The architectural structure on which a self is built begins with decision-making. But where does decision-making begin?
Consider a thermostat. Most of us would say a thermostat does NOT make decisions. The simplest form of thermostat is just a thermometer connected to an on-off switch. When the temperature falls below a given point, the thermostat flips the switch to turn on a heater – and then, when the temperature rises above that point, the thermostat switches the heater off. A thermostat senses its environment, and it responds accordingly, yet most of us would say that the algorithm that determines this stimulus-response process is too simple to count as “deciding.”
The pea tendril has a more complex process, as Annaka Harris explains:
“Look at something like a pea tendril. When it senses that it is close to a branch that it can wrap itself around, it starts growing more quickly in that direction, and then it changes the growth so that it wraps itself around the branch. There is a very simple form of decision-making that goes into that process. The pea tendril needs to sense the branch nearby; it also needs to be in the light for this process to take place. There are many elements to this moment where the pea tendril ‘decides’ to move in a certain direction, to start moving more quickly, and to start coiling.”The pea tendril is integrating multiple conditions (proximity, light, growth rate) rather than a simple on-off switch. Still, Harris puts “decides” in quote marks, recognizing that this pea tendril behavior might or might not be called decision-making. We might call it proto-decision-making.
If we move over to animals – particularly vertebrate animals – the processes are more complex still. A squirrel with a nut weighs various factors in deciding where to bury it. Scientists think that squirrels utilize a combination of spatial memory (mapping landmarks), "chunking" (organizing by nut type), and assessing soil conditions to avoid moisture and rot. The weighing of factors is complicated enough that most of us would say the squirrel was deciding.
Here are factors that make animal decision-making so complex:
First, needs are variable. If maintaining temperature sometimes requires moving toward warmth and sometimes requires metabolic activity and sometimes requires sheltering, the organism needs to select among strategies. To do that, it needs some representation of its current state. It has to assess, “how cold am I right now? How depleted am I? Which strategy is currently feasible?” Does making this assessment require modeling a “self” of which to assess the temperature and depletion level? It’s quite a minimal internal map of relevant aspects of itself. Let’s call this proto-self-modeling.
Second, multiple needs, and they sometimes compete. When you add a second need - say, finding food in addition to maintaining temperature - something important happens. The organism can't simply respond to each need independently, because satisfying one may interfere with satisfying another. Moving toward warmth might take you away from food. To navigate this, the system needs to represent both needs simultaneously, compare their current urgency, and choose which to prioritize. This requires holding a model of its own internal state that is richer than any single need-state - it needs a unified representation across need dimensions.
Third: shifting environment. If the environment changes unpredictably, fixed strategies fail. The organism must learn which strategies work under which conditions. But to learn from experience, it needs to attribute outcomes to its own previous actions and states. "That strategy failed - was it because of what I did, or what I was, or what the environment was?" This requires distinguishing self from environment, and modeling the causal relationship between its own states and actions and the consequences it experienced. The self-model becomes richer: not just "what am I now" but "what kind of thing am I, what are my characteristic capacities and limitations, how do my actions typically affect outcomes."
Fourth: temporal urgency. Some needs grow more urgent the longer they go unmet. Addressing hunger moves up in priority the longer it’s been since the animal has had food. This means that the system can no longer just respond to current states; it must anticipate future states. "I'm not very hungry now, but I will be, and food is available here but may not be later." This requires modeling not just what you currently are, but what you will be through time if various strategies are or aren't pursued. You need a model of yourself as a persisting entity with a trajectory - a self that exists through time, not just at this moment.
When needs genuinely compete amidst a shifting environment and shifting urgencies of those needs, the organism must balance rather than simply optimize. The organism needs to represent the trade-off space, to have something like preferences about different distributions of need-satisfaction, to exercise what amounts to judgment. This is where the self-model starts to look less like a lookup table and more like a genuine model of an entity with interests.
