A parrot can say “fire” without warning anyone of one, and a road sign can direct traffic only within a practice that gives its marks a rule. Some things mean on loan. Philosophers call this derived intentionality and contrast it with original, or non-derived, content: content whose correctness standard is not exhausted by an interpreting practice.

Original here means neither novelty nor self-creation. Language models can compose sentences no one has written before, and inherited vocabulary does not bar a learner from acquiring world-answerable uses its teachers never fixed. The question asks what now fixes the standard that makes a state accurate or mistaken, not whether the output appeared before or the system made itself. Novelty belongs to the product; originality concerns the relation among history, present use, world, and standard.

Across thermostats, animal nervous systems, trained mechanisms, inscriptions, and whole minded systems, the question stays the same: what fixes the standard by which the state gets things right?

---

14.1 The Thermostat Objection: Registration Without Perception

The thermostat Chapter 13 left on the wall has a second job. Early functionalists liked it because it made the mystery of mind look solderable: mentality as a sort of complicated thermostatting. The example still tempts the AI conversation.

Calling both a thermostat and a trained model representational leaves a question open: what kind of representation has each earned? The thermostat can carry derived content without perceiving the room. A learned model may do much more than relay a signal. Its richer organization deserves examination, but complexity alone does not tell us whether it perceives anything.

Tyler Burge saw why most cleanly. His Origins of Objectivity runs six hundred pages on essentially one question: what distinguishes a system that merely registers its environment from one that genuinely represents it? Burge treats the distinction as real, principled, and developmental. Birds, frogs, and infants cross it; thermometers, weather vanes, and detection circuits do not. The line matters for how we talk about machines.¹

Sensory registration occurs whenever a physical system’s inner state varies systematically with some feature of its surroundings. A thermometer’s mercury column varies with temperature. A photodiode’s voltage varies with light. A frog’s retinal ganglion cells fire when a small dark spot moves across the visual field. These registrations carry what Fred Dretske called natural information about the environment² — the kind Chapter 13 showed comes cheap. Correlation alone has not established perceptual representation.

Perceptual representation requires a further distinction: between the stimulation reaching the sensors and the condition of the world beyond them. Burge calls that achievement objectification. In its own processing, a system must treat the proximal energy hitting its sensors as evidence about something beyond them. That distal condition could stay the same while the sensors change, or change while the sensors stay the same. This tests perception, not every form of content: a meaningful road sign need not perceive what it directs us toward.

14.1.1 Constancy as the Mark of Perception

The diagnostic signature of the achievement, Burge argues, is perceptual constancy. Constancies let the visual system hold a feature of the distal world fixed across wildly varying proximal stimulation. A white piece of paper in shadow reflects less light to your eye than a black piece in sunlight. Yet you perceive the first as white and the second as black. Color constancy tells you about the paper’s reflective properties, not the photons currently arriving. A cat ten feet away projects a smaller retinal image than a cat at three feet, yet you perceive the same cat at two distances. Size constancy delivers the cat, not the angular size of its projection. Alison Springle puts the point cleanly: perceptual systems exhibiting constancy achieve a stable perception of actual macro-level physical properties of distal stimuli out of varying proximal stimulation. Burge uses constancy as the mark of perception proper: perception begins where the system starts solving for what’s out there rather than merely responding to what’s coming in.³

Does the thermostat distinguish its own bimetallic strip from the temperature of the room? It does not. Within the device, the strip exhausts the question; the thermostat solves for no distal fact and bridges no proximal-distal gap. Blow warm air directly onto the strip while the room stays cold, and it will report the room as warm. No level of its processing distinguishes the sensor from what the sensor is of. The thermostat instantiates a causal relation that we find useful to describe in representational terms.

John Searle made this last point, and made it with characteristic bluntness. In The Mystery of Consciousness and in the older “Is the Brain a Digital Computer?”, Searle distinguishes between intrinsic intentionality, the genuine aboutness of a thought or perception, and observer-relative intentionality, which we attribute when a system merely admits of a useful interpretation.⁴ His free-mapping argument gets that far: the same physical pattern can be assigned one syntax, another, or none if interpretation alone does the work. It does not show that selection and use can never privilege one content over its rivals. The simple thermostat remains the easy case. Its strip controls the heating within an arrangement we built to regulate room temperature. That gives the state real derived content within our regulatory practice, without making the device a perceiver or establishing an original correctness standard.

