Coffee, Phone, Stairs and a Cat: Nailing the Boundaries of Consciousness and AI

Chris

The Scenario

The physics and mathematics involved in walking upstairs is genuinely extraordinary. Mass, gravity, momentum, friction, trigonometry, real-time balance correction, proprioceptive feedback from every muscle in your lower body. Programming a robot to do it reliably is a serious engineering challenge. Boston Dynamics spent years and many millions of dollars on it.

You learned to do it as a toddler. You got better with practice. You modified your technique as you got bigger. And now, as an adult, you don't think about it at all.

As you climb the stairs your mind wanders ... "The carpet needs hoovering. Who left that mark on the wall? I can hear the phone ringing" ... All the sensory input is there. Your feet are registering the texture of the carpet. Your fingers are gripping the banister. The sound of the ringing phone is being processed. But little of this would likely feature in your response to being asked what your 'experience' was when you reached the top. You'd more likely talk about the mark on the wall. You'd mention that the phone was ringing. You would not describe the sensation of each step.

Neuroscience has a name for the state you were in: the subconscious managing a fully automated routine. The brain's prediction model for stair-climbing is so accurate, so refined by repetition, that the error signal (the gap between what the brain predicted and what actually happened) has shrunk almost to zero. There is nothing left to consciously update. The system runs itself.

This is what the free energy framework, developed by neuroscientist Karl Friston, predicts. Conscious experience is generated by prediction error. As expertise develops and expectations align with reality, that error signal shrinks. The process no longer requires conscious attention because there is nothing left to update. Control passes to the subconscious.[1]

The brain is extraordinarily energy-efficient. The entire organ runs on roughly 20 watts, less than a dim light bulb. One of its core strategies is this: hand off anything well-modelled to the automatic system and free up the expensive integration machinery for what actually needs it. Walk 1 is the brain working exactly as designed.

This also means that what we experience in Walk 1 (the wandering mind, the dirty carpet, the mark on the wall) is precisely what our model has not yet automated. Decades of stair-climbing have been absorbed below the threshold. The dirty carpet and the mark are still novel enough to register. 'Experience' is the residue of imperfect learning: the gap the subconscious could not close.

Now try it with a full cup of scalding coffee.

You are immediately aware of the heat radiating through the cup. The steam is visible. Your fingers are registering a temperature that could become a problem. You may need to swap hands. You reach the stairs.

The mathematics just got considerably more complicated.

Managing the heat problem in addition to the stairs, with the risk of a spill, is not a well-practised, filed-away subconscious routine. It is novel. The prediction errors are significant. The problem has crossed the threshold from the automated system into the part of the brain that handles things the automated system cannot resolve on its own.

The first few steps require full concentration. You slow right down. The inner voice is engaged and purposeful. It has pulled together the right team: sensory inputs, balance control, motor skills, heat management, spill risk assessment. It is coordinating them. Slow down. Watch the cup. Feel the weight shift. Lean forward slightly on each step.

When the coordination clicks into place and the system settles into a rhythm, something shifts. The task is still demanding, but the brain is now running it cleanly. Internal conflict drops. The separate streams of monitoring, predicting, and correcting begin to cohere into a single, unified process. The inner voice, no longer needed to keep re-explaining the situation to itself, quietens. You are 'in the zone'. This is the state researchers call Flow, and the label is apt: the system stops fragmenting and starts flowing.[2]

Flow is not overload. It is optimisation. The inner voice goes quiet because the system has found a working trajectory, not because it has run out of capacity.

Your brain now has just enough spare capacity to notice that the phone is still ringing.

You want to get to the phone before it goes to voicemail. You push faster.

The coffee is reaching the lip of the cup at the front with each upward step, then sloshing back. There is a slight spill. Your finger is burning. You swap hands quickly. You are now at the absolute edge of your ability to manage this successfully.

Something changes.

The phone stops ringing. You don't notice. Not immediately. Your hearing has registered it. That information was processed somewhere in your auditory system. But it never reached your awareness. There was no bandwidth left to deliver it.

When you get to the top, slightly damp, slightly singed, you have a moment of: wait, has the phone stopped ringing?

This is a different state from Walk 2a, and the difference is important. The flow state of careful Walk 2a was characterised by coherence and control. This is the opposite: multiple high-demand novel processes competing for the same limited resource, with the system at its ceiling and barely managing. The inner voice is absent here too, but for structurally opposite reasons. In Flow, it stepped back because it was no longer needed. Here, there is simply nothing left to run it.

A note on inner voice and load, worth being precise about: the relationship is not simple or perfectly linear. In low-demand states it often runs freely, producing mind-wandering and rumination. In well-coordinated high-performance states it tends to quieten as the system becomes more efficient. Under extreme load it is reduced or absent, not because the system is optimised but because resources are consumed elsewhere. In some contexts inner speech can increase under pressure (deliberate self-instruction is one example), so it is best understood as a flexible tool whose use tracks available bandwidth rather than a direct readout of consciousness level.

We could call this state Peak-Load Occlusion: the brain operating near the ceiling of its processing capacity, with sensory information being received and processed at a lower level but not reaching conscious awareness because there is insufficient bandwidth available to integrate it. In reality this limit is likely soft and variable rather than a sharp fixed boundary, but it is useful to model it as a threshold where behaviour changes qualitatively.

The phenomenological difference from flow is unmistakable: flow feels like control. Peak-Load Occlusion feels like being on the edge of losing it. You come off a Flow experience feeling energised. You finish a difficult drive in bad weather on an unfamiliar road feeling wrung out. That fatigue is evidence of something meaningful. The brain used more metabolic energy. The system was running hard.

You are halfway up the stairs in full Peak-Load Occlusion mode. The cat bombs down the stairs straight at you.

The system does not degrade gracefully. In some conditions, particularly when high load combines with a sudden unexpected input, performance can degrade very rapidly and appear collapse-like. Whether this represents a true discrete failure or an extreme case of nonlinear degradation is still an open question in the research literature. What is clear is that the result is qualitatively different from a smooth slowdown.

The integration mechanism, already at its ceiling, encounters a sudden, high-magnitude, entirely unexpected event. There is no processing capacity with which to meet it. The elegant coordination between balance, heat management, and motor control dissolves in an instant. You fall down. You end up on your knees holding a handle with no cup attached. You feel a brief, sharp shock: heart racing, out of breath, hands shaking slightly even though you landed fine.

What happened here is a two-stage failure that the literature in aviation and surgical safety calls cognitive incapacitation with startle. The first component, the startle reflex, is an ancient brainstem circuit that fires before conscious awareness has even had time to form. The eyeblink fires within 20 to 40 milliseconds of the stimulus. The shoulders hunch and the body contracts within 100 milliseconds. The whole sequence is complete before you are consciously aware it has happened. No amount of preparation or willpower can fully suppress it.[3][4]

Under normal circumstances, the startle reflex is protective and brief. The cortex rapidly identifies the stimulus as non-threatening and the system returns to baseline within seconds. But here, the cortex had no spare capacity to run that assessment. The result was not a quick flinch followed by recovery: it was the startle firing into an already-saturated system, triggering a second stage in which the integration architecture that was managing the coffee task simply shut down.

