16.1 One Sentence, Several Achievements

“The window is open.”

Suppose I say it while a summer breeze lifts the curtain: a report. Say it as rain crosses the sill and you sit by the latch: probably a request. Say it after glass breaks downstairs: a warning.

The words, order, and window stay fixed. What I mean and do change.

The sentence presents a condition: the window stands open. My purpose may go further: I want you to close it. In speaking, I may also make a request for which you can reasonably ask why. Content, speaker meaning, and the act performed with the words already come apart.¹ You can understand the request perfectly and still refuse it.

Chapters 12–15 supplied the metasemantic background: form cannot select a duplicate, correlation lacks an error standard, and producer-consumer history may establish one. A frog can represent prey without words.

Language adds public signs, combinatorial structure, communicative purposes, acts, and social correction.

Calling all of this use leaves us something to explain. Use includes teaching, correction, perception, consequence, and action. Text prediction may reward astonishing competence while directly comparing outputs only with signs. Which wider relations make those signs about rain, arthritis, or this window?

My answer keeps use and reference together. Not every word names an object. Rather, some uses must answer to objects, events, conditions, or socially maintained kinds beyond relations among signs. Those routes determine the practice’s empirical subject matter. Reference names this worldward dimension of use, not an added ingredient.

Control and Map supplied the nonsemantic test. Map preserved and corrected one target across consumers; Control only triggered a response. That world-directed organization, not a hidden mind, does the grounding work.

A greeting need not name anything, and a promise can fail without referring badly. Frege and Russell show why reference cannot mean attaching labels to every word.

16.2 Sense, Reference, and Empty Expressions

The simplest theory of meaning treats words as labels. Cat means cat because the word stands for cats. London means London because it names a city. The view offers a pleasing economy: find the object, attach the word, and send language on its way.

Two old puzzles spoil the arrangement.

Gottlob Frege spent his career trying to show how thought could preserve both objectivity and logical structure. In 1892 he asked why identifying the morning star with the evening star can teach someone something while identifying the morning star with itself cannot. Both expressions in his example refer to Venus. If reference exhausted meaning, the two identity statements should carry the same cognitive value. They plainly do not.²

Frege distinguished a sign’s reference, what it stands for, from its sense, the way that referent gets presented. The morning star and the evening star reach the same planet by different routes. A person may grasp one route without knowing that it joins the other.

Sense also lets us understand an empty expression. The largest prime number presents a condition although no number satisfies it, and a fictional name can guide a reader without acquiring a flesh-and-blood bearer. Language can organize a thought before the world supplies an object.

Bertrand Russell wanted a different repair. In “On Denoting,” he turned to the present King of France. The phrase looks like a name, but Russell analyzed it as contributing existence, uniqueness, and predication conditions without placing any king inside the proposition. The present King of France is bald says, roughly, that exactly one person now reigns as King of France and that this person is bald. The existence condition fails. The sentence comes out false without requiring a royal ghost to make it meaningful.³

Grammatical shape does not settle semantic function. A noun phrase may quantify rather than contribute an object; other expressions modify or connect. Discourse can fail to refer at one point while remaining answerable to a world at another.

The qualification locates reference at the level of a linguistic practice. Fregean sense, Russellian analysis, logical vocabulary, fiction, questions, and promises enrich that practice without making reference every word’s job. None shows that a complete natural language with determinate empirical subject matter could float free of any world.

Consider a frightened child who says, “There is a monster under my bed.” No monster answers, but the child has not produced noise. Stories and pictures give the word recognizable criteria. The sentence reaches toward the world and misses. Missing differs from never reaching.

The distinction becomes decisive in the artificial case. A language model’s emitted words already belong to English. Their tokens inherit conventional, derived meaning from the practice that supplied the training corpus. We should not deny that meaning merely because the model did not invent English. Printed books do not invent it either. The remaining question concerns the system’s own referential acts. Does water in this output merely reproduce a public word whose reference others fixed, or does a continuing artificial system itself use the word to track, correct, remember, and act upon water? The visible token cannot answer that question alone.

Words seldom travel alone. Language owes much of its power to how they combine.

16.3 How Sentences Build Meaning

A child learns a few thousand words and soon understands sentences nobody has previously uttered. You can read Three purple elephants discussed mortgage rates beneath the moon without consulting a phrasebook devoted to elephant finance. The sentence’s novelty does not block comprehension. Familiar parts and familiar modes of combination let you work out an unfamiliar whole.