14.1.2 The Frog’s Subpersonal Machinery

The frog makes the more interesting case, because frogs and toads possess sensory systems sophisticated enough to make the question serious. When a frog’s tongue flicks at a passing fly, what is the visual system doing? I find this the most instructive case in the chapter, precisely because the answer stays so hard to pin down. The classic ethological work — Lettvin and colleagues on the frog, J.-P. Ewert on the toad — is often read as suggesting something humbling: that the prey-capture circuitry responds to small moving dark contours rather than to flies as such.⁵ That deflationary reading is contested, and the chapter does not need it to win. Ewert’s own worm/anti-worm results show the circuitry responding selectively to a configuration, not a bare moving speck. Burge treats frogs’ distance constancies as evidence that they perceive distal bodies. Whether they represent those bodies as flies, a natural kind, remains a different question. Ruth Millikan anchors representation in the evolutionary function of just such mechanisms.⁶ Burge asks what the animal perceptually objectifies; teleosemantics asks how the frog’s history and consumer economy determine what an internal state can get right or wrong for the frog. The circuit need not become a miniature subject. Its producer-consumer history may give its states subpersonal content, while the animal’s wider organization determines what perceptual and cognitive capacities those states help realize. The next chapter isolates a producer-consumer organization without assuming that this circuit alone achieves objectification. It earns the thinner representational claim without denying perception to the frog.

Suppose a signal carries direction and distance to several consumers: one turns the animal, another aims its tongue. A blocked strike need not make the direction wrong. A lucky catch need not make it right. If the animal turns off target despite catching food, correction of the shared signal can reveal what its history established it to track. Chapter 15 develops that comparison without assuming that every frog actually has this organization.

Burge takes frogs to perceive distal bodies through constancies; he doubts that they represent flies as a natural kind. Our isolated circuit comparison assumes neither achievement. It asks what fixes a signal’s content, while the animal’s perceptual organization answers the further question of objectification. Burge’s and Millikan’s thresholds remain distinct even when an actual animal clears both.

14.1.3 Corpus Invariance and Perceptual Constancy

Bring this back to large language models trained and used only within a corpus, without a perceptual channel to a scene. Their tokens arrive as public signs produced by minds that solved their own proximal-to-distal problem. A model finds invariances across paraphrase, language, and coreference, but these are invariances of a corpus, not evidence of constancies that hold a perceived property fixed through changing stimulation. This corpus-only organization has not established Burgean perceptual objectification, though its mechanisms may carry thinner content. Textual encoding alone settles nothing: sensor readings could also arrive as character strings, while a larger coupled system recovers distal properties across changing stimulation.

The reply is that humans also pattern-match on inputs. But human inputs are proximal, and the visual system uses them to recover distal properties stable across changes in stimulation — properties it can misperceive and whose detection was shaped by engagement with a physical environment. Byrne and Hilbert make the point about color: appearance tracks a distal surface property more closely than the photons currently arriving.⁷ Without that separation of sensing from sensed, there is registration rather than perception.

The simple thermostat fails this distal test. The frog’s circuit can carry subpersonal content within an animal that perceives. Corpus invariances alone do not establish the further perceptual achievement for a model. Chapter 15 asks how selection and causal use can give a mechanism correctness conditions without granting it every capacity of the containing system.

---

14.2 Derived vs. Original Intentionality: Borrowed Meaning

A well-trained language model produces a sentence about a maple tree. The sentence parses, predicts, and pleases. Did the model say something about a maple tree?

A growing line of argument answers yes — not by claiming the model perceives leaves or understands botany, but by locating content at a different grain. Its outputs belong to a public linguistic practice whose token-types have histories and uses. Its internal mechanisms have also been shaped by training. Those facts explain different achievements, and they should not be rolled together. The practice supplies borrowed public meaning; training supplies an acquired causal organization. Neither fact by itself makes the whole model a source of meaning.

