22.1 The Move the Last Chapter Left Open

Chapter 21 closed on a diagnosis: neither operator-level rule-following nor fluent output establishes understanding by the implemented whole. A computationalist can grant all of that and still think the argument has missed him, because (he will say) the mind was never supposed to be the symbols. The symbols are just the visible spelling. The mind is the realized computation: recurrence, memory, integration, planning, and control, implemented as one causal organization whether built from neurons or silicon. Attack the operator all you like; the candidate system remains untouched.

That reply forces a distinction the slogan the program usually blurs. An abstract program is a formal type, available for indefinitely many implementations and carrying no token history merely by existing. A running implementation is a physical process with determinate causal powers, a provenance, and whatever relations it bears to an environment. This chapter denies mentality to neither by definition. It argues that the abstract type leaves the relevant worldly and system-level facts open, while the running token must be inspected to learn which of them it realizes. A wide computationalist who includes the complete content-bearing and integrative organization can therefore agree with the positive account and disagree only about its name.

Take the question in its rudest form, the one philosophers play as a parlor game. There is always someone (there is always someone) who announces, with the calm of a person about to enjoy themselves, that the radiator against the far wall is, at this very moment, running Microsoft Word. Not metaphorically. Not “could be wired to.” Running it, now. The radiator’s molecules occupy a physical state too complicated ever to write down, and that state grinds, moment to moment, into other states. Write down the sequence of computational states Word passes through as it runs, then pair each one, in order, with one of the radiator’s physical states. Given enough physical states to draw on, and enough freedom in how you pair them, the pairing exists. So the radiator runs Word. It also, by the identical trick, runs a chess engine, and the flight software of a 747, and the chess engine’s exact opposite, and it dreams of Provence on alternating Tuesdays.

The pairing works on these permissive terms. The question is whether pairing states in a sequence amounts to implementing a program. A proper account must distinguish the radiator from the computer without assuming that only one can think. Causal structure supplies that distinction.

22.2 The Wall That Allegedly Runs Any Program

Functionalism defines a mental state by its causal role and permits the same mind to run on different materials. The triviality argument takes that promise and runs it off a cliff. If a mind is the right functional organization, then having the mind is instantiating the organization, and organization comes cheap. To say a physical system “implements” a finite-state automaton is only to say that some mapping from physical states to automaton states makes the physical transitions line up with the transition table. Take any system with enough internal states changing fast enough (a wall warming unevenly, a bucket of water, the radiator) and, with no rule against gerrymandering, such a mapping always turns up. Putnam proved a sharp form of it: every ordinary open physical system implements every inputless finite-state automaton.¹

And Putnam inflicted the wound himself: he wanted minds to be organization precisely so they could be multiply realized. The triviality result says the realization relation is so unconstrained that implementing-this-automaton achieves nothing at all — a permission slip you write yourself. So the radiator already instantiates the right automaton, and already has my mind, and yours, and the mind of a reader this argument persuades, and the mind of one it leaves cold, all at once, settled only by which mapping a bored interpreter lays over it. A theory on which the radiator already understands everything has not explained understanding. It has mislaid the word.

The point arrives from the other direction too. Searle reached it through physics, not automata. Whether a lump of matter counts as a “symbol,” a “computational state,” a case of “running a program” is not, he argued, an intrinsic feature of it the way mass and charge are intrinsic. An observer assigns that feature to it by reading the computational description on. Syntax, in his flat phrase, is not intrinsic to physics.² You can describe the same wall as implementing any program, or its negation, or nothing, depending entirely on the interpretation projected onto its molecular states — and a physical system has the causal properties it has, full stop, whether or not anybody reads them as a computation.

Putnam and Searle, who agreed about almost nothing in the philosophy of mind, reached the same wall from opposite sides — Putnam from inside functionalism, Searle as its standing critic.³ The natural reply is that the mappings are gerrymandered: genuine implementation must respect causal structure and counterfactuals, not merely trace an arbitrary sequence through matter.⁴ That is the strongest repair, and it deserves a clean hearing.

The defect is real and fixable. A repair that respects counterfactuals and the organization of components defeats the wall; a different question survives: whether intrinsic computation fixes meaning.

