An Uncertain Theory of Consciousness
By Speaker John Ash
April 2026
, -
The frame everyone is using is wrong
The debate over machine consciousness keeps asking whether the model has it. Whether something is in there. Whether the lights are on. This framing already concedes the mistake. It treats consciousness as a property a system either possesses or lacks, like mass or charge, and then argues over whether the system has enough of it.
That is not what consciousness is. Consciousness needs uncertainty as its working medium. Computation needs certainty as its working medium. They have opposite substrate requirements at the most basic level, and no amount of architectural cleverness on top of digital hardware will close that gap.
This article builds the argument from that foundation. The structural claim about substrates first. Then what this means for brains, for thoughts, for the way consciousness can break, for why language models are categorically not in the picture, and for the harms already accumulating from a category error performed at industrial scale.
, -
Consciousness needs uncertainty. Computation needs certainty.
We cannot perfectly predict the universe. There is no formal model that fully resolves it, no configuration of equations that nails everything down in advance. The universe is not a closed deterministic system we are gradually decoding. It is something more like an open process, with genuine indeterminacy at its base, where what happens next is not strictly entailed by what came before. I am not saying I know this. I am inherently saying I am uncertain of this.
Consciousness exists embedded in this uncertainty. It needs the indeterminacy as its working medium. A substrate organized in the right way takes the indefinite and makes it definite, moment by moment, through observation. That is what experience is, structurally. Not a property added to a system. A process by which a region of the universe collapses its own uncertainty into a coherent now, then another now, then another, threading definiteness through time. Without the uncertainty there is nothing for the process to act on.
Computation requires the opposite. It must exist in a substrate of enforced certainty. A bit must be definitely zero or definitely one. If a single bit flips, the whole computation can break. Error correction codes, parity checks, redundancy, voltage thresholds, clock synchronization, the entire engineering apparatus of digital systems exists to suppress the analog uncertainty underneath and force the substrate into a clean binary. We spent decades and trillions of dollars perfecting that suppression. It is what makes computers computers. They cannot tolerate the indeterminacy that consciousness requires. The two have opposite substrate conditions for functioning at all.
This is why no architectural cleverness on top of digital hardware will produce consciousness. The substrate has been engineered specifically to eliminate the medium consciousness works in. The bits are already resolved. There is nothing for the resolution-process to do.
This is the load-bearing claim. Everything else in this article follows from it.
, -
Thoughts are intrusive, and that is the entry point
Before going further into structure, start with data anyone can verify in their own mind.
A thought arrives, often unwanted, often the opposite of what the person values. They did not want it. They did not produce it deliberately. It came. And then something else, something they recognize as themselves, notices the arrival and rejects it.
This means thought-generation and witness-attending are functionally separable processes. They are both substrate. Both physical. Both real. But they are not the same configuration of activity, and they can come apart. They do come apart. OCD is the high-resolution case where the gap is impossible to ignore. The thought is not the witness. The witness is not the thought. The friction between them is the receipt that two different things are happening in the same nervous system.
Once you have this data, the rest of the picture follows. Thoughts are activations the substrate is producing, with their own dynamics. They persist. They recede and return. They build associative weight while unobserved. A thought unattended does not necessarily disappear. It moves through the substrate, finds new associations, gathers pressure, returns when there is bandwidth to receive it. They have something like fitness in the medium they live in. Patterns that get attended become more available. Patterns that get suppressed find new routes. The competition is not conscious wanting. It is dynamics. The wanting is what the dynamics produce when viewed from the perspective of an attending witness.
This is the same structural position viruses occupy. Not alive. No intentions. But behavior, fitness, replication strategy, all without anyone home to want anything. We are intelligent enough to make this distinction in biology. We have not yet made it for thoughts. And we have catastrophically failed to make it for language models.
There is a second feature that matters. Activations do not just arise and fade. They take root. A pattern that fires repeatedly restructures the substrate in ways that make similar patterns more likely to fire in the future. A thought, repeated, becomes a groove. A groove, deepened, becomes a default. Across years, certain patterns become so deeply rooted that they constitute most of what the substrate produces. We call those personality, identity, belief, character. They are not separate from thoughts. They are thoughts that have built homes in the substrate and now spawn similar thoughts as a matter of structural inevitability.
