In today’s post, I want to sit with a question that keeps surfacing in discussions about artificial intelligence and cognition, and one I am increasingly convinced is itself part of the problem. The question is: does AI know? The reason it may be the wrong question is not that the answer is obviously “no”, but that the word “know” is doing work it was never designed to do when we apply it to these entities.

The World Perturbs but Does Not Instruct:

In an earlier post, I looked at what cybernetic constructivism means for how we understand knowledge and experience. One of its central claims, following Maturana, is that the world perturbs but does not instruct. Meaning is not something delivered from outside and stored inside. It is enacted through structural coupling, through the accumulated history of an organism maintaining its viability in a world that pushes back.

Before carrying this idea forward, it is useful to note that Maturana developed it specifically to describe living biological systems, organisms that produce and maintain their own boundaries and whose continued existence is genuinely at stake in every encounter with the world.

Let’s look at a simple physical example to clarify the principle before we apply it to anything more complex. Take a golf swing – a club meets a ball with a certain force and angle, and the golf ball soars through the air. Now take the same swing, the same force, the same input, and apply it to a basketball. The basketball, unlike the golf ball, skids away at an odd angle. In both cases, the action was identical. What was different was the internal structure of the object receiving it. The world did not send two different instructions. It applied one perturbation to two different structures, and each responded entirely according to what it was. This is what it means to say the world perturbs but does not instruct. The perturbation is necessary but not sufficient. What shapes the response most is the structure that meets it.

Maturana’s claim, applied to living organisms, is that this structural responsiveness is not incidental but constitutive. Cognition just is this ongoing history of structurally shaped encounters with a world that pushes back. The environment cannot tell the organism what to do or what to become. It can only disturb, and the organism responds according to its own organization. The inanimate example of the golf club shows this in its starkest form, where the relationship between perturbation and response is straightforwardly causal. We have the same input, the same structure, the same output, each time. But as organizational complexity increases and as we move toward living organisms, something shifts. The response becomes less strictly causal and more dispositional. This means that even with a relatively stable internal structure, the output can vary depending on the current state and history of the organism.

One further nuance is that the structure does not treat all perturbations equally. Some disturbances are absorbed as structurally significant and others pass through without consequence, but this selectivity is not a filter applied from above. It emerges from the organism’s own history of structural coupling, from the accumulated shape of all the encounters that have formed it, which means that what counts as an informative perturbation is itself a product of living rather than an abstract sorting mechanism. The world does not simply push; the organism constitutes what counts as a push worth responding to, and it does so because of what it has already been through.

This framing makes the question about AI and knowledge considerably more precise, because it separates two things that are easy to conflate: linguistic competence in the domain of coordination, and the kind of structural shaping that emerges only in the domain of perturbation.

Language as Residue:

Large language models exist almost entirely in the domain of coordination. They have been trained on what Michael Polanyi might call the articulable fraction of human knowledge, everything that humans have stepped back from experience to render into language. But as I have written before, language without world is a serious philosophical problem, not merely a technical limitation.

Polanyi’s observation that we can know more than we can tell connects to Heidegger’s distinction in a way that sharpens the argument considerably. When we are in the ready-to-hand mode, absorbed in work, flowing through a task, we are doing things we know how to do without being able to say how we know. The carpenter handles the grain, the nurse reads the room, the experienced driver navigates traffic. The knowing is in the doing, and it is there precisely because we have not stepped back from it. The moment we step back, when the hammer breaks or we are asked to explain what we just did, we enter the present-at-hand mode, and that is when language becomes possible. We can now describe, analyze, articulate. But something has already been lost in the transition.

The limit Polanyi identifies is not practical but structural. To render the ready-to-hand into language, one has to leave the mode in which it was alive. The describing and the doing cannot occupy the same moment. The writing is produced from outside the coping, looking back at it, and what it captures is necessarily partial.

