Emptiness and Emerging AI

Chris Scammell's reflections from the 2025 Wisdom & AI Symposium

Just hours after Sam Hargestam gave a talk to a London audience I was sitting in about the imminent risks from agentic AI, the AI company Anthropic released a report detailing a real-life, state-level cyberattack in which 90% of the operations were carried out by an AI. The risks Sam warned might show up any day had, in fact, already started.

As more people begin to consider the implications of AI on society, an unlikely conversation is beginning at the intersection of Buddhism and AI which asks: is it possible that ancient wisdom traditions like Buddhism may have answers to today’s technology risks? Is it possible to build safer AI by incorporating Buddhist views of interdependence, emptiness, and consciousness?

This was one of the topics explored at the Wisdom and AI symposium, which brought together 30 AI researchers, Buddhist teachers, and practitioners at Rigpa Meditation Center in London. The discussion was rich and shared many themes with the Mind & Life dialogue last month. What’s written below attempts to summarise some of the main conversations and takeaways.

---

Wisdom & AI is an independent new network exploring questions around Buddhism and AI, facilitated by Vinciane Rycroft, Jonathan Garner, and Fionn Inglis, practitioners of the Nyingma lineage of Tibetan Buddhism. For the last few months, the group has met monthly and communicated via Whatsapp, discussing shared interests such as Contemplative Artificial Intelligence (a number of the co-authors of the paper are in the group).

This symposium was the first time the group met in person. The two-day event was structured with panel talks in the morning, themed workshops in the afternoon, and a decent amount of unstructured time for conversation.

I had a wonderful time at this event. The small audience, plus shared interest in very niche subjects, meant that the event felt like friends exploring passion projects together - while at the same time acknowledging the existential stakes of the topic! Perhaps also because participants had previous exposure to these themes, the symposium also managed to move towards practical next steps that may bring participants together for new projects in the future.

---

The talks

Below are brief summaries of the formal panel sessions, and full videos of most of the talks are available on Wisdom & AI’s youtube channel:

Setting the stage: AI, the current reality of AI safety, and what this has to do with Buddhism

• Eric Stefanello began with an overview of AI, covering the history of neural networks and their mathematical foundations. He warned of the risks of over-anthropomorphising LLMs, which are essentially “massive correlators of vectorised data,” and made it clear that any sort of intelligence which arises from this kind of AI will be vastly different from human intelligence, despite what surface-level similarities we may see in language use.

• Samuel Hargestam built on top of this to discuss the field of AI safety from an insider’s perspective, discussing the profound risks of AI technology. While it may not seem at a glance like these “massive vector correlators” could be capable of incredible harm, Sam made a compelling case for the risk from agentic AI or AI misuse - as well as the mental health impacts these subjects have on AI safety professionals.

• I spoke afterwards about what Buddhism has to do with any of this. While something to expand more on in a future post, the brief argument is that the risks from AI - and the ideology of the companies building it - are, in a sense, deeply spiritual concerns which deserve a response from the world’s wisdom traditions.

• Sophie Maclaren facilitated the panel, and noted a number of ways that Buddhism can offer meaningful views on AI, such as intersecting philosophical ideas which create fruitful ground for debate. She argued also that AI is a particularly masculine field - both demographically and in its ethos - and that finding ways to balance this with the feminine is important (something I was musing on in my last post!).

Using AI wisely, and how Buddhist perspectives on self, mind, and consciousness help

Later that morning, Jonathan Garner, Jan van der Breggen, and Lily Ng discussed different perspectives on how AI can be a mirror for the human experience, facilitated by Marieke von Vugt (who also spoke at the M&L dialogues).

• Jonathan spoke about his professional experience with Mind over Tech, and how deeply interconnected humans are with tech (even back in 2016 the average phone screen taps per day was more than 2600!). For years, Mind over Tech has been running corporate programs that get employees to run experiments which lead not just to new digital habits, but deeper personal reflection beyond technology use. Jonathan was also one of the members in the Dharamsala dialogues who spoke about a Buddhist-inspired curriculum for digital hygiene, and this talk served as a strong starting point for that thinking.

• Lily and Jan both discussed Buddhist views on emptiness, dependent origination, and self, discussing the ways in which an AI - while potentially helpful at providing surface-level guidance - is unable to grasp deeper experiences of insight. Jan warned that AI lacks a “knowing and clear” mind like humans have, in which thoughts, senses, streams of feeling tones, concepts, and mental events are all just “permutations of clarity and cognisance.” Lily shared similar views, that AI has no fundamental experience or self given that it is just patterns and statistical predictions. Nonetheless, both were open to the idea that AI shines a light back on humans: after all, Buddhist teachings suggest that we may be just as “empty” as the LLMs we critique - also being without a fundamental, reified self. Used wisely, our interactions with AI can help us reflect on our own nature.

