A physicalist owes an account of how a physical state can acquire a correctness standard not exhausted by what an interpreter assigns it. The account need not locate a first meaning without a cause. The most developed attempt runs through biology: teleosemantics, from the Greek telos, purpose or end. Meaning is read from what a state is for. A frog’s eye may represent a fly, airborne prey, or something nutritive at a location because of the work its signals do in the frog’s evolved economy. Which description the signal earns is part of the problem, not a premise. Teleosemantics has its strength and its limit at the same point: function can constrain content without always fixing its exact grain.

---

15.1 Why Causal Grounding Is Not Enough

Start with the most direct answer: causal connection. A mental state represents water because water reliably causes it and it, in turn, causes water-seeking behavior. On this proposal, a representation gets its content from what it tracks along a causal chain from object to state.

Fred Dretske developed this idea most rigorously. On his account, a signal carries information that an environmental condition obtains when, given the relevant background, its occurrence guarantees that condition under the laws of nature. The guarantee runs from signal to source: water causing a signal does not suffice if other things produce the same signal too. Learning then recruits the connection to a use. A belief becomes a belief about water once the system has learned to form that state when water is present and use it to guide behavior.¹

But causal connection alone does not suffice. A calibrated thermometer tracks temperature precisely without acquiring a mental state about it. Reliability does not by itself generate aboutness. Dretske therefore gave learning history the work of selecting content. Millikan went deeper.

She proposed biological function, specifically the function a producer-consumer organization acquired through evolutionary or developmental history.² Causal sensitivity and selected use together help determine what a state can get wrong. Return to Chapter 14’s frog. Nothing here reverses Burge’s verdict that its detection circuit may fail to perceptually objectify a stable distal fly.³ Teleosemantics asks a thinner question: can the signal represent what its consumers need and therefore misrepresent a BB? The claim must earn a prey-directed correctness condition; its exact grain remains unsettled.

---

15.2 The Frog That Couldn’t Be Wrong

So: the frog flicks its tongue, and this time it catches a small black BB someone fired across the pond. It swallows, looks satisfied, and waits for the next. The biologist writes the frog mistook the BB for food; the philosopher, watching from a folding chair, writes did it, though?

This is a paradigm of occasional misrepresentation: was this tokening accurate? The easy gloss says the state means fly rather than fly-or-BB because ancestral uses succeeded when flies, not BBs, were present. The example earns less. It may establish only moving nutritive prey at this location, with fly as our convenient label. The orthodox story excludes the BB branch only if pellets played no role in stabilizing the mechanism and satisfy none of the normal conditions under which its consumers succeed. Taxonomic grain does not come free. The harder case concerns systematic error: a whole perceptual apparatus reliably presenting the world in some respect as it is not.

That story is teleosemantics. A representation gets its content from the proper functions of the mechanisms that produce and consume it.⁴ The frog’s signal acquired a prey-directed role because tokenings in relevant ancestral conditions guided tongue strikes and feeding in ways that explain why this producer-consumer arrangement persisted. Its neurons know nothing about flies. Neither does selection. The history matters because it explains why this mechanism exists with this role rather than another.

The account extracts aboutness from causation, organization, and time, with no outside interpreter. Catching something nutritive may explain persistence where catching something black does not. Selection narrows the candidates; organization and use must do the remaining work.

Systematic error does not refute teleosemantics. It shows that selection history cannot be the whole story. Evolution explains why states exist and can fail; task functions, tracking relations, and consumer use must help fix their content.

Selection supplies equipment, task, and a history from which failure can be explained. It has not yet supplied a signal that says something. What turns failing into falsehood? And at what level should the representation be attributed?

