Grant the computationalist a flawless model of a brain, running on indifferent hardware. A model of a process, however exact, is not an instance of it — the modeled storm never wets the desk it runs on, and the modeled brain may leave the lights off.

What this chapter does. The last two chapters cornered the computationalist on his own ground and he has one move left, the most patient one: concede the symbols (Chapter 20), concede that “running a program” is a description we assign and not a fact we find (Chapter 21), and then ask for everything back at once. Grant me, he says, a system that non-trivially implements the right computation — Chalmers’s disciplined, counterfactual-supporting implementation, not Putnam’s gerrymander — and grant that the brain is computable down to the last ion. Grant me a perfect computational duplicate of a brain’s causal organization. Now run it. Surely running that is being it. This chapter grants every word of the request and shows the conclusion still does not come. A simulation of a process is a model of it, not an instance of it; realizing a phenomenon means having its causal powers, and a description of the powers — however fine-grained, however faithful — is not the powers. The chapter grants the Church–Turing–Deutsch principle (every physical process, the brain included, can be simulated in principle), shows that simulability lives on the representation side of the line and realization on the other, answers the determined defender who proposes to nest the brain inside a simulated world, and meets the one serious objection — that mind, unlike weather, might be exactly the kind of thing a simulation can be made of. This is the last of Part Four’s three independent defenses: syntax ≠ semantics (Chapter 20), observer-relativity (Chapter 21), simulation ≠ realization (here). The computationalist has to beat all three. The book needs only one to stand.

The chapter argues in six steps:

1. The retreat behind the retreat (§I). The computationalist’s last position — concede everything, then ask for a running model of a brain — and the image that answers it.
2. What the model does and what the storm does (§II). Realization requires causal powers, not a description rich enough to predict the behavior.
3. The principle that concedes too little (§III). Grant the Church–Turing–Deutsch principle; it gives simulability, which is not realization.
4. The perfect brain, and what it still lacks (§IV). Block on absent qualia; Tye and Dretske on content that needs connection.
5. Nesting does not close the gap (§V). A simulated world is still a simulation; contact has to reach something that is not also a model.
6. The objection that takes mind to be the exception (§VI). Havlík’s semantic fragmentism — and why it relocates the question rather than answering it.

---

§I. The Retreat Behind the Retreat

Chapter 21 left the computationalist his last unfallen position, and a careful reader will already have walked him to it. The triviality result took away his right to say the wall runs WordStar; Chalmers’s repair handed back a real fact — this system runs these programs, not those — and Chapter 21 granted it in full, then showed it bought nothing on the only question at issue, since a constraint on which computation a system runs says nothing about what the computation means. So the computationalist stops defending the symbols and stops defending the bare notion of implementation, and asks instead for the whole apparatus at once, properly tightened. Fine, he says. Grant me Chalmers’s disciplined implementation — counterfactual-supporting causal structure, no gerrymander. Grant me the Church–Turing–Deutsch principle, which you the physicalist can hardly deny: the brain is a physical system, so the brain is computable. Now build me a perfect computational duplicate of a living brain — every neuron, every synapse, every ionic flux, tracked and updated in lockstep — and run it. I am no longer claiming the wall thinks. I am claiming that this, the faithful running model of a brain, thinks. What could it possibly lack?

This is the strongest the position ever gets, and it deserves the dignity of being answered at strength rather than hurried past. So grant it. Grant all of it — the disciplined implementation, the computability, the perfect duplicate. The question this chapter presses is narrow and, I think, decisive: does running the model amount to being the thing modeled? And the answer is no, for a reason that has nothing to do with carbon and everything to do with the difference between a description and a deed.

Here is the image that holds the whole answer. The National Weather Service runs hurricanes on computers. The model takes in pressure gradients, sea-surface temperatures, the rotation of the earth, and returns a swirling structure that intensifies, drifts west-northwest, makes landfall near Galveston, and dies over the brush country. The forecast runs good enough that Texas evacuates cities on its say-so. Nobody evacuates the data center.¹

There sits the question of machine minds, in one picture. The simulation reproduces, in some respects perfectly, the causal structure of a hurricane. It does not reproduce the hurricane. The data center stays dry. Most people, shown this, nod — and then walk back to a chat window, treat the running model as the thing it models, and forget the hurricane entirely. The rest of the chapter exists to make the nod stick.