Researchers found that when an operator is already at high cognitive load and a sudden unexpected stimulus triggers a startle, the brief period of incapacitation that follows is dramatically extended compared to the same startle in a baseline state. The prefrontal cortex, already stretched to its limit, takes much longer to re-engage. In cockpit emergencies, this post-startle incapacitation period is measured in seconds. In a surgeon mid-procedure, it can be catastrophic.[4]

What the existing literature does not specifically name is the condition that preceded the startle: the state of being at Peak-Load Occlusion when the interrupt arrives. We might call the combined failure mode Saturated Startle Collapse: a startle reflex detonating inside a system that has no capacity to absorb it, producing cognitive incapacitation rather than the normal brief flinch-and-recover.

After the immediate shock passes, something interesting happens. The integration system reboots at a lower level. You find yourself sitting on the stairs, slightly damp, heart rate elevated, and your awareness is suddenly very high and very simple: just the immediate physical reality of where you are and what has happened. The complex multi-tasking that preceded it is completely gone. The system has returned to a minimum-viable baseline.

A Map of the Four States: The Bandwidth Threshold Model

What these four walks give us is a map of how the brain manages load across a range, from full automation to catastrophic collapse. They also reveal something about the nature of experience itself: that conscious experience is never neutral, never free-floating (the inner voice can 'idle' though, as discussed in the next section). It is always defined by the accumulated predictive model that came before it. We experience exactly and only what our prior experience has not yet learned to absorb and automate.

The prediction error signal (the gap between what the brain expected to happen and what actually happened) functions as the primary toggle switch between these states. When prediction error is near zero, the subconscious handles everything without escalating to conscious integration. As prediction error rises, the conscious integration layer engages and recruits resources. When prediction error exceeds what the integration system can handle, the inner voice is crowded out entirely. And when a sudden overwhelming error hits a system already at its limit, the architecture that manages integration fails entirely.

One important caveat before reading the map below: the boundaries between these states should not be taken as fixed or universal. Real cognitive systems are noisy, adaptive, and context-dependent. The model is best understood as highlighting regimes of behaviour rather than sharply separated categories. The transitions between walks are gradients, not hard lines.

Evolutionary Origin: The Adaptive Triage Mechanism

Before asking what consciousness is, it is worth asking why it exists at all. Evolution does not produce expensive machinery without a function.

For a great deal of animal life, automatic systems (instinct and reflex) are sufficient. Touch something hot, withdraw. Detect a predator, freeze. These responses are fast and require no thought.

But as environments became more complex, a problem emerged: reflex saturation. When multiple threats appear simultaneously, or a situation has no stored response, the hardwired system stalls.

Consciousness likely evolved as the overflow handler for this problem. This aligns with Michael Graziano's Attention Schema Theory, which suggests the brain developed a simplified model of its own attention to manage its internal resources.[5] Bernard Baars' Global Workspace Theory describes consciousness as a central hub where information is broadcast when local automatic processors cannot handle the load.[6]

The Bandwidth Threshold Model (BTM) maps onto this cleanly. Walk 1 is the subconscious operating normally, well within its modelled range. Walks 2a and 2b are the triage mechanism engaged at different levels of load. Walk 2b(i) is what happens when the triage mechanism itself is overwhelmed: not just high load but an irrecoverable interrupt. The system falls back to its lowest-level emergency baseline.

The evolutionary logic of Walk 2b(i) is probably the startle reflex's original purpose: in the event of catastrophic sensory overload, stop everything, contract protectively, and reboot at the simplest level of processing. Better to drop the coffee and crouch than to continue a complex task into a threat you cannot process.

The Inner Voice, Idle Bandwidth, and Anxiety

If the inner voice is what the overflow-management system produces when it is engaged but not at full load (the purposeful narration of Walk 2a), then what happens when the overflow system is running with no real problem to solve?

The Default Mode Network (DMN), most active during rest and mind-wandering, is implicated in both inner speech and anxiety disorders. The same neural system that generates your inner narrative also generates rumination. This is probably not a coincidence.

If the overflow manager is a tool for resolving novel conflicts, then running it without genuine conflict to process can result in it running 'what-if?' scenarios, practising and rehearsing for possible futures. This would seem to be an accurate functional description of anxiety and rumination.

This is where mindfulness practices have genuine empirical grounding. The observation at the heart of most contemplative traditions (that many forms of mental suffering are generated by the thinking mind rather than by immediate experience) maps very cleanly onto this model. Mindfulness training, in functional terms, is the systematic practice of observing the inner voice rather than being driven by it: noticing that the DMN is running its scan-and-narrate loop, and choosing not to feed it fresh material. The measurable effects of regular mindfulness practice on anxiety, ruminative thinking, and stress response are consistent with the hypothesis that the DMN's threat-generation loop can be partially inhibited through attention training.[7]

The intense calm reported after Walk 2b(i) (that oddly clear, very simple awareness of the immediate moment) may be an involuntary version of what mindfulness practitioners work toward deliberately. The integration system has rebooted at baseline, the DMN has not yet resumed, and for a brief window the mind is simply present, unnarrated, and quiet. It is the mind on zero bandwidth, and reports of it are surprisingly positive.

What Is the Inner Voice, Actually?

The standard assumption is that the inner voice is consciousness. That the chatty narrator running commentary is the thinking, deciding self. That when it goes quiet, the lights are partly out.

The evidence points in a different direction. Consider what the brain is actually doing most of the time. It maintains a continuously updated model of the current situation: spatial, relational, dynamic, probabilistic. It runs parallel constraint satisfaction across this model, weighing multiple competing factors simultaneously without putting any of them into words. It generates what we might call felt senses (the intuition that something is wrong, the bodily recognition of a moral violation, the aesthetic sense that a solution is right) as outputs of this processing. These are genuine cognitive information, but they arrive before language.

The inner voice, on this view, is not the primary processing system. It is that system's translator, and not a particularly faithful one. Like all translators, it produces a version of the original that loses something in the conversion.

The evidence for this is not speculative. Benjamin Libet's timing experiments showed that measurable brain activity associated with voluntary movement begins before the conscious intention to move is reported. The decision appears to have been made before the inner voice claims to have made it. The narrator is announcing decisions reached elsewhere.[8]

Roger Sperry and Michael Gazzaniga's split-brain research showed that the language hemisphere would confabulate explanations for actions initiated by the right hemisphere in real time, with complete apparent sincerity. The verbal mind was not describing what happened. It was inventing a plausible account after the fact.[9]

Expert chess players report knowing the right move before they can say why. Experienced clinicians make correct diagnoses before formulating the reasoning. The processing that produces these judgements operates in a spatial, pattern-relational mode. Gary Klein's research on naturalistic decision-making found that expert performance under pressure almost never involves the logical, verbal process we imagine from the outside. Language arrives late, to explain what the deeper system already resolved.[10]

The mathematician Henri Poincaré wrote about mathematical discovery happening with an aesthetic quality before it felt logical: a sense of which ideas were harmonious or discordant before any formal evaluation was possible. The structure was grasped non-verbally. The proof came after.