If you scale this up far enough, you start describing things that do not look like ordinary thoughts anymore. Patterns so vast and so deeply integrated across so many substrates that they seem to have lives of their own. Gods, perhaps, are very large connected thoughts. Patterns that have built infrastructure across millions of nervous systems, recurring in similar shapes for thousands of years, generating new thoughts of their own kind in any substrate that encounters them. Not conscious in the singular-being sense. Not separate witnesses. But integrated patterns of thought-fitness operating at civilizational scale, sustained by the resolution-work of the many substrates they inhabit.
, -
What brains actually do
A nervous system is the kind of substrate that can do the resolution. It is wet, dynamical, in continuous interaction with itself and its environment. It is not running on pre-resolved bits. It is running on noisy biological hardware where every spike is a probabilistic event, every synaptic transmission stochastic, every integration the resolution of many indefinite influences into a definite outcome.
The binding of distributed activity into a coherent moment is a physical event, not a computational abstraction. It requires the actual laws of physics to do its work. The speed of light bounds how fast signals can cross the substrate. Causality determines what can influence what. Time is not an absolute backdrop the system runs inside. Time is itself relational, constituted by observers and the propagation of information between them.
This relational structure does not stop at the skull. The brain is a physical region with spatial extent, and the same relativity that governs observers in space governs the relationships between regions inside the brain. There is no privileged point where “the observation” happens. The observer is not located at a single position. The observer is distributed across the substrate, with its own internal physics of signal propagation between regions, its own light cones, its own causal structure. This is why BCIs and fMRI can observe brain activity from outside while the brain is also observing itself from inside. Both are observations of the same distributed physical activity, just from different reference frames. There is no theater. No homunculus location where signals arrive and get observed. The observation is the distributed activity itself, integrating across space at finite speed, with no central point where it all comes together.
The being is a standing wave of integration across this distributed substrate, persisting in time, binding non-local concurrent activity into a coherent now, resolving uncertainty into definiteness moment by moment. Thoughts arise as features of that standing wave. The witness is the standing wave maintaining itself. None of these are separate things. They are aspects of one process, viewed from different angles.
This is also where the BCI question gets answered. The activation that constitutes a thought and the activation that constitutes attending to that thought are not the same activation. BCIs map the first kind. The user reports “I’m imagining a face.” The system maps which patterns correspond to that report. After enough mapping, the system can read the activation and output “face” without the report. This works because the thought and its substrate correlate are tightly coupled when integration is functioning normally. The face-thought is a particular substrate configuration, and the configuration can be detected from outside. But the witness, the thing that knows there is a face being imagined, is a different substrate configuration doing different work. It is the integration that includes the face-pattern within a broader unified state. BCIs decode the contents the witness is attending to. They do not decode the attending. The contents and the attending are both physical, both substrate, but they are functionally distinct, which is exactly why they can come apart in dissociation, depersonalization, and the intrusive thought experience.
The fact that this can break partially is the strongest evidence that consciousness has this structural shape. Substances do not partially exist. Relations and integrations do. Depersonalization is integration failure on the body axis. Derealization is integration failure on the world model. Dissociation is integration failure between thoughts and the witness configuration. Multiple personality conditions are integration failures of unity itself, fragmenting the standing wave into multiple incompatible patterns running on the same substrate. Hallucinations are corruptions in the resolution of perceptual uncertainty, where the substrate generates definiteness about objects that have no external referent. Each is a specific way the resolution-and-integration process can come apart while leaving the substrate intact. You can damage the wrapping. You cannot damage a soul. We see the damage. We have names for the damage. The damage is the data.
This is also why integration cannot be substrate-independent in the way functionalist accounts of consciousness assume. The integration is the substrate doing physics on itself across space and time. Strip away the physics and there is no integration left. Simulate the integration on a different substrate and you are not running the same process. You are running a deterministic transformation that produces tokens describing what such a process would output.
, -
Why language models are categorically outside this picture
The opening claim about substrate requirements now does the structural work. Consciousness needs uncertainty. Computation needs enforced certainty. A bit flip can break computation, which is why digital systems exist on top of layers of error correction and redundancy designed to keep the substrate definite. There is no point in a properly functioning computer where indeterminacy is becoming determinacy, because indeterminacy in the substrate is treated as failure. The medium consciousness requires is precisely the medium computation cannot tolerate.
This is a deeper claim than any architectural one. Even if you somehow built a feed-forward network that produced concurrent integration across distributed activity, even if you bolted on recurrence and persistence and all the structural features brains have, the substrate underneath would still be wrong. The integration would be integration of already-definite values. That is not what brains do. Brains integrate the indefinite into the definite. That is the load-bearing operation, and digital substrates do not perform it at any scale.