Jumbly Grindrod’s “Large Language Models and Linguistic Intentionality” works from Gareth Evans on naming practices and Ruth Millikan on teleosemantics.⁸ Pointing out that no one has shown the model to be conscious misses his target. Grindrod sets sentience and consciousness aside and argues for linguistic intentionality that requires no mind on the speaker’s end. On his account, public meaning lives in the practice, and any device that successfully participates in the practice bears it.

Words really do have public lives. “Water” refers to H₂O whether the speaker can recite the formula; lineage matters. Millikan built her account to capture this: a token gets its content from the cooperative history that selected it.⁹

14.2.1 Three Questions That Keep the Achievements Apart

The trouble starts when a meaningful output-token, a content-bearing mechanism, and a system that understands something get treated as one achievement.

Three questions keep them apart:

1. First, what fixes the content: a public practice, a training history, individual learning, or world-involving consumer use?
2. Second, at what level does the representational explanation operate: token, mechanism, circuit, model, or whole agent?
3. Third, how far does the content reach: a linguistic context, a distal kind, or a particular thing in the world?

No answer on one axis settles the other two. A token can carry public meaning while its producer lacks understanding; a mechanism can represent linguistic contexts while the larger system refers to no particular maple; and a content-bearing component need not become a believer.

14.2.2 What Training Selects

Millikan’s mechanism has two halves. Producers make signs; consumers take them up; selection happens because consumer uptake helps determine which producer-tokens persist.¹⁰ Artificial training can reproduce that pattern. The question is whether a proposed mapping helps explain the history or merely supplies an observer’s gloss.

Fintan Mallory develops the strongest case for a genuine semantic result. On his consumer-based teleosemantics, word-embedding mechanisms acquire original content about linguistic contexts because training selects them to guide context-sensitive prediction.¹¹ He identifies a producer, a consumer, parameter changes under loss, and a contrast between successful and unsuccessful outputs. On the account used here, that described relation warrants original content about word-types and their contextual distributions when it obtains. Specific implementations and transfer to new consumers still require their own evidence.

Mallory supplies more than causal influence. His output is checked against word contexts found in the data, and the mismatch changes both producer and consumer. That gives the proposed content a worldly target. The remaining question asks what fixes the learned state’s accuracy when that target and an assigned verdict part company.

Use the same comparison for a word model and a leak detector:

1. Specify the target and the grain. A distribution of word contexts differs from the things those words name. A coolant leak differs from the maintenance class assigned to it.
2. Trace how the proposed mapping shaped the mechanism and how a consumer uses it. Both sides of a comparison must have selection, correction, and consumer use. A trained system versus an inert chart proves too little.
3. Separate target, verdict, and payoff. Hold the target fixed while a grader changes its answer or a bonus hides failure. Then vary the target while holding those verdicts fixed. Ask which mismatch the established correction route registers and which state it retunes.
4. Follow that correction into consumer use. Does the revised state track the target across new cases, or only the assigned answers? Check the history as well as the response. A broken feedback channel can corrupt an established representation without changing what it represents.

Consider two context learners. Both receive the same word tokens, form embeddings, learn from error, and use the embeddings to predict nearby words. Their ordinary answers match. In one training history, target entries come from counts of the sampled contexts. In the other, they come from a schedule of approved answers that normally agrees with those counts. Now let the schedule favor engine beside cat while the sampled contexts still favor pet. Keep the word key and context window fixed. A count-based mismatch can retune the first learner even if the displayed score praises its old answer. Correction in the second follows the schedule. We have described different sources of correction, not simply counted how often each learner succeeds.

That difference supports distinct readings. The first history supports a claim about actual contexts. The second supports the approved schedule; our further reading of its answer as a claim about actual contexts needs a separate basis. It might even have original content about the schedule. Failure at one grain cannot settle every other one. Nor can a single corrupt update erase an earlier history of tracking contexts. The comparison concerns what established the mapping, not which signal won on Tuesday.