22.3 The Best Repair, Granted and Set Aside

A stronger objection says a careful account of implementation defeats the triviality argument, so the wall does not run every program after all.

Chalmers supplied functionalism’s strongest defense two chapters ago; he supplies the implementation repair here too. He grants that unconstrained mappings trivialize, then argues that genuine implementation was never unconstrained. For a physical system to implement a computation, its causal structure must mirror the computation’s formal structure: its states must divide into components that transition under counterfactual-supporting causal laws in an isomorphic pattern. He formalizes this with what he calls combinatorial-state automata. Ma and Kanai press the repair further: the organization must be specifiable without an observer’s labels and displayed in the system’s intervention profile. An ordinary rock does not thereby implement an arbitrary complex computation: its components must support the specified causal transitions, including those not exercised on a given run.⁵ Putnam’s mapping laid a story over the system instead of finding one in it. A single system may implement many computations, but choosing a mapping cannot make it implement any computation we please. The wall is innocent.

The repair wins. Intrinsic, counterfactual-supporting causal structure fences the gerrymander out; the radiator, disciplined, does not run Word. Grant all of it.

Chalmers has shown that a system’s causal structure constrains which computations it implements — a real fact, not a free interpretive gift. The constraint has not shown that fixing the computation fixes a meaning. It delivers a pattern in the shape of causal flow. The same abstract transition structure can serve chess, a tax calculation, or nothing anyone cares about. Causal organization fixes the implemented pattern; history and world-relations fix what, if anything, the pattern is about.

Re-instantiation repeats a pattern without deciding whether the later token inherits the first token’s content, continues the same subject, understands, or feels. Provenance and world-involving history bear on content; the realized cognitive organization bears on attribution.

22.4 What the Wall Lacks

With implementation established, a different question begins: what makes a computational state mean something? A system can really compute while its abstract transition pattern leaves the content open.

Not the formal structure alone. The structure has to come tethered: selection, learning, and causal traffic must constrain which worldly conditions a state answers to.⁶ A trained producer-consumer mechanism can thereby acquire content through that history. A nervous system may earn predator rather than sandwich crumbs while leaving finer distinctions unsettled; no interpreter assigned the pairing, but neither did selection carry a taxonomist’s pencil. The radiator dreaming of Provence on Tuesdays cannot dream of Provence because it has neither the relevant history nor an integrated cognitive organization in which such content could figure. A brain has both. Content depends on mechanisms developing and operating in worlds, not on formal organization alone.

The brain’s chemistry supplies determinate causal joints, but causal specificity alone is not content; a radiator has determinate chemistry too. Formal computation groups systems by transition structure and omits the selection, learning, and world history that do the tethering.

An artificial system might develop the relevant history too. The implementation repair gives us a genuine computation to investigate; it does not bar its states from acquiring content.

22.5 What the Model Does and What the Storm Does

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 Deutsch’s physical principle: the brain is a finitely realizable physical system, so a universal model computing machine can perfectly simulate it. Give me more than a machine with matching inputs and outputs: build 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?

Grant disciplined implementation, physical simulability, and the perfect duplicate. Which features does the running system merely describe, and which does it actually realize? We have to inspect the deed.

Imagine a hurricane forecast. The model takes in pressure gradients, sea-surface temperatures, and the rotation of the earth. Its projected storm intensifies, drifts west-northwest, and makes landfall near Galveston. Suppose the forecast proves good enough that Texas evacuates cities on its say-so. Nobody evacuates the data center.⁷

The computationalist can grant all of that: a simulated calculator really adds, and a simulated chess engine really plays chess. Perhaps mentality, too, consists in organization that the running computer actually realizes. If so, the dry data center tells us nothing against its thinking. I agree that the hurricane cannot decide which sort of capacity mind involves.

Watch what the weather model lacks. It updates numerical values for temperature, pressure, and wind velocity through atmospheric equations. The hurricane moves air, lifts heat off a warm ocean, and drives water ashore. The computer’s voltages do not produce those atmospheric powers. Refining the grid improves the forecast without turning the calculation into a storm. This comparison blocks an automatic transfer of powers from model to modeled thing; it supplies no general barrier to artificial realization.

