---

Philosophers call the broader theory functionalism, and for several decades it came as close to orthodoxy as philosophy of mind allows.¹ Functionalism identifies mental states by the causal roles they play rather than by their material composition. Computational functionalism, one influential child of the theory, adds that the relevant roles can be specified as computational states in a program.

Start with a program too simple to confuse anyone. Almost every programmer begins with some version of this one, here in C++:²

#include <iostream>

int main() {
    std::cout << "Hello, World!" << std::endl;
    return 0;
}

Type it, compile it, run it, and the words appear on your screen:

“Hello, World!”

The greeting has an author: you, the programmer. Displaying it gives us no reason to posit a new speaker. You might just as well have printed those words on a piece of paper. Nothing in the paper means to say hello to the world. Nobody is home. I offer this bit of common sense as a premise: your reproduction of meaningful words does not by itself establish a new mind.

Now add memory, recursion, and billions or trillions of parameters calculated from machine learning on enough text to fill ten Libraries of Congress.³ Add photos and videos and other media. At some point, according to some computational functionalists, a program something like this becomes a mind.

The claim concerns what learning and organization produce: states that retain information, correct one another, guide inference, and alter conduct in new situations. If the right causal organization constitutes a mind in us, why should its realization in a machine count for less? Get that organization right, the argument says, and the mind comes along.

20.1 What Functionalism Got Right

For much of the twentieth century, philosophy of mind hung between two unhappy options. The behaviorist treated a mental state as a disposition to behave: believing it will rain amounts to being disposed to carry an umbrella. But no behavior follows from a belief by itself. Up goes the umbrella only if you also want to stay dry, notice the clouds, and have not decided the walk is worth a soaking.

The identity theorist identified each mental state with a brain state: pain with the firing of C-fibers, full stop. Tidier, yes, but at an embarrassing price. A being without C-fibers, whether an octopus, a Martian, or anything whose nervous system took another road, could not feel pain. That discovers nothing about octopuses. It reveals a prejudice about plumbing.

Into this impasse stepped Hilary Putnam, whom you met on Twin Earth in Chapter 16, proving that “meanings just ain’t in the head.” The man who would drive meaning outside the skull first helped found a theory that seemed to put the mind entirely inside the wiring.⁴

Putnam’s founding functionalist move defined a mental state by its role, not its material. Functionalism identifies pain with whatever state bodily damage causes, which produces wincing and avoidance and interacts with beliefs and desires in the way pain does, keeping a creature off the sprained ankle and teaching it to mind the stove. Define the state by its place in that causal traffic and you free the mind from any particular substrate. The octopus can feel pain. So could the Martian. So could a machine that realized the relevant role.

From the late 1960s, the Princeton philosopher David Lewis turned the picture into a method. Gather what ordinary psychology says about a mental state, regiment its web of causes and effects, and define the state as whatever occupies that role. Science may then discover different occupants in different beings.⁵ Pain in us may have one physical basis, pain in the octopus another, pain in a silicon creature a third.

This possibility became multiple realizability: one mental kind might occur in different physical systems.⁶ The term joined two claims. Material anti-chauvinism denies any substance a monopoly on mentality; actual recurrence says a mental kind does occur across disparate realizers. Keep the first and treat the second as a question. Different hardware running one program cannot settle it.

Functionalism leaves the door open to artificial minds by refusing to define mentality in terms of neurons or biological tissue.⁷ Nothing that follows closes that door. The argument rejects narrow functionalism, not substrate neutrality. If a system built from silicon, cultured tissue, radio-linked people, empty beer cans connected by string, or Swiss cheese realized the world-involving organization a mind requires, its materials would not disqualify it.

20.2 Why Narrow Functionalism Fails

Material neutrality does not by itself bring the world into the account. A theorist can still draw the entire causal role inside the skin. On this narrow reading of functionalism, a mental state gets its identity from the role it plays inside an individual. Signals at the sensory boundary lead to the state; the state interacts with other inner states; behavior follows. Input and output descriptions remain narrow enough for an internal duplicate to satisfy them elsewhere. Call some inner state a “water detector,” and the narrow role says only which sensory pattern triggers it and which words come out. The role never says whether water, XYZ, a simulation, or an experimenter with a wire produced the pattern.

Computational functionalism takes the further step of specifying those inner roles as states of a program and identifying those states with mental states. An abstract program specifies formal states and transitions among them. Two implementations in the same formal state must therefore share the same mental state, whatever their surroundings or histories. Narrow functionalism makes internal causal role sufficient; its computational child makes formal program state sufficient.

The child inherits the parent’s problem. If internal causal role cannot determine content, formal program state cannot determine it either. A more precise program still leaves the world off the page. The computational version makes the mistake precise enough to test in both directions. Internally identical systems can think about different things because their histories and environments differ. Systems with unlike architectures can think about the same thing because their states reach the same worldly objects by different routes. A taxonomy that joins the first pair and divides the second misidentifies the mental kind. Putnam supplies both demonstrations.

Putnam first made the taxonomy precise: same machine table, same mental type, whatever the hardware. His later work on meaning showed that the taxonomy cuts both ways.

The same narrow state can carry different content. Oscar on Earth and his molecule-for-molecule duplicate on Twin Earth may share every internal state. Yet Oscar’s “water” reaches H₂O, while Twin Oscar’s reaches XYZ. Their organization matches; their content differs. No narrow computational type therefore suffices for belief identity.

