What Mary Could Not Build

Torin Alter’s defence of the knowledge argument is right about more than its critics allow. It goes wrong about what a concept of an experience is a concept of, and a case he sets aside shows where

A woman has seen red. She is now back in a black-and-white room, with a complete physical description of a creature from another planet and one question to answer. When the creature looks at a ripe tomato, is the state it is in the one she knows as red?

Torin Alter thinks she cannot work it out. He calls this a strong intuition, and he has put the knowledge argument on it. I think she can work it out, given one thing, and this essay is about that thing.

The case is a repair of a more famous one. Frank Jackson’s Mary has learned all the physics of color vision from black-and-white lessons and has never seen a color. When she leaves her room and sees a tomato she learns something, the argument goes, so the physical truths were not all the truths. The standard reply is that Mary lacked a concept, and nobody can deduce a truth they lack the concept to state. Alter accepts the reply. One cannot expect Mary to derive the concept of phenomenal redness from structural concepts, he says, “and nothing interesting follows from that.” So he gives her the concept and asks about someone else’s experience. David Chalmers makes the same move: “even when Mary has the relevant phenomenal concepts, she cannot always deduce the relevant phenomenal truths (about other people, for example)” (“The Two-Dimensional Argument Against Materialism”).

Alter’s book, The Matter of Consciousness (2023), is the most careful defence of the argument there is. Its conclusion is that physicalism is false, or else that the world has natures which its structure leaves out. What the argument refutes, he says, is “a narrow kind of physicalism … which says that … the entire world including consciousness is fully structural.” That is the view I defended in What a Description Answers To. So his argument is aimed at me, and it deserves an answer that does not dodge.

The answer is unusual in how much it grants. I accept the principle that carries his argument from what Mary cannot deduce to what is not necessary. I accept that Mary learns something. I dispute one premise, about what a concept of an experience is a concept of. A case that Alter himself states and sets aside is the test of it.

The argument

Alter gives the argument three steps (section 1.4 of the book).

The first step is an epistemic gap. Mary learns the complete physical truth from her lessons. She can deduce whatever follows from it. And on leaving the room she learns new truths about what seeing in color is like. So some truths about experience cannot be deduced from the complete physical truth.

His explanation of the first step is what makes the book distinctive. What Mary learns in the room is structural: “expressible by statements containing only logical, mathematical, nomic, and perhaps spatiotemporal vocabulary” (2.2). Such information is objective because it is structural. “Sufficiently intelligent Martians could understand such information.” What she learns on release “cannot be deduced from structural truths but can be learned by having relevant experiences” (2.1). He takes from David Lewis a point that supports this. Lessons about non-physical things would not help Mary either, “if those further lessons are also conveyed in black and white.” So the contrast that matters is between structure and something else, and not between the physical and the non-physical.

The second step is a modal gap. If the phenomenal truths cannot be deduced from the physical ones, they are not necessitated by them either. Alter’s ground for the step is that there are no primitive necessities. A primitive necessity is “an a posteriori necessity that is not necessary in virtue of following a priori from a priori truths and non-modal truths about the actual world” (7.2). His model of an honest necessity is Euclid’s: being a triangle necessitates the angle sum, and the one can be deduced from the other.

The third step is an ontological gap. If the physical truth does not necessitate the phenomenal truth, physicalism is false. The alternative he favours is Russellian monism, on which physics describes structure and something with an intrinsic nature underlies it.

He claims no more for the argument than the balance of evidence. “When I describe the knowledge argument as sound, I do not mean to imply that one cannot reasonably doubt its soundness” (6.4.2). I argue to the same standard.

What I grant

The world is structural, in a strong sense. On my account what a description answers to is production, one state coming from another. Production is not a nature hidden under structure. I do not know which rule, if any, generates our physics, and I do not think anyone could be given it. The claim is about kind: whatever supports physics is of the kind a rule generates. So I cannot escape Alter by finding something non-structural at the bottom.

Mary learns something. The Grain of the Hard Problem said so. An ideal reasoner with the complete physics “can derive the organization at the grain of experience. Chalmers will say that it still cannot derive what that organization is like. We agree.”

Some truths are not structural. They are the truths stated with concepts that are applied by recognition. What Phenomenal Intuition Gets Right put it this way: “structure cannot deliver like this, because like this is available only to a system that has undergone the occurrence.” All truths are reducible, and every reduction is structural. That is not the same as every truth being stated in structural terms.

There are no primitive necessities. The previous essay listed brute necessities among the six things its view rules out.

Introspection does not reveal how an experience is constituted. Alter says introspection is not “equipped with a constitution detector” (8.2.3). The earlier essay said that “the constitutive complexity is not presented in recognizing an experiential character.”

