Shannon information is everywhere and cheap. Meaning is earned. The distinction the whole back half turns on.
What this chapter does. A thermostat carries information about the room and understands nothing about cold. That gap — between carrying information about the world and meaning it — is the one the whole back half of this book turns on, and this chapter opens it cleanly before the next three chapters fill it. Carrying information is cheap: thermometers, calculators, and falling barometers all do it. Meaning is earned. The chapter argues that no amount of information-carrying, and no amount of formal symbol-shuffling, adds up to aboutness on its own; it names the precise, engineering sense of information that the looser talk keeps borrowing against; and it shows that even the most disciplined information underdetermines what a state is about. Something more is needed, and naming what that something is occupies the rest of Part Three.
The chapter runs in four moves:
- The gap between carrying information and meaning it (§I). The thermometer and the doctor share the reading and not the understanding; and a bare correlation, unlike a belief, cannot be wrong — the possibility of error is the mark of aboutness.
- The symbol grounding problem (§II). Symbols defined only by other symbols never make contact with what they are about — Harnad’s dictionary-go-round, the divide between intrinsic and borrowed meaning, and why piling on more symbols never breaks out.
- Information in Shannon’s sense (§III). The precise, semantics-free notion that built the digital age — bits, not aboutness — and Dretske’s serious attempt to build meaning from it, defeated by the way information underdetermines what a state is about.
- Why the gap is generic (§IV). What follows for a brain said to “process information” and a machine said to “represent the world” — both quietly trading on a word this Part will not let them keep for free.
There is a thermostat on the wall of nearly every room I have ever worked in. It reads the temperature. It compares the reading to a set point. If the temperature falls below the set point, it turns the heat on; if it rises above, it turns the heat off. The thermostat, in a loose and ordinary sense, “knows” what temperature the room is. It “responds” to the room’s state. It “represents” the temperature in its internal register.
Does the thermostat understand what cold means?
The question sounds almost too simple to ask. Of course it doesn’t. It responds to temperature the way a door responds to being pushed — mechanically, without any understanding of what it is doing or why. The thermostat has no idea that warmth matters, that cold can be dangerous, that the people in the room have preferences about their thermal environment. It processes a signal and produces an output. Nothing is home — and about a thermostat almost no one will dispute that, which is exactly what makes it a clean place to begin.
But here is the puzzle: what exactly is missing? The thermostat carries information about the temperature; it represents the temperature in some minimal sense. If that minimal sense were all that mind required, the thermostat would have a mental state and would understand what cold means. It clearly doesn’t. So information-carrying is not enough, and something more distinguishes genuine representation — the kind that constitutes understanding — from mere signal processing.
This chapter is about what that something more consists in.
§I. The Gap Between Carrying Information and Meaning It
Consider the difference between a thermometer and a doctor taking a patient’s temperature. The thermometer reads 102.6 degrees Fahrenheit. The doctor reads the same number. Both have, in some sense, the same information. But what the doctor has, and the thermometer lacks, is an understanding of what 102.6 degrees means — what it implies about the patient’s condition, what it calls for in terms of response, how it fits into a broader picture of health and illness. The thermometer carries a signal. The doctor understands a situation.
The gap between these two has nothing to do with information. They have the same information, in the narrow sense. The gap lies in what philosophers call intentionality — the reaching-beyond-itself toward the world that Chapter 1 identified as the mark of the mental. The doctor’s belief that the patient has a fever is about the patient’s condition in a way that the thermometer’s reading is not about anything. The belief constitutes a mental state that takes the patient’s fever as its object; the thermometer’s reading constitutes a physical state that correlates with the patient’s temperature. Correlation is not yet aboutness — and what must be added to a bare correlation to make it reach its object will occupy the chapters ahead.
There is a sharp test for how far the two come apart, and it turns on the possibility of error. A thermometer cannot misread. Let its mercury column separate, and it will show 180 while the water boils — but the 180 is not a falsehood the thermometer is telling about the water, a claim it gets wrong. It is simply a new state of the glass, covarying now with something other than the water’s heat. The thermometer has no stake in getting the water right, so nothing it does can count as getting it wrong. The doctor can be wrong. Her belief that the patient has a fever can be false — and that it can be false is the surest sign that it is about the fever in the first place, for only what reaches toward a fact is the kind of thing that can miss it. Aboutness brings a standard the state can fail to meet; covariation brings no standard at all, only tighter or looser correlation. This is why a correlation, however reliable, is not yet aboutness: a correlation cannot lie, and a representation is precisely the kind of thing that can.1 What a mind adds to a covarying signal is the very possibility of misrepresenting — and with it the whole normative dimension that makes a state answerable to the world rather than merely caused by it.