A simulation of metabolism burns no calories — or rather burns only the watts the hardware draws, which is precisely the wrong expenditure, the cost of the substrate being itself rather than the cost of a body doing its living. 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.⁸

Coupling can change the answer. A simulated controller wired through transducers to a boiler really controls the boiler. Its variables now participate in the activity previously only modeled. Calculation, control, digestion, and weather have different realization conditions; the word simulation tells us none of them.

That shifts the burden to the theory of mind. If consciousness depends only on intrinsic causal-computational organization, a system realizing it is no counterfeit merely because we call it a simulation. If consciousness requires world-involving content, those powers must be realized too. Wire a model into the right traffic and it may acquire a power it previously depicted. The noun settles nothing; what the running system does settles the case.

For more than forty years Searle has pressed the comparison, usually with rainstorms and five-alarm fires: “The brain simulator program by itself no more causes consciousness than the fire simulator program burns the house down or the rain simulator program leaves us all drenched.”⁹ The hurricane is mine; the analogy is his. Its application to consciousness requires the further argument about what a mind realizes.

22.6 Simulability Is Not Realization

The original Church–Turing thesis concerns functions calculable by an effective procedure.¹⁰ Deutsch’s physical principle adds a claim about nature: every finitely realizable physical system can be perfectly simulated by a universal model computing machine operating by finite means.¹¹ His perfect simulation matches the target’s input–output behavior under specified labels; his proposed universal machine uses quantum computation. Neither a classical digital machine nor neuron-by-neuron duplication follows just from that principle. The brain duplicate in Section 22.5 therefore adds a stronger stipulation. Physical simulability gives that proposal something it needs, but it supplies no inference from a model’s fidelity to mentality. A brain may be perfectly simulable without every power of the brain passing to its model.

22.7 The Perfect Brain, and What It Still Lacks

Give the computationalist the limit case: a simulation tracking every neuron, synapse, neurotransmitter release, and ionic flux in a human brain. Its modeled neurons reproduce the patterns of a living brain having lunch, reading Proust, or falling in love. Hold fixed the fine-grained causal organization inside the brain, including its responses to alternative inputs at the sensory boundary. Leave the new implementation’s provenance and environmental relations unspecified. Surely, the intuition says, somewhere in there the lights come on.

That stipulation fixes internal organization, not the original subject’s entire world-involving history. Two questions remain.

The first gap opens from the side of meaning, the gap Part Three prepared: tracking, selection, and learning relations constrain what a state is about—the causal-historical relations Dretske naturalized and Tye used to secure wide content, which Chapter 6 then identifies with phenomenal character in the conscious case.¹² A match in internal organization does not by itself settle whether a new token inherits those relations. An accidental duplicate has no claim on the donor’s coffee-history; a simulation produced by causally copying a particular brain may stand at the end of that history, just as a speaker can inherit a name’s reference through a chain she did not begin. Any inheritance would come from the copy’s provenance, not from the firing pattern alone. A separately trained copy may instead earn content through its own history. Internal duplication settles neither route. Losing reference to that particular cup would not, however, show that the copy lost the look or taste of coffee. Chapter 6 distinguishes the object referred to from the nonconceptual properties presented in experience; the latter fix phenomenal character. Changed object-reference need not change those properties. A different presented content, no such content, a different experience, and no experience require different arguments.

Ned Block reached the second opening from the functionalist’s own side, by way of the homunculus-nation the reader already met in Chapter 20.¹³ Here it takes the shape Block gave it in 1995, tailored to precisely this case: the separation of phenomenal consciousness (what it is like to undergo an experience) from access consciousness, the availability of information for reasoning and control.¹⁴ An exact reproduction of the relevant access functions delivers access consciousness by the functional characterization. Whether it delivers the phenomenal remains the point under dispute, not a conclusion contained in the vocabulary. The perfect duplicate therefore gives the functionalist her strongest case. It does not get to borrow the donor brain’s history for free, and the word simulation cannot deny it consciousness for free either.

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 cannot be dismissed by repeating the word simulated. A virtual environment contains real digital objects and real causal relations among the physical states that implement them. A collision can genuinely redirect a process; a virtual predator can genuinely alter what its agent does next. Chalmers is right to insist on that much. The harder question is whether the agent’s relations to that environment fix content for the agent, or whether the whole ecology remains a drama whose semantic roles come from its makers. More model is not automatically more grounding. Neither is it automatically none.