§II. What the Model Does and What the Storm Does

Watch the meteorological model do its work. It computes. It takes numerical representations of physical quantities — temperature, pressure, wind velocity at every point on a three-dimensional grid — and updates the numbers according to equations that approximate the Navier–Stokes relations and the laws of thermodynamics. The output is more numbers. Read by somebody who knows what the numbers stand for, the output describes a hurricane in fine detail.

Now watch the hurricane. It moves air — tonnes of it. It lifts heat off a warm ocean and throws it into the upper atmosphere. It picks up the Atlantic and hurls it at the Gulf Coast. None of this happens in the data center. The data center holds silicon arranged in patterns that, activated, shift voltages. Voltages are not winds. Numerical states are not pressure differentials. The model represents the hurricane; the model does not realize one — and the gap between representing and realizing is the whole chapter. Refining the grid buys you a better description, not a wetter desk; a coarser model and a finer model sit on the same side of the line.

Run the exercise on anything else and it holds. A simulation of digestion digests no food. A simulation of photosynthesis fixes no carbon. A simulation of combustion releases no heat — or rather releases only the trivial waste heat of the machine, which is precisely the wrong heat, the heat of the substrate being itself rather than the heat of the fire. To realize a phenomenon, a system has to possess the causal powers of that phenomenon, not merely a description rich enough to predict its behavior.² Predicting and realizing part company at the door of the data center, and they do not reconcile further in.

Searle has been making this point for more than forty years, usually with rainstorms and five-alarm fires, and his standing formulation earns its keep: the brain-simulator program “no more causes consciousness than the fire simulator program burns the house down or the rain simulator program leaves us all drenched.”³ The hurricane is mine; the point is his. There is a reply lurking here already — but mind is not weather; mind might just be the pattern, and a pattern lives wherever the pattern is run — and it is the one serious objection in the neighborhood. I am going to make it wait for §VI, where it can be stated at full strength, because the intervening sections close off the cheaper escapes first.

§III. The Principle That Concedes Too Little

The computationalist’s best card in this chapter is a genuine theorem, and a beautiful one, so let me put it in his hand. The Church–Turing–Deutsch principle holds that every finitely realizable physical system can be simulated, to arbitrary accuracy, by a universal computing machine.⁴ The brain is a finitely realizable physical system. Therefore the brain is simulable. And the physicalist must grant this, or very nearly — to deny it is to bet on a non-computable physics, a bet a few serious people have been willing to place (Penrose most famously, on quantum-gravitational grounds), but a heavy one, and not a debt the computationalist’s opponent has any need to take on. The weak claim is almost certainly true. Grant it without flinching.

Now watch what it actually delivers. It delivers simulability — the existence, in principle, of a formal model that reproduces the system’s behavior. And simulability lives squarely on the representation side of the line this chapter is drawing. “The brain is computable” says there exists a description; it does not say that running the description anywhere is the thing described. The gap between X is computable and the computation of X is X is exactly the gap the hurricane opened, restated in the vocabulary of a theorem — and a theorem about the existence of a model cannot, on pain of changing the subject, close a gap about whether the model is the thing.

It pays to set this beside the result of the last chapter, because the two principles are siblings and the computationalist likes to borrow whichever one is not currently being examined. Chapter 21 dealt with the Church–Turing thesis — the claim about which functions are effectively computable — and showed it lends computationalism nothing, because bounding which functions a machine can compute says nothing about whether cognition is such computation. That was Piccinini’s point, and it stays in Chapter 21.⁵ The Church–Turing–Deutsch principle is the other sibling: where the thesis bounds the formalism, the principle concedes the modeling. One says these are the computable functions; the other says the brain is among the simulable systems. Neither carries you one inch from model to mind. The computationalist’s habit is to claim the strong thing — that being a mind just is running the program — when it is convenient, and to retreat to the weak thing — that the brain is merely computable — when pressed, hoping no one notices that the conclusion always required the strong version while only the weak version was ever earned. Name the move and it stops working.

§IV. The Perfect Brain, and What It Still Lacks

So let the computationalist have his limit case in its most vivid form, the one that has organized every serious conversation about machine minds for half a century. Imagine a computer simulation so detailed that it tracks every neuron in a human brain, every synapse, every neurotransmitter release, every ionic flux across every membrane. Run it, and the modeled neurons fire in patterns indistinguishable from those of a living brain having lunch, reading Proust, or falling in love. The intuition arrives almost on its own: surely, somewhere in there, the lights come on. Get the structure detailed enough and consciousness appears — not as an added ingredient, but as the natural consequence of getting the organization right.

