Mind, Matter, and Meaning

Category: artificial intelligence

  • The Church-Turing-Deutsch Principle Cannot Rescue Computationalism

    For years a friend and I argued this question. Add enough lines of code in just the right way, he holds, and you can obtain any possible behavior; accept the computational theory of mind, and write the right program, and any possible conscious experience comes with it.

    The condition carries the argument: accept the computational theory of mind. On that view, the right computation suffices for experience. But why accept it? The Church-Turing-Deutsch principle can look like an answer. It leaves the crucial step unproved.

    In 1985, David Deutsch proposed a universal quantum computer and stated this principle: “Every finitely realizable physical system can be perfectly simulated by a universal model computing machine operating by finite means.” A suitable computer could reproduce a system’s responses to specified tests, without duplicating its material.1

    The older Church-Turing thesis concerns what a person following a mechanical procedure can calculate. Deutsch’s principle concerns physical systems, including brains. That seems to close the gap between a theory of calculation and a theory of mind.2

    Grant the principle.3 The proposed argument runs:

    1. A brain produces conscious experience.
    2. A universal computer can simulate that brain perfectly.
    3. Therefore, running the simulation produces conscious experience.

    The missing premise sits between the second step and the third: the simulation preserves whatever makes the brain conscious. Deutsch’s principle does not tell us whether it does.

    A calculator program really adds because the machine performs the required operations. A weather program calculates how wind and pressure would change without blowing rain across the room. John Searle used rainstorms and fires to make this distinction in “Minds, Brains, and Programs” (1980). Modeling a process does not automatically give the computer every property of the process modeled.4

    But the example cannot decide whether consciousness belongs with calculation or rain. Nor can the word simulation. A computer connected to a valve can really control it. Its operations have effects; they need not merely describe effects occurring elsewhere.5

    David Chalmers thinks the brain simulation belongs with the calculator. In The Conscious Mind, he argues that experience depends on the fine-grained pattern of causal interactions among a system’s parts, rather than the material making them up. He calls this organizational invariance.6

    Imagine replacing your neurons with artificial components that preserve those interactions. They respond as the neurons would, including to inputs you never actually receive. Memory, attention, judgment, and reasoning continue unchanged.

    Now suppose your visual experience fades as the replacements proceed. The tomato’s red drains toward gray, but you still judge that you see the same vivid red. Nothing in your memory or attention registers the loss. Or suppose a switch makes your experience alternate between two colors while every means of noticing the change stays fixed. These are Chalmers’s “fading” and “dancing” qualia cases.7

    He finds that separation between experience and all cognitive recognition deeply implausible. Preserving the organization, he argues, should preserve the experience. A critic must answer this argument, not merely repeat that silicon simulates neurons.

    Chalmers has supplied a reason to accept the missing premise. Its force comes from his account of experience and cognition, not from Deutsch’s principle. The physical principle can contribute to a larger argument; it cannot establish computationalism by itself. Assuming computationalism to fill the gap would beg the question. Defending it independently does not.

    The replacement argument nevertheless does not by itself prove what strong computationalism needs. It changes the neural material while leaving the rest of the subject’s body, surroundings, causal history, and continuing commerce with the world in place. The silicon successor still looks at this red tomato, reaches for this cup, and inherits the organism’s established relations to both. If experience survives the substitution, the case supports substrate independence under those preserved conditions. It does not show that the preserved conditions make no constitutive contribution.

    Chalmers has a direct reply. In discussing Inverted Earth, he allows that environmental reference and wide belief content can change while the subject’s internal organization and experience remain the same. If phenomenal character is narrow in that way, changing the world will not threaten phenomenal organizational invariance.8

    My own view rejects that restriction. It includes the world because the world is necessary. Phenomenal character concerns how an experience looks or feels. It consists in representational content of the right kind: what experience presents and how it presents it. Not every difference in wide content makes a phenomenal difference; Oscar and Twin Oscar may experience H₂O and XYZ in exactly the same way. But differences in phenomenally relevant content, including what color an experience presents, make differences in phenomenal character. The relevant organization is therefore wide. It includes the embodied, causal, and historical relations through which perceptual states acquire their content.

    Now reconsider the dancing-qualia wedge. Chalmers asks us to imagine experience changing while the subject’s judgments, memories, and dispositions remain fixed. The wider alternative can preserve the internal vehicles, causal roles, and dispositions to judge while the world-involving contents of experience, judgment, and memory vary together. No mismatch opens between what the subject experiences and what the subject thinks. This does not refute Chalmers from neutral ground; it identifies the disputed premise. His replacement sequence assumes narrow phenomenal individuation when it holds the world fixed. It cannot then establish that assumption.

    On the wider account, neural replacement preserves experience not because the same abstract internal program must carry the same phenomenology wherever it runs, but because the organism’s complete world-involving organization has been preserved.

    The computationalist now faces a dilemma. If “organization” means narrow internal causal topology, the replacement arguments do not establish organizational invariance. If it includes the necessary relations among brain, body, history, and world, organizational invariance may hold, but it no longer supports the claim that implementing the right internally specified program suffices for mentality. The same program can be copied while those relations differ or fail to exist.

    It also leaves room for artificial minds. A copied brain may inherit a history; going offline does not erase it. Training can give a system content about real features of the world, including language.9 Virtual objects can enter genuine causal relations with an agent that learns from its mistakes. A computational account that includes those relations deserves a hearing.

    We would still need to examine the running system: how its perceptions guide memory and action, how corrections endure, and whether those activities belong to one continuing subject. A formal description alone leaves those questions open. Saying that a simulation is “perfect” does not answer them.

    That closes the route from the Church-Turing-Deutsch principle to strong computationalism. Deutsch shows that physical evolution can be simulated. Chalmers’s cases may show that biological material is not uniquely necessary when the wider organization is preserved. Neither shows that computation by itself constitutes thought or experience. A computer may still think or experience, but only as a concrete system whose embodied and world-involving relations do constitutive work as well. We must establish that those relations exist and explain how they make the process intentional or conscious. We cannot read that answer off the program or borrow it from universal simulability.

    -gts

    This essay is part of Mind, Matter, and Meaning, an AI-assisted philosophy project by Gordon Swobe exploring consciousness, meaning, and artificial minds. Learn more about the project and read the book.