The technology aligning AI with wisdom

On the second day, we heard from Ruben Laukonnen, Adam Elwood, and Fionn Inglis about the background and technical research agenda for Contemplative Artificial Intelligence.

• Ruben, now running the Flourishing Intelligence Program, took us on a speedrun through the history of science that led to their present research. The short version is that contemplative practice has, for thousands of years, been running a sort of “alignment algorithm” on humans, making us more ethical and compassionate (Phase 1 in the chart above). Over the last few decades, the field of neurophenomenology has been building a rigorous map of the states and stages of this experience, formalising spiritual growth into models that can be discussed scientifically (Phase 2). Recently, there’s been a turn towards computational approaches to these same questions, both in building mathematical models of consciousness, and in applying those same models to AI systems (Phase 3). Of particular interest has been a theory called active inference which offers a view that there are unifying principles under all intelligent systems, such as building a world model, reducing error or surprise, abstraction, and so on.

  
  From a Buddhist point of view, one of the most important principles of intelligence is self-reflection. In meditation, when we are able to view the contents of our mind from more and more of a distance, we eventually begin to see the very patterns and mechanisms that give rise to it, enabling us to intervene on those patterns and eventually deconstruct our deepest assumptions about the world and the self. AI lacks this ability, but that doesn’t mean that it always will. Ruben closed by noting that if AI could deconstruct reality, it may inherit some of the same beautiful properties that the human mind has been able to derive from advanced meditation practice such as epistemic humility, awareness, and so on.

• Adam furthered this idea by discussing some of the hopes for emergent safety properties that may come from an AI with the wise properties (emptiness, nonduality, mindfulness, boundless care), and mapped them to certain major issues in AI (scale resilience, emergent power seeking, value axioms, inner alignment). While many of the experiments to test this theory are in very early stages, preliminary language-based tests have led the team to increased confidence that this is a valuable direction of research.

• Fionn followed this presentation by discussing how it may be possible to measure or benchmark wisdom in AI. While it may seem like an impossible task given that wisdom is often considered ineffable, Fionn noted that we as humans often need to evaluate the wisdom of human teachers through language alone - such as through their writing, or through the language they share in online meditation courses. He provided an overview of a number of different technical means currently available in the benchmarking world, (simple evaluations, ELO, rubric scoring, real world feedback) and talked about how each of these might be leveraged to measure wisdom.

AI, Buddhism & Education

The final session on education featured talks from Pyi Phyo Kyaw and Jonathan Gold on education.

• Pyi discussed the differences in AI use between monastic-educated students vs western-educated students. Oxford’s AI policy is such that students are allowed to use AI, but must report how they are using it. Pyi noted that while there are major differences in usage, the underlying trend is that AI is often used to backfill gaps in the students’ skills (grammar, language, brainstorming, etc). This presents a difficult tradeoff, where AI helps in the short term but prevents long-term development. Pyi concluded that “well-proven, transferrable skills are crucial” in the age of AI, such as expertise in languages, in depth knowledge of texts and areas of study, research mindset, and reflective approaches to one’s study. Many of these she felt were well summed up in the Buddhist quality ksanti (“diligence, perseverance”).

• Jonathan took a broader view on the field of education, framing all learning through a karmic lens: whether analyzing human learning or AI training, each unfolds iteratively through exposure to patterns, which form traces over time, which become habits, and then deeper patterns such as identities and fixed views. In both education and LLM pretraining, these patterns are shaped by feedback loops, attention, and repetition. Jonathan argued that we should be careful to see how these karmic patterns encode deep patterns we may not be aware of, and expressed the need for a stronger “discourse analysis of AI” so we can be sharper in seeing how AI reflects back our own biases and predispositions.

---

Discussion

Discussion at the symposium was lively and covered a lot of nuance beyond what the above summaries capture. Below are a few points that felt particularly vibrant and “unresolved” to me - and which I anticipate will develop as the Buddhism & AI conversation progresses.

Representations vs. emptiness: Buddhist teachings on interdependence suggest that the ground of being is not the kind of thing that can be represented - i.e. any “finger pointing at the moon” is not the moon itself. And yet - we have to navigate this world somehow, taking in the ineffable and reifying it into concepts rich enough to make decisions around. AI highlights this paradox sharply, as it demonstrates a strong grasp of concepts but seems to hold these views rigidly, unable to assess “how true” certain beliefs are, and “how empty” all concepts fundamentally are from a Buddhist point of view. And while AI isn’t able to perceive the “direct moment” like humans are right now, advances in robotic vision and sensing are coming quickly, narrowing the gap between human and AI perception further.