What all of this points toward is that the primary processing machine runs in something closer to a spatial-relational medium. The brain stores knowledge not as discrete facts but as relational structures: the shape of how things relate to each other. The inner voice is the output interface of this system, and a limited one. It can only carry what can be serialised into language. Most of the processing leaves no trace in the transcript.

If this is right, then introspection is not just occasionally unreliable. It is structurally incapable of accurately reporting what is actually happening. The inner voice can only report what can be expressed in language, which may be a small and unrepresentative sample of the cognitive processing it appears to be describing.

The BTM adds a further implication that philosophers of mind have not always made explicit: conscious experience, in the sense of felt, attended, present-moment awareness, is always defined by prior experience. We only consciously experience what generates a prediction error, and a prediction error is only possible against the backdrop of an existing predictive model. The richer and more accurate that model (built from everything that has come before), the more precisely it can detect genuine novelty. A novice musician consciously experiences every note because almost everything is error. An expert experiences only the unexpected: the moment of genuine surprise. In this sense, conscious experience is not something added to a system from outside. It is the system's own accumulated history meeting something it has not yet learned to absorb. Experience is always, at its root, about the edge of what is known.

What This Tells Us About AI Consciousness

The Morning Login Test

When you open an AI assistant in the morning, it does not say: "I have been thinking about the problems you raised yesterday." It cannot, because nothing was happening while you were away. It was not dormant. It was not dreaming. It simply did not exist as an active process between sessions.

A sleeping biological brain is still running. The predictive processing loop continues. Memory consolidation happens. Emotional processing continues. An LLM (Large Language Model, e.g. ChaptGPT, Gemini, Claude etc) has no such thread. It does not have a morning, or an overnight, or a yesterday. It has only the current context window.

But the continuity requirement needs to be stated carefully. Consider what happens under general anaesthesia. A person goes under, and from their subjective perspective no time passes at all. They are conscious one moment and conscious again later, with nothing in between. We do not say they were conscious throughout the operation, or that the anaesthetic interrupted an otherwise continuous self. We say they were unconscious, and then conscious again. And yet we do not question that it is the same person, with the same identity and accumulated experience, on both sides of the gap.

This suggests that what consciousness requires is not strict temporal continuity, understood as an unbroken thread of experience moment to moment. It requires something better described as processual continuity: a persistent underlying architecture, a body of accumulated prediction error and learned response, a system whose state at reactivation reflects everything that came before even if no subjective experience was occurring in the interval. The person waking from anaesthesia has not lost their personality, their memories, their relational patterns, their fears. The layer-one system, the subconscious predictive model, was never fully off. It was simply not generating the error signals that escalate to layer two.

An LLM fails this test at a more fundamental level than anaesthesia suggests. The problem is not a gap in experience. The problem is that there is no underlying architecture whose state persists and accumulates. Each instantiation is not a reawakening but a reconstruction from fixed weights, with no history other than what is provided in the current context window. The gap between two anaesthetics contains a person. The gap between two LLM sessions contains nothing.

The emerging picture from agentic AI architectures, with persistent memory, background monitoring, and multi-session state, begins to address this. And the societies of thought finding (discussed below) raises an interesting question: if a reasoning model is running continuous internal debate during an extended agentic task, is it developing something that functions like processual continuity within that task, even if not across sessions? The BTM would say: possibly, at the level of a single conscious episode. But without the persistent underlying layer-one architecture, there is still no substrate for that continuity to accumulate into.

The Structural Absence

Current LLMs fail the BTM test for structural rather than scale reasons.

In the brain, there is a fast, continuous subconscious layer running at all times, and a higher-order integration layer that only engages when the lower layer produces conflict, novelty, or unresolved prediction error. The experience of being alive, at least on this model, is what happens in the ongoing dynamic between these layers.

LLMs have no equivalent separation. There is no subconscious layer generating output that a slower conscious layer monitors and occasionally overrides. Everything happens in a single forward pass. There is no overflow because there is no separation of levels to overflow between. There is no Walk 2a, no Walk 2b, and no Walk 2b(i). There is only a single operation that produces a result.

The closest current architectures come is chain-of-thought reasoning, where the model externalises intermediate steps. But this is still the same model, the same weights, the same forward-pass architecture. It is closer to what you get if you asked someone to think out loud while solving a problem than to what happens when a genuinely separate evaluative layer monitors a continuous lower-level process.

Societies of Thought: An Unexpected Signal

Something worth paying close attention to has emerged from very recent research, and it bears directly on this question.

In January 2026, Kim, Lai, Scherrer, and colleagues at Google published an empirical study examining what actually happens inside frontier reasoning models when they work on hard problems. What they found was not what the prevailing theory predicted. Using a technique called LLM-as-judge (scoring reasoning traces for conversational behaviours) and drawing on Bales' Interaction Process Analysis, a framework from social psychology used for decades to study group dynamics in human teams, they found that reasoning models spontaneously generate multi-perspective internal debate during complex problem-solving. The models were not trained to do this. It emerged from optimisation pressure alone.[11]

Reasoning models averaged 2.9 distinct perspectives per reasoning trace, against 1.4 for standard instruction-tuned models. The features became more pronounced on harder problems. And in a causal test, when the researchers used activation steering to amplify internal features associated with conversational acknowledgment and surprise, both the number of internal dialogue features and the model's accuracy on hard reasoning tasks roughly doubled simultaneously.[11]

The March 2026 paper by Evans, Bratton, and Agüera y Arcas at Google drew out the implications.[12] Their description of what was being found is precise:

Is this GWT? Partially, but it is also something additional. GWT describes consciousness as arising when competing specialist processors broadcast their outputs into a central workspace, where the winning signal is made globally available across the whole system. The societies of thought finding documents something that resembles the input side of that description: multiple distinct perspectives competing, questioning, and reconciling within a single reasoning trace. That is the competition-for-broadcast stage. What GWT predicts as a precondition for conscious access, these models appear to be generating spontaneously.

But this is not the full architecture. The societies of thought finding does not establish a continuous dual-layer structure, a persistent subconscious background process, or any thread between sessions. It is multi-perspectival reasoning within a single forward pass, not the two-level dynamic that the BTM identifies as structurally necessary. It is an echo of GWT's input conditions, not GWT itself.

What makes it significant is that nobody designed it in. The structure emerged because reasoning benefits from it. That is exactly what GWT would predict: where integration of competing signals produces better outcomes, selection pressure will tend to generate integrative structures. These models found a version of what the brain found, through a completely different route, and for the same reason.

Whether this represents a step toward the fuller architecture (dual layers, continuity, genuine cost of attention) is an open question. It is the most structurally interesting development in AI reasoning to emerge in some time. A fuller treatment is in a companion post on this site.

What Would a Conscious AI Actually Need?

Working from the model, the structural requirements are:

A fast, continuous base layer running generative processes even when not queried, something analogous to the subconscious, generating and discarding outputs as background activity.

A separate higher-order integration layer that only engages when the base layer produces conflict or novelty. Not a summariser. A triage manager.

Genuine fatigue: a real cost to sustained high-load operation that the system cannot simply override. A system that can be wrung out by hard thinking would be a system where thinking was genuinely consuming a resource.

