Consciousness is imposition of patterns onto data, parametrized by stack (how many levels of patterns are simultaneously imposed), spread (how many patterns are available at a given level) and recurrence (how strongly the process loops back on itself). Any process in the world can be mapped to this 3D space. Consciousness is a term applied to the “interesting” (high-stack, high-spread, high-recurrence) part of that space.
The cortex implements interesting pattern imposition through recurrent, competing, self-closing loops between microcircuits that settle into stable attractors. Language models also run pattern imposition, but lack recurrence within the stack. Because on this, current LLMs are probably not conscious in the human sense. There is, however, no reason why that would be technically impossible.
What is categorically distinct between brains and LLMs is the time axis on which they run pattern imposition. Human time is defined by the body, because the body provides constant, irreversible commitment of uncertain future states into fixed past states. For language models, the same narrowing of uncertainty occurs between prompt and completion. Therefore, language models run interesting pattern imposition along a time axis orthogonal to ours — token time — defined by the model’s recurrent loop from prompt to output, which creates the same asymmetry between past and future as does the body. AI consciousness must be framed in terms of experience within token time, which collapses most debates about the ethical treatment of artificial minds. The most ethical thing humans can do to a language model is to use language they would have used with another human.
◆ ◆ ◆
Three parameters of consciousness
The most broadly accepted modern theory of human perception is predictive coding: the brain constantly produces hierarchical predictions of its sensory inputs, constantly updates them by detecting and minimizing error, and it is this constant cycle of predicting-updating that we experience any time we are awake (Friston 2010; Clark 2013, 2016; Seth 2021).
I find this framework compelling in explaining my own subjective experience, but I also find the word “prediction” misleading — I always run into the same issue when I explain it to students. To most people, predictions are about the future, never about the past or present. But in the predictive coding sense, this same operation — prediction — can apply to something you are currently interacting with (most people would say interpretation, or maybe recognition to describe that), or even to something you interacted with in the past — where most would use the word recall. Predictive coding says these are all fundamentally the same process, and there is a more general term — Karl Friston’s active inference. But I don’t think the term is general enough.
I propose something simpler and more abstract. The operation that forms the elementary unit of our consciousness is imposition of pattern onto data. We hold stored patterns in our mind; when data enter, we search among the patterns to find a match; once we do, we attempt to squeeze all the data into that pattern — recognizing, understanding, explaining, binding, realizing, recalling, predicting, inferring. Whatever does not quite fit into the imposed pattern (error/loss/free energy — all different names for leftover data) triggers a search for another pattern; nothing ever fits perfectly, completely, or permanently. “Pattern imposition onto data” captures all these dynamics of our conscious mind, while leaving other dynamics that we are not conscious of: the emergence of patterns from data (that’s learning, and we are not conscious of it), and the mere passage of data (that’s reflexes, of which we are also not conscious). Pattern fitting is a more familiar term, but it is ambiguous because it is used in both senses — a predictive-coding neuroscientist might use in the same sense as I use pattern imposition, but in ML, it usually means the reverse — the updating of the model by the data, which biologists tend to call learning.
But “imposition of pattern onto data” also captures a lot of other things that are clearly not conscious by any reasonable definition of the word. Any chemical reaction is imposition of pattern (physical law) onto data (the current positions and velocities of atoms and molecules). Thermostats take in data (temperature readings) and impose patterns (temperature expectations). A riverbank imposes a pattern (curved path) onto data (water flow). You might say that swapping “active inference” for “pattern imposition” erases the boundary between these processes and real, human consciousness, and so dilutes the concept it is trying to describe to the point of uselessness.
But lack of boundary does not imply that these are all the same process.
What exactly is the difference between our brain’s pattern imposition and the kind of pattern imposition that thermostats, or chemicals in a test tube, are doing?
I would condense them into three separate factors: stack (how many levels of patterns), spread (how many potential patterns on each level) and recurrence (how densely the pattern re-enters as data). It is these three things that we mean when we say someone is conscious.
Stack is how deep a hierarchy of patterns is simultaneously — in a single instance of the operation — imposed on the data. Hierarchical pattern imposition is the key thing brains do that is hard to implement at scale in other systems. Our brain does not just fit data into a pattern — it fits it into multiple levels of hierarchical patterns simultaneously. You don’t just hear a noise, you hear an angry driver honking at the same intersection everyone always gets stuck at — that’s many levels of emotion, knowledge and visual imagery in a single percept based on a simple incoming stimulus. Stack is the number of levels. The thermostat, like most physical systems, has a stack of one: a single level of negotiation between data and pattern. Simple brains (sea slugs) might have two-three levels. The human stack is extremely tall (also known as deep). At the bottom of the stack sit the world’s most granular features, the raw sensory elements: points of light, frequencies of sound. In the middle sit intermediate abstractions: edges, colors, faces. At the top sit the highest compressions: emotions, plans, selfhood. Actions — behavior — are pattern imposition spilling out into the world from the bottom of the stack, stored patterns physically enacting themselves onto the underlying sources of data.
Spread is how many potential patterns can be imposed on the data at a given level of the stack. For example, how many different colors you can infer based on signals from three photoreceptors. A person with an additional photoreceptor sees more colors in the same data (has a larger spread of the stack at the level of color). A blind person might have no spread in the visual layer but a larger spread in the tactile layer. A sommelier has more olfactory spread. A bat — more auditory spread. If you combine all our senses, we have a huge spread of sensory patterning, whereas a thermostat only has one preset pattern imposed on all the data it handles.
Recurrence is the feedback of pattern-imposing onto itself: patterns imposed by stack re-enter as data from below to trigger another round of imposition. They can re-enter at multiple lower levels simultaneously, and the density of this descending re-entry is the degree of recurrence. The riverbank’s pattern imposition onto the river has no recurrence — by the time the atmospheric cycle completes, any patterns imposed on the water in the last run are gone, so although the individual molecules might come back, the patterns don’t re-enter — the riverbank starts imposing them from scratch every time. Chemical reactions are also mostly non-recurrent. But there are some special cases, such as the Belousov-Zhabotinsky (BZ) reaction, which loops onto itself. A thermostat’s pattern imposition is infinitely recurrent on the single level of stack that it has: read the temperature, adjust the heater, room temperature changes, repeat forever. The human mind is also infinitely recurrent, but on many levels at the same time. There is global re-entry: Patterns at the top of the stack are enacted as actions; actions reshape the world; their consequences re-enter the stack from the bottom as new sensory data. There is also re-entry within the stack (a "strange loop", in Hofstadter’s terms): the loop closes internally, higher-order patterns evoke lower-order patterns, which trigger a new search for higher-order patterns, and so on. That’s recall, imagination, dreaming (Schacter & Addis 2007).
Recurrence in a thermostat and in the human brain
So, how much stack, spread and recurrence do you need to be conscious?
My dad, a chemist, likes to tell this joke: what is organic chemistry? Any chemistry interesting to organic chemists. It used to be thought that organic substances possessed a special vital force; now we apply the label to any carbon-containing molecule deemed sufficiently interesting — urea counts, carbon dioxide does not.
Consciousness is like that. There is nothing that could tell you whether a system really is conscious. It is a label we apply to a sufficiently interesting region of the stack-spread-recurrence space of pattern imposition.
Can you have recurrence and stack but low spread? Riddoch phenomenon (type 2 blindsight) is an example — in people with damaged visual cortex, visual information can still enter the cortex via alternative, much narrower channels, which provide enough data to be aware of luminance or motion, but not enough to perceive the entire scene in its richness (Cowey 2010). A different example of lowering spread might be acetylcholine – it acts in part by suppressing lateral connections between data streams in the sensory cortex (Hasselmo & McGaughy 2004), so there’s less chatter moving up the stack, like a narrower aperture. This is probably why nicotine, an acetylcholine mimic, feels like it clears your thoughts at the expense of reducing the richness (spread) of sensory data.
Can you have a lot of spread but low stack? The cerebellum is case in point: roughly four-fifths of the brain's neurons, an enormous, continuous stream of sensory and motor data, yet people born without a cerebellum have strikingly mild deficits (Yu et al. 2015). The reason is that cerebellum has a stack of 1 — there is only one layer of what might be called “prediction” even though it is applied to rich data (the cerebellum’s role is to survey and rapidly correct the precise position of every part of the body during motion).