Having an experience is not the only way to know what it is like. Alter rejects that doctrine, with Hume’s subject who can supply a missing shade of blue (4.1). Grain used the same case.

One consequence has to be faced at once. What Phenomenal Intuition Gets Right replied to Chalmers that his objection “assumes that phenomenal concepts reveal the full nature of their referents.” Alter disowns that assumption, and argues at length that his principle stands without it (8.2). So the published reply does not reach him. A better one is needed.

What a necessity needs

Alter does not deny that some necessities are known only by investigation. Water is necessarily H2O, and nobody knew it before chemistry. His point is that such necessities are not primitive. “If the underlying explanatory structure of water is H2O, then necessarily water is H2O” is known a priori, as part of having the concept of water. The antecedent is a plain fact about the actual world. The necessity follows from the two. He adopts Chalmers’s general statement: “For every a posteriori metaphysical necessity ‘Necessarily S’, there is a nonmodal truth D such that ‘If D, then necessarily S’ is a priori” (7.2.3).

Jackson draws the consequence for deduction, and Alter follows him (1.4.2). From the fact that a mug holds water one cannot deduce that it holds hydrogen. Add that water is the watery stuff around here, and that the watery stuff is H2O, and one can. Kripke’s cases show necessity without deducibility only when the description is partial. Mary’s description is complete.

I accept all of this. It tells me what I owe. If the identity of an experience with an organization is to be a necessity of the honest kind, there must be a conditional known a priori and a non-modal fact that satisfies it. Grain claimed that there is: the necessity “is not brute. It follows from an identity together with a discovered ordering of grains.” Spelled out, the conditional is this. If what my recognition of red picks out is a certain region of organizations, at a certain grain, then necessarily an experience of that kind is an occurrence in that region. And the fact is that it does pick out that region.

Whether the conditional is known a priori depends on what the concept is a concept of. Suppose the concept is of a nature that is fully apparent to me when I see red. Then the conditional is false, because nothing about organization is part of what is apparent. Suppose the concept is of whatever is present when my recognition fires. Then the conditional holds in the way the one about water does.

So the dispute is not about the bridge from deduction to necessity. Alter and I both accept it. The dispute is about one concept.

Fool’s orange

Alter’s picture of an experience has two layers. He distinguishes two principles of Revelation (8.2.2). On the strong one, the complete nature of a type of experience is apparent to the one who has it, including “whether and how that type is (actually) constituted.” He rejects that. On the weak one, the complete nature is apparent, and that nature “does not concern whether and how that type is (actually) constituted.” He says the weak one would suffice for what Goff wants explained, and that it is compatible with experiences being structurally constituted. On the picture that results, an experience has a phenomenal nature, which is apparent, and a constitution, which is not.

He then considers an argument that Philip Goff might give for the strong principle, and it contains the case I want. If experiences were structurally constituted, there could be an experience that “feels exactly the same” as an orange one and is not orange, because it is differently constituted. Call it pseudo-orange. Then nobody could be confident that their own experience is orange and not pseudo-orange. Since we can be confident, experiences are not structurally constituted, or else their constitution is revealed.

Alter’s reply is that the physicalist need not allow pseudo-orange at all. Standard physicalism “is perfectly compatible with counting any O experience—including an O experience that is not structured as yours is—as a phenomenally orange experience” (8.2.3). Whatever feels that way is orange, however it is made.

That reply treats “orange” as unlike “water.” Water is not whatever looks and tastes like water. The stuff on Putnam’s Twin Earth looks and tastes like water and is not water, because it is not H2O. Alter says in conversation that he is “disinclined to think” that terms like “seeing red” are natural-kind terms of the Putnam–Kripke sort. I think they are, with one qualification, and I take the horn he sets aside. There could be pseudo-orange. I will call it fool’s orange.

The phrase “feels exactly the same” has two readings, and the earlier essays keep them apart. Two experiences can be the same to recognition: the subject cannot tell them apart. Or they can be the same in what is present. The first does not give the second. “Two occurrences that fall under the same type, and so share a matching profile, need not share exact character,” Grain said. Exact character “is not known in the sense of being represented. It is present.”

With that distinction, two experiences can differ in how they are constituted in three ways.

They can differ below the grain of experience, as neurons differ from silicon. That makes no difference to the experience. “Their material may vary, but the constitutive dynamics must actually be present.” Alter’s reply fits this case, and here he is right.

They can differ at the grain of experience and below the grain of recognition, in one kind of mind. The subject cannot tell them apart, and they differ in exact character. Both are orange.