---

15.3 From Failing to Falsehood

Put the thermostat and the frog beside each other. Both have jobs; both can fail them. The easy contrast says that the thermostat merely relays while the frog represents. A sophisticated thermostat spoils that answer. It may have graded sensors, several downstream policies, predictive control, and a long history of design revisions. Complexity, variation, and feedback do not draw the semantic line. The job must be stated without indicates, accuracy, or content. That is the job description challenge.⁵

Try a constructed comparison before stating the job:

Build two pond controllers matched in detector sensitivity, a signal carrying direction and distance, feeding function, selection history relevant to the strike, and ordinary tongue-strike success. In the Control system, each coordinate pair feeds one tongue trajectory. In the Map system, the same pair also feeds orienting, obstacle avoidance, and deferred pursuit; consumers recombine it with body position and barriers. Block the straight strike. Control stalls. Map preserves the target variable and changes the route. Bias the detector and let a lucky deflection keep delivering prey. Map’s orienting consumer now stops off target; that cross-consumer mismatch retunes the detector even while feeding continues. Jam the tongue and the relation may remain undisturbed despite failure.

The systems match in ordinary feeding, not in everything they would do under intervention. Their additional consumers and histories of use differ. That difference gives us something to investigate; it does not prove that adding parts creates meaning. Map illustrates three features of the proposed job:

1. A producer generates a family of states whose internal differences vary systematically with independently changing distal conditions.
2. Several consumers preserve those differences while selecting different responses.
3. The same state variable remains usable when the task, body position, or available route changes; it combines with separately changing inputs to produce responses never individually selected; and feedback retunes the producer when the state–world mapping drifts, even where some other part of the system rescues the final outcome.

Ward develops this decouplability test for directives. Its descriptive partner asks whether one distal-condition-sensitive variable can guide several possible behaviors rather than merely trigger one.

That description contains no semantic premise, yet it predicts a contrast. A relay tied to one response dies with the link. A family variable available across consumers transfers, recombines, and can be corrected at the mapping while downstream success stays fixed. These tests identify what the state does before deciding what to call it.

Millikan’s correspondence rule names the structure that unifies those counterfactuals.⁶ Each producer variant maps onto a distal condition, and that mapping figures as a Normal condition for the proper performance of several consumers — “Normal” here means historically explanatory, not statistically common. History privileges the rule that explains why these consumers preserve and exploit the same differences. A token’s satisfaction condition comes from the world-side condition linked to its variant by that historically privileged rule. Nothing semantic entered the job description; mappings that fail its intervention tests never reach this step.

Correctness therefore cannot collapse into payoff. The mapped condition may obtain while the tongue jams, or fail to obtain while a lucky ricochet puts food in the frog’s mouth. The rule also projects to variants no ancestor produced. Selection explains why this mapping governs the signal family; the world determines whether a present token satisfies it.

Jerry Fodor puts the sharpest pressure on the content’s grain. A frog whose state means fly and one whose state means fly-or-any-small-dark-speck fare alike in an ancestral pond where the only specks were flies. Fitness alone cannot choose between them. The question is no longer whether the frog can fail, but what would make fly the thing it can fail to detect.

The consumer side of the loop constrains the answer without dictating a philosopher’s favorite label.⁷ Start with the task function whose performance causally explains why the producer-consumer mechanism stabilized. Then ask which tracking relation figures directly in the mechanism’s performance of that task and how the organism’s wider suite of capacities divides the work. If one circuit targets flying insects while different circuits guide the frog toward worms and other food, fly at this location may provide the most revealing explanation. If the same circuit and the same downstream use treat every small moving nutritive item alike, moving prey may exhaust the content. Taxonomy does not outrank the animal’s organization.