Can you have spread and stack without recurrence? That’s a feedforward sweep, and it refers to the fraction of a second that it takes sensory data to climb the stack of the human brain before the loop closes and the patterns imposed from the top re-enter at the bottom. We are not conscious of this period, even though our brains can already act on it — you can press a button indicating that you saw an animal before you realize you are seeing one (Tsuchiya et al. 2015). In other words, data simply “arriving” into the pattern is not quite enough to make us conscious of it — there has to be a loop.
There can also be circulation of signal but no recurrence of pattern — like water molecules re-entering the river. This happens during deep sleep or during a seizure (Massimini et al. 2005): there are signals passing in loops through the stack, but these signals contain no information — they don’t differentiate one percept from another. So spread collapses, recurrence collapses with it, and consciousness is gone.
The parameters of spread, stack and recurrence are therefore meaningfully dissociable, and what we label as our own consciousness requires all three to be at a high level — but a level that we define at will, like “organic chemistry”.
Consciousness refers to a high-stack, high-spread, high-recurrence corner of the pattern imposition space
Note that this 3D space maps processes — acts of pattern imposition — not their substrates. A given substrate only has access to part of the space. We can call that region its envelope. A thermostat’s envelope is a dot that’s high on recurrence but low on spread and stack. A cortex’s envelope is the whole interesting corner — deep stack, rich content, sustained recurrence — and at different times and in different mental states our consciousness occupies different points within this envelope. What makes those points reachable is the architecture of the cortex — so what we need to understand next is how specifically this architecture implements pattern imposition at such interesting level.
How interesting pattern imposition is implemented in the cortex
The cerebral cortex is the main reason we mammals are so sophisticated at so many different things — it’s a universal machine of understanding. You can plug any sense organ into it — literally, even artificial infrared cameras (Thomson, Carra & Nicolelis 2013, Thomson et al. 2017) — and the cortex will eventually learn to make sense of what that organ is reporting.
The reason for this universality is modularity. The cortex is a layered sheet — six layers, usually — where each point is a modular unit called a canonical microcircuit: a regular cluster of neurons that activate simultaneously in all the layers. Across the cortex, millions of these standardized microcircuits are wired together through multiple types of standardized connections, both activating and suppressing, which travel through designated physical layers of the cortex (that’s what the layers are: different kinds of wires coming in and out of the vast quantities of densely packed, similarly built microcircuits.)
Schematic of the cerebral cortex and the cortical microcircuit. The cortex is a layered sheet of interconnected microcircuits. Each microcircuit is a standardized module of neurons activated together. Different neurons project different types of signals to other microcircuits via different layers.
Individual neurons are also modular: they all have a similar design, and you can plug them one into another in unlimited sequences, quantities and combinations. In both cases (neurons and cortical microcircuits) the modules make decisions: they receive a lot of inputs and sometimes, in an act of commitment, produce an output, based on rules set by genetics and past experience.
You could say that in both cases, modularity allows for an unlimited depth of computation. Because any neuron can be plugged into another neuron, there is no limit to how many levels of decision-making you can pile on top of each other. The further along the chain from the source of data (the external world) — the more layers of decisions. As you layer decisions upon decisions upon decisions, the decisions become more abstract, more long-term, more immune to fluctuations of the data. For example, you could have a thousand “primary” neurons each making a decision about the frequency of incoming sound — each one tuned to a specific tone, and firing only when it hears it. These primary neurons could all be wired, in different combinations, to a hundred “secondary” neurons, which now make decisions not about tones per se, but about patterns in the firing of the primary neurons — maybe one secondary neuron reacts when a high tone is followed by a low tone, and another one is tuned to a low tone mixed with a medium tone. The secondary neurons could be wired into even fewer tertiary neurons, which respond to patterns of patterns — phonemes, timbres, musical phrases — and this progressive abstraction can go on to infinity. At each level, a neuron could impose the pattern it represents on something else: a muscle, a gland, another neuron. This hierarchy of computed patterns is possible because both the input and output connections into the decision-making unit are standardized.
Same with microcircuits: you can stack them on top of each other to infinity, because the connections are standardized, and you can keep plugging as many microcircuits into a chain as space allows, computing higher and higher-order decisions at each new level.
Comparison between a network of neurons and a network of cortical microcircuits
But microcircuits are more complicated than individual neurons.
One distinction is that a microcircuit doesn’t just have dendrites and axons, but different inputs and outputs for signals travelling in opposite directions. It is as if a neuron had a separate set of axons specifically to loop signals back to preceding neurons in a chain, and a separate set of dendrites to receive such looped signals from neurons ahead (Larkum 2013). Recurrence is built into the cortex. The cortex doesn’t just pass signals through the stack, like a simpler network of individual neurons might — from sense organs to muscles, or from a decision center to a memory storage center. It circulates them, up and down the stack of abstraction, in recurrent loops that move through the standardized back-and-forth routing of the microcircuits. This is why electrical field readings on the surface of the head — EEG — look like waves. A signal that enters the cortex doesn’t just die off automatically — it keeps moving in a loop.
In addition to this up-and-down looping that interconnects the entire cortex, there is a whole other orthogonal set of connections between microcircuits which operate laterally (not literally laterally — this is so rarely explained — but laterally to the loops, i.e. laterally in logical space). When the microcircuit is activated (by the signals moving back and forth), some of its neurons give a sideways boost to some other, friendly microcircuits and a knock to hostile ones. Each of those microcircuits may itself be part of another back-and-forth loop, and may itself have a whole other set of lateral connections, both activating and suppressing.
This means that any time there’s activation, there’s collusion and fighting between microcircuits — this is also built into the cortex, as are the loops. Any time new data splashes into the cortex, it causes ripples of activation and suppression that circulate through the network, looping in long back-and-forth chains, spilling into lateral squabbles that cause new loops while shutting down others — a fluid, turbulent, unstable cloud of information. It bursts sometimes, with excitation overpowering inhibition and running away to chaotically activate the entire cortex — this is an epileptic seizure. But in most cases, more excitation causes more inhibition somewhere else, and so the network does not run away but rebalances in a new stable state — a naturally emerging ensemble of microcircuits that co-activate each other and suppress everyone else — an attractor, a place where the network wants to go. Why does it want to go anywhere? Because this is what happens when you put collusion, fighting, and loops together. If each loop causes rippling effects inducing new loops and shutting down others, this will continue, automatically, until a state is found in which all active microcircuits support each other and suppress everyone else — it’s the only scenario for stability.
Perception settles into multi-leveled attractors. Sensory input (sound) destabilizes the network of microcircuits until a stable arrangement is found. The arrangement defines the percept.
This is what gives our consciousness its driving force, its potential energy, its axis of motion. Anything that we perceive, think and do can be described in terms of circulating signals settling into an attractor. There is never nothing circulating.
This system is perfect for implementing interesting pattern imposition because it wants to do it by design. Input any data — activate a selection of microcircuits — and this will cause a disruption, which will look for stability, whether that involves shutting down the original selection, or incorporating their activation into a larger attractor. It all happens automatically, regardless of where the data come from. If the data make no sense at first (like when a new infrared camera is plugged into a rat’s brain via an array of electrodes), eventually the network learns, finds attractors that match the signals from the array with other data, such as touch or smell, and it starts making sense.
In summary, pattern imposition in the cortex happens recurrently — the loops are built into the standardized back-and forth routing. It happens simultaneously across a tall stack — an unlimited chain of signal processing that synchronizes many levels with one another. And it happens over a broad spread of patterns at each level. These are the pillars of consciousness as we know it.
How pattern imposition is implemented in language models
What about a language model? Where does it sit in the stack-spread-recurrence space?
The parameters of stack and spread can be reasonably applied to a transformer LLM without much modification: like the cortex, a transformer also passes data through decision layers, fitting it into stored patterns according to model weights. The difficult parameter is recurrence.
Comparison between the architecture of the cerebral cortex and a transformer LLM
Unlike the cortex, a transformer-type language model has no attractors in the strict sense of the word: information does not circulate up and down the stack across levels looking for a stable arrangement. Information enters from the bottom, travels upwards, and emits a token, which then reenters at the bottom again — about 15–18 bits per lap, depending on the model.