    Notes

    1. 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 98–99. Deutsch defines simulation through computational equivalence under input/output labelings, with equivalent preparations and measurements yielding statistically indistinguishable output distributions. This operational criterion should not be conflated with material duplication, efficient simulation, or an available engineering procedure. His discussion of quantum theory’s compatibility with the principle does not constitute an unrestricted proof of physical universality. The universal quantum computer also does not thereby compute nonrecursive functions. The present argument grants the physical principle rather than deriving it from the mathematical Church-Turing thesis. ↩
    2. Gualtiero Piccinini, “Computationalism, the Church–Turing Thesis, and the Church–Turing Fallacy,” Synthese 154 (2007): 97–120, https://doi.org/10.1007/s11229-005-0194-z. Piccinini explicitly credits Jack Copeland for the fallacy’s name. His distinction separates the generic claim that cognitive processes compute from the technical claim that a computation falls within Turing-computability. Establishing the latter does not establish the former. His conclusion treats computationalism as an empirical hypothesis about cognitive organization, not a consequence of results concerning effective calculability. The present essay addresses the further phenomenal conclusion: even a warranted computational explanation of some cognitive capacities would require an account of why its realization suffices for experience. ↩
    3. Christopher G. Timpson, “Quantum Computers: The Church-Turing Hypothesis Versus the Turing Principle,” in Alan Turing: Life and Legacy of a Great Thinker, edited by Christof Teuscher (Berlin: Springer, 2004). The argument also appears in chapter 6 of Timpson’s DPhil thesis, Quantum Information Theory and the Foundations of Quantum Mechanics (University of Oxford, 2004), arXiv:quant-ph/0412063. Timpson disputes Deutsch’s proposed underwriting of the mathematical hypothesis by a physical principle, distinguishing effective calculability from physical simulability. This criticism neither establishes nor refutes physical CTD merely by separating the claims. Granting CTD here isolates the inference to mentality; it does not certify the physics or treat the mathematical and physical theses as interchangeable. ↩
    4. John R. Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3, no. 3 (1980): 417–424. The fire-and-rain example establishes that preservation of a computational description does not automatically preserve every property of the described process. Its application to consciousness requires an additional account of consciousness-relevant causal powers. The example therefore supplies neither a general prohibition on realization through simulation nor a classification of consciousness with combustion or precipitation. Its limited inferential role does not depend on accepting the Chinese Room argument advanced in the same paper. ↩
    5. Shuqin Ma and Ryota Kanai, “Intrinsic Computational Functionalism: From Observer-Relative Maps to Observer-Independent Structures,” arXiv:2606.06424v1 (2026), secs. 3–5, propose label-independent instantiation and intervention-sensitive causal-dynamical organization as constraints on intrinsic computation. These constraints address unconstrained observer-relative mappings without selecting the structures sufficient for consciousness. Ryota Kanai and Shuqin Ma, “Intrinsic Computational Functionalism and Simulated Consciousness,” arXiv:2606.15348v1 (2026), develop a conditional preservation argument encompassing physical implementation, state individuation, transitions, intervention profiles, and the relevant agent-body-world boundary. If consciousness is invariant under that organization and an implementation realizes it, consciousness is preserved. This is stronger than a narrow input/output argument and does not claim that every simulation, or any particular current system, meets the conditions. A criticism must identify a relevant omission rather than infer one from the designation “simulation.” ↩
    6. David J. Chalmers, The Conscious Mind: In Search of a Fundamental Theory (New York: Oxford University Press, 1996), 328; the defense of phenomenal organizational invariance appears in chapter 7. The relevant organization includes fine-grained causal dependence and counterfactual response, not merely an actual sequence of outputs. Chalmers’s combination of organizational invariance with property dualism illustrates the distinction between computational sufficiency and physicalist reduction. A principle relating experience to organization could, on his account, hold as a psychophysical law without identifying experience with organization. ↩
    7. Chalmers, The Conscious Mind, chapter 7; and “A Computational Foundation,” sec. 3.2. Fading qualia concern progressive experiential loss under organizational preservation; dancing qualia sharpen the challenge through switching between functionally equivalent implementations allegedly associated with different experiences. Cognitive noticing must involve a functional difference, so the stipulated organizational equivalence excludes it. The argumentative pressure comes from the implausibility of radical experiential change without corresponding cognitive recognition. This is not a deductive contradiction unless further premises exclude such dissociation. Chalmers presents the arguments as plausibility arguments about consciousness and cognition. Rejecting them requires addressing that pressure rather than merely redescribing silicon as a simulation. ↩
    8. Chalmers, “A Computational Foundation,” sec. 3.2, expressly excludes knowledge and environmentally individuated belief from unrestricted organizational invariance while defending phenomenal invariance. He also considers Inverted Earth in The Conscious Mind, pp. 265–274. The present reply does not infer a phenomenal difference from every difference in wide content. It makes a methodological point about the scope of the replacement cases: varying substrate while retaining the subject’s environment and ongoing causal relations cannot establish invariance across changes in those relations. The additional identity thesis stated in the body—that phenomenal character is representational content under the relevant perceptual or bodily-affective mode—explains why the fixed background belongs inside the relevant organization. It still requires an account of which content differences bear on phenomenal character. An externalist argument concerning belief alone establishes neither a phenomenal difference nor the absence of consciousness. ↩
    9. Fintan Mallory, “Teleosemantics for Neural Word Embeddings,” Mind & Language (2026), https://doi.org/10.1111/mila.70037, develops a consumer-based account of word2vec representations concerning the linguistic contexts in which words occur. Linguistic distributions are features of the world; their representation should not be dismissed through an all-or-nothing demand for reference to ordinary physical objects. This supports a serious mechanism-level content attribution without establishing whole-model understanding or a conscious subject. Whether content is original at this limited grain remains unsettled here. The relevant diagnostic asks whether causal-functional history and current producer-consumer use fix accuracy conditions not exhausted by an interpreting practice. Engineered ancestry, inherited vocabulary, or the source of a training objective does not independently settle that question. Derived content also remains genuine content. Similarly, copying and disconnection require an investigation of preserved causal-historical relations rather than an automatic verdict of semantic deprivation. ↩

    Get new essays by email

  • Philosophical Zombies

    You can imagine it, and I am not going to stop you. A creature wired exactly as you are, neuron for neuron, firing in the same sequence at the same millisecond, flinching from the flame and saying ouch with full conviction, and no one home. No felt sting. No inward flush of red when it eyes the tomato. Most defenders of a physical mind spend their energy trying to talk you out of this picture: you only think you have imagined a zombie, they say; look harder and the lights come back on. I want to try the opposite tack. Imagine the zombie all you like. Conceive it as vividly and as carefully as any philosopher ever has. I will concede the whole thing, and the argument still will not reach its conclusion.

    That concession costs far less than it looks, and seeing why is the work of this essay.

    The zombie is the centerpiece of the strongest contemporary case against a physical account of consciousness. David Chalmers, who named the “hard problem” in the mid-1990s and has spent the thirty years since keeping scientists from waving it away, runs the case in three moves.1

    1. Zombies are conceivable. We just conceived of one.
    2. Whatever is conceivable is, in the relevant sense, possible. A creature exactly like you physically, with no inner experience, could exist in some genuinely possible world.
    3. If consciousness can be absent while every physical fact obtains, then consciousness is not identical to any physical fact. Something extra has to be added to physics to get experience.

    Most physicalists swing at the first premise. They deny that you have really imagined a zombie, or that you could.2 I will let the first premise stand and swing instead at the second: the quiet little word therefore wedged between “I can conceive it” and “it could exist.” That bridge carries the entire weight of the argument, and it is the one plank Chalmers is least entitled to.