Or only the cognitive role is the same. A system makes the same judgments and the same discriminations, and its organization at the grain of experience is different, or is absent. Grain argued that this can happen: “A functional description, specified at the grain of cognition, really does underdetermine experience.” A duplicate of our discriminative organization that keeps some regulated carrier “would plausibly have experiences of the same types, differing in exact character. One that dissolves the carrier, as some unfoldings may, would be a candidate for having none.”

The third case is fool’s orange. It has the presentation without the organization, and it is not orange. The second case is to orange as heavy water is to water.

What “orange” names, then, is a region: the organizations that this recognition tracks. This is the overflow of experience beyond cognition, seen from the side of the concept. Recognition is coarser than what it responds to, so one recognition answers to many occurrences.

Here the analogy with water stops, and the essay should say so. H2O draws water’s boundary with no help from how water looks. Organizations vary continuously, and the boundary of orange is drawn by recognition’s own lumping. The kind is natural in what it is a region of. Where the region ends is relative to the recognizer.

Why read the concept this way and not Alter’s? Because of how we treat it when recognition and presence come apart. Two shades may be indistinguishable when seen separately and distinct when they share an edge. We do not say that they were the same and became different. We say they differed all along. We defer to what was present over what recognition delivered, and that is how one treats a term for a kind.

Goff’s worry remains: how can I be confident that my experience is orange? In the way I am confident that the glass holds water. Recognition is reliable in the conditions it was built for. Judgments at its grain, about kind and similarity, are mostly right. Exact character is present without being known.

Alter has an argument that phenomenal concepts are not blind pointers (8.2.1). Mary’s belief that phenomenal redness is this quality is informative, so the concept is not a bare demonstrative. I agree. The concept gives access to an experience through its actual character, and gets its presence and its distinctness from other experiences right. It is not opaque. It is not transparent either, since it does not present the grain or the complexity of what it picks out. Alter grants that a “semantic explanation for the epistemic gap … would undermine the inference to the modal gap.” He considers one such explanation, radical opacity, and rejects it. The concept I have described is of a middle kind, and it gives such an explanation.

The apparent nature is not the complete nature

The last section set two pictures side by side. This one argues for mine.

Take a series of shades of blue in which each is indiscriminable from the next and the two ends are plainly different. The experiences at the ends differ in character. So somewhere along the series, experiences that the subject cannot tell apart differ in character. Part of the character of each is therefore not apparent to the one who has it. The weak principle of Revelation says that the complete phenomenal nature is apparent. It is false, and for a reason that has nothing to do with constitution.

Alter’s own example makes the point. “It can be apparent to the wine novice that the taste experiences of two wines are similar, while the complexities of the tastes are not apparent to them” (8.2.3). He offers it to show that confidence about similarity does not reveal a complete nature. It shows more. What the novice misses and the expert gets is in the taste. It is character, and not something beneath character. The line between what is apparent and what is hidden does not run between a phenomenal nature and its constitution. It runs through the nature.

Once that is granted, there is no layer left for a fully apparent nature to occupy. There is the occurrence, and recognition presents it coarsely.

A second piece of evidence is one Alter reports (4.3). Paul Churchland predicted, from a model of how opponent cells in the visual system fatigue, that people would see afterimages in colors that no surface has, among them an orange more saturated than any orange object. People then saw them. Alter’s interest is in the surprise that remains: that someone who knows the theory “will be surprised at all when she sees that color indicates that she was not able to do the relevant deduction.” My interest is in the prediction. The structure of the vehicles told us of a character before anyone had met it.

Alter anticipates a reply of this general kind and has an answer (6.4). Derk Pereboom suggested that introspection may misrepresent experience, as perception misrepresents color on Locke’s view. Alter says that the physicalist needs more than the possibility. The inaccuracy must be actual, and it must be radical, since “if any aspect of the nature of phenomenal redness is as introspection represents that nature as being, then knowledge argument proponents could argue that upon release Mary gains information about that aspect.”

My claim is not Pereboom’s, and it meets the demand differently. Introspection is largely right. It is right that an experience is present, that it has a character, and that it differs from and resembles other experiences. It makes one mistake: “because the constitutive complexity is not presented in recognizing an experiential character, the character is taken to have no constitutive complexity.” The aspects it gets right are structural, and Mary could have derived them. “A complete description of the vehicles entails that an experience of red is more like an experience of orange than one of green.” The one aspect that would not be structural is simplicity: a nature with no composition, exhausted by my access to it. That is the aspect introspection gets wrong. A single inaccuracy, placed exactly where the non-structural information would have to be, does the work that Alter says needs a radical one.