This also answers the gruesome disjunction without treating absence from the ancestral pond as magic. Fly-or-speck is excluded only if the speck branch figures in no normal condition of consumer success, lacks the fly branch’s nomological and counterfactual support, and improves no causal explanation of the system. Mere historical non-occurrence does not exclude it. Edible moving thing, prey, and fly pose the harder empirical choice. The supported content is the property, or structured family of properties, that makes a difference to performance across actual and relevant counterfactual tasks. If two candidates survive every consumer-relevant intervention equally well, content remains indeterminate at that grain. A frog need not carry a sharper concept than its life can use.⁸

Chapter 13 distinguished having a correctness condition from fixing its grain. A tie between prey here and fly here leaves the BB token false on either reading. A tie between nutritive moving prey and small dark moving thing matters more, since the BB counts as an error only on the first. Intervention can break the tie: hold darkness fixed and vary nutritiveness, then reverse the test. Which correlation makes a difference to consumer success is a counterfactual fact about the animal, not our description of it.⁹ Where no asymmetry exists, the account accepts indeterminacy.

A tie that survives every consumer-relevant intervention lies below every discrimination the wider organization can use, and therefore below any phenomenal difference on the identity claim. Mechanism-level content supplies the semantic ingredient; perceptual mode, integration, and poising locate the larger world-presenting state.

15.3.1 Radical Enactivist Objection

Hutto and Myin press two claims. Much basic cognition does not need content: catching, reaching, navigating, and prey capture may consist in historically shaped organism–world coordination. And teleosemantics cannot create content from covariance plus proper function. Those facts explain sensitivity and malfunction, they argue, but never a condition of satisfaction. Calling the normal relation a correspondence rule merely gives coping a semantic title.¹⁰ They are right about the one-link controller. It needs no representation, and selection does not care about truth.

15.3.2 Reply: What the Correspondence Rule Explains

Return to Map’s lucky catch. Feeding continues while the orienting error retunes the shared signal. A description in terms of food obtained alone misses that correction; a richer contentless description might capture it.

Map’s portability, recombination, and mapping-targeted correction distinguish it from direct control and satisfy Ward’s job description for representational explanation. A radical enactivist can still describe the same common variable as part of a richer control organization. If that description preserves every consumer-relevant counterfactual, content has not earned its keep.

The remaining question asks what best explains the invariant. Several changing behaviors remain answerable to one distal condition; history privileges that relation; and whether the condition obtains can come apart from success in any response. I take that organized relation to constitute a satisfaction condition because it unifies portability, recombination, and correction under one rule. Hutto and Myin can call the same organization teleosemiotic rather than semantic. The physical account no longer borrows a semantic premise; the disagreement now concerns whether its history-privileged rule warrants the semantic verdict.

The frog does not discharge Chapter 7’s debt about pain and mood. Apply the same test to that chapter’s proposed pain content: a threat at this bodily site, bad for the subject. A history of location-sensitive protection could fix the bodily address. The relevant condition concerns the threatened tissue, not whether withdrawal happens to succeed. Protection can fail despite an accurately located threat; a reflex or another person can prevent damage despite a mistaken location. Phantom pain then preserves an established bodily address without the corresponding limb. This explains how the address could be wrong, not how much the patient suffers.

The evaluative dimension needs its own constraint. Why does the state present a threat as bad rather than merely locate it? On the proposed account, its history and use must distinguish conditions that impair bodily functioning from harmless changes, and assign them different priorities across attention and protection. A mere increase in avoidance would not establish that content. The account needs evidence that this mapping explains the organized response better than a neutral damage signal followed by aversion.

Mood makes the unfinished work more visible. For Chapter 7’s body-and-field account, candidates include bodily depletion and a surrounding situation offering few manageable possibilities for action. Their accuracy would depend on the body and available possibilities, not on how convincingly the day feels bleak. Which candidate a mood represents requires evidence about its history, use, and correction. The general theory supplies a test; it does not yet fix every mood’s diffuse content. That remains a substantive limit on the promised naturalization, not a reason to make felt character fundamental.

15.3.3 Content Before Agency

Nothing in that verdict requires a free chooser. A frog’s subpersonal mechanism may carry a correctness condition though the frog cannot compare a mortgage with a daughter’s last year at home. A human deliberator may use words inherited from parents, teachers, and a public practice. Semantic inheritance does not disqualify the reason, and freedom does not retroactively mint the content.

