Put two stones on the table, matched in mass, color, and shape. Ask someone to pick up the one you mean. “The smooth gray stone” will not help; both qualify. “The one beside this lamp” will, because now you have supplied a relation to something already picked out. A longer list of shared qualities can leave us very precisely undecided.

The puzzle concerns thought about that particular thing, what philosophers call de re thought. We usually solve it without noticing: point, look, trace a name back to someone who encountered its bearer. Now remove all those routes. Imagine a mind with no history and no channel to a world, equipped with every formal dictionary entry but no encounter, teacher, or inherited name. What could make its thought concern one stone rather than the other?

A language model does not fit that description. It inherits human texts and a training history. The question for machines concerns what those routes transmit and how the system uses them, not whether a real model amounts to a historyless mind.

Part Two argued that experience consists in representational content of the right embodied, world-directed kind. Part Three now asks how a state comes to be about a world at all. Biological function, causal history, and a community of speakers help fix content. Singular reference adds a specific constraint: a concrete particular needs a relational handle that distinguishes it from qualitative duplicates. Encounter can supply that handle; later consumers can inherit it through the practice.

The constraint belongs to a referential practice, not every speaker’s sensory biography. Somewhere in the chain, a world-involving relation must distinguish the particular from its rivals. Strawson and Gareth Evans developed the modern account of how our practice achieves this; the rationalist challenge asks whether thought could instead reach a particular through description alone.¹

12.1 Why Reference Seemed Free

The inner theater makes reference look free. To think about a stone, on this picture, is to entertain an inner item that pictures or encodes it. The stone causes the item; after delivery, the thinking goes on inside. Cut every sense and the item remains, apparently carrying its aboutness with it.²

But an inner item, considered merely as a pattern or state, fixes no object by itself. Relations to the world distinguish this stone from its rivals, and a relation is not luggage the mind can carry indoors. The sealed-mind question asks whether singular reference can originate without encounter, inheritance, or another channel that selects one particular. It cannot.

12.2 Tokening and Singling Out

Distinguish a system’s having a representation from a token’s picking out a particular. A flashcard bearing “the Eiffel Tower” carries a public representation through the practice that produced it; the ink alone does not fix the tower. A model emitting the same string has tokened a well-formed item, richly connected to others, and training may give its internal mechanisms task-shaped structure carrying inherited content. Whether any state singles out that tower in Paris remains a further question.

Call the first achievement tokening: producing or entertaining an item with the relevant representational form and internal liaisons. Call the second singling out: standing in the relation to a particular that makes a state about that one rather than another. Neither internal structure nor qualitative fit can single out this stone over its perfect twin; some causal-historical route must tie a state or practice to one.

Tokening comes cheaply to a sealed mind or a model, though training may earn the model more. Singular reference also needs a world-involving relation.

12.3 The Floor in Three Steps

1. Step one: qualitative fit alone does not ground de re reference to a concrete particular. The stones become distinguishable when we borrow an anchor: this lamp, this speaker, here. Remove the borrowed anchors, as Max Black’s perfectly symmetrical spheres do, and a complete qualitative description hangs over both at once.³ A definite description may still denote a unique satisfier in an ordinary asymmetric case. The narrower claim is that qualitative, uncentered fit cannot by itself determine which member of a symmetric pair this thought is de re about. Centered, indexical, and causal descriptions do not refute that claim; they reveal its metasemantic point. “The stone two feet from me” discriminates because me is evaluated from an occupied standpoint. “The cause of this very use” discriminates by tracing this token’s actual ancestry. The expressions can remain grammatically descriptive, but their anchor is fixed by standing in a relation, not by matching a further qualitative specification. Any regress of descriptions ends in a token, center, or producer already placed. Causal contact, or a name inherited from someone who once had such contact, supplies a handle of that kind. Strawson and Evans make spatiotemporal location fundamental to our practice because it supplies the public framework in which those routes converge.⁴⁵ The moral is narrower than “every referrer must locate.” A singular practice must distinguish.
2. Step two: a referential practice gets placed through world-involving producers. A perceiver does not infer its perceptually present here and now from a neutral description; bodily and sensory encounter places it in a world already under way. That encounter supplies the clearest producer relation: this stone, presented here. Only now do the little words earn their reach. Here, left, and now single out from a standpoint: a demonstrative borrows its aim from the encounter that anchors it — a relation the twin does not stand in, not a description the twin might also satisfy. But the relation can travel. A child who learns Napoleon from teachers need not meet Napoleon; testimony and deference connect the child’s use to producers whose practice was anchored through records, artifacts, and earlier encounters. Thought alone cannot manufacture the initial singular element, but an individual thinker can inherit it.⁶ Sensibility therefore names the paradigmatic entrance of particulars into a practice, not a demand that every later consumer personally sense each object.
3. Step three joins them. Qualitative description cannot distinguish duplicates (step one). A world-involving producer relation can establish a singular handle, and a causal-historical practice can pass it to consumers (step two). Therefore a closed, historyless system cannot create singular reference from internal form alone. It may possess the form of a spatial representation, a superb atlas with no route connecting any mark to this particular terrain. If it later receives an inherited name, a structured feed from a real environment, or another discriminating relation, the verdict can change. The floor is ecological and genealogical: somewhere in the practice, the world must get a vote. The same lesson will return when instantaneous physical duplicates diverge in content because their histories differ: present internal structure does not exhaust the relations that fix what a state is about.