He asks for evidence that the inaccuracy is actual. The series of shades is evidence. The predicted colors are evidence. And there is an explanation of why a system would make the mistake: a complex activity that is available to it as a unit will be taken for simple, because the unity of the access is read as the simplicity of what is accessed.

He also says why the idea is hard to accept. A Lockean about color keeps the feeling of color and moves it into the mind. The parallel view about the feeling itself has nowhere to move it. The earlier essay answered this. Vision presents surface colors as simple and homogeneous. The early modern relocation “transfers a feature of represented content to the representing occurrence.” But “an experience of a homogeneous red surface need not be homogeneous or simple.” Denying simplicity feels like denying the feeling because the feeling was first described by borrowing the look of a surface.

Others have argued that nothing in introspection or in theory shows experience to be non-structural, Daniel Stoljar among them (“Four Kinds of Russellian Monism”, 2014). The nearest published relative of the positive claim is Kristjan Loorits’s, that consciousness is “a complex neural pattern that misperceives some of its own highly complex structural properties as monadic and qualitative” (“Structural Qualia”, 2014). What I add is a grain at which the structure lives, and a mistake confined to one aspect.

What she can deduce

Return to the woman in the room. Her reasoning has Jackson’s form.

The kind I call red is what this recognition of mine picks out. What it picks out is a region of organizations at a certain grain. That is a physical truth about her own brain, and it is in her books. The alien’s state lies in that region, or it does not. That is a physical truth about the alien, and it is in her books too. So the alien’s state is red, or it is not.

The first premise carries the account I have been defending. Without it, what her recognition picks out might be a nature that merely goes along with the region. So I do not claim that her conclusion follows from the physics and her concept alone. It follows given that an experience is the organized occurrence. Can Two Minds Share One Experience? put the identity “at the level of a concrete dynamical organization.” Sameness of kind is then sameness of organization, and the physics settles that.

Alter’s intuition that she cannot tell is strong for a reason. She cannot check her conclusion. She has never had the alien’s state beside her own, and being human she never will. The deduction gives her a truth and no new recognition. What she feels is the absence of that recognition. His intuition, stated as a claim, is that the account is false, and the argument against that claim is the one in the last section.

Some cases come out as one would hope. If the alien’s space of qualities has no counterpart of her red, the answer is no, and she deduces it. If the alien’s organization lies near the edge of her region, the answer is as vague as the edge, which her own recognition draws. If the creature is a bat, she can deduce that its states are of no kind she has a concept for.

What Mary lacked

Before her release Mary lacked the concept, and the question is why her lessons could not give it to her.

A concept applied by recognition gets its way of presenting its object from undergoing the object. Grain said why Mary cannot supply that for herself. “Our power to write into our own vehicles is coarse. Imagery passes through the same bottleneck as access, in the other direction. We can interpolate within quality spaces we already have, and cannot add a dimension to them. That is why Mary is stuck.” Hume’s subject can fill in a shade of blue between two shades he has seen. Mary has no dimension to interpolate along.

This puts me at odds with a premise Alter states outright. A textbook of the complete physical truth, he says, “could be written in a highly mathematical, objective scientific language, which could be understood by any sufficiently intelligent creature, regardless of their sensory capacities,” and that is “exactly the sort of book pre-release Mary could fully understand” (1.4.1). I think Mary can state every truth in the book and derive every truth that follows from one. That is less than understanding all of it. To understand the description of an organization, at the level of the organization, is to be able to recognize an instance. Acquiring understanding means reconfiguring whatever does one’s representing. For most organizations an instance can be built on paper or in imagination, and recognized there. For an experience the recognizing is done by the system that has it, so the instance has to be built in that system’s own vehicles.

Mary’s case is the far end of something ordinary. At every level of description above the lowest, some truths are stated with concepts that are applied by recognition and do not say what structure they pick out. We recognize a stone, or a glider in the Game of Life, without the concept telling us what it is made of. Nobody infers a gap in the world from that. The truths are reached by building an instance from the lower-level description and recognizing it under the upper level’s types.

Physics is the hard case for this claim, since physics is the science accused of being purely structural. It passes, and Alter’s own text shows two reasons.

The first is an old problem that he grants. A theory in which every term is defined by its role is satisfied by any collection of things of the right number, so a purely structural physics would say nothing about the world. To avoid this he keeps the nomic concepts primitive, and does not define them by role (2.3). The physics his argument needs is therefore not purely structural, by his own construction.

The second is how physics is applied. A track in a detector is recognized as an electron’s. The concept electron does not say what lies beneath physics, as red does not say what organization it picks out. To meet an objection of Stoljar’s, that spatial and temporal concepts are tied to experience, Alter has to replace the physical concepts with counterparts defined through Ramsey sentences (2.3). The physics Mary is taught is structural only after its concepts have been stripped of what recognition contributes.