The argument does not say no machine or virtual agent could realize a mind. It says that reproducing internal organization does not by itself reproduce the history content may require. A system becomes a candidate understander when content-bearing states participate in a unified, persistent organization of perception, memory, inference, correction, planning, and action. Its environment may contain trees or digital objects; actual means causally actual for the system, not pleasingly non-virtual to us.

The concession’s best advocate is Chalmers, who has argued at book length that virtual worlds deserve better than their scare quotes: virtual objects are genuine digital objects, virtual events genuinely occur, and a life lived in a well-built virtual world would be a real life in a real environment. He has aimed the point at the present juncture. In his discussion of virtual embodiment, he argues that virtual environments can supply genuine grounding rather than merely stand in for physical ones.¹⁵ He is right. The relations inside a simulated ecology can be causally real, and no general principle makes their identity forever borrowed from a designer. Designers also built the laboratory, named the stimulus, and chose the frog; none of that prevents the frog from earning content in the resulting traffic. The semantic question turns instead on whether selection, learning, and causal history distinguish among equally structure-preserving readings. A bare web cannot (Section 22.8); a mechanism with the right consequential history may. Cups and predators can become genuine contents because the history fixes what the relevant states are for, not because the maker labels the sprites.

Dangerous-for-it marks a separate achievement. A pattern that threatens an organization the system must maintain can ground a practical interest, but it does not tell us what any state is about. Chapter 24 takes up that question separately from meaning and consciousness.

The stronger patch makes the system model its own continued existence and act to preserve it — a self-evidencing agent holding its states within viable bounds.¹⁶ Modeling viability is not enough. The regulation has to maintain the organization whose loss would end it; otherwise the self-concern is depicted along with everything else.

22.8 The Objection That Takes Mind to Be the Exception

The organizational reply beside the hurricane has a specific semantic version. Vladimír Havlík, writing in Synthese in 2025, argues that linguistic understanding has realization conditions a language model can meet.¹⁷ His position, semantic fragmentism, grounds an expression’s meaning in its relations within a sufficiently rich fragment of language. A corpus reaches what he calls semantic saturation when adding relevant expressions no longer changes those meanings. On his account, acquiring and using that web realizes bounded linguistic understanding. Reference to things outside language need not enter the account at all. A data center need not move air to achieve that.

Havlík and I need not disagree about every bounded achievement. Training can give mechanisms genuine content about linguistic contexts. When several content-guided capacities correct and constrain one another, the book also grants a basis for partial understanding. Neither adult human breadth nor conscious awareness sets the minimum. But success on a language task, content about its patterns, and understanding through the coordinated use of that content remain distinct claims. A system that uses a correction to revise an inference, a paraphrase, and its later answers gives us more evidence of bounded understanding than one whose successes remain isolated.

A structural limit bears on reference, though it does not refute Havlík’s narrower claim. Draw three nodes with an arrow from the first to the second and another from the second to the third, and no other arrows. Read them as three towns joined by direct one-way roads. Now read them as three tanks joined by direct pipes. Stipulate that each domain has exactly those connections. Both readings satisfy the same abstract pattern, although roads and pipes differ. Each node has a distinct position; no symmetry has to swap the nodes within the original graph. What varies here includes both the things assigned to the nodes and the concrete relation assigned to the arrows. Nothing in the bare pattern says road rather than pipe.¹⁸

That example deliberately leaves out names, causal history, and evidence that the represented connection really involves roads. Adding those facts may select the first reading. Merely renaming nodes would establish no competing content for a physically trained network. The modest point is that matching a relational form does not, by itself, select its distal interpretation. Newman and Putnam developed related structural and model-theoretic problems; their arguments require attention to which constraints stay fixed. They supply no theorem that a trained mechanism, with its actual history included, must lack determinate content.

Havlík can reply that changing the distal reading while preserving the relevant linguistic relations leaves meaning unchanged in his sense. That reply reaches the argument. A proof that bare structure leaves distal reference open cannot defeat a theory that does not require distal reference. The present result limits what fragmentism establishes about reference; it leaves its narrower understanding claim to further argument and evidence. Our externalist account gives reasons to ask how words connect to the world, but citing that account cannot make Havlík accept its criterion. Nor can I refute bounded linguistic understanding by asking instead for a conscious subject with a human range of concerns.