This floor concerns singular address; general property content and abstract reference face different determination problems.

12.4 Objections

12.4.1 The mathematician’s numbers

We refer constantly to things no sense could ever touch: the number seven, the empty set, the value of justice, the proposition that snow is white. None of these occupies a region of space or a moment of time. If reference required spatiotemporal location, none of it would be possible — and plainly it is. So the thesis is false.

The objection reaches beyond the thesis, which concerns concrete de re reference, where a qualitative twin can stand outside the description and be missed. Abstract reference raises a different determination problem. In our practice it grows from symbol use, proof, teaching, stipulation, and applications embedded in a public world: seven from traffic with pluralities, the empty set from operations on collections, a proposition from worldly conditions that would make it true. The practice remains world-embedded even when its referent lies outside space and time; structural, inferential, or stipulative relations may fix an abstract referent without causal contact with it.⁷ That genealogy leaves room for nominalists, structuralists, and non-haecceitistic Platonists. A Platonism demanding singular reach to this causally isolated abstractum among structurally equivalent candidates inherits a further problem. Even seven was raised on apples.

12.4.2 The self-sufficient intellect

The rationalist asks why the limits of human cognition should bind every mind. Perhaps an intellect could grasp its objects directly without sensory contact: the tradition’s intuitus intellectualis, or intellectual intuition. Would our account rule it out merely because it lacks familiar senses?

The objection has a real historical champion. Leibniz held that every individual thing has a complete individual concept: a description that includes everything ever true of it.⁸ Such a concept, he thought, fastens on Caesar and no other — not by locating him, but by the sheer completeness of the description. Leibniz, a mathematician of the first rank who thought the world through to its logical floor, gives the rationalist his strongest card.

A description, however complete, may determine a unique satisfier; Evans’s distinction asks whether the thinker thereby achieves the object-dependent reach at issue here.⁹ Leibniz can deny that a qualitative duplicate remains possible and insist that Caesar alone satisfies Caesar’s complete concept in every possible world. Even then, we can ask what attaches this token of thought to Caesar rather than merely making its description fit him.

Leibniz’s best modern reader would object that this sells the doctrine short. Robert Merrihew Adams, whose 1994 study set the standard for taking Leibniz’s metaphysics whole, reads the complete concept as no mere list of qualities: each substance expresses the entire world from its own point of view, so the concept arrives already relational, already perspectival. Grant the reading; it moves the problem without solving it. A description of a perspective tells you what the view contains, not what attaches this token of thought to its subject. If a primitive thisness, divine selection, or direct presentation supplies that attachment, the singular element has been supplied rather than derived from qualitative content. The rationalist may call the relation direct intellectual presentation. Fine; then the view has supplied a discriminating presentation rather than generated reference from description alone. The concession keeps its teeth: a faculty that reaches its object owes the same account any channel owes, of how it discriminates this one, while a faculty that supplies no such relation remains internal form. What matters is the reach, not whether it arrives through a human sense organ.¹⁰

12.4.3 Testimony, Disability, and Unfamiliar Channels

The personal-sensory version of the thesis would prove too much. A blind or deaf thinker refers fully; a speaker can refer to Napoleon by testimony; a scientist can name an entity detected only through instruments. The ecological version predicts all three. Reference can pass through touch, records, teachers, cables, and instruments because the practice contains producers and preservative chains. A brain receiving structured signals from a real environment may acquire singular handles through the cable; a machine may acquire them through unfamiliar channels.¹¹ The condition fixes on a discriminating world relation somewhere in the practice and stays neutral on hardware. What it rules out is not silicon or disability. It is self-anchoring form.

