Insight Editions · No. 293
The Mechanical Mind
A Philosophical Introduction to Minds, Machines and Mental Representation (Third Edition)
Tim Crane
Crane argues that the mind is a natural causal mechanism whose states represent the world and cause behaviour, and that the deepest problem for this picture is explaining how a mechanism can represent anything at all.
A study guide to The Mechanical Mind by Tim Crane. Shared free for personal reading; not for sale and not endorsed by the authors or publishers. Please read and buy the original.
The Book on One Page
The Mechanical Mind (third edition, Routledge, 2016) is Tim Crane's introduction to the philosophy of mind organised around one question: how can a mere mechanism, a causal system in the natural world, represent the world? He builds the answer from pictures and words through common-sense psychology, computation, the language of thought and connectionism, causal and biological theories of content, externalism and the extended mind, to consciousness. Crane's verdict is that the mind is a causal mechanism with effects in behaviour, but that representation and consciousness must be understood in their own terms, because every attempt to reduce them to something else has failed.
The mind is a causal mechanism, and its thoughts are representations that cause behaviour; but nothing thinks simply by computing, computation presupposes representation, and the best account of content is non-reductive, modelled on measurement and scientific models.
Philosophy of mind and cognitive science from Descartes, Hobbes and Leibniz to Turing, Searle, Fodor, Dretske, Millikan, Putnam, Clark and Chalmers, Nagel and Jackson: representation, intentionality, mindreading, computation, AI, connectionism, content, modularity, externalism, embodiment and consciousness.
Turing's machine (1936) and test (1950); Searle's Chinese room (1980); Fodor's language of thought and Fodor-Pylyshyn (1988); Twin Earth, Otto's notebook (1998), Swampman; the zombie and knowledge arguments.
Anyone who builds or governs AI and wants to know what it would take for a machine's states to mean something, and students who want a clear map of the debates, with arguments laid out premise by premise.
- The puzzle (Chapters 1 to 3). Words and pictures are physical patterns that do not represent intrinsically: resemblance fails as a theory of pictures, pictures cannot express "if", "and", "not", and convention or ideas presuppose representation. Interpretation is bestowed by minds, so mental representation (intentionality) is fundamental. Thoughts are states of the form "A ψs that S", individuated by content; they can be unconscious; and Brentano's thesis survives once we separate original from derived intentionality.
- Thoughts as causes (Chapters 4 and 5). We understand others by reasoned conjecture about the beliefs and desires that make sense of behaviour; behaviourism fails because thoughts connect to behaviour only via other thoughts; Davidson's argument and the mechanical world picture support the causal picture of thought. Common-sense psychology is a theory (or a simulation skill); eliminativism's case for its falsity is weak. Its core: thoughts represent the world and cause behaviour.
- Computation (Chapters 6 and 7). A computer is a device that processes representations systematically according to rules; Turing machines make algorithms unmysterious; but computing a function differs from instantiating one, since computation needs representation. Against AI, Dreyfus says common sense is know-how no rule-base captures; Searle's Chinese room shows syntax is not sufficient for semantics. Crane's verdict: nothing thinks simply by being a computer.
- Mechanisms of thought (Chapter 8). Fodor's language of thought explains the systematicity, productivity and rationality of thought by sentences in the head whose syntax mirrors their semantics and drives their causal role. Connectionism separates the objects of computation from the objects of representation; Fodor and Pylyshyn reply that it either fails systematicity or merely implements a classical mechanism. Either way, computation takes representation for granted.
- Explaining content (Chapters 9 and 10). Reductive theories try to fill "X represents Y iff ___" with natural facts. Causal (indication) theories cannot account for error: the misrepresentation and disjunction problems survive ideal conditions, Dretske's functions (the magnetotactic bacteria are indeterminate) and Fodor's asymmetric dependence. The success theory needs desire content; teleosemantics supplies it from natural selection but is embarrassed by Swampman. Evolutionary psychology is at its best for modules, Fodor's encapsulated input systems; massive modularity undermines itself.
- The extent of the mind (Chapter 11). Twin Earth externalism depends on denying narrow content, which Crane defends; the extended mind (Otto's notebook) is functionalism drawn out; enactivism's insights fit a functionalist "sandwich" unless it insists on our particular bodies. None overturns the mechanical mind.
- Crane's positive view (Chapter 12). Representation is a theoretical concept, assigned by an interpretation function (Cummins's Tower Bridge) as numbers are assigned to weights; contents are abstract objects used to model brain processes. The price: sub-personal states lack unique, essential contents.
- Consciousness (Chapter 13). Phenomenal consciousness (what it is like) is not the same as qualia; the zombie and knowledge arguments make the mind-body problem a precise dilemma: causation drives us towards physicalism, consciousness away. No argument defeats the mechanical mind, but reductive explanation and scientific knowledge have limits.
| Ch | Title | One-line payload |
|---|---|---|
| 1 | Introducing the Mechanical Mind | The mind as a natural causal mechanism; the problem is how such a thing can represent (and be conscious); mechanism is not reductionism. |
| 2 | The Puzzle of Representation | Resemblance fails, pictures cannot express logical relations, words need interpretation: all roads lead to mental representation. |
| 3 | Mental Representation | Thoughts as "A ψs that S" with content; unconscious thought; intentionality vs intensionality; original vs derived intentionality. |
| 4 | Understanding Thinkers and Their Thoughts | Representation survives any mind-body answer; behaviourism fails by holism; Davidson and the causal picture of thought. |
| 5 | Common-Sense Psychology and Science | The Theory Theory, eliminative materialism (Churchland) rebutted, simulation theory; thoughts represent and cause. |
| 6 | Computation and Representation | Algorithms, Turing machines, Church's thesis; a computer processes representations by rules; computing vs instantiating a function. |
| 7 | Can a Computer Think? | Dreyfus's know-how objection and Searle's Chinese room; syntax is not semantics; nothing thinks merely by computing. |
| 8 | The Mechanisms of Thought | Fodor's language of thought, systematicity, syntax-semantics-causation; connectionism and the Fodor-Pylyshyn dilemma. |
| 9 | Explaining Mental Representation | Causal and indication theories; the misrepresentation and disjunction problems; Dretske's functions; asymmetric dependence. |
| 10 | The Mechanical Mind and Its Biological Basis | Success theory, teleosemantics and Swampman; evolutionary psychology's limits; Fodor's modules and massive modularity. |
| 11 | The Extent of the Mind | Twin Earth and narrow content; the extended mind as functionalism; enactive cognition and the sandwich model. |
| 12 | A Non-Reductive Conception of Mental Representation | Representation as a theoretical concept: interpretation functions, measurement, scientific models; no unique contents. |
| 13 | Consciousness and the Mechanical Mind | What it's like, two senses of qualia, zombies and the knowledge argument; the mind-body dilemma; limits of reduction. |
Introducing the Mechanical Mind
The book's question is how anything, and above all a mind, can represent anything at all: how a state of my mind can be "about" Nixon and China. Crane frames it inside the mechanical mind, the view that the mind is a natural causal mechanism working in a systematic, law-governed way. The puzzle is sharp precisely because it is hard to see how a mere mechanism can "reach outside" itself and direct its states upon the world.
The Problem of Intentionality
My belief that President Nixon visited China is about Nixon and China. Crane asks how a state of my mind manages that, and, more generally, what it is for anything to represent anything else. Contemporary philosophers call this the problem of intentionality. It is ancient, but developments in philosophy of mind, linguistics, psychology and artificial intelligence have re-posed it in a new form.
Two live questions hang on it. Whether a computer could think is tied to whether its states can be about anything. And whether there can be a "science of thought" (whether the mind can be explained by science, or needs its own non-scientific mode of explanation) depends, Crane will argue, on the nature of mental representation.
The mechanical mind is the view that the mind is a causal mechanism, a natural part of natural organisms, functioning in a systematic and regular way. Representation becomes a problem for it because machines push and pull; they do not obviously point at things.
From the Organic to the Mechanical World Picture
The idea that the mind is a mechanism grows out of thinking of nature itself as a mechanism, a view born in the Scientific Revolution of the seventeenth century with Galileo, Francis Bacon, Descartes and Newton.
Before that, in the Middle Ages and Renaissance, the world was understood in organic terms. Leonardo da Vinci could write that the earth has a "vegetative soul", that its flesh is the land, its bones the rocks, its blood the pools of water, and its breathing and pulse the ebb and flow of the sea. This organic picture owed most to Aristotle, so dominant he was simply called "the Philosopher". In his system everything had a natural place and acted to reach it: stones fall because their place is the ground, fire rises to the heavens. Every thing had a final end or goal, a picture in harmony with a universe whose ultimate driving force is God.
In the seventeenth century this collapsed. Explanation by final ends and "natures" gave way to explanation by the regular, deterministic behaviour of matter in motion, and the method of discovery switched from interpreting Aristotle to observation, experiment and precise mathematical measurement. Galileo said the "grand book" of the universe is "written in the language of mathematics", its characters triangles, circles and other figures.
| Feature | Organic world picture | Mechanical world picture |
|---|---|---|
| Era | Middle Ages and Renaissance | Seventeenth century onward |
| Authority | Aristotle, "the Philosopher" | Observation, experiment, measurement |
| Why things move | To reach their natural place or final end; God's will | Caused to move according to laws of nature |
| Explanatory model | The inorganic understood as if organic (the earth as an animal) | The organic understood as if inorganic (the body as a watch) |
| Language of nature | Natures, purposes, souls | Mathematics: equations and laws |
| Ruling metaphor | The animal | The watch |
What "mechanical" means here
Crane uses "mechanical" in a deliberately broad sense. Historically the word meant something narrower: systems that interact only on contact and deterministically. Later science refuted that narrow picture twice over: Newton's gravitation apparently acts at a distance, and fundamental physical processes turn out not to be deterministic. Neither discovery touches the general picture of a world of causes working according to natural laws or regularities, and that general picture is all Crane means.
- Narrow (historical) sense. Contact-only, deterministic interaction of matter in motion; refuted by action at a distance and quantum indeterminism.
- Broad (Crane's) sense. Things happen because they are caused, in accordance with natural laws or regularities; untouched by those refutations.
Animals, watches and life
The reversal is the key move. The organic picture modelled stones on animals; the mechanical picture models animals on stones. René Descartes (1596-1650) held that non-human animals are machines, lacking consciousness, their behaviour explicable entirely mechanically. The watch replaced the animal as the dominant metaphor: Julien de La Mettrie wrote that "the body is but a watch", man "a collection of springs which wind each other up".
The great remaining mystery was life itself. Thomas Hobbes asserted in 1651 that "life is but a motion of limbs", and the assumption was that a mechanical explanation existed and merely had to be found. Discovery after discovery made life look like a mechanical process, culminating in the structure of DNA in 1953. Reproduction became explicable, in principle, in chemical terms: the organic explained by the inorganic.
Small slips in the history. Crane says life remained a mystery "until the middle of this century", wording carried over from the 1995 first edition; by 2016 it means the twentieth century. He credits DNA to Watson and Crick (the book's chronology adds Maurice Wilkins), omitting Rosalind Franklin, whose X-ray diffraction images, especially Photograph 51, were essential evidence. The chronology also dates Copernicus to 1473, which is his birth year; De revolutionibus appeared in 1543.
- 1632 Galileo publishes the Dialogue on the two chief world systems.
- 1641 Descartes's Meditations sets out the principles of his new science.
- 1642 Pascal builds the first purely mechanical adding machine.
- 1651 Hobbes's Leviathan argues for a materialist, mechanistic picture of human beings.
- 1694 Leibniz completes a calculating machine that can also multiply.
- 1748 La Mettrie publishes L'Homme Machine; Hume publishes his Enquiry.
- 1786 Galvani reports stimulating a frog's muscles with electric current.
- 1854 Boole publishes The Laws of Thought.
- 1879 Wundt founds the first psychological laboratory, in Leipzig.
- 1953 The structure of DNA is discovered.
The Mind as the Last Mystery
Descartes regarded animals as machines but not the human mind: the mind (or soul) has effects in the physical world yet sits outside the mechanical universe of matter. Later mechanists rejected this, so the mind became the one remaining mystery for the mechanical world picture. They assumed an explanation would come, and coined slogans to say so: La Mettrie, "the brain has muscles for thinking as the legs have muscles for walking"; the physiologist Karl Vogt, "the brain secretes thought just as the liver secretes bile". Crane calls these rallying cries, not theories.
Reductionism versus supervenience
One influential answer of the last forty years is that explaining the mind means showing that it is really just matter, that mental states are chemical states of the brain. This physicalist view usually rides on reductionism: every science must be vindicated by physics, meaning its contents are deducible from physics plus "bridge" principles linking physical to non-physical concepts. Ernest Rutherford's quip captures the attitude: "there is physics; and there is stamp-collecting".
Crane finds extreme reductionism "very implausible". Scientific practice does not conform to it, very few sciences have actually been reduced to physics, and science seems to be diversifying rather than unifying. So he separates two theses: that the mind can be mechanically (causally) explained by some science or other, and that that science must reduce to physics. One can hold the first without the second.
God creating the world. Imagine God making this world exactly as it is, and ask what God would have to do. Physicalism says: create the underlying physical nature and stop. Everything else (chemistry, cells, minds, stock markets) comes "for free", because it is fixed by the physical. This is supervenience: everything supervenes on the physical. The image shows why supervenience is weaker than reduction: knowing that the mental is fixed by the physical tells you nothing about how the mind works or whether physics could ever explain it.
This book is not about physicalism. Much of it is compatible with physicalism, but merely saying that the mental supervenes on the physical does not explain how the mind works. A mechanical explanation must do two things:
- Locate the mind in the causal order: show how it is part of the world of causes and effects.
- Articulate causal generalisations: state regularities in the mind, something like natural laws of psychology. Just as physics finds the laws of the non-mental world, on the mechanical picture there can be a natural science of the mind.
Two Obstacles: Consciousness and Thought
The mechanical view is widely accepted in outline, but its application to the mind is deeply problematic. Two phenomena block it, generating recent philosophy of mind's two preoccupations:
- Consciousness. How can a mere mechanism be conscious?
- Thought (intentionality). How can a mere mechanism think about and represent things?
The book's central theme is the second question, how a mere mechanism can think about and represent the world, though Crane holds that intentionality and consciousness cannot be neatly separated, so much of what he says about one carries over to the other. A theory of the mechanical mind silent on consciousness could not count as complete, so the final chapter returns to it.
- Chapter 2. The general puzzle of representation (pictures and words).
- Chapter 3. Mental representation and intentionality.
- Chapter 4. Understanding minds from the common-sense point of view.
- Chapter 5. Understanding minds from the scientific point of view.
- Chapter 6. The basic ideas behind the computer.