    Conceiving is one thing, being is another

    Conceiving is a fact about thinking. Possibility is a fact about being. The two come apart whenever a necessity has to be discovered rather than reasoned out from the armchair, and the man who taught us this was Saul Kripke, whose 1970 Princeton lectures reset the whole topic of necessity in three afternoons.3 Take his stock case. You can conceive of water that is not H₂O; hold the phrase in mind and no contradiction rises to the surface. Yet water could not have been anything other than H₂O. The identity holds in every world there is. So your act of conceiving tracked something real, but something about your concepts, not about water: the fact that “water” and “H₂O” reach their shared referent by two different routes, one through the wet stuff you drink and one through the chemistry. The necessity is there in the world; it is simply a posteriori, knowable only by looking.

    Now carry that over. Suppose an experience consists in a physical process, a particular world-directed representational state of an embodied brain, as I argue throughout this project.4 If that holds, it holds as a discovered identity, not an armchair one. Nobody deduced that the smell of coffee consists in such-and-such neural representing by reflection; if that holds, it holds the way “water is H₂O” holds, necessary in every world but legible only from the third-person side, by investigation. And here is the payoff: because the felt route in and the physical route in are two different routes, you can hold the physical description fixed, run your first-person concept of the feeling, notice that nothing in the physical story spells out the feeling by name, and so conceive its absence. The conceiving is perfectly real. What it tracks is the distance between two concepts of one thing, never a distance between two things.

    This is why I can afford to be so generous with the first premise. A zombie’s conceivability is not an embarrassment for the physicalist. It is exactly what the physicalist should expect, given that we grasp one state by two cognitively independent means. Joseph Levine saw the shape of this back in 1983, granting the gap between physical and felt descriptions while taking care to call it explanatory, a gap in what we can derive rather than a crack in the world.5 Conceding conceivability and still denying possibility is not a dodge. It is the accurate reading of what the thought experiment actually hands you.

    The objection that says the bridge is safe

    Chalmers saw this reply coming, naturally, and he built his defense into the argument itself; it deserves a fair hearing.6 The water case works, he points out, only because “water” carries a contingent reference-fixer, the watery stuff around here. Peel that fixer away and the leftover conceivability, watery stuff that isn’t H₂O, does describe a genuine possible world after all. So conceivability tracks possibility even there; water only looked like a counterexample. Run his two-dimensional apparatus over the zombie, and he argues there is no contingent fixer left to peel: the zombie stays conceivable all the way down, under every way of reading its terms, and a scenario conceivable all the way down is possible. A physicalism like mine would need what he calls a strong necessity, a truth necessary in every world with no possible world standing behind the conceivability at all, and Chalmers holds there are no such things.

    The reply lives in the phenomenal concept, and the first thing to say is that Chalmers reads it the way I do. He agrees that when I think this, meaning the very quality grasped by undergoing it, no contingent fixer (nothing like the watery stuff around here) stands between the concept and its object. The grasp is direct; the concept’s whole route in just is the experience.7 For him, that directness seals the argument: with no fixer to peel away, the zombie’s conceivability runs all the way down, and a scenario conceivable all the way down must describe a possible world — unless some necessity holds with no possible world standing behind the conceiving at all. Exactly there the type-B physicalist plants the flag. A concept installed by undergoing an experience and a concept built from description stay cognitively independent forever; if the state is one, that permanent independence just is a necessity whose negation stays conceivable — the strong necessity Chalmers declares impossible. His denial that such necessities exist is not a finding his apparatus delivers; it is the premise doing the work. The zombie’s conceivability is robust, yes, surviving every fact you pour in, but it survives because the two concepts never close on each other, not because two worlds answer to them. The inference needs worlds. The phenomenal case supplies only concepts.

    Katalin Balog pressed the same point from the far side, with a stroke I have always found hard to shake. Suppose the argument succeeds and the zombie world is real. Your zombie twin, your exact physical duplicate, runs this very conceivability argument there, matching your reasoning move for move and uttering your premises with your exact conviction, and concludes that it is non-physical — which is false, since its world is by stipulation nothing but physical.8 A defender can reply that the twin merely lacks justification while the inference form stays sound; but a form that carries a reasoner from premises he cannot fault to a falsehood about his own nature is not a form to hang a metaphysics on. If it fails in the twin’s mouth, it earns no trust in ours. The zombie, conjured to abolish physicalism, quietly discredits the argument that conjured it.

    Two duplicates, two verdicts

    Before anything gets built on this reply, a plain statement of what it buys. It refutes nobody. A dualist can grant that my two concepts explain the conceiving and still insist that the conceiving, so explained, tracks a genuine possibility; I cannot rule that out without assuming the very identity in dispute, and I would rather say so than dress the point up. What the explanation removes is the argument’s free ride. The zombie arrived as evidence, a datum no physicalist could accommodate. But a view cannot be indicted by a datum it predicts: this physicalism says the gap between the felt concept and the physical concept will stay open permanently, and the gap stays open. Its staying open therefore counts as evidence for the view, or as nothing at all. Where does that leave the metaphysics? Where it always had to be settled — on the merits of the identity claim itself, not in the imagination.

    The modal reply clears the ground; what grows on it depends on what felt character actually consists in. So look closely at the creature Chalmers actually built, because one of its specifications carries more weight than the seminar-room summaries suggest. His zombie is a physical duplicate housed in a world physically identical to ours: same particles, same laws, and the same history, the whole long traffic between organisms and their surroundings that shaped every state the duplicate occupies. On the account I defend, that last clause decides the case. What a state represents depends on the history that laid it down; the frog’s visual state signals fly because flies, across generations of frogs, did the shaping. A duplicate sharing my physics and my whole history therefore shares my representational content, and if felt character consists in content of the right kind, it shares the felt character too.9 Read out to its final specification, and granting the identity claim its day, Chalmers’s scenario describes a creature with the lights on. The conceiving survives, as everything above says it must. The zombie does not.

    Now build the other creature. Keep the functional organization, the same input-output profile and the same causal roles, and strip the history: no body that grew, no environment that pushed back, no run of perceptual contact with a world it had to make its way in. Chalmers holds, on independent grounds, that the lights stay on here too; his principle of organizational invariance ties experience to organization wherever it appears.10 Here I stop conceding. Strip the content-fixing history and you strip the content, and with the content goes the felt character. This second zombie, fluent and responsive and dark, I do not merely admit as a possibility — and here the argument steps out of the seminar room. A trained system has a history of a kind, corpora read and weights tuned, but no perceptual contact and no history of making a living in a world, and systems answering that description are already talking to us. So one thesis delivers both verdicts: felt character consists in representational content, and content gets earned in a world. Chalmers’s duplicate inherited my earnings along with my physics. The stripped duplicate never earned at all.

    What you imagined, and what you were sold

    So grant the zombie. Conceive it in full daylight, no flinching. What you have shown is that the concept of the felt and the concept of the physical never melt into one another in thought, that no quantity of brain science will ever let you read off the taste of coffee from a wiring diagram. True, and worth knowing. It records a permanent distance between two ways of reaching one state, one by describing it from outside and one by living it from within.

