Collaboration for the Preservation of Dharma
Notes from the Tibetan Translation & AI Workshop + Upcoming Community Dialogue
Hi everyone. This month we’re sharing notes from the Tibetan Translation and AI workshop that we co-organized in Padua, Italy. Before that, a quick flag: we’ve got a virtual community event coming up in early June. And to the new subscribers who joined us this last month – welcome, we’re glad you’re here!
Join the Conversation: Community Dialogue with Dolma Gunther (Khyentse Vision Project) | June 2nd 3:30-4:45pm PT
In April, we featured a guest post by Dolma Gunther (founder and creative director of the Khyentse Vision Project) titled “Magic or Mayhem? - Translating Dharma in the Age of Intelligent Machines.” The post sparked a lot of interest, so we’re making more space for a community conversation to explore together some of the themes and questions that it raises. In what ways can AI contribute to the translation, study, and practice of Buddhism? What does wise engagement with AI look like?
Join us for a virtual gathering on June 2nd 3:30-4:45pm PT to hear more from Dolma and get into conversation with each other on these questions and more. Register to join here!
Update from Padua
On April 16-19, we co-hosted the Tibetan Translation and AI workshop, held in Padua, Italy. The conversation was facilitated by Mind & Life Europe (MLE), Italian Buddhist Union Research Center (UBI CS), and us (Buddhism & AI Initiative), with generous financial support from Khyentse Foundation.
About 25 people gathered in person and another 10 joined remotely, all of whom are contributors to the fields of Tibetan translation, Buddhist studies, and related AI/NLP efforts.
Over the course of four days, we discussed a number of existential questions related to the future of these fields:
Why do we still need translators, scholars, and teachers? What is the role of human beings in the field of Buddhist studies in the age of AI?
What challenges does AI pose to this ecosystem and to Buddhism overall?
What boundaries should be set around the types of beings who can genuinely transmit Dharma? Could AI meet certain qualifications for transmitting authentic Dharma? What would those qualifications be?
What opportunities exist with AI? How can Buddhist teachings flourish in the age of AI?
None of these questions have easy answers, and from a year of networking we’ve seen that Buddhists’ first reactions for how we ought to respond to AI vary greatly, from “we should stay far away from this technology” to “we need to build Wise AI.”
But at least among this group, what we found is that with the opportunity to dialogue, in person, across multiple days–what looks on the surface like plausible disagreement can be synthesized into a collaborative path forward.
Should Buddhists share data with Big Tech?
Consider the case of whether or not Buddhist communities should collaborate with or eschew Big Tech and AI Labs, particularly as it relates to sharing Buddhist data.
One of the main reasons to share data with Big Tech is that top AI models are quickly becoming the primary way that newcomers learn about Buddhism and meditation. Recognizing this, Buddhists may want to intervene to help ensure that AI Lab’s models are built with high-quality data and respond to questions wisely. In addition, the Buddhist community doesn’t have the resources to compete with these labs, and the fact of the matter is that the companies have probably already vacuumed up all the Buddhist data they can–meaning it’s possible the “moat” for Buddhist independence is quite small.
But the case against Big Tech is also strong: Big Tech’s motivations of profit (and deep embeddedness in the military-industrial complex) are questionable at best and largely antithetical to Buddhist values. If one takes the risk of advanced AI seriously, then these labs are also ushering in danger for all of society. And for Buddhism’s continuity, it’s critical for Buddhist communities to have sovereignty over their data, tools, and technology; and Big Tech could take away AI models, change pricing, or “pull the rug” in any number of ways.
One could reasonably take either side of this debate. And indeed, when we ran a survey among workshop participants on whether the ecosystem should aim to collaborate with Big Tech by sharing data, the results showed a range of viewpoints:
But once we actually began discussing, it was clear that the seeming disagreement wasn’t a disagreement over facts (i.e. nobody denied the risks; nobody denied the opportunities) – it was a different weighting of the tradeoffs.
By the end of the dialogue the group managed to find a synthesis between the viewpoints: it seems likely that Buddhists should share data to the degree that it can help support the Dharma (from AI models being many people’s first encounter with Buddhism, to models helping with tasks in the Buddhist ecosystem like philology and translation). And at the same time, Buddhists should proceed with extreme caution in their relationships to Big Tech–via backup plans with open-source technology, building internal competency, holding out some data as private, and never forgetting the human element at the center of this work (e.g. always linking projects back to the lineage of transmission, and never losing sight that the purpose of technology should also be liberative). With this more nuanced view, individuals that want “nothing to do with AI” are still contributing to the overall picture–by protecting the most human elements of study and practice; and those that are “getting their hands dirty with AI” are also contributing–by ensuring that this new container of knowledge supports wisdom.