- Chapter 7. Can computers think? Crane answers no.
- Chapter 8. Are any mechanisms of our own minds computational?
- Chapters 9, 10 and 12. Since computation presupposes representation, how can the mechanical mind explain representation?
- Chapter 11. A digression: do externalist, extended and embodied/enactive views refute or replace the mechanical mind? Crane argues they do not.
- Chapter 13. Consciousness.
For anyone building AI systems. Crane's two questions map neatly onto two debates about current models: do their internal states represent anything (and how would we know), and is there anything it is like to be them. His separation of causal explanation from reduction is also the stance most engineers take implicitly: we explain a system by its functional organisation, not by deriving it from the physics of silicon.
The Puzzle of Representation
Representation feels effortless, yet words and pictures are only physical patterns, and nothing in a mark tells you what it stands for. Crane tests the obvious solutions and finds each one moves the problem rather than solving it: resemblance cannot explain how pictures represent, pictures cannot capture logical relations, and convention or ideas explain words only by presupposing representation. The trail ends in the mind: interpretation is something minds bestow, so mental representation is the fundamental case.
The Pioneer 10 Plaque
A letter to nobody in particular. In 1972 NASA launched Pioneer 10 with an engraved metal plate: a diagram of a hydrogen atom, a diagram of the relative sizes of the planets marking where the probe came from, and, largest of all, line drawings of a naked man and woman, the man's right hand raised in greeting. Some intelligent alien might find it in millions of years and grasp our science and our peaceful intent. Grant the improbable chain (the probe reaches a star with a life-bearing planet whose intelligent inhabitants can perceive the plate). The aliens would still have to grasp that the marks are symbols at all rather than scratches or decoration; what kind of symbols they are (the atom a scientific diagram, not a picture); what sort of thing each stands for (life forms, not chemical elements); and that the raised hand means greeting rather than aggression, impatience, contempt, or just the arm's normal resting position. Crane's verdict: their understanding the plate is even less likely than the probe arriving in the first place.
The story shows the puzzle. No amount of scrutiny of the marks reveals that these stand for a man, those for a hydrogen atom. The marks can be understood in many ways and nothing in the marks themselves fixes how. Ludwig Wittgenstein, whose philosophy was dominated by questions of representation, put it in a line: "Each sign by itself seems dead; what gives it life?"
The plaque's real authors and critics. Crane tells the story without names. The plaque was designed by Carl Sagan and Frank Drake, with the artwork drawn by Linda Salzman Sagan. His sceptical reading echoes contemporary criticism: the arrow marking the probe's path, for instance, presupposes a convention (arrows as projectiles) that only creatures with a particular history would share.
So the puzzle is simply: how is it possible for one thing to represent something else? It is hard to see at first because representation is woven into everyday life (speech, writing, pictures, gestures, facial expressions). Two facts pull against each other:
- Representation is natural. When we talk or look at a picture, what is represented is immediate; we do not work it out.
- Representation is mysterious. Words and pictures are just vibrations in the air, marks on paper, stone, film or metal. Children who repeat a word until it "loses" its meaning, and anyone who has learned a foreign language, discover that words do not have their meaning intrinsically.
Like time, truth and existence, representation is a characteristically philosophical puzzle: the obvious becomes, on reflection, deeply mysterious. Many other problems cluster around it: the place of mind in nature, thought and language, mutual understanding, consciousness, and thinking machines.
The Idea of Representation
A representation is something that represents something. Crane avoids "something else" because a representation can represent itself: the Liar Paradox sentence "This sentence is false" represents itself. Normally, though, one thing (the representation) represents another, the object of representation. That yields two questions.
What can be a representation
Almost anything: words and pictures; the Pioneer diagram of the atom (neither words nor picture); numerals representing numbers, lengths, or (as a triple) a colour's hue, saturation and brightness; computer data structures representing text, numbers or images; tree rings representing age; a flag a nation; a demonstration aggression; music a mood of unbearable melancholy; flowers grief; a glance irritation. Above all, a belief, hope, desire or wish can represent almost anything. Crane sticks to simple cases of language and thought because the problems already arise in the simplest cases.
What can be an object of representation
Again almost anything: physical objects (a person, a house); properties (a person's shape, a house's colour); facts, situations or states of affairs ("Someone is in my house"); non-physical objects such as numbers (where in the physical world is the number 3?); moods, feelings and emotions; and things that do not exist: unicorns, dragons, the greatest prime number.
Representation is a strange relation. "X represents Y" looks like a two-place relation, and relations normally require both relata to exist. If I kiss Santa Claus, Santa Claus must exist; his non-existence is exactly why I cannot kiss him. Yet I can think about Santa Claus, and his non-existence is no obstacle at all. Many philosophers take this to be central to the nature of representation.
Two questions for any theory
- How does a particular kind of representation represent? What about pictures, or words, makes them play their representing role? The story for one kind need not carry over to another: pictures do not represent the way music does.
- Is one kind more fundamental than the others? Can, for example, language be explained by the mind's representational powers, or must mental representation be explained by language? If one kind is basic, we are on our way to understanding the whole.
Crane's own answer to the second question, defended through the rest of the chapter, is that mental representation is the most fundamental form.
Pictures and Resemblance
Pictures look like the easy case. Nothing intrinsic to the word "dog" makes it represent dogs, but a picture of a dog looks like a dog, in virtue of intrinsic features such as shape and colour. Hence the resemblance theory of pictorial representation: pictures represent things by resembling them.
A toolkit: necessary and sufficient conditions
- Necessary condition. A is necessary for B when B is true only if A is true. Being in England is necessary for being in London.
- Sufficient condition. A is sufficient for B when if A is true, B is true. Being in London is sufficient for being in England.
- Necessary and sufficient. A holds if and only if (iff) B holds. Being in London is not necessary for being in England, and being in England is not sufficient for being in London.
Stated precisely, the resemblance theory says: picture P represents X iff P resembles X. Precision exposes its problems.
Resemblance is not sufficient
The sufficiency half says: if X resembles Y, X represents Y. Crane's argument against it, set out as a reductio:
- If resemblance were sufficient, whatever resembles something would represent it.
- In the loose sense of "having some feature in common", everything resembles everything: I resemble my parents, but also my desk (both physical objects) and the number 3 (both objects of some kind).
- Resemblance is symmetrical: if the picture resembles Napoleon, Napoleon resembles the picture.
- Resemblance is reflexive: everything resembles itself.
- So by (1)-(4), I would represent my desk, Napoleon would represent his portrait, and everything would be a picture of itself.
- These are absurd. Therefore resemblance is not sufficient for representation.
Restricting resemblance to "some respect" does not help (I resemble my father in character traits without representing him), and restricting it to "the respects in which X represents Y" makes the theory circular and uninformative: true, perhaps, but no analysis.
Resemblance as merely necessary
The fallback claims only that if P represents X, P resembles X in some respects. A portrait of the Queen surely must resemble her; that is what a "good likeness" is. But much representation runs on very little resemblance: cubist pictures, caricatures and stick figures. A caricature of the Queen may resemble her less than a detailed drawing of someone else and still be a picture of her. Wittgenstein: a picture might be a striking resemblance of one person and yet be a portrait of someone else it resembles less.
Loosen resemblance enough to cover cubism and corporate logos, and it stops doing explanatory work. Crane does not deny that pictures often resemble what they represent, or that this may help make them pictures rather than sentences or graphs. His point is that factors other than resemblance must be doing the representing.
| Claim tested | Objection | Verdict |
|---|---|---|
| Resemblance sufficient | Everything resembles everything in some respect (me and my desk, me and the number 3) | Fails |
| Resemblance in some respect sufficient | I resemble my father's character without representing him | Fails |
| Resemblance in the represented respects sufficient | Circular: presupposes representation | Uninformative |
| Resemblance sufficient (structural) | Resemblance is symmetrical (Napoleon would depict his portrait) and reflexive (everything depicts itself) | Fails |
| Resemblance necessary | Caricatures, cubism, stick figures, logos represent with minimal resemblance | Possibly true, explains little |
Pictures need interpretation
In Michelangelo's The Last Judgment in the Sistine Chapel we see the damned struggling in agony beneath Christ raising his hand in judgement. We could see Christ welcoming them up with friendly encouragement ("come on up, it's cooler here"), just as the Pioneer hand could be read many ways. We don't, because we interpret the picture in the light of its context: we know it depicts the last judgement.
Wittgenstein's drawing makes the same point. What makes it a man walking up a slope rather than sliding down? Nothing in the picture. Our everyday experience and the contexts we are used to seeing such pictures in fix the reading: the picture does not interpret itself; interpretation comes from outside the picture, and ultimately from the minds of those who use it.
Why Pictures Cannot Be the Basic Form
Could all representation reduce to pictures, given a complete theory of pictures? No. First, interpretation cannot itself be another picture, because that picture would need interpreting too (a regress). Second, and decisively, pictures cannot represent logical relations.
Draw a conditional. Try to picture "If it doesn't rain this afternoon, we will go for a walk." A rain-free scene with us walking in it fails at once: "this afternoon" cannot be pictured (a clock in the corner represents time, but not by picturing it). Two pictures, one rain-free scene and one of us walking, fail too: why read them as if...then rather than and, or either...or? Even "not raining" defeats us: a rain-free scene is equally just a sunny scene. Drawing rain with a cross through it, like a No Smoking sign, smuggles in a non-pictorial symbol, the cross.
So pictures cannot express the relations we mark with "if...then", "and", "either...or" and "not". A picture may be worth a thousand words, but a thousand pictures cannot represent some things one sentence can. For this reason alone, other forms of representation cannot be reduced to pictorial representation.
Linguistic Representation
Convention
A natural thought: words represent by convention, not naturally; communication succeeds when speakers' conventions converge. Convention is surely involved, but what is the convention? If it is "a convention to use 'dog' to represent dogs" (literally and truthfully), it is stated using the very idea of representation, so it takes representation for granted and cannot explain it.
Words as signs of ideas
An equally natural and historically distinguished thought: words represent by being conventionally linked to the ideas speakers intend to express. It goes back to Thomas Hobbes (1588-1679) and especially John Locke (1632-1704), for whom words are "the sensible marks of ideas".
If ideas are mental images, the theory falters three ways:
- Particular or general. If my image accompanying "dog" is of Fido, why does "dog" mean dog rather than Fido?
- No image of "dogness". It is hard to imagine what an image of dogs in general would look like.
- Logical words again. Many words, such as and, or, not, if, have no associated image, the very problem that sank pictures.
Drop the image theory (we often think in words, not pictures) and keep only the claim that words represent by expressing ideas, whatever ideas are. This escapes the convention objection but cannot explain representation, because an idea is itself a representation: a dog-idea represents dogs just as the word "dog" does. One kind of representation is explaining another.
Why not stop at ideas?
One might say we must take something for granted, so take the representational powers of ideas as primitive. Crane's reply:
- If we may take the mind's representational powers for granted, we may equally take language's for granted.
- The mind is not better understood than language; in recent philosophy the reverse may hold, since ideas and thoughts seem even more mysterious than words and pictures.
- That would favour explaining ideas by language, but that cannot be done (as later chapters argue).
- So the representational nature of ideas must itself be explained.
Crane adds that convention and idea-expression are only illustrations, not the only options for a theory of language; the point is that any theory of linguistic representation will in the end appeal to mental representation.
Borrowed arguments and live rivals. The symmetry and reflexivity objections are the classic case made by Nelson Goodman in Languages of Art (1968). Resemblance theory did not die there: later philosophers defended experienced resemblance (Christopher Peacocke, Robert Hopkins) or replaced it with Richard Wollheim's "seeing-in", accounts that grant interpretation and context a role yet keep perceptual similarity central. Crane's quick dismissal targets only the crudest version.
Conclusion: Pictures, Words and Interpretation
Words, like pictures, do not represent intrinsically; they need an interpretation assigned to them. The natural account is that interpretation is bestowed by the mind: words and pictures represent what they do because of the states of mind of those who use them. Those states are representational too, so understanding linguistic and pictorial representation requires understanding mental representation.
Meaning from form alone? Large language models have turned Crane's puzzle into an engineering argument. Emily Bender and Alexander Koller (2020) argued, with their "octopus" thought experiment (an octopus taps a telegraph cable between two islanders and learns to mimic their messages), that a system trained only on linguistic form cannot learn meaning, the relation of words to the world and to speakers' intentions. Critics reply that text carries a great deal of world structure, that multimodal models (CLIP in 2021, and later vision-language models) tie words to images, and that interpretability work finds structured internal representations in these systems. The dispute is Crane's: whether the model's states have only derived intentionality, borrowed from the humans who wrote and read the text, remains open.
Data has no meaning on its own. A field called risk_flag = 3 is Pioneer's raised hand: its meaning lives in the schema, the documentation and the people who agreed on it. Enterprise data governance (glossaries, ontologies, lineage) is the institutional machinery for fixing interpretations, and integration failures are almost always failures of shared interpretation, not of the bits.
Mental Representation
If words and pictures borrow their meaning from minds, the question becomes how states of mind represent. Crane builds the basic vocabulary: thoughts (all representational mental states), propositional attitudes of the form "A ψs that S", and content. He argues that thought can be studied apart from consciousness, distinguishes intentionality from intensionality, and tests Brentano's thesis that intentionality is the mark of the mental, concluding that the first half is nearly true and the second is plausible once we separate original from derived intentionality.
Thoughts: Representational States of Mind
By a state of mind Crane means a belief, desire, hope, wish, fear, hunch, expectation, intention, perception and the like, and he holds that all of these represent the world. The sense is directedness: you cannot simply hope without hoping for something, believe without believing something, or desire without desiring something. What you believe or desire is what your belief or desire represents.
Crane adopts "thought" as the most neutral everyday word for all representational mental states. Belief, desire, hope and love are all thoughts in his sense. Whether all mental states are thoughts is left for Brentano's thesis at the end.
Attitude and situation
When I hope there will be bouillabaisse on the menu at my favourite restaurant tonight, I am thinking of bouillabaisse, the menu, the restaurant and tonight, but not in a random, disconnected way: I am thinking of a possible situation. If I walk past and read the menu, I may instead believe there is bouillabaisse tonight. The situation is the same; what differs is that belief takes it to be the case, while hope would like it to be the case.
So every such thought has two aspects: the situation represented and the attitude taken to it. I can expect that I will visit Budapest, hope that I will, or believe that I have: one situation, three attitudes. Conversely one attitude (belief) ranges over countless situations. Thoughts of this kind are uniquely picked out by:
- The attitude (belief, hope, expectation, etc.);
- The situation represented.
Many attitudes also come in degrees (wanting more or less strongly, believing with more or less conviction), which complicates but does not change the picture.