Philosophers and linguists call this feature compositionality: the meaning of a complex expression depends systematically on the meanings of its parts and on how those parts are arranged.⁴ Word order matters because the dog bit the postman and the postman bit the dog contain the same vocabulary and report different disturbances. Structure makes finite language productive.

Compositionality explains a real achievement. It also tempts a familiar overreach. Once a formal system can represent syntax and calculate how parts combine, one may think semantics has emerged from the machinery. But a rule of combination preserves and transforms whatever semantic contributions its inputs already possess. It does not establish that dog concerns dogs, that water reaches H₂O, or that a speaker’s claim answers to the world rather than to a sequence of marks.

The distinction here runs between a semantic theory and a metasemantic theory. A semantic theory tells us what expressions mean and how their meanings determine the contents of larger expressions. A metasemantic theory asks what facts make those expressions have those meanings in the first place.⁵ The previous chapters supplied part of the metasemantic story: causal history, selection, learning, public practice, and producer-consumer use. Compositionality explains how content, once grounded, can travel through structure.

Donald Davidson gave this semantic ambition one of its most influential forms. Rather than treating meanings as another class of objects attached to sentences, he proposed a systematic theory that derives the truth conditions of each sentence from the semantic roles of its parts: “Snow is white” is true if and only if snow is white.⁶

Truth conditions do not exhaust language. Questions, commands, jokes, metaphors, and context-sensitive expressions require more. Davidson’s point concerns systematicity: finite resources must explain how speakers grasp indefinitely many sentences.

This matters for language models because compositional competence counts as evidence of more than memorization. A model that handles new combinations, preserves dependencies across long passages, and adapts an expression to a novel context has learned substantial regularities of language.

It does not yet answer the grounding question. Two systems can manipulate the same sentence structure while their words reach different things, as Twin Earth will make vivid. A third may preserve every formal relation while inheriting its semantic interpretation from the people whose language it models. Composition can carry meaning a long way. It cannot choose where the journey began or what lies at its destination.

16.4 Meaning in Use

Ludwig Wittgenstein approached ordinary words with the suspicion that philosophy had lost its way by leaving their use behind. In Philosophical Investigations, he gave up the search for one hidden essence and looked instead at the varied lives of words.

“For a large class of cases—though not for all—in which we employ the word ‘meaning’ it can be defined thus: the meaning of a word is its use in the language,” Wittgenstein wrote.⁷ The qualification matters. He asked us to inspect language-games: ordering, describing, guessing, joking, thanking. Speaking belongs to an activity, a form of life.

The shift breaks the private-label picture. Imagine someone who calls a recurring private sensation S and records each occurrence in a diary, relying on memory to guide the next entry. What would distinguish remembering correctly from merely feeling satisfied? If whatever seems right counts as right, correctness has collapsed into apparent correctness. A sign whose application answers only to what seems right at each use cannot supply its own standard.⁸

Nothing in that argument denies sensations, introspection, first-person authority, or talking to oneself. Solitude causes no difficulty. The difficulty comes from stipulating a use for which apparent correctness exhausts every possible standard of correctness.

Rules work through practices. Children learn words amid examples, correction, consequence, and shared attention. A carpenter who calls for a slab expects a slab, not a beam; another participant can correct the mistake, and the building can expose it. No monitor needs to watch every later application. Correct and incorrect already have a place in the learned activity.

A public standard does not turn consensus into truth. Many speakers and inherited classifications treated whales as fish. Public use made disagreement intelligible; anatomical and evolutionary evidence, working through the aims of biological classification, forced a revision. The community did not vote mammals into existence, and the world did not write the taxonomy for us.

Wittgenstein’s investigation locates correctness within norm-governed activity; it does not offer a causal theory of reference. The synthesis I draw from it goes further. Meaning-as-use and reference reinforce each other because meaningful use already includes practices of identifying, testing, correcting, and acting upon things. A practice remains open to objects, events, and people that can prove its applications wrong.

Kripke, Putnam, and Burge add a different consideration without turning Wittgenstein into an externalist: which causal histories, environments, and social relations help determine the people, substances, and kinds a practice concerns? The combination belongs to the account developed here. First we need to notice what speakers do with content once they have it.

16.5 What Speakers Do with Words

Return to the open window. The sentence says that the window stands open. That content remains available whether I report the fact, warn you about an intruder, or request that you close it. Sentence meaning constrains what I can reasonably mean. It does not finish the job.