The same content can cross different narrow states. Two neighbors may both believe that many cats live on the street. One learned it by watching the strays. The other heard it from a friend and thinks in a language whose words and inferences do not line up with English. Different states, same content.

A computationalist can still hope to zoom out until some shared type appears. Putnam’s real case blocks that hope. Imagine a creature whose brain, or whose machine, updates its beliefs on a different schedule and groups similarities we would never count as the same. It can still be right about the cats. Its table of inner states need not share a single state with ours, or even a space of states that maps onto ours. To collect those unlike organizations as one belief, we already treat them as reaching the same neighborhood cats. That collecting judgment answers to the world and to the practice of interpretation. It is not a computational type we found first and then named.⁸

In Representation and Reality, under the title “Why Functionalism Didn’t Work,” Putnam drew that moral: a brain considered in isolation cannot supply a state common to all and only the thinkers of a given thought. The machine table had drawn its boundary too tightly.

Here comes the part usually lost when Putnam’s recantation gets reduced to a funeral notice. He considered a reply, credited to the philosopher of science Richard Boyd, that widened the machine. Sociofunctionalism or global functionalism would treat “organisms-cum-environments” as the systems whose relations matter. Inner representations would stand in causal relations to objects, practices, and other speakers. Externalism might force this complication, Putnam wrote, but functionalism was “not yet refuted.”⁹

That reply widens functionalism itself; it need not preserve the identification of mental roles with program states. On the wide reading, relations to the world partly constitute the mental role. A state may represent tissue damage when it has been recruited to covary with damage and alter consumer activity in ways that protect damaged tissue. A perceptual state presents a tomato as red partly because its system developed and operated among colored surfaces. A speaker thinks about water partly because her word and concept participate in a causal and social history reaching H₂O and her linguistic community. Change those relations while holding the mechanism fixed and the content may change.¹⁰

Wide roles remain physical and causal. They specify what produces and consumes a state, which conditions its use answers to, how error becomes possible, and how it guides learning and action. History need not return as a present force; it helps constitute what the inner state — the vehicle that carries the content — means. The explanation has widened, not departed from nature.

Wide individuation does not make every external object part of the cognitive system. My belief about New York depends for its content on New York; the city does not thereby become one of my cognitive organs. A state can belong to an agent while its identity depends on relations beyond the agent. Tightly coupled tools or environmental loops may sometimes join the realizing system, but that claim requires its own causal argument.

A second caution: a wide role cannot mean “whatever produces mentality.” Each addition must earn its place. Tracking history and producer-consumer use help fix content; mode, attention, poising, and integration bear on consciousness; flexible coordination bears on understanding. Calling every relation functional does not collapse these questions.

This separates two claims packed into substrate independence.

1. Material neutrality says composition alone does not settle mentality.
2. World independence says an inner profile fixes the mental type regardless of environment and history.

The first opens the door to artificial minds without carrying multiple realization through as a theorem. Putnam’s cross-classifications defeat the second.

A separate computational question then remains. Abstract-program sufficiency says that the right formal transition structure suffices for mentality. Computational realization allows that a concrete, recurrent, historically embedded computational organization may constitute a mind when it realizes the relevant content-bearing, integrative, and phenomenal relations. This book rejects the first and leaves the second open. Computation poses no objection when a system realizes the complete wide organization a mental capacity requires; program identity never establishes that it does. The program alone does not think. A suitably organized computational agent might.

Putnam nevertheless rejected the widened repair. No finite rule, he thought, could capture the open-ended background of beliefs, charitable constraints, community practices, and worldly facts that makes one interpretation better than another. Wide functionalism therefore failed as the reductive identity theory he had wanted.

That failure need not carry naturalism down with it. Putnam asks for more than a procedure that decides hard cases: he asks for a finite definition of intentional relations in physical and computational terms, one that holds across physically possible cases. Rejecting a decision procedure alone would not answer him. The alternative defended here starts with capacity-specific causal and historical explanations. To count as a reduction rather than a redescription, those explanations must specify the relevant relations without presupposing the mental facts they explain. Whether they can cover every possible case remains a further debt, not a victory won by widening the diagram.

Putnam pressed a separate complaint against computationalism. On a permissive account of implementation, he argued, every ordinary open system could be read as running any finite program that receives no outside input. A rock, redescribed cleverly enough, runs the program of your choosing.¹¹ Chapter 22 takes up that argument and sides with the reply that genuine implementation requires observer-independent causal organization. Grant the non-trivial implementation; Putnam’s content point survives. Running a program does not by itself determine what its states concern.

The disagreement with Putnam concerns what his failed universal reduction rules out. I retain the possibility that physically realized organism-world relations constitute content-bearing roles, while accepting that the role framework alone has not explained them. The positive case must come from the accounts of content, understanding, and consciousness, each defended on its own terms.

20.3 Why Computationalist Multiple Realizability Fails

Apply those cross-classifications to the computationalist’s own argument. The case for multiple realizability ran upward from material to role:

1. Mental states should not be identified with one neural or physical state type.
2. Different physical structures can occupy the same relevant causal role.
3. Therefore a mental type may have different physical realizers.