A ψs that S, where A is the person in the mental state, ψ (psi, for "psychological") is the attitude, and S is the situation represented.
- Vladimir (A) believes (ψ) that it is raining (S).
- Renata (A) hopes (ψ) that she will visit Romania (S).
Bertrand Russell (1872-1970) named such states propositional attitudes (in The Analysis of Mind, 1921). His "proposition" is roughly Crane's "situation": what the attitude is directed at, not a piece of language.
Content
The now-universal term for Russell's proposition is content. When I believe there is beer in the fridge, the content of my belief is that there is beer in the fridge; desires, hopes and so on have content too. In current philosophy the problem of mental representation is usually phrased as "What is it for a mental state to have content?" For now, content is what distinguishes states involving the same attitude: beliefs are individuated by their contents, and so are desires.
Not all thoughts are propositional attitudes
- Love. You cannot love without loving something, so love is representational. But it is directed at a person, place or thing, not a situation, and cannot be put in the "A ψs that S" form. Love is a thought but not a propositional attitude.
- Desire. On the surface desire resembles love (I desire a cup of coffee, a thing). Many philosophers hold this under-describes it: really I desire that I have a cup of coffee. On that view all desires are desires that so-and-so, and desire is a propositional attitude.
- Unconscious attitudes. If Oedipus really desires to kill his father and marry his mother, those desires fit the schema and count as thoughts, though they are not conscious.
Thought and Consciousness
Consciousness is what makes waking life seem as it does and is arguably the source of all value; Einstein told the philosopher Herbert Feigl that without this "inner illumination" the universe would be "nothing but a heap of dirt". Many people treat "thought" and "conscious" (even "mental" and "conscious") as synonyms. Crane nonetheless separates the questions, with a simple argument:
- Some thoughts are unconscious: there are things we think without being aware that we think them.
- If something can be a thought without being conscious, then being conscious is not essential to being a thought.
- So it should be possible to explain what makes a thought a thought without explaining consciousness.
Three kinds of evidence for unconscious thought
The President's socks. Asked "Does the President of the United States normally wear socks?", you would answer yes, and what people say is good evidence of what they think. Yet the sentence has probably never crossed your conscious mind. Did you only start believing it when asked, having had no opinion a moment earlier? It makes more sense to say the unconscious belief was there all along.
Meno's slave. In Plato's Meno, Socrates defends the view that all knowledge is recollection of truths known in the soul's previous life. He questions Meno's slave about squares, asking leading but non-revealing questions, until the slave reaches the correct geometrical answer. Socrates: were these opinions his own? Meno: yes. Socrates: yet he did not know them, but "these opinions were somewhere in him". Crane rejects recollection but keeps the lesson: knowledge can lie "somewhere" in a mind without being explicitly conscious, can be recovered, or can stay buried forever. Knowledge is a kind of thought, so unconscious knowledge implies unconscious thought.
The geometry is simplified. In the Meno the slave starts with a square of side 2 (area 4) and is asked for the side of a square of double the area (8). He first guesses that doubling the side doubles the area, discovers it quadruples it (16), and eventually sees that the answer is the square drawn on the diagonal. Crane's summary compresses this to "the area of a square with sides 2N"; the philosophical point is unaffected.
- Freud. The third influence is Sigmund Freud's insight that much of what we do is explained by unconscious beliefs and desires, some buried so deep that psychoanalysis is needed to dig them out. We can accept the general claim without the specifics (the Oedipus complex, "penis envy"). It also fits ordinary talk of people who do not "know their own minds" or who deceive themselves, which would be impossible if thoughts were essentially conscious.
Descartes would have found unconscious thought a contradiction in terms, and some still agree; Crane thinks far more philosophers now take it seriously. Note that each of his examples involves a thought brought to consciousness, and he will often use conscious examples. What interests Crane is what makes thoughts thoughts, not what makes them conscious.
Penrose's challenge
Roger Penrose, in The Emperor's New Mind, claims that "true intelligence requires consciousness". Crane neither agrees nor disagrees. Even if consciousness is a necessary condition for thinking, it does not follow that understanding thought requires understanding consciousness.
Thought and life. Suppose everything that thinks must be alive. Then "true intelligence requires life". Yet nobody concludes that a theory of thought must also be a theory of life; we simply presuppose that thinkers are alive. The same goes for consciousness. Crane keeps an open mind on whether a creature could have wholly unconscious thoughts, and the book's argument does not depend on the answer.
Intentionality
Philosophers call the representational character of mental states intentionality, and states that have it intentional states. The word comes from the medieval scholastics: intentio meant concept, and St Thomas Aquinas (c.1225-1274) used esse intentionale (intentional existence) for the way things are conceptually represented in the mind. Franz Brentano (1838-1917) revived the term in Psychology from an Empirical Standpoint (1874), characterising mental phenomena by "the intentional inexistence of an object", which he glossed as "relation to a content", "direction upon an object" and "immanent objectivity".
These phrases all say one thing: mental phenomena involve the representation or presentation of the world. "Inexistence" means the object exists in the act of thinking itself: not that a dog is "in" my mind when I think of my dog, but that my dog is intrinsic to my thought, since what makes it the thought it is, is having my dog as its object. Crane reads intentionality as directedness on something. Many contemporary writers prefer "aboutness"; Crane prefers "directedness".
Brentano's thesis: all mental phenomena exhibit intentionality and no physical phenomenon does. Intentionality is "the mark of the mental".
Confusion one: intentionality and intention
The words suggest a link with intending and acting intentionally. There is a connection: to intend to do A is plausibly to represent A to oneself, so intentions are intentional states. Beyond that there is no substantial link: intentions are intentional states, but most intentional states have little to do with intentions.
Confusion two: intentionality and intensionality
Intensionality-with-an-s (as John Searle labels it, against intentionality-with-a-t) is a feature of linguistic and logical contexts. A context is intensional when it is non-extensional, and a context is extensional when two principles hold:
- (A) Intersubstitutivity of co-referring expressions (salva veritate, "saving truth"). If an object has two names, swapping one for the other cannot turn a truth into a falsehood. George Orwell was born Eric Arthur Blair (he took "Orwell" from the River Orwell in Suffolk), so "Orwell wrote Animal Farm" and "Blair wrote Animal Farm" are equally true.
- (B) Existential generalisation. From a statement about something, infer that it exists: from "Orwell wrote Animal Farm" infer "there exists someone who wrote Animal Farm".
Belief sentences break both.
| Sentence | Principle tested | Result |
|---|---|---|
| Orwell wrote Animal Farm → Blair wrote Animal Farm | (A) substitution | Holds: extensional |
| Vladimir believes Orwell wrote Animal Farm → Vladimir believes Blair wrote it | (A) substitution | Fails: Vladimir doesn't know Orwell's real name |
| Orwell wrote Animal Farm → someone wrote Animal Farm | (B) existential generalisation | Holds |
| Vladimir believes Santa Claus lives at the North Pole → there is someone Vladimir believes lives there | (B) existential generalisation | Fails: Santa doesn't exist |
| Nine is necessarily > five; the number of planets is nine → the number of planets is necessarily > five | (A) in a modal context | Fails, with no mind involved |
| Vladimir sees Orwell → Vladimir sees Blair; there is someone he sees | (A) and (B) in a perceptual report | Both hold, though seeing is intentional |
The principles can fail in belief contexts, not must; often substitution and existential inference go through. The point is that there is no guarantee.
Linked, but not the same
The link is tempting. Substitution fails in belief reports because their truth depends on how the believer represents the object: Vladimir represents Orwell as Orwell, not as Blair. Existential generalisation fails because we can think about what does not exist, a defining mark of intentionality. So perhaps intensionality in belief sentences is a consequence of intentionality in the beliefs. That is as far as it goes, for two reasons:
- Intensionality without intentionality. Modal sentences about necessity and possibility are intensional though they concern no mind: "nine" and "the number of planets" co-refer, but it is not necessary that there are nine planets (there could have been four, or none).
- Intentionality without intensionality. "X sees Y" describes an intentional state yet appears extensional: if Vladimir sees Orwell he sees Blair, and the author of The Road to Wigan Pier, and there is someone he sees. Crane admits this is controversial.
Argument (1) alone suffices: intensionality is not sufficient for intentionality, and may not be necessary, so it cannot serve as a criterion of intentionality.
Intentionality (with a t). A property of mental states: their directedness on objects or situations, including non-existent ones. The subject matter of philosophy of mind.
Intensionality (with an s). A property of sentences and contexts: the failure (or possible failure) of substitution of co-referring terms or of existential generalisation. Found in belief reports, but also in modal talk that has nothing to do with minds.
A dated example. The nine-planets case comes from W.V. Quine's classic discussion of modality. Since the International Astronomical Union reclassified Pluto in 2006, the number of planets is eight, so the second premise is now false. The argument still works with "eight", and the change itself illustrates the contingency that makes the modal context intensional.
Brentano's Thesis Tested
Crane splits the thesis in two:
- Do all mental states exhibit intentionality? (Is mentality sufficient for intentionality?)
- Do only mental states exhibit intentionality? (Is mentality necessary for intentionality?)
Are all mental states intentional?
Introspection suggests not. Two counterexamples:
- Pain. A sharp pain at the base of my spine is mental (only a conscious being can have it) yet seems just a feeling, not about anything. Crane notes that bodily pain differs greatly from the "mental pain" of bereavement, but a toothache is no less mental for being called physical.
- Objectless depression. One can be depressed without being able to say what about.
Crane's replies:
- Pains have felt location. I could have two pains exactly alike "pain-wise", one felt in my right hand and one in my left. The only difference is where they are felt to be, which is plausibly a difference in what the state is directed on. So pains are not wholly non-intentional, even though they are not propositional attitudes ("Oswaldo feels pain" doesn't fit "A ψs that S"). Not being a propositional attitude does not make a state non-intentional, as love showed.
- Depression is directed on the world. First, being able to identify a state's object cannot be a criterion of intentionality, or certain kinds of self-deception would be impossible. Second, depression is, in Lewis Wolpert's phrase, a "thoroughly negative view of the external world": everything seems bad, nothing is worth doing, the world "shrinks". The world in general can be an object of a state of mind.
So no state of mind is obviously wholly non-intentional. But there may be non-intentional properties of states: my toothache is directed on my tooth, yet its naggingness may not be directed on anything; it is just there. Such properties are called qualia. If they exist, Brentano's thesis survives for whole states but leaves a non-intentional element in mental life, which Crane calls "something of a Pyrrhic victory".
Are only mental states intentional?
The obvious counterexample is books: their sentences mean and represent, yet a book has no mind. The reply recycles the last chapter: the book's sentences have intentionality only because readers interpret them, and readers' states have intrinsic intentionality.
- Derived intentionality. Intentionality that depends on the intentionality of something else: a book's sentences derive theirs from the thoughts of writers and readers.
- Original intentionality. Intentionality that does not derive from anything else: the intentionality of minds.
The question becomes: can anything other than a mind have original intentionality? Crane finds it baffling. If we met something with original intentionality, it is hard to see how whether it had a mind could be a further question. If computers turned out to have original intentionality, we could say "A computer can have a mind!" or "Something can have original intentionality without a mind!", and the choice looks largely terminological. The second half of the thesis seems plausible in outline, but judgement waits on a better account of what it is to have a mind.
The random library. Imagine a machine that strings letters at random, and by sheer chance one page reads as a correct account of the Battle of Hastings. Nobody meant anything by it. Does it have intentionality? If derived intentionality needs only readers who interpret, then yes, as soon as a historian reads it; if it needs a writer's intention, then no. The case separates two sources of derived meaning, and it is no longer hypothetical: text generated by a model sits somewhere between the random page and a human author.
Do language models have original intentionality? Crane's terminological fork has become a live debate. One camp holds that an LLM's outputs have only derived intentionality, meaningful because human readers and the human-written training text make them so. Another argues that a model's words can refer by inheriting the causal-historical chains of the usage it learned from, in the way a child's first words borrow reference from a community. Interpretability research adds that these networks build internal states that track features of their inputs and can be causally manipulated. Whether that tracking amounts to original intentionality, rather than information-carrying that we interpret, is exactly the question this chapter leaves open.
Conclusion: From Representation to the Mind
The Pioneer plaque exposed the puzzle; the resemblance theory proved a slim basis for pictorial representation; pictures and words both need interpretation; and interpretation derives from mental representation, or intentionality. To understand representation at all, we must understand representational states of mind, which is where the next chapter goes, by asking how we understand thinkers and their thoughts.
Understanding Thinkers and Their Thoughts
We already know a great deal about minds without any brain science: we predict our friends better than the Met Office predicts the weather. Crane shows that the problem of representation survives whichever answer we give to the mind-body problem, rejects behaviourism because thoughts can only be tied to behaviour via other thoughts, and defends the causal picture of thought: thoughts lie behind behaviour as its causes, a claim supported by Davidson's argument about acting for a reason and by the broader mechanical world view.
We Are All Experts on Minds
If understanding minds required understanding the brain, most people would know almost nothing about minds, and even experts find parts of the brain mysterious. Yet we plainly know a vast amount: that we and others have thoughts, experiences, memories, dreams, sensations and emotions, and fine distinctions such as hope versus expectation, regret versus remorse. We put this to use constantly. Arrange lunch with a friend tomorrow and you will be more confident she turns up than you are of the weather forecast, a prediction resting on knowledge of her mind: that she understood you, knows where the restaurant is, wants to meet.
This does not show the mind is something other than the brain. Our mental knowledge might be knowledge of biochemical brain states without seeming so. Happily, Crane argues, we need not settle the mind-body problem to say what we know about minds.
The Mind-Body Problem
The mind-body problem is how mind and body are connected. That they are connected is obvious: brain damage transforms thinking; drugs and alcohol affect the brain and thereby thought. The difficulty is that we are dragged two ways at once:
- We must be just matter. Humans evolved from lower forms of life made entirely of matter, so minds first evolved from complex matter; and if all my matter were removed bit by bit, nothing of me would be left.
- We cannot be just matter. It seems incredible that we are "a few dollars' worth of carbon, water and some minerals"; any bodily injury makes it seem absurd that fragile, messy matter constitutes a thinking, conscious agent. Talk of the "chemistry" between lovers is metaphor; love as literally "nothing but a chemical reaction" seems absurd.
"Motherlove is magnesium." P.J. Fitzpatrick told of 1940s researchers who found that female cats deprived of magnesium stopped caring for their offspring; a newspaper ran the headline "Motherlove is magnesium". Whether true doesn't matter: the joke works because reducing conscious life to chemicals feels as absurd as reducing motherlove to magnesium. Yet scientists keep finding correlations between psychological disorders and brain chemicals, and it would be a desperate last resort to insist, from near-total ignorance, that there must be a limit. The truth is not "motherlove is magnesium", but it may not be too far away.