The order runs the other way. First a world-involving causal-functional history fixes what a state gets right or wrong. Then integration may make that content available to a persisting subject across memory, comparison, correction, planning, and action. When differences in the represented consideration can bend conduct through that economy, the content functions as a reason. Chapter 18 calls the resulting graded capacity reasons-responsive agency. Moral answerability adds a further relation: a standing to receive a demand, understand it, and revise. Correctness, reason, agency, and answerability belong on four lines, not in one definition.

Making free will necessary for original content would close a circle around the theory. Free agency requires reasons; reasons require content; content would then require free agency. The circle explains nothing. The compatibilist lesson applies on both sides without making the sides identical. Causal inheritance does not prevent conduct from belonging to an agent when causes work through the agent’s reasons-responsive organization. Semantic inheritance does not prevent a state from acquiring world-answerable content when learning and use establish a correction standard no longer exhausted by an external interpreter. Neither achievement escapes causation, and neither requires self-creation.

Free will does not mint content. It describes what an integrated subject can do with content: recognize considerations as reasons, weigh them, act through them, and revise when the world or another person answers back.

---

15.4 Representation and Its Level of Attribution

The last section earned a correspondence rule: a family of states varies in a way its consumers depend on, and the history of that dependence privileges some accuracy conditions over others. It does not yet tell us at what level to attribute the result — to a circuit, or to the animal the circuit serves.

The producer-consumer account already identifies a bearer at the level relevant to its explanation. A frog’s visual circuit can carry content because variations in its states guide consumers whose proper functions depend on a mapping to prey-related conditions. The circuit need not become a tiny frog.^9b Subpersonal representation concerns what a state does within a mechanism; personal-level belief concerns what the organized animal can perceive, learn, infer, and do by means of it.

The same split governs the artificial case. Chapter 14 grants the embedding mechanisms causally active content that may genuinely answer to worldly linguistic distributions. Whether that content counts as original remains unsettled. Either way, the semantic finding does not settle the level of attribution. Even an original representation would not put a speaker behind the vectors.

Integration can alter what we should attribute to the whole. A state used only by one reflexive circuit may support a thin subpersonal content. Coordinated across memory, learning, inference, conflict resolution, and action, the same content may help justify saying that an animal or machine believes or understands something. An external notebook or prosthesis can enter that organization when the coupling does the relevant cognitive work.

What none of it licenses is the leap from one content-bearing mechanism to a whole understanding subject. A representation need not own itself, and a subject cannot be assembled by changing the noun on every useful component.

15.5 What Embodiment Contributes

Descartes treated the body as incidental to the mind. The res cogitans, the thinking thing, is pure intellect attached to a body that brings hunger, fatigue, and misperception. Its essence remains independent, and in principle it could exist disembodied.

This gets things backwards. Embodiment supplies much of what makes a mind possible. Here embodiment names a functional achievement, not a carbon clause. A biological body realizes it one way; an artificial or distributed system could realize it another if the relevant loops actually close.

1. Embodiment first supplies sensorimotor traffic: cycles of perception and action through a resistant environment in real time. Reach for a cup and vision guides the hand, proprioception registers the arm, touch registers contact, and movement changes what vision receives next. Perception and action form one loop, with the world running through it.
2. It may also supply autonomy, when that traffic helps maintain the organization doing the sensing and acting. An organism lives exposed: it can be depleted, damaged, or killed. Reciprocal self-maintenance may help explain why the organism counts as an individual agent and where its practical interests lie.^9a It does not supply the content of its states.
3. Embodiment supplies a history of encounters that tunes sensitivities and fixes addresses. How you understand a forest path turns partly on paths that resisted and yielded, on falls, recoveries, and learned footing. A nonbiological system could acquire an equally consequential history through other sensors and effectors. What matters is not skin but traffic that selects among possible readings of its states.
4. Finally, embodiment supplies a located point of view: a here from which things can be near or far, reachable or blocked.¹¹ Not every later user of a name must stand where its referent once stood. I can refer to Napoleon because testimony and deference pass along a handle that someone in the chain acquired through contact. Chapter 16 asks what that chain must contain. For now, the locating work need occur somewhere in the practice, not under every speaker’s feet. A functionally embodied system may inherit reference through a community and acquire new addresses through its own traffic. It cannot conjure a distal particular from internal form alone.