Human practices determine what counts as a word and help choose the corpus. Once those are fixed, a judge cannot make engine common beside cat by approving the prediction. A child can learn that fact about language just as a machine can. Independence from a verdict does not require independence from the people who made the target.

Mallory identifies actual distributions as the source of training error. His account supplies the relevant dependence for the described algorithm: comparison with sampled contexts shapes the producer and consumer. That warrants the first reading at this limited grain. He does not report the matched experiment just proposed, but a fresh experiment supplies no missing semantic ingredient when historical and mechanistic evidence already establishes the relation. The comparison tests doubtful cases; it does not suspend the concession. A given implementation or a new consumer still requires its own evidence, and neither engineered loss nor linguistic ancestry counts against originality.

Relabeling demands a different control. Rename indices and adjust the matrices and word key together, preserving all links between actual words and contexts. The notation changes, not the candidate mapping. Switch which actual word feeds an index without that adjustment, and the input relation changes. Alter token boundaries or the context window, and the target may change too. None of these changes isolates an interpreter’s verdict while leaving the content-fixing relations intact.

The concession remains narrow in reach. Hold the corpus fixed while the maple outside changes, and training on that corpus does not register the change. That limits a claim about the maple, not a claim about distributions in the corpus. The corpus belongs to the world too. Human uses of “maple” acquired wider address through practices answerable to maples. Statistical traces of those practices do not by themselves single out the tree outside my window.

Grindrod’s public-practice argument presses at another level. He argues that the model can consume a naming practice in Evans’s sense and inherit reference without the producer’s demonstrative grip.¹² Grant that type-level inheritance: the practice fixes a public reference, and the model’s token descends from it. What remains unearned by corpus training alone is occasion-level reference to this maple in a shared world. A consumer can inherit such a route through testimony and correction without going to look. We still need to establish that route for this use. Content about word distributions, even if original, does not supply it by itself.

14.2.3 Mechanism Content and Whole-System Understanding

Nothing in that concession makes an embedding a miniature speaker. Neural states can represent edges, motion, or color without believing anything; a frog’s prey circuit can represent prey without becoming a second frog. The same distinction applies to an artificial mechanism. Content at one level does not automatically become belief, understanding, or reference by the larger system that contains it. A gearbox transmits torque without driving to the grocery store. Philosophy occasionally needs the same permission to distinguish a contribution from the achievement it serves.

Whole-system attribution depends on how the larger organization uses the content: integration across learning and inference, conflict resolution, continuing correction by what the state concerns, perceptual presentation, and flexible control. Such conditions may justify attributing a capacity to a whole agent. They do not create the component state’s content after the fact.

The prayer analogy belongs here.¹³ A transcription can preserve the words of a prayer without worshipping. Likewise, a mechanism can carry causally active content about textual distributions without the whole model understanding, asserting, or caring about any of it. That attributional limit holds whether the content remains derived or earns independence.

14.3 Dependence Without a Semantic First Mover

Calling content derived does not demand one uncaused first meaning. Every representational system has a causal pedigree. The contrast concerns explanatory dependence: does an interpreting practice exhaust the correctness standard, or do the system’s world-involving producer-consumer history and present use help fix it?

Evolution, learning, culture, and training can all help fix this relation. Both context learners inherit a task. Both learn through error. Their histories differ in what drives that change: actual word occurrences or an approved-answer schedule. The source of their words settles neither case. Mallory’s described producer-consumer relation warrants the limited word-context concession; the content of other implementations or transferred embeddings depends on their history and use.

Dennett offers the strongest reason to discard the distinction. On his real-pattern view, we earn literal talk of beliefs through robust, economical prediction. The pattern must hold across new cases and changes in conditions. No semantic spark or first interpreter needs to enter the story.¹⁴ The book agrees that intentionality can emerge by degrees and that objective patterns need no ghost behind them. It asks one further question. Can actual contexts correct a prediction despite an approving verdict, or does the relevant history track the approved schedule alone? Dennett may treat that contrast as one more feature of the pattern. Nothing here forces him to adopt the original/derived vocabulary. The terms name a dependence that can vary while learning, useful prediction, and real content occur on both sides.