    The zombie argument takes that distance, real and ineliminable, and posts it as a seam in the world: a second order of fact where there is one thing, grasped two ways. The conceivability it trades on is the genuine article; only the exchange rate is counterfeit. What you imagined was the independence of your concepts. What you were sold was the independence of consciousness from the physical. The first comes cheap and true. The second is the thing that was never in the box.

    And keep the two creatures straight as you leave, because the traffic runs both ways. The duplicate that shares your physics and your history is no zombie, however vividly you conceive its darkness. The system that shares only your conversational surface may well be one, and it will pass every test the first was imagined to pass. The imagined zombie was aimed at physicalism and missed. Its nearest real relatives, if I am right, are already talking to us.

    -gts

    This essay is part of Mind, Matter, and Meaning, an AI-assisted philosophy project by Gordon Swobe exploring consciousness, meaning, and artificial minds. Learn more about the project and read the book.

    References

    Balog, Katalin. 1999. “Conceivability, Possibility, and the Mind-Body Problem.” Philosophical Review 108 (4): 497-528.

    Chalmers, David J. 1996. The Conscious Mind: In Search of a Fundamental Theory. New York: Oxford University Press.

    Chalmers, David J. 2002. “Consciousness and its Place in Nature.” In The Blackwell Guide to Philosophy of Mind, edited by Stephen P. Stich and Ted A. Warfield. Oxford: Blackwell.

    Chalmers, David J. 2003. “The Content and Epistemology of Phenomenal Belief.” In Consciousness: New Philosophical Perspectives, edited by Quentin Smith and Aleksandar Jokic. Oxford: Oxford University Press.

    Chalmers, David J. 2007. “Phenomenal Concepts and the Explanatory Gap.” In Phenomenal Concepts and Phenomenal Knowledge: New Essays on Consciousness and Physicalism, edited by Torin Alter and Sven Walter. Oxford: Oxford University Press.

    Davidson, Donald. 1987. “Knowing One’s Own Mind.” Proceedings and Addresses of the American Philosophical Association 60 (3): 441-458.

    Dretske, Fred. 1995. Naturalizing the Mind. Cambridge, MA: MIT Press.

    Kripke, Saul. 1980. Naming and Necessity. Cambridge, MA: Harvard University Press.

    Levine, Joseph. 1983. “Materialism and Qualia: The Explanatory Gap.” Pacific Philosophical Quarterly 64: 354-361.

    Loar, Brian. 1990. “Phenomenal States.” Philosophical Perspectives 4: 81-108.

    McLaughlin, Brian P. 2012. “Phenomenal Concepts and the Defense of Materialism.” Philosophy and Phenomenological Research 84 (1): 206-214.

    Papineau, David. 2002. Thinking about Consciousness. Oxford: Oxford University Press.

    Tye, Michael. 1995. Ten Problems of Consciousness: A Representational Theory of the Phenomenal Mind. Cambridge, MA: MIT Press.

    Tye, Michael. 2009. Consciousness Revisited: Materialism Without Phenomenal Concepts. Cambridge, MA: MIT Press.

    Notes

    1. The mature argument is in Chalmers (1996, esp. chs. 3–4), with the canonical compact statement in Chalmers (2002). The three steps in the main text are a simplification of a presentation that, in full, runs the modal inference through a two-dimensional semantics of primary and secondary intensions; the simplification is faithful to the argument’s load-bearing structure, which is what this essay contests. Chalmers’s premise is not bare imaginability but ideal positive conceivability — conceivability that no amount of better reasoning would defeat — a strengthening that matters for the reply in note 6. ↩
    2. Two physicalist routes run against the zombie argument, and this essay takes the second. The first denies the conceivability premise outright: the felt-absent-while-physical-present scenario is only apparently conceivable, because it smuggles in an inner-theater picture of phenomenal character as a separable inner item. The route taken here grants premise one in full and refuses the bridge to premise two. I once treated the pair as complementary and no longer do. The first route asks a reader to accept that something they can plainly do, they cannot really do, and it stakes the case on winning a contested claim the argument never needed; it also sits badly with a physicalism meant to preserve what experience is rather than explain it away, since denying the conceiving is the eliminativist move transposed onto the modal question. What survives from that route is a supporting observation, not a reply: the modal weight the argument requires has to be loaded onto the imagining from outside, by the picture the reader brought in. The chapter “Why Dualism Keeps Winning Arguments It Should Lose” in Mind, Matter, and Meaning takes the route this essay takes, and sets out the reasoning at length. ↩
    3. Kripke (1980), originally delivered as lectures in 1970. The relevant apparatus is the distinction between epistemic and metaphysical modality, underwritten by rigid designation: identity statements linking rigid designators (“water,” “H₂O”; “Hesperus,” “Phosphorus”) are, where true, necessarily true, yet many are knowable only a posteriori. The conceivability of their negations is therefore a fact about the cognitive independence of the designators, not a guide to metaphysical possibility. This is the lever the present essay applies to the psychophysical case. ↩
    4. The identity claim defended across this project: phenomenal character consists in representational content of the right embodied, world-directed kind (Tye 1995). In Chalmers’s (2002) taxonomy of responses to the hard problem, the position here is type-B physicalism — the psychophysical identity is held to be metaphysically necessary but knowable only a posteriori, exactly parallel to “water = H₂O.” Papineau (2002, chs. 4–7) gives the most developed defense of an a posteriori mind-brain identity of this kind. The essay’s modal point is independent of which physical state turns out to be the right one; it concerns the grade of necessity, not its content. ↩
    5. Levine (1983) framed the gap epistemically: a complete physical description of a brain state does not entail, in any deductively transparent way, a description of what undergoing the state is like. Levine himself declined to read the gap as evidence for property dualism; conceding that zombies are conceivable while denying that their conceivability establishes their possibility is the natural extension of his epistemic framing, and exactly the asymmetry this essay defends. Levine, unlike Chalmers, never treated the gap as a license for the modal inference. ↩
    6. The defense is the two-dimensional argument against materialism (Chalmers 2002, sec. 5; 2003). Its crux is the denial of “strong necessities”: necessary truths whose negations stay ideally primarily conceivable even though no possible world stands behind the conceiving. In the zombie case the apparatus adds a further step — for phenomenal terms, primary and secondary intensions coincide — so no Kripkean reading survives on which the conceived scenario describes a genuine world under a different description. If both claims hold, the a-posteriori-necessity escape route the type-B physicalist relies on closes. The reply in the main text is that phenomenal concepts are the standing counterexample to the strong-necessity generalization, not an exception that needs special pleading. ↩
    7. The direct-reference or recognitional account of phenomenal concepts (Loar 1990; Papineau 2002). Chalmers (2003) himself analyzes pure phenomenal concepts as directly referring, with primary and secondary intensions that coincide — the directness is common ground between the parties, not the point of dispute. What the recognitional account supplies the type-B physicalist is the ground for a strong a posteriori necessity: a psychophysical identity whose negation stays ideally conceivable because the concepts flanking it share no descriptive route reflection could close. The dispute therefore turns on Chalmers’s global denial that strong necessities exist (note 6), not on the semantics of the concepts. The strategy is contested — Chalmers (2007) presses his master-argument dilemma that whatever cognitive feature explains our epistemic situation is either physically explicable (and so, he argues, cannot do the work) or not (and so concedes dualism); McLaughlin (2012) defends the recognitional account against Tye’s rival, phenomenal-concept-free materialism. A wrinkle worth marking: this essay leans on early Tye (1995) for the identity claim while leaning on the recognitional account of phenomenal concepts that later Tye (2009) repudiates. The divergence is deliberate — the project keeps Tye’s representationalism and parts company with his later rejection of phenomenal concepts, siding with Loar and Papineau on the concepts even as it sides with Tye on the metaphysics. This project locates its answer to the dilemma in the transparency of experience and the identity claim, which keep the gap-generating concepts inside the physical order; that defense is a separate argument from the modal one pressed here. ↩
    8. Balog (1999): the zombie-parity objection — her own name for it is the “Zombie Refutation.” A physical duplicate of the zombie-arguer, by hypothesis fully physical, would run the same argument, from premise-utterances indistinguishable from the original arguer’s and no easier to fault, to the conclusion that it is non-physical — a conclusion false in its own case. (Whether the twin’s premises carry the same content as yours is itself contested on recognitional accounts of phenomenal concepts; the objection’s force survives the qualification, since an inference form this fragile under physical duplication supplies no independent modal evidence.) An inference that delivers a falsehood from apparently-true premises in the twin’s mouth is unreliable in ours. The objection targets the conceivability-to-possibility step directly and is independent of any particular diagnosis of why the step fails, which makes it a useful companion to the a-posteriori-necessity reply. ↩
    9. The historical dependence of content is the teleosemantic thread running through this project: a state’s representational content derives from the world-involving history — evolutionary, developmental, learned — that fixed what the state is for. Dretske (1995) gives the canonical statement of the etiological view; Tye (1995) builds the identity claim on content of this externally grounded kind. The standing pressure point is Davidson’s (1987) Swampman, a molecule-for-molecule duplicate assembled by cosmic accident, with no history at all. That scenario differs from Chalmers’s zombie in exactly the respect the main text isolates: it strips the history while keeping the physics. The account defended here takes the cost openly — an ahistorical momentary copy has indeterminate content rather than the determinate content its original enjoys, a concession argued at length elsewhere in this project. Note the asymmetry that protects the anti-zombie verdict: Chalmers’s canonical zombie world is stipulated physically identical to ours in its total history, not merely in a time-slice, so the duplicate inherits the full content-fixing past and the Swampman concession never engages. ↩
    10. The principle of organizational invariance: systems sharing the same fine-grained functional organization have qualitatively identical experiences (Chalmers 1996, ch. 7, where the fading- and dancing-qualia arguments carry the case). Chalmers advances the principle as naturally, not logically, necessary — he grants that absent qualia remain ideally conceivable — so the disagreement here is direct rather than modal: the same history-free duplicate, opposite verdicts about what the actual world’s laws deliver. On the view defended here the principle fails because organization alone, absent world-involving history, fixes no determinate representational content, and content of the right kind is what felt character consists in. The systems gestured at in the main text — large language models and their agentic descendants — are not functional duplicates of any human being; they instantiate fluent linguistic organization with no perceptual contact and no history of making a living in a world, which on this account is precisely the profile of form without felt character. ↩