The problem has seemed so intractable that some think it insoluble. Joseph Glanvill (1636-1680): "How the purer spirit is united to this clod is a knot too hard for fallen humanity to untie." The main proposed solutions:
- Materialism (physicalism). The mind is just the matter of the brain organised in a complex way; today the orthodox view in one variety or another.
- Dualism. There are two main kinds of thing, material and mental. Believers in immaterial souls that survive death must be dualists: if the mind were the body, how could it survive the body's annihilation?
- Idealism. Everything is ultimately mental; rare today.
Why the representation problem survives either answer
Crane's argument that he can set the mind-body problem aside:
- If materialism is true, the question becomes how the brain can represent anything. Picture it: about 100 billion cells, roughly a kilogram of grey and white watery stuff resembling yoghurt. How this "clod" constitutes thoughts about yourself, your life and the world is just as puzzling as how a mind represents.
- If dualism is true, the question becomes how an immaterial soul represents anything. Descartes said the soul's essence is to think, but saying it is in a thing's essence to do X does not explain how it does X. That is the doctor in Molière's Le Malade imaginaire explaining that opium sends you to sleep because of its virtus dormitiva, its "dormitive virtue".
- So both physicalism and dualism still need a theory of representation: the physicalist must say what distinguishes thinking from non-thinking matter, the dualist what enables the non-material mind to think.
- Therefore the problem of representation can be pursued without deciding the mind-body problem.
Idealism dissolves the problem (everything is thought anyway), but it is much harder to believe than most philosophical views, so it trades one mystery for another.
| Answer to mind-body problem | What remains to explain | Status of representation puzzle |
|---|---|---|
| Physicalism | How this matter thinks and other matter doesn't | Untouched |
| Dualism | How the immaterial soul manages to think | Untouched; "essence" is a dormitive virtue |
| Idealism | How a wholly mental world appears physical | Dissolved, at a heavy price |
Brain numbers. Crane's "about 100 billion brain cells" and "about a kilogram" are textbook round figures. Counts by Suzana Herculano-Houzel's group (2009) put the human brain at about 86 billion neurons with a comparable number of non-neuronal cells, and an adult brain typically weighs about 1.3 to 1.4 kg. Nothing in the argument depends on the figures.
Understanding Other Minds
Asking how we find out about minds sheds light on what we know. One asymmetry is obvious: we know our own minds partly by introspection, concentrating on the contents of consciousness, but we cannot introspect your mind. Sometimes I need a friend or therapist to tell me what I think, but learning about others' minds is always like that. The asymmetry parallels bodily knowledge: I must look to see whether your legs are crossed, but I know without observation whether mine are.
Scepticism set aside
With others, all we seem to have is observable behaviour, what they say and do. Scepticism about other minds says this cannot yield knowledge.
Robots and vats. Everyone around you might be robots programmed by a mad scientist to behave as if conscious, leaving you the only real mind; this is compatible with all your evidence. Its twin, scepticism about the external world, imagines your brain in a vat of nutrients, its nerves stimulated by a mad scientist so that you seem to experience everyday objects. Neither is meant as a tenable position (few philosophers have seriously denied that others have minds); each is a challenge to say what we really know and how.
Answering scepticism belongs to epistemology, the theory of knowledge, and Crane leaves it there. Whatever knowledge turns out to be, it had better allow the obvious fact that we know a lot about other minds. His question is how we know, not whether. He accepts one premise of the sceptic: we do not perceive others' thoughts directly; we perceive words and actions (some philosophers disagree).
Behaviourism and Its Internal Failure
One once-serious answer: observable behaviour is all there is to having a mind; being in pain is just pain-behaviour (crying, moaning, complaining). Behaviourism was briefly popular in twentieth-century psychology and philosophy. It makes knowledge of other minds easy but self-knowledge baffling, since we know our own minds without watching ourselves; hence the joke repeated "ad nauseam": two behaviourists meet, and one says, "You're feeling pretty well today, how am I feeling?" It also ignores or denies subjective, conscious experience. Crane sets those familiar objections aside to press behaviourism's internal inadequacy: even on its own terms, it cannot define mental states purely in terms of behaviour.
From behaviour to dispositions
I believe Riga is the capital of Latvia but have never shown it in behaviour. The behaviourist replies that belief requires a disposition to behave, like a sugar lump's solubility: soluble even if never put in water, because disposed to dissolve if it were.
- Suppose the relevant behaviour is verbal: saying "Riga is the capital of Latvia" when asked (the question plays the role of the water).
- But I will answer that way only if I understand English, and monoglot Latvians hold the belief without understanding English. So understanding English is a separate mental state needing its own behavioural analysis.
- Understanding cannot be a disposition merely to make the sounds: a parrot could have that. It must be a disposition to utter them appropriately.
- When is it appropriate? Not simply when I believe it: I might utter the sentence to make my audience believe it while myself (wrongly) believing the capital is Vilnius.
- And appealing to the belief to fix appropriateness would be circular, since the utterance-disposition was meant to explain the belief.
- So thoughts cannot be defined in terms of behaviour alone; other thoughts must be mentioned.
The holism of the mental. Each attempt to pair one thought with one piece of behaviour fails unless other mental states are in place, and pairing those with behaviour hits the same problem. A single thought may go with many different behaviours depending on which other thoughts you have.
The umbrella man
A man looks out of the window, goes to a closet and takes an umbrella before leaving. The obvious reading is that he thought it was raining. Yet that thought would move him only if he also wants to stay dry, believes the umbrella will keep him dry, and believes this object is his umbrella, thoughts probably never conscious, without which the act would be mysterious.
We learn about the thoughts of others by making reasoned conjectures about what makes sense of their behaviour, guided by general assumptions about what it is to be a thinker. The same act fits many conjectures:
| Conjecture | Beliefs and desires attributed |
|---|---|
| Rain | Thinks it is raining; wants to stay dry |
| Sun | Thinks it is sunny; wants protection from the heat |
| Magic | No view on weather; thinks the umbrella wards off evil spirits |
| Murder | Plans to kill an enemy; thinks the umbrella hides a weapon |
Given that it is actually raining and we know it, the first is by far the likeliest, because we assume he sees what we see, that getting wet when clothed is undesirable, and that people avoid the undesirable when it is cheap to do so. We attribute the thoughts it is reasonable for him to have, given his view of his surroundings, his mental faculties and his degree of rationality.
Common-sense psychology
Many philosophers and some psychologists call these assumptions a theory of other minds: "common-sense psychology" or "folk psychology", by analogy with "folk physics", our common-sense principles about how physical objects behave. Crane accepts that it is a theory in a loose sense but prefers "common-sense psychology": "folk psychology" connotes homespun truisms of the "many hands make light work" kind, an invidious attitude that prejudges later debates.
Philosophers also say the purpose of common-sense psychology is "the explanation of behaviour". True, in that we make sense of behaviour by attributing mental states, but misleading: often we want to know what people think out of pure curiosity, interested in the psychological facts that organise and "lie behind" behaviour rather than the behaviour itself.
The behaviourist can accept that the umbrella man "took it because he thought it would rain" as the explanation we find natural, while denying that anything inside produces the behaviour. Crane finds this very unsatisfactory: we want to know what is true of others, and so what makes our explanations true and what justifies preferring one over another, which means what brings the behaviour about.
Super-Spartans. Hilary Putnam (1963) imagined a community trained from birth to suppress every outward sign of pain: no wincing, no complaint, no avoidance behaviour, though they feel pain exactly as we do. If such people are conceivable, pain cannot be a disposition to pain-behaviour. Combined with Crane's holism argument (no thought maps to behaviour on its own), it leaves behaviourism attacked from both sides: mind without behaviour, and behaviour that underdetermines mind.
The Causal Picture of Thought
That thoughts cannot be directly perceived does not make them mysterious: microbes are too small to see, black holes too dense even to let light escape, and neither is mysterious merely because unseen. Saying thoughts "lie behind" behaviour means more: thoughts produce behaviour, and we know them through their effects. The relation between thought and behaviour is causal.
Causation pervades ordinary thought: economic policy causes inflation, smoking causes cancer, HIV causes AIDS, excess carbon dioxide causes global warming, which will cause rising sea levels. David Hume (1711-1776) called causation "the cement of the universe". If my desire for a drink caused me to go to the fridge, the relation is fundamentally the same as that between smoking and cancer. Philosophers dispute what causation is, or even whether it exists, so Crane uses only three uncontroversial features.
- Counterfactuals. If A caused B, then had A not occurred, B would not have occurred. If my desire for a drink caused my drinking, then without the desire there would have been no drink. Causes work only in the circumstances: a bacterium causes an epidemic only given poor vaccination, sanitation and medical care, just as a desire causes action only given other beliefs (that I can get a drink).
- Explanation. To explain is to answer a "Why?" question, and citing a cause is one way: "Why did he get cancer?" "Because he smoked." "Why did the man take his umbrella?" "Because he thought it was raining."
- Regularity. Hume: a cause is "an object followed by another, and where all the objects, similar to the first, are followed by objects similar to the second". We expect similar causes to have similar effects; a thrown ball that failed to fall would make us look for an intervening cause. If thought causes behaviour, there should be regularities connecting them, a topic for the next chapter.
Regularities differ in strength. Crane has never eaten a pizza more than 20 inches across, but only modesty stops him; nature stops unsupported objects flying into space. Philosophers distinguish accidental regularities from laws of nature.
Note what the causal picture does not say. Causation may be mental or physical depending on its relata, so the causal picture does not entail physicalism. It is, however, a key element of the mechanical mind: the mind is part of the causal order just as the liver and heart are, and we learn about other minds as we learn about the rest of nature, by their effects.
The Opposition: Wittgenstein and Anscombe
The causal picture is not a trivial truth that anyone who grasps mental concepts must accept. Wittgenstein and followers deny it, holding that describing mind in terms of causes and mechanisms imposes a model of explanation fit only for non-mental things. G.E.M. Anscombe, Wittgenstein's student: the mistake is to think that doing something in execution of an intention is "a causal relation between act and intention".
The causal picture. Being amused is an inner state that causes laughing and smiling. We need the inner/outer distinction because people can be silently amused or fake amusement convincingly. The inner state is known by its effects.
The expressive picture. Genuine laughter is the expression of amusement, partly constituting what being amused is, not an effect of it. Treating the inner state as a hidden cause is like looking for hidden facts a picture or piece of music expresses. Wittgenstein: speech with and without thought compares to playing music with or without thought.
Two Arguments for the Causal Picture
Davidson's argument from acting for a reason
The first argument derives from Donald Davidson ("Actions, reasons and causes", 1963). A reason here is a collection of thoughts that make sense of a plan of action, good or bad, moral or not.
Boleslav and his brother. Boleslav is jealous of his brother, who he feels frustrates his progress, and wants to kill him. Case 1: in a bar-room brawl that has nothing to do with his plot, he accidentally kills a man who, unknown to him, is his disguised brother. Case 2: he walks into the bar, recognises his brother and shoots him dead. In both he has a reason to kill his brother and kills him; only in case 2 does he kill him for that reason.
- The two cases obviously differ: only in the second does Boleslav act for his reason.
- Any account of acting for a reason must explain this difference.
- The counterfactual test separates them: in case 1, holding circumstances fixed, without the fratricidal thoughts he would still have killed his brother in the brawl; in case 2, not.
- The explanatory test separates them: in case 1, "because he was jealous" is no answer to why he killed his brother; in case 2 it is.
- These are marks of causation. So treating the reason as the cause of the action in case 2, and not in case 1, explains the difference.
- Therefore to act on a reason is to have that reason as a cause of one's action.
Crane rates this persuasive but not compelling. Its structure is inference to the best explanation: two situations differ, causation explains the difference. That leaves room for an opponent to offer a better non-causal account, so the argument will not convert a Wittgensteinian.
Historically, it belongs to a movement against the Wittgensteinian view that found an essential causal component in many mental concepts:
- Perception as a causal relation between perceiver and object perceived.
- Memory as a causal relation between the memory and the fact remembered.
- Knowledge, and the relation between language and reality, as based on causal relations.
Deviant causal chains. The causal theory faces a problem Davidson himself raised (1973): a climber wants to rid himself of the weight of his partner on the rope and knows that loosening his grip would do it; the thought so unnerves him that his grip loosens. The desire causes the movement, yet he does not let go for that reason or intentionally. So causation by a reason is not sufficient for acting on it; the cause must operate "in the right way", and no agreed account of the right way exists. Crane's chapter does not mention this weak point.
The ideological argument
The second argument depends on accepting a whole picture of the world, the mechanical/causal world picture, rather than on indisputable premises. All of nature obeys general causal laws (of physics, chemistry, biology), and psychology has its laws too. Throughout nature we find the regular succession of events and the determination of one by another; why should the mind be exempt?
- Everyone accepts that mental states are effects of physical causes: the colours you see, smells, tastes and sounds result from mechanistic physical processes; stimulants, antidepressants, narcotics and alcohol act on thought with regular, law-like connections.
- Nothing in the concept of a mental state suggests it can be an effect but never a cause.
- So, absent a reason to the contrary, mental states are causes as well as effects.
Conclusion: Toward a Science of the Mind
Neither argument is a knock-down proof. Together they are an invitation to place thoughts in the causal order of the world, as things with causes and effects outside the mind. Crane distinguishes this from three stronger claims: that mental concepts can be analysed in causal terms, that only things with causes and effects are real, and physicalism. But the causal picture raises the question the next chapter takes up: to what extent can thought be studied scientifically, given that, while not all sciences reduce to physics, there are scientific accounts of many causal phenomena?
Common-sense psychology in machines. In 2023 Michal Kosinski reported that large language models pass classic false-belief tasks, the standard test of attributing mental states, at rates comparable to young children. Tomer Ullman replied the same year that small, trivial alterations to the stories (making a container transparent, say) made the models fail, suggesting pattern-matching to the familiar test format rather than robust mental-state attribution. Later studies found strong models do well on many such tasks yet remain brittle off-distribution. The debate replays Crane's umbrella man: success on the test is behaviour, and which network of attributed states, if any, lies behind it is the open question.
Explaining an agent's actions. Audit trails for agentic AI in banking face the Boleslav problem: an agent can produce a stated rationale and an action without the rationale being what actually caused the action. Regulators asking "why did the system do this?" want the causal reason, which is why counterfactual testing (would the decision change without this input?) matters more than post-hoc explanations.
Common-Sense Psychology and Science
If common-sense psychology is a theory, then beliefs and desires are theoretical entities, and a theory can be false. That opening lets eliminative materialists like Paul Churchland argue that there are no beliefs or desires at all, to be replaced by neuroscience. Crane finds their arguments weak, examines the rival simulation theory, and suggests the two accounts of mindreading may differ only in emphasis. Common-sense psychology tells us that thoughts represent the world and cause behaviour, which opens the door to the idea that minds compute representations.