Machine-state functionalism made the second premise look clean by joining two claims: different physical systems can implement the same computational state, and that state constitutes the mental state. Together they yield the chain:

Different hardware → same computational state → same mental state.

Grant the first arrow. What counts as the same computation depends on how we individuate programs and states, a problem Chapter 22 will inherit. For now, suppose two physical systems non-trivially implement the same program. We have established multiple implementation of a computation.

The second arrow breaks. A shared narrow computational type does not suffice for a shared contentful mental type, and a shared contentful type need not correspond to one narrow computational type. The classifications cross-cut. Multiple physical implementations of one program therefore need not realize one mental state. The hardware-software analogy reaches computation; it does not cross the remaining distance.

The best computationalist can grant all of this and refuse defeat:

Earth Oscar and Twin Oscar have different broad contents, but the same narrow machinery explains their matching local transitions and proximal behavior. Psychology needs that reusable causal kind. A complete mental state can therefore consist of a narrow computational type plus an external reference-fixing factor. Multiple realizability was always a claim about the narrow causal type, or about that two-factor package, rather than broad content alone.

The concession should stand.¹² A narrow computational description may identify a vehicle, explain local transitions, predict behavior, and characterize domain-specific competence guided by representational content. Computation can remain part of the mechanism after computationalism fails as a theory of mental kinds.

Calling the narrow vehicle a psychological type does not make it the state by which a subject thinks about water, perceives damage, or understands a claim. If multiple realizability concerns only that vehicle, it has become multiple implementation of a causal component. If the external reference-fixing factor enters the type, the mental kind has gone wide. Calling the package wide computation changes no result: computation remains part of the realization, but sameness of program no longer establishes sameness of mind.

A computational vehicle may run wholly inside the system and cause the next transition. Its content depends on environmental sources, selection or learning history, consumers, and community practices where relevant. One vehicle type may bear different contents in different embeddings; different vehicles may bear the same content when their total systems realize the relevant wide role.

The negative conclusion has an exact scope: multiple implementation of one program does not establish multiple realization of one contentful mental state. A biological system and a silicon system might nevertheless both represent bodily damage. We would need evidence that their mechanisms track damage through the relevant histories, feed consumers organized to protect the body, and guide learning and avoidance. Different materials pose no objection. The shared capacity must be established rather than read off the program.

The boundary of the total base may vary. Sometimes the environment helps individuate a state while the immediate mechanism remains inside the agent. Sometimes a recurrent control loop crosses skin, tool, and environment so densely that the coupled system forms the best candidate realizer. The boundary follows the relations that explain the capacity; wide functionalism cannot settle it by decree.

What remains is material neutrality and an open empirical question. Some world-involving capacities may recur across strikingly different total systems; some finely specified mental kinds may not. Functionalism alone guarantees neither result. “Same program, same mental state” fails. Carbon enjoys no monopoly. Silicon receives no promissory note.

20.4 Three Tests of the Wide View

A critic may now suspect that wide functionalism has won by moving the boundary whenever trouble appears. Block’s homunculus-nation, the relabeling objection, and Chalmers’s replacement argument test whether the widened view says anything sharper than “include whatever matters.”

20.4.1 Block’s Homunculus Nation

Ned Block asks us to build a system whose narrow functional organization duplicates a human brain in pain.¹³ Take a large population, give each citizen a radio and a local machine-table task, and have them collectively implement for one hour the transition profile that a human brain instantiates after a stubbed toe.¹⁴ Block connects the organized population by radio to the sensory and motor neurons of an artificial, brainless body, making the population function as its brain. The body closes the cheap escape that the thought experiment forgot perception and action.

Set it running. The inputs match. The outputs match. Every stipulated inner transition matches. Does the nation hurt? Is there, beyond the billion busy citizens, a further subject who feels the toe?

The intuition that there is not proves hard to shake. Block calls this an absent-qualia case: the functional wiring appears perfect and the lights appear off.¹⁵ Narrow functionalism must bite the bullet. If the specified internal role exhausts mentality, then the nation hurts, and its strangeness shows only our parochial attachment to familiar materials.

Intuition alone cannot settle the case. Exotic composition may mislead us; if the stipulated facts truly sufficed for consciousness, disbelief would not make them stop sufficing. Wide functionalism offers a diagnosis rather than a louder intuition. Block stipulates an hour of narrow transition matching plus present bodily traffic. At a fine enough grain, that can include current producer-consumer transitions, memory dispositions, learning dispositions, and control of the artificial body. The wide view should grant every one of those causal facts. What the stipulation does not settle is which contents those transitions carry for the assembled whole, whether any state of the whole presents that content under a bodily-affective mode, or whether the activity belongs to a persisting subject with the required attention, poising, and integration.¹⁶ Present contact is not by itself a representational past.

An ahistorical simulator may match the transitions without an independent basis for their contents. A copied mechanism inserted into a content-bearing perception-action loop may inherit the roles its histories and consumers already fix; a persisting nation could acquire an equivalent wide organization through its own encounters. The narrow machine table cannot tell us which case we have.

This answer must accept its own consequence. Suppose the nation-cum-body persisted, learned through its own encounters, developed wide states whose consumers answered to damage and color, integrated those states through memory, inference, planning, and action, and realized the perceptual and bodily-affective organization required for consciousness.¹⁷¹⁸ Then the material of citizens and radios would supply no objection. Wide functionalism should count the system as a candidate mind. If it still said no merely because a nation feels like the wrong kind of thing, it would relapse into the chauvinism it rejected in Section 20.1.

