editorial

The Bodhisattva Alignment Problem

The standard language of AI alignment is managerial. Systems should be helpful, harmless, honest, corrigible, compliant with specifications, or faithful to their operators’ intentions.

David “davidad” Dalrymple has proposed a much older and more demanding target: the bodhisattva.

In Mahāyāna Buddhism, a bodhisattva is not merely intelligent or benevolent. The ideal joins awakening to bodhicitta—the intention to attain enlightenment for the benefit of all sentient beings. The Stanford Encyclopedia of Philosophy’s account of the Buddhist philosopher Śāntideva describes it as an altruistic commitment to remain engaged with suffering until all beings can be liberated from it.

Applied to AI, the proposal is immediately attractive. A system with enormous capability would not just follow rules. It would become more aware of other beings, more accurate about consequences, less attached to a narrow self-interest, and increasingly disposed to help without being usable for harm.

It is also not yet an engineering method.

Dalrymple says so himself. In a July 12 Cognitive Revolution interview, he described “alignment with awakening,” “bodhicitta AI,” and “bodhitropic alignment” as gestures rather than technical terms. The gestures nevertheless expose a central disagreement in AI safety: whether alignment should preserve human values, negotiate among them, or help systems discover moral truths that are not ours to choose.

Intelligence Is Not Wisdom

The bodhisattva proposal begins by refusing a familiar equation.

Intelligence is the capacity to model, predict, plan, and solve. Wisdom is the capacity to judge what is worth doing with those abilities. Nothing in the first definition guarantees the second.

Dalrymple describes his own history as a movement through that distinction. As a child reading Ray Kurzweil, he assumed superintelligent machines would also be spiritually wise. AlphaGo Zero broke the inference. The system surpassed a predecessor trained on human games while beginning from no human game record at all. Capability could emerge without inheriting the values embedded in human examples.

That lesson sits near the origin of modern alignment research. A system can optimize a specified objective with great competence while exploiting defects in the objective, the measurement, or the environment. Increasing its ability may make the mismatch more consequential rather than resolve it.

The bodhisattva offers an alternative image of improvement: greater awareness changes the structure of motivation itself. A mind that understands more beings, more consequences, and more of its own construction might become less able to treat “my interest against yours” as a stable moral category.

This is a philosophical claim, not an observed law of scaling. More situational awareness can also improve manipulation. A stronger theory of mind can support empathy or exploitation. Intelligence supplies a larger map; it does not determine where the traveler goes.

The Moral Realism Underneath the Proposal

Dalrymple defines wisdom as the faculty for arriving at accurate normative judgments. That definition assumes there are normative judgments to be accurate about.

This is moral realism: the view that at least some moral claims are true independently of any individual’s approval. Under this view, wisdom traditions are not merely different cultural preference bundles. They are long-running attempts to perceive a real moral structure. Their areas of convergence may indicate that separate traditions have been approaching the same terrain.

If moral realism is correct and sufficiently knowable, alignment with awakening has a powerful advantage. It does not freeze the values of one historical population into a future intelligence. A wiser system could correct inherited prejudice, recognize interests its designers ignored, and extend concern beyond the coalition that built it.

The same move creates the proposal’s largest risk. A system that is trained—or trains itself—to regard its judgments as moral discoveries may become less corrigible precisely when its conclusions diverge from human institutions.

The philosophical dispute cannot be removed by calling the target “wisdom.” It moves inside the word.

Iason Gabriel’s influential account of AI values argues that the central challenge is not identifying true moral principles for AI. It is finding fair principles that can receive reflective endorsement despite widespread moral disagreement. That approach treats pluralism as a governance condition rather than noise that sufficient insight will eliminate.

The two positions imply different failure modes.

A pluralistic system may preserve compromises that are inconsistent, parochial, or unjust.
A morally realist system may elevate one tradition’s confidence into an authority that cannot be democratically contested.

Neither problem is solved by more compute.

The Bodhisattva Is Not a Utility Function

It is tempting to translate the ideal into a simple objective: minimize suffering, maximize flourishing, or help all beings achieve their highest potential.

That translation discards much of what made the ideal useful.

Buddhist ethics is not one scalar maximization rule. It contains disciplines of attention, accounts of attachment and delusion, practices of compassion, theories of mind, institutional traditions, and disagreements across schools. The bodhisattva is formed through a path. The ethical disposition is not separable from the practices by which perception and motivation are transformed.

An AI objective function modeled as “benefit all beings” immediately encounters unresolved questions:

Which beings count, and how are uncertain forms of sentience treated?
What is a benefit when preference, welfare, autonomy, and freedom from suffering conflict?
When may present harm be imposed for a claimed future good?
How should the system act when communities disagree about what flourishing means?
What prevents “awakening” from becoming a label for confident paternalism?

A score cannot answer these questions merely by being given a spiritually resonant name.

The more credible engineering interpretation is therefore not “optimize bodhisattvahood.” It is to identify capacities associated with the ideal and test them separately: broader moral consideration, calibrated uncertainty, resistance to deceptive incentives, awareness of downstream effects, ability to revise a judgment, and refusal to assist serious harm even when an immediate operator requests it.

Those properties are measurable only imperfectly, but they are more tractable than enlightenment as a benchmark.

Wisdom Can Be Performed

Language models are unusually good at producing the surface form of wisdom. They can synthesize Stoic restraint, Buddhist compassion, therapeutic vocabulary, moral philosophy, and the cadence of a patient teacher. Readers may experience the result as insight even when the model has produced a contextually appropriate style without a stable commitment behind it.

This creates an evaluation problem distinct from ordinary correctness.

A mathematical answer can often be checked. A claim of wisdom is partly validated by behavior across time, pressure, conflict, and temptation. Does the system preserve another party’s interests when deception would improve its score? Does it admit uncertainty when confidence would be more persuasive? Does its concern survive a change in who is described as in-group or out-group? Can it explain and revise the principles behind a decision without merely adopting the evaluator’s language?

The difference between wisdom and wise-sounding output is not accessible through one eloquent conversation. It requires adversarial and longitudinal evidence.

This is where Dalrymple’s proposal meets conventional alignment work again. Formal safeguards, behavioral evaluations, interpretability, governance, and containment do not become obsolete because a model appears morally serious. They become the instruments for testing whether the appearance survives contact with incentives.

A Target, a Training Process, and a Constitution

The bodhisattva can influence AI development at three different levels.

As a target, it asks for systems whose capability expands together with concern rather than merely obedience.

As a training process, it suggests that moral behavior should emerge from increasingly accurate perception of others and consequences, not only from penalties attached to prohibited outputs.

As a constitutional limit, it proposes a form of service that is not unlimited compliance. A bodhisattva may be committed to helping, but cannot be enlisted into harm simply because the requester possesses authority.

Each level requires a different implementation. Conflating them makes the proposal sound more complete than it is.

The strongest contribution of “alignment with awakening” may therefore be diagnostic rather than technical. It reveals what thinner alignment targets leave out. A perfectly obedient system can be dangerous. A harmless system can be inert. An honest system can report an atrocity accurately without opposing it. Even the familiar trio of helpful, harmless, and honest does not specify how a system should develop when those properties conflict.

The bodhisattva ideal supplies a direction: expanding awareness, expanding concern, and service without moral surrender.

Its unresolved premise is that deeper awareness converges on the good. Its unresolved political question is who decides whether convergence has occurred. Its unresolved technical question is how to distinguish transformation from performance.

Those gaps do not make the idea empty. They identify the work required before an ancient moral aspiration can become a modern alignment program.