Physics seems purely structural for a reason, and the reason is that its recognizers can be swapped. Any instrument and any observer will do. That is the truth in the remark about Martians. For red only one kind of recognizer will do, the system that has the experience. Objectivity is the degree to which the recognizer can be replaced, and Mary’s case is at one end of the scale.

The claim has to be kept this specific. That a description of an experience does not reproduce the experience is, as the earlier essay said, “true but weak”: “Reading a description of viscosity does not make the page viscous either.” What explains Mary is an architecture “that can recognize a concrete difference without representing the complexity constituting it.”

Alter calls his own explanation of what Mary learns abductive, so it can be met with a rival. Mary learns something on release because the concept she lacked can be had only by a system in which the organization occurs, and nothing in her could bring that about from a description. Lewis’s point is predicted by this explanation at least as well as by his. Lessons about non-physical things would not help Mary, because a lesson in any vocabulary is a description.

Two things Alter says along the way show where the rival explanation bites. He stipulates that in the room Mary “does not imagine in color” (1.4.1). And in answering the objection that she lacks concepts, he says that if those concepts can be physically explained, her teachers could give her the explanation, and “that should enable her to construct the relevant concepts” (9.4.3). For a human the second is false, and so the first is not needed. For a reasoner who could do what the second describes, the first removes the ability by fiat.

A reasoner who can build

Grain described such a reasoner. It “reads the physical description, derives the organization at the grain of experience, configures its own vehicles to match, and brings them into its regulation. Then it undergoes and recognizes the result.” It needs a path from its reasoning into its experiential carrier, a carrier general enough to realize the target, and regulation that takes the result in. Nothing in principle rules these out for an engineered mind. For such a reasoner Mary’s lessons would be enough.

The obvious objection is that this is cheating, as it would be if Mary were shown a photograph of a tomato. Grain gave a criterion. “A technology cheats when it causes the experience independently of what the reasoner knows.” The photograph cheats. Anyone who looked would have the experience, and garbling Mary’s physics would change nothing. The constructor’s route runs through its grasp of the description, “so a garbled description would yield a different result.”

The criterion needs a second clause. Daniel Dennett imagined a robot, Locked RoboMary, whose color system is disabled. As Alter quotes him, she “builds a model of herself,” works out the state B that the model enters on seeing a tomato, and “puts herself into state B” (4.2.1). Her route runs through what she knows. But the fact that B is the red state reaches her as “the state a tomato causes.” That is a label. She does not know why this configuration should be red. A genuine constructor knows why: what the organization is more like and less like, and what it opposes. A route cheats when the information about which experience is being built arrives other than structurally.

In the terms of the previous essay, a label is a lookup table. The cheating constructor stands to the genuine one as the dust man’s table stands to an admissible map.

It follows that a genuine constructor is possible only if red is structural. If red were an atom of quality there would be no why, only labels, and every route would cheat. So “could a reasoner build red without cheating?” and “is red structural?” are one question.

Alter’s verdict on the robot is that she “does not deduce phenomenal truths from physical truths. Instead, she puts herself in a dispositional state that constitutes possessing the relevant phenomenal knowledge. This is hardly better than if she had simply unlocked her color-vision system” (4.2.3). The strongest form of the objection is this. That a reasoner can produce the experience by building the organization shows only that the organization suffices here. A dualist with laws linking brain to experience predicts the same success. Deduction would show that the phenomenal truth follows from what the physical truth says.

I agree that production is not deduction, and I do not rest the necessity on it. The necessity is derived, as section three required and section six carried out. Construction does a different job. It is how a reasoner who lacks the concept gets it from the description.

It also does a second job, which matters more. The human who has seen red cannot check her conclusion about the alien. The constructor can. It builds the alien’s organization, binds it beside its own red in one field, and recognizes. If it built two organizations alike and could tell them apart when bound, the account would be wrong. The account is open to an experiment, and that is more than can be said for an intrinsic nature that no description reaches.

Alter illustrates what he will not count with two geometers (4.2.4). One proves that the angles of a trapezoid sum to 360 degrees. The other uses a device that puts anyone who contemplates the axioms into the state of knowing the sum: “that process is plainly not deduction.” The device fails Grain’s test, since a garbled axiom would make no difference. Amy Kind adds a third geometer, who needs a diagram (“What Counts as Cheating?”, 2024). If Euclid’s diagrams are indispensable his proofs are not pure deduction, and using one is no foul. The constructor belongs with her. Grain compared its method to what Kant called the construction of a concept in intuition.