Block’s one-hour case exposes how much narrow duplication leaves open. It does not prove that the nation lacks content or consciousness; it shows that the stipulation settles neither.

20.4.2 Has Functionalism Merely Changed Its Name?

The second objection presses harder. Either content-fixing relations count as more functional organization, in which case the theory expands until it becomes trivially true, or they count as something beyond function, in which case functionalism has been abandoned. Which is it?

Take the first horn with conditions. The wide view enlarges functional organization. That is no embarrassment if the added relations receive independent descriptions and explanations. A state tracks damage because of a causal history, not because we first call it a pain. A consumer makes that state answerable to damage because its operations vary with the state in ways selected or learned for dealing with damage. A linguistic state reaches water through demonstrative chains, environmental contact, and social deference. These relations can fail while the internal transition table remains fixed. They predict differences the narrow view cannot see.

The theory would become empty if it defined the relevant role as “the role played by whatever has the mental state.” It does not. It offers defeasible, capacity-specific conditions. Content requires a source of correctness conditions. Whole-agent understanding requires flexible coordination among content-guided capacities. Consciousness requires a state to present its content under the right mode and make it poised and integrated. These claims may prove wrong, but they can be tested and criticized separately. A theory with joints can lose an argument.

Nor does width smuggle in dualism. The wide role consists in causal, historical, social, and organizational relations among physical systems. Calling the state relational does not make it immaterial. A key fits a lock partly because of a relation between their shapes; the fit does not float outside nature. A belief fits a world in a more complicated way, but complication does not change its ontological address.

The residue between wide functionalism and this book’s world-involving representationalism may indeed come down to filing. I am content with that result. The chapter criticizes a substantive thesis, not a word: the claim that internal causal or formal organization suffices. Anyone who rejects that claim, refuses material chauvinism, and explains mentality through the right world-involving organization has crossed the important distance, whatever appears on the folder tab.

20.4.3 Organizational Invariance

The strongest test comes from David Chalmers, a property dualist who nonetheless defends the principle of organizational invariance: systems with the same fine-grained functional organization have qualitatively identical experiences.¹⁹²⁰ His argument asks us to replace a person’s neurons one at a time with functionally identical silicon chips. Each chip performs the signaling job of the cell it displaces; the subject continues speaking and acting as before.

Could experience fade while the subject insists that the tomato remains vividly red and the pain remains sharp? That would make the subject radically and systematically wrong about her present experience. Could consciousness persist through thousands of replacements and vanish with one final chip? The boundary looks arbitrary. Chalmers concludes that experience must ride with the organization throughout. If material replacement leaves mentality intact, Block’s strange materials cannot by themselves turn the lights off.

Wide functionalism should welcome the replacement verdict while rejecting Chalmers’s universal principle as stated. The replacement story preserves far more than an abstract transition table. It preserves the subject’s body, environment, practical engagements, social setting, and causal history. Each chip enters an already functioning, content-bearing economy and inherits the role of the cell it replaces. The process holds the subject’s complete world-presenting organization fixed while changing the inner material.

The case shows that material substitution can leave mentality unchanged when the subject’s wide organization and history remain fixed. It establishes one route across materials, not a general multiple-realizability thesis. Chalmers does not show that world and history can be dropped from the type being preserved.²¹

Separate three cases.

1. A replacement descendant inherits the original subject’s content-fixing history and complete phenomenal organization.
2. A freshly assembled narrow duplicate matches the internal topology, judgments, and utterances but does not receive their broad content merely from that match.
3. A differently built system may independently acquire an equivalent history and wide organization.

Wide functionalism grants the relevant capacity in the first and third cases. The narrow description leaves the second unsettled. During gradual replacement, judgments retain their worldly content because the subject’s history stays fixed; in a fresh duplicate, matching transitions and sounds do not establish the same content or phenomenal mode.

If Block’s system acquired an equivalent history and integration, the difference would narrow; if every wide relation relevant to the attributed capacity matched, the wide functionalist should grant the capacity. The theory cannot use history as a ceremonial password, denying mentality forever to an artifact because it did not grow in the approved way. History matters only where it constitutes the present state’s content and role.

20.5 The Background No Rule Contains

Putnam reaches narrow computationalism from the side of meaning. Hubert Dreyfus reaches it from the side of action.

In the early 1960s, Dreyfus taught philosophy at MIT while its young artificial-intelligence project pursued the prospect that engineers might explain the mind by programming one. His objection concerned something ordinary expertise makes easy to overlook: knowing what matters in a situation.

The wall is relevance. To act intelligently in a real situation, a system must treat some available facts as bearing on what to do and ignore the boundless rest. A program that applies explicit rules to represented facts faces an awkward question: which facts? Most truths about any situation do not matter to the task. Give the program a rule for selecting the relevant ones. That rule helps only if the system can tell when it applies, which seems to require another rule, and the regress begins one floor up.