So the question is not how to get more information into the thermostat. It is how aboutness arises at all — how a physical state comes to reach beyond itself and be answerable to what it reaches for. That is among the deepest questions in the philosophy of mind, and the rest of Part Three is the attempt to answer it.
§II. The Symbol Grounding Problem
Stevan Harnad — a Hungarian-born cognitive scientist who founded the journal Behavioral and Brain Sciences and spent a career on how minds attach symbols to the things they pick out — put the problem sharply in 1990 and gave it a name: the symbol grounding problem. Imagine looking up an unfamiliar word in a Chinese-Chinese dictionary. You find the entry, but the definition consists entirely of other Chinese characters you don’t know. You look those up. More characters. You look those up too. You could spend forever in the dictionary without ever making contact with anything the characters actually mean — because meaning, Harnad argued, cannot arise from symbols that are defined solely in terms of other symbols.2
The point generalizes, and the generalization is what makes it dangerous. Any purely formal system — one that manipulates symbols by their shapes, according to syntactic rules — faces the same wall. Pile on more symbols and you have only lengthened the dictionary; sharpen the rules and you have only quickened the shuffling. Neither move reaches outside the system to the things the symbols are supposed to be about, because nothing inside a rulebook for transforming shapes is a shape’s being about anything. Whatever meaning such a system seems to have is meaning we lend it from outside — the meaning in our heads, projected onto its tokens because we can read them. The meaning a system has on its own, owing nothing to an outside reader, is what philosophers call intrinsic; the meaning it merely borrows from us is derived, or parasitic. A formal system has, at most, the derived kind.3 (That distinction — borrowed meaning against the original sort — is the hinge of the next chapter, which presses the thermostat until it yields the difference.)
Harnad’s own remedy points where Part Three is headed. Ground the elementary symbols bottom-up, he proposed — in a system’s sensorimotor traffic with the world, its seeing and reaching and sorting — so that the higher symbols built from them inherit a connection to what they pick out. Whether such grounding is enough for meaning, or only necessary for it, is a question the rest of this Part will press, and not concede cheaply. The lesson to carry forward is the firmer, narrower one: meaning requires something a closed system of symbols cannot supply from within — a connection to the world the symbols are about that is not itself merely more symbols.
This restates, in computational dress, the point Franz Brentano pressed in 1874 — the mark of the mental from Chapter 1: intentionality, the way a mental state stands always of or about something, directed beyond itself toward an object.4 Formal symbol systems perform no such reaching; they shuffle shapes by rule, and shuffling is not directedness.
What provides the ground that genuine directedness requires? Before the next chapters answer, one more confusion has to be cleared — because the word that has been doing the most work in this chapter, information, turns out to name two different things.
§III. Information in Shannon’s Sense
A single word has been carrying two loads, and the time has come to separate them. When an engineer says a channel carries information, the word means something exact and quantifiable. Claude Shannon — the Bell Labs mathematician whose 1948 paper founded the discipline and, with it, the digital age — defined information as the reduction of statistical uncertainty across a channel: how far a signal narrows the range of what you did not yet know, measured in bits. It is a magnificent theory, and it built the machine you are reading this on. But Shannon drew its boundary himself, almost at the outset: the semantic aspects of communication, he wrote, are irrelevant to the engineering problem.5 Whether the message means anything, whether it is true, whether it concerns the weather or the war — none of it enters the mathematics. A page of random noise and a page of Shakespeare can carry the very same number of bits.
So information in Shannon’s sense is not aboutness, and never pretended to be. A thermostat carries all the Shannon information you like about the room and means nothing by it. This is the cheap thing the chapter began with, now named precisely: covariation measured in bits, with the question of meaning set to one side by definition.
Could meaning be built from that raw material — aboutness assembled out of information, given enough further structure? The most serious attempt is Fred Dretske’s, and it deserves naming here because it is the version that does not cheat. In Knowledge and the Flow of Information (1981), Dretske took Shannon’s notion as the floor and tried to raise genuine content on it: a state carries the information that the world is thus-and-so when it lawfully covaries with the world’s being that way, and a learning history selects, from among everything a state covaries with, the one condition the state comes to mean.6 The theory matters here precisely because it concedes the chapter’s point in the act of advancing it — granting that bare information is not yet meaning, and that a further condition has to be added to reach content.
And the charge that bites a theory as careful as Dretske’s is not that it puns on a word. It is that even disciplined, covariational information underdetermines aboutness. One and the same lawful correlation can be read a dozen ways. The state that covaries with a fever covaries just as lawfully with the infection that caused the fever, the inflammation that drives it, the thermometer reading that records it, and the worried frown that follows it — a chain of reliable correlates running outward in both directions along the causal line. Nothing in the correlation picks which link the state is about. Something beyond covariation must do the picking — a function the state was selected to serve, a use an organism puts it to, a history that makes one correlate its target and the rest mere company. Naming that something is the work the next chapters inherit. Shannon gave us a perfect theory of the channel; he did not give us, and never claimed to give us, an account of what travels through it meaning anything.