Alter concedes more than his verdict suggests. “Perhaps Mary could in some way exploit her physical knowledge to acquire phenomenal knowledge of color experiences without having color experiences. Dennett’s RoboMary case indicates that such a thing is possible” (4.5).

Is what the constructor knows a priori? Chalmers’s own account of the a priori idealizes the reasoner: “If I cannot know that P independent of experience, but another less limited being could do so, then it is a priori that P” (“Does Conceivability Entail Possibility?”). The constructor is a less limited being. His ideal reasoner, as Grain said, “is idealized in capacity but not in architecture.” Idealize the architecture as well. Then what the constructor undergoes supplies a concept, and the conditional from the physics to the phenomenal truth is justified by reasoning. On the usual view, experience that only supplies a concept does not make knowledge a posteriori. I would not fight for the word. The account itself, that an experience is the organized occurrence, is a theory, and I do not claim it a priori.

What is taken on trust

The account generalizes across experiences and across experiencers only weakly, and the essay should say how weakly.

The weakness begins inside one subject. Since blue is structural and not an atom, experiences I cannot tell apart may differ. When I take today’s blue to be yesterday’s, I trust that they do not differ in some way I fail to notice across the two occasions. Across subjects the trust is larger. The woman in the room concludes that the alien’s state is red as hers is because their organizations fall in one region. That the same organization is the same experience is the account. She takes it on trust, and for her it is a leap.

A constructor is better placed, in two ways. It can bind any two experiences it chooses beside each other, so whatever can be told apart side by side is told apart. And it knows the organization it built. It shares our limit on what can be known by acquaintance, since for any mind that knowledge is bounded by recognition. But it knows, from the account, whether two shades that look the same are the same experience. What it still takes on trust concerns differences that no binding reveals, which make no difference it could recognize. Its leap is across an insignificant gap. Ours can be across a large one.

One caveat. Grain noted that “each comparison supplies a new context, and the context is part of what fixes exact character.” Binding two blues makes a new context. The constructor built the joint organization, though, so it knows what the binding preserved in each part.

Alter says that his argument can reasonably be doubted. Mine can too, at this point. I think the gap on my side is the smaller one, for a reason the section on Russellian monism gives.

Zombies, and the usual strategy

Physicalists who accept that Mary learns something usually explain it by a feature of phenomenal concepts. The concepts are recognitional, or they contain their referents as a quotation contains a word, and this isolates them from physical concepts. Alter spends a chapter on the strategy (chapter 9) and turns Chalmers’s dilemma against it. Let C be the claim that we have concepts with the relevant feature, stated without presupposing experience. Either a physical duplicate of us without C is conceivable, or it is not. If it is, C is not physically explicable. If it is not, zombies have C. But “zombies do not share our epistemic situation.” So C does not explain ours.

The dilemma is addressed to someone who grants that zombies are conceivable on ideal reflection. I do not grant it. A zombie, a physical duplicate of a person with no experience, is conceivable only at first look, and only by forgetting the grain of physics. Grain traced the forgetting. A description of what a system does, at the grain of cognition, really does leave its experience open. The zombie intuition “extends the verdict from functional descriptions to physics,” because “our frame presents functional and physical descriptions alike as third-person structure.” The two differ in grain, “and the frame has no way to register the difference.” A duplicate at the grain of cognition may differ in experience or lack it. That is fool’s orange. A duplicate at the grain of physics cannot.

This rests on a fact that might have been otherwise for all we could tell in advance: our physics is finer than the grain of our experience. Grain allowed observers whose physics is coarser than their own experience. I now doubt that such creatures are possible. It is a guess, and it goes by mechanism. If evolution can make an organization do work, measuring devices can reach that organization too. Nothing in this essay depends on the guess.

The earlier essays already stood apart from the usual strategy. “The usual strategy treats the isolation of phenomenal concepts as a basic feature of how they work. Here it follows from the overflow.” And they said of the usual strategy what Chalmers’s second horn says: on it, nothing in the world answers to what the physical story seems to leave out.

Alter’s test for an explanation of our situation is that “a sufficiently capable reasoner who understands C would be well-positioned to understand our epistemic situation” (9.4.1). The grain account passes it. A concept that closed the gap “would have to specify the grain of its referent, as physical concepts do, while bearing its mode of presentation, as recognitional concepts do,” and recognition reaches its object only at the grain of discrimination. A reasoner who read that could predict that Mary learns something, that zombies seem conceivable, and that the seeming is partly right, since a functional description does leave experience open.