The Theory Theory
Adam Morton labelled the view that common-sense psychology is a theory the "Theory Theory". A theory, generally, is a principle or set of principles devised to explain phenomena. Common-sense psychology's principles may be truisms: people generally try to achieve the objects of their desires (other things being equal); a person looking at an object in good light will normally believe it is in front of them.
The Theory Theory claims more than that such principles exist. It claims mental states are theoretical entities: the true, complete theory of mental states tells us everything there is to know about them. Compare atoms: everything there is to know about atoms is contained in the true complete theory of the atom, so their nature is exhausted by the theory's description. Contrast colours: arguably we know something about colours (what they look like) that no physical theory of colour supplies.
King Lear's son. (The analogy follows David Lewis's account of theoretical terms.) Treat a theory like a story: "Once upon a time there was a man called King Lear, who had three daughters, Goneril, Regan and Cordelia..." Asked "Who was King Lear?", you paraphrase: the man who divided his kingdom, disinherited his favourite daughter, went mad, ended up on a heath. Asked "Did Lear have a son? What was his hairstyle?", the story is silent, and not because it omits a fact: everything there is to know about Lear is in the story. On the Theory Theory, belief is like Lear: list all the familiar truths ("states that causally interact with desires to cause actions...") and you have a theoretical definition that exhausts what beliefs are.
The Theory Theory is independent of the causal picture. One could accept that thoughts have effects without treating common-sense psychology as a theory, or (unusually, as a sceptic about causation might) accept the Theory Theory while denying that thoughts cause anything.
Rationality and generalisation
Running for the bus. A woman carrying several bags runs along an empty pavement as a bus overtakes her towards a stop. Obviously she is running for the bus. She might think she is being chased, or be exercising, but the empty pavement and the fact that people don't exercise carrying large bags rule those out. We assume a degree of rationality: she pursues an immediate goal (catching the bus) for a longer one (getting home) by reasonable means (not lying in the road hoping the driver will pick her up).
Irrational and crazy behaviour exists, but if all behaviour were crazy, interpretation would be impossible: we could not choose between wild hypotheses. So interpretation requires regularity between thought and behaviour, and if common-sense psychology is a theory, it should contain generalisations describing those regularities.
The "dull truisms" objection
If there were such generalisations, shouldn't we be able to state them? We are bad at it. Morton: principles like "anyone who thinks there is a tiger in this room will leave it" are almost always false, and the true ones ("people try to achieve what they desire") provoke "Tell me something I didn't know!" Common-sense psychology combines powerful, versatile explanation with an absence of daring hypotheses: in general we find only dull truisms, yet in particular cases we produce acute, brave hypotheses about why someone acted.
Crane grants something to this but offers two replies:
- Familiar is not feeble. If we apply the theory constantly, its principles will feel truistic. Folk physics says solid objects resist pressure and penetration: a truism that informs all our dealings with objects.
- Tacit knowledge. Only knowledge of a psychological theory can explain how we interpret one another so successfully, but that knowledge need not be conscious. Like the geometry in Meno's slave, or tacit knowledge of linguistic rules that explains how we understand language, it can be there without being statable.
What the Generalisations Say Thoughts Are
Asked why we read the woman as running for the bus, we might say "the bus is coming". Not quite: what moves her is that she thinks it is coming. If it came and she didn't realise, she wouldn't run; if she mistook a truck's sound for the bus, she would.
- Motivating reason. The agent's beliefs and desires and how they produce action (Davidson's usage; the distinction is drawn by Michael Smith): she ran because she thought the bus was coming.
- Normative reason. A fact in the world that counts in favour of thinking or doing something; T.M. Scanlon: "a consideration that counts in favour of something". She ran because the bus was coming.
Whatever we call them, what people do is determined by how they take the world to be, which is not always how it is. Taking the world a certain way is representing it. So, according to common-sense psychology, the thoughts that determine behaviour are representational, and it matters how things are represented, not just which objects are. Someone running for the bus must represent it as a bus, not just as some motorised vehicle (why chase one of those?). In the brawl, Boleslav represented his victim somehow, but not as his brother, which is why his fratricidal desire did not cause the killing. This is the intensionality of the previous chapters doing real explanatory work.
The belief-desire generalisation
The second ingredient, on the causal picture, is that "because" in psychological explanation is causal, just as in "he got cancer because he smoked". Combining the causal picture with the Theory Theory yields generalisations about the effects of thoughts, such as:
Beliefs combine with desires to cause actions which aim at the satisfaction or fulfilment of those desires.
I desire a glass of wine and believe there is wine in the fridge, that the fridge is in the kitchen, and that the kitchen is over there; these cause me to move towards the fridge. I may not, if I have a stronger desire to keep a clear head, or believe the wine is someone else's. That does not refute the generalisation, which says only that one acts to satisfy one's desires, whatever they are; the stronger desire causes a different act (avoiding the fridge, a bracing walk in the country). Such trains of thought are not conscious ("if this is what they are consciously thinking, it is probably unwise to have another drink"). They are the causal "springs" of action, not the occupants of consciousness.
Two questions follow for the Theory Theory:
- Is it a good theory? Even if our everyday understanding is a theory, is it true?
- Is it a theory at all? Or should mindreading be understood some other way? (Rejecting the Theory Theory is not rejecting the causal picture.)
Elimination or Vindication?
Common-sense psychology needs only a fairly mature mind, some imagination and familiarity with people; in that sense we are all psychologists. Scientific psychology uses technical concepts and quantitative methods. Both claim to be theories of the mind. We cannot dodge the clash by saying science studies the brain and common sense the person, for three reasons:
- Two theories, one thing. The mind could just be the brain. Common sense says the table is solid wood; particle physics says it is mostly empty space. Arthur Eddington inferred two tables; that is a bad inference from two theories to two things.
- Shared subject matter. Scientific psychology asks how memory works, how we see objects, why we dream, what mental images are, all phenomena familiar to common sense; no reason to think "vision" names different things in each.
- Psychology without neurons. Much scientific psychology proceeds without reference to the brain, not out of Cartesian dualism but because it makes sense to study the mind in large-scale, macroscopic terms before seeking its neural implementation.
So the relationship is urgent, and the options boil down to two.
Vindication. Common-sense generalisations are largely true, and science will explain why. We know that normal perceivers looking at an object in good light believe it is there; the science of vision explains what about brains, eyes and light makes that so. Analogy: people knew thrown objects fall before Newton explained why.
Elimination. There are good reasons to think common-sense psychology false, so science owes it no explanation and should develop without its categories. Analogy: alchemy's philosopher's stone, which science did not explain but discarded.
Eliminative materialism
Because eliminativists are always materialists, the view is called eliminative materialism. Paul Churchland (1981): our common-sense conception of psychological phenomena is "a radically false theory", so defective that "both the principles and the ontology" will be displaced by completed neuroscience. The ontology is what the theory says exists (beliefs, desires, intentions), so eliminativists normally hold that there are no beliefs, desires, intentions, memories, hopes or fears.
It sounds incredible, but it flows smoothly from the Theory Theory:
- Common-sense psychology is a theory, and mental states are theoretical entities whose whole nature the theory describes.
- So thoughts are whatever the theory of thoughts says they are.
- So if the theory is false, there is nothing for thoughts to be (as there would be no atoms if atomic theory were false).
- Common-sense psychology is false (Churchland's three reasons below).
- Therefore there are no thoughts.
Churchland's three reasons, and Crane's replies
| Churchland's charge | What he means | Crane's reply |
|---|---|---|
| Explanatory failure | Common sense leaves mental illness, creative imagination, sleep, perceptual illusions, memory and learning "wholly mysterious" | Why should a theory of belief and desire have to explain sleep? This demands too much of vindication |
| Stagnation | Little development in its long history (Churchland says, rather arbitrarily, twenty-five centuries) | It has changed: the acceptance of unconscious drives and desires is now part of common sense. And stability may signal a well-established theory: "unsupported bodies fall" hasn't changed either |
| Irreducibility | Belief and desire are unlikely to reduce to physical categories ("motherlove is magnesium") | This assumes science must reduce to physics, an extreme assumption Crane rejected in Chapter 1; vindication needs no reductionism |
Is eliminativism self-refuting?
To assert something is to express a belief in it. If there are no beliefs, nobody can assert eliminativism; eliminativists are just "vibrating the airwaves with meaningless sounds".
Churchland's vitalist. Churchland parodies the objection with nineteenth-century vitalism, which explained life by a life-spirit or "entelechy". Imagine a vitalist arguing: "My learned friend says there is no life-spirit. If that is true, he has no life-spirit and must be dead; if he is dead, his statement is a string of noises. So antivitalism, if true, cannot be true. QED." The argument is plainly bad; Churchland says the self-refutation charge against eliminativism is equally bad, since it assumes the very theory under dispute (that assertion is expressing belief).
Crane is unpersuaded by the analogy. We easily make sense of life without entelechy, but it is very hard to make sense of assertion without belief: assertion is itself a common-sense psychological notion (to assert is to claim true; to believe is to hold true). Eliminate common-sense psychology and assertion goes too, and with it the assertion of materialism. Churchland can reply that science's future should not be dictated by the limits of our imagination, as with vitalists who could not imagine life without entelechy. Crane agrees caution is wise but notes it does not get us anywhere near eliminativism. The self-refutation question need not be settled anyway, because the positive arguments are weak.
Still, the mere possibility of elimination disturbs many philosophers, because it is implicit in the Theory Theory. An empirical theory, true of the ordinary world of experience, must be falsifiable in principle, however confident we are, as with evolution or relativity. One escape is to deny that common-sense psychology is a theory at all.
Theory Versus Simulation
Many philosophers think the Theory Theory misrepresents what we do. Understanding others is imaginative projection into their minds, called "replication" or "simulation": we put ourselves in their shoes.
- Jane Heal: I try to replicate or recreate his thinking, placing myself in his initial state by imagining the world from his point of view, then deliberating to see what decision emerges.
- W.V. Quine (1960): attributing attitudes involves something like quoting one's imagined verbal response to an imagined situation; we cast "our real selves thus in unreal roles", even projecting into a mouse's state of mind, "such is our dramatic virtuosity".
On every version, mindreading is not applying a theory to behaviour but a skill: imagining ourselves into others' minds and predicting and explaining their behaviour as a result.
- Theory Theory (Morton's label). We understand others by applying a body of generalisations (largely tacit) about how perception, belief, desire and action interrelate; mental states are theoretical entities defined by that theory. Exposed to elimination if the theory is false.
- Simulation theory (Heal, Quine's remark; also Robert Gordon and Alvin Goldman). We understand others by running our own decision-making "offline" from their imagined starting point; no theory is consulted. Escapes the eliminativist argument's first premise, and explains why we cannot state powerful generalisations: there are none to state.
- Crane's resolution. Simulation describes what we are explicitly aware of in interpreting; the Theory Theory can say simulation is possible only because of tacit theory. Each can absorb the other's insight; "maybe there is no real dispute here, only a difference of emphasis".
Simulation does not rescue the mind from elimination
Simulation theory blocks one premise of the eliminativist argument: if common-sense psychology is not a theory, it does not compete with science. But this is a poor reason to adopt it, since simulation is equally congenial to eliminativism: if common-sense psychology doesn't even present itself as a proto-science, we needn't think it true at all (assuming only science tells us what exists). Quine held exactly this combination: in an "ideal last accounting of everything", positing mental entities "is not efficacious". So avoiding eliminativism is no motive for simulation theory. Its independent motive is Morton's dull truisms: we cannot produce powerful generalisations.
Is there a real dispute?
Simulation is obviously part of waking life; "seeing things from someone's point of view" is nearly synonymous with understanding them. So a Theory Theorist can say: agreed, that is how it seems, but you couldn't simulate without tacit knowledge of an underlying theory that makes simulation work. Since simulation concerns what we are explicitly aware of, it does not refute tacit knowledge of generalisations; simulation theorists need independent arguments. Crane concludes the best elements of both can be combined.
Missing options and later turns. Crane frames the debate as two-sided, but by 2016 there were well-developed alternatives. Daniel Dennett's "intentional stance" (1987) treats belief-desire attribution as a predictive strategy whose truth needs no inner theoretical entities, a middle path between vindication and elimination. Hybrid theories (Goldman's Simulating Minds, 2006) assign simulation and theory different jobs. And the embodied "interaction theory" of Shaun Gallagher holds that we often perceive others' intentions directly in expression and joint activity, the minority view Crane mentions only in a passing aside (that some deny we perceive only behaviour).
Mindreading science since 2016. Two developments reshape the theory/simulation debate. Cecilia Heyes's Cognitive Gadgets (2018) argues mindreading is a culturally learned skill, transmitted like literacy, not an innate module or pure simulation. And predictive processing (Andy Clark, Jakob Hohwy, Karl Friston's active inference) models the brain as minimising prediction error in a hierarchy; on its active inference variant, goals are recast as prior predictions the agent acts to make true. If that science matures, it would neither straightforwardly vindicate belief-desire psychology nor eliminate it, but re-describe desire and belief in a different vocabulary: a third outcome Crane's binary leaves out. 4E cognition (embodied, embedded, enactive, extended) has meanwhile grown into a research programme that rejects treating mindreading as inference from observed behaviour at all.
Conclusion: From Representation to Computation
How do we find out about minds? By applying conjectures, or a theory, to explain behaviour. What do we learn? That thoughts are states which represent the world and have effects in it. Since our evidence is behaviour, and not necessarily verbal behaviour, these ideas apply to other animals too. How far down?
The Tunisian desert ant. C.R. Gallistel describes a long-legged, fast-moving ant leaving its humid nest on the featureless Tunisian desert, looping tortuously ever further away, until it finds a scorpion carcass, gouges out a chunk nearly its own size, then turns to within one or two degrees of the straight line to the nest entrance, a 1-millimetre hole 40 metres away. It runs straight for 43 metres, holding course by its angle to the sun; three metres past where the entrance should be, it breaks into a search pattern and finds it. Gallistel finds it "hard to resist the inference" that the ant represented at every moment its position relative to the nest, enabling it to compute the solar angle and homeward distance.
The ant is explained by representations of locations, and Gallistel adds that it computes them. What could that mean, and why is the inference hard to resist? Many think the same is true of us: as we move about and think, we compute representations. That is the next chapter's subject.