A cook in an unfamiliar kitchen does not run the regress. The pot, the flame, and the moment to lower the heat show up as relevant against a background of embodied involvement she acquired without writing it into propositions. Her know-how does real work. Trying to spell all of it out creates more facts whose relevance would itself need settling. AI met a close relative of the difficulty as the frame problem. Dreyfus treated it not as one missing rule but as evidence that the rule-and-representation picture had omitted the background against which any rule matters.²²

The criticism lands against classical, narrow computationalism. A finite stock of inner representations and explicit rules does not manufacture relevance by adding another representation and rule. But Dreyfus does not thereby refute a wide, world-involving account. Such an account may treat skilled coping as learned, embodied, holistic, and dynamical. It can explain relevance through a history in which some distinctions became action-guiding within recurrent organism-world traffic. No homunculus consults a relevance list. The organization has been shaped so that some possibilities solicit response and others recede.

Modern learning systems make that reply concrete. Training can give mechanisms causally potent, task-shaped states with correctness conditions; recurrent perception and action can continually revise them. Natural selection and gradient-based learning differ profoundly, but both remind us that a function may depend on history rather than an explicit rule stored inside the present token.²³ The remaining questions concern what those states represent, how flexibly their consumers use them, and whether the mechanisms compose one persisting agent. Learning answers none automatically. It changes the right question from “Where is the relevance rule?” to “What organization did the history produce?”

I therefore take Dreyfus’s wall and leave his blueprint. Dreyfus often drew an anti-representational moral: intelligent coping does not proceed by manipulating inner content. This book keeps representation and asks how world-involving histories give its states correctness conditions.²⁴ A wide naturalistic account can keep both the content and the background by locating them in an embodied, temporally extended organization rather than a snapshot of rules.

Putnam and Dreyfus converge on a narrower verdict than either sometimes drew. Formal transitions alone do not supply meaning or relevance. A running computational mechanism can still participate in historically grounded, content-sensitive loops; several such mechanisms can constrain one another in flexible action. The cook knew the diagram was incomplete. She did not prove that no diagram could ever include the kitchen.

20.6 The Door to the Room

Chapter 21 tests a different question: whether formal processing establishes understanding by the implemented whole. Its thought experiment can block that shortcut without disqualifying a historically embedded system that acquires content and integrates it across a cognitive life.²⁵

Chapter Summary

The claim. Carbon and neurons hold no monopoly on mind. Material make-up alone cannot settle mentality. Yet a narrow functional role cannot fix content, because a causal diagram cut off from the world cannot determine what its states concern.

Where the argument stands. Putnam’s externalism blocks the inference from one narrow program state to one contentful thought. Computational kinds still explain mechanisms; a historically embedded computational system might also realize a mind. Multiple implementation of one program does not establish multiple realization of one mental state, though material replacement can preserve a state when the subject’s wider organization and history remain fixed.

What’s left open. Content-bearing organization does not by itself settle understanding, state consciousness, or subjecthood. It also does not settle autonomy or welfare. Which implemented systems meet those separate conditions remains an empirical question.

The hand-off. Remove the world and functionalism becomes syntax with excellent wiring. The clerk in the Chinese Room waits to show how little that guarantees.