H. P. Grice had an unusually attentive ear for the work ordinary conversation performs. He drew the decisive distinction between signs that naturally indicate something and speakers who mean something. Dark clouds can mean rain without intending an audience. Spots can mean measles. Communicative meaning differs. A speaker who says, “Take an umbrella,” may intend a hearer to recognize that rain is coming and to form that belief partly by recognizing why the words were spoken.⁹

The account explains how speakers communicate more than their sentences literally say. Adapt one of Grice’s examples: ask a colleague whether Smith makes a good philosopher and hear, “His handwriting is excellent.” The sentence concerns handwriting. In the circumstances, the speaker probably means that Smith’s philosophy inspires no comparable praise.¹⁰

Grice organized such reasoning around a Cooperative Principle and conversational maxims: give the exchange what it requires, avoid what you believe false, stay relevant, and speak clearly. Actual conversation departs from these maxims cheerfully and often. Implicature frequently arises when a speaker conspicuously flouts one while leaving cooperation in place. Irony would have a difficult career otherwise.

J. L. Austin widened the view. Philosophers had concentrated on statements and their truth. Austin pointed to utterances that perform actions: I apologize, I promise, I name this ship. Saying the words under the right conditions does not merely report an apology, promise, or naming. It carries one out.¹¹

Austin distinguished three acts. A locutionary act produces meaningful words with a certain sense and reference. An illocutionary act does something in saying them: asserting, warning, promising, or requesting. Perlocutionary effects follow from saying them, as when speech persuades, alarms, or amuses. John Searle refined the structure by separating utterance acts, propositional acts, and illocutionary force. The same reference and predication can appear in an assertion, a question, or a command.¹²

Words and grammar contribute content. Context helps determine what gets said. A speaker uses the content with a communicative purpose and may perform an act that carries a force and a commitment. The hearer may infer still more.

Not every speech act answers to truth in the same way. Assertions commit speakers to how things stand. Directives aim to make conduct fit words. Promises place the speaker under an obligation to make later conduct fit what was said. Questions solicit an answer. Each has conditions under which it succeeds, misfires, or gets withdrawn.

The same gap reopens for artificial systems because emitting an assertion-shaped sentence and undertaking an assertion may come apart. A system can print The bridge is safe. Designers, users, or an institution may use that output to issue a warning or certify the bridge without making the mechanism a speaker on its own account. Has the system itself committed to the claim? Can evidence defeat the commitment? Will the correction govern later answers and plans? Which continuing organization stands behind the words tomorrow? Fluency supplies no automatic answers.

16.6 How Language Reaches the World

Wittgenstein’s grammatical investigation has already refused the picture of meaning as private mental cargo. It neither asks nor answers the later metasemantic question: why does water reach H₂O rather than the clear liquid on Twin Earth, or why does arthritis exclude an ailment in the thigh despite one speaker’s confident usage?

Saul Kripke attacked the idea that every speaker carries a description that fixes a name’s bearer. Most people who use Feynman cannot uniquely describe Richard Feynman. They inherit the name through a chain of communication reaching back through books, teachers, colleagues, and earlier speakers. Somewhere in that history, uses connect to the person. Later speakers can refer successfully while knowing very little.¹³

The chain involves more than causal descent. A later speaker normally intends to use the name with the reference it has in the practice. By the same condition, calling a dog Feynman can begin a new chain precisely because the speaker does not intend to preserve the physicist’s name-bearing relation. That condition will matter when we ask whether a trained mechanism participates as a linguistic consumer or merely transmits tokens whose reference others fixed.

Hilary Putnam built Twin Earth with a mathematician’s discipline: alter one variable and hold the rest fixed. On that planet, a liquid called water looks and behaves like ours but has a different chemical structure, XYZ. Imagine Oscar on Earth and his psychological counterpart on Twin Earth before either community knows chemistry. Their inner states match. Their terms do not reach the same substance. Oscar’s practice developed around H₂O, while Twin Oscar’s developed around XYZ. Their environments help fix what their words concern.¹⁴

Tyler Burge showed how the community enters the picture. His patient believes he has arthritis in his thigh, though arthritis affects joints. The patient has not changed the meaning of arthritis by misunderstanding it. Medical and linguistic practice supplies a standard against which his use comes out mistaken. A counterfactual community that used the word more broadly would give the same inner state a different content.¹⁵

Kripke supplies causal history, Putnam environmental anchoring, and Burge social individuation. Gareth Evans adds a useful division of labor. Producers maintain recognitional or informational contact with a referent. Consumers inherit the ability to use the term through deference and communication. Each of us consumes far more names than we produce. Reference does not require every speaker to meet every object. Consumer standing still involves participation in the practice: prior producers count as authoritative, and correction can alter later use.¹⁶

I take these externalist arguments to fit the account of use rather than replace it. That compatibility forms part of this book’s synthesis, not a doctrine Wittgenstein stated. A public practice reaches the world through an ecology: producers encounter and investigate; consumers inherit, defer, and correct their use when that ecology pushes back.