§IV. Why the Gap Is Generic
The thermostat and the dictionary look like special cases — a dim little device, a closed book. They are not. The same gap opens under two of the most ordinary things we say, one about brains and one about machines, and each trades on exactly the word §III just pulled apart.
Take the brain first. Open almost any book about it and somewhere in the opening pages you will read that the brain processes information — draws it in through the senses, works on it, ships it out as behavior. The sentence goes down as smoothly as the kidneys filter blood. But look at what actually crosses into the skull. Not the photons; they stop at the retina. Not the pressure waves; they die in the cochlea. What travels inward is ion flux across membranes, a few billion times a second. The neurons carry no little envelopes stamped information. They fire. Between the neurons fire and the brain processes information, a word has been quietly added — and it names something we bring to the brain from outside, a description we find useful, not a stuff we have found inside. Searle made the point sharply: whether a lump of matter counts as processing information holds only relative to an observer who reads its states that way.7 Intrinsic to the physics there sits causal traffic and nothing more. But mark what this leaves alone. It makes computation and information-processing observer-relative; it leaves entirely open whether the brain has meaning of its own, intrinsic to it, that no outside reader confers. That it might, and how a stretch of matter could come by such meaning, stays the open question of this Part. The narrow claim here runs only this far: calling the brain an information processor does not, by itself, answer it — the phrase borrows the very interpreter whose meaning was the thing to be explained.
The machine runs the same trick the other direction. We say a language model represents the world — and at the scale these systems now operate, the claim can feel almost forced upon us. A model trained only to predict the next token defines its tokens in terms of other tokens: Harnad’s merry-go-round, spun very fast over a civilization’s worth of text, and speed is not contact. The harder case, and the one worth taking seriously, is the model trained by feedback rather than by text alone — where the corrective signal arguably reaches past the words to the world that rewards them. Whether that feedback earns genuine grounding or only builds a subtler merry-go-round is the argument of Chapter 20; here I only flag that represents the world, in the easy reading, still describes what we do when we read the output rather than what the system does when it produces it.
So the gap the thermostat opened turns out to be no shortage of information at all. It is the absence of the one thing information was never going to supply on its own — aboutness, the reaching-toward that a covarying signal does not yet do — and no amount of brain-talk or machine-talk closes it by trading on the richer word while having paid only for the thinner one.
The symbol grounding problem names the hole; it does not fill it. To say that meaning requires a connection to the world only relocates the question — what kind of connection, and what does it take to have one? The next three chapters supply the answer, in order. Chapter 15 pushes the thermostat harder and draws the line between meaning that is borrowed and meaning that is original. Chapter 16 asks where original meaning could come from in a physical world, and finds the beginning of an answer in biological function. Chapter 17 follows the word itself out of the skull and into the world it reaches for.
Chapter Summary
Carrying information is cheap and meaning is earned: a thermometer and a doctor share the reading and not the understanding, because aboutness consists in a reach toward the world that mere covariation never makes — and no pile of symbols, and no quantity of Shannon’s bits, supplies it from within. The gap opens alike under the brain we call an information-processor and the machine we call a representer. The chapter names the hole; Chapters 15 through 17 take up what fills it.