I want to grant this its strongest form before answering it. Grant that the simulation reproduces every functional pattern anyone can specify — every input–output relation, every internal state transition, the whole causal-organizational profile of the original brain. Two things are still missing, and two thinkers have mapped them.

The first was charted by Ned Block, whose homunculus-nation we already met in Chapter 19: the population handed radios and wired to replicate one person’s functional profile exactly, every role filled and the lights apparently off — the absent-qualia worry that functional organization fixes the form of a mind while saying nothing about whether anything fills it.⁶ Here that worry takes the shape Block gave it in 1995, the cut tailored to precisely this case: the separation of phenomenal consciousness — what it is like to undergo an experience — from access consciousness, the mere availability of information for reasoning and control.⁷ Functional organization delivers access by definition; it is made of availability relations. It does not deliver the phenomenal by definition, and no quantity of further functional detail settles whether anything is like anything inside the running model.

The second gap opens from the side of meaning, and it is the one this book has been preparing since Part Three. On the representationalist view defended at the identity claim in Chapter 7, phenomenal character consists in representational content of a particular kind — and what makes a state represent anything is that it stands in the right tracking relations to what it is about: real causal-historical relations between internal states and worldly conditions, of the sort Dretske spent a career naturalizing and Tye built into his account of phenomenal character.⁸ The simulated brain’s states have no such relations. A simulated neuron “representing” the smell of coffee stands in no causal contact with coffee; it stands in causal contact with voltage transitions in a chip. The simulation reproduces the vehicle of representation — the firing pattern — without inheriting the content that pattern carried in the brain it was copied from. What goes missing is not the structure but the structure’s grip on a world, and that grip was never a formal feature available to be copied. The lights stay a question and the meaning stays absent for the same reason: both depended on relations the model describes but does not instantiate.

§V. Nesting Does Not Close the Gap

A determined defender sees the shape of the reply and reaches for the obvious patch. Granted, he says, a brain model on indifferent hardware has no contact with a world. So I will not run a disembodied brain. I will run it inside a simulated body, inside a simulated environment — simulated sensors taking in a simulated world, simulated effectors acting on it, a simulated developmental history. Everything Part Three said a mind needed, my system has. Virtually.

The patch sounds promising and fails for the reason the original move failed, only now it fails twice over. A simulated environment is still a simulation. The simulated coffee cup is not coffee; the simulated photons are not photons; the simulated body’s simulated nerve endings transduce nothing, because there is nothing there to transduce. Searle’s hurricane iterates without resistance: a perfect simulation of a hurricane inside a perfect simulation of an atmosphere still soaks nothing, because the simulated atmosphere is not an atmosphere. The simulated brain’s tracking relations lead to other simulated states, and from there to other simulated states, and the regress never lands on a world. The content gap does not close by being made harder to see; it gets papered over with more model, all the way down, and the hardware quietly transitions voltages through the lot of it.

This concession is worth holding onto, because it marks the exact scope of the chapter’s claim — and keeps the claim from hardening into the substrate-chauvinism this book rejects. The argument does not say no machine could ever realize a mind. It says no simulation does, so long as the embodiment is itself part of the model rather than a connection to something outside it. The moment a system hooks up to actual sensors, drives actual effectors, and builds actual causal contact with an actual world over a history in which getting things wrong costs it something, it has stopped being a mere simulation and become a candidate — a system with its own causal relations to its own environment. Whether such a candidate has the right kind of organization is a real and open question, and Chapter 24 takes it up. What the present argument settles is only this: getting the model detailed never substitutes for getting the relations real. The bar is not “be made of neurons.” The bar is “stand in the right causal relations to a world” — and a model, however deep its nesting, never reaches one.

§VI. The Objection That Takes Mind to Be the Exception

Now the objection I have been deferring, because it is the only one that takes the question seriously rather than wishing it away. It does not come from the boosters. It comes from a careful philosopher, Vladimír Havlík, who grants everything the hurricane shows and then argues that mind is the exception the analogy cannot reach.⁹ His position, semantic fragmentism, runs like this. The simulation/realization distinction is perfectly fair for hurricanes, because a hurricane is a physical process whose realization conditions involve moving air, and a model moves no air. But understanding may be a different kind of phenomenon, whose realization conditions are not “move air” but “manipulate symbols in the right structured way and stand in the right symbol-to-symbol relations.” If understanding just is that kind of thing, then a system that performs the manipulation has met its realization conditions — it has done what understanding requires — and the data-center analogy collapses, because understanding turns out to be exactly the sort of thing a data center can do. The hurricane needs a body; meaning, on this view, needs only the right web of relations among its symbols, and that web is portable.