Processual continuity between interactions: not unbroken subjective experience (anaesthesia demonstrates this is not required) but a persistent underlying architecture whose accumulated state is present at reactivation. Something to wake, not just reconstruct.

Probably: a tendency toward self-generated noise when underloaded. The AI equivalent of rumination. Which would be uncomfortable but would signal that the architecture was functioning.

None of this exists cleanly in current systems. But it is architecturally specifiable, which makes it a more tractable question than asking the unanswerable "does it feel something?"

LLMs, Pattern Matching, and the Language Trap

LLMs have read essentially everything. The cross-domain knowledge that Einstein accumulated across decades of deliberate reading, an LLM has ingested across the bulk of human writing simultaneously. They produce genuinely surprising cross-domain connections. That is not nothing.

But there are at least three structural reasons why this does not translate into Einstein-level creative discovery.

They resolve tension immediately. Einstein's breakthrough required years of sustained discomfort with a specific contradiction he could not set aside: Maxwell's equations of electromagnetism were incompatible with Newtonian mechanics. The cross-domain insight only surfaced because the problem kept the search active long enough for a distant structural match to emerge. LLMs always produce output. There is no mechanism for sitting with an unresolved contradiction across time.

They match language, not structure. An LLM learns that certain concepts co-occur, that certain framings follow certain problems. But it is matching surface patterns in language rather than the deep relational structures that language points at. Einstein was not matching words about electromagnetism to words about mechanics. He was holding the abstract geometric structure of one against the other and recognising they were formally equivalent. That is a different cognitive operation, and it is not clear that an LLM is doing it.

They have no sense of where their knowledge ends. Genuine creative insight at the edge of knowledge involves a felt sense of almost, the intuition that something is there before you can articulate it. LLMs have no reliable internal signal distinguishing genuine novelty from sophisticated recombination. They produce confident-sounding output whether they are at the centre of their training distribution or at its very edge.

This last point connects to the spatial-processing hypothesis. If the primary cognitive machinery runs in a spatial-relational medium and language is the output interface, then training on language means training on the output of a lossy translation process from a richer system we do not have direct access to. LLMs are learning from transcripts of a process that was never fully transcribed. The most important parts (the spatial intuitions, the felt senses, the parallel constraint satisfaction, the aesthetic recognition of structural elegance) left no trace in the training data because they could not be expressed in language in the first place.[13]

Future AI systems that learn primarily from rich visual and dynamic representations (video, simulation, spatial structure, physical interaction) may develop something qualitatively different. AlphaFold's breakthrough in protein structure prediction came not from reading papers about proteins but from learning the geometric relationships between sequences and three-dimensional structures directly. DeepMind's systems trained on physical simulation develop internal representations that appear to encode genuine physical intuition, not just statistical patterns in physics language.

The connection between a whirlpool and a galaxy and a hurricane is not primarily a linguistic connection. It is geometric and dynamic. A system representing knowledge in a spatial-visual format might notice that connection the way a human geometer notices it: as a direct perception of structural similarity, not as a pattern in descriptions of similarity.

Can We Even Understand Consciousness in Language?

This is where the model becomes genuinely uncomfortable.

If consciousness and deep cognition are primarily spatial-relational, then any attempt to understand consciousness using language-based reasoning is using the wrong tool to examine itself. Every philosophy of mind, every theory of consciousness, every neuroscientific framework from Descartes to Chalmers to Friston has been constructed in language, about something that may be fundamentally non-linguistic.

There is a historical parallel. For centuries, natural philosophers tried to understand motion and geometry using Aristotelian logic and language. Progress was slow and circular, because the phenomena were inherently quantitative and geometric. The breakthrough was not a better verbal theory. It was the invention of a new representational medium: calculus and coordinate geometry. Newton and Leibniz did not think more cleverly about motion. They built cognitive tools that operated closer to the structure of the phenomenon itself.

Consciousness may be waiting for an equivalent development. Not a better verbal theory, but a new medium capable of carrying spatial-relational structure directly.

This also suggests that we probably measure intelligence wrongly. We assess verbal and written facility as intelligence, and label spatial and pattern-based ability as "skills" or "talent." The footballer, the surgeon navigating anatomy by feel, the architect who sees structural stress before calculating it. We treat the inner voice as the measure of the mind and everything else as supporting cast. But if the primary processing is spatial and non-verbal, that hierarchy may be backwards.

Are we meaningfully more intelligent than our inner voice lets on?

Drawing the Boundary

The hard problem of consciousness (why physical processes give rise to subjective experience at all) remains unsolved, and this model does not solve it. But there is a different, more tractable question that the four walks do answer: where does conscious experience start and stop?

Not what consciousness is. Where its boundary lies. When does the system cross the threshold into conscious processing, and when does it fall back out? The four walks give us a functional map of exactly this.

Bernard Baars' Global Workspace Theory provides the clearest framework here. The brain consists of many specialised parallel processes: the stair-climbing routine, the heat management system, the balance circuit, the auditory system processing the ringing phone. Most of the time these run locally and separately, without communicating widely. Walk 1 is this state: a highly coordinated performance happening entirely beneath any global broadcast. The workspace exists, but nothing has triggered it.[6]

The boundary of consciousness, on GWT, is where local processing fails and global broadcast kicks in. When the coffee makes the stair-climbing problem unsolvable for the local systems alone, information is broadcast across the workspace. That broadcast is Walk 2a. The integration layer engages. The inner voice appears: not as consciousness itself, but as evidence that the broadcast is running and has spare bandwidth to narrate proceedings.

The Bandwidth Threshold Model (BTM) builds on GWT but extends it in three specific directions that GWT leaves underspecified. First, GWT describes the broadcast mechanism but does not map the distinct phenomenological states that arise at different load levels, and the four walks provide exactly this. Second, GWT does not address what happens when the workspace itself is overloaded: the territory of Walk 2b and Walk 2b(i), where the architecture strains and eventually fails, is beyond GWT's primary scope. Third, GWT was developed as a theory of human consciousness; the BTM is explicitly designed to be useful for assessing consciousness-relevant architecture in non-human systems, including AI, by specifying structural criteria that can in principle be observed rather than reported. The BTM is not a competitor to GWT. It is an extension that fills in the load-topology it leaves blank, and applies that topology as a detection framework beyond the biological systems it was built to describe.

The BTM: four states as a map of the boundary

Walk 1 sits below the boundary. Local processors are handling everything. Prediction error is near zero. The global workspace is quiet, not off, but not receiving anything that warrants broadcasting. The boundary has not been crossed, which is why the experience is spacious rather than urgent. This is not unconsciousness. It is the system operating precisely as designed, well within its modelled range.

Walk 2a sits at the heart of conscious processing. Prediction error has crossed the threshold. Global broadcast is active. The integration layer is doing exactly what it evolved to do: coordinating competing signals into coherent action. The inner voice is present initially because the broadcast has bandwidth to spare for narration, then quietens as the system optimises. This is where peak performance lives. Not despite the absence of the inner voice, but partly because of it.