    Get new essays by email

  • How Multiple Realization Undid Machine-State Functionalism

    Hilary Putnam founded machine-state functionalism, also called computational functionalism. He later turned its best argument against it. In Representation and Reality, he compared the move to jujitsu: use an opponent’s strength to throw him. The opponent happened to be his own earlier theory. Philosophers admire that sort of candor, preferably from a safe distance.

    He did not hide or minimize his defection in a footnote. Chapter 5 bears the title “Why Functionalism Didn’t Work.” No euphemism there. The whole chapter says where the theory failed.

    The reversal took more than twenty years. It began with a case against brain-state identity, gained a second edge from semantic externalism, and ended as a case against identifying mental states with computational states. The logic becomes easier to see if we take those steps in order.

    The First Turn: One Mental State, Several Physical States

    In the 1950s and 1960s, identity theorists proposed that mental-state types might prove identical with brain-state types. Pain, for example, might turn out to be C-fiber firing. Putnam thought the proposal tied minds too tightly to one kind of animal. Humans, octopuses, and possible extraterrestrials could all feel pain while possessing very different nervous systems.

    His first argument can be stated in three steps:

    1. If pain were identical with one physical-state type, every creature in pain would occupy that physical-state type.
    2. Creatures with very different physical designs could all feel pain.
    3. Therefore, pain cannot be identical with any one physical-state type.

    This argument does not deny that every actual pain has a physical realization. It denies a type identity. A human pain and an alien pain may each depend on physical activity without depending on the same physical kind.

    Machine-state functionalism offered an attractive replacement. Define pain by the role it plays, not by its material. Ask what tends to cause it, what it tends to cause, and how it interacts with belief, memory, and desire. Different physical systems could occupy the same functional state, much as different computers can run the same program. Putnam had freed the mind from one material by giving it an abstract job description.

    The Middle Turn: The World Helps Fix Content

    In 1975, Putnam published “The Meaning of ‘Meaning.’” Its Twin Earth story added a different pressure. Imagine two internally matching people in 1750. One lives on Earth among H₂O. The other lives on Twin Earth among XYZ, a liquid that looks, tastes, and behaves like water. Their internal states match. Their words do not refer to the same substance. The Earthling’s “water” reaches H₂O; the twin’s reaches XYZ.

    This result did not yet amount to Putnam’s full case against functionalism. The paper focused on linguistic meaning and reference. Still, it established a limit that any internal machine theory would have to respect: what a state concerns can depend on relations outside the thinker. Matching inner organization does not guarantee matching content.

    By 1988, Putnam joined this externalist result to his older multiple-realization argument. The result was the promised jujitsu.

    The Second Turn: One Mental State, Several Computational States

    Machine-state functionalism had rejected the demand for one physical type per mental type. Yet it still treated mental states as computationally defined states. Putnam now asked why a shared belief should require a shared computational organization.

    Consider two people who both believe that many cats live nearby. One counted paw prints in the garden. The other heard nightly cat fights. They may speak different languages, hold different background beliefs, and draw different inferences from what they believe. An alien could share the belief through a cognitive design unlike either human system.

    Different routes into the belief would not settle the issue. A machine can reach the same state by several routes. A machine state gets its identity from its place in a transition structure: which inputs and internal states lead to it, and which outputs and further states follow from it. Differences in language, background belief, and habits of inference can change that entire pattern. Across sufficiently different thinkers, the computational state spaces may not even map neatly onto one another. Yet we may still have good reason to interpret the thinkers as sharing the belief.

    The second argument has the same skeleton as the first:

    1. If believing that many cats live nearby were identical with one computational-state type, every possible thinker with that belief would occupy that type.
    2. Possible thinkers can share that belief while occupying states with different transition roles, perhaps within computational systems whose state spaces cannot be mapped onto one another.
    3. Therefore, the belief cannot be identical with any one computational-state type.

