come.

Lastly, 84000 is a community project at heart. With the rapid developments of AI, we will surely benefit from frequently surveying our organization, our readership, and the wider Buddhist community to understand how scholars, practitioners, monastics, and laypeople perceive the integration of AI in Buddhist translations. Receiving community feedback in this way could help us navigate this transition transparently, honoring the concerns and expectations of those we serve.

Appendix A: What is AI Capable of?

Disclaimer: The information below is a snapshot from May 2025. AI develops rapidly and our assessment of its capability from May 2025 will likely soon be outdated. Moreover, we only describe AI’s capability specifically in terms of contributing to the current translation workflow that we have adopted at 84000. Apart from this unique workflow AI can perform a variety of other natural language processing tasks that may be helpful to others, even others at 84000, such as named entity recognition extraction, textual alignment, optical character recognition, etc. However, these functions are not evaluated here. 

AI’s Current Capacity in Relation to Translating the Kangyur and Tengyur 

As of May 2025, AI can draft text summaries and simple glossary definitions. AI can also compare Tibetan source texts and perform basic glossary extraction. AI also has potential to correct human translations by checking for consistency, omissions, and general errors. But technical solutions need to be developed first to make this possible (see below). 

AI is also able to produce draft translations of Kangyur texts, although frequently the translations contain mistakes (both minor and major). These translations can be helpful to a non-specialist in providing a general understanding of a text. Depending on one’s approach to the use of AI, they could also form a basis on which translators can produce a final translation. 

The Quality of Materials Produced by AI 

In the translation workflow at 84000, apart from drafting text summaries, comparing Tibetan source texts, and performing basic glossary extraction, AI is currently unable to perform any task as well as an experienced human scholar.

However, several tasks can be performed by AI at the level of a junior scholar—producing work that is often acceptable, at times even excellent, but which also contains frequent errors and hallucinations.

AI is currently performing especially poorly in terms of writing scholarly introductions and annotations to the translations. When it comes to conducting research on the philosophical, historical, and philological aspects of the texts in the Kangyur with an awareness of existing scholarship, the quality of the work drafted by AI is currently so low, and with frequent hallucinations,  that it has no practical use, not even as a first draft.

Looking ahead, it seems likely that AI's capacity to translate and edit Kangyur texts will continue to improve. However, the extent and timeline of this improvement remain difficult to predict.

Technical Challenges

AI is only able to translate within its context window, a limited amount of text at a time, which is defined differently for each LLM This complicates the workflow and makes it difficult for AI to hold the whole text, and have a comprehensive relationship to the text it translates, which in turn limits its ability to resolve issues that depend on context. However, the current context window will likely increase over time as the technology is further developed.

The interface for working with AI through the commercial LLMs by itself, does not offer an environment in which translations can be edited with a track-changes function. This makes it difficult to review the changes made to the translations through further prompting or revisions. This challenge underscores the rationale for creating workflow tools that include  translation editing with scholarly annotations that can be used to track comments and versions during the entire lifecycle of a translation.

Appendix B: 84000 Policy for Translators and Editors on the Use of AI 

1. Purpose

This policy sets out 84000’s position on the use of artificial intelligence (AI) in its translation, editorial, and research activities. It aims to ensure that AI is used ethically, responsibly, and in full alignment with 84000’s mission: to translate and make accessible the words of the Buddha with clarity, accuracy, and spiritual integrity.

2. Guiding Principles

• Human-led translation remains central. The translation of sacred texts will continue to be undertaken by qualified human scholars. AI will not be used to generate first drafts of canonical texts for publication.
• AI is a tool, not a translator. AI may be used to assist scholars and editors in specific tasks that support—but do not replace—the human-centered translation process.

3. Approved Use Cases for AI

AI may be used at 84000 in the following contexts:

• Comparing variant editions of source texts
• Generating draft summaries of translations (including for introductions), term definitions, and metadata for public outreach or internal research
• Detecting errors, inconsistencies, or omissions in human translations
• Developing translation memory tools
• Assisting in the alignment of Tibetan and English texts
• Helping design or test digital tools for scholarly editing and collaboration.

4. Prohibited Use Cases for AI

AI should not be used to:

• Produce first-draft translations of canonical texts intended for publication
• Independently translate texts (even if the texts are not being published) without direct scholarly guidance and review
• Make decisions, as the primary resource, between divergent possible interpretations in a translation.

5.  Approved Tools

Only those AI tools and APIs that have been approved by 84000 can be used. A list of the approved tools can be found here.

The list of approved tools is maintained by the Translation Team’s AI working group. Suggestions for new tools can be submitted to the working group via [email protected].

6. Scholarly Oversight and Transparency

All AI-supported inputs and outputs must be reviewed by qualified human translators and editors.

AI involvement must be documented transparently in internal workflows and, where relevant, disclosed in public-facing materials, by submitting the AI Use Disclosure Form at the completion of a translation or editorial project.

Scholarly prompts (inputs) used to initiate AI-supported outputs must be shared and must include the full prompt, the model and tool being used, the date of the prompt and the resulting output. This information should be provided via the AI Prompt Documentation Template.

7. Training and Capacity Building

Staff, in-house translators, and editors using AI tools will receive appropriate training to ensure ethical and effective use. Ongoing assessments of AI tools will be conducted to evaluate their impact and relevance to 84000’s evolving needs.

8. Ethical and Spiritual Considerations

84000 recognizes that Buddhist texts are not merely documents of linguistic, historical, or cultural interest but carriers of a spiritual lineage. All AI use will be evaluated in light of its impact on the authenticity and sanctity of the texts.

9. Community Engagement

84000 commits to engaging with our readership—scholars, practitioners, monastics, and laypeople—on matters related to AI use. Community input will inform future developments and refinements of this policy.

9. Review and Updates

This policy will be reviewed annually, or as needed, to respond to technological advances, community feedback, and developments in the field of AI and Buddhist studies.