Chapter Summary

The claim. Derived content depends on an interpreting practice; original content is world-answerable in a way not exhausted by one. Originality is compatible with causal and semantic inheritance, can vary by content and grain, and never means self-creation. Selection, learning, and artificial training may establish it when their history fixes a standard rather than merely internalizing one.

Where the argument stands. A system can carry content without perceiving a distal object. The matched learning test asks what drives correction: the target, an assigned verdict, or payoff. Mallory’s described relation warrants original word-context content when it obtains; other implementations and transferred uses still require evidence. None of this turns a circuit into a perceiver or a model into a speaker who understands.

What’s left open. Integration, perceptual mode, attention, the distributed consciousness profile, and continuing world traffic may govern which wider system believes, perceives, understands, acts for reasons, or is conscious. They must earn those attributions directly rather than serving as hidden semantic conditions. The original/derived distinction diagnoses one kind of dependence; it does not rank every kind of mind.

The hand-off. A correctness standard can become world-answerable without any system making itself. The next chapter asks how producer-consumer history and a correspondence rule can establish that independence — and how much remains before content becomes a reason in a mind.

---

Notes

1. Tyler Burge, Origins of Objectivity (Oxford: Oxford University Press, 2010), esp. chs. 8–9. Burge’s long argument distinguishes sensory registration — a state’s covarying systematically with an environmental feature — from perceptual representation, which requires that the system itself treat proximal stimulation as evidence about a distal world it could be right or wrong about. The distinction is comparative and developmental: Burge sets the threshold of representation low enough to include many animals and human infants, but well above thermometers, weather vanes, and detection circuits. His disjunction-problem discussion, 323–324, specifically attributes distance constancies and distal-body perception to frogs while questioning perceptual representation of flies as a natural kind. The isolated circuit comparison in the text does not withdraw that animal-level attribution. ↩
2. Fred Dretske, Knowledge and the Flow of Information (Cambridge, MA: MIT Press, 1981), esp. chs. 3 and 7–8. Dretske’s natural information is a nomic-correlational notion, and the dependency runs from signal to source: a signal carries the information that a source is a certain way when, under the relevant conditions, the signal’s occurring as it does is nomically sufficient for the source’s being that way (equivalently, the conditional probability of the source’s being that way, given the signal, is one). The step from carrying information to representing requires something further — for Dretske, the use of a semantic structure in shaping behavior — which is exactly where covariation and representation come apart. ↩
3. Alison Springle, “Perception, Representation, Realism, and Function,” Philosophy of Science 86, no. 5 (2019): 1202–1213, esp. sec. 2, develops the thesis that perceptual constancy delivers a stable perception of macro-level distal properties out of varying proximal stimulation; the main text paraphrases that thesis rather than quoting it. Burge makes constancy the criterial mark of perception proper (Origins of Objectivity, 2010, esp. ch. 9; see note 1): the constancies are the mechanisms by which a system represents an objective, distally located property rather than merely registering proximal stimulation. ↩
4. John Searle, “Is the Brain a Digital Computer?” Proceedings and Addresses of the American Philosophical Association 64, no. 3 (1990): 21–37, esp. 25–29; and The Mystery of Consciousness (New York: New York Review Books, 1997). The distinction between intrinsic (or original) intentionality and observer-relative (or as-if) intentionality is Searle’s. His slogan “syntax is not intrinsic to physics” exposes the freedom of an interpretation that maps formal states onto arbitrary physical patterns. Chapter 22 accepts that objection to unconstrained maps while rejecting Searle’s stronger conclusion: a physical mechanism can realize computational counterfactuals observer-independently. The semantic question remains separate; intrinsic computation does not, by itself, fix content. ↩