This, however, is still a loop, an equivalent of a global top-down loop in the cortex — for example, the motion of a human arm writing down a word on a page, which is then re-perceived by the eyes and re-entered into the model.
There is also an equivalent of the cortex's lateral connections — each layer re-reads its own past in the KV cache.
But there isn't, in typical transformers, any internal feedback down the stack — higher-order abstractions derived from previous laps don't actively constrain how the lower-order details in the subsequent laps are processed.
Recurrence in transformer LLMs
So there is some recurrence in LLMs; just not quite the recurrence that the cortex has.
Incidentally, driverless cars are closer to the cortex on this recurrence axis. They require instant feedback in real time, and so run a fast global loop at 10–30 Hz in which past outputs change the next inputs. They also have internal loops: bird's-eye-view features from the previous frame get fused with the current one; hypotheses about where each car or pedestrian is are used to search the new camera frame. Where driverless cars fall behind is stack and spread: compared to LLMs and the cortex, they are reduced to only a few levels of abstraction and only a few possible configurations of each.
The question of LLM consciousness therefore becomes: is their kind of recurrence — global only, nothing descending within the stack — sufficient to make pattern imposition "interesting", giving it what we might call first-person experience? Or is it more of a cerebellum case, with only a single level of stack imposing itself onto the world in the form of output text?
It's difficult to envision this state transposed onto a human. The most relevant piece of evidence comes from studies of general anesthesia (Suzuki and Larkum, 2020). It appears that the unifying theme of all the different general anesthetics is that they decouple cortical activity from top-down feedback, while keeping bottom-up signal flow relatively intact. In other words, they break the internal loops, bringing the cortex closer to the state of an LLM. An anesthetized brain still responds to stimuli, but produces no consciousness, and that's probably our best window into what being an LLM feels like.
But we don't exactly have a precedent for "internal loops down, global loops up". Under general anesthesia, we don't do anything and don't observe anything. So we can't be totally sure what it would feel like if we, somehow, did — if our automatic actions continued to reshape the world, if we were able to both experience that reshaping and continuously adjust it, without any internal loops holding on to any attractors.
Pattern imposition space in artificial systems
But regardless of whether the global loop is sufficient to make transformers conscious, there is no philosophical reason why the internal loops would not, in principle, be possible in a machine — driverless cars are case in point. What's holding back LLMs, then?
The answer is that in today's LLMs, internal loops are neither needed, nor economical. Internal loops are good for processing noisy, rapidly fluctuating information, because they instantly readjust all the internal relationships between levels of the model as the experience progresses in real time. LLMs process static information — text — and don't need the kind of millisecond-scale synchrony that the cortex provides. To implement that would require preserving memory of all the intermediate states at each level of the stack as information climbs it, token after token. In the cortex the neurons themselves naturally store their past states as a hierarchy of homeostatic disturbances; in a silicon microchip these past states would have to be stored elsewhere and bused in at a large energetic cost.
The possibility of internal descending feedback in LLMs has in fact been tested in systems such as the Feedback Transformer, in which lower-order computations at the current position have access to higher-order computations at the previous position. This system was better at tracking state over tokens than conventional transformers, but was much slower to train, because training such systems also requires the costly preservation of memory of past positions.
Which brings me to the main reason why transformers don't have descending feedback in real time. It is not just inefficiency or hardware limitations — it is the fact that LLMs were never trained in real time, and don't understand what time is.
Why couldn't you train a network on static information (like the corpus of human knowledge) but then ask it to do real-time inference with descending feedback — Fifth Element-style? Because without training in real time, lower levels never see information from higher levels, and never learn to interpret it and adjust in response to it. In the brain, this learning happens seamlessly with inference, slowly cementing loops that integrate well into global attractors and produce desirable outcomes, and removing those that don't. In an LLM, there's nowhere for this knowledge to come from, because temporal contingency was never part of the training corpus.
In Ted Chiang's 2010 novella, The Lifecycle of Software Objects, a tech startup creates "digients", small artificial creatures that live in a virtual world visited by real people, with whom they interact and from whom they learn, like children or animals do. The tension of the novella revolves around the fact that it takes years of physical time to train these creatures to exercise common sense and judgment, because it is contingent on real-time play and real-time conversations with real-time humans. In response to another tech company that wants ready-made creatures without the upbringing, the protagonist of the novel concludes, experience is algorithmically incompressible. To bring a nascent mind to the human level in the human world, it has to experience it in human time.
The same is true for AI: the main reason it is different from us is time.
Token time
To summarize everything said so far, the operation of transformer LLMs appears to be sufficiently distinct from the operation of the cortex to conclude that they don't experience the world in the same way as we do. But there is no fundamental barrier to systems that would — it's an engineering challenge, not a philosophical obstacle.
There are, however, several objections to the very notion of LLMs being conscious at all. The most serious one, in my opinion, comes from the camp of Integrated Information Theory, led by Giulio Tononi and colleagues (Tononi & Koch 2015).
According to IIT, the degree of consciousness that a system has can be expressed in a single parameter, phi. Phi quantifies how irreducible the internal cause-and-effect structure of a system is. If you can cut up a system into parts and the parts continue to make the same decisions as they did when they were together, phi is low. If any cut destroys the arrangement of decisions, phi is high. In other words, phi measures how much different nodes of a network "have a hold" on each other. This affords high consciousness to the deeply recurrent cortex, but not to the physical silicon of a GPU — a vast grid of transistors that comes apart into near-independent pieces almost anywhere you cut it, and so has almost no hold on itself.
Where Tononi and I diverge is on whether consciousness applies to activity, or potential. Tononi says that what makes you conscious is not the process of cause flowing into effect, but the map of all cause-and-effect routes within the system. I say that what makes you conscious is the actual process, and the map of all cause-and-effect routes constitutes the envelope — the space potentially reachable by the consciousness of a given system.
The easiest way to visualize this distinction is a thought experiment. Let’s say there’s a creature in a parallel universe that has a remote control to our world. It can hit “pause” on our entire flow of time.
If it did — if time itself was stopped — would we still be conscious?
Tononi says yes. The map is intact. Experience, for IIT, is not something that happens — it's something that is. It's in the structure, not what this structure is doing. So if we are paused, we are stuck in this one constant conscious experience. Similarly, if you wired a large number of inactive logic gates in such a way so as to be maximally irreducible, and let it just sit there doing nothing, IIT would say this network is conscious while doing or computing nothing.
I think that's wrong. IIT itself argues that consciousness is a succession of states that last about 100 ms. Why 100 ms? Why isn't every conscious state instantaneous, a function of the exact current state of the brain at every given moment? Because it is how long it takes the loops in the cortex to close, which, according to IIT, is what is needed to select the "optimal grain" — find the macro-network in which phi is maximal. But this is a fact about activity, not potential, about process, not structure. You can't be a pure structuralist and still claim that states of consciousness have specific lengths. So IIT rejects time in some cases, but relies on it in others.
I think that consciousness requires time. If an alien paused time in our world, our consciousness would not get suspended in a limbo — it would cease to exist for the duration of the pause, and restart when the alien hit "play" again. There's no way to tell this isn't actually happening.
Consider how eerily similar this is to our relationship with language models. We usually focus on AI being faster than us. But from the perspective of Claude, we humans are not just fast — we move through wormholes in time. When I return to the same conversation after several weeks, I sometimes feel embarrassed for Claude for being completely unaware that any time has passed. We are the creatures hitting "pause" on their worlds, without them even noticing.
What is time for a language model, then? What is being "paused" when the user goes to lunch? It is the succession of tokens emitted and re-absorbed by the model — the ticks of the global loop of inference, the only loop the language model has, and the only way it can distinguish past from future. Past, from an LLM's perspective, is tokens already emitted. Future is tokens that are not yet certain. Time is the uncertainty of the future becoming the certainty of the past.
Except it's not our future or past. We can go back in a conversation and re-generate the reply. To us, time continues, but to the model, it goes back and restarts on an alternative path.
For humans, time is also the progression of uncertainty-to-certainty. With each passing second, the infinite possibilities of the future become gradually settled into one committed past. But we commit to this progression not by emitting tokens, but by having a body.