Rich internal structure still deserves its credit. Researchers have recovered a working map of a game board from a model trained only to predict moves, and “world model” is not an unfair name for what they found. Training made the geometry a network achievement and the mechanism uses it causally. Its history may also fix content, including original content where an interpreting practice does not exhaust the correctness standard. Neither the structure’s richness nor the fact that a task came from humans settles that last question. We must examine the history and use rather than rerun the relabeling trick. The familiar merry-go-round objection marks the gap between relations among expressions and distal reference; it cannot decide whether those relations support linguistic understanding.¹⁹

David Lewis supplied another way to constrain reference: structure plus eligibility. Some properties carve the world at its joints, so interpretations assigning those properties stand ahead of gerrymandered rivals.²⁰ No literal magnet tugs at a symbol. Eligibility constrains the best overall interpretation and may help fix reference. That concession still leaves us to inspect how the system uses its content. A narrow linguistic achievement may be real without establishing an integrated cognitive life or phenomenal consciousness. The hurricane tells us to investigate the powers actually realized; it cannot deliver the verdict.

Chapter Summary

The claim. A wall does not run every program because an observer can map its states. A strict causal test can show that a system truly computes, but causal form alone fixes neither meaning nor mind. Calling the system a simulation settles nothing else.

Where the argument stands. Putnam exposes how loose mappings trivialize implementation; Chalmers’s repair ties computation to the system’s own causal structure. Church–Turing results concern computability rather than every modeled power, while the hurricane blocks automatic transfer without deciding the mind case. History and use may fix original content where correctness standards are not exhausted by interpretation. A bare relational pattern leaves distal reference open; that limit does not refute Havlík’s narrower claim of linguistic understanding. Content, partial understanding, and consciousness still require separate assessments.

What’s left open. A real system may combine its own computation, world-directed content, and a unified cognitive life; neither a virtual setting nor an engineered origin bars it. Self-maintenance may add autonomy and interests, but it creates neither meaning nor consciousness.

The hand-off. The wall has left the witness stand. Actual systems now require judgment by their histories and organization, not by a loose mapping or the word simulation.