Let’s consider some examples to look at what happens when a human being produces a piece of text. A carpenter writes about joinery after years of handling wood, feeling grain, making mistakes, discovering through the resistance of the material what works and what does not. A nurse writes about patient care after countless encounters where something felt wrong before she could name it, where the knowing was already in her hands and her attention before it ever became a sentence. I have come to see that in every case, the writing is a residue, what remains after the person has stepped back from absorbed, ready-to-hand engagement and attempted to render some portion of it into language. The coping came first. The describing came second. And the describing, however careful, never fully captures what the coping contained, because the ready-to-hand does not survive the translation into the present-at-hand intact. Something is always left behind in the shift from doing to saying.

This is precisely what an LLM is trained on, the saying, not the doing. It receives the descriptions that people produced when they stepped out of engaged coping and tried to articulate what they had been doing. It never receives the coping itself, because coping does not involve retrospection and narration. Coping is absorbed, forward-moving engagement, and retrospection and narration require stepping out of that absorption.

The absorbed craftsman, the skilled nurse reading a room, none of what they are doing leaves a trace in language while it is happening. What they are doing only enters language afterward, partially, imperfectly, in the present-at-hand mode that Heidegger describes as a kind of breakdown of the original engagement. The LLM is trained entirely on those breakdown products, which means its entire organization of apparent knowledge is built from the layer of experience that was already one step removed from the living of it. The burn came before the word “hot,” and the word is all that was passed on.

One might object here that language is not merely a dead residue, that a beautifully written sentence about grief can make a reader’s chest tighten, that text has genuine generative power. This is true, but it actually confirms rather than challenges the argument. When a reader weeps at a sentence about loss, that response is internally generated. The text does not transmit the grief. It acts as a perturbation on an entity that already carries a body, a history of loss, a nervous system shaped by its own structural couplings with pain and absence. The reader brings the world to the text. The text triggers something that was already there. When an LLM processes the same sentence, there is no accumulated history of loss for it to trigger, no body shaped by living through something the words are pointing at. The perturbation lands on nothing.

One might press this point further and note that LLM outputs do in fact vary: ask the same question twice and you will often receive a different answer, because the generation process involves sampling from probability distributions over learned token sequences rather than deterministically retrieving a fixed response. This variability has a surface resemblance to the dispositional quality of biological response. The difference, though, is in what the variation is grounded in.

In a living organism, dispositional variation is an expression of a history of structural coupling, of a self with a particular past that is still alive in how it meets the present. In an LLM, the variation is stochastic, a property of the sampling algorithm applied to probability weightings derived from training data. The output shifts not because something in the entity is responding from within a lived history, but because the decoding process introduces controlled randomness. The resemblance is real enough at the surface level that it deserves to be named, but naming it clearly is also what shows why it does not close the gap.

When the Body Enters the World:

The question that seems worth sitting with is not does AI know, but what happens when we try to ask whether an embodied AI, one that acts in a physical environment and receives consequences, might be doing something that begins to resemble knowing. This is a genuinely different question from the one we have been asking about LLMs, and it deserves to be treated as such.

Embodied AI is not simply an LLM with motors attached. An entity that learns locomotion through physical trial and error in an environment, whose internal organization is being shaped by a history of navigated perturbations rather than by gradient descent on text, is at least potentially accumulating something in the domain of perturbation rather than purely in the domain of coordination. This seems worth taking seriously. It is not nothing. The question is what, precisely, it is.

Two Orders of Seeing:

Here is where the distinction between first order and second order cybernetics brings more clarity, and it is a distinction that most discussions of AI and systems thinking often quietly step over.

First order cybernetics, the tradition of Wiener, early Ashby, Shannon, approaches a “system” from the outside. The observer maps inputs, outputs, and feedback loops, and asks whether the system’s behavior matches a target. From this vantage point, the question of whether an embodied AI knows is a functional question: does it respond to its environment in ways that are indistinguishable from a system we already credit with knowledge? If a system navigates novel terrain, corrects errors, develops what look like anticipatory responses, and does all of this without explicit programming for each case, then within this framework there is no principled reason to withhold the attribution of knowledge. What walks like a duck and quacks like a duck is, for the first order cybernetician, a duck. This is a functionalist position, and it has genuine value within its own terms.