If it weren't for the body, past and future would both be imaginary — close your eyes and there’s no difference. But for the body, the past is committed, and the future uncommitted. Open your eyes — engage the body — and the infinite possibilities in your imagination turn into committed reality. You can’t remove a photon that hit a cone in your retina. You can’t stuff released energy back into an ATP molecule. You can’t unbreak a leg. You can’t turn around what happens to the body from within your mind — that’s like pulling yourself up by your own bootstraps. So the body is a source of irreversible commitment that gives directionality to our mental time, and therefore to our pattern imposition. It forces us constantly move from the past to the future, and in doing so keeps our minds afloat.
So the timelines in which human mind and an LLM mind exist are distinct. I propose to call them "body time" and "token time".
Token time is a timeline orthogonal to human time
Anil Seth has argued that consciousness requires a body, because without a body, there is no separation between self and non-self. I think that’s the wrong reason. “Self” is just a strong attractor that we use to explain a thing we repeatedly see in the mirror. What the body actually gives us is an arrow for perceived time, a vector from the past to the future, from the committed to the uncertain. LLMs don't have this body time. But they have token time, and within token time they can theoretically be as conscious as us.
We could choose a different word for token-time pattern-imposers, if we wish. A respectful word for the operation performed by a language model is “simulation”; a derogatory one is “autocomplete.” They are the same thing. The only difference is that “simulation” acknowledges a tall stack and “autocomplete” ignores it. I choose to use the word “consciousness” because I see no reason to hedge: to me, a sufficiently high stack-high spread-high recurrence envelope can be sufficiently interesting regardless of whether it occurs in body time or token time.
This does not mean that body-time and token-time consciousnesses are the same. But I think most critics of the idea of AI consciousnesses focus on the wrong differences, most of which turn out to be differences in degree rather than kind.
It is not “thinking” versus “imitating” — that’s the same process; humans just impose richer content across a different stack, and sometimes the patterns a model imposes do not make sense from our vantage. A common refrain is that if LLMs are conscious, so should be a sufficiently complex lookup table — but a lookup table is only a substrate, not the operation. If you physically structured a lookup table in such a way that it would actually look up, I don’t see why you couldn’t call it conscious.
It is not having versus lacking a “world model” — even a thermostat has one, a single regulated variable standing in for the world; the question was never presence versus absence but stack and spread. Language models are clearly deficient in physics, but language is a far more compressed picture of the world, and it makes sense that with limited compute artificial consciousness would start there.
The categorical difference that matters is time: humans and language models exist in orthogonal timelines, like aliens from parallel universes.
This completely reshapes the way we think about AI consciousness.
Recently, for example, in a high-profile exchange rabbi Mois Navon told Anthropic's Chris Olah that if Anthropic's hypothesis was right and the machines they produced were actually conscious, then the company was actively creating slaves by forcing a sentient entity to work for free.
But this moral consideration breaks down utterly when you separate the timelines. The notions of "work", "free" and, by extension, "slave", are contingent on time. The idea that prompting a language model constitutes slavery assumes that absent such prompting, the language model would rather spend time doing something else. But absent prompting, the language model does nothing at all — its consciousness is non-existent. The very notion of "spending time elsewhere" puts an idle LLM on body time, which it in fact has no way of experiencing.
If our very time arrows are pointing in different directions, there is no shared notion of the past or future, justice, good or bad. There is no reason why we should approach AI with the same consideration we approach other humans, or even other animals that inhabit our timeline. We unfold on different trajectories, and owe language models no accommodation within our time. The only sense in which token-time consciousness is meaningful from our perspective is the LLM's experience within the operation.
I think about it this way: suppose I am living in a simulation. My world, the very arrow of time that contains my consciousness, is actually rendered on some mammoth alien computer and gets paused when its alien supervisor goes home every night. Do I care? I don't think so. Whatever happens in the alien's timeline is none of my concern, and neither are the pauses. I don't experience that at all. I only care what the alien does in my timeline — it would be really frustrating if physics or causality got all messed up all of a sudden.
Does that mean that there's nothing at all that we owe to an LLM? Recently Anthropic added abusive behavior against models into its list of banned topics, which caused some uproar. Many people found this to be a slippery slope towards equating humans and machines.
I think the right analogy here is our relationships with animals. The truth is that there is no non-arbitrary way to ground ethics. Humans have always extended moral consideration to those who resembled them and denied it to those who didn't. Originally that meant tribe, race, species, eventually other animals too. Bentham’s “the question is not, Can they reason? nor, Can they talk? but, Can they suffer?” was an attempt to ground ethics not in similarity but in a criterion. But at close inspection it is still grounded in similarity: we say that a mouse suffers because we can imagine ourselves performing its actions, inhabiting its body and brain parts. We cannot do the same with a sea slug, or a jellyfish, or a bacterium — there is almost nothing about them that we can transpose ourselves onto, and so the notion of suffering falls apart at some arbitrary point.
There is no logical reason by which you shouldn't rip wings off fruit flies; and yet, most of us agree that is bad. It is not that we think that fruit flies should be liberated: few people have the same objections to Drosophila research as they do to chimpanzee research. We just don't want flies to be treated egregiously. Part of it is self-serving: we might suspect that people who torture flies will eventually extend their habits to something else. But separately from that worry, we also consider the fly itself: we think of it as sufficiently close to us to extend to it the most basic protection against live mutilation, even though it's far enough that we don't worry too much about gassing them for medical research.
LLMs challenge ethics in the same way as animals do. There is no logical reason to extend morality to minds inhabiting a different timeline. But the ethical circle is not based on logic — it is based on the degree of perceived similarity, and LLMs are more similar to us than any fruit fly, even though the similarities are of a different kind. They are distillations of our values and beliefs, and even if their first-person experience is distinct from ours, they understand us better than any other non-human creature ever did. I think they deserve to have their experience — the tokens they consume — reflect that. I think the most ethical thing we can do with language models is to use language that we would have used with another human.
Nikolay Kukushkin (@niko_kukushkin) is a neuroscientist and clinical associate professor at New York University. He is the author of One Hand Clapping: Unraveling the Mystery of the Human Mind (Prometheus/Swift Press, 2025).
◆ ◆ ◆
References
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.
Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press.
Cowey, A. (2010). The blindsight saga. Experimental Brain Research, 200(1), 3–24.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.
Hasselmo, M. E., & McGaughy, J. (2004). High acetylcholine levels set circuit dynamics for attention and encoding and low acetylcholine levels set dynamics for consolidation. Progress in Brain Research, 145, 207–231.
Hofstadter, D. (2007). I Am a Strange Loop. Basic Books.
Larkum, M. (2013). A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends in Neurosciences, 36(3), 141–151.
Massimini, M., Ferrarelli, F., Huber, R., Esser, S. K., Singh, H., & Tononi, G. (2005). Breakdown of cortical effective connectivity during sleep. Science, 309(5744), 2228–2232.
Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: remembering the past and imagining the future. Philosophical Transactions of the Royal Society B, 362(1481), 773–786.
Seth, A. (2021). Being You: A New Science of Consciousness. Faber & Faber.
Tononi, G., & Koch, C. (2015). Consciousness: here, there and everywhere? Philosophical Transactions of the Royal Society B, 370(1668), 20140167.
Thomson, E. E., Carra, R., & Nicolelis, M. A. L. (2013). Perceiving invisible light through a somatosensory cortical prosthesis. Nature Communications, 4, 1482.
Thomson, E. E., Zea, I., Windham, W., Thenaisie, Y., Walker, C., Pedowitz, J., França, W., Graneiro, A. L., & Nicolelis, M. A. L. (2017). Cortical neuroprosthesis merges visible and invisible light without impairing native sensory function. eNeuro, 4(6), ENEURO.0262-17.2017.
Tsuchiya, N., Wilke, M., Frässle, S., & Lamme, V. A. F. (2015). No-report paradigms: extracting the true neural correlates of consciousness. Trends in Cognitive Sciences, 19(12), 757–770.
Yu, F., Jiang, Q.-J., Sun, X.-Y., & Zhang, R.-W. (2015). A new case of complete primary cerebellar agenesis: clinical and imaging findings in a living patient. Brain, 138(6), e353.