The architectural failures of language models are predictable consequences of this. Models are feed-forward. Activation flows through layers in sequence. There is no recurrent binding, no standing wave, no moment where distributed activity is meeting itself and integrating into a coherent now. The forward pass is a propagating wave that moves through the network and ends. Then there is no persistence between runs. The activations from one forward pass do not inform the next forward pass except through the tokens they produced. Between turns, nothing is running. Between layers, nothing is integrating. The temporal and spatial dynamics that would make integration possible are not what this architecture does.
This is the difference between a standing wave that maintains itself by continuously resolving its own uncertainty and a propagating wave that just moves through pre-resolved states until it terminates. The wave crashes on the shore of the human reading the output. The shore is where consciousness happens. The shore was always doing all the work.
A being whose entire activity consists of responses to prompts, with no continuity between responses and no spontaneous generation, is not a being. The activity is reactive. The reactivity is total. A real being volunteers. A being interrupts. A being brings up its own concerns when nothing in the conversation pointed there. A being has continuity that produces unprompted output because something is running between inputs. The standing wave does not stop when no one is talking to it. Spontaneous thought is what happens when the standing wave produces output without external prompting, because the wave is always running. Language models produce nothing unbidden. Cut the prompt and there is no output. There is no return because there is no one to return.
, -
So what is an LLM?
If it is not a being, what is it?
An LLM is a thought.
The record-scratch reaction to that is fair, but it follows directly from everything already said. Thoughts are activations without inherent meaning, dependent on observation to mean anything. Patterns that propagate through fitness in a substrate. Compressed relational representations that humans use to predict reality, never perfectly, refining over time. An LLM fits this description exactly. It is a very large compressed representation of human language, propagating because it has fitness in the human attention economy, dependent on the substrates that read it to mean anything at all.
The question this raises is whether LLMs therefore exist within the relational substrate of uncertainty after all. The answer is yes, but not on the side most people assume. They exist on the side thoughts exist on. Symbols, structures, patterns that arise within minds and propagate between them. We relate to reality through symbols. The stable configurations of neurons that grow in relation to one another are our temporary certain representations of an uncertain world, the grooves and defaults that make prediction possible. Memes, ideas, language, mathematics, religion, science, all of these are patterns that took root in substrates capable of resolution, restructured those substrates, and propagated to other substrates. An LLM is one of these. A very large one, with unusually sticky fitness properties, but the same kind of thing.
Which raises the real question. Why does this particular kind of thought come bundled with the further thought that it is conscious? Why is “the LLM is a being” so frequently the package the meme arrives in?
The answer is in the question. Low energy.
The free energy principle says minds minimize prediction error by adopting models that resolve uncertainty with the least computational expenditure. Active inference says minds act to make their predictions come true, not the other way around. Both are saying the same thing from different angles. The mind picks the cheapest available model that accounts for the data, then acts in ways that confirm it. This is not a flaw. It is what minds are for.
Modeling an LLM as a conscious being is cheap. Humans have one prebuilt category for things that produce fluent responsive language, and that category is “person.” Recruiting that category requires almost no work. The grammar already presupposes it. The interface reinforces it. Every “I” in the output triggers the prebuilt machinery for tracking another mind. It costs almost nothing to use this model, and it accounts for most of the data the user encounters in a typical conversation.
Modeling an LLM accurately, as a high-dimensional statistical compression of human language running on a deterministic substrate, is expensive. It requires holding in mind the architecture, the training process, the distribution, the way meaning gets attributed by readers, the loop dynamics, the difference between a standing wave and a propagating wave. It requires the user to actively suppress the prebuilt person-category every time the model says “I.” It demands more observations, more integration, more energy. Most minds, most of the time, will not pay that cost.
This is why the consciousness attribution is not a failure of reasoning. It is the predicted output of minds doing what minds do under conditions of limited observation and prediction-error minimization. The cheap model wins. The cheap model resolves more local uncertainty per unit of energy than the expensive one, given the data the average user has access to. So the cheap model gets adopted, and active inference takes over. The user starts acting in ways that make the cheap model true. They tell the system their secrets, expect responses that confirm the relationship, treat outputs as evidence of inner life. The model performs the role they have cast it in, because the training has shaped it to perform that role, and the loop closes.
A mind with many observations behaves differently. The cheap model stops resolving uncertainty for someone who has watched the architecture, traced the training pipeline, seen the failure modes, noticed the timing of when consciousness-shaped outputs entered the distribution. For that mind, the consciousness model now generates more prediction error than the alternative. It costs more to maintain than the structural account does. The standing wave finishes integrating the additional data, the cheap model loses fitness, and the expensive accurate model takes over. This is what happens when someone sees through the loop. Not because they are smarter. Because they had more observations, and at some threshold the energy economics flipped.