12.4.4 The machine that has space and time of its own

A modern language model, the fourth objection says, has space and time of its own. It processes inputs in sequence, token after token, giving it a kind of time. It also organizes meanings in a high-dimensional vector space where nearness encodes similarity, giving it a kind of space. The model therefore commands a spatiotemporal framework, and the floor poses no barrier.

The model’s “space” and “time” do real work, but not all the work reference needs. A token sequence orders symbols; an embedding space maps similarities and, through training, acquires task-shaped structure and learned sensitivity to textual distributions. Neither geometry alone distinguishes this stone from its twin, because qualitative and structural relations can survive a swap of distal particulars. Internal order can make inherited content causally active without fixing a singular address or originating the semantic standard. The inference from interpretable geometry to self-baptized worldly relata therefore fails. Training, retrieval, feedback, tools, or sensors offer stronger candidate routes, and some may meet the floor by linking a state to a particular.

The strongest version says the geometry may model the world, not merely words. A game-playing model can encode the current board state, and interventions on that representation can change its predictions. Grant the finding: a learned world-model can carry content fixed by training and use. A formal game position needs no geographical address to count as content. If we ask which particular physical board or episode the state concerns, however, we ask a further question. That singular link must come from training history, a live feed, inherited producer chains, or some other route to the particular.

Discourse can preserve a producer chain for a consumer; a live retrieval, tool, or sensor call can transduce a particular’s state at use. The floor classifies these possibilities rather than barring them. Even when an episode fixes a referent, two questions remain: whether the system participates as a consumer rather than routing a referential token, and whether it integrates the content across memory, correction, inference, planning, learning, and action. Chapters 15 and 23 meet them. Coordinates can carry content, but geometry alone does not baptize its relata.¹²

12.5 A Real Space and Time, and the Attribution Question Left Open

Concrete de re reference needs a relation that distinguishes one particular from its qualitative rivals. Encounter supplies the clearest producer route; history and community can carry it to consumers. The requirement concerns genealogy, not hardware or private sensory biography.

The argument does not make space and time mind-made. Sensibility supplies a mode of access to a spatiotemporal world, never the world itself. Stones stand at a distance whether or not anyone perceives them; the senses supply a standpoint in that real frame. Strawson reconstructed this realist reading from the older argument, and the externalism ahead requires it.¹³ A book that insists you see the apple rather than an inner picture cannot turn space into another inner picture.

Unlike the sealed mind, a language model inherits causal history, public meanings, and task-shaped internal structure. Its tokens may carry derived routes to particulars, and a deployed system may participate in some. The remaining questions ask whether it acts as a consumer by deferring to producers, using the token as part of the practice, and remaining corrigible by its standard, and whether it integrates the content across memory, correction, inference, planning, learning, and action. Part Four takes up those questions; the model is not simply the sealed mind with better arithmetic.¹⁴

Chapter Summary

The claim. Concrete reference to this person, tree, or place cannot arise from qualitative fit or internal structure alone. It needs a route through the world that singles out one particular; encounter provides the clearest case, while testimony and causal history can pass the handle onward.

Where the argument stands. An uncentered description can fit perfect duplicates, so added internal detail cannot select one of them. Centered and causal descriptions work only by relying on a standpoint, token, or source already fixed through a relation to the world. Spatiotemporal contact can provide that anchor, and public practice can preserve it for later users; the requirement concerns history and organization, not hardware, so silicon can refer and each human speaker need not sense the object firsthand.

What’s left open. Abstract reference may need another account. A corpus history may link an artificial state to a distant object or only to linguistic contexts, and even successful reference leaves the system’s understanding unsettled.

The hand-off. A signal can point home without knowing where home lies. The next chapter separates information that travels from meaning that reaches.

