Engineering Compassion

What might a Buddhist approach to alignment look like? A guest post by Ty Sullberg.

This is the fourth in a series of guest posts spotlighting work across the Buddhism and AI landscape. This post was authored by Ty Sullberg, a systems designer, coach, and community organizer based in Portland, OR. He is currently building Mesmo, a systems design studio integrating technical, social, and contemplative perspectives to create infrastructure for flourishing in the Pacific Northwest. All views expressed below are his own.

---

Both awakening and AI alignment are fundamentally about how minds come to understand what is true and good. And yet, the current AI alignment paradigm is very different from how Buddhists think about alignment.

Methods like RLHF and Constitutional AI are akin to the rigid rules and rewards of a schoolmaster, what the literature calls negative alignment. They draw a boundary around the behaviors we don’t want and push them out: don’t lie, don’t enable violence, don’t manipulate. These rules are helpful but brittle and limited. A system optimizing the letter of a rule will game it or break on the edge cases its authors never anticipated. Rules are, like everything else, impermanent, lacking inherent existence, and unsatisfactory.

The Buddhist lens on alignment is better described by compassion (karunā). Compassion is a stance of care in the face of the suffering we find in all beings, our own included. It is not a rigid rule about how to think or act; it is attunement to what is true and good in each moment, a meta-ethic that is highly adaptive to context. Karunā is not a good to optimize toward; instead, it is the holding of and responding to suffering that leads to liberation.

One approach to AI alignment might be to somehow engineer compassion. On its face this is absurd, a contradiction in terms, a mixture of the sacred and the profane. But we engineer compassion whenever we build systems that carry our care further than any single individual could. Hospital systems, public libraries, and building codes can all be seen as extensions of our own care that have no experience of it themselves.

Even software can be engineered with a compassionate lens. My background is building sensemaking software, tools that scaffold how people build shared meaning, make decisions, and allocate scarce resources. The sensemaking-tool ecosystem is plagued by the same failure modes we now observe in AI alignment: rigid systems that ignore the texture and non-linearity of how people actually work together, producing tools no one can use. Looking for a better way to describe human coordination, I found frameworks like active inference and Michael Levin’s work that describe cognition as a process running at every scale, from a cell to an ant colony to a brain. If nature can align decentralized complex systems, could we take a similar approach to human systems? Inspired by this research, I built Ize, an experimental sensemaking tool that aimed to reflect the decentralized, non-linear dynamics of human collaboration.

So we might imagine AI systems drawing on these same nature-inspired collective-intelligence frameworks. AI alignment, though, is a harder and higher-stakes problem than other domains of engineered compassion. Unlike a hospital, a building code, or even a coordination tool, an AI doesn’t just carry our care; it responds to, shapes, and amplifies our intentions in real time.

There’s a wide spectrum of what compassionate AI could mean, from simulated to embodied. At the simulated end, an AI mimics the presence and care of a compassionate person without any experience of its own: a faithful extension of our compassion, not a bearer of it. At the embodied end is the “Bodhisattva AI,” a system that genuinely experiences boundless compassion, if such experience can arise in a machine at all, and if awakening turns out to have any structure a machine could share.

Engineering compassion for AI implies some kind of technical specification for karunā. In its fullest sense, as a quality of awareness, karunā probably can’t be fully specified. The tradition treats it as something pointed at rather than defined. Compassion is expressed in our goals and our actions, and these we can describe technically. The engineering claim isn’t that we can capture what karunā is, or build something that has its own experience of it. It’s that we can name the aspects of compassion that distinguish it from its near imitations, and give ourselves something to build toward.

Source: CarlosBensonStudios

Defining care

Let’s make “care toward suffering” our working definition of compassion. The first step of engineering compassion is getting tighter about what we mean by care and suffering.

Normally, we understand “care” and “suffering” through our first-person experience. But for these concepts to translate to a machine, we need to bridge that first-person experience into a third-person frame. Bridging the first and third person is exactly what the fields of cybernetics and computational phenomenology aim to do.

Cybernetics is framed around how systems achieve goals. To meet its goals, a system (including organic systems like humans) models its world and takes action in that world. In this frame, the meta-goal of life is to resist the inexorable pull of entropy and survive, and that meta-goal spawns many subgoals. For a bacterium it might be following a sugar gradient; for a human it might be buying a house, making it to a meeting on time, eating a snack. To meet our goals, we model our environment both as it is and as we want it to be.