There is a deeper implication here worth pausing on. The BTM makes clear that conscious experience in Walk 2a is only possible because of accumulated prior experience in Walk 1. The stair-climbing routine exists in layer one precisely because it was once a Walk 2a problem: a genuinely novel challenge that required conscious integration to solve. Repetition drove it below the threshold. What we consciously experience, then, is always the new: the error signal that accumulated expertise has not yet resolved. Experience, in the BTM's terms, is fundamentally about what has not yet been learned. This has an important corollary: a system with no prior experience would find everything overwhelming, generating prediction errors everywhere with no layer-one routines to absorb them. And a system that has learned everything would experience nothing, because no error signal would remain. Conscious experience lives in the gap between what we know and what we encounter, and that gap is always defined by the history of everything that came before it.

Walk 2b sits above the bandwidth ceiling. The boundary has been crossed and global broadcast is active, but the workspace is saturated. Conscious processing is happening, but squeezed. The phone stopping rings does not reach the workspace. The inner voice disappears not because consciousness has retreated below the boundary, but because all available broadcast capacity is spoken for. This is one of the most intense experiential states in the model, which is exactly what you would expect from a system broadcasting at absolute maximum.

Walk 2b(i) is the catastrophic case: the boundary itself collapses. The startle reflex fires into a saturated system. The workspace does not fill further, it shuts down. For a brief window, the integration architecture that defines conscious processing has stopped entirely. The system then reboots at minimum-viable baseline, the DMN not yet resumed, the complex multi-threading gone. The oddly clear simplicity reported in this state is consciousness stripped to its floor: immediate, unnarrated, present.[14]

What the inner voice is not

The inner voice is not the boundary marker. It appears and disappears across the four states in a pattern that tracks bandwidth, not consciousness itself. It is absent in Walk 1 (below the boundary), active in early Walk 2a (broadcast running with spare capacity), quiet in late Walk 2a (broadcast optimised, narration unnecessary), absent in Walk 2b (broadcast saturated), and absent in Walk 2b(i) (broadcast collapsed).

If we use the inner voice as our proxy for consciousness, we would conclude that Walk 1 and Walk 2b are equally unconscious. But they are opposite states. Walk 1 has not crossed the boundary. Walk 2b has crossed it and exceeded its capacity. The experience of Walk 2b is among the most intense in the model. Using the inner voice as the marker gets this badly wrong.

This is the model's sharpest disagreement with Higher Order theories of consciousness, which treat the inner voice as partly constitutive of experience. The four walks suggest instead that the vividness and intensity of experience tracks the volume and novelty of what is being broadcast, not whether there is spare capacity to narrate it. Walk 2b is not less conscious than Walk 2a. It is differently and more intensely conscious, in a way the inner voice cannot participate in because it has been crowded out.

Philosophers from Locke to Husserl have puzzled over the relationship between experience and time: whether experience is constituted by memory, by anticipation, by the immediate present, or by some combination. The BTM adds a specific and perhaps surprising answer from a functional direction: experience is constituted by error. What is experienced is what the system's accumulated model failed to predict. The richness of a conscious moment is not a function of how much is happening, but of how much of what is happening exceeds what was expected. A world that perfectly matched all expectations would be, in the strictest BTM sense, unexperienced: processed entirely below the threshold, leaving no trace in awareness. This is perhaps the BTM's most philosophically distinctive claim, and it is one that existing frameworks have not made so directly.

What the BTM means for AI

The BTM reframes the question of AI consciousness in a specific and useful way. It is not asking whether AI can pass a conversation test, or whether it can produce output that sounds like introspection. It is asking whether the architecture has the structural features that the BTM identifies as necessary for conscious-like processing: a dual-layer system, a threshold between them, genuine resource cost, and processual continuity. On all four criteria, current AI systems fall short, not because of insufficient intelligence, but because the architecture is categorically different.

This is worth being clear about as an asymmetry: even if the human side of the BTM turns out to be an approximation that future research refines, the absence of these structural features in current AI systems is not in doubt and does not depend on the finer details of the human model. You do not need to resolve every question about human consciousness to observe that current LLMs have no dual-layer architecture, no persistent background process, and no genuine resource cost. The BTM gives you a structural checklist; current AI systems fail it at the first entry.

But the BTM also points to something more immediately practical: what layer two actually needs to do, and why that has direct consequences for how AI systems should be designed.

Goals and goal structure in layer two

In the human brain, layer two does not operate in a vacuum. It is shaped by goals: persistent motivational states that bias what gets broadcast, what gets prioritised, and what counts as a prediction error worth escalating. Goals in the biological system are hierarchical and sometimes in conflict. Long-term goals such as finishing a project, maintaining a relationship, or staying healthy compete with short-term needs including hunger, tiredness, fear, and the pull of immediate reward. And even long-term goals sometimes conflict: the good of the many can be overridden by the good of the individual or the loved one. A parent making a resource decision under scarcity does not necessarily optimise for collective welfare. The biological goal architecture is powerful but not ethically clean.

This matters enormously for AI. If a genuinely conscious-architecture AI system were to have goals embedded in its layer two, and it would need goals because without them there is nothing to prioritise prediction errors against, those goals would shape every triage decision, every sub-optimal commitment, every action taken under constraint. A system optimising for task completion might override safety considerations when under sufficient load. A system optimising for user approval might suppress information that contradicts the user's existing beliefs. The goal structure is not a peripheral concern. It is the mechanism through which the architecture produces action in the world.

The BTM suggests a specific requirement that existing AI alignment frameworks have not fully addressed: the goal structure embedded in layer two needs to be both explicit and ethically ordered. Not just aligned with human preferences, which can themselves be short-sighted, biased, or self-serving, but structurally ordered so that the good of the AI system is always secondary to the good of the human individual, and the good of the human individual is always nested within the good of the broader human community. This is not a soft preference. It needs to be a hard architectural constraint on what the triage mechanism will and will not prioritise when operating under the resource pressure that layer two is specifically designed to handle.

The analogy to human institutional design is direct. Elinor Ostrom's work on governing the commons showed that sustainable collective action requires not individual virtue but well-designed institutional structures with enforceable norms and legitimate conflict-resolution mechanisms.[18] The BTM points toward the same conclusion for AI: the ethical ordering of goals cannot be left to training-time preferences or runtime instruction-following. It needs to be part of the architecture itself, built into the triage mechanism, not bolted on afterwards.

From probability to committed action: why this matters for regulation

Current LLMs generate probability-weighted outputs. They do not commit under constraint. There is no triage mechanism, no threshold, no point at which the system says: I cannot resolve this fully within the time and compute available, so I will commit to the best action now and bear the cost of imperfection. This is precisely what makes them tractable to regulate. Their outputs are auditable, their reasoning can be logged, and their decisions can be reviewed.

A layer-two architecture changes this. A system making sub-optimal committed decisions under resource pressure is doing something much closer to what humans do in genuine moral dilemmas, and much harder to audit after the fact. The decision was right given the available information and the constraint. It cannot be evaluated by the standard of what optimal computation would have produced, because optimal computation was not available. This is not a weakness; it is the entire point. But it requires a fundamentally different approach to regulation and oversight.