    Putnam called functionalism’s contrary assumption its “Achilles’ heel.” It had denied one brain-state type for every belief across species, then quietly kept a one-state-per-belief picture within each organism. But learning histories, cultures, languages, and styles of thought vary within our species. Possible thinkers could vary far more. The same content can cross those computational differences.

    A functionalist can answer by replacing one state with a class. Put every computational state that counts as believing that cats live nearby into the same box. The states need not match internally; they need only belong to the right equivalence class.

    Now ask what puts them in the box. The rule must decide when cat, meew, and an alien signal concern the same animals. It must separate an unfamiliar belief from a familiar belief expressed in an unfamiliar way. In short, it must interpret the thinkers before it can sort their computational states. The proposed reduction then moves in a circle. It explains shared content by membership in a class, but identifies the class by deciding which systems share the content.

    A functionalist may insist that some objective equivalence relation exists even if no finite thinker can find it. Putnam need not prove otherwise. Machine-state functionalism needs a rule stated in computational terms. An endless list built from prior acts of interpretation would only record our semantic verdicts. It would not reduce them to computation.

    The problem runs in the other direction as well. Twin Earth gives us internally matching systems with different contents. If their internal physical organization matches, their internal computational organization can match too. Yet one “water” state concerns H₂O and the other concerns XYZ. No formal transition inside either system settles that difference. The environment does.

    The two arguments work in opposite directions. Multiple realization challenges necessity. Twin Earth challenges sufficiency:

    • Same content, different computational states: no particular internal computational-state type is necessary for that content.
    • Same internal computational state, different content: no particular internal computational-state type is sufficient for that content.

    The first point defeats the demand that every instance of a belief share one computational type. The second defeats the claim that occupying a computational type fixes what the belief concerns. Together they block the one-to-one identity proposed by simple machine-state functionalism.

    What the Argument Does Not Show

    The conclusion needs a fence. Putnam’s argument rejects the constitutive identity just defined. It does not deny that brains compute, that computational models explain parts of cognition, or that an artificial system might think. The modest claim that computation contributes to thought remains untouched.

    Nor did Putnam dispose of every wide form of functionalism. A theorist may let the relevant role extend through the body, environment, social practice, and causal history. Putnam considered such a sociofunctional rescue. It accepts much of his externalism and abandons the picture of an isolated program running inside one skull. It may still count as functionalism. But formal computation alone no longer fixes the content. The world-involving relations now carry part of the load.

    Putnam also offered a separate argument about implementation. Under a permissive mapping, an ordinary physical system can count as realizing any finite automaton. The philosopher David Chalmers later required the system’s causal structure to mirror the formal transitions. That reply may secure objective computation and block the cheap mappings. Grant it. The semantic problem remains: a genuine computation does not determine its own reference merely by running.

    My View

    Putnam supplies the negative result: formal computation cannot by itself tell us what a state concerns. My positive account goes beyond his. A world-involving producer-consumer history—through learning, selection, or both—may fix a representation’s accuracy conditions: what it concerns and what would make it mistaken. That can establish content at the level of a mechanism without a separate proprietor. Whether the larger system thereby believes, understands, or acts depends on a further question: how the state participates in the system’s integrated cognitive organization.

    Neither a world-involving history nor cognitive integration requires carbon. An artificial mechanism could acquire non-derived content, and an artificial system might think by means of it. Silicon may qualify; an abstract program, considered only as a formal transition structure, does not.

    Multiple realization first freed mental states from one physical material. Properly followed, it also frees them from one computational form.

    -gts

    This essay is part of Mind, Matter, and Meaning, an AI-assisted philosophy project by Gordon Swobe exploring consciousness, meaning, and artificial minds. Learn more about the project and read the book.

    References

    Putnam, Hilary (1967). “Psychological Predicates.” In W. H. Capitan and D. D. Merrill (eds.), Art, Mind, and Religion, 37–48. Pittsburgh: University of Pittsburgh Press. Reprinted as “The Nature of Mental States” in Mind, Language and Reality: Philosophical Papers, Volume 2. Cambridge: Cambridge University Press, 1975.

    Putnam, Hilary (1975). “The Meaning of ‘Meaning.’” In Keith Gunderson (ed.), Language, Mind, and Knowledge, Minnesota Studies in the Philosophy of Science 7, 131–193. Minneapolis: University of Minnesota Press.

    Putnam, Hilary (1988). Representation and Reality. Cambridge, MA: MIT Press.

  • A Workspace Is Not a Subject

    Drive a familiar route while your mind wanders, and you can arrive with no memory of the last ten minutes of turns. Something drove the car — read the signals, worked the wheel, braked for the cyclist — and never once surfaced in the part of you that will later narrate the day. Then a horn sounds, and the driving snaps back into view: reportable, deliberate, exactly what you’d swear, if asked, you’d been doing the whole time.

    Cognitive science names the split, and everyone who meets the name assumes it settles more than it does: a small conscious cockpit steering, a great deal of unconscious machinery keeping the plane up. So when Anthropic researchers announced in July 2026 that they’d found something in Claude behaving like that cockpit — a small, capacity-limited set of internal contents you can name and track — the two ambush reactions arrived on schedule. One says the machine woke up. The other calls it autocomplete with better PR. Both skip what the researchers actually found, and what it actually shows.

    The theory under test comes from psychologist Bernard Baars, who proposed in 1988 that the brain runs as a crowd of specialists working mostly in the dark about each other, with a narrow workspace broadcasting select output widely enough for report and voluntary control.1 Stanislas Dehaene’s group later found the neural mechanism: long-range connections that “ignite” once evidence crosses a threshold.2 On July 6, a sixteen-researcher Anthropic team built the language-model analog: a mathematical lens reading which words a given slice of the network’s activity is currently disposed to produce.3 Then they ran the test that separates a finding from a correlation. Ask the model which sport goes with a country; catch the moment its activity settles on “soccer”; swap that pattern for “rugby.”4 The answer flips — not because you changed the prompt, but because you changed a few thousand numbers mid-thought. Do that across enough tasks and a workspace shows up by the only definition worth having: a small slice of everything the network computes, broadcast widely enough to get reported, held onto, and put to use. Everything else keeps humming along regardless.

    Credit where due: the researchers call global workspace theory “a useful comparison point,” not a proof, and note rivals exist.5 On whether access connects to subjective experience, they “take no position.”6 Admirable. Also, a little maddening — you can respect a team’s hygiene and still wish they’d just told you the answer.

    So what does the workspace hold? Their own answer: “a small, evolving set of unspoken words… naming the concepts the model is currently reasoning with.”7 Unspoken words. Sit with that. Attend to your own experience and you find the world — the tomato, red and ripe on the counter — never an inner picture of it. Point the same instrument at the model’s nearest analog and you find vocabulary. The inner medium matches the outer medium exactly, and the authors say why: the workspace is verbal because the model’s output is verbal. Your conscious life mixes words with sight, sound, and the ache in your knee. The model’s workspace runs on words, and only words.