How the ant measures distance. The ant is a Cataglyphis desert ant, studied for decades by Rüdiger Wehner. In 2006 Matthias Wittlinger, Wehner and Harald Wolf glued stilts onto some ants' legs and trimmed others' to stumps after they reached a feeder: on the way home, stilted ants overshot the nest and stumped ants stopped short, showing the ant measures distance by counting steps, a living pedometer feeding its path-integration computation. It is a vivid case for Gallistel's claim that the ant computes over representations of distance and direction.
Computation and Representation
Thoughts must be both representations of the world and causes of behaviour, and a computer is the clearest example of a causal mechanism that contains representations. Crane builds the idea from the ground up: functions, algorithms, Turing machines, coding, and automation, arriving at the definition a computer is a device that processes representations systematically according to rules. The chapter's key distinction is between instantiating a function (as planets do) and computing one (which requires representations), and so between being modelled by a computer and being one.
Asking the Right Questions
What people do is caused by how they represent the world to be. To explain thought we must explain how a state can at once represent the world and cause behaviour. Many psychologists, cognitive scientists and philosophers think the mind is a kind of computer, and the link to this theme is that a computer is a causal mechanism that in some sense "contains" representations. The idea excites some and strikes others as preposterous or degrading; Crane promises a fair hearing.
Start with the right questions. "Can the mind be modelled on a computer?" is the wrong one: the British Treasury builds computer models of the economy, yet no one thinks the economy is a computer. Crane separates:
- Can a computer think? More precisely, can anything think simply by being a computer? (Chapters 6 and 7.)
- Is the human mind a computer? More precisely, are any actual mental states and processes computational? (Chapter 8.)
To prepare, abandon your presuppositions: keyboards, screens, metal, plastic and silicon chips are inessential; a computer need not even be electronic. The target definition: a computer is a device which processes representations in a systematic way. Getting there needs two ideas: computation, and how computation can be automated.
Functions and Algorithms
Functions
Arithmetic gives the paradigm. Feed 7 and 5 into addition and get 12. Strip the numerals from "7 + 5 = 12" and you have a pattern with gaps, written with variables as x + y. The inputs are the function's arguments; the output is its value. The defining feature: for any given arguments, there is exactly one value.
The notion is far more general than arithmetic. "The natural father of x" is a function from people to their fathers (everyone has exactly one). In logic, "and" and "or" are truth-functions: P&Q takes the value True when both arguments are true and False otherwise. We rely on functions constantly, every time we total prices in a supermarket.
Algorithms
Knowing what a function is differs from knowing how to find its values. Long multiplication of 127 × 21 (127 + 2540 = 2667) seems banal, yet it is enormously powerful: a method that is entirely general (works for any two numbers) and entirely unambiguous (at every stage you know what to do next). Contrast the rough-and-ready rules for meeting new people (introduce yourself, smile, shake hands), which yield no definite answers and sometimes backfire.
An algorithm is a method for finding the value of a function. Algorithms are also called effective procedures (applied correctly, they are wholly effective) or mechanical procedures (Crane avoids this term, since he uses "mechanical" more loosely). A function is computable when there is an algorithm that gives its value for any argument, and the mathematical theory of computation is the theory of computable functions.
Three clarifications:
- Algorithm is not function. One function can have many algorithms: 127 × 21 by long multiplication, or by adding 127 to itself 20 times.
- An algorithm needn't yield a number. Testing whether one number divides another without remainder may return "no"; what the algorithm guarantees is a procedure for finding out whether there is an answer.
- Two conditions define an algorithm. (1) At each stage there is a definite thing to do next, with no guesswork, insight or inspiration; (2) the procedure can be specified in a finite number of steps.
A worked algorithm: multiplication by repeated addition
The example (from Ned Block) uses three pieces of paper: X (first number), Y (second number) and ANSWER.
| Pass | X after step (iii) | ANSWER after step (iv) | Step (ii): X = 0? |
|---|---|---|---|
| Start | 4 | 0 | No |
| 1 | 3 | 5 | No |
| 2 | 2 | 10 | No |
| 3 | 1 | 15 | No |
| 4 | 0 | 20 | Yes: stop |
4 × 5 = 20. Laborious, but entirely effective: at each stage it is completely clear what to do next, and for any pair of finite numbers it terminates in finitely many steps, however many.
Note that steps (iii) and (iv) use other algorithms (subtraction, addition). That is legitimate provided those have algorithms too, which they do; long multiplication likewise uses addition. Most algorithms are built from simpler algorithms plus simple moves between steps, an idea central to how computers work.
Because flow charts can describe almost any procedure, algorithms can be written for all sorts of things, including recipes.
- Turn on the stove.
- Fill the pan with water.
- Place the pan on the stove.
- When the water boils, add one egg and set the timer (if not boiling, wait).
- When the timer rings, turn off the gas.
- Remove the egg from the water.
- Result: one boiled egg. Finite steps, a definite next action at each, no inspiration required.
Turing Machines
Algorithms are at least as old as Ancient Greek mathematics, but only in the 1930s did mathematicians try to make the concept precise. Intense work on the foundations of mathematics since the late nineteenth century raised the question: is there a general way to tell, for any proposed method of calculation, whether it is an algorithm? If we cannot say what algorithms are, we cannot say what mathematics is.
Alan Turing (1912-1954) answered by asking: what is the simplest possible device that could perform any computation whatsoever? His answer, now called a Turing machine, is not a physical machine but an abstract specification of a possible machine, designed to expose the general properties of algorithms (people have built them, but that is not the point).
Turing, Enigma and some dates. Crane says Turing published in 1937 and "cracked the Nazis' Enigma code". The paper "On computable numbers" was submitted in 1936 and appeared in the Proceedings of the London Mathematical Society across 1936-37; the book's own chronology says 1936. Enigma was first broken by Polish cryptanalysts led by Marian Rejewski in the early 1930s; at Bletchley Park Turing led the attack on naval Enigma and designed the Bombe (improved by Gordon Welchman), one contribution among many. The U-boats were devastating Allied merchant shipping in the Atlantic more than "the British Navy" as such.
The anatomy of a Turing machine
Every Turing machine has an infinite tape divided into squares and a device that reads and writes symbols, one square at a time, while in one of a finite set of internal states. Assume just two symbols, 1 and 0. The device can do only four things:
- Move the tape one square, left or right.
- Read the symbol on the current square.
- Write a symbol, on a blank square or over another symbol.
- Change its internal state.
Its possible operations are given by its machine table, in effect its program: instructions of the form "if in state X reading symbol S, then write/erase, change to state Y (or stay), and move the tape". What the machine does at any moment is entirely fixed by two things: its current state and its current input symbol.
Interpreting the tape
To get the machine to compute anything, the symbols must be interpreted. Crane uses a prisoner's tally: a string of n 1s represents the number n (111 is 3), and 0s are mere punctuation separating numbers. So the tape ...000011100111111000100... represents the sequence 3, 6, 1, however many 0s sit between.
A worked example: the "add 1" machine
The machine needs just two internal states, A (initial) and B.
| State | Reading 1 | Reading 0 |
|---|---|---|
| A | Change to B; write 1; move tape one square right | Stay in A; write 0; move tape one square right |
| B | Stay in B; write 1; move tape one square right | Change to A; write 1; STOP |
Feed it the tape 00011000, representing 2. It starts in state A, reading the rightmost square. Because the tape moves right under a fixed reader, the reader works its way leftward along the symbols:
| Line | State | Symbol read | Action | Tape afterwards |
|---|---|---|---|---|
| (i) | A | 0 (8th square) | Write 0, stay A, move tape | 0 0 0 1 1 0 0 0 |
| (ii) | A | 0 (7th) | Write 0, stay A, move tape | 0 0 0 1 1 0 0 0 |
| (iii) | A | 0 (6th) | Write 0, stay A, move tape | 0 0 0 1 1 0 0 0 |
| (iv) | A | 1 (5th) | Write 1, change to B, move tape | 0 0 0 1 1 0 0 0 |
| (v) | B | 1 (4th) | Write 1, stay B, move tape | 0 0 0 1 1 0 0 0 |
| (vi) | B | 0 (3rd) | Write 1, change to A, STOP | 0 0 1 1 1 0 0 0 |
| (vii) | Halted | Output | 0 0 1 1 1 0 0 0 = 3 |
The machine has computed the function x + 1 for the argument 2, using only the four primitive operations. That is the whole lesson: any computation reduces to combinations of read, write, move and change state.
Run it yourself on 4. Write 0 0 1 1 1 1 0 0 and start in A on the rightmost square. Two 0s pass in state A; the first 1 flips the machine to B; three more 1s keep it in B; the next 0 becomes a 1 and the machine halts with five 1s. Now try starting on the leftmost square instead: in state A the reader moves off leftward over endless blank 0s and never halts. The same machine table computes x + 1 or computes nothing, depending on where the reading starts, which shows how much a computation depends on the conventions surrounding the table, not just the table.
Internal states are not hidden
It is natural to suspect that the "internal states" smuggle in the real calculating. They do not. State B just is the state such that reading 1 makes the machine do so-and-so and reading 0 makes it do such-and-such. That is all there is to it; "internal" misleads by suggesting a hidden nature. Complex computations (such as the multiplication algorithm) need more states, more tape and richer notation, but no more sophisticated basic operations.
Notation and binary
The tally notation has drawbacks: it cannot represent 0, and big numbers are slow (adding 1 to 7,000,000 would need more squares than London has inhabitants). Binary (base 2) represents every natural number with 1s and 0s, the column values being powers of 2 instead of powers of 10.
| Denary | Binary |
|---|---|
| 1 | 1 |
| 2 | 10 |
| 3 | 11 |
| 4 | 100 |
| 5 | 101 |
| 6 | 110 |
| 7 | 111 |
| 8 | 1000 |
Binary makes 0 a numeral rather than mere punctuation, yet the machine still needs only the same basic operations (read 1, write 1, read 0, write 0, move). Punctuation between instructions can itself be coded in 1s and 0s, with ingenuity.
The universal Turing machine
Now the exciting step. The labels "A" and "B" were arbitrary (%, @, * would do), so states and operations can be numbered, and in binary written on a tape. Since a Turing machine is completely defined by its machine table, and any table can be coded numerically, any Turing machine can be written on the tape of another. Code the "add 1" machine in binary, append the number 127, and feed both to a second machine whose program interprets the first: it mimics the add-1 machine and adds 1 to 127.
There is a Turing machine that can mimic any other Turing machine: the universal Turing machine. We need not build a separate machine for each operation, only one that can imitate them all. This is the idea behind the general-purpose digital computer, and Crane does not think it an exaggeration to say it has affected the character of all our lives.
Church's thesis
What can particular Turing machines do? Turing claimed that any computable function can in principle be computed on a Turing machine, given enough tape and time, so any algorithm can be executed by one. Most logicians and mathematicians now accept that to be an algorithm is to be executable by some Turing machine. This is Church's thesis, after Alonzo Church (1903-1995), who independently reached similar conclusions; it is also called the Church-Turing thesis. Its job is to give precise sense to "algorithm". The point of the tedious tapes is that Turing reduced anything we recognise as an effective procedure to steps so simple no one could think them mysterious: he made effective procedures unmysterious.
Coding and Symbols
A Turing machine is an input-output device: a tape of 1s and 0s goes in, another comes out, transformed according to the machine table. But how can it run an algorithm for boiling an egg or finding which key opens a lock, when only numbers go on the tape? It cannot boil an egg, but the algorithm is a description of how to boil one, and descriptions can be coded.
- Letters to numbers. Assign A = 1, B = 2, C = 3 and so on, so TIM CRANE becomes 20 9 13 / 3 18 1 14 5.
- Punctuation. Mark where letters, words and whole texts end; no problem of principle (old telegrams spelled out "STOP" between sentences).
- Binary. Rewrite the numbers in base 2 and put them on the tape.
Real programmers don't use this scheme; the abstract point is that any text, hence any algorithm stated in any language, can be coded as numbers and run on a universal Turing machine.
Digital representation. Each element is represented in an on-off way: every square holds 1 or 0, nothing in between. A digital clock shows distinct numbers for each second and nothing between them. The computers discussed in debates about thought are all digital.
Analogue representation. Representation by continuous variation: an analogue clock marks time by the smooth sweep of a hand. Analogue computers exist but are not directly relevant to Crane's questions.
The definition is now nearly complete. "Processes in a systematic way" is explained by algorithms, made precise by Turing machines. "Representations" is implicit in the Turing machine: for it to compute a function, its symbols must be taken to stand for something, and other representations (English sentences) can be coded into those. Calling computers "information processors" or "symbol manipulators" says the same thing: they process representations, which carry information because they "say" or are interpretable as saying something, and they do it by effective procedures.
Instantiating a Function Versus Computing a Function
Functions also describe nature. Newton's second law, F = ma, uses the mathematical function multiplication to represent a relationship among masses, forces and accelerations, itself a function in nature ("Newton's function"). But a body accelerating under a force does not compute Newton's function; if it did, every physical object would be a computer. It instantiates the function: it is an instance of it, or describable by it. Likewise the planets move in ways described by Kepler's laws, but they do not compute their orbits; they just move.
Compare a physical adding machine. We say it computes addition, but strictly numbers cannot be fed into machines; you don't destroy the number 3 by destroying every written "3". What the machine takes and gives are numerals, representations of numbers. That is the difference from the planets.
Computing a function requires representations: representations as input and representations as output. In Jerry Fodor's slogan: "No computation without representation!"
Church's thesis implies that any natural process describable step by step can be modelled by a Turing machine, which merely specifies transitions between the process's states. That no more makes the process a computation than representing my body temperature by a number makes my temperature a number. If a theory of a phenomenon can be represented algorithmically, the theory is computable, a fact about theories, not phenomena. Hence the opening distinction:
- A system can be modelled on a computer when a theory of that system is computable (the economy, the weather, the solar system).
- A system performs computations when it processes representations by using an effective procedure.
Is representation really essential to computation? Crane follows Fodor in building representation into the definition. That is contested. Gualtiero Piccinini's mechanistic account (Physical Computation, 2015) defines computation as the manipulation of medium-independent vehicles according to rules, with no semantic content required; others (Stephen Stich's "syntactic theory of mind", Frances Egan) treat content as a gloss we add to a purely formal process. If they are right, "computing versus instantiating" cannot be drawn by representation alone, and the worry that every physical system computes something (pressed by Hilary Putnam and John Searle) returns.
Automatic Algorithms
Turing machines describe the abstract structure of computation, but "moving", "reading" and "writing" were taken for granted. How does any effective procedure run without a human at every step? Real computers use automated algorithms: algorithms and input/output representations "embodied" in the physical structure of the machine.
Black boxes and functional analysis
The mousetrap. A mousetrap takes live mice as input and gives dead (or trapped) mice as output. Treated as a black box, what's inside doesn't matter, only the input-output task. Break into it and find two components, bait and trapping device (the mouse moves from bait to trap, not the reverse), each itself a black box defined by what it does. Break into the trapping device of a comic-book trap and find a spring and a metal bar released when the bait is taken, also black boxes. What is one black box at one level decomposes into others at the next, until we understand the whole.