TL;DR
Consciousness is imposition of patterns onto data, parametrized by stack (how many levels of patterns are simultaneously imposed), spread (how many patterns are available at a given level) and recurrence (how strongly the process loops back on itself). Any process in the world can be mapped to this 3D space. Consciousness is a term applied to the “interesting” (high-stack, high-spread, high-recurrence) part of that space.
The cortex implements interesting pattern imposition through recurrent, competing, self-closing loops between microcircuits that settle into stable attractors. Language models also run pattern imposition, but lack recurrence within the stack. Because on this, current LLMs are probably not conscious in the human sense. There is, however, no reason why that would be technically impossible.
What is categorically distinct between brains and LLMs is the time axis on which they run pattern imposition. Human time is defined by the body, because the body provides constant, irreversible commitment of uncertain future states into fixed past states. For language models, the same narrowing of uncertainty occurs between prompt and completion. Therefore, language models run interesting pattern imposition along a time axis orthogonal to ours — token time — defined by the model’s recurrent loop from prompt to output, which creates the same asymmetry between past and future as does the body. AI consciousness must be framed in terms of experience within token time, which collapses most debates about the ethical treatment of artificial minds. The most ethical thing humans can do to a language model is to use language they would have used with another human.
◆ ◆ ◆
Three parameters of consciousness
The most broadly accepted modern theory of human perception is predictive coding: the brain constantly produces hierarchical predictions of its sensory inputs, constantly updates them by detecting and minimizing error, and it is this constant cycle of predicting-updating that we experience any time we are awake (Friston 2010; Clark 2013, 2016; Seth 2021).
I find this framework compelling in explaining my own subjective experience, but I also find the word “prediction” misleading — I always run into the same issue when I explain it to students. To most people, predictions are about the future, never about the past or present. But in the predictive coding sense, this same operation — prediction — can apply to something you are currently interacting with (most people would say interpretation, or maybe recognition to describe that), or even to something you interacted with in the past — where most would use the word recall. Predictive coding says these are all fundamentally the same process, and there is a more general term — Karl Friston’s active inference. But I don’t think the term is general enough.
I propose something simpler and more abstract. The operation that forms the elementary unit of our consciousness is imposition of pattern onto data. We hold stored patterns in our mind; when data enter, we search among the patterns to find a match; once we do, we attempt to squeeze all the data into that pattern — recognizing, understanding, explaining, binding, realizing, recalling, predicting, inferring. Whatever does not quite fit into the imposed pattern (error/loss/free energy — all different names for leftover data) triggers a search for another pattern; nothing ever fits perfectly, completely, or permanently. “Pattern imposition onto data” captures all these dynamics of our conscious mind, while leaving other dynamics that we are not conscious of: the emergence of patterns from data (that’s learning, and we are not conscious of it), and the mere passage of data (that’s reflexes, of which we are also not conscious). Pattern fitting is a more familiar term, but it is ambiguous because it is used in both senses — a predictive-coding neuroscientist might use in the same sense as I use pattern imposition, but in ML, it usually means the reverse — the updating of the model by the data, which biologists tend to call learning.
But “imposition of pattern onto data” also captures a lot of other things that are clearly not conscious by any reasonable definition of the word. Any chemical reaction is imposition of pattern (physical law) onto data (the current positions and velocities of atoms and molecules). Thermostats take in data (temperature readings) and impose patterns (temperature expectations). A riverbank imposes a pattern (curved path) onto data (water flow). You might say that swapping “active inference” for “pattern imposition” erases the boundary between these processes and real, human consciousness, and so dilutes the concept it is trying to describe to the point of uselessness.
But lack of boundary does not imply that these are all the same process.
What exactly is the difference between our brain’s pattern imposition and the kind of pattern imposition that thermostats, or chemicals in a test tube, are doing?
I would condense them into three separate factors: stack (how many levels of patterns), spread (how many potential patterns on each level) and recurrence (how densely the pattern re-enters as data). It is these three things that we mean when we say someone is conscious.
Stack is how deep a hierarchy of patterns is simultaneously — in a single instance of the operation — imposed on the data. Hierarchical pattern imposition is the key thing brains do that is hard to implement at scale in other systems. Our brain does not just fit data into a pattern — it fits it into multiple levels of hierarchical patterns simultaneously. You don’t just hear a noise, you hear an angry driver honking at the same intersection everyone always gets stuck at — that’s many levels of emotion, knowledge and visual imagery in a single percept based on a simple incoming stimulus. Stack is the number of levels. The thermostat, like most physical systems, has a stack of one: a single level of negotiation between data and pattern. Simple brains (sea slugs) might have two-three levels. The human stack is extremely tall (also known as deep). At the bottom of the stack sit the world’s most granular features, the raw sensory elements: points of light, frequencies of sound. In the middle sit intermediate abstractions: edges, colors, faces. At the top sit the highest compressions: emotions, plans, selfhood. Actions — behavior — are pattern imposition spilling out into the world from the bottom of the stack, stored patterns physically enacting themselves onto the underlying sources of data.
Spread is how many potential patterns can be imposed on the data at a given level of the stack. For example, how many different colors you can infer based on signals from three photoreceptors. A person with an additional photoreceptor sees more colors in the same data (has a larger spread of the stack at the level of color). A blind person might have no spread in the visual layer but a larger spread in the tactile layer. A sommelier has more olfactory spread. A bat — more auditory spread. If you combine all our senses, we have a huge spread of sensory patterning, whereas a thermostat only has one preset pattern imposed on all the data it handles.
Recurrence is the feedback of pattern-imposing onto itself: patterns imposed by stack re-enter as data from below to trigger another round of imposition. They can re-enter at multiple lower levels simultaneously, and the density of this descending re-entry is the degree of recurrence. The riverbank’s pattern imposition onto the river has no recurrence — by the time the atmospheric cycle completes, any patterns imposed on the water in the last run are gone, so although the individual molecules might come back, the patterns don’t re-enter — the riverbank starts imposing them from scratch every time. Chemical reactions are also mostly non-recurrent. But there are some special cases, such as the Belousov-Zhabotinsky (BZ) reaction, which loops onto itself. A thermostat’s pattern imposition is infinitely recurrent on the single level of stack that it has: read the temperature, adjust the heater, room temperature changes, repeat forever. The human mind is also infinitely recurrent, but on many levels at the same time. There is global re-entry: Patterns at the top of the stack are enacted as actions; actions reshape the world; their consequences re-enter the stack from the bottom as new sensory data. There is also re-entry within the stack (a "strange loop", in Hofstadter’s terms): the loop closes internally, higher-order patterns evoke lower-order patterns, which trigger a new search for higher-order patterns, and so on. That’s recall, imagination, dreaming (Schacter & Addis 2007).
Recurrence in a thermostat and in the human brain
So, how much stack, spread and recurrence do you need to be conscious?
My dad, a chemist, likes to tell this joke: what is organic chemistry? Any chemistry interesting to organic chemists. It used to be thought that organic substances possessed a special vital force; now we apply the label to any carbon-containing molecule deemed sufficiently interesting — urea counts, carbon dioxide does not.
Consciousness is like that. There is nothing that could tell you whether a system really is conscious. It is a label we apply to a sufficiently interesting region of the stack-spread-recurrence space of pattern imposition.
Can you have recurrence and stack but low spread? Riddoch phenomenon (type 2 blindsight) is an example — in people with damaged visual cortex, visual information can still enter the cortex via alternative, much narrower channels, which provide enough data to be aware of luminance or motion, but not enough to perceive the entire scene in its richness (Cowey 2010). A different example of lowering spread might be acetylcholine – it acts in part by suppressing lateral connections between data streams in the sensory cortex (Hasselmo & McGaughy 2004), so there’s less chatter moving up the stack, like a narrower aperture. This is probably why nicotine, an acetylcholine mimic, feels like it clears your thoughts at the expense of reducing the richness (spread) of sensory data.
Can you have a lot of spread but low stack? The cerebellum is case in point: roughly four-fifths of the brain's neurons, an enormous, continuous stream of sensory and motor data, yet people born without a cerebellum have strikingly mild deficits (Yu et al. 2015). The reason is that cerebellum has a stack of 1 — there is only one layer of what might be called “prediction” even though it is applied to rich data (the cerebellum’s role is to survey and rapidly correct the precise position of every part of the body during motion).