---

15.6 Swampman

One objection cuts deeper, and it arrives dripping. Donald Davidson’s Swampman is a lightning-struck duplicate assembled by cosmic accident: molecule for molecule yourself, his visual cortex and color circuitry matching yours to the last synapse, but with no history whatever.¹² At the first instant his organization and dispositions match yours. He lacks any etiological or learning history that could privilege one content over another. Swampman can preserve himself and still lack teleosemantic content at t = 0.

Strict Millikan and strict Dretske make history necessary for content. This chapter accepts the price. Present contact can begin a causal history, but it cannot retroactively supply one. Swampman’s first state may covary with a tomato and cause reaching, sorting, and eating. Until learning recruits such relations into functions, they remain causal dispositions rather than representations. His consequential traffic can begin the history from which content later emerges.

Duplication settles neither content nor identity. A manufactured copy of a selected artifact may inherit a function through its lineage; an accidental duplicate inherits none merely by matching the structure. Copying makes succession wonderfully ugly. Philosophers managed that well enough before the lightning arrived. What matters for content is causal provenance and producer-consumer history, not whether the resulting system repairs itself.

Under strong representationalism, a system with no content of the right perceptual kind has no phenomenal character. At t = 0 there is nothing it is like to be Swampman, a phenomenal externalism Dretske accepted explicitly.¹³ This follows from the account; it supplies no independent evidence for it. The internalist can instead make phenomenal content supervene on present organization, but that gives up the book’s wide answer to Inverted Earth. I accept the other horn: complete phenomenal content includes the missing history. Millikan and Neander resisted the common Swampman verdict with open eyes; intuition does not decide the issue.¹⁴

The comparison with a trained language model now concerns history, not manufacture. Swampman lacks a content-fixing history. A text-trained model has a training history that may make its submechanisms answer to actual linguistic distributions. Those distributions belong to the world, even though public practice produced the words. Originality at that grain remains unsettled; inherited materials do not decide it. Model-level understanding remains open too. A history that establishes linguistic-context content does not by itself integrate it into belief, planning, reference to particulars, or a point of view.

Chapter Summary

The claim. A family of signals, its correspondence rule, and its producer-consumer history can establish original content at the grain the explanatory organization supports when that history fixes a world-answerable standard not exhausted by an interpreting practice. This independence requires no semantic self-creation.

Where the argument stands. Selection, learned sensitivity, and consumer use explain how a state can get something wrong and how interventions can choose among candidate contents. A frog’s subpersonal circuitry may represent prey without becoming a separate subject. Chapter 14’s narrower verdict on embeddings carries over: content may genuinely answer to linguistic distributions, while originality at that grain remains unsettled. Neither conclusion establishes whole-system understanding or distal reference beyond the history that fixes the content.

What’s left open. The general model still owes fuller accounts of pain’s evaluative content, mood’s diffuse accuracy conditions, and why pain itself warrants relief. Integration can make content a reason; reasons-responsive control and moral answerability require more. A state’s perceptual mode and conscious role also remain distinct from the organization that bears it.

The hand-off. A distinctively human reach runs through words, which refer to a world that helps fix what they mean outside the skull. Chapter 16 turns to that reach.