5. Karen Neander, A Mark of the Mental: In Defense of Informational Teleosemantics (Cambridge, MA: MIT Press, 2017), esp. ch. 5. The prey-capture case — descending from Lettvin et al. on the frog and J.-P. Ewert on the toad — is the standard battleground for theories of content. It is worth flagging that Neander herself does not use it deflationarily: she argues the relevant states are low-level perceptual representations of visible features (a small, dark, moving thing), not mere registrations, as part of her solution to the indeterminacy problem. The main text needs less than either the “moving dark contour” deflation or Neander’s own reading — only that a mechanism can carry consumer-imposed correctness conditions without clearing Burge’s objectification threshold. Ewert’s worm/anti-worm discrimination cautions in any case against reading the circuitry as a bare moving-speck detector. ↩
6. Ruth Millikan (b. 1933), the American philosopher who built teleosemantics into a systematic theory of content — work the Royal Swedish Academy recognized with the 2017 Rolf Schock Prize — in Language, Thought, and Other Biological Categories: New Foundations for Realism (Cambridge, MA: MIT Press, 1984), esp. chs. 1–2; compactly restated in “Biosemantics,” Journal of Philosophy 86 (1989): 281–297, esp. 284–88. Millikan grounds the content of a state in the proper function its ancestors performed — what the state was selected for — so that a mechanism can misrepresent precisely because it can fail to perform the function that fixes its content. The account is developed at length in Chapter 15. ↩
7. The point belongs to the representationalist and reflectance-physicalist literature on color — Michael Tye, Consciousness, Color, and Content (Cambridge, MA: MIT Press, 2000), esp. ch. 7, 147–51; Alex Byrne and David R. Hilbert, “Color Realism and Color Science,” Behavioral and Brain Sciences 26 (2003): 3–21, esp. 7–12. Color appearance tracks a distal surface property far more closely than it tracks the proximal photon flux currently reaching the eye — which is why color constancy holds the perceived color fixed as illumination changes. ↩
8. Jumbly Grindrod, “Large Language Models and Linguistic Intentionality,” Synthese 204:71 (2024). Grindrod sets aside general intelligence, sentience, and consciousness and recovers meaningful use at a different level: the linguistic practice itself. Drawing on Gareth Evans’s distinction between producers and consumers of a naming practice (Evans, The Varieties of Reference, ed. John McDowell [Oxford: Clarendon, 1982], ch. 11), and on Millikan’s teleosemantic account of conventional signs, Grindrod argues that an LLM can stand in the consumer role even though it lacks the demonstrative-recognitional capacities Evans required of producers. The argument is the strongest version of the LLM-meaning move currently in print; the response developed here grants inherited public type-level meaning while denying that inheritance alone establishes occasion-level reference or whole-model understanding. ↩
9. Ruth Garrett Millikan, Language, Thought, and Other Biological Categories: New Foundations for Realism (Cambridge, MA: MIT Press, 1984), chs. 1–2; and “Biosemantics,” Journal of Philosophy 86 (1989): 281–297. Millikan’s “proper function” is the function a trait has in virtue of the evolutionary or learning history that selected for its predecessors. Applied to signs, the proper function of a token-type is what its ancestors did that made it persist. The water-as-H₂O case is most directly developed in Hilary Putnam, “The Meaning of ‘Meaning,’” in Mind, Language and Reality: Philosophical Papers, Volume 2 (Cambridge: Cambridge University Press, 1975), 215–271; the convergence between Putnam’s causal-historical externalism and Millikan’s teleosemantics is one of the more underappreciated agreements in twentieth-century semantics. ↩
10. The producer-consumer asymmetry distinguishes Millikan’s account from a purely informational starting point, not from a Dretskean theory confined to correlation. Dretske, Knowledge and the Flow of Information (Cambridge, MA: MIT Press, 1981), ch. 8, esp. 192–195, explains how learning establishes an information-carrying role that subsequent tokens can fail to fulfill. Information alone does not establish misrepresentation; his account introduces the additional functional history. Millikan locates the content-fixing work in consumers and the historically Normal conditions for their proper performance (“Biosemantics,” sec. II, 284–290). The bee-dance example is Millikan’s own (Language, Thought, and Other Biological Categories, 96–98). Peter Godfrey-Smith, “Mental Representation, Naturalism, and Teleosemantics,” in Teleosemantics: New Philosophical Essays, ed. Graham Macdonald and David Papineau (Oxford: Oxford University Press, 2006), 42–68, provides the cleanest critical overview. ↩