---

Notes

1. Hilary Putnam, Representation and Reality (Cambridge, MA: MIT Press, 1988), ch. 5 and the Appendix, esp. pp. 120–125 (the page range at which Chalmers locates the theorem): every ordinary open physical system realizes every abstract inputless finite-state automaton. The result requires the system be open (interacting with an environment so its states do not cycle) over the relevant interval; the proof constructs the needed state-to-state mapping by exploiting the density of distinct physical states. For automata with specified inputs and outputs, Putnam offers a weaker result: any physical system with the right input–output profile realizes indefinitely many compatible automata, not literally every one. The strength of the inputless claim — every automaton, not merely many — is what makes it a triviality result rather than a mere restatement of multiple realizability. ↩
2. John R. Searle, “Is the Brain a Digital Computer?,” Proceedings and Addresses of the American Philosophical Association 64, no. 3 (1990): 21–37, at 26–28, where Searle argues that “syntax is not intrinsic to physics” — that computational description is observer-relative, no fact intrinsic to a physical system settling whether it is computing at all. Searle’s 1990 paper does work distinct from the Chinese Room of Chapter 21: the Chinese Room argues that syntax does not suffice for semantics; “Is the Brain a Digital Computer?” argues that nothing in a physical system is intrinsically syntactic in the first place. The restatement in less technical register appears in The Mystery of Consciousness (New York: New York Review of Books, 1997), ch. 1. The observer-relativity of syntax is the governing premise; for its extension to semantics see Teodor Negru, “Intentionality and Background: Searle and Dreyfus against Classical AI Theory,” Filosofia Unisinos 14, no. 1 (2013): 18–34. ↩
3. Putnam, Representation and Reality, esp. chs. 5–6. The reversal is genuine and unhedged: the multiple realizability that recommended functionalism in the 1960s is, by Putnam’s 1988 lights, exactly what dissolves it once the realization relation is seen to be unconstrained. On the philosophical and rhetorical weight of functionalism’s founder arriving at this verdict, see Oron Shagrir, “The Rise and Fall of Computational Functionalism,” in Hilary Putnam, ed. Yemima Ben-Menahem (Contemporary Philosophy in Focus; Cambridge: Cambridge University Press, 2005). This chapter treats the Putnam–Searle convergence as evidential, not merely rhetorical: independent derivations of one conclusion from incompatible starting points. ↩
4. Peter Godfrey-Smith, “Triviality Arguments against Functionalism,” Philosophical Studies 145, no. 2 (2009): 273–295, esp. secs. 4–5. Godfrey-Smith’s inventory separates the versions of the triviality argument that succeed from those that overreach. His response requires realizers to possess the right lower-level causal organization: appropriately similar physical states, independently variable components, and suitably localized role occupants. That retreat from wholly autonomous functional description converges with the position Section 22.4 defends. Godfrey-Smith explicitly brackets teleological versions of functionalism in this paper. ↩
5. David J. Chalmers, “Does a Rock Implement Every Finite-State Automaton?,” Synthese 108, no. 3 (1996): 309–333; the compressed version is in The Conscious Mind: In Search of a Fundamental Theory (New York: Oxford University Press, 1996), ch. 10. Chalmers requires a combinatorial-state automaton structure: internal states factor into components whose transitions are reliable, counterfactual-supporting, and isomorphic to the formal computation’s state-transition structure. Ron Chrisley, “Why Everything Doesn’t Realize Every Computation,” Minds and Machines 4, no. 4 (1994): 403–420, presses a kindred constraint. For the subsequent debate see Matthias Scheutz, “When Physical Systems Realize Functions…,” Minds and Machines 9, no. 2 (1999): 161–196; Mark Sprevak, “Three Challenges to Chalmers on Computational Implementation,” Journal of Cognitive Science 13 (2012): 107–143; and Michael Rescorla, “The Computational Theory of Mind,” Stanford Encyclopedia of Philosophy. Shuqin Ma and Ryota Kanai push the repair further in “Intrinsic Computational Functionalism: From Observer-Relative Maps to Observer-Independent Structures,” arXiv:2606.06424 (2026): their criteria require observer-independent structure grounded in causal-dynamical organization displayed under intervention. The chapter grants disciplined intrinsic realization; it does not claim Chalmers’s particular formal account is exceptionless. Sections 7–8 acknowledge logically possible false implementations and multiple computations within one system. The relevant constraint concerns implementation of a specified complex computation, not a small total number of computations per system. Their companion paper, Ryota Kanai and Shuqin Ma, “Intrinsic Computational Functionalism and Simulated Consciousness,” arXiv:2606.15348 (2026), rightly shifts the burden: a critic must identify a consciousness-relevant intrinsic structure a purported simulation fails to realize. Chapter 24 accepts that burden without treating thermodynamic self-maintenance as a consciousness requirement; it separates world-involving content, whole-system integration, consciousness-relevant poising and identity, and the autonomy and practical interests a realized stake may ground. ↩