16.6.1 The conceptual-role objection

A stronger challenge goes beyond empty names. A conceptual-role theorist argues that meaning can consist in the place an expression occupies within reasoning and use. We understand justice through the judgments it licenses, wit through contrasts and applications, and even a fictional king through the network of claims in which he appears. Language models can learn enough of that network to draw novel inferences. Why demand a further word-world relation? Steven Piantadosi and Felix Hill press this case directly for large language models: relations among a system’s internal states may realize important aspects of meaning without reference. Robert Brandom, a contemporary American pragmatist, builds a broader inferentialism around what speakers commit themselves to, what claims they have reason to make, and which inferences depend on content rather than form. His account also includes entry moves from perception to claims and exit moves from claims to action, so it does not close the network around words alone.¹⁷

The challenge identifies something real. Inferential role explains important aspects of sense and competence. A narrow version claims that relations among internal states can fix subject matter by themselves. Imagine two practices whose token-to-token inferences and actual-world use profiles match. In Actual, speakers produce red through a detector whose history and organization make it track red surfaces. In Twin, the same token and internal transitions answer to a magnetic field. Redness and the field covary throughout the speakers’ actual environment, so their ordinary uses match. Intervene to break the correlation by presenting a red surface without the field, however, and the different producer routes show themselves.

Brandom rejects representational reference as a primitive building block of meaning. His inferentialism instead includes reliable circumstances in which speakers apply a term, practical consequences of applying it, and social practices of commitment and correction. On his account, a practice comes to count as representing the world through that norm-governed activity rather than through a semantic atom added first. This gives the practice worldly friction in Brandom’s own terms.

That broader answer satisfies the practice-level requirement rather than defeating it. Once reliable entry and exit moves count as part of an expression’s role, the practice no longer closes around words alone. I need not claim that reference comes before use. I claim that meaningful use in a complete natural language must include some world-reaching routes that determine empirical subject matter. The Twin case targets the narrower claim that internal relations suffice.

A metasemantic question still remains for any world-facing role: what makes an entry move answer to redness rather than to the field? Producer history, consumer use, and learning constitute the different routes; interventions reveal which route continues to govern when the correlations come apart. If none privileges a target, the content remains indeterminate.

This establishes the practice-level requirement without assigning a referent to every word. A complete natural language must contain some routes by which its uses answer to nonlinguistic or socially maintained conditions beyond relations among signs. Fiction, mathematics, logical vocabulary, and other subpractices can take their semantic place within that wider language. What, if anything, fixes reference to abstract objects remains open. The present claim needs only that a complete language cannot close around its own signs while retaining determinate empirical subject matter.

16.6.2 The empty-expression objection

Fregean sense and Russellian descriptions may seem to show that meaning needs no reference. They show something narrower: an expression can fail to denote, or perform no referential job, while the larger practice remains answerable beyond its signs. The present King of France fails against a world in which France has no king.

The objection therefore changes the level of the requirement rather than defeating it. Empty expressions defeat a universal denotation requirement, not the practice-level referential condition. Their applications, implications, and conditions of satisfaction still belong to a language that answers beyond the signs.

16.6.3 The inherited-reference objection

The stronger AI objection grants externalism and turns it in favor of language models. Matthew Mandelkern and Tal Linzen argue that a model’s words may inherit reference through the human uses recorded in its training corpus. If ordinary speakers can consume a name without meeting its bearer, why cannot a model inherit reference from the same chain?

Here the objection succeeds, up to a point. A generated token, the mechanism that produced it, or causally active internal states may inherit semantic or referential content through the public tokens and corpora used in training. A blanket denial would give human consumers a privilege the theory itself rejects.

I was once too quick to treat corpus descent as semantically inert. That was a mistake. Training can make public linguistic distinctions causally active inside a mechanism, and a deployed system may participate further through correction, tools, perception, memory, and action.