Notes
- The possibility of misrepresentation as a constraint on any naturalistic theory of content is a central theme of Fred Dretske’s work, from Knowledge and the Flow of Information (Oxford: Blackwell, 1981) to the sharper treatment in “Misrepresentation,” in Belief: Form, Content, and Function, ed. Radu J. Bogdan (Oxford: Oxford University Press, 1986), 17–36: a state that merely covaries with a condition cannot, by that fact alone, misrepresent it, since under a pure covariation reading the state simply indicates whatever it correlates with and can never get that thing wrong. Genuine representation requires a state that can be false — that has a determinate content it can fail to match — which is why the bare information relation is insufficient for aboutness. The difficulty of recovering error from covariation (the “disjunction problem”) drives the appeal to function and selection history taken up in Chapter 16; for the sharpest statement of how reliable misrepresentation can nonetheless occur, see Angela Mendelovici, The Phenomenal Basis of Intentionality (New York: Oxford University Press, 2018), ch. 5, and “Reliable Misrepresentation and Tracking Theories of Mental Representation,” Philosophical Studies 165 (2013): 421–443. ↩
- Stevan Harnad, “The Symbol Grounding Problem,” Physica D: Nonlinear Phenomena 42 (1990): 335–346. Harnad’s formulation draws explicitly on Searle’s Chinese Room as a demonstration that formal symbol manipulation is insufficient for semantics, and extends the point to connectionist systems as well as classical symbol processors. His proposed remedy — grounding elementary symbols bottom-up in sensorimotor interaction with the world (iconic and categorical representations), so that symbols built from them inherit a connection to what they pick out — anticipates the embodied, world-involving account this book develops in Chapters 16 and 16. Harnad is careful to cast his proposal as an account of symbol grounding rather than a full theory of meaning: grounding is necessary for content without being obviously sufficient for it. The strongest recent challenge to applying his verdict to contemporary systems is Dimitri Coelho Mollo and Raphaël Millière, “The Vector Grounding Problem,” Philosophy and the Mind Sciences 7 (2023) (preprint arXiv:2304.01481), who distinguish five notions of grounding and argue that reinforcement-trained large language models may achieve the referential variety; the present book resists that conclusion and engages it directly in Chapter 20. ↩
- The distinction between intrinsic (or original) intentionality and derived (or as-if) intentionality is Searle’s: a sentence, a map, or a thermostat reading has only the meaning its interpreters confer on it, whereas a thinker’s states mean what they mean on their own account. See John Searle, Intentionality: An Essay in the Philosophy of Mind (Cambridge: Cambridge University Press, 1983), 6, and The Rediscovery of the Mind (Cambridge, MA: MIT Press, 1992), ch. 5, where the intrinsic/derived/as-if taxonomy is set out in full. Harnad’s “symbol grounding” picks out the same divide from the engineering side. The chapter uses the distinction only to mark the gap; the question of how a physical state could possess intentionality of the intrinsic kind is the explicit business of Chapter 15, which presses the thermostat until it yields the difference. ↩
- Franz Brentano, Psychology from an Empirical Standpoint (1874), trans. Antos C. Rancurello, D. B. Terrell, and Linda L. McAlister (London: Routledge, 1995), esp. Book II, ch. 1, where intentionality (“the intentional inexistence of an object”) is identified as the mark distinguishing mental from physical phenomena — the thesis introduced at length in Chapter 1. Brentano’s formulation seeded both the phenomenological tradition (through Husserl) and the analytic intentionalism this book inhabits (Crane, Tye, Harman, Dretske); for the development and the subsequent misreadings, see Tim Crane, “Brentano on Intentionality,” in The Routledge Handbook of Franz Brentano and the Brentano School, ed. Uriah Kriegel (London: Routledge, 2017), 41–48. The point recruited here is narrow: a formal symbol system exhibits none of the directedness Brentano made criterial, so symbol-shuffling alone cannot constitute a mental state. ↩
- Claude E. Shannon, “A Mathematical Theory of Communication,” Bell System Technical Journal 27 (1948): 379–423, 623–656. Shannon measures information as the reduction of statistical uncertainty across a channel; in the paper’s opening he is explicit that the “semantic aspects of communication are irrelevant to the engineering problem.” The technical sense of information and the ordinary sense of meaning are two different notions wearing one word — a conflation this Part works to undo, and one Chapter 24 takes up in full. ↩
- Fred Dretske, Knowledge and the Flow of Information (Oxford: Blackwell, 1981), is the most serious attempt to build a theory of semantic content on Shannon’s foundation: a signal carries the information that s is F when the conditional probability of s‘s being F, given the signal, is 1 (while less than 1 on the prior probabilities alone), and Dretske adds a learning period in which a state acquires its meaning by being recruited to indicate one of the many conditions it covaries with. The view is refined in “Misrepresentation” (cited n. 1) and Naturalizing the Mind (Cambridge, MA: MIT Press, 1995). Burge singles it out as the most rigorous development of “information-theoretic notions alone” (Origins of Objectivity [Oxford: Oxford University Press, 2010]). The underdetermination objection pressed here — that lawful covariation runs along an entire causal chain and so cannot, by itself, fix which link a state is about — is the standard pressure point that motivates the teleosemantic supplement (function, selection history) developed in Chapter 16, where Dretske’s project is taken up at the place where it succeeds and the place where it needs more than information can give. ↩
- John Searle, “Is the Brain a Digital Computer?”, Proceedings and Addresses of the American Philosophical Association 64 (1990): 21–37. Searle argues that computation and information-processing are observer-relative: whether a physical system counts as computing, or as carrying information about anything, depends on an interpreter who assigns its states that role, because syntax is not intrinsic to physics. The brain’s intrinsic operations are biophysical — ion flux, neurotransmitter release, membrane potentials — and “information processing” is a description we find predictively useful rather than a feature discovered in the tissue. The point is taken up at length in Chapter 21; here it does narrower work, showing that “the brain processes information” cannot ground meaning because it already presupposes the interpreter whose meaning was to be explained. ↩