---

Notes

1. Fred Dretske, Knowledge and the Flow of Information (Cambridge, MA: MIT Press, 1981), esp. ch. 3; and Explaining Behavior: Reasons in a World of Causes (Cambridge, MA: MIT Press, 1988), esp. chs. 3 and 7–8. Dretske’s account develops from an information-theoretic foundation: a signal carries information about a state of the world when the signal’s occurrence is nomically sufficient for the world’s being in that state (under certain idealized conditions). The move from information-carrying to representing requires the additional step of explaining how information gets put to use — how the organism learns to exploit the information to guide behavior. This move involves the organism’s learning history and is where the connection to teleosemantics begins to emerge. ↩
2. Ruth Millikan, Language, Thought, and Other Biological Categories (Cambridge, MA: MIT Press, 1984), esp. chs. 1–2; and White Queen Psychology and Other Essays for Alice (Cambridge, MA: MIT Press, 1993). The core claim of teleosemantics is that the content of a representation is determined by its proper function — the function it has in virtue of its evolutionary or learning history — rather than by what it actually tracks on any given occasion. This explains misrepresentation: a detector misfires when it responds to a pellet because its proper function is to detect prey, not pellets, even though on this occasion it is causally responding to a pellet. On the frog: the neurophysiology is Lettvin, Maturana, McCulloch, and Pitts’s (1959), and the philosophical use of the frog case in the disjunction debate is canonically Fodor’s (see note 8); Millikan’s own paradigm cases in Language, Thought, and Other Biological Categories and “Biosemantics” are the beaver’s splash, the bee dance, the hoverfly, and the toad’s bug-detector. For a useful critical overview, see 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. ↩
3. Tyler Burge, Origins of Objectivity (Oxford: Oxford University Press, 2010), esp. chs. 8–9. Burge’s constraint is that perceptual representation requires objectification — constancies that present distal properties across changing proximal stimulation. Nothing here contests that threshold for perception. The separate teleosemantic question concerns thinner, subpersonal content fixed by producer-consumer organization; granting such content does not turn the circuit into a perceptual subject. ↩
4. The thesis that content derives from selected function is Millikan’s; see note 2 above for Language, Thought, and Other Biological Categories (1984) and “Biosemantics” (1989), and note 1 for Dretske’s learning-based variant. The revised account adds no ownership condition. Selected producer-consumer function may fix what a mechanism can misrepresent when the proposed mapping, rather than an observer’s gloss, helps explain the mechanism’s stabilization and survives the relevant interventions. ↩
5. William Ramsey, Representation Reconsidered (Cambridge: Cambridge University Press, 2007), ch. 1, poses the job description challenge: say what a state does as a representation rather than merely as a causal relay. Zina Ward, “Directive Representation and the Job Description Challenge,” Philosophy and the Mind Sciences 7 (2026), doi:10.33735/phimisci.2026.12005, argues that a directive earns representational status when its specific contribution remains decouplable from any one behavior. Variation, a consumer-presupposed correspondence rule, and the history that makes that rule explanatory supply the present reply. ↩
6. Ruth Millikan, Language, Thought, and Other Biological Categories (Cambridge, MA: MIT Press, 1984), 96–102; “Biosemantics,” Journal of Philosophy 86 (1989): 281–297, at 283–290; and Varieties of Meaning (Cambridge, MA: MIT Press, 2004), 71–86. Every representation belongs to a set of interrelated representations a producer can send to a consumer, and it is Normal for consumer performance that the set map onto world conditions under a rule. The rule and its explanatory history, not a further owner, make the content non-derived. “Normal” means historically explanatory, not statistically usual. ↩