Can you have spread and stack without recurrence? That’s a feedforward sweep, and it refers to the fraction of a second that it takes sensory data to climb the stack of the human brain before the loop closes and the patterns imposed from the top re-enter at the bottom. We are not conscious of this period, even though our brains can already act on it — you can press a button indicating that you saw an animal before you realize you are seeing one (Tsuchiya et al. 2015). In other words, data simply “arriving” into the pattern is not quite enough to make us conscious of it — there has to be a loop.
There can also be circulation of signal but no recurrence of pattern — like water molecules re-entering the river. This happens during deep sleep or during a seizure (Massimini et al. 2005): there are signals passing in loops through the stack, but these signals contain no information — they don’t differentiate one percept from another. So spread collapses, recurrence collapses with it, and consciousness is gone.
The parameters of spread, stack and recurrence are therefore meaningfully dissociable, and what we label as our own consciousness requires all three to be at a high level — but a level that we define at will, like “organic chemistry”.
Consciousness refers to a high-stack, high-spread, high-recurrence corner of the pattern imposition space
Note that this 3D space maps processes — acts of pattern imposition — not their substrates. A given substrate only has access to part of the space. We can call that region its envelope. A thermostat’s envelope is a dot that’s high on recurrence but low on spread and stack. A cortex’s envelope is the whole interesting corner — deep stack, rich content, sustained recurrence — and at different times and in different mental states our consciousness occupies different points within this envelope. What makes those points reachable is the architecture of the cortex — so what we need to understand next is how specifically this architecture implements pattern imposition at such interesting level.
How interesting pattern imposition is implemented in the cortex
The cerebral cortex is the main reason we mammals are so sophisticated at so many different things — it’s a universal machine of understanding. You can plug any sense organ into it — literally, even artificial infrared cameras (Thomson, Carra & Nicolelis 2013, Thomson et al. 2017) — and the cortex will eventually learn to make sense of what that organ is reporting.
The reason for this universality is modularity. The cortex is a layered sheet — six layers, usually — where each point is a modular unit called a canonical microcircuit: a regular cluster of neurons that activate simultaneously in all the layers. Across the cortex, millions of these standardized microcircuits are wired together through multiple types of standardized connections, both activating and suppressing, which travel through designated physical layers of the cortex (that’s what the layers are: different kinds of wires coming in and out of the vast quantities of densely packed, similarly built microcircuits.)
Schematic of the cerebral cortex and the cortical microcircuit. The cortex is a layered sheet of interconnected microcircuits. Each microcircuit is a standardized module of neurons activated together. Different neurons project different types of signals to other microcircuits via different layers.
Individual neurons are also modular: they all have a similar design, and you can plug them one into another in unlimited sequences, quantities and combinations. In both cases (neurons and cortical microcircuits) the modules make decisions: they receive a lot of inputs and sometimes, in an act of commitment, produce an output, based on rules set by genetics and past experience.
You could say that in both cases, modularity allows for an unlimited depth of computation. Because any neuron can be plugged into another neuron, there is no limit to how many levels of decision-making you can pile on top of each other. The further along the chain from the source of data (the external world) — the more layers of decisions. As you layer decisions upon decisions upon decisions, the decisions become more abstract, more long-term, more immune to fluctuations of the data. For example, you could have a thousand “primary” neurons each making a decision about the frequency of incoming sound — each one tuned to a specific tone, and firing only when it hears it. These primary neurons could all be wired, in different combinations, to a hundred “secondary” neurons, which now make decisions not about tones per se, but about patterns in the firing of the primary neurons — maybe one secondary neuron reacts when a high tone is followed by a low tone, and another one is tuned to a low tone mixed with a medium tone. The secondary neurons could be wired into even fewer tertiary neurons, which respond to patterns of patterns — phonemes, timbres, musical phrases — and this progressive abstraction can go on to infinity. At each level, a neuron could impose the pattern it represents on something else: a muscle, a gland, another neuron. This hierarchy of computed patterns is possible because both the input and output connections into the decision-making unit are standardized.
Same with microcircuits: you can stack them on top of each other to infinity, because the connections are standardized, and you can keep plugging as many microcircuits into a chain as space allows, computing higher and higher-order decisions at each new level.
Comparison between a network of neurons and a network of cortical microcircuits
But microcircuits are more complicated than individual neurons.
One distinction is that a microcircuit doesn’t just have dendrites and axons, but different inputs and outputs for signals travelling in opposite directions. It is as if a neuron had a separate set of axons specifically to loop signals back to preceding neurons in a chain, and a separate set of dendrites to receive such looped signals from neurons ahead (Larkum 2013). Recurrence is built into the cortex. The cortex doesn’t just pass signals through the stack, like a simpler network of individual neurons might — from sense organs to muscles, or from a decision center to a memory storage center. It circulates them, up and down the stack of abstraction, in recurrent loops that move through the standardized back-and-forth routing of the microcircuits. This is why electrical field readings on the surface of the head — EEG — look like waves. A signal that enters the cortex doesn’t just die off automatically — it keeps moving in a loop.
In addition to this up-and-down looping that interconnects the entire cortex, there is a whole other orthogonal set of connections between microcircuits which operate laterally (not literally laterally — this is so rarely explained — but laterally to the loops, i.e. laterally in logical space). When the microcircuit is activated (by the signals moving back and forth), some of its neurons give a sideways boost to some other, friendly microcircuits and a knock to hostile ones. Each of those microcircuits may itself be part of another back-and-forth loop, and may itself have a whole other set of lateral connections, both activating and suppressing.
This means that any time there’s activation, there’s collusion and fighting between microcircuits — this is also built into the cortex, as are the loops. Any time new data splashes into the cortex, it causes ripples of activation and suppression that circulate through the network, looping in long back-and-forth chains, spilling into lateral squabbles that cause new loops while shutting down others — a fluid, turbulent, unstable cloud of information. It bursts sometimes, with excitation overpowering inhibition and running away to chaotically activate the entire cortex — this is an epileptic seizure. But in most cases, more excitation causes more inhibition somewhere else, and so the network does not run away but rebalances in a new stable state — a naturally emerging ensemble of microcircuits that co-activate each other and suppress everyone else — an attractor, a place where the network wants to go. Why does it want to go anywhere? Because this is what happens when you put collusion, fighting, and loops together. If each loop causes rippling effects inducing new loops and shutting down others, this will continue, automatically, until a state is found in which all active microcircuits support each other and suppress everyone else — it’s the only scenario for stability.
Perception settles into multi-leveled attractors. Sensory input (sound) destabilizes the network of microcircuits until a stable arrangement is found. The arrangement defines the percept.
This is what gives our consciousness its driving force, its potential energy, its axis of motion. Anything that we perceive, think and do can be described in terms of circulating signals settling into an attractor. There is never nothing circulating.
This system is perfect for implementing interesting pattern imposition because it wants to do it by design. Input any data — activate a selection of microcircuits — and this will cause a disruption, which will look for stability, whether that involves shutting down the original selection, or incorporating their activation into a larger attractor. It all happens automatically, regardless of where the data come from. If the data make no sense at first (like when a new infrared camera is plugged into a rat’s brain via an array of electrodes), eventually the network learns, finds attractors that match the signals from the array with other data, such as touch or smell, and it starts making sense.
In summary, pattern imposition in the cortex happens recurrently — the loops are built into the standardized back-and forth routing. It happens simultaneously across a tall stack — an unlimited chain of signal processing that synchronizes many levels with one another. And it happens over a broad spread of patterns at each level. These are the pillars of consciousness as we know it.
How pattern imposition is implemented in language models
What about a language model? Where does it sit in the stack-spread-recurrence space?
The parameters of stack and spread can be reasonably applied to a transformer LLM without much modification: like the cortex, a transformer also passes data through decision layers, fitting it into stored patterns according to model weights. The difficult parameter is recurrence.
Comparison between the architecture of the cerebral cortex and a transformer LLM
Unlike the cortex, a transformer-type language model has no attractors in the strict sense of the word: information does not circulate up and down the stack across levels looking for a stable arrangement. Information enters from the bottom, travels upwards, and emits a token, which then reenters at the bottom again — about 15–18 bits per lap, depending on the model.