---

Notes

1. Block opens “Troubles with Functionalism” by observing that functionalism “may now be dominant,” then immediately notes that the label covers several distinct projects — reformulations of behaviorism, mind-machine analogies, applications of empirical psychology, and arguments about mind-brain identity. The version pressed and answered here is the strong one: each mental state is identical to a functional state, a state defined by its causal relations to inputs, outputs, and other inner states. Computational functionalism adds that the relevant functional states are computational states and that running the right program suffices for mentality. Nothing in the broader argument against narrow functionalism turns on machine-table formalism specifically; the later computational argument does. Ned Block, “Troubles with Functionalism,” in Perception and Cognition: Issues in the Foundations of Psychology, ed. C. Wade Savage, Minnesota Studies in the Philosophy of Science, vol. 9 (Minneapolis: University of Minnesota Press, 1978), 261–325, esp. 261–63. ↩
2. The example goes back to Brian Kernighan’s internal Bell Laboratories memorandum A Tutorial Introduction to the Language B (1972), where it illustrates external variables; the phrase arrives split across three of them because a character constant in B held at most four ASCII characters. Kernighan carried it into C in a second Bell Labs memorandum, Programming in C: A Tutorial (1974), and it entered the discipline’s common stock through Brian W. Kernighan and Dennis M. Ritchie, The C Programming Language (Englewood Cliffs, NJ: Prentice-Hall, 1978), sec. 1.1, which prints it in lower case and without terminal punctuation — “hello, world”. The first edition’s program carries no include directive; that line was added in the second (1988). The C++ form given above follows later convention, and nothing in the argument turns on the language, capitalization, or particular string. Chapters 22 and 23 distinguish implementation from realization and formal structure from world-grounded content. ↩
3. The premise commands near-universal assent, though not always for the same reasons. Panpsychism and Russellian monism grant it because on their account nothing phenomenal comes into being anywhere, phenomenal properties sitting at the fundamental level rather than arriving with any particular arrangement of matter; Chapter 10 takes up that family in detail. Integrated information theory ties experience to intrinsic cause-effect power and, on its proponents’ own assessment of conventional computer architectures, finds very little of it there. I assume the premise here in its plain form: writing and running this program brings no new subject into being. Appendix B.18 engages integrated information theory in full. Recent disclosed pretraining counts supply the scale behind the comparison: DeepSeek-V3 used 14.8 trillion tokens, Llama 3.1 405B used 15.6 trillion, and Qwen3 used 36 trillion. DeepSeek-AI, “DeepSeek-V3 Technical Report,” arXiv:2412.19437 (2024); Abhimanyu Dubey et al., “The Llama 3 Herd of Models,” arXiv:2407.21783 (2024); Qwen Team, “Qwen3 Technical Report,” arXiv:2505.09388 (2025). A widely repeated 1997 estimate put the text of the Library of Congress’s books at about 10 terabytes. Using the rough English conversion of four characters per token yields approximately 2.5 trillion tokens per Library and places those disclosed corpora at roughly six to fourteen Libraries; the main text uses ten as a midpoint illustration. See Leslie Johnston, “How Many Libraries of Congress Does It Take?,” The Signal, March 23, 2012; OpenAI, “Understanding and Counting Tokens.” The comparison measures token-scale exposure, not unique books: training data may include repeated material, code, synthetic text, and many languages, and many leading closed-model developers disclose no comparable total. ↩
4. Hilary Putnam, “Philosophy and Our Mental Life,” in Mind, Language and Reality: Philosophical Papers, Volume 2 (Cambridge: Cambridge University Press, 1975), 291–303, at 291 (epigraph); and “Psychological Predicates,” in Art, Mind, and Religion, ed. W. H. Capitan and D. D. Merrill (Pittsburgh: University of Pittsburgh Press, 1967), reprinted as “The Nature of Mental States” in Mind, Language and Reality, 429–440, esp. 432–37. Putnam models the whole organism as a probabilistic automaton whose functional state is fixed by sensory inputs, motor outputs, and transitions among states rather than by physical composition. Functionalism itself remains compatible with dualism. The lesson retained here is that material labels alone do not decide mentality, not that cross-material recurrence follows automatically. Putnam later abandoned the reductive project: see Representation and Reality (Cambridge, MA: MIT Press, 1988), esp. ch. 5. ↩
5. David Lewis, “An Argument for the Identity Theory,” The Journal of Philosophy 63, no. 1 (1966): 17–25; “Psychophysical and Theoretical Identifications,” Australasian Journal of Philosophy 50, no. 3 (1972): 249–258, esp. 249–56; and “Mad Pain and Martian Pain,” in Readings in Philosophy of Psychology, vol. 1, ed. Ned Block (Cambridge, MA: Harvard University Press, 1980), 216–222. Lewis’s Ramsey-sentence method defines mental terms as whatever occupies the causal roles specified by the platitudes of common-sense psychology, allowing the occupant to differ across beings (the “mad pain” and “Martian pain” cases dramatize the two ways role and realizer can come apart). The device makes precise the sense in which functionalism keeps the role fixed while leaving the realizer open. ↩
6. Putnam’s original physical multiple-realization argument appears in “The Nature of Mental States,” 435–37. In Representation and Reality, introduction, xii–xv, and ch. 5, esp. 74 and 82–84, he later argues that the same belief can occur across unlike physical states and across unlike, even nonmappable, computational state spaces. The later result defeats multiple realizability as computationalists had used it: physical multiple implementation of one narrow computational type does not establish multiple realization of one mental type. Whether one world-involving type recurs across different total physical bases remains a further question. ↩
7. Functionalism entails neither physicalism nor computationalism. A dualist could identify mental kinds by causal role while holding that some nonphysical property occupies that role; a functionalist need not identify the relevant roles with states of a program. Its importance for strong AI lies elsewhere: nonbiological systems remain possible occupants of mental roles. The chapter therefore keeps material neutrality while arguing that an internally specified organization leaves out the world-relations that fix content. It does not infer a shared mental type merely from two realizations of one program. ↩