7. Millikan’s consumer emphasis appears in Language, Thought, and Other Biological Categories, chs. 1–2 and 6, and “Biosemantics.” Nicholas Shea develops an explanatory version in Representation in Cognitive Science (Oxford: Oxford University Press, 2018), especially chs. 3, 4, and 6. Task functions inherit the grain of the causal explanations that stabilize them, and content depends on which correlations play an unmediated role in performance. Kevin J. Mitchell, “The Origins of Meaning: From Pragmatic Control Signals to Semantic Representations,” Philosophy and the Mind Sciences 7 (2026), doi:10.33735/phimisci.2026.11674, traces how action-bound signals become usable across wider cognitive organization. Wider integration may warrant system-level attribution; it does not create the component content. ↩
8. Fred Dretske, Explaining Behavior: Reasons in a World of Causes (Cambridge, MA: MIT Press, 1988), chs. 3–4, and Naturalizing the Mind (Cambridge, MA: MIT Press, 1995), ch. 1, where the merely indicating properties of a state (what it nomically correlates with) and its representational properties (what it has the function of indicating, acquired through a learning history that recruits the indicator to a use) can come apart. The “thickening” image follows Dretske’s account of learned indicator-functions, but no naturalistic theory owes a distinction where rival contents remain explanatorily equivalent across every relevant counterfactual. The frog-and-fly disjunction in its influential form is Fodor’s: see Jerry Fodor, A Theory of Content and Other Essays (Cambridge, MA: MIT Press, 1990), and, on the charge that his asymmetric dependence presupposes the semantic and functional facts it was meant to ground, Peter Schulte and Karen Neander, “Teleological Theories of Mental Content,” Stanford Encyclopedia of Philosophy, substantive revision 26 May 2022, secs. 4.1–4.2. The detailed fly, prey, and disjunction round appears in Appendix B.15. ↩
9. The objection is that an appeal to which correlations play an unmediated role in explanation may track the interests of the explainer rather than a joint in the mechanism. Versions of it are pressed against Shea’s framework by Todd Ganson, who argues the distinction carries no explanatory significance, and by Frances Egan, who is deflationary about the utility of representational explanation generally; both are surveyed in Schulte and Neander, “Teleological Theories of Mental Content,” sec. 4, with Shea’s own reply in Representation in Cognitive Science, ch. 6. The reply pressed in the body does not turn on the explainer at all, and it does not merely observe that the history happened. Which correlation makes a difference to consumer success with the others held fixed is itself a counterfactual fact about the mechanism, so the selection among historical facts that the objection targets is as objective as the facts selected. ↩
10. Mark Bickhard grounds representation directly in self-maintenance in “Representational Content in Humans and Machines,” Journal of Experimental and Theoretical Artificial Intelligence 5, no. 4 (1993): 285–333, and “The Interactivist Model,” Synthese 166, no. 3 (2009): 547–591. Wayne Christensen and Cliff Hooker connect autonomy to anticipative learning in “An Interactivist-Constructivist Approach to Intelligence,” Philosophical Psychology 13, no. 1 (2000): 5–45. Alvaro Moreno and Matteo Mossio’s Biological Autonomy (Springer, 2015) systematizes organizational closure. These remain important theories of autonomy and agency. Closure bears here on autonomy and agency; any semantic work must come from the producer-consumer mapping and its history. ↩
11. Teleosemantic accounts routinely locate producers and consumers below the organism without turning each subsystem into a subject. Nicholas Shea treats vehicles as constituent parts of a larger system in Representation in Cognitive Science, ch. 1; Peter Schulte and Karen Neander make the same allocation explicit in “Teleological Theories of Mental Content,” Stanford Encyclopedia of Philosophy, revision 26 May 2022, sec. 2. Andy Clark and David Chalmers extend cognitive processes across reliably coupled resources in “The Extended Mind,” Analysis 58, no. 1 (1998): 7–19. These cases support a distinction between mechanism-level content and the wider level at which belief, agency, or understanding should be attributed. Nicholas Shea’s unity-of-purpose requirement addresses that later threshold (“Realist Representational Explanations of Agency Should Require Some Unity of Purpose,” Philosophy and the Mind Sciences 7 [2026], doi:10.33735/phimisci.2026.12098). For the artificial mechanism case, see Fintan Mallory, “Teleosemantics for Neural Word Embeddings,” Mind & Language (online 2026), doi:10.1111/mila.70037, esp. secs. 3–5.1. Mallory establishes differential retention, consumer use, and a privileged causal mapping. ↩