6. The teleosemantic grounding gestured at here is developed in Part Three; see Chapter 15 (teleosemantics) and Chapter 16 (semantic externalism), resting on Ruth Garrett Millikan, Language, Thought, and Other Biological Categories (Cambridge, MA: MIT Press, 1984), chs. 1–2 on proper functions, and Fred Dretske, Explaining Behavior: Reasons in a World of Causes (Cambridge, MA: MIT Press, 1988), for the parallel project through learning history. Artificial training can in principle supply producer-consumer functions and task-relative correctness conditions. Jumbly Grindrod, “Large Language Models and Linguistic Intentionality,” Synthese 204 (2024): article 71, https://doi.org/10.1007/s11229-024-04723-8, presses the further possibility that corpus-mediated causal-historical chains can support linguistic reference. The text grants both routes while keeping distal reference and whole-system understanding distinct. ↩
7. The hypothetical hurricane forecast is mine; the simulation/realization distinction it dramatizes is Searle’s (see n. 9). Improving the numerical resolution of a conventional atmospheric model does not give its implementation the atmospheric powers represented. This blocks an inference from predictive fidelity alone to possession of those powers. It does not show that fidelity supplies no evidence about any modeled capacity, or that computation cannot realize mentality. The realization conditions of the particular capacity must be established independently. ↩
8. John Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3 (1980): 417–424, at 420–423: the distinction between a system that has the relevant causal powers and a system that merely models them is central to his reply to the brain-simulator version of strong AI. “Causal powers” is Searle’s term of art. The book retains the demand to realize the powers relevant to the attributed capacity without accepting a biological restriction or inferring that a model cannot also be a realization. Sections 22.5–22.7 assess that further question separately. ↩
9. 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. ↩
10. Gualtiero Piccinini, “Computationalism, the Church–Turing Thesis, and the Church–Turing Fallacy,” Synthese 154, no. 1 (2007): 97–120, esp. 99–101 and 114–15. Piccinini distinguishes computationalism proper — the empirical thesis that cognition consists in the brain’s computation of certain Turing-computable functions — from the far weaker claim that brain functions are Turing-computable. Even the weaker claim requires a physical computability premise; physicalism alone does not supply it. The inference from computability to computationalism is the Church–Turing fallacy, and Piccinini’s paper shows several historically influential versions of it (Newell, Pylyshyn, and others) to be unsound. The coinage is B. Jack Copeland’s — Piccinini writes of “what Jack Copeland has called the Church–Turing fallacy” — see Copeland, “Narrow Versus Wide Mechanism: Including a Re-Examination of Turing’s Views on the Mind-Machine Issue,” Journal of Philosophy 97, no. 1 (2000): 5–32. The thesis and Deutsch’s physical principle are distinct claims doing distinct work — and the computationalist tends to draw on whichever one is not currently under examination. ↩
11. 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, at 99. Deutsch’s physical principle says that “every finitely realizible physical system can be perfectly simulated by a universal model computing machine operating by finite means.” It is an additional thesis about physics, not a consequence of physicalism. The crucial point is that the principle concerns simulation: it is silent on whether running the model instantiates every causal power of the system modeled. Deutsch defines perfect simulation as computational equivalence under specified input/output labels, not fine-grained organizational duplication; the latter is a separate stipulation here. His universal model is quantum, not simply a classical digital computer. The principle remains contested in its strongest form; granting it supplies physical simulability without settling whether a running model realizes mentality. See sec. 1, pp. 3–4 of the author’s later typeset version, https://www.daviddeutsch.org.uk/wp-content/deutsch85.pdf, for the definition and quantum qualification. ↩
12. 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. Tye’s abstract-content requirement excludes particular concrete objects from phenomenal content (with limited exceptions for the subject); see Ten Problems, pp. 137–39. Chapter 6 likewise distinguishes singular reference and accuracy from the nonconceptual property content that fixes character. Its strong representationalist identity concerns content of the appropriate kind under the relevant conscious organization, not every difference in reference. World-involving history bears on which properties a state represents; poising and the book’s further account bear on whether the state is conscious. Failure to inherit reference to the donor’s particular object proves neither different phenomenal content nor absent experience. A merely formal description does not establish these conditions, but a running artificial system may realize them. ↩
13. The homunculus-nation and the absent-qualia objection are built and sourced in Chapter 20, Section 20.4.1 (notes 13–18); 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. 14). ↩
14. Ned Block, “On a Confusion about a Function of Consciousness,” Behavioral and Brain Sciences 18, no. 2 (1995): 227–247, esp. 227–30. 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. Reproduction of the stipulated access functions secures the first under that characterization; it does not, by the distinction alone, settle the second. Block’s taxonomy is not a proof that functional duplicates lack phenomenal consciousness. ↩