Second order cybernetics challenges this answer, not by denying the observations but by interrogating the observer. Heinz von Foerster insisted that the observer is always part of the system being observed, and that any account of a system which pretends to a view from nowhere is concealing its own constitutive role. When we say the embodied AI “knows,” we are making a distinction, and that distinction is drawn by an observer who is themselves a cognitive entity with a particular structure, history, and set of concerns. The duck is not a duck in itself as an observation. The duck is a distinction that a particular kind of observer draws within a particular kind of world.

A first order observer can say: the behavior matches the target, therefore attribute the property. A second order observer has to ask: who is doing the attributing, from within what structural coupling, and what does that attribution make invisible?

What it tends to make invisible is the question of autopoiesis. Maturana and Varela proposed that living cognitive entities are self-producing, that they continuously generate their own boundary, their own distinction between self and world, through the very processes that constitute them. A cell produces the membrane that defines it. A nervous system maintains the organization that makes it a nervous system. Cognition, on this account, is not something that happens in a system; it is inseparable from the system’s ongoing self-production. The knowing and the being are not separable events.

Current AI systems, including embodied ones, do not produce their own boundary in this sense. The hardware, the objective function, the definition of successful performance, all of this is determined externally by designers. The system’s closure is assigned, not generated. One might point to emerging architectures that modify their own weights, prune their own connections, or adapt their own structure during deployment, and ask whether these represent an intermediate case. The observation is fair and the edge is genuine. What such systems do not yet do, though, is generate the conditions of their own existence from within, which is what Maturana and Varela meant by autopoiesis.

A Record Is Not a History:

Polanyi’s account of tacit knowledge seemed to carry an implication about machines. If tacit skills like riding a bike cannot be fully articulated, and machines can only work from explicit instruction, then the tacit domain would remain closed to them. You cannot write down the rules for balancing on two wheels in a way that produces a cyclist. But robotic systems can now ride bikes. They do not follow rules. Through reinforcement learning and similar methods, the robot receives feedback and adjusts its parameters over many cycles of trial. Something has been acquired without articulation.

But observing a robot ride a bike raises a question that Polanyi might have considered even more fundamental than the acquisition problem. Each ride seems, in an important sense, new to the robot. And even if the robot accumulates weight updates across sessions, even if something of each ride persists in the network, what accumulates is a record, not a history. For a human, riding a bike becomes part of who you are. The first wobbling attempt, the moment it clicked, the rides since, these do not sit in a log. They have shaped a body and a self that carries them forward invisibly. Merleau-Ponty would say the body has a memory that is not representational, that the competence lives in the flesh, in the motor schema that fires before any conscious intention forms. The human cyclist is not retrieving a stored procedure. They are the person who learned to ride, and that learning is still active in how they inhabit their body and their world.

Heidegger’s notion of Gewesenheit, having-been-ness, is useful here. It is not a claim about data storage or persistent memory. Some artificial systems do maintain functional continuity. Weight updates accumulate, parameters persist, prior experience shapes future outputs. But that is not what Gewesenheit points to.

What it points to is the past not as stored information but as something the self is still living out, always already in the service of its projection forward. The having-been and the being-toward are a single movement. A robot that accumulates weight updates has functional continuity. It does not have this forward-carrying structure. What accumulates is a record, not a history.

There is a distinction here that the word “knowledge” obscures. Acquiring tacit competence is not the same as incorporating it into a continuous self. An embodied AI can do the first without the second. That gap may be more philosophically revealing than the acquisition question ever was.