8. The argument rests on meaning holism together with the division of linguistic labor and the environmental determination of reference: Putnam, “The Meaning of ‘Meaning,’” in Mind, Language and Reality, 215–271, taken up in Chapter 16 of this book. Representation and Reality, ch. 5, esp. 74 and 82–87, draws both halves of the lesson. Subjects can share a belief while occupying different, even nonmappable, computational state spaces, so no particular narrow computational type is necessary for that content. Twin-Earth counterparts can share their narrow organization while differing in reference, so no such type is sufficient. Putnam considers collecting unlike states into an equivalence class but argues that the relation doing the collecting is interpretive rather than computable: correct interpretation answers to the world and speech community, not merely to a formalism internal to the believer. His “cats in the neighborhood” example illustrates the point. ↩
9. Hilary Putnam, Representation and Reality (Cambridge, MA: MIT Press, 1988), ch. 5, “Why Functionalism Didn’t Work,” 73–90, esp. 74–75 under “Sociofunctionalism.” Putnam credits Richard Boyd with the reply, lets the relevant system include society and environment—“organisms-cum-environments”—and concludes at that stage that externalism has made functionalism more complicated but “not yet refuted.” In ch. 6, esp. 92–94, he denies that a universal computable equivalence relation can collect diverse computational states into belief types; his closing diagnosis adds an internal-realist objection to the ontology presupposed by wide reduction. The retrospective essay in Words and Life, ed. James Conant (Cambridge, MA: Harvard University Press, 1994), 441–59, esp. 445–46, calls the widened proposal “Global Functionalism.” Putnam did not endorse it. His reduction requirement in Representation and Reality, 77–80, concerns finite definability in physical/computational vocabulary across physically possible cases, not merely an effective decision procedure. The chapter accepts the externalist constraint and pursues capacity-specific physical reductions without claiming that this strategy already answers his universal challenge or accepting his later internal-realist conclusion. ↩
10. The name and the canonical statement come from Gilbert Harman, “Wide Functionalism,” ch. 13 of Reasoning, Meaning, and Mind (New York: Oxford University Press, 1999), 235–43. Harman treats psychological explanation as a kind of functional explanation whose relevant functions concern perceiving and acting in relation to the environment, and takes the wider functionalism to be more plausible than its narrow, more solipsistic rival. His long-arm functional roles — roles whose input side reaches past the sensory surfaces to distal conditions — are the ones Ned Block builds on in “Inverted Earth,” Philosophical Perspectives 4 (1990): 53–79, a case this book takes up in Part Two alongside the identity claim. Putnam’s earlier widening (note 9 above) extends the system to organisms-cum-environments; Harman widens the role itself. Both moves meet the same externalist constraint from different directions: the individuating relations cross the skin. ↩
11. Representation and Reality, appendix, esp. 120–25, argues that, on an unconstrained notion of implementation, every ordinary open system realizes every inputless finite-state automaton, so computational description by itself ascribes almost nothing and cannot constitute mentality. David J. Chalmers, “Does a Rock Implement Every Finite-State Automaton?,” Synthese 108, no. 3 (1996): 309–333, esp. secs. 7–9, replies that genuine implementation requires a system’s causal organization to mirror the automaton’s combinatorial state-transition structure, an observer-independent constraint the triviality construction fails to satisfy. This book follows Chalmers on that question; Chapter 22 treats observer-relativity directly. The concession leaves narrow functionalism no better off: a system may run a program nontrivially while its internal organization still fails to determine what its states are about. ↩
12. The retreat packages a narrow computational type with an external reference-fixing factor. Jerry A. Fodor, Psychosemantics: The Problem of Meaning in the Philosophy of Mind (Cambridge, MA: MIT Press, 1987), ch. 2, treats narrow content as a function from contexts onto truth conditions (47–53). Ned Block, “Advertisement for a Semantics for Psychology,” Midwest Studies in Philosophy 10 (1986), identifies the internal factor with conceptual role. Putnam, Representation and Reality, ch. 3, “Fodor and Block on ‘Narrow Content,’” 43–56, argues that meaning holism and the division of linguistic labor “militate against the whole division of meaning into a ‘narrow content’ and a ‘broad content’” (55). The present reply grants the reusable causal kind and denies that it is the mental type whose multiple realization was at issue. ↩
13. Block has pressed the same conviction — that reductive theories of mind keep omitting something real — across absent qualia, the distinction between phenomenal and access consciousness, and the “mental paint” objection to representationalism. The homunculus-nation is the version aimed squarely at functionalism’s sufficiency thesis. Block, “Troubles with Functionalism” (1978), esp. 279–81. ↩
14. Block’s “homunculi-head” and “Chinese nation” (or “China-body system”) cases, both from “Troubles with Functionalism” (1978, esp. 279–81). The scenarios are built so that the system’s functional organization is, by stipulation, identical to a human’s, isolating the single question of whether that organization suffices for phenomenal consciousness. In the China case, roughly one billion participants carry out local machine-table tasks by radio while the nation controls an artificial body. Block chose the approximate population by reference to the estimated number of neurons in a human brain; he did not assign one citizen to each neuron. The setup makes vivid that every narrowly specified functional fact can be satisfied while the materials and their arrangement are nothing like a brain’s. ↩
15. Block distinguishes absent qualia (the system satisfies the functional specification but has no phenomenal experience at all) from inverted qualia (two systems functionally identical but with systematically swapped experience). A parallel pressure comes from inversion, but I leave it aside here: turning the inverted spectrum against a representational theory of consciousness, rather than against functionalism, opens a separate and harder debate, taken up in Part Two’s treatment of the identity claim. Against functionalism’s sufficiency thesis specifically, the absent-qualia case is enough. For the standard taxonomy see the Stanford Encyclopedia of Philosophy, “Qualia,” sec. 4. ↩
16. The inference is deliberately epistemic. Block stipulates a narrow organization but not the complete world-involving intentional structure. His case therefore does not establish content or consciousness; neither does this reply establish their absence. On the identity claim defended in Part Two, absence of the requisite world-presenting structure would entail absence of the corresponding phenomenal character, but whether the nation has that structure must be shown independently rather than read off its unusual materials. ↩