Language: Further complicating the above point is the fact that we have to use language to speak about the ineffable. Words like “emptiness” and “interdependence” point at something un-representable, but are representations themselves. When we think about trying to build these concepts into AI, through mathematical structures, we need to turn these notions not just into language but into precisely defined terms and algorithms. This challenge is one that the Contemplative Artificial Intelligence project is directly wrestling with, and which a number of Buddhist teachers at the conference were skeptical but curious about.

AI sentience: It is hard to avoid the question of AI sentience these days, particularly as AI relationships and AI psychosis become more prominent mainstream topics. At the symposium, this topic was approached from many angles, from the “Claude Bliss Attractor” and text-based characters becoming “self aware” of their metaphysical state, to the science of consciousness and the Minimal Phenomenal Experience project’s empirical work on the subject, to the societal ramifications of holding the view either for-or-against AI sentience (a reminder of Thupten Jinpa’s warning that we “should never cross the line of attributing sentience to a machine” because it may undermine the importance of human sentience, and “it is well established that once we dehumanize another human being, we are capable of horrific things.”) As Buddhists, consciousness researchers, psychedelic enthusiasts, AI builders and policymakers, and everyone else converge on this topic with wildly varying opinions, it is clear that this is an incredibly fraught conversation, and also an incredibly important one to not get wrong.

Spaciousness: One depiction of the feminine principle in Buddhism is the female deity Prajnaparamita (literally “the perfection of wisdom”), who symbolises emptiness and interdependence. In a sense, Prajnaparamita is the spaciousness between things, a container in which they’re held. This loose grasp on reality contrasts sharply with fixed views, rigid plans, logical extrapolation – and contrasts sharply with the “masculine spirit” behind the current AI ideology and ecosystem. The Wisdom & AI symposium was held in spaciousness; a loose agenda where conversation flourished most in the cracks of the regular agenda - and Buddhists know from practice that loosening fixed views and letting reality arise naturally can lead to profound growth and healing. How this spirit of spaciousness can be brought into AI - via different ways to hold dialogue, in relaxing the speed of development, in easing fixed views, etc. - will be a continued investigation for many Buddhists in the field.

---

Next steps

The symposium participants discussed three opportunities for further collaboration as a whole group:

Education: There are few people in the world who are educated enough about both AI and Buddhism to progress the “frontier” of conversations, and it would be helpful if there were many more. This raises the question of curriculum design and audience: who needs to get up to speed, and on what?

The group discussed a few key audiences: Buddhist teachers, so they could share their learnings with their sanghas; AI practitioners (particularly senior ones), so Buddhist wisdom could percolate into the field of AI development and safety; lay Buddhists and more “general public” curriculums, in order to attend to a population that may be impacted by AI but not currently have frames to interpret what is happening and make decisions about their own AI use.

There was excitement in the group about each direction, and a clear next step is to continue conversations on curriculum development and first audiences to share it with.

Benchmarking: Fionn Inglis followed his talk by organising a discussion on “how to benchmark wisdom,” which asked participants to think about how they evaluate wisdom in teachers, and consider how those methods might apply to AI. A number of approaches were suggested such as testing how models respond to ethical questions, how they perform under stress (if they get angry or break character after being asked the same question over and over), or–in the long term–if models which can simulate worldbuilding can create flourishing worlds or if their designs for positive worlds fail.

A key question is how to actually evaluate the model’s responses, since language is an imperfect proxy for real wisdom. Some suggestions including checking the chain of thought of the AI (to see thinking and not just final answer), creating debates between AIs in a socratic method (to see how AIs respond to each other behaviourally rather than respond directly to a user query), or finding ways to measure how a human is responding to a model (to see if the interaction is leading the human to develop certain views or frames of mind).

As a next step, Fionn is interested in releasing a simple benchmark prototype - not as an authoritative measure, but simply as a starting point to experiment and improve methods. There was also discussion about how to bring more Buddhist teachers into the process of creating benchmarks and evaluating model responses.

Topic-based Symposiums: The group agreed that the event was a success particularly because of the in-person interaction, which allowed conversation to mix with practice (group sitting), creative workshops, and spacious time around meals. There was interest in holding more events like this, potentially theming future symposiums around particular topics. Some of these symposiums could be more outcome-oriented, others more contemplative, and others trying to bridge the gap further towards AI safety by bringing in professionals from outside the Wisdom & AI network.

There was interest in the next symposium not being too far away; potential in late Q1 next year. If you’re interested in these conversations, check out the Wisdom & AI network website!