12. Daniel D. Hutto and Erik Myin, Radicalizing Enactivism: Basic Minds without Content (Cambridge, MA: MIT Press, 2013), esp. chs. 3–4, argue that covariance and proper function never yield truth-conditional content. The disagreement rests entirely on Section 15.3’s correspondence rule. The rule specifies a satisfaction condition for each signal variant, including unused variants; accuracy can diverge from present success; and producer-consumer history makes the mapping explanatory rather than optional. Hutto and Myin may deny that this suffices. Adding a proprietor would not answer them, because it would identify a candidate subject without specifying what any state says. ↩
13. The argument that embodiment is constitutive of cognition rather than incidental to it draws on a broad tradition in phenomenology and cognitive science. Key sources include Maurice Merleau-Ponty, Phenomenology of Perception, trans. Colin Smith (London: Routledge & Kegan Paul, 1962 [1945]), whose account of the lived body as the ground of perception anticipates much contemporary embodied cognition research; Francisco Varela, Evan Thompson, and Eleanor Rosch, The Embodied Mind: Cognitive Science and Human Experience (Cambridge, MA: MIT Press, 1991); and Andy Clark, Being There: Putting Brain, Body, and World Together Again (Cambridge, MA: MIT Press, 1997). The specifically philosophical argument about what embodiment contributes to semantic grounding is developed in Gareth Evans, The Varieties of Reference, ed. John McDowell (Oxford: Clarendon Press, 1982), esp. ch. 6, whose account of demonstrative thought and spatial cognition shows how perceptual and action capacities underlie the ability to think about particular objects. ↩
14. Donald Davidson, “Knowing One’s Own Mind,” Proceedings and Addresses of the American Philosophical Association 60 (1987): 441–458, at 443–444, where the Swampman case is introduced — not as an objection to teleosemantics, which was not its target, but to press the point that Davidson’s own externalism about content makes causal history essential to what a state means. Teleosemantics inherits the case because it makes a stronger and more specific historical claim. ↩
15. The consequence is phenomenal externalism: if phenomenal character is representational content of the right kind, and content is fixed in part by world-involving relations, then phenomenal character fails to supervene on the intrinsic physical state of the subject’s body taken alone — molecular duplicates can differ phenomenally. Fred Dretske accepted the commitment explicitly and defended its compatibility with physicalist supervenience, once the supervenience base is widened to include the world-involving facts that fix content: Naturalizing the Mind (Cambridge, MA: MIT Press, 1995), ch. 5, “Externalism and Supervenience,” 123–168; and “Phenomenal Externalism or If Meanings Ain’t in the Head, Where Are Qualia?”, Philosophical Issues 7 (1996): 143–158. Alex Byrne and Michael Tye defend the same wide reading against the charge of absurdity in “Qualia ain’t in the Head,” Noûs 40 (2006): 241–255. The physicalism at issue never promised intrinsic-duplicate invariance; it promised no mental difference without a physical difference, and a missing causal history is a physical difference. Ned Block presses the contrary intuition — that felt character must be narrow — as part of the mental-paint package; see “Mental Paint,” in Reflections and Replies: Essays on the Philosophy of Tyler Burge, ed. Martin Hahn and Bjørn Ramberg (Cambridge, MA: MIT Press, 2003), 165–200. Appendix B.9 sets out the pressure and the reply. ↩
16. Karen Neander, “Swampman Meets Swampcow,” Mind and Language 11 (1996): 118–129, argues the intuition should be overturned, not accommodated; Ruth Millikan, “On Swampkinds,” Mind and Language 11 (1996): 103–117, argues that a swamp creature falls under no real (historical) kind, so generalizations from humans to swampmen carry no support — a case her Language, Thought, and Other Biological Categories (cited in full at note 2 above) had anticipated with a swamp-replica of its own (93), three years before Davidson’s. The dialectic, including the “real nature” strategies and Papineau’s actual-world alternative, is surveyed in Schulte and Neander, “Teleological Theories of Mental Content,” sec. 4.2; Appendix B.9 works through the case in full. ↩