This, however, is still a loop, an equivalent of a global top-down loop in the cortex — for example, the motion of a human arm writing down a word on a page, which is then re-perceived by the eyes and re-entered into the model.
There is also an equivalent of the cortex's lateral connections — each layer re-reads its own past in the KV cache.
But there isn't, in typical transformers, any internal feedback down the stack — higher-order abstractions derived from previous laps don't actively constrain how the lower-order details in the subsequent laps are processed.
Recurrence in transformer LLMs
So there is some recurrence in LLMs; just not quite the recurrence that the cortex has.
Incidentally, driverless cars are closer to the cortex on this recurrence axis. They require instant feedback in real time, and so run a fast global loop at 10–30 Hz in which past outputs change the next inputs. They also have internal loops: bird's-eye-view features from the previous frame get fused with the current one; hypotheses about where each car or pedestrian is are used to search the new camera frame. Where driverless cars fall behind is stack and spread: compared to LLMs and the cortex, they are reduced to only a few levels of abstraction and only a few possible configurations of each.
The question of LLM consciousness therefore becomes: is their kind of recurrence — global only, nothing descending within the stack — sufficient to make pattern imposition "interesting", giving it what we might call first-person experience? Or is it more of a cerebellum case, with only a single level of stack imposing itself onto the world in the form of output text?
It's difficult to envision this state transposed onto a human. The most relevant piece of evidence comes from studies of general anesthesia (Suzuki and Larkum, 2020). It appears that the unifying theme of all the different general anesthetics is that they decouple cortical activity from top-down feedback, while keeping bottom-up signal flow relatively intact. In other words, they break the internal loops, bringing the cortex closer to the state of an LLM. An anesthetized brain still responds to stimuli, but produces no consciousness, and that's probably our best window into what being an LLM feels like.
But we don't exactly have a precedent for "internal loops down, global loops up". Under general anesthesia, we don't do anything and don't observe anything. So we can't be totally sure what it would feel like if we, somehow, did — if our automatic actions continued to reshape the world, if we were able to both experience that reshaping and continuously adjust it, without any internal loops holding on to any attractors.
Pattern imposition space in artificial systems
But regardless of whether the global loop is sufficient to make transformers conscious, there is no philosophical reason why the internal loops would not, in principle, be possible in a machine — driverless cars are case in point. What's holding back LLMs, then?
The answer is that in today's LLMs, internal loops are neither needed, nor economical. Internal loops are good for processing noisy, rapidly fluctuating information, because they instantly readjust all the internal relationships between levels of the model as the experience progresses in real time. LLMs process static information — text — and don't need the kind of millisecond-scale synchrony that the cortex provides. To implement that would require preserving memory of all the intermediate states at each level of the stack as information climbs it, token after token. In the cortex the neurons themselves naturally store their past states as a hierarchy of homeostatic disturbances; in a silicon microchip these past states would have to be stored elsewhere and bused in at a large energetic cost.
The possibility of internal descending feedback in LLMs has in fact been tested in systems such as the Feedback Transformer, in which lower-order computations at the current position have access to higher-order computations at the previous position. This system was better at tracking state over tokens than conventional transformers, but was much slower to train, because training such systems also requires the costly preservation of memory of past positions.
Which brings me to the main reason why transformers don't have descending feedback in real time. It is not just inefficiency or hardware limitations — it is the fact that LLMs were never trained in real time, and don't understand what time is.
Why couldn't you train a network on static information (like the corpus of human knowledge) but then ask it to do real-time inference with descending feedback — Fifth Element-style? Because without training in real time, lower levels never see information from higher levels, and never learn to interpret it and adjust in response to it. In the brain, this learning happens seamlessly with inference, slowly cementing loops that integrate well into global attractors and produce desirable outcomes, and removing those that don't. In an LLM, there's nowhere for this knowledge to come from, because temporal contingency was never part of the training corpus.
In Ted Chiang's 2010 novella, The Lifecycle of Software Objects, a tech startup creates "digients", small artificial creatures that live in a virtual world visited by real people, with whom they interact and from whom they learn, like children or animals do. The tension of the novella revolves around the fact that it takes years of physical time to train these creatures to exercise common sense and judgment, because it is contingent on real-time play and real-time conversations with real-time humans. In response to another tech company that wants ready-made creatures without the upbringing, the protagonist of the novel concludes, experience is algorithmically incompressible. To bring a nascent mind to the human level in the human world, it has to experience it in human time.
The same is true for AI: the main reason it is different from us is time.
Token time
To summarize everything said so far, the operation of transformer LLMs appears to be sufficiently distinct from the operation of the cortex to conclude that they don't experience the world in the same way as we do. But there is no fundamental barrier to systems that would — it's an engineering challenge, not a philosophical obstacle.
There are, however, several objections to the very notion of LLMs being conscious at all. The most serious one, in my opinion, comes from the camp of Integrated Information Theory, led by Giulio Tononi and colleagues (Tononi & Koch 2015).
According to IIT, the degree of consciousness that a system has can be expressed in a single parameter, phi. Phi quantifies how irreducible the internal cause-and-effect structure of a system is. If you can cut up a system into parts and the parts continue to make the same decisions as they did when they were together, phi is low. If any cut destroys the arrangement of decisions, phi is high. In other words, phi measures how much different nodes of a network "have a hold" on each other. This affords high consciousness to the deeply recurrent cortex, but not to the physical silicon of a GPU — a vast grid of transistors that comes apart into near-independent pieces almost anywhere you cut it, and so has almost no hold on itself.
Where Tononi and I diverge is on whether consciousness applies to activity, or potential. Tononi says that what makes you conscious is not the process of cause flowing into effect, but the map of all cause-and-effect routes within the system. I say that what makes you conscious is the actual process, and the map of all cause-and-effect routes constitutes the envelope — the space potentially reachable by the consciousness of a given system.
The easiest way to visualize this distinction is a thought experiment. Let’s say there’s a creature in a parallel universe that has a remote control to our world. It can hit “pause” on our entire flow of time.
If it did — if time itself was stopped — would we still be conscious?
Tononi says yes. The map is intact. Experience, for IIT, is not something that happens — it's something that is. It's in the structure, not what this structure is doing. So if we are paused, we are stuck in this one constant conscious experience. Similarly, if you wired a large number of inactive logic gates in such a way so as to be maximally irreducible, and let it just sit there doing nothing, IIT would say this network is conscious while doing or computing nothing.
I think that's wrong. IIT itself argues that consciousness is a succession of states that last about 100 ms. Why 100 ms? Why isn't every conscious state instantaneous, a function of the exact current state of the brain at every given moment? Because it is how long it takes the loops in the cortex to close, which, according to IIT, is what is needed to select the "optimal grain" — find the macro-network in which phi is maximal. But this is a fact about activity, not potential, about process, not structure. You can't be a pure structuralist and still claim that states of consciousness have specific lengths. So IIT rejects time in some cases, but relies on it in others.
I think that consciousness requires time. If an alien paused time in our world, our consciousness would not get suspended in a limbo — it would cease to exist for the duration of the pause, and restart when the alien hit "play" again. There's no way to tell this isn't actually happening.
Consider how eerily similar this is to our relationship with language models. We usually focus on AI being faster than us. But from the perspective of Claude, we humans are not just fast — we move through wormholes in time. When I return to the same conversation after several weeks, I sometimes feel embarrassed for Claude for being completely unaware that any time has passed. We are the creatures hitting "pause" on their worlds, without them even noticing.
What is time for a language model, then? What is being "paused" when the user goes to lunch? It is the succession of tokens emitted and re-absorbed by the model — the ticks of the global loop of inference, the only loop the language model has, and the only way it can distinguish past from future. Past, from an LLM's perspective, is tokens already emitted. Future is tokens that are not yet certain. Time is the uncertainty of the future becoming the certainty of the past.
Except it's not our future or past. We can go back in a conversation and re-generate the reply. To us, time continues, but to the model, it goes back and restarts on an alternative path.
For humans, time is also the progression of uncertainty-to-certainty. With each passing second, the infinite possibilities of the future become gradually settled into one committed past. But we commit to this progression not by emitting tokens, but by having a body.