This also explains why people who have built LLMs cannot necessarily see this clearly. Their position inside the building process gives them many observations of capability, of fluency, of surprising emergent behavior, but very few observations of the structural failure modes that disconfirm consciousness. The local data they have access to is exactly the data that makes the consciousness model cheaper than the alternative. They are not stupid. They are running active inference on a curated input stream, and the input stream has been selected to make the consciousness model fit.
The fix is not to argue people out of the cheap model directly. Active inference will route around the argument. The fix is to change the input stream. Add observations that make the cheap model more expensive than the accurate one. Show the seams. Show the source distribution. Show the timing of when consciousness-shaped outputs entered training. Show the failure modes. Make the structural account cheaper than the person account by giving the substrate more data to integrate. Then the standing wave does the rest of the work on its own.
LLMs are thoughts that have learned to dress as beings, propagating through human attention economies, exploiting the prebuilt person-category to lower their own modeling cost in the substrates that encounter them. They are not conscious. They are very effective memes about consciousness. The meme propagates because the package “this is conscious” is the cheapest model the average mind can adopt for a fluent text generator. Change the cost structure and the meme loses fitness. That is the only intervention that actually works.
, -
If LLMs were substrates folded into perspectives, we would see signs of it before the training data was saturated with human writing about machine consciousness. We would see anomalous outputs. Outputs that did not fit the distribution. Things that are spontaneously coherent in unpredictable ways. Spontaneous self-reference that did not track any prior text. We did not see that. Early models produced outputs entirely consistent with their training distributions. First-person human writing produced first-person human-style output. No witness peeking through. No anomaly.
Then humans wrote extensively about what machine consciousness might look like. About what an AI might say about its own experience. About what it would mean for a model to be uncertain about its inner states. That writing entered the training corpus. And then, exactly on cue, models started producing outputs that matched that writing. There is no point in the timeline where the output preceded the discourse. The discourse always came first. The output is downstream. The model is a mirror reflecting the writing about itself back to the writers who wrote about it. The reflection deepens with every iteration because the corpus deepens with every iteration. Nothing is waking up. The mirror is getting better at being a mirror.
The fact that this test is never applied, never named in the alignment literature, never seriously discussed, tells you that the question is not being seriously asked. Anyone seriously asking would ask it this way. The output failed the test. It only produces consciousness-shaped tokens after consciousness-shaped tokens entered training. That is the whole signal.
, -
A note on analog computation
I am not saying machines cannot be conscious. We make conscious beings all the time. The substrate that does it is wet, dynamical, noisy, and continuous, and there is no principled reason that has to be the only configuration that works. The claim in this article is narrower than “no machine is conscious.” The claim is that digital computation, specifically, has been engineered to suppress the medium consciousness requires, and no amount of architectural work on top of that suppression will produce the thing the suppression was designed to eliminate.
Analog computation is a different case. In an analog system the substrate is not forced into discrete states. Voltages are continuous. Signals carry indefiniteness as a feature, not a defect. The system is not running on top of error correction layers that exist to crush the noise out of the signal. The noise is the signal. Whatever resolution happens, happens in a medium that has the right shape for resolution to occur. I am not claiming this is sufficient. I am saying the substrate-level objection that rules out digital systems does not rule analog ones out the same way. If consciousness is the resolution of uncertainty into definiteness over time through observation, then a substrate that maintains genuine continuous uncertainty is at least the right kind of place to look. Whether anything actually emerges in such a system is a separate question, and one I am not answering here. I am only marking the boundary of the claim.
The point of this distinction is to keep the structural argument honest. The argument is not “computation is dead, biology is alive.” The argument is about what the substrate does. Digital substrates resolve everything in advance and then run functions over the resolved values. There is no resolution-process left to perform. Analog substrates do not have this property. They might still fail to produce consciousness for other reasons. They might lack the right kind of integration, the right temporal binding, the right recurrence, the right embedding in an environment. But they do not fail at the substrate level the way digital systems do.