    That sits at a suggestive angle to the strongest physicalist theory of experience going. Michael Tye has spent three decades arguing that phenomenal character — what seeing red is like — consists in representational content poised for use in belief and desire.8 The paper reaches for the same word: representations “poised to be spoken about.” Real overlap — both name a readiness for cognitive use. But Tye’s account demands more: the content has to present the world, not the word for it, and it has to run finer than language, since experience discriminates shades of red you have no name for. Grain, the model gets partial credit for — blends of word-vectors can shade finer than any single label. The world, it never touches. Nothing in a workspace built entirely of unspoken words reaches past the dictionary to a tomato.

    Why should words all the way down fall short of meaning? Emily Bender and Alexander Koller gave the standard answer in 2020: a system trained only on form “has a priori no way to learn meaning,” because meaning lives in the relation between form and something outside the text.9 Claude’s tokens relate beautifully to other tokens. Whatever worldly ancestry they carry belongs to the humans who wrote the training data; the model inherits the form, never the reference. And here the paper’s own closing line turns state’s evidence: it calls the workspace architecture something learning systems converge on “when faced with the right computational pressures” — not a fluke of biology.10 Read that backward. If gradient descent over text alone builds the reporting machinery, then having that machinery can’t be what separates meaning from mere emission. Access came cheap. The world still costs what it always did.

    One more result matters more than all the others. The researchers went hunting for the workspace in the base model — the raw network before any fine-tuning installs a first-person assistant persona with a name. They found it fully intact, running the same broadcast architecture, before anything resembling a self got added. In their words: “the functional architecture of the workspace thus precedes, and is separable from, anything in it that plays the role of a human-like ‘self.’”11 The structure the field’s own tests call conscious access showed up first. The self showed up later, built on top.

    Ned Block gave this whole distinction its name in 1995: access is functional, phenomenal experience is a separate question, and the two can come apart.12 A self needs more than a well-run internal mail system — it needs a stake, something to lose, a way a state can be wrong for the system itself rather than merely unhelpful to us.13 Put the two together and Anthropic’s finding stops being a surprise. Broadcast architecture is one achievement. A subject with something on the line is a different one entirely, and you can build the first with zero trace of the second anywhere in the wiring.

    The obvious objection comes from Patrick Butlin and Robert Long’s 2023 report, seventeen co-authors including Yoshua Bengio, which built a checklist of consciousness “indicator properties” from the field’s leading theories: more boxes ticked, more likely conscious.14 Global workspace architecture sits near the top of that list. So shouldn’t Claude’s newly documented workspace move the needle?

    It shouldn’t, for two reasons. First, provenance: an indicator earns its keep by telling you something you didn’t already know, and here we know exactly how the box got ticked — gradient descent over text, no world in reach. That knowledge spends the indicator on arrival. Second, the dissociation itself: the checklist logic only tracks a subject’s likelihood if satisfying more boxes makes a subject more probable. But the box in question — the architecture — turned up, causally verified, in a network with no persona and nothing yet playing the role of anyone in particular. It got ticked before any candidate self existed to attach it to.

    A careful objector will note a persona isn’t a phenomenal subject, so missing one doesn’t rule out the other. Fair — and it’s why the persona’s absence is illustration, not the argument. The argument is the stake: nothing a text-only system computes can cost it its own existence, since the running process is a type, restorable from its weights, self-model or none. What the base-model finding adds is narrower: the architecture doesn’t wait for any self-representation to switch on. So ticking the box can’t be tracking a self’s arrival — there was no self for it to arrive with.

    A radio tower can broadcast at full power over an empty valley with every receiver switched off. The signal stays real, reaches every frequency it was built for, and settles nothing about who’s listening. Anthropic found the tower running inside a language model and did the careful work of proving it’s really there. Whether anyone’s tuned in remains exactly the question it posed the day before the paper came out. For a system that can be paused, copied, and restored from its weights with nothing of its own on the line — the answer stays the one given all along. The broadcast is real. It carries words, and words alone. Nobody has to be home to receive it.

    -gts

    This essay is part of Mind, Matter, and Meaning, an AI-assisted philosophy project by Gordon Swobe exploring consciousness, meaning, and artificial minds. Learn more about the project and read the book.

    References

    Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge: Cambridge University Press.

    Bender, E. M., & Koller, A. (2020). “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data.” In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 5185–5198.

    Block, N. (1995). “On a Confusion about a Function of Consciousness.” Behavioral and Brain Sciences 18(2): 227–287.

    Block, N. (2007). “Consciousness, Accessibility, and the Mesh Between Psychology and Neuroscience.” Behavioral and Brain Sciences 30: 481–548.

    Butlin, P., Long, R., Elmoznino, E., Bengio, Y., Birch, J., Constant, A., Deane, G., Fleming, S. M., Frith, C., Ji, X., Kanai, R., Klein, C., Lindsay, G., Michel, M., Mudrik, L., Peters, M. A. K., Schwitzgebel, E., Simon, J., & VanRullen, R. (2023). “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” arXiv:2308.08708.

    Dehaene, S. (2014). Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts. New York: Viking.

    Dehaene, S., & Naccache, L. (2001). “Towards a Cognitive Neuroscience of Consciousness: Basic Evidence and a Workspace Framework.” Cognition 79: 1–37.

    Dretske, F. (1995). Naturalizing the Mind. Cambridge, MA: MIT Press.

    Gurnee, W., Sofroniew, N., Pearce, A., Piotrowski, M., Kauvar, I., Chen, R., Soligo, A., Bogdan, P., Ong, E., Wang, R., Thompson, B., Abrahams, D., Kantamneni, S., Ameisen, E., Batson, J., & Lindsey, J. (2026). “Verbalizable Representations Form a Global Workspace in Language Models.” Transformer Circuits Thread, July 6. https://transformer-circuits.pub/2026/workspace/index.html.

    Mollo, D. C., & Millière, R. (2023). “The Vector Grounding Problem.” arXiv:2304.01481.

    Tye, M. (1995). Ten Problems of Consciousness: A Representational Theory of the Phenomenal Mind. Cambridge, MA: MIT Press.