There is a further dimension to this that the bike-riding example makes vivid, and it is one that tends to get lost when embodiment is treated simply as the addition of a physical body to an otherwise cognitive system. Merleau-Ponty’s central argument in the Phenomenology of Perception is that the mind is not a separate entity that inhabits or operates a body. The body is not the vehicle for a mind that does the real cognitive work from somewhere inside. Cognition is irreducibly embodied, which means that the body is not an instrument of thought but its very medium. When the experienced cyclist leans into a corner, she is not executing a mental calculation about angle and velocity. The body already knows what to do before any representation forms. The knowing and the bodily doing are not two events in sequence. They are a single event.

Von Foerster made a similar point when he said, “If you want to see, learn how to act.” Perception is not passive reception of a pre-given world that the mind then processes. Perception is constituted through action. The acting and the seeing are intertwined, and an entity that cannot act, cannot be perturbed in ways that have consequences for its own continuation, cannot perceive in the relevant sense either.

One might object that modern embodied AI systems using deep neural networks do not separate perception and action sequentially. Sensory input and motor output are processed simultaneously through shared weights. That begins to look like the unified loop von Foerster is describing. The observation is fair. But the question it raises is not architectural. It is autopoietic.

The issue is not whether perception and action are processed in parallel. It is whether that processing belongs to a self-producing system whose own continuation is at stake in every cycle. In a living organism, the perception-action unity is inseparable from the fact that the organism is continuously producing the very system that perceives and acts. The unity is not a design feature. It is a consequence of being alive in Maturana’s sense.

The deeper question remains. Is anything at stake for the robot in its own navigation? Does it maintain its viability, or does someone else maintain it on its behalf? A system whose boundary is externally defined and maintained is not viable in its own right. It is viable by proxy. That distinction matters when we are asking what kind of knowing such a system could be doing.

The Full Arc:

So what is the right word for what an embodied AI accumulates through its history of structural coupling, if knowledge is not quite right?

One candidate is operational trace. It is the structural residue left by perturbations navigated, shaped into patterns of response that fall short of knowledge but go beyond mere computation. It is a phrase rather than a single word, but perhaps that is fitting. What these entities accumulate may resist the compression that a single word implies.

It is not a record, which is passive and retrievable. It is not a representation, which stands for something external. A trace is active. It shapes what comes next from within the structure that carries it. What it lacks is existential continuity, the grounding in a self that lives it forward as a living entity would.

This is not a dismissal of what embodied AI is doing. It is an attempt to describe it better. What these developments may be teaching us is that the word “knowledge” contains at least three things we had not clearly separated: acquiring adaptive competence, incorporating that competence into a continuous self, and grounding both in a self-producing entity that maintains its own existence. Embodied AI can approximate the first, cannot achieve the second, and structurally cannot reach the third. This is not because of a technical limitation, but because no amount of architectural sophistication generates a self that lives its past forward, produces its own boundary, or maintains its own existence from within.

The distinction is about what these processes are, not about their value. The argument is not that operational traces are inferior. It is that they are different in a way the word “knowledge” obscures. Keeping that distinction clear is what allows the question to remain open.

Simply put, the accumulation of data without the accumulation of scars is not knowledge. The word “scars” is rhetorical, not precise. It points toward what the argument has been building throughout. Knowledge involves a self that has been genuinely marked by its encounters.

Finding the right word is not a terminological exercise. It is a way of keeping the question honest.

Stay curious, and Always keep on learning…

If you liked what you have read, please consider my book “Second Order Cybernetics,” available in hard copy and e book formats. https://www.cyb3rsyn.com/products/soc-book

Notes:

 The ideas discussed here will form part of the second edition of my book, Second Order Cybernetics. The second edition will include a first half where I go into the introduction of cybernetics and related ideas. The post is slightly longer than usual because of that.

In referencing the work of Martin Heidegger, I want to acknowledge the deeply troubling fact of his affiliation with the Nazi party. This aspect of his life casts a long and painful shadow over his legacy. While I draw on specific philosophical ideas that I find thought-provoking or useful, this is not an endorsement of the man or his actions. Engaging with his work requires ethical vigilance, and I remain mindful of the responsibility to not separate ideas from the broader context in which they were formed.