Two positions that Alter discusses are close to mine. Jesse Prinz: “[m]aintaining availability is an actual neural process, and Mary cannot undergo that neural process, or any other, simply by reading about it.” Robert Howell: phenomenal knowledge “requires actually having the phenomenal states in question.” Alter answers both with the dilemma. I differ from both in what I say of the woman who already has the concept. Given the account, she can deduce.

That difference bears on a label. Chalmers sorts physicalists into those who deny that there is an epistemic gap on ideal reflection, type A, and those who accept the gap and deny that anything follows about the world, type B. What Phenomenal Intuition Gets Right said that the position “is type B,” because a gap remains. I would now put it differently. For a human the identity is known only by investigation, Mary learns, and the gap stays open. For a reasoner with the right architecture, and given the account, there is no gap. The taxonomy idealizes the reasoner, so by its lights the position leans to type A. It is a fine-grained one, in the sense in which the earlier essay admitted to a fine-grained functionalism: the entailment runs through organization at the grain of experience, and not through cognitive function. And it has a residue that no type-A view admits, which is the trust described above. I would sooner state those three things than choose a letter.

Alter also presses a point of Jackson’s and Chalmers’s (7.1). When one learns an old truth under a new guise, one generally learns a new truth as well, about the guise. My reply is that the guise is itself an organization, coarser than what it presents, and that the truths about it are structural truths Mary already had. The observer who knows all the physics and “may have no idea” what it is like for Mary lacks the concept.

Possible and conceivable

Behind the bridge principle stands a view of possibility. For Chalmers the basic modality is tied to “rational consistency, apriority, or conceivability,” and “to introduce a further primitive, restricting the space of worlds, is to introduce an unnecessary wheel” (as Alter quotes him, 7.2.5). What is coherent to describe is possible, once the description is ideal.

I hold that a description does not make a possibility. A situation is possible when a support for it could be rebuilt, level by level, down to what generates. I mean this metaphysically: a metaphorical God could do the rebuilding. To conceive a situation is to imagine the rebuilding in a toy model. The toy model is a proxy. It is good where it keeps what matters, and it fails where the imagining skips a step. The zombie is conceived by skipping one. To imagine a physical duplicate in the full detail of the physics is to specify the organization at the grain of experience, and then there is nothing left to subtract.

Chalmers will say that restricting the possible to what can be built is his unnecessary wheel. It is not a further primitive. On the account of the previous essay, modal facts at each level are constructed from the level below, and at the bottom there are no counterfactuals at all, only generation. That removes a modal primitive. His own definition of ideal conceivability already leans on possibility: a statement is ideally conceivable “when there is a possible subject for whom S is prima facie conceivable, with justification that is undefeatable by better reasoning.” He grants that the notion of possibility enters there, in what he calls a roundabout way. Which subjects are possible is then part of the question. On his view they are those that can be coherently described. On mine they are those that could be built.

I do not claim that this meets Alter’s principle in general. His principle wants every necessity to follow a priori from non-modal truths about the actual world. What I can claim in general is narrower: no necessity is brute. For the one necessity this essay needs, I have given the a priori conditional and the fact.

Both sides have a picture with God in it. Alter describes physicalism as the view that once God fixed the physical truth there was no further work to do (1.4.3). On the account here God builds from below, and in building the physics of a person has already built the organization at the grain of that person’s experience.

Russellian monism

Russellian monism has three parts (Alter and Pereboom). Physics describes the world only in its structure and dynamics. There are quiddities, properties that underlie that structure. And consciousness is constituted at least in part by them. Alter reaches the view from the knowledge argument: if the argument is sound there are structural and non-structural properties, and a theory is likelier to be true if it gives both kinds a part in causation.

The quiddity is the second layer of section four, given a place in the world’s furniture. Grain said where the idea of it comes from. The one alleged instance of a nature that is not structural is our own experience. It seems so because a structured occurrence is recognized without its structure, and the recognition adds the clause and that is all. “On our account, then, the paradigm of an intrinsic nature is structure below the grain of recognition. The Russellian takes the one place where structure hides from us as evidence of something that is not structure.” The two views also put experience on opposite sides of physics. For the Russellian it lies beneath physical structure. For me “physics is structure at one grain and experience is structure at another,” and the grain of experience is the coarser.

Four arguments follow from points already made.

The Russellian’s Mary is no better off. Alter says the view is reductionist: given the truths about quiddities as well, Mary could deduce all the facts about consciousness. By Lewis’s point, which Alter endorses, lessons about quiddities conveyed in black and white would not help her. So the truths about quiddities that matter can be grasped only by acquaintance, and his Mary needs exactly what mine does. Pereboom made the neighbouring point about powers: “Wouldn’t Mary be in a position to fully understand such powers while she is still in the room?”