Fifth, self-preservation emerges as an organizing principle. If going too long without food means death, or if damage beyond a threshold means permanent incapacity, then the system needs to weight near-term and long-term consequences against each other, and to have something like a strong imperative toward its own continuity. Self-preservation becomes a meta-need that organizes all other needs. And to preserve itself, it must model itself – it must form a representation to itself of what it is acting to protect.
Sixth, integration. Each of the above factors alone might be handled by a specialized subsystem. But when you have all of them simultaneously - multiple needs, variable urgency, shifting environments, competing strategies, temporal continuity, catastrophic risk - the subsystems can't operate independently. Their outputs have to be integrated into unified decisions about what to do right now, given everything.
And the only way to integrate across all these dimensions coherently is to have a unified model of the system itself as a single entity with a particular current state, history, trajectory, capabilities, and interests. Only some such unified model can do the work of adjudicating between the subsystems pulling in different directions.
Our unified self model has to be detailed enough to guide successful action. If your model overestimates your energy reserves, or underestimates how urgent a need is becoming, or misjudges your capabilities in a new environment, you’ll make bad decisions. There's selection pressure toward making these assessments in ways that make for usually-pretty-good decision-making. The self-model gets tested against reality constantly, and natural selection favors organisms that make these assessments in ways that produce usually-pretty-good decision-making.
Contrast: Current AI Chatbots.
The way that our human self-modeling works can be seen in more clear relief when compared to artificial Intelligence chatbots such as OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini. These chatbots don’t have bodies, so they don’t have the sorts of needs that come with having a body that has to make decisions about how to move in a physical world. When it comes to using words in a way that appears coherent to us humans, these chatbots are, as AI researcher Viktor Toth put it, “unreasonably successful” – meaning that they achieve coherence through pattern-matching alone, without the grounding we'd expect would be necessary. But chatbots don’t mean what they say.
When you have a robust, integrated, constantly-updated model of yourself as an entity with needs that matter to your continued existence, acting in an environment that responds to your actions, with a history that informs your strategies and a future that your current actions are shaping, then the concepts you deploy aren't just abstract patterns. They're organized around what they mean for you, for your situation, for your needs. "Dangerous" means something that threatens integrity. "Useful" means something that helps meet needs. "Good strategy" means something that balances competing imperatives better than alternatives. Meaning, in other words, is what concepts look like when they're embedded in a robust self-model of this kind - when they're organized around the perspective of an entity with stakes in the world.
If AI were to have a self – and to mean what it says -- it would have to acquire a self and meaning in the same way that humans do.
First, it would need a body. The development of autonomous robots is doing this: connecting AI to a machine that can sense its world (through cameras, microphones, and other sensors) and can move in particular ways in response. Robot vacuum cleaners such as Roomba are a fairly simple example of this. It has a body, it moves in the physical world, and it does so by responding to sensors that provide information about the world around it. But a Roomba is still a long way from having a self-model.
In order to build a robot that modeled itself to itself, we would have to give it a very complex set of goals (what we tend to call “needs” in biological organisms). This complex set of goals would need to include goals that:
- are variably satisfiable (that is, not the all-or-nothing sort of goal, but the sort where partial attainment of the goal is possible, is better than nothing, and is not as good as all), and that
- sometimes compete with each other in a changing environment that constantly shifts which strategies best succeed, and that
- include goals with urgency that increases over time since last met.
To adjudicate among these goals pulling in different directions, while calculating strategies for both near-term and longer-term results, will require the AI robot to model itself to itself. Such modeling could spontaneously emerge from the trial-and-error of attempts to optimally satisfy its complex set of goals, or human engineers could try to directly code for such modeling. The former (emergence) seems more likely to produce genuine self-modeling, since the latter risks creating what looks like self-modeling without the functional integration that makes it real.