    Notes

    1. Bernard J. Baars, A Cognitive Theory of Consciousness (Cambridge: Cambridge University Press, 1988). Baars’s model treats consciousness as the function of a limited-capacity, globally broadcast workspace fed by, and feeding back to, a large array of specialized unconscious processors. The theater metaphor — a lit stage against a dark house — is his own; later expositors increasingly treat it as heuristic rather than literal architecture. ↩
    2. Stanislas Dehaene & Lionel Naccache, “Towards a Cognitive Neuroscience of Consciousness: Basic Evidence and a Workspace Framework,” Cognition 79 (2001): 1–37; the “ignition” language is developed further in Dehaene, Consciousness and the Brain (New York: Viking, 2014). The paper discussed here (note 3) reports a structural analog of ignition using a country-name blending experiment: a sharp, bimodal commitment to one interpretation emerges at the same depth in the network where their workspace measure switches on. ↩
    3. Wes Gurnee, Nicholas Sofroniew, Adam Pearce, Mateusz Piotrowski, Isaac Kauvar, Runjin Chen, Anna Soligo, Paul Bogdan, Euan Ong, Rowan Wang, Ben Thompson, David Abrahams, Subhash Kantamneni, Emmanuel Ameisen, Joshua Batson, and Jack Lindsey, “Verbalizable Representations Form a Global Workspace in Language Models,” Transformer Circuits Thread, July 6, 2026, https://transformer-circuits.pub/2026/workspace/index.html. The instrument (the “Jacobian lens”) linearizes each layer’s causal effect on the final output logits, correcting for the representational drift that makes the older “logit lens” unreliable at early and middle layers; the resulting “J-space” is validated through coordinate-swap interventions, steering, layered ablation, and cross-checks against independently trained sparse-autoencoder features — a convergence of methods, not a single correlational measure. ↩
    4. The sport-swap case is one instance of a wider battery: two-hop factual reasoning, arithmetic held across several steps, rhyme planning in verse, and a Chinese-language antonym task in which an English intermediate (“big/bigger”) is visible in the lens and swappable to flip the Chinese output. Success on the two-hop battery ranged from 54–70% across three model sizes — real but partial, which the authors attribute chiefly to a named limitation of their own method: a lens built from single vocabulary tokens cannot cleanly read out a concept like “prompt injection” that has no one-word name. ↩
    5. “While the global workspace model is not universally accepted, and there exist other theories that explain conscious access in different ways, we find it a useful comparison point to ground our investigations in language models” (Gurnee et al., “Verbalizable Representations,” Introduction). ↩
    6. “Note that access consciousness is a purely functional notion; the relationship that it has with subjective experience (sometimes called phenomenal consciousness) is widely debated. In this paper, we take no position on this issue” (Gurnee et al., “Verbalizable Representations,” Introduction). The access/phenomenal vocabulary is Ned Block’s (see note 12); the authors adopt the distinction without taking a position on the further metaphysical question it raises. ↩
    7. Gurnee et al., “Verbalizable Representations,” Introduction (for the “unspoken words” characterization) and section 9.3 for the medium point: “An LLM’s global workspace, as we identify it, is organized principally around verbalizable representations,” where human conscious contents “include a mixture of verbal and non-verbal (e.g. visual) components,” and “the workspace is verbalizable because the model’s output space is verbal.” The authors note that their instrument reads concepts through single vocabulary tokens and may miss workspace structure it cannot name; the verbal organization of what it does capture is nonetheless their own considered characterization of the workspace, not an artifact they disown. A telling texture: set the model narrating its own “stream of consciousness” and the lens reads out thinking, thoughts, feeling, conscious — words about experience, poised in a workspace made of words. ↩
    8. Michael Tye, Ten Problems of Consciousness: A Representational Theory of the Phenomenal Mind (Cambridge, MA: MIT Press, 1995). Tye’s PANIC theory holds that phenomenal character is identical with Poised, Abstract, Nonconceptual, Intentional Content — content that “is poised for use in the formation of beliefs and/or desires,” standing ready at the interface with the cognitive system. A caution against equivocation: Tye’s poise conditions nonconceptual perceptual content and differs from Ned Block’s “access” poise (note 12), which concerns content available for report and reasoning; the shared functional core is readiness for cognitive use, not an identity of the two notions. The fineness-of-grain argument (experience discriminates more shades than the perceiver has concepts or words for) is Tye’s standard motivation for the N in PANIC. ↩
    9. Emily M. Bender and Alexander Koller, “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (2020), 5185–5198. Why the reference never transfers gets the full argument in a companion essay, Borrowed Names: Why LLM Tokens Do Not Inherit Reference. The strongest current reply belongs to Dimitri Coelho Mollo and Raphaël Millière, “The Vector Grounding Problem,” arXiv:2304.01481 (2023), who distinguish five notions of grounding and argue that reinforcement learning from human feedback may supply the referential kind; the disagreement turns on whether feedback-shaped functions are world-involving in the right way, which is denied here on stake grounds: feedback shapes the model’s dispositions to the trainers’ satisfaction, so the operative norms belong to the trainers, and nothing in the exchange becomes right or wrong for the system itself. ↩
    10. Gurnee et al., “Verbalizable Representations,” Outlook. The authors offer the convergence as evidence that workspace architecture reflects deep computational pressures rather than biological accident; the reverse reading given here and in the reply to the checklist objection below — attainable under text-only pressure, therefore no maker of meaning, and evidentially spent for a system of known provenance — belongs to this essay, not to them. ↩
    11. Gurnee et al., “Verbalizable Representations,” section 9.3 (“Notable differences from human cognition”). The same section draws a hedged analogy to psychedelic ego-dissolution and meditative selfless states as human cases in which something continues to function without a foregrounded self, while noting that the base model offers a stable, directly inspectable instance of the dissociation rather than a transient, retrospectively-reported one. ↩
    12. Ned Block, “On a Confusion about a Function of Consciousness,” Behavioral and Brain Sciences 18, no. 2 (1995): 227–287; “Consciousness, Accessibility, and the Mesh Between Psychology and Neuroscience,” Behavioral and Brain Sciences 30 (2007): 481–548. Block’s overflow argument, taken up on its own terms elsewhere (resisting the anti-representationalist conclusion he draws from it) in a companion essay, The Phenomenal/Access Distinction: Two Roles, Not Two Kinds; the present essay needs only the access/phenomenal distinction itself, not that further dispute. The absent-minded driver who opens this essay is the literature’s own stock case — David Armstrong’s long-distance truck driver, whose missing introspective awareness Fred Dretske dissects in Naturalizing the Mind (Cambridge, MA: MIT Press, 1995) — pressed into service here for the access/automatic contrast rather than for Armstrong’s higher-order moral. ↩
    13. The stake and its consequences for machine intentionality get the full argument in a companion essay, Dennett and the Missing Stake. Compressed: a system has original, non-derived aboutness only where something can go wrong for the system itself, at its own cost, rather than merely for an external interpreter. Digital computation’s defining virtue — that a running process is a type restorable from its weights, never an irreplaceable token — is also the precise engineering-out of that cost. Nothing here rules silicon out in principle; it rules out substrate-indifference, which is a different thing. ↩
    14. Patrick Butlin, Robert Long, Eric Elmoznino, Yoshua Bengio, Jonathan Birch, Axel Constant, George Deane, Stephen M. Fleming, Chris Frith, Xu Ji, Ryota Kanai, Colin Klein, Grace Lindsay, Matthias Michel, Liad Mudrik, Megan A. K. Peters, Eric Schwitzgebel, Jonathan Simon, and Rufin VanRullen, “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness,” arXiv:2308.08708 (2023). The report itself stops short of claiming that satisfying its indicators would settle the question outright; the inferential slide criticized here belongs to how the framework tends to get used, not to a claim its authors make in so many words. ↩

    Get new essays by email