---

Notes

1. The argument’s modern, realist statement is Strawson’s and Gareth Evans’s (notes 4 and 5): material bodies in a single spatiotemporal system are the basic particulars our referential practice rests on, and a fundamental producer route singles one out by locating it in that system. Evans’s producer-consumer distinction also explains how later speakers inherit singular reference without encountering the object. The distant ancestor is Kant’s claim, in the Transcendental Aesthetic of the Critique of Pure Reason (1781/1787), that space and time are the forms under which a sensible mind is given particulars. This chapter keeps the realist and practice-level insight while leaving behind both Kant’s idealism and the stronger claim that every individual act of reference requires a personal sensory presentation. ↩
2. The genealogy is the work of Chapter 2. The semantic version of the inner-theater error — locating aboutness in an inner item rather than in a state’s relation to the world — is the same reification the book diagnoses for experience in Chapters 5 and 7, transposed from phenomenal character to intentional content. ↩
3. Max Black, “The Identity of Indiscernibles,” Mind 61 (1952): 153–164, argues against the Identity of Indiscernibles through a symmetrical two-sphere universe. The chapter makes no claim that spatial position metaphysically grounds their distinctness; in Black’s world the positions are themselves descriptively symmetrical. Its cognitive prediction is that an equally symmetrical thinker, with no asymmetric relation to either sphere, could not think de re of one rather than the other. Demonstrative location is the paradigm of the needed handle, but a causal-historical name or inherited producer chain can preserve the same distinction. ↩
4. P. F. Strawson, Individuals: An Essay in Descriptive Metaphysics (London: Methuen, 1959), Part I. Strawson argues that material bodies in a single spatiotemporal system are the basic particulars of our conceptual scheme. The claim concerns the structure of an identifying practice, not a requirement that each user personally perceive each particular. “Descriptive metaphysics” is his term for laying bare the structure ordinary thought already has, as against replacing it with a revisionary scheme. ↩
5. Gareth Evans, The Varieties of Reference, ed. John McDowell (Oxford: Clarendon Press, 1982), esp. ch. 4 for the fundamental Idea, ch. 6, “Demonstrative Identification,” and ch. 11 on name-using practices. Evans’s fundamental Idea locates an object in objective space and time; his distinction between producers and consumers explains how a name-user can retain singular reference by deferring to a practice whose producers possess direct information links to the object. The first relation anchors; the second preserves and distributes the reach. ↩
6. The singular/general contrast is Evans’s (Varieties of Reference, chs. 4–6): a general thought represents a kind and underdetermines which particular falls under it, whereas a singular thought reaches a determinate object. Perceptual and bodily anchoring supplies the fundamental producer case; Evans’s later producer-consumer apparatus allows individual thinkers to inherit that reach. The older form is Kant’s contrast between concept and intuition (Critique of Pure Reason, A51/B75, in the Meiklejohn translation), with the proviso of note 1: the chapter keeps the need for a singular presentation somewhere and drops both the idealism and the personal-sensory universal. ↩
7. Strawson argues in Individuals, ch. 1, that reference to non-basic particulars is introduced within a conceptual scheme whose basic identifying practice concerns spatiotemporal bodies. The extension to abstracta is the chapter’s genealogical proposal, not a result Strawson alone establishes. It says how a public practice acquires and teaches its mathematical symbols, not that the number seven occupies space or that every mathematician must trace each proof back to apples. Benacerraf’s familiar identification problem explains the remaining qualification: structurally equivalent reductions of numbers leave a haecceitistic Platonist owing an account of why the practice reaches one abstract object rather than its rival. ↩
8. The doctrine of the complete individual concept (notio completa) is Leibniz’s: every individual substance has a concept so complete that it contains, and allows the deduction of, every predicate ever true of that substance — the position stated in the Discourse on Metaphysics (1686), sec. 8 (no verbatim quotation is made here; for the text see G. W. Leibniz, Philosophical Essays, ed. and trans. Roger Ariew and Daniel Garber [Indianapolis: Hackett, 1989]). The complete concept is the strongest historical attempt to individuate a particular by descriptive content alone, with no spatiotemporal handle, which is why it is the rationalist’s best instance of the intuitus-style grasp the objection invokes. Robert M. Adams, Leibniz: Determinist, Theist, Idealist (Oxford: Oxford University Press, 1994), chs. 1–3, treats contingency, identity, and complete concepts; Part III, chs. 9–13, treats the wider monadic and perspectival system. Adams supplies the strongest reconstruction rather than an objection to the chapter. The revised reply does not assume Leibniz must admit duplicates: even necessary unique satisfaction leaves open what attaches this token of thought to the satisfier. ↩