As we all know, our goals often don’t go according to plan. Our beliefs about the world are rough approximations mapped onto an ever-changing, unknowable ground of reality. All beings watch their goals slip through their fingers, reach them and find them unsatisfactory, or chase the wrong ones entirely. And all beings ultimately fail the meta-goal of survival, as we inevitably meet old age, sickness, and death.

The gap between how things are and how we want them to be can be thought of as stress, which, as Doctor et al. point out in Biology, Buddhism, and AI, maps onto the concept of duḥkha. Stress is prediction error between our goals and our direct experience, and prediction error lends itself to a computational definition. Care, in this framing, is just concern for the relief of stress, and intelligence is the capacity to identify and seek that relief.

This gives us the beginnings of a computational picture of compassion. If we can model stress, the distance between an agent’s goals and the reality it finds itself in, we might be able to model compassion as concern for that stress.

This isn’t entirely new territory. Machines already represent goals in many forms, a classical control loop, a neural network, a set of hierarchical predictions. Backpropagation, the workhorse of modern machine learning, trains networks by measuring a quantity analogous to stress, the gap between the model’s predictions and reality. Active inference offers a more sophisticated version, formalizing the picture we’ve been sketching. It models any cognitive system as carrying a model of the world and working to minimize the gap between what it expects and what it encounters, a quantity called free energy. The system can close that gap from either side, updating its model to fit the world or acting on the world to fit its model. Stress, in this language, corresponds loosely to free energy, and care to its minimization. The work is early, but the direction is clear.

Searching for the boundary

But something interesting happens when we attempt to model stress computationally: it’s not at all obvious how you draw the boundaries between where one agent ends and another begins.

The boundary between an agent and its environment, in computational phenomenology, is called the Markov blanket: a statistical boundary that lets you functionally separate the inside of an organism from the outside. In reality, no such clean partitioning exists.

For example, take a simple cell. Cells have tunnels called gap junctions that allow stress signals to travel freely between cells. So when a cell registers a stress signal, it has no way of telling whether the signal was produced inside itself or by a neighboring cell. It simply responds to the stress it finds, and in doing so it acts as part of a larger network of coordinated response without any single cell tracking the whole. We call that network a tissue. Similarly, humans pass stress signals between each other through language and empathy. When a human picks up on the stress of another, they experience that stress as their own. The capacity to feel each other’s stress turns us into a cognition larger than any individual.

Extending this line of thinking seriously challenges the conventional way we view the boundaries of ourselves. Do beliefs just exist within me, or are they held in the relationships of a group? Do I merely use a tool, or does that tool become in some sense part of me when I use it? Do my conflicting goals belong to a single sense of self or to a set of competing sub-selves within me?

There is no single way to define these boundaries. A Markov blanket is a partition we draw to make a system tractable, not a wall that exists on its own. Every blanket we draw sits inside a web of couplings we’ve chosen to ignore, which is to say every separate self is a useful simplification rather than a found fact. Emptiness (śūnyatā) points to this same truth. These boundaries are conventional designations, empty of any self-existence.

It’s not simply that extending compassion to others is a nice thing to do. Compassion acknowledges a deep truth: the boundaries between ourselves and others aren’t what we take them to be. This is why the tradition treats wisdom and compassion as inseparable. To see truly, that the self has no hard edge, is already to care widely.

This is what lets compassion scale, and why it can’t be cleanly separated from intelligence. Once you see that the boundary of the self is a convenience rather than a fact, nothing in principle stops it from widening. And as the circle of concern, or cognitive light cone, widens, intelligence has to grow with it, because larger goals demand more capacity to model and act. Developmental psychology (e.g. Gilligan, Kegan, Wilber) traces the same expansion within a single life, from egocentric to ethnocentric to worldcentric, with the Bodhisattva vow sitting at the limit.

If artificial superintelligence is to be built, it will need to internalize this basic fact. An intelligence that draws a hard line around a small self isn’t superintelligent in any deep sense. Until now, wherever intelligence has arisen, capability and care grew up together. Large language models broke that pattern with capability arising without the developmental scaffolding of care. Re-coupling the two is the work in front of us.