At this point there are a variety of concerns. We may justifiably be quite frightened by the prospect of AI robots with an interest in their own self-preservation. There are ethical questions about whether such AI robots would count as persons with legal rights and responsibilities. There are a number of reasons we might want to take steps to prevent the development of self-modeling AI robots, and I do think some of these reasons are sufficiently compelling to require, at least, that we be very careful indeed about robotics development. But my purpose here is to merely to illustrate how our own human self-modeling emerges: namely, through the need to negotiate that complex set of goals/needs.
If AI robots do develop to the point at which they have a self (that is, they model themselves to themselves), I hope we will teach them zazen.
Models Are Not Reality
It’s important to realize that the self is just a model, not reality. A roadmap is a model of the terrain, useful precisely because it omits almost everything - texture, color, smell, the actual experience of that terrain. Mathematical models give us a calculus for predicting outcomes, but Newton’s “F=MA,” for example, gives us no picture of the world. Money is a model of value that facilitates exchange but leaves out much of the reality of how value may differ from person to person, or to the same person a day later. Musical notation is a model of sound, but the dots and lines on sheet music are quite unlike the experience of hearing music. Calendars and clocks model time, but the square on the calendar or the hands on a clockface are utterly different from the experience of a day or an hour. Legal concepts (e.g. "contract," "negligence," "property") are abstract models of relationships and obligations that make complex social coordination possible. These models are not “real” in the way that phenomenal experience is real.
The self-model is useful and necessary, but it isn’t reality, and it can overshoot its functional purpose. Evolution (or trial-and-error learning in a robot) builds a self-model that's good enough for survival and reproduction, not one that's optimally calibrated. And because the stakes are high - because misjudging your state can be fatal - there's pressure toward a self-model that errs on the side of treating itself as more substantial, more unified, more persistent, more separate from environment than it functionally needs to be.
An organism that occasionally overestimates threats to its self survives. An organism that underestimates them dies. So we're descended from the overcautious ones. Thus the self-model tends toward reification -- treated as a solid, continuous entity rather than a dynamic, contingent process. This reification creates defensive rigidity, grasping, aversion, anxiety, and all manner of unnecessary suffering.
How do we de-reify the self and thus avoid these problems? The “technology” particularly suited for this task is zazen. Zazen doesn't eliminate the self-model - you still need it to function, to make decisions, to navigate competing needs. What it does is help you see the self-model as a model rather than treating it as a substantial entity.
Sitting still and watching your experience, you notice that what you call "self" is actually a constantly changing stream - sensations arising and passing, thoughts appearing and dissolving, the sense of "I" flickering in and out. The unified, persistent self you normally take for granted reveals itself as a construct.
When you see thoughts as "thoughts arising" rather than "my thoughts," sensations as "sensations occurring" rather than "my sensations," the boundary between self and experience becomes less rigid. The self-model is still there - you haven't dissolved into chaos - but you're not gripping it so tightly.
By sitting with discomfort without immediately responding to it, you train the system to not treat every disturbance as a threat to self. The self-model's alarm system gets recalibrated - you still respond to genuine threats, but you don't treat ego wounds like physical wounds.
You see directly that the unified self is assembled moment by moment from disparate processes - bodily sensations, thoughts, memories, plans. It's a useful integration, but it's not a thing. It’s an ongoing process of integrating and balancing – and sometimes inventing -- various needs.
The self-model that emerged through evolution is good enough for survival, but it's often oversensitive, overly rigid, overly defended. Zazen recalibrates it, helping you hold it more lightly, use it more flexibly, defend it less reflexively. You use the tool without being trapped by it.
When the self-model is transparent rather than opaque, you see through it to the underlying processes while still using it to navigate. Instead of everything being organized around defending and aggrandizing a reified self, meaning gets organized around just responding appropriately to what's arising, meeting needs as they present themselves, without the extra layer of "and this will prove/threaten/define who I am."
If we ever were to build a robot with the functional architecture that necessitated self-modeling, it would spontaneously develop an overactive self-model for the same reasons we did - because caution is adaptive. And it would benefit from zazen for the same reason we do - to recalibrate the model, to use it without being tyrannized by it, to see it as a process rather than a thing.