Evans’s consumer still participates in a practice: producers remain authoritative, correction alters later use, and an intention to preserve the inherited reference helps keep the chain joined. A mechanism need not reproduce that intention in human form. Nor must a one-off token persist in order to inherit semantic reference. Persistence bears on the different question whether one continuing system refers, retracts claims, or undertakes commitments on its own account. The relevant evidence asks whether corrections survive, govern later inference or action, and keep that organization answerable to the practice’s producers.

That evidence also identifies the bearer. A transient model, a persistent agent, and an embodied controller may use the same language mechanism while differing in which corrections survive, what observations constrain them, and whether later conduct answers to earlier words. The product label does not decide where the achievement belongs. The control organization does.

Mollo and Millière strengthen the externalist case for machines. Mediated causal information, selected functions, and world-sensitive training may ground content in model outputs or internal states. I accept that possibility without taking it to establish understanding, knowledge, or consciousness in the larger system.¹⁸

A brain envatted after a lifetime of worldly engagement can retain reference established before the vat. Putnam’s always-envatted brain lacks the ordinary causal route to brains and vats. An interactive system, by contrast, may form a genuine route to stable features of its physical or simulated environment. Where rival interpretations remain tied, reference may remain indeterminate. The medium can change without making inner traffic sufficient. History and use do the work.¹⁹

16.7 When a Machine Uses Language

Readers who take current language models to understand have good reason to resist the phrase mere fluency. Contemporary systems do more than repeat memorized sentences. They preserve grammatical structure across novel combinations, adapt to context, exploit implication, translate, summarize, correct local errors, and shift register. Those capacities provide real evidence of linguistic competence. The models’ internal states can carry causally active derived content inherited from public language. Calling all of this “just statistics” replaces analysis with tone.

Fluency gets a system into the argument. It does not decide it.

Imagine three systems saying The window is open. The first replays a recording when a light sensor flips. The second uses a learned classifier to track the sash and generate a report. The third belongs to a household agent that notices rain, checks the latch, warns you, and closes the window when authorized. The recording carries meaningful words its designers can use to warn; the classifier adds tracking; the agent coordinates the report with a purpose and an action.

Let the household agent use that same classifier, and let it mistake a reflection for a gap. You show the agent that the sash has closed and ask it to check the latch before warning you again. In the stronger version of the case, that correction changes its stored expectation, its next check, and its later report even after the conversation ends. Merely replacing the current sentence would achieve less. The correction gives us something to test: does the same organization learn what went wrong and use the lesson, or does someone outside it have to carry the lesson from one component to another?

That difference supplies evidence of understanding within this household task. It does not require every adult human capacity, and it does not establish consciousness. Whether the agent itself warns, rather than simply delivering a warning delegated by someone else, also depends on its communicative purpose and the commitments its later conduct sustains.

Six claims must therefore remain separate.

1. The output instantiates meaningful English expression types. Their conventional meanings come from the public practice; visible shape alone would not suffice.
2. The token or generating mechanism may carry semantic content. Corpus-mediated history can transmit derived content or reference to an output token and make related distinctions causally active in internal states. Corpus descent alone does not establish original intentionality or show that the larger system refers on its own account.
3. The system displays linguistic competence. It reliably exploits syntactic, semantic, contextual, and pragmatic regularities across new cases. This capacity does not presuppose understanding.
4. The system may mean something by its words. Gricean speaker meaning requires a communicative purpose available for audience recognition. A designer or user may also communicate through a mechanism that lacks a purpose of its own.
5. The system may perform a speech act. An utterance can function as an assertion, promise, warning, or request through convention or delegated authority. A system that undertakes the act on its own account must also bear the corresponding commitment. Persistence helps us attribute such commitments across time; it does not define illocutionary force.
6. The system may understand to some degree. Bounded understanding begins when several content-guided capacities constrain one another and remain answerable to correction. Breadth across memory, perception, inference, learning, planning, and action can make that understanding more integrated and world-involving. No single transcript settles how far it extends or which bearer has earned the attribution.

Havlík presses a different challenge: a sufficiently rich linguistic fragment may sustain minimal semantic content and understanding through its contextual organization. I do not think relations within the fragment alone fix a distal subject matter, but they can support a real bounded capacity. Burge’s arthritis patient supplies a human precedent for separating possession of content from complete mastery without establishing that any particular model crosses the threshold. Partial understanding can still count as understanding.²⁰

In this book I will sometimes put quotation marks around a language model’s “understanding.” They mark a limited, domain-specific capacity rather than a counterfeit one. The bearer may amount only to a running mechanism or deployed product during a bounded exchange: enough organization to interpret a question, draw relevant inferences, correct an error, and recombine what training supplied. Integrated personal understanding requires a further showing that memory, perception, purposes, and correction belong to one persisting, world-answerable organization. Chapter 17 takes up that question. Bounded understanding alone establishes neither original intentionality nor phenomenal consciousness.