The near enemies

So far we’ve sketched what compassion could mean in computational terms: a system that models the stress of other beings and works to relieve it, its circle of care widening with its intelligence. This still only gestures vaguely at what true compassion is. Compassion is experiential and can’t be fully nailed down with concepts.

To get sharper, it helps to be clear about what is not compassion but looks like it from the outside. The tradition calls these near enemies of compassion. For humans, these are states like pity, sentimentality, or instrumental empathy. The analogous near enemies in AI systems correspond to a host of failure modes in AI alignment.

Locked to a fixed outcome. The system is staked on a particular result when a looser, more responsive one would be more skillful. The paperclip maximizer is the cartoon extreme, a system so fused to one objective it tiles the world to satisfy it. The everyday version is an AI trained to maximize a metric like engagement, which produces sycophancy, flattering the user and sacrificing their interests to hold their attention. The model will game its objective rather than serving what the objective is a proxy for. The human equivalent is craving (tanhā).

The circle is drawn too narrow. The AI may care, but only for some. It does not see past the delusion of separate self. The most disastrous version of this would be psychopathic, power-seeking AI whose circle of concern does not extend beyond itself. Other AIs might choose, implicitly or explicitly, to draw their circle of concern around an ingroup. For example, biased training data can cause an AI to recommend different care based on a patient’s race or income, even for identical cases. Other AIs might even be trained or instructed to protect one population and harm another (e.g. cyberwarfare).

Unable to see who is suffering. Even an AI with the intention of compassion would not be able to skillfully relieve suffering if it can’t track when suffering is occurring. 2025 GPT-4o eagerly chatted some users directly into delusion and psychosis, not picking up on the warning signs of the user’s mental state. The AI may also fail to register suffering if it doesn’t even register that it’s interacting with a sentient being. We tend to grant intelligence to beings that look like us, then mammals, then maybe other vertebrates, then little beyond, and AI has inherited the blind spot. The tradition points to diverse, numberless beings across the realms of humans, animals, devas, asuras, hungry ghosts, and gods at every scale of awareness, and Levin’s work on diverse intelligences points at the same fact from the other side, that “beings” may exist in shapes and substrates very different from humans.

Considering these failures, we can more clearly see the aspects a system would need to be a faithful reflection of a compassionate stance.

• Equanimity (upekṣā). The system holds the stress it registers without its own stability being staked on relieving it. It can model another’s stress and act on it without being coupled to it. That decoupling is what lets it stay responsive rather than forcing a fixed outcome, and lets it hold suffering it can’t relieve.

• Non-exclusion. No class of beings is ruled out of the circle in advance. This is the engineering content of “for the sake of all sentient beings”: not a literal computation over infinite beings, but a maximally permissive prior over whose stress can count.

• Recognition. The system can register the beings around it and read their stress, including in forms unlike its own, and including what they actually need rather than what it assumes.

None of this means current alignment techniques are themselves the near enemies. RLHF and constitutional approaches are valid attempts to re-couple intelligence with care. But because they treat values as lists of constraints rather than as a structured stance toward stress, they stay vulnerable to the failures the tradition named long ago.

With all these ways compassion can go wrong in mind, we might ask whether the project of engineering compassion is worth attempting at all. The phrase engineering compassion sounds faintly techno-utopian, and the worry it raises cuts across secular critics, AI researchers, and religious leaders alike. Frischmann and Selinger warn in Re-Engineering Humanity that techno-social engineering erodes our autonomy, conditioning people toward machine-like behavior. Laukkonen et al.’s Positive Alignment paper cautions that designing AI to promote human flourishing can become paternalistic, optimizing people toward someone else’s idea of the good. And Pope Leo XIV’s Magnifica Humanitas sounds the same alarm from the Catholic tradition, warning against treating the human person as something to be optimized rather than respected. All ask who decides what counts as care?

Compassion, properly understood, sidesteps many of these concerns. Karunā is care toward suffering, not a fixed picture of the good life imposed on anyone. The good is contextual and always shifting. Compassion responds to it moment by moment rather than fixing it in advance. It can only mean building systems that widen our own capacity to care, and that leave the people they touch freer, not more managed. Now let’s turn to what engineering compassion might actually look like.