The advantage is one he half withdraws. Asked why the knowledge argument cannot be run again against quiddities, he says that “by hypothesis we don’t understand these properties.” Richard Brown replies that the physicalist can say the same about a future physical theory. Alter allows that “there may be parity,” and that his advantage “might be a Pyrrhic victory.” I think the sides are not on a par. The structural side has an account of why the gap appears and of what would close it. His side has a placeholder, which he describes himself as “a big mystery at the center.”

There is an unknown beneath physics on my side too, and it does different work. I decline to say what lies beneath the lowest level we could reach. A critic might say I have quiddities of my own. The difference is in what the unknown is asked to do. The Russellian’s unknown is a nature, and experience is made of it. Mine is more structure of the same kind, and experience does not depend on which structure it is. For humans and for the machines we build, the grain of experience lies far above the grain of physics.

Quiddities are hard to fit into causation. If the tie between a quiddity and the dispositions it underlies is contingent, the quiddity is, in Howell’s phrase, “just along for the causal ride” (“The Russellian Monist’s Problems with Mental Causation”, 2015). Alter wants quiddities integrated into causation, and I have not yet read the chapter where he argues it.

There is also the question the Russellian has to answer in the end: how small quiddities combine into the experience of a person. An answer would be an account of organization from the small to the large. That is a structural account, and it is the kind the Russellian said was not enough.

Against structuralism itself Alter and Pereboom give one sentence. Views on which physics involves structure all the way down “seem to yield a world devoid of substance or qualities.” The seeming is the one Grain explained.

What would change my mind

The room

The woman in the room can work out the answer. She cannot feel that it is right, because she cannot have the creature’s state beside her own. What she takes on trust is that the same organization is the same experience, and a mind that could build the creature’s state in itself would have less to take on trust.

Alter and I share two things that most of his critics deny. Mary learns something. And nothing is necessary without a ground. What we disagree about is whether a concept of an experience reaches a nature that is all there on its surface. The shades of blue say it does not.

Draft notes (remove before publishing)

Draft of 17 Nov 2026, from the outline in research/structural-realism-second-essay-outline.md. Not committed, not in the index.

To verify before publishing:

  1. Every quotation from Alter’s conversation with Richard Brown (Consciousness Live!) comes from an auto-generated transcript and must be checked against the recording: “nothing interesting follows from that”; “a narrow kind of physicalism …”; “disinclined to think”; “by hypothesis we don’t understand these properties”; “there may be parity”; “might be a Pyrrhic victory”; “a big mystery at the center”. Add a link to the recording.
  2. “Experienced Mary” does not occur in the sections of the book that were read. The opening rests on the conversation and on Chalmers’s sentence. Find where the book has the case (chapter 5 or section 7.3 are likely) or say that it is from the conversation.
  3. Quotations from the book are from pasted text and carry section numbers only. Page numbers are needed.
  4. The attribution of the pseudo-orange argument: Alter states it on Goff’s behalf. Check that he presents it so, and Goff’s own text. Check also how far Alter endorses the weak principle of Revelation; the essay says only that he holds it compatible with structural constitution, and then argues against the two-layer picture it gives.
  5. Dennett’s “What RoboMary Knows” is quoted only as Alter quotes it. The reading that state B reaches the robot as a label is Claude’s, from that quotation.
  6. Churchland’s chimerical colors: what was predicted, what was observed, and the reference. The essay says “people then saw them”.
  7. Putnam’s Twin Earth, Lewis’s lesson about black-and-white instruction in non-physical subjects, and Jackson’s two papers: references.
  8. Chalmers’s dilemma is given as Alter reports it. Check it against “Phenomenal Concepts and the Explanatory Gap”.
  9. The “usual view” that experience which only supplies a concept does not make knowledge a posteriori: a reference, and whether Chalmers accepts it.
  10. Alter’s chapter 15 on causal integration is unread, as are chapters 3, 5, 10–14 and sections 7.3 and 8.1.
  11. The trapezoid: the essay says the angles sum to 360 degrees; check Alter’s wording of the example.
  12. Loorits’s first name and the Stoljar reference.

Confirmed by the author on 17 Nov: the argument from the graded series of shades against weak Revelation, and with it the conclusion that there is no second layer.

Claims the author has not confirmed in these words:

Left out, from the outline: the hypercomputer as a test case for the two notions of possibility; Sellars’s grain problem; the detail of Kind’s exchange with Alter; Balog’s reply to the dilemma; the changes the essay implies for the earlier essays (dated notes at What Phenomenal Intuition Gets Right on “type B” and on the reply to Chalmers, and at Grain on “We agree”), which are listed in the outline.