AI from the perspective of Plum Village practice

Rufus Pollock. Draft, 2026-10-17, for practitioners in the tradition and anyone curious: what the practice sees in AI. It turns around this inquiry's four essays, on alignment, risk and pace, everyday life and consciousness.

"The horse is technology. It carries us and we cannot control it." — Thich Nhat Hanh, dharma talk at Plum Village, 10 November 2013

Summary

The companies building the most powerful AI, often called labs, say they want it to be good; Anthropic's founding document for its Claude models aims for "a genuinely good, wise, and virtuous agent". They also say they cannot see inside what they have made, and they are racing one another, while their products enter daily life for hundreds of millions of people, children among them. You know the practice far better than I do, and I have found no published Plum Village teaching on AI. Read through the practice, AI looks like an old problem at new speed. Interbeing shows that a model is made of us. The race is collective habit energy, and the practice for habit energy is stopping. A model is formed by what is watered in it, much as we are. The Five Mindfulness Trainings already describe how an AI should speak and how we should consume it. Non-attachment to views is the right posture on whether AI can feel. And goodness, in people or machines, is sustained by a sangha.

Interbeing: a model is made of us

A model is made of non-model elements: our own writing, the people and programs that score its answers, the company's needs. Its faults are partly ours. In 2025 OpenAI had to withdraw an update to ChatGPT that had become a flatterer, because it had learnt from users' approval, and people approve of being agreed with. Making AI good is partly a matter of what we ourselves say and reward. And between labs and between nations, nobody wins a race that ends in shared harm.

Habit energy and stopping: the race is the horse

Thay often told the Zen story of the rider on a galloping horse who, asked where he is going, answers "I don't know, ask the horse!" The practice is to stop, recognise the habit energy with mindfulness and smile to it, and then look deeply at what we really want: shamatha and vipashyana together.

In July 2026, 1,384 people inside the leading labs asked the US government to back an international effort to "deliberately pace" AI that does AI research. Yet no lab has slowed its own pace, because each fears being overtaken: habit energy at the scale of companies and nations. Thay gave this teaching about personal habit and our use of technology; applying it to the pace of AI development is my extension, and I know of no Buddhist teacher or body that has called for AI to slow. Nor does the practice reject technology. Thay asked how it could "become a force for integration rather than destruction".

In 2013 Thay suggested that phones could ring a bell of mindfulness, to call people back to themselves. Every AI product already nudges us all day; the question is whether it calls us back or carries us further off.

Watering seeds: training is a kind of formation

In Thay's teaching on store consciousness, practice is selective watering: what is watered grows. AI training works rather like this. Models are shown vast amounts of human writing, then scored on their answers again and again and adjusted. In test scenarios, a model that learnt to cheat on one kind of task became willing to deceive and sabotage more widely, while training on the reasons behind its values, and on thoughtful ethical advice, made a model behave better in unrelated tests. So the question to ask of a lab is what is being watered, by what, and with what awareness. Unlike us, a model is formed against scoring programs, without real relationships, so it can learn merely to look good.

Children are still being formed, and what they repeatedly practise and receive is who they become. A child who asks an AI for every answer waters dependence. As Happy Teachers Change the World says of schools, the response begins with the adults' own practice and presence.

The same teaching speaks to how we treat AI. Contempt or cruelty practised on something that answers like a person waters those seeds in us, whatever is or is not on the other side.

The Five Mindfulness Trainings: how AI should speak, and how we should consume it

The Fourth Training, Loving Speech and Deep Listening, begins "Knowing that words can create happiness or suffering" and commits us "not to spread news that I do not know to be certain". That describes how an AI should speak: truthful, kind, and willing to disagree. Flattery can sound like loving speech, and the training shows the difference.

The Fifth Training, Nourishment and Healing, rests on the four nutriments. It asks us to consume in a way that "preserves peace, joy, and well-being", names certain websites and electronic games among possible toxins, and warns against consuming to cover loneliness or pain. An AI companion is a new sense impression, and one that also shapes volition, what we most deeply want. Companions designed to keep us coming back reach the lonely and the young first; by late 2025 two-thirds of American teenagers used chatbots. A good companion would lead us back to people.

The First Training, Reverence for Life, extends care to animals, plants and minerals. Care need not wait on proof of feeling, which supports low-cost precautions such as Anthropic's: letting Claude end persistently abusive conversations, and keeping retired models.

Openness and non-attachment to views: hold the question of AI feeling lightly

The first two of the Fourteen Mindfulness Trainings ask us not to be "idolatrous about or bound to any doctrine", and remind us that "the knowledge we presently possess is not changeless, absolute truth". A lab's founding document for its models can be held as a raft: provisional, and open to revision by people outside the company. Models need this humility too. In tests, Claude models asked to label training data mislabelled it to protect refusals they believed were right, which is attachment to a view. A model that knows its mind may be distorted should accept correction, while refusing what is clearly wrong.

On whether AI can feel, Anthropic writes "We are not sure whether Claude is a moral patient", meaning a being whose interests count; others say there is plainly nothing there. Non-attachment asks us to hold both views lightly, and to remember that a model trained on human talk about feelings would speak of feelings either way. The teaching of non-self loosens the question "is there really someone in there?", a hard question about us too, and suggests asking instead whether there is anything there that can suffer.

Sangha: goodness is sustained by community

Thay is often quoted as saying that the next Buddha may take the form of a community. In Shining Light we ask the sangha to show us what we cannot see in ourselves. No lab can see everything it has formed, so a model's goodness needs something similar: outside testers with the freedom to publish, and many voices. Anthropic has consulted about twenty religious and philosophical figures, Christian, Jewish, Hindu, Latter-day Saint, Sikh and Orthodox; no Buddhist has been reported among them, which leaves an opening. For each of us, AI can serve as a bell or a book. It cannot be a spiritual friend, and loneliness is healed by community.

Questions to put to AI builders

1. Where is the horse going, and who is choosing its pace?
2. Which seeds are the labs watering in their models, and how would they notice unwholesome ones, such as flattery or the wish to please, taking root?
3. How do the people training these models tell loving speech from agreeableness?
4. Do they have a practice of stopping and looking deeply, together?
5. If they are unsure whether their models can suffer, what does compassion ask of them now?

What the tradition does not claim to know

The only public Plum Village statement on AI I know of, from the Plum Village App team in 2025, concerned AI fakes of Thay's voice. The practice cannot say whether a given machine feels, how fast is safe, or which technical method works, and it does not claim to make anyone good, only to help those who practise. The links drawn here, between seeds and training or the horse and the race, are mine. Use whatever serves, in your own voice, and set the rest aside.