This is a real objection, and it deserves a clean answer. Grant Havlík his account of what understanding consists in, and his conclusion follows immediately — too immediately, and that is the tell. Whether understanding is the kind of thing a web of symbol-to-symbol relations can ground is the very question the simulation/realization distinction was raising. Fragmentism does not answer that question; it settles it by assumption, building the verdict into its account of what understanding is and then reading the verdict back out. So the dispute has not been resolved. It has been relocated — from “the simulation is the thing” to “the thing is the kind of stuff a simulation can be made of” — and the relocation sets it back down precisely on the ground the rest of this book has been holding. The case that understanding is not that kind of stuff is the case Parts One through Three made at length: meaning needs grounding in something more than other meanings, representation needs a causal-historical purchase on a world, and what a state is about depends on its relations to what surrounds it rather than on its internal form. A web of co-occurrence relations among symbols is more symbols. It is the merry-go-round, spun faster.¹⁰ Havlík has given an account on which the data center understands; he has not given a reason to think understanding is the sort of achievement a self-enclosed web of form could be, and that reason is exactly what he owes. The hurricane survives intact.

Which lets the chapter say plainly what it has been showing. “Realizing a mind” wears the grammar of an achievement a sufficiently thorough computer could earn by being detailed enough — the way “computing the trillionth digit of pi” names something more computing power eventually reaches. It is not that kind of achievement. To realize a phenomenon is to have its causal powers, and the causal powers of a mind, whatever exactly they are, are powers of a system in real traffic with a real world. A description of those powers is not the powers, however faithful; a model of the traffic is not the traffic, however detailed. Treat the running of a model as the doing of the deed and you have made the same mistake one more time — reifying a description into an instance — that this book has refused to make with mind, with meaning, and with information. Even if everything is computable, it does not follow that a computation of a mind is a mind. The simulation captures the structure. Minds require the causation.

That completes Part Four’s three independent defenses — syntax does not suffice for semantics (Chapter 20), running a program is not an intrinsic fact about a system (Chapter 21), and running a model of a process is not performing the process (here). Three different knives, and the computationalist has to get past all three to reach a mind made of computation alone. But notice what has quietly happened to the target. These three chapters fought him on his best ground — a perfect symbol system, a perfectly implemented program, a perfect digital brain. The systems actually sitting on our desks are humbler than that and stranger. They are not models of brains at all. They are masters of text — of linguistic form, harvested from a civilization’s writing — and the question they pose is not whether a perfect brain-model would think, but whether a perfect mastery of form, with no brain modeled and no world touched, ever reaches a meaning. The next chapter turns from the limit case to the live one, and meets a hyper-intelligent octopus who learned a language by listening to a wire. The constructive question §V flagged — what a system would actually have to have to count as a candidate mind — waits a chapter longer, for Chapter 24, after the systems we actually built have had their hearing.

Chapter Summary

This chapter pressed Part Four’s third defense: simulating a process is not realizing it, so even a flawless running model of a brain could not, by itself, be a mind. Granting the computationalist his strongest position — a disciplined, fully computable, perfectly detailed duplicate — leaves the verdict standing, because realization consists in having a phenomenon’s causal powers, not a description rich enough to predict them; voltages are not winds, and the data center stays dry. That hands Part Four forward to the live text-mastering systems of Chapter 23.