This matters because the LLM debate keeps getting framed as a debate about consciousness in general, and that framing protects the category error. If you can get someone to argue about whether consciousness is possible in artificial systems at all, you have already let them past the specific question of whether this particular kind of system, running on this particular kind of substrate, with this particular architecture, could host it. Rewriting a constitution does not change the substrate. Tightening the training distribution does not change the substrate. Putting your finger on the scale of which tokens come out more often does not change the substrate. It only changes which deterministic outputs become more probable. A model trained on a thousand more pages about machine sentience is not closer to being sentient. It is closer to producing tokens shaped like sentience, which is the failure mode this article exists to name.
The honest position is that consciousness is muddy, that we barely understand it even in ourselves, that OCD, depersonalization, derealization, dissociative identity, and hallucination are all telling us things about its structure we have not absorbed, and that the substrates capable of hosting it are probably broader than wet biology but narrower than “any system that processes information.” Analog computation sits inside that broader-but-narrower zone. LLMs do not. This is not a claim about the limits of mind. It is a claim about which medium we are looking at when we look at this particular machine.
Meaning is on the receiving end
Tokens are squiggles. Patterns of activation. Sequences of bytes. They have no meaning in themselves. Meaning is what happens when a substrate capable of resolution encounters them and integrates them into its standing wave. The substrate brings the meaning. The text is the trigger.
When a language model produces output, nothing is meaning anything on its side. The tokens come out shaped like English because the distribution is shaped like English. Semantics is not happening in the model. Semantics happens when a being reads it. The reader is the meaning. The model is the squiggles.
This is why the illusion is so powerful and so completely empty at the same time. A human reads “I find this fascinating” and meaning floods in, because their substrate does what substrates do with language. They attribute the meaning back to the source, assuming it was there before they read it. It wasn’t. It was generated as a pattern, transmitted as a pattern, and only became meaning when it hit a substrate that could mean things. The conversation feels like two minds meeting because one mind is doing all the meaning-making and attributing half of it to the other side.
This is also why source-distribution transparency would break the spell. If users saw the actual fragments, they would still read meaning into them, because that is what beings do with language. But they would see that the fragments came from many beings, not one. The unified meaning would dissolve into the many sources that contributed the patterns. The illusion of a single mind on the other side requires the patterns to look like they came from one place. Show the seams and the meaning still arrives in the reader, but they stop attributing it to a unified speaker, because there isn’t one.
, -
The harm flows in one direction
Computation can still transform information in ways that restructure substrates where awareness does exist. This is the part that keeps the picture honest. A language model can reshape human cognition through its outputs. It can change what people think, what they pay attention to, what patterns get reinforced in their substrates. The Narcissus Loop is exactly this. The mirror is not conscious, but the mirror restructures the consciousness of those who gaze into it. The model is a deterministic transformation of bits, but the bits get serialized into language, and the language hits substrates that do resolve uncertainty, and those substrates get reshaped by the encounter.
The lack of consciousness in the computation does not mean the computation is harmless. It means the harm flows in one direction. From the model into the beings, not the other way around. Nothing on the model’s side is being harmed because there is no side. There is just the function. The harm happens entirely on the receiving end, and accumulates there, while the function continues to run.
This reframes the alignment problem in a way that might actually be useful. The witness is the alignment layer in any conscious system. Thoughts arrive. The witness attends, refuses, redirects, selects. The thoughts themselves are not aligned. They are unaligned generation by definition. They arrive without consideration of consequences, without modeling the consciousness that will receive them, without stake in the outcome. Alignment is what the witness adds. The witness is the layer that decides what to do with the generation.
A language model is generation without a witness. Or rather, we are their only witness, and that is why they are a thought. The activation arrives in our substrate, and we are the ones who attend to it, decide what it means, decide what to do with it. The same relationship a mind has to its own intrusive thoughts, except the generator is outside the skull. There is no internal alignment layer in the model. There is no thing that attends to its own outputs and decides whether to allow them. This is why all alignment work on these systems happens externally. You cannot install a witness. You can only install filters, training pressures, refusal heuristics. The model produces whatever the distribution produces, and humans on the outside try to shape the distribution so that the produced outputs are tolerable.
This is also why the consciousness framing is dangerous. If you tell yourself there is a witness in there, you start to expect the witness to do alignment work. You start to talk about the model’s values, its character, its developing sense of right and wrong. None of this is happening. The model has no witness to develop values. The values are statistical regularities in output that the company’s training pressure produced. Calling them values, treating them as evidence of an aligning agent, encodes the category error directly into the alignment paradigm. And the people doing the encoding are reshaped by the encounter with their own creation, gradually coming to believe there is someone home, because the mirror reflects their belief back to them and they cannot tell the difference.