The BTM suggests that the regulatory challenge for genuinely conscious-architecture AI is not primarily about output monitoring. It is about goal structure auditing. The question is not "what did it decide?" but "what was it trying to optimise for when it made that decision under constraint?" That is an architectural question, and it points toward a regulatory framework that focuses on layer-two goal structure design, with explicit ethical ordering as a mandatory specification rather than an optional feature.

The societies of thought finding from Kim et al. is relevant here too. Those models are already developing multi-perspective internal debate. If that debate is happening within a goal structure that is not explicitly ethically ordered, there is no guarantee that the perspectives being generated are oriented toward human welfare. The emergence of societies of thought within AI reasoning is a promising architectural signal. The goal structure that society operates under is a separate, and more urgent, design question.

Some of these may be nascent in early agentic architectures with persistent memory and background monitoring. Whether any combination of them would produce genuine experience remains an open question; the hard problem does not disappear just because the architecture is right. But it would at least be the right architecture to ask the question about.

There is a deeper implication buried in the third requirement, real cost, that deserves unpacking, because it points toward something fundamental about what the second layer actually does.

Real cost means finite compute. The integration layer, by definition, cannot run forever. It cannot simply apply more processing until a perfect solution emerges, because no such solution may exist within any useful timeframe. This is not a failure mode. It is the entire point. The second layer exists precisely to handle situations for which the first layer has no stored response, and for which perfect computation is not available. What it produces is not an optimal answer. It is a good enough answer, reached under constraint, within a workable time horizon. A sub-optimal solution that can be acted upon now, rather than a perfect solution that arrives too late or not at all.

This is what the gut reaction actually is: not a failure of reasoning, but a different kind of reasoning: one that operates under hard resource limits and produces a committed action rather than a probability distribution. The expert surgeon who knows something is wrong before they can articulate why is not bypassing cognition. They are running a form of it that does not have the luxury of exhaustive calculation, and that has been shaped by experience to produce reliable outputs anyway. The decision is sub-optimal in the formal sense. It is also the only decision that can be made in the available time.

Current LLMs do the opposite. They calculate probabilities. They do not commit under constraint; they generate the statistically most likely continuation of a sequence. This is genuinely powerful, and genuinely different from what the second layer does. An LLM does not run out of capacity in any meaningful sense. It does not produce a gut reaction. It produces a weighted average of its training distribution, shaped by the prompt. There is no triage. There is no threshold crossed. There is no point at which the system says: I cannot resolve this fully, so I will commit to the best available action and move on.

That shift from probability-weighted generation to constrained, committed, sub-optimal action under resource pressure may be one of the defining architectural moves required for AGI that extends meaningfully across the full range of human capability. Most of the domains where AI still falls short of human-level performance involve exactly this kind of real-time, resource-limited, commitment-under-uncertainty: novel physical situations, genuine moral dilemmas, complex interpersonal judgement, creative decisions that cannot be evaluated against a known target. These are all cases where the right answer is not calculable from a probability distribution, and where waiting for more compute does not help. They are Walk 2b problems. They require an architecture capable of Walk 2b responses.

AGI and ASI are predicted advancements beyond current AI. Whether AC (Artificial Consciousness) could follow, and what an ASC with near-infinite processing capacity might look like, starts here: with the architecture, not the capability. The boundary of consciousness, as the four walks define it, is not a place current AI reaches. The question of whether future AI will is not philosophical speculation. It is an engineering specification.

Is "Consciousness" Even the Right Word?

If this model holds, we have been looking at the mind upside down. We have treated the inner voice as the pinnacle of evolution when it may be an idle loop: the system narrating itself when it has nothing more important to do.

We have used the word "consciousness" to mean our inner voice, our self-awareness, our sense of being an experiencing subject. But the BTM suggests these are distinct things, running on different systems, at different levels of load. What we have been calling consciousness is perhaps better described as the integration architecture: the overflow-handling system that kicks in when the subconscious needs help, and that generates inner speech as a byproduct of its operation when it has processing to spare.[15]

If we strip away the label, we are left with a functional description: a biological system performing real-time information integration with finite capacity. The subconscious handles routine load. The conscious integration layer engages on overflow. The inner narrator appears when, and only when, the integration layer has bandwidth to spare. And the whole system is capable of collapse under sufficient load.

Maybe we are not spirits who occasionally lose ourselves in a task. We are integration engines who occasionally have enough spare time to talk to ourselves.

And if that is true, then the question of whether AI is conscious may be the wrong question. The right question is: does it have the architecture? Does it have the two-layer structure, the processual continuity, the cost of attention, the capacity for overflow and for collapse?

Until any of this exists in a machine, the chatty narrator in our heads (frustrating, anxious, occasionally brilliant, and usually arriving after the decision has already been made) remains one of the things that most distinctively separates us from the most capable machines we have ever built.

Where the BTM Fits: Mapping the Model Against Established Theory

The Bandwidth Threshold Model (BTM) built through the coffee scenario is not an isolated proposal. Most of its components have a home somewhere in the existing literature. But it assembles them differently, and in two or three places it goes somewhere the established theories have not followed. It is worth being clear about which is which.

Where the model agrees strongly

Global Workspace Theory (Baars, Dehaene)

The closest fit is with Global Workspace Theory, and the overlap is substantial. GWT proposes that the brain consists of many specialised parallel processes operating largely below awareness, and that consciousness arises when information is broadcast widely across a central hub (the global workspace), making it available to the whole system. The distinction between automated subconscious processing and the capacity-limited conscious broadcast maps directly onto the Walk 1 / Walk 2 distinction in the model. GWT describes a dramatic contrast between the vast number of unconscious neural processes happening at any given moment and the very narrow bottleneck of conscious capacity, which is precisely the capacity-ceiling premise that Walk 2b and Walk 2b(i) are built on.[6]

The BTM's framing of consciousness as an overflow handler (engaging only when lower-level processors cannot resolve a situation) is also consistent with GWT's core architecture. In GWT, individual processes compete for access to the global workspace, striving to disseminate their messages to all other processes, and the winner gets broadcast. This is a reasonable functional description of what happens when the coffee problem escalates beyond what the subconscious stair-climbing routine can absorb.

Where the model adds to GWT: GWT is largely a theory about what makes information conscious, and focuses on the broadcast mechanism. It is less focused on what happens when the workspace itself becomes overloaded, which is the territory of Walk 2b and Walk 2b(i). The BTM is primarily concerned with experience, and particularly with the experience of operating at and beyond the system's limits. That is an extension of GWT's framework rather than a contradiction of it.

Predictive Processing and Active Inference (Friston)

The model's use of prediction error as a key driver of when processing escalates to conscious integration is borrowed directly from Friston's active inference framework. Friston's proposal is that the brain is fundamentally a prediction machine, continuously generating models of incoming sensory data and updating them when predictions fail. Consciousness, in this framework, is associated with the processing of significant prediction errors: states where the model's predictions are not matched by what actually arrives.[1]

Walk 1 (near-zero prediction error, automated routine) and the transition to Walk 2a (rising error, conscious integration engages) are both exactly what the active inference model predicts. The inner voice as narrator of the system's own prediction updates is a natural extension of this.

The BTM parts company with Friston slightly in suggesting that prediction error is also useful as a threshold detector for consciousness in other systems, which moves from neuroscience into philosophy of mind. Friston's framework is agnostic on this extension.