This is functional analysis ("functional boxology"): explaining a machine by the functions of its component parts. Here "function" means the causal role a part plays in the system ("what's the function of this bit?"), a different sense from the mathematical function computed.
Dennett's homunculi
Apply this to a machine running the multiplication algorithm. At the top it is a multiplier black box (numerals in, product out). Inside is the flow chart; each step (adding, subtracting, reading) is itself a step-by-step procedure with its own flow chart, and so on down to the simplest tasks. Daniel Dennett offers a vivid image:
The intelligent multiplier is functionally explained by postulating progressively stupider homunculi. If we can build a physical device that does the job of a stupid homunculus, we can combine such devices into a multiplier, since the multiplier is nothing more than simple devices arranged as the flow chart specifies. Turing showed that any algorithm breaks down into the primitive Turing machine operations, so all we need are devices that move, read and write.
Think of the tape as an array of switches, on for 1 and off for 0. A machine that steps along the switches, registers their position ("reading") and turns them on or off ("writing"), programmed with which Turing machine to mimic, is a computer built out of switches. Real computers are built from switches too, though less simply: one of the earliest (1944) used telephone relays; ENIAC used valves; relays and valves are in effect switches. The real advances came with semiconductor switches, computing faster than Turing ever dreamed, and high-level programming languages that make basic operations perform complex ones.
- 1642 Pascal's mechanical adding machine.
- 1822 Babbage begins the difference engine for polynomial functions.
- 1854 Boole's Laws of Thought.
- 1936 Turing's "On computable numbers" describes the Turing machine.
- 1941 Konrad Zuse builds a computer to aid aircraft design.
- 1943-44 Colossus at Bletchley Park.
- 1944 Harvard Mark I (Howard Aiken with IBM).
- 1945 John von Neumann's EDVAC design: stored program plus CPU, the von Neumann architecture.
- 1946 ENIAC (Eckert and Mauchly, University of Pennsylvania), about a thousand times faster than Mark I.
- 1950 Turing's "Computing machinery and intelligence" proposes the imitation game.
- 1958 Jack Kilby's integrated circuit.
- 1971 Intel 4004, a whole processor on one chip; Longuet-Higgins coins "cognitive science".
- 1997 Deep Blue defeats world champion Kasparov.
- 2011 IBM's Watson wins Jeopardy!
Corrections to the book's computing chronology and text. The chronology calls Harvard's Mark I "the first fully electronic calculator"; it was electromechanical (relays, shafts and clutches). ENIAC was the Electronic Numerical Integrator and Computer (not "calculator"), completed in 1945 and unveiled in February 1946, and was built to compute artillery firing tables, not "missile trajectories". Colossus, designed by Tommy Flowers of the Post Office, attacked the Lorenz teleprinter cipher, not Enigma. The transistor was invented at Bell Labs in December 1947 (Bardeen, Brattain, Shockley). Kilby's first integrated circuit was germanium; silicon chips followed with Robert Noyce (1959). The world champion's name is usually spelled Garry Kasparov.
Conclusion: What Is a Computer?
- Computer. A machine that processes representations systematically according to rules.
- Function. A mapping giving exactly one value for each set of arguments.
- Algorithm (effective procedure). A finite, step-by-step method for finding a function's value, with a definite next step at every stage.
- Computable function. One for which an algorithm exists.
- Turing machine. An abstract tape-and-reader device whose machine table fixes what it does from current state plus current symbol.
- Universal Turing machine. One that mimics any Turing machine whose table is coded on its tape.
- Church's (Church-Turing) thesis. Any computable function can be computed by a Turing machine.
- Coding. Systematic mapping of any representation (text, instructions) into numbers and then binary.
- Instantiating versus computing. Being describable by a function versus transforming representations of its arguments into representations of its values.
- Automatic algorithm. An algorithm embodied in physical structure, its primitive steps performed by simple devices such as switches.
The definition's consequence: it doesn't matter what a computer is made of. What matters is what it does, the computational tasks it performs or the program it runs. Today's machines use microscopic circuits on silicon, but the same tasks could in principle be done by switches, beads, matchsticks and tin cans, "and even perhaps by the neurochemistry of the brain". This is variable (or multiple) realisation: the same program (software) can be realised by different hardware.
Heuristics: where algorithms run out
Not all real computing is strictly algorithmic. In chess the rules fix legal moves unambiguously, and since a game ends in finitely many moves, a machine could in principle look ahead to every consequence of every move, but this would take even the most powerful computer unimaginably long. Designers therefore add heuristics, rules of thumb such as "castle as early as possible", which suggest good courses of action but, unlike algorithms, guarantee no outcome. Heuristics have been very influential in artificial intelligence, sometimes described as the science of thinking machines. What it means for a computer to think is the next chapter's question.
A garbled number. The printed text reports John Haugeland's estimate as a look-ahead of "10 moves", which could not exceed the quantum states of the universe. An exponent has evidently been lost: the comparison only makes sense for a number of the order of 10^120, Claude Shannon's classic (1950) estimate of the number of possible chess games, and the scale of number Haugeland was invoking.
Representations inside neural networks. Since 2016 the computers that matter most are not hand-programmed symbol processors but trained neural networks, and Crane's question of whether they compute over representations or merely instantiate a function has become empirical. Mechanistic interpretability finds internal representations directly. Kenneth Li and colleagues (2023) trained a transformer only on legal Othello move sequences and found it built an internal model of the board that probes could read and that, when edited, changed the moves the model predicted; Neel Nanda and colleagues showed the board state is linearly encoded (as "my piece / their piece"). Anthropic's dictionary-learning work (2023-2024) used sparse autoencoders to extract millions of interpretable features from Claude 3 Sonnet, including one for the Golden Gate Bridge whose amplification made the model talk obsessively about the bridge. These states are causally efficacious and content-bearing in Crane's sense, though whether their content is original or derived from training data and human interpretation remains open.
Chess after heuristics. Crane's picture of chess programs as search plus hand-written heuristics describes Deep Blue (1997). In 2016 DeepMind's AlphaGo beat Lee Sedol at Go, and AlphaZero (2017) learned chess, shogi and Go from self-play alone, replacing hand-coded rules of thumb with a learned evaluation network guiding a Monte Carlo tree search. The "heuristics" are now learned representations, not rules a programmer can read off.
Architects already think in Crane's terms. Variable realisation is why the same service contract can run on mainframe, cloud or GPU; black-box functional analysis is how we decompose systems into components with defined interfaces; and the model versus computation distinction is the difference between a risk model that describes a portfolio and a decision engine that acts on representations of it. The interpretability results above matter for banking because they make what a model represents auditable in principle, not just what it outputs.
Can a Computer Think?
Computers process representations and thoughts are representations, but that overlap proves nothing: books contain representations and cannot think. Crane clears away bad arguments on both sides, defines AI as the view that a computer can think simply by being a computer, and tests it against the two classic objections: Hubert Dreyfus's claim that thinking rests on common-sense know-how that rules cannot capture, and John Searle's Chinese room, which argues that syntax is never sufficient for semantics. His verdict: Searle's premise is right, so nothing thinks merely by computing, yet computation may still be part of how we think.
Bad Arguments For, and a Bad Argument Against
The question is why anyone would think that processing representations systematically could constitute thinking. There are tempting links: representation figures in the definitions of both thought and computers, and in both cases the representations seem to cause what happens. An adding machine outputs 5 after 2, +, 3, = partly because those symbols represent what they do. Crane warns against three bad inferences from such parallels.
- The book fallacy. Representation also defines books, and books cannot think, so sharing the notion of representation shows nothing about whether computers think.
- The equivocation on "information processing". Thought processes information in a loose, everyday sense; computers are information processors in a precise technical sense. Moving from one to the other is a fallacy of equivocation.
- The machine-table fallacy. That there is a Turing machine table for thinking would only show that the theory of thinking is computable. If astronomy were computable, that would not make the universe a computer: again, computing a function is not instantiating it.
The opposite mistake is too-quick debunking. John Searle mocks the habit of modelling the brain on the latest technology: in his childhood it was a telephone switchboard; Sherrington thought it a telegraph system; Freud compared it to hydraulic and electromagnetic systems; Leibniz to a mill; some ancient Greeks to a catapult. Why should a brain evolved over millions of years be explained by ideas from sixty or seventy years of speculation on the foundations of mathematics?
Why computers are different from mills and catapults
Crane's reply is historical: the invention of the computer always went hand in hand with attempts to systematise human thought, which is not true of hydraulics or switchboards.
- Leibniz (1646-1716). He proposed a "universal character" (characteristica universalis), a precise, unambiguous language into which ideas could be translated so that disputes could be settled by calculation; he expected it to strengthen the mind more than telescopes and microscopes strengthen the eye. He never built it, but he did develop binary notation. The combination is telling: strip thought of ambiguity, then build a calculus to process the skeleton.
- George Boole (1815-1864). In The Laws of Thought (1854) he expressed logical relations between propositions algebraically, using 1 for true and 0 for false: 1 × 0 = 0 reads "a truth and a falsehood gives a falsehood"; 1 + 0 = 1 reads "a truth or a falsehood gives a truth". He claimed to have found the "fundamental laws of those operations of the mind by which reason is performed".
- Logic gates. Boole's algebra became central to digital circuit design. An and-gate outputs a current when and only when it receives current from both inputs X and Y, so it implements the truth-function "and"; similar logic gates implement the other Boolean operations.
- Babbage, Turing, Church. With Charles Babbage (1792-1871), these ideas led to the general-purpose programmable digital computer, realised through the theoretical work of Turing and Church and post-war electronics. No wonder the public hailed computers as "electronic brains".
| Inputs X, Y | Boole: × ("and") | And-gate output | Boole: + ("or") |
|---|---|---|---|
| 1, 1 | 1 | current | 1 |
| 1, 0 | 0 | no current | 1 |
| 0, 1 | 0 | no current | 1 |
| 0, 0 | 0 | no current | 0 |
Two historical simplifications. Boole's own "+" was defined only for mutually exclusive classes (closer to exclusive "or"); the inclusive "or" with 1 + 1 = 1 came from William Stanley Jevons and later algebraists, so the tidy table above is modern Boolean algebra rather than Boole's 1854 system. Leibniz published binary arithmetic in 1703 and gave it its fullest early treatment, but binary systems had been described before him (by Thomas Harriot and Juan Caramuel, among others), so "invented" overstates it.
What Artificial Intelligence Is
"AI" names several projects. Alan Garnham calls it "the science of thinking machines"; Margaret Boden, more ambitiously, "the science of intelligence in general". "Intelligence" misleads newcomers: much AI research targets things we would never call intelligent, such as identifying and manipulating simple 3D objects or understanding simple text. Systems for drug diagnosis, or Google's search and translation, are sophisticated but not intended to be thinking computers. The philosophically interesting project has always been building a thinking machine, which Boden's view says should teach us what thought is in general.
The aeroplane problem
Birds fly and so do aeroplanes, but building aeroplanes teaches little about how birds fly. A thinking computer might likewise think differently from us. Yet thinking is what we do, so how could anything think except our way? This generates a dilemma:
- Horn one: thinking means thinking as humans do. Then a thinking computer is possible only if human thinking is computational, which requires detailed psychological investigation first. This collaboration of psychology and AI is cognitive science (Chapter 8).
- Horn two: things could think differently. Then AI need not study human psychology; it should build machines that do tasks which would require thought if done by people. This was the method of early AI in the 1950s.
The second horn invites the objection that it produces only a simulation of thought. Some shrug ("if it does the job intelligently, who cares?"), but that undermines AI's claim to be the science of intelligence, since blurring real thought and simulation can tell us nothing about our own real thought.
Turing's redefinition
Crane traces the blurring to Alan Turing's "Computing Machinery and Intelligence" (1950). Finding "Can a machine think?" too vague, Turing replaced it with the question of when a machine would be mistaken for a thinking person. In the Turing test, an interrogator communicates at a distance with a machine and a person; if the interrogator cannot tell which is which, the machine thinks. Crane holds the test's assumptions are behaviouristic, hence inadequate as an account of thinking. His point here is narrower: if passing the Turing test suffices for thought, then only performance matters, not whether the machine performs the way humans do, and that is what let AI blur the line between real thought and simulation.
This sharpens the two questions from Chapter 6:
- Can a computer think? That is, can something think simply by being a computer?
- Is the human mind a computer? That is, do we think, in whole or part, by computing?
They are independent. One can say yes to 1 and stay agnostic on 2 ("I don't know how we think, but here's a computer that can!"), or yes to 2 and no to 1 ("nothing thinks simply by computing, but computing is part of how we think"). For clarity, Crane uses "AI" for the thesis that a computer can think. Two classic objections stand against it.
Dreyfus: the competence objection. Computers must obey rules (algorithms or heuristics), but thinking cannot be captured in any system of rules. It needs active engagement with life, participation in a culture, and know-how that cannot be formalised. Conventional AI will never even pass for general intelligence.
Searle: the understanding objection. Grant that a program could pass the Turing test. It would still only simulate understanding, because computers manipulate symbols by their form alone and are insensitive to meaning. Success in behaviour is not sufficient for thought.
Can Thinking Be Captured by Rules? Dreyfus's Critique
The rookie bus driver. The Arizona Daily Star of 31 May 1986 reported that a new driver was suspended for failing to act when a girl had a heart attack on his bus. A union official blamed overly strict rules forbidding drivers to leave their route without permission; a company spokesman defended them: "You give them a little leeway, and where does it end up?" The story, used by sociologist Harry Collins, illustrates the worry: by sticking to the rule, the driver could not respond intelligently.
The quick objection runs: computers must stick to strict rules, thinking is flexible, so computers cannot think. Crane finds it too quick: perhaps the fault lies with this rule, not with rules as such. So amend it: leave your route with permission unless there is a medical emergency, then drive to the nearest hospital. But what if the hospital is under siege by terrorists, or a doctor is on board? Each patch spawns new exceptions, and "we don't even know how to fill in the dots". The real problem is to find a rule specific enough to direct action ("drive to the nearest hospital", not "do something sensible") yet general enough to cover all eventualities.
George Orwell's rules for writing ("Never use a long word where a short one will do") end with: "Break any of these rules sooner than say anything outright barbarous." The bus company's equivalent would be "break any of these rules sooner than do anything stupid", in other words, "Use your common sense!" Humans can be relied on for that. A computer, which works by rules and representations, would need common sense stored as explicit rules and representations, so AI needs a way to program computers with explicit representations of common-sense knowledge.