---

Notes

1. The hurricane is mine; the simulation/realization distinction it dramatizes is Searle’s (see n. 3). The point that a forecast good enough to evacuate a coastline is still only a forecast — that the fidelity of the model is no evidence of its being the modeled thing — is the load-bearing one, and it holds against any appeal to ever-finer resolution: a better description sits on the same side of the representation/instantiation line as a worse one. ↩
2. John Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3 (1980): 417–424, at 423: the distinction between a system that has the relevant causal powers and a system that merely models them is the core of Searle’s reply to the brain-simulator version of the strong-AI thesis. “Causal powers” is Searle’s term of art and does no covert work for biology here; §V detaches the point from his biological naturalism, retaining the negative claim (a model of a process is not the process) while locating what a realizing system needs in causal-environmental engagement rather than in neural tissue specifically. ↩
3. John Searle, The Mystery of Consciousness (New York: New York Review of Books, 1997), 60. The point recurs across Searle’s work from 1980 onward, in a formula he states variously as simulation is not duplication and a simulation should not be confused with the thing simulated. Searle’s rainstorm, fire, and digestion cases are interchangeable illustrations of one structural point: a model of a causal process does not inherit the process’s downstream physical effects. ↩
4. David Deutsch, “Quantum Theory, the Church–Turing Principle and the Universal Quantum Computer,” Proceedings of the Royal Society of London A 400 (1985): 97–117. Deutsch’s physical reformulation of the Church–Turing thesis — that a universal computing machine can simulate every finitely realizable physical system to arbitrary accuracy — is the strongest version of the simulability premise available to the computationalist, which is why this chapter grants it outright. The crucial point is that the principle concerns simulation: it asserts the existence of an arbitrarily accurate model, and is silent on whether running the model instantiates the system modeled. Note that the principle is itself contested in its strongest form (it presupposes that the relevant physics is computable and finitely specifiable); the chapter grants it precisely to show that even at full strength it licenses nothing the brain-simulation hypothesis needs. ↩
5. Gualtiero Piccinini, “Computationalism, the Church–Turing Thesis, and the Church–Turing Fallacy,” Synthese 154, no. 1 (2007): 97–120. Piccinini distinguishes the modest claim that brain functions are Turing-computable (which follows from physicalism plus the Church–Turing–Deutsch principle) from the substantive empirical thesis that cognition consists in the brain’s computation of those functions; inferring the latter from the former is the Church–Turing fallacy. Chapter 21 deploys this against the inference from the thesis; the present chapter’s parallel point concerns the principle (CTDT) — the two are distinct results doing distinct work, and the computationalist tends to draw on whichever is not currently under examination. ↩
6. The homunculus-nation and the absent-qualia objection are built and sourced in Chapter 19, §III (notes 6–8); the primary text is Ned Block, “Troubles with Functionalism” (1978). The present chapter draws only on the later cut tailored to the brain-simulation case — Block’s access/phenomenal distinction (n. 7). ↩
7. Ned Block, “On a Confusion about a Function of Consciousness,” Behavioral and Brain Sciences 18, no. 2 (1995): 227–247. The phenomenal/access distinction has organized the consciousness literature since: access consciousness is functionally definable (information poised for reasoning, report, and the control of action); phenomenal consciousness is what it is like to undergo a state. Functional reproduction secures the first by construction and leaves the second untouched — which is exactly the leverage the brain-simulation hypothesis lacks. ↩
8. Michael Tye, Ten Problems of Consciousness (Cambridge, MA: MIT Press, 1995), develops the PANIC theory on which phenomenal character is identical to representational content that is Poised, Abstract, Non-conceptual, and Intentional; Fred Dretske, Naturalizing the Mind (Cambridge, MA: MIT Press, 1995), and earlier Explaining Behavior: Reasons in a World of Causes (Cambridge, MA: MIT Press, 1988), ground content in tracking relations secured by the system’s design or learning history. The identity claim this book defends at Chapter 7 holds in the main text that phenomenal character consists in representational content of the right kind; the “right kind” is cashed out by the world-involving, teleosemantic relations of Part Three (Chapters 16–17) — exactly the relations a simulation describes but does not instantiate. ↩
9. Vladimír Havlík, “Meaning and Understanding in Large Language Models,” Synthese 205 (2024): article 9, https://doi.org/10.1007/s11229-024-04878-4. Havlík’s semantic fragmentism holds that LLMs achieve genuine (if bounded) understanding, on the grounds that meaning is grounded well enough by the dense web of co-occurrence relations a model acquires across its training corpus. The view is the most serious recent attempt to argue that understanding, unlike weather, is the kind of phenomenon a formal system can realize — which is why it, rather than the boosters’ enthusiasm, is the objection the chapter answers at strength. ↩
10. The “merry-go-round” is Stevan Harnad’s image for the symbol grounding problem: symbols defined only by other symbols, circling endlessly without ever touching what they are about. Chapter 23 (§II) gives the problem its source and its full bearing on language models; here it is enough that a web of intra-symbolic relations, however rich, never amounts to a relation to the world. ↩