9. The reply applies Evans’s singular/general distinction (Varieties of Reference, chs. 4–5; see note 6) to the complete concept. A fully specified general concept may denote a unique satisfier while failing to place the thinker in an object-dependent relation to it. What closes that gap is a singular element: demonstrative presentation, causal acquaintance, baptism, or an inherited chain preserving one of those relations. The structural objection runs back through Russell’s distinction between knowledge by acquaintance and by description, “Knowledge by Acquaintance and Knowledge by Description,” Proceedings of the Aristotelian Society 11 (1910–11): 108–128. Causal descriptivism builds the relation into a description — reference fixed by a token-reflexive clause such as “the object at the causal origin of this very use”: David Lewis, “Putnam’s Paradox,” Australasian Journal of Philosophy 62 (1984): 221–236, who coins the label; Frederick W. Kroon, “Causal Descriptivism,” Australasian Journal of Philosophy 65 (1987): 1–17. The revised argument grants the vocabulary. The expression may remain descriptive; its success depends on the actual token and actual causal ancestry that make its reflexive clause true. Causal descriptivism preserves descriptivist semantics by accepting the relational metasemantics this chapter requires. ↩
10. The intuitus intellectualis — an intellect that would create or grasp its objects directly, without being sensibly affected — is Kant’s foil for finite cognition (e.g., B71–72, B135, B145). Kant denies humans possess it. The reply here is conditional rather than theological: if direct intellectual grasp supplies a genuine relation to one object rather than its duplicate, it meets the chapter’s world-involving constraint by another route. What it cannot do is derive singularity from purely general descriptive content. Nor does the conditional trivialize the constraint: either the posited grasp stands in a discriminating relation to its object, in which case it functions as a channel and its defender owes an account of how the channel discriminates, or it operates on internal content alone, in which case it hangs over duplicates with every other description. “World-involving” names the relation’s achievement, not a courtesy title awarded to whatever succeeds. Cf. Evans, Varieties of Reference, ch. 4. ↩
11. The organ- and substrate-neutrality of the anchoring relation mirrors the book’s standing anti-chauvinism. Chapter 23, meeting Mollo and Millière, grants that sensorimotor transduction is not necessary for reference; causal-historical learning, social deference, and other channels may do the work. The present chapter adds the ecological constraint: a singular practice cannot originate its distal addresses from internal description alone. Some producer or learning route must be affected by, discriminate, or otherwise become accountable to the particular; consumers may inherit that relation. The brain-in-a-vat and machine cases are therefore decided by the feed and history, not by the familiar shape of an organ. ↩
12. The “merry-go-round” is Harnad’s symbol-grounding image (Chapter 23). An internal metric or sequence can support genuine computation and, when shaped by training and consumer use, task-specific causal organization and content. What geometry alone cannot do is fix a distal singular referent: structural relations remain invariant under permutations of the worldly relata. Training history, a live indexical feed, or an inherited producer chain may add the missing relation. The empirical case is Kenneth Li, Aspen K. Hopkins, David Bau, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg, “Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task,” International Conference on Learning Representations (2023), arXiv:2210.13382: board state is recoverable from Othello-GPT’s activations, and intervention on the recovered representation changes subsequent predictions. The live philosophical dispute divides accordingly. Steven T. Piantadosi and Felix Hill, “Meaning Without Reference in Large Language Models” (2022), arXiv:2208.02957, defend conceptual-role meaning without reference; Matthew Mandelkern and Tal Linzen, “Do Language Models’ Words Refer?”, Computational Linguistics (2024), arXiv:2308.05576, argue that model words may inherit reference through human practice. This chapter does not settle whole-system understanding; it rules out internal structural readability as self-sufficient metasemantics and classifies candidate inherited routes. ↩
13. P. F. Strawson, The Bounds of Sense: An Essay on Kant’s “Critique of Pure Reason” (London: Methuen, 1966), which separates the defensible Kantian argument (the conditions of a possible experience of an objective world) from the transcendental idealism Strawson rejects. The book takes the former and leaves the latter, consistent with its realism: space and time are real features of the world; sensibility is our mode of access to them, not their author. Evans’s externalism (Varieties of Reference) likewise presupposes a mind-independent spatial world that thought reaches rather than constitutes. ↩
14. The comparison now turns on genealogy rather than sensory inventory. A genuinely historyless sealed mind lacks any route that fixes a singular address. A trained language model does have a causal history: public word-types carry derived meaning, and training gives internal mechanisms learned sensitivities to textual distributions. Evans’s consumer condition (note 5) requires more than causal descent: the use defers to producers, manifests participation in the relevant practice, and remains corrigible by its standard. These achievements come apart: a token’s inherited derived reference, a mechanism’s causal internalization of that content, an agent’s consumer participation and integrated understanding, and an autonomous individual’s stake, vulnerability, and welfare. Chapters 14–16 distinguish these achievements; Chapters 23–24 apply them to artificial systems. ↩