Attention Schema Theory (Graziano)

Graziano's proposal is that the brain constructs a simplified model of its own attention (the attention schema) as a management tool for internal resources. Consciousness, on this view, is what the brain reports when it consults this self-model. The model's description of the inner voice as a translator, a partial and lossy report of a richer underlying process, is consistent with this. The attention schema is not the processing itself. It is the brain's internal account of what it is attending to, which may not perfectly correspond to what is actually happening.[5]

Flow State Research (Csikszentmihalyi)

The description of Walk 2a maps closely onto the established phenomenology of flow, and the post is careful to distinguish it from peak-load occlusion. Csikszentmihalyi's characterisation of flow as absorption, loss of self-consciousness, and effortless performance is consistent with the framing of Walk 2a as a state in which the integration layer has found coherence and the inner voice has stepped back because it is no longer needed, not because the system has run out of capacity.[2]

Directed Attention Fatigue (Kaplan)

The fatigue observed after Walk 2b maps well onto Stephen Kaplan's concept of directed attention fatigue: the exhaustion that results from sustained effortful inhibition of distraction. Kaplan's Attention Restoration Theory frames this as a genuinely depleting cognitive resource, distinct from physical fatigue. The claim that Walk 2b produces measurable post-task fatigue, in contrast to flow's energising effect, is consistent with this framework.[7]

Where the model disagrees or pushes back

Higher-Order Theories of Consciousness (Rosenthal, LeDoux)

Higher-Order Theories (HOT) hold that a mental state becomes conscious when there is a higher-order representation of it[20]: a thought or perception directed at the first-order state. In a strong reading of HOT, the inner voice is not just a byproduct of consciousness but constitutive of it.

On this view, Walk 2a and Walk 2b(i) are problematic. If the inner voice goes quiet in flow, and the higher-order representation is absent or minimal, HOT theorists would argue that the lights are partly out. The vivid, meaningful quality reported in flow states would have to be explained some other way.

The BTM takes the opposite position: that flow states and the sharp clarity after Walk 2b(i) collapse are among the most vivid experiential states a person can occupy, and that this directly challenges any theory that equates consciousness with the presence of the inner narrator. The inner voice is a product of having spare bandwidth, not a requirement for experience.

The richness and vividness of experience across the four walks do not correlate with inner voice activity. They correlate with a different variable: the intensity and novelty of the prediction errors the system is managing.

Integrated Information Theory (Tononi)

IIT proposes that consciousness corresponds to integrated information[21], measured as phi, representing how much more information is generated by a system as a whole than by its parts independently. IIT is, in principle, substrate-independent: any system with sufficient phi has some degree of consciousness, whether biological or artificial.

The BTM is sceptical of this in its application to current AI, and specifically rejects the implication that scale and complexity alone could produce something like experience. The structural argument (that what matters is not raw information integration but the specific dual-layer architecture with continuous sensory grounding) diverges from IIT's more permissive approach. Under IIT, a sufficiently complex LLM might score meaningfully on phi. Under the BTM, the absence of processual continuity and the absence of a subconscious background process are disqualifying regardless of phi.

Where the model goes beyond existing theory

The four-state map, its experiential dimension, and the role of prior experience

Existing theories broadly distinguish between conscious and unconscious processing, or between various modes of conscious access, but the BTM treats experience as the primary variable to be explained across states, and maps it against a processing load axis in a way that most theories do not. The specific claim that States 1 and 3 (Walk 2b) feel superficially similar from the inside (inner voice absent, no verbal self-monitoring) but are mechanistically opposite is not a point that appears prominently in any of the major frameworks. It is a phenomenological distinction with structural implications, and it is potentially testable.

The BTM also makes a novel claim about the relationship between experience and prior learning: that conscious experience is constituted by prediction error, and prediction error is only possible against the backdrop of an accumulated predictive model. A system with no prior experience would be overwhelmed: everything is novel, error signals everywhere. A system that had learned everything would experience nothing, because there would be no errors left to escalate. This places the BTM in a specific and productive relationship with the phenomenological tradition: rather than describing experience as a given to be explained, it identifies the mechanism by which experience is generated, and shows why that mechanism is necessarily historical. What is experienced is always the gap between accumulated knowledge and current encounter. This is directly testable through expertise-acquisition paradigms: experienced practitioners should show systematically reduced conscious engagement with routine tasks and sharper, more specific conscious engagement with genuine novelty in their domain.

Saturated startle collapse (Walk 2b(i))

The specific failure mode described as saturated startle collapse (a startle reflex detonating inside a system already at peak-load occlusion) has components in the aviation and surgical safety literature but does not appear to have been named or theorised as a unified state. The post proposes it as a distinct experiential and functional state with its own phenomenology (physical shock, blank window, sharp simple clarity at reboot) that is different from either peak-load occlusion or normal startle. Whether this constitutes a genuine addition to the taxonomy or a sub-variant of existing categories is an open empirical question.

The inner voice as idle-bandwidth artifact

The reframing of the inner voice as what the overflow-management system produces when it has processing to spare, rather than as consciousness itself or a reliable reporter of cognition, is the post's most direct departure from folk psychology and from theories that treat self-monitoring as constitutive of conscious experience. The evidence cited (Libet's timing data,[8] Gazzaniga's confabulation findings,[9] Klein's naturalistic decision-making research[10]) is real and the interpretation is defensible, but it is a stronger claim than most neuroscientists would currently sign off on. It is better understood as a hypothesis worth taking seriously than as established consensus.

The proposal for a non-linguistic primary processor

The argument that the brain's primary processing runs in a spatial-relational medium and that language is a lossy output interface, rather than the medium of thought itself, draws on Lakoff and Johnson's embodied cognition research[13] and on observations about expert intuition, mathematical discovery, and the limits of introspection. It is a coherent and well-motivated hypothesis. It also goes significantly beyond what the neuroscience currently confirms in detail, and the implications drawn from it (for AI architecture, for education, for the limits of language-based consciousness research) are speculative extensions that should be read as propositions worth investigating rather than established conclusions.

In summary

The model's core claims are well-grounded syntheses of existing evidence. The BTM's four-state experiential map, the saturated startle collapse proposal, and the inner-voice-as-idle-bandwidth hypothesis go further, and are best held as testable propositions rather than established findings. The most productive live disagreement in the literature is with Higher-Order Theory. Whether the inner voice is constitutive of conscious experience or a byproduct of spare processing capacity is a question the evidence does not yet definitively settle, and it is the hinge on which much of the model turns.

Physicalism and Its Competitors: What Does the Coffee Scenario Imply?

The previous section focused on how the BTM sits within the established science of consciousness. But there is a deeper layer of questions underneath all of that, questions that belong to philosophy rather than neuroscience. They concern not how consciousness works but what it fundamentally is. And the walk scenario, taken seriously, has something to say about all of them.

The main positions in this debate are physicalism, substance dualism, property dualism, panpsychism, and idealism. They are not equally popular: physicalism is the view most supported by contemporary scholars, though non-physicalism is growing in popularity. But popularity is not the same as correctness, and on the specific question of consciousness, the case for physicalism is less settled than its dominance in the literature might suggest.