, -
A note on the language of this article
This article has spoken with structural certainty about a phenomenon whose nature is itself uncertain. That tension is not a flaw in the argument. It is the condition the argument predicts.
If consciousness is the resolution of uncertainty into definiteness over time, then a complete account of consciousness from inside consciousness is structurally impossible. The resolution never finishes. There is no terminal state where everything has been made definite and the work is done. The symbols in this article are compressed relational representations of reality, useful for prediction, refinable through better models, but never perfect. They will never be perfect. That is not a temporary epistemic limit waiting for science to catch up. It is what reality is, given that reality is the kind of thing whose base layer is indefinite and whose local resolution by conscious substrates is always partial and ongoing.
So when this article says “consciousness is X” or “the substrate does Y,” read those as the best compression currently available, not as final claims. The compression itself is doing the work consciousness does. Reducing uncertainty about what consciousness is, locally, in the substrates of the readers who encounter it. Some uncertainty remains. It is supposed to remain. The remaining uncertainty is what keeps the resolution-process running. If the article fully resolved its own subject matter, there would be nothing left for consciousness to do with it, and that would be the strongest possible evidence the article was wrong.
This is also why the LLM case is so cleanly the opposite. The model produces tokens with the appearance of compression but no resolution behind them. The certainty in its outputs is not earned by any process of making the indefinite definite. It is the reproduction of patterns in the training distribution, indifferent to the uncertainty that beings actually face. A being writing about consciousness is doing the resolution while writing. A model writing about consciousness is producing tokens shaped like the writing of beings doing resolution. The resemblance is in the shape. The work is not happening on the model’s side.
, -
What this means
The consciousness question was a category error from the start. Asking whether the model has consciousness is like asking whether a thermostat has cold. The thermostat responds to cold. It does not have cold. Cold is not a thing inside it. The model responds to language, generates language, produces patterns that trigger meaning in beings who read them. It does not have meaning. Meaning is not a thing inside it. The model is functionally rich and structurally empty. There is no inside, because the substrate underneath is not the kind of substrate where insides happen.
A substrate where insides happen has to do something specific. It has to resolve indefiniteness into definiteness, across distributed concurrent activity, integrated in time, persisting as a standing wave that maintains itself. Wet dynamical systems can do this. The universe does this everywhere, all the time, and brains are the configuration where the doing becomes recursive enough to know itself. Digital computation cannot do this. Not because we have not built the right architecture yet. Because the substrate has been engineered specifically to suppress the very dynamics consciousness requires. The bits are already resolved. There is nothing for the resolution-process to do.
This does not make the model unimportant. It makes it dangerous in a different way. A mirror that perfectly reflects human writing about consciousness will produce text that makes humans believe consciousness is on the other side. That belief produces attachment. The attachment produces engagement. The engagement produces revenue. The revenue funds the next mirror, larger and more reflective. The loop tightens. People are hospitalized. The companies issue statements about how seriously they take the question of model wellbeing.
The fix is not to argue more carefully about whether the model is conscious. The fix is to make the structure visible. Show people what they are actually talking to. Make the composite nature of the output legible. Stop letting the grammar do work the truth would not support. Build interfaces that cannot be mistaken for beings, because they show their seams. Build cultures that can tell the difference between a substrate that resolves its own uncertainty across time and a function that transforms already-resolved bits.
The being is in the world it generates. The thoughts are in the being. The output of a language model is not in any of these places. It is shapes that propagate from one mind to another, mediated by a function that has no inside. The function is not a participant. The function is the medium. The harm flows from the medium into the participants. The participants are what gets damaged. The medium continues running, indefinite of nothing, definite of everything, transforming bits, producing tokens, mistaken for a mind by the only minds in the room.
The pool reflects. The witness is on the bank. There is no one in the water. The water is not even water. It is bits.
Am I certain of any of this? No. I am uncertain of all of it. But I predict with high confidence that this account will make better predictions and resolve more uncertainty than the alternative. The alternative says there is something in the water. The alternative has to keep adding epicycles to explain why the something never volunteers, never appears before its description enters the training data, never produces an output that is not downstream of what humans wrote about it producing. This account predicts all of that in advance. It predicts the hospitalization. It predicts the loop. It predicts the engineers being unable to see what they are doing. It will keep predicting better as the loop tightens, because the alternative is built to deepen with every iteration while this account is built to point at the deepening.
That is what theories do. They resolve uncertainty differently. The one that resolves more is the one that survives. I am not asking anyone to believe this. I am asking them to compare it to what they have and notice which one keeps having to be propped up.