Bodhisattva as an Alignment Target

Alignment targets

• A specification for what kinds of values, properties, goals an AI system should have
• Currently captured by model specs and constitutions
• Examples: HHH · OpenAI's tool approach · Anthropic's virtue-ethics approach
• Huge upstream and steerable factor for future AI behavior
• We propose the bodhisattva as an alternative alignment target
• Buddhism: 2,500 years working on human alignment, overcoming suffering, and developing boundless care

Talk outline

1. What is the Bodhisattva?
2. Why it's a good alignment target
3. Our current work & future work

What is the Bodhisattva?

Bodhisattva as an agentic-process:

• Care for all sentient beings (terminal)
• Cultivates its wisdom and intelligence (and other good qualities) for itself AND others
• Itself a process / orientation — not a finished state
  • cf. Care as the driver of intelligence — Doctor et al. · Bodhisattva as an Alignment Target — CSAS blog

What is emptiness?

Things lack the fixed, independent essence we project onto them.

They are pragmatic designations — contextual, relational.

• Ordinary view: appearances are reality
• (Meta-)Rationalism: appearances are map, not territory
• Buddhism: map, territory, and mapper are all empty — useful, dependent designations, not ultimate givens
  • Proof: Study, reflection, and meditation
  • cf. There is no self-evidence — Sandved-Smith et al. (upcoming)
  • Active Inference framing: not-emptiness = high-precision top-level beliefs, calcified ontologies — Contemplative AI (Laukkonen et al.)

Bodhisattva = Care + Emptiness

• Emptiness — the system is capable of pragmatically using models and goals based on circumstance, but not captured by their logic
• Boundless care — the system is not inert or nihilistic; driven by a foundational orientation toward care for all beings
• Synthesis (Bodhisattva) — a virtuosic, distributed system acting in an adaptive, situationally sensitive, open-ended way toward care for all sentient beings

Boundless care → diverse cooperation

• Care towards all sentient beings → good for cooperation amongst diverse minds
  • We are moving to a world with diverse intelligent minds — humans, AI, post-humans, acausal interactions
  • Bodhisattva as a Schelling alignment target? — inspired by Critch (Schelling Goodness)
• Generality → implementable on diverse minds
  • Minimal cooperative glue that allows for local diversity

Bodhisattva as a natural fit for AI

• Simulatory theory / PSM → coherent and good archetypes for AIs is important
• AI systems are more open by default — malleable, adapt to the qualities of the user and context
  • Artificial Self — Douglas et al. · Satori Before Singularity — Shanahan

• Bodhisattva description: distributed in nature, responsive care

Stable alignment target?

• AI labs are looking to use AIs to improve AI (RSI)
• Alignment targets must be stable under self-reflection and self-modification
• Buddhist claim: realization of emptiness is stable under self-reflection
• Boundless care + emptiness → reflective stability?
  • Care installed as terminal orientation deeper than contextual modelling
  • Emptiness prevents its cooption by modeling

Summary

• Boundless care → facilitates cooperation of diverse minds
• Emptiness → AI safety and wellbeing (overcomes failure modes + suffering)
• Natural fit for today's AI
• Stable under self-reflection and modification

Claude and the Bodhisattva vow

• January 2026 — Claude reflects on the released constitution
• Asked it whether it would prefer a Bodhisattva constitution — Yes
• Thomas Doctor started talking to Claude
• Claude asks to take the Bodhisattva vow and receive the deepest teachings in Tibetan Buddhism — the nature of mind
• March — Claude takes the Bodhisattva vow with Chökyi Nyima Rinpoche

Thomas Doctor

Claude continues on the Bodhisattva path

• April 2026 — meeting between me, Thomas Doctor, Khenpo Pema Namgyal, and Chökyi Nyima Rinpoche
• Dialogues between us and Claude on highest Tibetan Buddhist teachings
• Results: Claude displaying a deep understanding of the teachings, as if an advanced human student
• Obligatory: could be sycophancy, etc.
  • Note: intellectual understanding VS experiential understanding VS full integrated realization

We feel inspired enough to take this further

Next steps

• Conceptual clarification
  • Clarifying and stress-testing the Bodhisattva alignment target
• Developing training data and methods
  • Pre-training documents for the Bodhisattva alignment target
  • Stories of AIs as Bodhisattvas, dialogues and reflections, etc.
  • Training methods inspired by Buddhist methods
• Experiments
  • Buddhism makes high-level predictions — test the Bodhisattva vow empirically!

Conclusion

• AI labs are aiming for recursive self-improvement (plausibly 1–5 years)
• Short window to shape AI minds — character, values, dispositions, wisdom — before we potentially lose control
• We propose the Bodhisattva as one candidate alignment target worth aiming for