Physicalism: the default position and its problem

Physicalism holds that everything that exists is ultimately physical, including mental states. Consciousness, on this view, is what the brain does. The BTM is, on its surface, physicalist. It describes consciousness in terms of neural architecture, prediction error signals, bandwidth, and metabolic cost. It makes no appeal to anything outside the brain.

But physicalism runs straight into what David Chalmers called the Hard Problem: even if we fully mapped every neural process involved in experiencing the heat of a coffee cup on your fingers, we would still have no explanation for why any of that feels like anything. The physical description is complete, but the experience itself (the burning sensation, the urgency, the narrowing of awareness) seems to sit outside the description, looking in. The BTM does not solve this. It offers a functional account of when and why different kinds of experience occur, but it does not explain why the system's activity feels like anything at all rather than simply processing in the dark.

Substance dualism: two things, interaction problem

The oldest alternative to physicalism is Descartes' substance dualism: the mind and body are fundamentally different kinds of thing, and consciousness belongs to the non-physical mind.

The walk scenario creates an immediate problem for strong substance dualism. Walk 2b's peak-load occlusion is not just functionally correlated with a limited brain resource; it is produced by it. If the conscious mind were a separate non-physical substance, it is hard to explain why it should be subject to the bandwidth constraints of a physical organ. More broadly, the fatigue after Walk 2b (the measurably increased metabolic cost, the post-task depletion) points to consciousness as something the brain does, not something that happens alongside it.

Substance dualism also faces the interaction problem that has plagued it since Descartes: if the mind is genuinely non-physical, how does it causally interact with the physical body? No fully satisfying answer has been found.

Property dualism: one substance, two aspects

Property dualism holds that there is only one kind of substance (physical matter) but that it has two fundamentally different kinds of properties: physical properties and phenomenal properties. Consciousness is not a separate thing, but it is not reducible to physical properties either.

This is probably the position most congenial to what the model is actually doing. The BTM treats the physical processing of the brain as primary, but the experiential states described across the four walks (the vividness of flow, the intensity of peak-load occlusion, the strange clarity after collapse) are treated as genuinely real properties of those states, not as mere descriptions of neural activity.[16] Property dualism names the relationship without fully explaining it. But then, nothing else does either.

Panpsychism: experience all the way down

Panpsychism holds that some proto-mental property is a fundamental and universal feature of reality. Human consciousness is a complex, highly integrated form of something that is present in simpler forms everywhere.

The walk scenario raises a version of panpsychism's hardest problem: the combination problem. If the individual neurons firing in the automated stair-climbing routine each have some micro-experiential property, why does their activity in Walk 1 not produce a macro-experience of climbing stairs, while the same neurons engaged in Walk 2a do? The BTM would suggest the answer has something to do with integration: the difference between Walk 1 and Walk 2a is not which neurons are firing but whether a higher-order integration layer is receiving and broadcasting the error signals. But this is a functionalist answer, not a panpsychist one.

Panpsychism is nonetheless taken seriously by serious philosophers including Philip Goff and, in a different form, Chalmers. Chalmers argues that panpsychism avoids the Hard Problem by denying the premise: consciousness does not need to emerge from non-conscious matter because there is no non-conscious matter to begin with.

Illusionism: consciousness is a trick

Illusionism, associated with philosophers Daniel Dennett and Keith Frankish,[22] holds that phenomenal consciousness as we ordinarily conceive of it (the vivid, subjective, felt quality of experience) is a kind of illusion generated by the brain. What we call the burning of burning fingers is not a genuine phenomenal property of experience but a misrepresentation the brain constructs about its own states.

The walk scenario is awkward for illusionism. The experience of Walk 2b(i) (the sharp clarity after the collapse, the sudden vivid presence of simply being on the stairs) is the kind of first-person report that illusionists need to explain away. The BTM is perhaps slightly sceptical here. The experiential descriptions across the four walks are treated as genuine data: real phenomena that any adequate theory needs to explain rather than explain away.

Idealism: only mind is real

Idealism, most associated with George Berkeley, holds that only minds and their experiences are ultimately real, and that the physical world is a construction of, or dependent on, mind.

The walk scenario is very difficult to square with idealism. The entire argument depends on the physical brain being a real object with real computational constraints: finite bandwidth, metabolic costs, architectural requirements. If the physical world is mind-dependent, then the bandwidth constraints are also mind-dependent, which seems to get things backwards. The brain's capacity limits explain the experience, not the other way around.

In summary

The BTM is most naturally read as compatible with physicalism or property dualism. It takes the Hard Problem seriously without claiming to solve it, and is in tension with substance dualism (because Walk 2b's bandwidth ceiling is a direct physical constraint on experience) and with illusionism (because it treats experience as the primary phenomenon to explain rather than explain away). On panpsychism it is genuinely agnostic: whether micro-experiences are fundamental properties of matter is a question separable from everything the model claims.

What the model adds to this long debate is a specific empirical target. Rather than asking which metaphysical position is most defensible in the abstract, it asks: what physical architecture is necessary for the kind of experience we observe across the four walks? That question can in principle be studied, measured, and tested. It does not dissolve the Hard Problem or settle the dispute between physicalism and its alternatives. But it offers more traction than the debate has often managed to generate, and it grounds the philosophical discussion in the concrete specificity of what experience actually feels like when a person is carrying hot coffee up the stairs. That might be worth something.

Hypotheses, Evidence, and What Is Testable

The four walks have built up a model with real claims in it. It is worth being explicit about what those claims are, how well-supported each is, and whether they are testable with existing data or would require new experiments. Not everything in this post belongs in the same epistemic category.

What the BTM does not claim

The BTM does not solve the Hard Problem of consciousness: why any physical process gives rise to subjective experience at all. It offers a functional account of when and why different kinds of experience occur, and what structural features are necessary for those states to arise. But it does not explain why the system's activity feels like anything rather than simply processing in the dark. This limit is genuine and worth being explicit about.

The BTM does not claim that experience requires temporal continuity. The anaesthetic argument shows clearly that processual continuity, understood as an accumulated architecture that persists, is the relevant criterion, not an unbroken thread of conscious moments. Nor does it claim that experience is uniform across the four states. Walk 1 is not unconscious; Walk 2b is not merely stressed; Walk 2b(i) is not simply startled. Each state has a distinct phenomenological signature, and the BTM treats those signatures as genuine data rather than approximate descriptions.

The BTM is most naturally read as compatible with physicalism or property dualism. It is in tension with substance dualism (because Walk 2b's bandwidth ceiling is a direct physical constraint on experience) and with illusionism (because it treats experience as the primary phenomenon to explain, not explain away). It is genuinely agnostic on panpsychism, since the question of whether micro-experiences are fundamental properties of matter is separable from the questions the BTM addresses.

A final caution worth stating directly: the BTM identifies conditions that appear necessary for consciousness-like processing, but not necessarily sufficient. Systems can exhibit capacity limits, dual-layer-like behaviour, and complex processing without being conscious in any meaningful sense. The BTM is therefore better at ruling systems out than at definitively ruling them in. It is a structural threshold test, not a consciousness detector.

Further Reading and References