15. David J. Chalmers, “Could a Large Language Model Be Conscious?,” Boston Review, August 9, 2023 (from a NeurIPS 2022 address); see the discussion of virtual embodiment. The book-length defense of virtual realism is Reality+: Virtual Worlds and the Problems of Philosophy (New York: W. W. Norton, 2022): virtual objects are genuine digital objects, virtual events genuinely occur, and virtual experience need involve no illusion. The text grants the metaphysics. Designer-assigned labels do not by themselves fix what an agent’s states represent, but selection, learning, and consequential traffic within a virtual ecology can in principle do so. Virtuality neither supplies grounding nor blocks it; self-maintenance bears separately on autonomy and interests. ↩
16. The “self-evidencing” idiom belongs to Jakob Hohwy, “The Self-Evidencing Brain,” Noûs 50, no. 2 (2016): 259–285: a system that embodies a model of itself gathers evidence for its own existence by acting to keep its sensory states within the bounds the model expects. The underlying formal framework is Karl Friston’s free-energy principle — organisms resist disorder by minimizing variational free energy, an information-theoretic bound on the surprisal of their sensory states: Karl Friston, “The Free-Energy Principle: A Unified Brain Theory?,” Nature Reviews Neuroscience 11 (2010): 127–138. Neither author draws the present distinction. The question is not whether self-evidencing can be implemented in software in the abstract, but whether a particular physical implementation’s regulation helps constitute the maintenance of the very organization doing the regulating. Chapter 24 calls that condition thermodynamic self-maintenance and treats it as substrate-general. ↩
17. Vladimír Havlík, “Meaning and Understanding in Large Language Models,” Synthese 205, art. 9 (2025), https://doi.org/10.1007/s11229-024-04878-4. See the discussions of referential grounding, semantic fragmentism, and the grounding of LLM meanings. Havlík’s semantic fragmentism holds that LLMs achieve genuine, bounded linguistic understanding because training embeds their internal states in dense contextual and inferential relations, supports minimal semantic contents, and can ground meanings without requiring conscious awareness or direct sensory contact. Havlík explicitly rejects referential grounding as necessary for linguistic meaning; on his view an expression represents its place in the corpus’s relational complexity. A limit on distal reference therefore does not refute semantic fragmentism’s narrower claim. The chapter preserves that distinction and leaves the sufficiency of his linguistic-understanding criterion open, rather than demanding conscious subjecthood or adult human breadth as an unargued additional condition. ↩
18. For related limits on pure structural description, see M. H. A. Newman, “Mr Russell’s ‘Causal Theory of Perception,’” Mind 37, no. 146 (1928): 137–148, and Hilary Putnam, Reason, Truth and History (Cambridge: Cambridge University Press, 1981), ch. 2 and the Appendix. These are related but distinct arguments, not one theorem that directly applies to every trained network. Putnam’s ch. 2 considers reinterpretations preserving whole-sentence truth values while predicate extensions change. The main-text example illustrates transfer of a relational structure between domains, including transfer of the relation’s interpretation; it does not infer a nontrivial automorphism of a fixed graph from the existence of an isomorphic copy. Actual names, training relations, and causal history are excluded from that toy example, not shown semantically idle. Selection history and corpus-mediated causal chains may fix content or distal reference. Chapter 23, note 8, makes the same limited distinction. For rich, behavior-controlling internal structure recovered by interventionist probing, see Kenneth Li et al., “Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task,” ICLR 2023 (arXiv:2210.13382). Such evidence concerns realized internal organization and its use, not a proof of semantic independence or whole-system understanding. Where an interpreting practice exhausts the correctness standard, content remains derived; inheritance of a human-designed task alone proves no such exhaustion. ↩
19. 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 (Section 23.2) gives the problem its source and its bearing on language models. The point here is limited to distal reference: a web of intra-symbolic relations does not by itself select a worldly interpretation, though training history and corpus-mediated causal chains may add what bare geometry lacks. This does not establish that corpus-relational meaning or bounded linguistic understanding is impossible. ↩
20. The reply is David Lewis’s, “Putnam’s Paradox,” Australasian Journal of Philosophy 62, no. 3 (1984): 221–236, esp. 227–29, which answers the permutation argument by adding a constraint of eligibility—reference biased toward the more natural properties—to the structural ones Putnam allows. J. R. G. Williams reconstructs and presses the eligibility view in “Eligibility and Inscrutability,” Philosophical Review 116, no. 3 (2007): 361–399, esp. 372–80. Eligibility constrains interpretations globally; it need not operate through a causal “pull” on a particular symbol. The text grants that naturalness can break some structural ties and may help fix reference when joined to selection history and causal engagement. That semantic concession leaves whole-system understanding as a separate attribution question. For sustained skepticism about treating reference magnetism as a Lewisian doctrine at all, cf. Wolfgang Schwarz, “Against Magnetism,” Australasian Journal of Philosophy 92, no. 1 (2014): 17–36. ↩