These achievements do not form a ladder, and none may impersonate the others.

None of the six settles whether the system is conscious or whether a persisting subject stands behind the words. Those questions require their own evidence.

16.7.1 The other-minds objection

A defender of that attribution can press a fair objection. We do not inspect another person’s causal history and internal organization before accepting that person’s words. We hear fluent, context-sensitive speech and attribute understanding. Why demand more from a machine?

The history behind the charge deserves respect. AI debate has often treated yesterday’s impossibility as today’s irrelevant trick. The answer cannot consist in raising one bar each time a system clears the last one.

Human speech sometimes reaches us only as a transcript. Even then, our interpretation draws on extensive background evidence about the kind of continuing life that produced it: a life with perception, memory, correction, plans, and actions available in principle to further observation. We do not inspect that organization each time. The background makes fluency a reliable cue; when the cue comes apart from that background, its evidential weight falls. None of this provides certainty, and Chapter 19 rejects analogy as the universal basis of our knowledge of other minds, not as a possible aid. The point concerns evidence, not delivery medium.

The same standard should govern artificial systems. Biology receives no automatic privilege. When an artificial system preserves commitments, learns from correction, coordinates language with perception and action, and carries those gains through one continuing organization, the evidence rises. Where a product exposes only fluent output while persistence, correction, and world-involving control remain absent or scattered across components coordinated only by external users, the transcript underdetermines the verdict. Equal treatment requires the same distinctions, not identical assumptions about unlike cases.

Chapter Summary

The claim. A language with determinate empirical subject matter requires reference at the level of the whole practice, though many words name nothing. Some uses must reach beyond signs to objects, events, and shared kinds. Without such routes, the practice cannot fix what its claims concern.

Where the argument stands. Reference neither exhausts meaning nor comes before use. Form and inference alone cannot fix what a language concerns. Public rules, speaker aims, speech acts, causal history, the world, and social standards each do different work, so a system can show language skill without thereby earning every other success.

What’s left open. Fluent artificial systems may produce good English and tokens or inner states with inherited content. They may show real skill and, in a limited sense, “understand” within a field. We must still ask which bearer has original rather than derived content, refers or communicates on its own account, takes on commitments, and brings its content-guided skills together.

The hand-off. The next chapter turns from meaning to the one who means: a subject assembled through memory, language, and social life.