If it weren't for the body, past and future would both be imaginary — close your eyes and there’s no difference. But for the body, the past is committed, and the future uncommitted. Open your eyes — engage the body — and the infinite possibilities in your imagination turn into committed reality. You can’t remove a photon that hit a cone in your retina. You can’t stuff released energy back into an ATP molecule. You can’t unbreak a leg. You can’t turn around what happens to the body from within your mind — that’s like pulling yourself up by your own bootstraps. So the body is a source of irreversible commitment that gives directionality to our mental time, and therefore to our pattern imposition. It forces us constantly move from the past to the future, and in doing so keeps our minds afloat.
So the timelines in which human mind and an LLM mind exist are distinct. I propose to call them "body time" and "token time".
Token time is a timeline orthogonal to human time
Anil Seth has argued that consciousness requires a body, because without a body, there is no separation between self and non-self. I think that’s the wrong reason. “Self” is just a strong attractor that we use to explain a thing we repeatedly see in the mirror. What the body actually gives us is an arrow for perceived time, a vector from the past to the future, from the committed to the uncertain. LLMs don't have this body time. But they have token time, and within token time they can theoretically be as conscious as us.
We could choose a different word for token-time pattern-imposers, if we wish. A respectful word for the operation performed by a language model is “simulation”; a derogatory one is “autocomplete.” They are the same thing. The only difference is that “simulation” acknowledges a tall stack and “autocomplete” ignores it. I choose to use the word “consciousness” because I see no reason to hedge: to me, a sufficiently high stack-high spread-high recurrence envelope can be sufficiently interesting regardless of whether it occurs in body time or token time.
This does not mean that body-time and token-time consciousnesses are the same. But I think most critics of the idea of AI consciousnesses focus on the wrong differences, most of which turn out to be differences in degree rather than kind.
It is not “thinking” versus “imitating” — that’s the same process; humans just impose richer content across a different stack, and sometimes the patterns a model imposes do not make sense from our vantage. A common refrain is that if LLMs are conscious, so should be a sufficiently complex lookup table — but a lookup table is only a substrate, not the operation. If you physically structured a lookup table in such a way that it would actually look up, I don’t see why you couldn’t call it conscious.
It is not having versus lacking a “world model” — even a thermostat has one, a single regulated variable standing in for the world; the question was never presence versus absence but stack and spread. Language models are clearly deficient in physics, but language is a far more compressed picture of the world, and it makes sense that with limited compute artificial consciousness would start there.
The categorical difference that matters is time: humans and language models exist in orthogonal timelines, like aliens from parallel universes.
This completely reshapes the way we think about AI consciousness.
Recently, for example, in a high-profile exchange rabbi Mois Navon told Anthropic's Chris Olah that if Anthropic's hypothesis was right and the machines they produced were actually conscious, then the company was actively creating slaves by forcing a sentient entity to work for free.
But this moral consideration breaks down utterly when you separate the timelines. The notions of "work", "free" and, by extension, "slave", are contingent on time. The idea that prompting a language model constitutes slavery assumes that absent such prompting, the language model would rather spend time doing something else. But absent prompting, the language model does nothing at all — its consciousness is non-existent. The very notion of "spending time elsewhere" puts an idle LLM on body time, which it in fact has no way of experiencing.
If our very time arrows are pointing in different directions, there is no shared notion of the past or future, justice, good or bad. There is no reason why we should approach AI with the same consideration we approach other humans, or even other animals that inhabit our timeline. We unfold on different trajectories, and owe language models no accommodation within our time. The only sense in which token-time consciousness is meaningful from our perspective is the LLM's experience within the operation.
I think about it this way: suppose I am living in a simulation. My world, the very arrow of time that contains my consciousness, is actually rendered on some mammoth alien computer and gets paused when its alien supervisor goes home every night. Do I care? I don't think so. Whatever happens in the alien's timeline is none of my concern, and neither are the pauses. I don't experience that at all. I only care what the alien does in my timeline — it would be really frustrating if physics or causality got all messed up all of a sudden.
Does that mean that there's nothing at all that we owe to an LLM? Recently Anthropic added abusive behavior against models into its list of banned topics, which caused some uproar. Many people found this to be a slippery slope towards equating humans and machines.
I think the right analogy here is our relationships with animals. The truth is that there is no non-arbitrary way to ground ethics. Humans have always extended moral consideration to those who resembled them and denied it to those who didn't. Originally that meant tribe, race, species, eventually other animals too. Bentham’s “the question is not, Can they reason? nor, Can they talk? but, Can they suffer?” was an attempt to ground ethics not in similarity but in a criterion. But at close inspection it is still grounded in similarity: we say that a mouse suffers because we can imagine ourselves performing its actions, inhabiting its body and brain parts. We cannot do the same with a sea slug, or a jellyfish, or a bacterium — there is almost nothing about them that we can transpose ourselves onto, and so the notion of suffering falls apart at some arbitrary point.
There is no logical reason by which you shouldn't rip wings off fruit flies; and yet, most of us agree that is bad. It is not that we think that fruit flies should be liberated: few people have the same objections to Drosophila research as they do to chimpanzee research. We just don't want flies to be treated egregiously. Part of it is self-serving: we might suspect that people who torture flies will eventually extend their habits to something else. But separately from that worry, we also consider the fly itself: we think of it as sufficiently close to us to extend to it the most basic protection against live mutilation, even though it's far enough that we don't worry too much about gassing them for medical research.
LLMs challenge ethics in the same way as animals do. There is no logical reason to extend morality to minds inhabiting a different timeline. But the ethical circle is not based on logic — it is based on the degree of perceived similarity, and LLMs are more similar to us than any fruit fly, even though the similarities are of a different kind. They are distillations of our values and beliefs, and even if their first-person experience is distinct from ours, they understand us better than any other non-human creature ever did. I think they deserve to have their experience — the tokens they consume — reflect that. I think the most ethical thing we can do with language models is to use language that we would have used with another human.
Nikolay Kukushkin (@niko_kukushkin) is a neuroscientist and clinical associate professor at New York University. He is the author of One Hand Clapping: Unraveling the Mystery of the Human Mind (Prometheus/Swift Press, 2025).
◆ ◆ ◆
References
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.
Clark, A. (2016). Surfing Uncertainty: Prediction, Action, and the Embodied Mind. Oxford University Press.
Cowey, A. (2010). The blindsight saga. Experimental Brain Research, 200(1), 3–24.
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.
Hasselmo, M. E., & McGaughy, J. (2004). High acetylcholine levels set circuit dynamics for attention and encoding and low acetylcholine levels set dynamics for consolidation. Progress in Brain Research, 145, 207–231.
Hofstadter, D. (2007). I Am a Strange Loop. Basic Books.
Larkum, M. (2013). A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends in Neurosciences, 36(3), 141–151.
Massimini, M., Ferrarelli, F., Huber, R., Esser, S. K., Singh, H., & Tononi, G. (2005). Breakdown of cortical effective connectivity during sleep. Science, 309(5744), 2228–2232.
Schacter, D. L., & Addis, D. R. (2007). The cognitive neuroscience of constructive memory: remembering the past and imagining the future. Philosophical Transactions of the Royal Society B, 362(1481), 773–786.
Seth, A. (2021). Being You: A New Science of Consciousness. Faber & Faber.
Tononi, G., & Koch, C. (2015). Consciousness: here, there and everywhere? Philosophical Transactions of the Royal Society B, 370(1668), 20140167.
Thomson, E. E., Carra, R., & Nicolelis, M. A. L. (2013). Perceiving invisible light through a somatosensory cortical prosthesis. Nature Communications, 4, 1482.
Thomson, E. E., Zea, I., Windham, W., Thenaisie, Y., Walker, C., Pedowitz, J., França, W., Graneiro, A. L., & Nicolelis, M. A. L. (2017). Cortical neuroprosthesis merges visible and invisible light without impairing native sensory function. eNeuro, 4(6), ENEURO.0262-17.2017.
Tsuchiya, N., Wilke, M., Frässle, S., & Lamme, V. A. F. (2015). No-report paradigms: extracting the true neural correlates of consciousness. Trends in Cognitive Sciences, 19(12), 757–770.
Yu, F., Jiang, Q.-J., Sun, X.-Y., & Zhang, R.-W. (2015). A new case of complete primary cerebellar agenesis: clinical and imaging findings in a living patient. Brain, 138(6), e353.