11. Fintan Mallory, “Teleosemantics for Neural Word Embeddings,” Mind & Language (online 2026), https://doi.org/10.1111/mila.70037, esp. secs. 3–5.1. Mallory argues that word2vec-style embeddings possess genuine, original rather than derived content under a consumer-based teleosemantics. Section 5 treats selection through parameter alteration under loss minimization, distinguishing it from literal survival of unchanged weights. Section 5.1 grounds the probability-distribution interpretation in training and consumer use, not softmax mathematics alone. It distinguishes learned embeddings from their externally keyed one-hot inputs and explicitly treats linguistic distributions as worldly targets. His note 8 already discusses switched encoding vectors. A coordinated permutation may preserve the word-to-context mapping as a reparameterization; changing the external pairing instead changes a candidate content-fixing relation. Neither operation by itself establishes derivation. The chapter accepts original word-type/context-distribution content for the producer-consumer relation Mallory describes, conditional on its obtaining. Specific implementations and transfer to new consumers require their own evidence; a fresh matched experiment is not a missing semantic ingredient for the described relation. The ancestry of the vocabulary alone decides neither, and lack of demonstrative reference to a maple does not answer Mallory’s proposed grain. No ownership condition resolves that dispute. The matched comparison in the text is a proposed evidential control, not an experiment Mallory reports or a necessary capacity for online self-correction. Its teleosemantic basis is Millikan, “Biosemantics,” Journal of Philosophy 86, no. 6 (1989): 281–297, sec. II, 284–290: selected and learned functions, historically Normal conditions, correspondence rules, and the distinction between proper success and accidental payoff. Mallory’s author manuscript, sec. 4, pp. 7–8, and sec. 5.1, pp. 11–13, identifies the actual-data comparison; that is positive evidence of the relevant dependence, not merely numerical geometry. Separately varying corpus contexts, supplied target entries, and the reported score would test which relation explains the learned mapping. Equivalent historical-mechanistic evidence can support the same conclusion without a new intervention. The test neither requires a second owner nor makes cross-task flexibility a condition of all content. ↩
12. Grindrod, “Large Language Models and Linguistic Intentionality,” sec. 4. Millikan’s framework explains content through selection-relevant producer-consumer feedback. Grindrod reads that framework as supporting the model’s participation in a naming practice. The response developed here grants inherited public type-level meaning while denying that inheritance alone establishes occasion-level reference or whole-model understanding. On Millikan’s consumer emphasis, see “What Has Natural Information to Do with Intentional Representation?” Royal Institute of Philosophy Supplement 49 (2001): 105–125. For an intralinguistic route to model meaning, see Vladimír Havlík, “Meaning and Understanding in Large Language Models,” Synthese 205, art. 9 (2025). ↩
13. The vehicle/content distinction at work here echoes Tim Crane’s deployment of it in “Is There a Perceptual Relation?”, in Perceptual Experience, ed. Tamar Szabó Gendler and John Hawthorne (Oxford: Oxford University Press, 2006), 126–146, and connects to the system-level challenge in Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3 (1980): 417–424. The prayer-transcription analogy is mine. Its limited point is attributional: a component may carry or transform genuine content without making the larger system a worshipper, speaker, or understanding subject. ↩
14. Daniel C. Dennett, “Real Patterns,” Journal of Philosophy 88, no. 1 (1991): 27–51, esp. 32–34 and 42–44; and The Intentional Stance (Cambridge, MA: MIT Press, 1987), chs. 1–2. Dennett’s texts defend realism about predictively indispensable patterns and treat intentional attribution as answerable to more than an interpreter’s whim. The formulation in the main text — especially its emphasis on novelty and counterfactual robustness — reconstructs the evidential lesson for the present argument rather than quoting a criterion Dennett states in those words. ↩