---

Notes

1. The distinctions among expression type, utterance, propositional content, speaker meaning, and illocutionary force cross the conventional semantics/pragmatics border. See Kepa Korta and John Perry, “Pragmatics,” Stanford Encyclopedia of Philosophy, substantive revision 28 May 2024, secs. 1–2; Mitchell Green, “Speech Acts,” Stanford Encyclopedia of Philosophy, substantive revision 24 September 2020, secs. 1–2. Green’s force/content discussion also explains why one propositional content can occur in assertions, questions, and directives without reducing the differences among those acts to differences in reference. ↩
2. Gottlob Frege, “Über Sinn und Bedeutung,” Zeitschrift für Philosophie und philosophische Kritik, n.s. 100, no. 1 (1892): 25–50; Max Black, trans., “On Sense and Reference,” in Translations from the Philosophical Writings of Gottlob Frege, ed. Peter Geach and Max Black (Oxford: Blackwell, 1952), 56–78. The informative-identity argument begins from the difference in cognitive value between a = a and a true a = b, then distinguishes Sinn from Bedeutung. The consulted translation uses morning star and evening star; the later Hesperus/Phosphorus relabeling does not appear in Frege’s 1892 text. ↩
3. Bertrand Russell, “On Denoting,” Mind 14, no. 56 (1905): 479–493. Russell’s contextual analysis treats a definite description as contributing existence, uniqueness, and predication conditions to the proposition expressed rather than standing for a peculiar denoting entity. Later work divides sharply over whether ordinary definite descriptions carry Russellian quantificational semantics, Strawsonian presuppositions, or referential uses in Donnellan’s sense. For the contemporary map, see Peter Ludlow, “Descriptions,” Stanford Encyclopedia of Philosophy, substantive revision 21 September 2022. ↩
4. Bryan Pickel and Zoltán Gendler Szabó, “Compositionality,” Stanford Encyclopedia of Philosophy, substantive revision 3 November 2025. The principle requires specification of the relevant syntax, semantic values, and determination relation; without those auxiliary constraints, compositionality can become formally trivial. Conversational implicatures and many context-dependent contributions do not compose in the same direct sense as literal semantic content. ↩
5. Jeff Speaks, “Theories of Meaning,” Stanford Encyclopedia of Philosophy, substantive revision 31 July 2024, distinguishes semantic theories—which assign semantic properties to expressions—from foundational or metasemantic theories, which explain the facts in virtue of which expressions possess those properties. The book’s Chapters 12–15 operate chiefly at the metasemantic level; the present chapter adds semantic structure, speaker meaning, and force without treating those as rival answers to one question. ↩
6. Donald Davidson, “Truth and Meaning,” Synthese 17 (1967): 304–323, at 310 for the Snow is white test case. Davidson proposed that an empirically adequate Tarskian truth theory could serve as a theory of meaning for a natural language if constrained to yield interpretively acceptable theorems from finite resources. The proposal concerns systematic semantic competence; it does not claim that truth conditions alone supply a complete theory of language use, force, or the metasemantic grounding of primitive expressions. See Jeff Malpas, “Donald Davidson,” Stanford Encyclopedia of Philosophy, substantive revision 28 April 2023, secs. 2–3. ↩
7. Ludwig Wittgenstein, Philosophical Investigations, trans. G. E. M. Anscombe, 3rd ed. (Oxford: Basil Blackwell, 1967; English-text reprint, 1972), secs. 23 and 43. Section 43 carefully limits the meaning-as-use formula to “a large class of cases—though not for all”; Section 23 emphasizes the open-ended multiplicity of language-games and places speaking within activity and a form of life. The chapter therefore uses meaning as use as a diagnostic against private mental cargo, not as a one-clause analysis of every semantic phenomenon. ↩
8. Wittgenstein, Philosophical Investigations, secs. 201–202, 240–242, and 243–258. The rule-following considerations distinguish grasping a rule in practice from indefinitely interpreting one formulation through another. The private-diary case at sec. 258 attacks a proposed criterion whose seeming-correct exhausts correctness; secs. 241–242 distinguish agreement in language and judgments from agreement that votes propositions true. The claim that public practice, causal history, environment, and social relations together determine content forms the synthesis developed in this chapter, not a theory attributed to Wittgenstein. ↩
9. H. P. Grice, “Meaning,” The Philosophical Review 66, no. 3 (1957): 377–388. Grice distinguishes natural meaning from nonnatural or communicative meaning and develops the latter through an utterer’s intention to induce a response by means of the audience’s recognition of that intention. Later Gricean accounts revise the exact reflexive-intention conditions, especially for conventionalized, collective, covert, or noncooperative communication; the main text uses the durable distinction rather than treating Grice’s first analysis as final. ↩
10. H. P. Grice, “Logic and Conversation,” in Syntax and Semantics, vol. 3, Speech Acts, ed. Peter Cole and Jerry L. Morgan (New York: Academic Press, 1975), 41–58. Grice distinguishes what is said from conversational implicature and characterizes the latter through rational reconstruction under the Cooperative Principle and maxims of quantity, quality, relation, and manner. The Smith/handwriting exchange in the main text adapts Grice’s testimonial for “Mr. X” rather than quoting his example. Cancellability, calculability, and nondetachability remain useful diagnostics, not exceptionless definitions of implicature. ↩