This pattern of apparent disagreement dissolving into shared tradeoff-weighting showed up throughout the workshop, and produced a set of concrete projects that the group is committed to carrying forward.
The whole thing, quickly:
For those who are interested in reading a full debrief on the workshop, a 20-page, anonymized summary can be found here: Summary of the Padua Workshop - Anonymized.
For those that want the short version, read on!
The workshop wrestled with a number of practical concerns that the Buddhist Translation/Studies/Tech ecosystem is grappling with. Chief among these is the need for greater collaboration between organizations to share data and evaluate how good AI actually is at critical tasks: translation, Buddhist studies research, Buddhist factual recall, and similar. The speed at which AI is moving is difficult to wrap one’s head around, not to mention keep up with technically–and so these steps towards evaluation and benchmarking, as well as data sharing, are built on the premise that Buddhists need to work together to navigate these challenges.
The need for collaboration arises from a small community looking at a juggernaut–the hundreds of billions of dollars flowing into frontier AI–and recognising it can’t keep pace without teaming up together. In addition to collaborating, it was also identified that the very limited number of individuals with technical expertise in this space need to be actively supported. Many technical individuals working within the ecosystem have faced funding challenges that have required shifts in plans or career paths to less-aligned but more stable opportunities. And funding instability also prevents forward-thinking work necessary for scaling, e.g., individuals have not been incentivized to create new organizations, recruit ambitiously, or think about long-term technical goals.
The problem above is particularly challenging for the Tibetan diaspora community in India. Many Tibetans–from Geshes and Khenpos to programmers and data annotators–consider immigration their highest priority and are moving to other countries to pursue careers in the gig economy. This skill attrition is one-way: once someone has left India it’s difficult to bring them back. As it stands, a significant amount of work in the NLP/tech space around Tibetan Buddhism is done by Westerners, not Tibetans, and so the conference participants really underscored the need to support and protect projects driven by Tibetans.
Discussion continued across many topics, but eventually began to cohere into a set of near-term projects the community was interested in bringing forward, including:
Benchmarking: Participants will coordinate a webinar on how to do benchmarking well, which will serve as the basis for future coordination on benchmarking/evaluation projects aimed at seeing which parts of Buddhist translation (and some Buddhist studies work) AI can and cannot do well.
Data Gathering: Participants will kick off an initial effort on data mapping to identify Buddhist datasets that are already “shared by the ecosystem” and find a central place to list them. This work will then extend into a data gathering effort to bring more datasets into shared access.
Documentation of Translation Tool Workflows: Participants are documenting a map of the translation workflow and how various tools/APIs fit together, to get a picture of modularity (how well different orgs’ technologies can fit together).
Trainings for Buddhist Studies scholars and students:There is a summer school opportunity among other training options, to teach AI skills to Buddhist studies experts and students. Participants are meeting to determine next steps.
Overall, the workshop was a great success, bringing together many long-time collaborators who haven’t seen each other in person for a few years. Coming together in this way helped to build consensus and revealed that certain disagreements were more surface level than they appeared, and was critical for building a shared plan for how to move the ecosystem forward. All of this reaffirmed our belief that alignment of goals is a relational matter, not just a technical one.
If you’re interested in being more involved with any of these projects, reach out to hello@engagedbuddhists.ai and if there’s a fit, we can help connect you with these broader ecosystem efforts.
I once took an “Introduction to Classical Tibetan” course which culminated in the “attempted translation” of various Buddhist materials. That exercise taught me that texts aren’t translated; they instead are interpreted. (I suppose the same applies to any foreign language.) My sense is you could give ten human translators the identical Sanskrit/Pali/Tibetan text and you would get ten different “translations.” I suspect that people would be shocked the extent to which those old materials are mediated and interpreted by the human mind. Could AI do material damage to the translation endeavor? I am not so sure. A case could be made that perhaps AI could provide a “translator-neutral” interpretation. I certainly wouldn’t dismiss AI out of hand for such roles.