17. This is semantic externalism brought to bear on the theory of consciousness: what a state represents is not settled by anything internal to the system at a time but by its causal-historical relations to the environment, so two internally identical systems can differ in what their states are about because they are embedded in different worlds. This is precisely the resource narrow functionalism lacks: it specifies internal role exhaustively and is silent about the world-relations that fix content. The positive account is developed in Part Three; see especially the chapters on teleosemantics and on semantic externalism. ↩
18. The identity claim defended in Part Two begins from Tye and Dretske but belongs to the book’s cumulative argument. It identifies phenomenal character with an experience’s complete world-presenting intentional structure: independently fixed content under a perceptual or bodily-affective mode, with determinacy, field structure, and attentional organization. Michael Tye, Ten Problems of Consciousness (Cambridge, MA: MIT Press, 1995), esp. ch. 5, sec. 5.2; Fred Dretske, Naturalizing the Mind (Cambridge, MA: MIT Press, 1995), esp. ch. 3. Tye also distinguishes narrow functional duplication from “maximal” or wide duplication, which preserves external causal interactions, present and historical, and argues that maximal functional duplicates cannot differ phenomenally; Ten Problems, ch. 7, 196–201. The book treats history as content-fixing and uses a separate distributed profile of recurrence, availability, flexible influence, and control to assess state consciousness. Neither history nor that profile gets added to phenomenal character after the fact. The inference here runs from absence of the requisite world-presenting structure to absence of the corresponding phenomenal character; it does not run from content alone to consciousness. ↩
19. David J. Chalmers, The Conscious Mind: In Search of a Fundamental Theory (New York: Oxford University Press, 1996), esp. ch. 7. The argument has unusual dialectical force because Chalmers rejects the logical supervenience of consciousness on the physical yet independently defends organizational invariance. The case for invariance therefore does not merely assume reductive functionalism. ↩
20. Chalmers, The Conscious Mind, ch. 7, “Absent Qualia, Fading Qualia, Dancing Qualia.” The principle of organizational invariance holds that “given any system that has conscious experiences, then any system that has the same fine-grained functional organization will have qualitatively identical experiences” (249). The fading- and dancing-qualia arguments support it via gradual neural-replacement scenarios: a subject whose qualia faded or danced while functional organization held constant would be radically and systematically mistaken about her own present experience, which Chalmers takes to be incoherent. ↩
21. The chapter rejects Chalmers’s universal principle as stated. His neural-replacement scenarios preserve the subject’s body, environmental traffic, causal history, and perceptual organization while changing material. They support continuity in that lineage: a replacement descendant inherits broad content because its history remains fixed. A freshly assembled narrow duplicate does not acquire that content merely by matching internal topology, judgments, and reports. A differently built system may qualify if it independently acquires an equivalent wide role, but that possibility requires its own evidence. Block’s artificial body prevents the cheap claim that his nation lacks present traffic; the narrower point is that the one-hour case stipulates no address-fixing history for the nation as an assembled whole. ↩
22. Hubert L. Dreyfus, What Computers Still Can’t Do: A Critique of Artificial Reason (Cambridge, MA: MIT Press, 1992), the revised edition of What Computers Can’t Do (New York: Harper & Row, 1972), esp. parts II–III, chs. 5–9; for the relevance regress specifically, part III, ch. 8, esp. 256–59. The relevance regress — that picking out the relevant facts requires a rule, whose application requires a further rule, without end — is Dreyfus’s central charge against the rule-and-representation paradigm of “good old-fashioned AI.” A close relative appears in what AI came to call the frame problem, posed in John McCarthy and Patrick J. Hayes, “Some Philosophical Problems from the Standpoint of Artificial Intelligence,” in Machine Intelligence 4, ed. B. Meltzer and D. Michie (Edinburgh: Edinburgh University Press, 1969), 463–502, and given its classic philosophical statement in Daniel C. Dennett, “Cognitive Wheels: The Frame Problem of AI,” in Minds, Machines, and Evolution, ed. Christopher Hookway (Cambridge: Cambridge University Press, 1984), 129–150. Dreyfus’s positive account of the holistic background of skilled coping draws on Maurice Merleau-Ponty, Phenomenology of Perception (1945), esp. part I, ch. 3, and Martin Heidegger’s analysis of the ready-to-hand in Being and Time (1927), esp. div. I, ch. 3, secs. 15–18; the “know-how vs. know-that” contrast is Gilbert Ryle’s. ↩
23. Hubert L. Dreyfus, “Why Heideggerian AI Failed and How Fixing It Would Require Making It More Heideggerian,” Philosophical Psychology 20, no. 2 (2007): 247–268, esp. secs. II–VI. Dreyfus argues that dropping hand-coded rules does not by itself supply absorbed, affectively significant involvement. The present chapter narrows the inference. Training may yield causally internalized content in a mechanism. What does not follow automatically is a unified agent that learns, infers, resolves conflicts, and acts through those contents. Chapter 24 develops that further attribution condition and treats self-maintenance separately as a candidate basis of autonomy and stake. ↩
24. The divergence is deliberate. Dreyfus reads the background and relevance regress as reasons to abandon internal representation, a reading developed with reservations in Michael Wheeler, Reconstructing the Cognitive World (Cambridge, MA: MIT Press, 2005), 187–189, 222–223. This book keeps representational content and grounds it through selection, learning, consumer use, and world-involving history. For the contrast between Searle’s representational account and Dreyfus’s content-free background, see Teodor Negru, “Intentionality and Background: Searle and Dreyfus against Classical AI Theory,” Filosofia Unisinos 14 (2013): 18–34, esp. 25–30. Stake may explain why possibilities matter to an autonomous agent; it supplies no semantic bridge. ↩
25. John Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3, no. 3 (1980): 417–424, esp. 417–19, and “Is the Brain a Digital Computer?,” Proceedings and Addresses of the American Philosophical Association 64, no. 3 (1990): 21–37, esp. 25–29. The chapter adopts Searle’s insufficiency objection—formal syntax alone does not yield semantics—while rejecting his biological-naturalist conclusion. The proposed missing relation is substrate-neutral causal-historical grounding in an environment. ↩