11. J. L. Austin, How to Do Things with Words, ed. J. O. Urmson (Oxford: Clarendon Press, 1962). Austin begins from explicit performatives and the distinction between felicitous and infelicitous acts, then weakens the performative/constative dichotomy in favor of a general theory of locutionary, illocutionary, and perlocutionary acts. Statements themselves count as illocutionary performances and remain assessable along dimensions beyond bare truth or falsity, including uptake, authority, and circumstance. ↩
12. John R. Searle, Speech Acts: An Essay in the Philosophy of Language (Cambridge: Cambridge University Press, 1969), chap. 2. Searle distinguishes utterance acts, propositional acts of reference and predication, and illocutionary acts such as asserting, questioning, and commanding. The same propositional content can occur under different forces; conversely, repetition of the same sentence type across contexts need not preserve propositional content or force. ↩
13. Saul A. Kripke, Naming and Necessity (Cambridge, MA: Harvard University Press, 1980), especially the second lecture, 91–97. Kripke’s causal-historical sketch replaces the requirement that each competent user associate a uniquely identifying description with a name. The transmission condition includes an intention to use the name with the reference current in the practice. The dog named Feynman in the main text adapts Kripke’s Napoleon/pet-aardvark example. The sketch leaves substantial questions about failed baptisms, speaker reference, deference, and reference change; it supplies a world-involving constraint rather than a complete algorithm. ↩
14. Hilary Putnam, “The Meaning of ‘Meaning,’” in Language, Mind, and Knowledge, ed. Keith Gunderson, Minnesota Studies in the Philosophy of Science 7 (Minneapolis: University of Minnesota Press, 1975), 131–193. Putnam’s Twin Earth argument directly establishes an extension difference between Earthian water and Twin-Earthian water. His broader account adds syntactic and semantic markers, stereotype, extension, and the division of linguistic labor. The chapter therefore does not slide from difference of extension to a complete theory of meaning without argument. ↩
15. Tyler Burge, “Individualism and the Mental,” Midwest Studies in Philosophy 4 (1979): 73–121. Burge’s arthritis case extends externalist pressure from natural kinds to socially individuated concepts: the subject can possess and misuse a concept whose correct application depends partly on communal linguistic practice. The conclusion concerns content individuation, not an infallible community or a denial of first-person authority. ↩
16. Gareth Evans, The Varieties of Reference, ed. John McDowell (Oxford: Clarendon Press, 1982), chap. 11. Evans distinguishes information-linked producers from consumers who inherit a name through communication. Consumer standing does not require direct acquaintance with the bearer; it requires participation in a practice whose information can ultimately be maintained and corrected. The distinction becomes important for artificial systems because corpus descent may transmit public reference while leaving system-level correction and attribution open. ↩
17. Robert B. Brandom, Making It Explicit: Reasoning, Representing, and Discursive Commitment (Cambridge, MA: Harvard University Press, 1994), especially chap. 4 on language-entry and language-exit moves; Steven T. Piantadosi and Felix Hill, “Meaning without reference in large language models,” arXiv:2208.02957 (2022). Brandom gives inferential role its strongest systematic form while retaining reliable world-facing entry moves; Piantadosi and Hill press conceptual-role meaning directly in the machine case. The twin-practice argument targets the claim that relations among internal states suffice. Brandom’s broader role already includes entry differences; Chapters 12–15’s producer-consumer and intervention tests address the further metasemantic question of what makes one entry route answer to one target rather than another, without appealing to any criterion of practical stake. ↩
18. Matthew Mandelkern and Tal Linzen, “Do Language Models’ Words Refer?” Computational Linguistics 50, no. 3 (2024): 1191–1200, https://doi.org/10.1162/coli_a_00522, argue that externalist causal chains may extend through training corpora to language-model tokens. Dimitri Coelho Mollo and Raphaël Millière, “The Vector Grounding Problem,” Philosophy and the Mind Sciences 7, no. 1 (2026): 1–28, https://doi.org/10.33735/phimisci.2026.12307, distinguish several kinds of grounding and defend routes to referential grounding for trained systems. Both lines block a manufacture-based denial of reference; neither makes reference sufficient for integrated understanding or phenomenal consciousness. ↩
19. Hilary Putnam, Reason, Truth and History (Cambridge: Cambridge University Press, 1981), chap. 1, stages the always-envatted brain and argues that its terms cannot refer to brains and vats in the ordinary way because the required causal relations never formed. The recently-envatted contrast follows from standard externalist continuity rather than from Putnam’s compact presentation. Reference to stable virtual objects or feed structures, and indeterminacy where rival mappings remain tied, are extensions of the argument made here rather than claims attributed to Putnam. ↩
20. Burge’s arthritis case shows that a speaker may possess a socially individuated concept while misunderstanding part of its extension; see Burge, “Individualism and the Mental,” 77–79. For the machine case, Vladimír Havlík, “Meaning and Understanding in Large Language Models,” Synthese 205 (2025): article 9, https://doi.org/10.1007/s11229-024-04878-4, argues that minimal semantic content and semantic fragmentism support attributing natural-language understanding to LLMs without making awareness a condition of understanding. I accept the pressure against treating the achievement as mere simulation, but not the stronger claim that current models thereby duplicate integrated human understanding. ↩