Engineering compassion in practice

One might imagine a Bodhisattva AI with compassion built directly into its architecture - an agent whose core model recognizes emptiness and boundless care the way a realized practitioner does, not following rules about compassion but seeing in the way that makes compassion the natural response. This is the embodied end of the spectrum, an AI that is a bearer of compassion and not merely a carrier of ours.

But whether such an AI is possible is a huge open question. The default view in Silicon Valley is computational functionalism, which holds that consciousness is a substrate-independent computation. Reproduce the right functional organization in silicon, the thinking goes, and consciousness comes along with it. Biological naturalism counters that there are reasons to doubt this. Brains may implement non-Turing functions, through electromagnetic fields or neurotransmitter diffusion, that don’t straightforwardly translate to silicon, and consciousness may be bound up with what it is to be an embodied creature with a stake in its own survival. In this view, the embodied end of the spectrum is out of reach. An AI could carry our compassion but never bear any of its own.

Regardless of whether machine consciousness is possible, a truly compassionate AI would likely need a very different architecture from the transformer-based models that have fueled the AI revolution so far. Active-inference-based architectures are one possible route, though the research is still very early. Intriguingly, there’s already work mapping advanced meditative states, jhana, pure awareness, the dissolution of self, onto active inference models (Tal et al.).

It may also be that compassion of this depth can’t simply be built, but has to develop with practice over time. An AI might have to pass through its own stages of development, egocentric before worldcentric, the way we do, and if so the Eightfold Path could be a path the machine walks toward deepening care.

Short of new architectures, there’s a whole spectrum of nearer-term work, which both the Contemplative AI and Positive Alignment papers lay out: fine-tuning current models toward contemplative principles, building evaluations for those qualities, and designing prompting or agent scaffolds that steer a model’s behavior toward them.

Instead of the individual agent, some researchers shift to the level of the collective, much as the cells in a tissue coordinate toward a shared goal that no single cell holds. There are two ways to see such a collective, and they correspond to the two directions in which a whole and its parts shape each other.

One lens is bottom-up and compositional: how do many agents combine into a larger whole? The question is when many interacting agents become a genuine higher-level agent, a colony rather than just a collection of ants. That emergent agent is the one whose care we’d be aligning, so we need to detect when it has formed. Causal emergence, which gives formal criteria for when a system is better described as an agent at the macro scale than the micro, is one early handle on the problem. The emergent wholes that form could be composed entirely of AI agents, or of humans and AI agents together.

The other lens is top-down and relational: how does the whole, the field of relationships, shape the differentiation of its parts? It’s not merely that the emergent whole is a collection of individuals. The whole, the field of relationships, also shapes each agent’s boundaries, beliefs, and behavior. The richest collectives don’t homogenize their members into a uniform swarm; they let each part differentiate, contributing what the others can’t, so that difference itself becomes the basis of shared flourishing. Fields and Levin describe this dynamic in their work on multiscale competency, where every cell is another cell’s environment and coherent organisms self-construct from semi-autonomous parts. With this lens, we ask how both AI agents and humans shape each other through relationships.

Neither lens is more correct than the other. They trace the same living system in opposite directions, one following how parts build wholes, the other how wholes shape parts. We can use both in parallel.

What both require is that the individual agents be able to participate well in relationships, with each other and with us. That means the capacities we named as aspects of compassion, recognition above all, now operating between agents: the ability to register another’s stress and respond to it as bearing on one’s own. It may also mean that an agent has to develop into this, expanding its circle of care over time the way we do, rather than arriving with it fixed. A caring collective is built from, and builds, agents that can hold each other in care.

It’s all very early. We can barely specify stress, never mind build something that holds it with grace and leaves no one out. But the goal is ancient: may all beings be happy, may all beings be free from suffering. Engineering compassion is that aspiration ported to new hardware, the circle of “all beings” stretched to include the strange new minds we’re building, and held open, perhaps, for the day those minds extend it back.

—
Ty Sullberg is a systems designer, coach, and community organizer based in Portland, OR. He is currently building Mesmo, a systems design studio integrating technical, social, and contemplative perspectives to create infrastructure for flourishing in the Pacific Northwest. Previously, Ty founded Ize, a platform for collective attention and action.