Know-how versus knowing-that
Hubert Dreyfus argues this cannot be done. Human intelligence needs "the background of common sense that adult human beings have by virtue of having bodies, interacting skilfully with the material world, and being trained in a culture", and that background cannot be a vast base of propositional knowledge.
The argument turns on the distinction between knowing that (facts, the stuff of books: Sofia is the capital of Bulgaria) and knowing how (skills: riding a bicycle). Many philosophers hold that know-how cannot be reduced to knowledge of rules: nobody rides by applying "when turning right, lean slightly right"; you get the hang of it by trial and error. Dreyfus says general intelligence is like that. Knowing what a chair is is not knowing a definition but knowing how to sit on chairs, get up from them, spot them in a room, improvise one when none is around: a repertoire of bodily skills "which may well be indefinitely large", given the variety of chairs and of graceful, comfortable, poised ways to sit in them.
- General intelligence requires common sense.
- Common sense is, or rests on, know-how (skills, not facts).
- A computer is a device that processes representations according to rules; representations and rules are not skills (a book can contain representations of rules but has no skills).
- So a computer's knowledge, if any, is knowledge that, never knowledge how.
- Conclusion: computers cannot have common sense, so AI cannot create a computer with general intelligence.
AI's defenders can reject premise 1, which is unpromising (what would general intelligence without common sense be?), or reject premise 2, the usual move: making the assumptions of common sense explicit is hard work, not impossible work. But even a complete store would be only the beginning: the machine must retrieve and use the knowledge intelligently. "It's not enough to have an encyclopaedia, one must know how to look things up in it."
Relevance and the frame problem
The key notion is relevance: a computer that cannot tell which facts bear on which will use its stored common sense badly. Relevance shifts with world-views: a person's sex was thought relevant to the right to vote two hundred years ago, and is not now. Relevance travels with a sense of what is out of place or exceptional. Dreyfus's test for a program meant to understand restaurant stories (in the style of Roger Schank's "scripts"): until it can say whether the waiter wore clothes, walked forwards or backwards, and whether the customer ate with his mouth or his ear, all its right answers look like "tricks or lucky guesses". We know these things, Dreyfus argues, because our way of living rests on skills and interaction with things, not stored propositions.
These problems are sometimes grouped as the frame problem, among the hardest for traditional AI. Crane lists the responses:
- The "Google response". Improve the algorithms, store more data, sharpen the machine's sense of relevance. Crane notes that in 1993, when he wrote the first edition, today's storage and speed were unimaginable, so predictions about what computers cannot do need care. Yet nobody concludes that Google search is really thinking, however good it gets. Computers now excel at what sceptics thought impossible (face and text recognition, automated translation), but there is less and less tendency to think AI is on the brink of a thinking machine; "AI" has become a branch of engineering, not part of the science of mind.
- Separate the facts from the PR. IBM's Watson won the $1M top prize on Jeopardy! in 2011. IBM's website said Watson was "built to mirror the same learning process that we have" (Observe, Interpret, Evaluate, Decide) and, a line later, that it "can read millions of unstructured documents in seconds". We observe, interpret, evaluate and decide without reading millions of documents in seconds. As machines get better, their serious claims to be "doing what humans do" retreat into the shadows.
- Concede and change approach. Admit that "classical" rules-and-representations AI has failed to capture thought's basic abilities and look to robotics or a more brain-based view of computation (Chapters 8 and 11).
- Give up on thinking machines altogether.
At minimum Dreyfus sets traditional AI a challenge: represent common sense in rules and representations; at most he signals the breakdown of the idea that thought is rule-governed symbol manipulation. Either way, Crane thinks, scepticism about building a thinking computer on this conception is licensed.
"Frame problem" is used loosely here. In its original sense (John McCarthy and Patrick Hayes, 1969) the frame problem is a technical puzzle in logic-based AI: how to represent what does not change when an action occurs without writing endless "frame axioms". Philosophers (Dennett's robot-and-bomb story, 1984; Fodor) widened it into the general problem of relevance and holistic updating, which is Dreyfus's concern. Crane's text uses the wide sense without flagging the narrow one. Note too that Crane wrote in 2015-16 that AI had stopped claiming to build thinking machines; that description has not survived (see the update below).
Dreyfus after large language models. Dreyfus (who died in 2017) aimed at hand-coded rules and representations, and on that target history largely vindicated him: the large common-sense project Cyc (begun by Douglas Lenat in 1984) never delivered general intelligence. The systems that now handle restaurant-style questions are trained networks, not rule bases. Large language models answer Dreyfus's waiter questions easily, and benchmarks built to require common sense, such as the Winograd Schema Challenge, have been largely saturated by them. Whether this refutes Dreyfus is disputed: his deeper claim was that understanding needs embodied, skilful coping, and LLMs still fail in revealing ways on physical and situational reasoning and on novel variants of familiar puzzles. They are also, notably, not rule-followers in Crane's sense: their "knowledge" is distributed in weights, closer to the connectionist picture of Chapter 8. Crane's remark that AI had become "just engineering" has been overtaken: leading labs now state artificial general intelligence as an explicit goal.
The Chinese Room
Dreyfus says conventional AI will never pass for general intelligence. John Searle (1980) grants, for argument's sake, that a program could pass the Turing test, and argues that it would still be simulation, not the real thing.
Searle's Chinese room. Searle imagines himself in a room with two windows, I (input) and O (output). Through I come papers with complex markings. The room holds a huge rule-book in English: "whenever you get a paper with these markings, do such-and-such and pass out a paper with those markings through O", plus piles of marked paper (the data). Unknown to Searle, the markings are Chinese characters: inputs are questions, outputs are sensible answers. Chinese speakers outside think they are conversing with someone who understands. But Searle understands no Chinese: he manipulates the symbols by their shape alone. The set-up mirrors a computer: rules (the program) operating on symbols. So running a program can never constitute understanding.
- Computer programs are purely formal or "syntactic": they are sensitive only to the shapes of the symbols they process.
- Genuine understanding (and by extension all thought) is sensitive to meaning ("semantics").
- Form (syntax) can never constitute, or be sufficient for, meaning (semantics).
- Therefore, running a computer program can never be sufficient for understanding or thought.
Premises 1 and 2 are meant to be uncontroversial; premise 3 carries the weight, and the Chinese room is its defence. ("Syntax" and "semantics" are explained in Chapter 8; for now read them as form and meaning.)
The replies and Searle's answers
Crane discusses the systems reply, the "failure of imagination" reply and a robot/interaction reply, then the Churchlands' luminous room. Searle's original 1980 paper also answered a brain simulator reply, a combination reply and an other minds reply; they complete the picture and are added here.
| Reply | What it says | Searle's answer | Assessment |
|---|---|---|---|
| Systems reply | Searle is only part of the computer (like the CPU). The whole system (man + rules + papers + room) understands Chinese. | Memorise the rules and data and work outside the room: you are now the whole system, and you still don't understand. | Searle's jibe that "a room" understands is philosophically irrelevant; the memorisation answer is his serious one. |
| Failure of imagination | Memorising the whole program is no trivial feat; who knows whether understanding would follow? | It is obvious that symbol-shuffling, however vast, adds no meaning. | Becomes the Churchlands' charge (below). |
| Robot / interaction reply | Let Searle act in the world: as a waiter in a Chinese restaurant, he would link one symbol with fried rice, another with shark-fin dumplings, and start to grasp meanings. | This concedes the point: understanding needs interaction with the world, not just a running program. | Crane: strictly, Searle is right. But the interacting system still manipulates symbols algorithmically, so cognitive science survives. |
| Brain simulator reply (Searle 1980) | Let the program simulate the actual neuron firings of a Chinese speaker's brain. | Imagine the man operating water pipes and valves that mirror the neurons: still no understanding; it simulates the brain's formal structure, not its causal powers. | Shifts the issue to whether biology or organisation matters. |
| Combination reply (Searle 1980) | Put a brain-simulating program in a robot body: the whole thing must understand. | We would attribute understanding until we learned it ran a formal program; then we would withdraw it. | Concedes the case rests on intuitions about what is inside. |
| Other minds reply (Searle 1980) | We credit people with understanding from behaviour; the same evidence should count for the machine. | The question is not how we know others understand but what we attribute when we do. | Ties back to Crane's charge that the Turing test is behaviourist. |
Crane's verdict
Searle is right about one thing: premise 3, syntax is not enough for semantics. Symbols do not "interpret themselves". This is really a bald statement of the problem of representation: if it were false, there would be no such problem. It does not follow that meaning cannot be explained (Chapters 9 and 10 try), but any explanation must avoid surreptitiously introducing understanding, meaning or semantics into its explanans. That is, for Crane, the main lesson of the Chinese room.
The luminous room
Paul and Patricia Churchland (Scientific American, January 1990) question whether Searle is even entitled to premise 3. They build a parallel:
- Electricity and magnetism are forces.
- The essential property of light is luminance.
- Forces cannot constitute, or be sufficient for, luminance.
- Therefore light is not electromagnetic radiation.
Premise 3 might be "supported" by a luminous room thought experiment: someone in a dark room waves a magnet, generating electromagnetic waves, yet the room stays dark however fast she waves. But light is electromagnetic radiation, so premise 3 is false, and the thought experiment cannot establish it. Likewise, the Churchlands say, whether syntax suffices for semantics is an empirical question, not one settled by imagination: Goethe found it inconceivable that small particles could constitute light, and people found it unimaginable that inanimate matter could constitute life. What people can imagine "often has nothing to do with what is or is not the case". Searle replies that it is not and could not be an empirical question whether syntax suffices for semantics, so the analogy fails; making sense of this needs the account of syntax and semantics in Chapter 8.
The China brain. Ned Block (1978) asked us to imagine the whole population of China, each person with a radio link, simulating one neuron each of a human brain for an hour, connected to a body. Functionally the system is organised like a brain in pain. Is anything feeling pain? Most people say no, which shows how the intuition behind the Chinese room generalises from understanding to experience. It also shows the intuition's weak point, which Dennett pressed: we imagine the system at human scale and speed, a few slow people with papers, when a real brain runs billions of operations at once. If our verdict changes when we picture the room running at the speed and scale of a brain, the thought experiment was measuring the limits of imagination, which is exactly the Churchlands' complaint.
The Chinese room meets large language models. The argument is now argued about with real systems. In a 2025 preregistered three-party Turing test, Cameron Jones and Benjamin Bergen reported that GPT-4.5, prompted to adopt a persona, was judged to be the human 73% of the time, more often than the real human participants; earlier, in 2022, Google engineer Blake Lemoine publicly claimed the LaMDA chatbot was sentient. Searle's followers say this confirms his point: behavioural success is exactly what the room was granted, and it settles nothing about understanding. Opponents note that an LLM is not a rule-book lookup but a trained network with structured internal representations (the interpretability results noted in Chapter 6), and that multimodal and robotic models supply something like the robot reply's grounding. The ground debate about meaning from form alone and derived versus original intentionality is the one set out in the boxes in Chapters 2 and 3; the Chinese room adds the question of whether any amount of performance could ever be evidence of understanding.
Conclusion: Can a Computer Think?
In 1965 Herbert Simon predicted that "machines will be capable, within twenty years, of doing any work that a man can do". Half a century on, the advances would surprise even Simon, yet the claim still looked fanciful in 2016.
Dreyfus and Searle both say no. Dreyfus's upshot: general intelligence needs common sense, which cannot be rules and representations, so a genuine thinking computer is impossible. Crane does not hold this against AI, because real AI is no longer the science of thinking machines; the failure belongs to the overblown rhetoric of AI's heady early days.
Searle's lesson is different. His argument begs the question, since it simply denies AI's central thesis that thinking is formal symbol manipulation, but his assumption is correct. The right reply is: Searle-in-the-room, or the room alone, cannot understand Chinese; let the outside world impact on the room and semantics might get a foothold. That reply concedes that thinking is not simply symbol manipulation. Nothing can think simply by being a computer, yet it may still be true that we think partly by computing, which is the question of the next chapter.
Read the Watson lesson into today's AI programmes. Crane's distinction between how a system actually works and the claims made for it is the core of model risk governance: an LLM copilot for relationship managers does not "understand the client" in Searle's sense; it produces text whose meaning is supplied by the people and records around it. Dreyfus's relevance problem is also the practical failure mode of retrieval-augmented systems: having the documents is the encyclopaedia, knowing which ones bear on this case is the hard part. Design controls (grounding, citation, human review of exceptions) at exactly the points where the bus driver's rule-book failed.
The Mechanisms of Thought
Even if nothing thinks merely by computing, our thinking might have a computational basis. Crane sets out the computational theory of cognition in its strongest form, Jerry Fodor's language of thought (Mentalese): beliefs are sentences in the head whose syntax mirrors their semantics and whose shapes drive their causal roles, which explains the systematicity and rationality of thought. He weighs two objections (the homunculus fallacy and following versus conforming to rules), then sets out the connectionist alternative, in which the objects of computation need not be the objects of representation, and the Fodor-Pylyshyn dilemma against it. The upshot: computation presupposes representation and so cannot explain it.
Cognition, Computation and Functionalism
The mechanical view says the mind is part of nature with a regular, law-governed causal structure. Saying that this structure is computational is a further step, but the association is as old as the mechanical world picture. Thomas Hobbes in Leviathan (1651) wrote that when a man reasons he does "nothing else but conceive a summe totall, from Addition of parcels", and that "REASON ... is nothing but Reckoning". John Haugeland reads this as a prefiguration of the computational view of thought.
The computational view differs from Chapter 7's thesis. One can deny that anything thinks just by computing and still hold that some of our mental states and processes are computational, without computation exhausting the nature of thought. The idea dominates current philosophy of mind and cognitive psychology, so it deserves a detailed look.
Scope: which states could be computational?
Computation is defined through representation, so only representational states can be computational. If, as some hold, bodily sensations have non-representational "qualia", the whole mind cannot be a computer. The rival view, representationalism or intentionalism, says all mental states in all their aspects are representational, removing the obstacle. Crane parks the dispute (until Chapter 13) and follows good method: assess a theory in its most plausible version, since "no-one is interested in a critique of a caricature". The best candidates are beliefs and desires: their essence is to represent, they need not be conscious, and a belief's nature is exhausted by how it represents the world.
Functionalism is not the computational theory
Readers who know 1960s functionalism may object: did it not classify mental states by Turing machine tables, with pain (input tissue damage, output moaning) as the paradigm? Crane's reading: machine tables were a way to give an abstract specification of mental state types that did not tie them to particular neural states, a reaction against the identity theory's "pain = C-fibre firing". Many physically different systems can be in the same mental state; the machine-table analogy showed how. But being describable by a machine table is not computing (the solar system lesson of Chapter 6).
Functionalism. The mind is defined by its causal structure: a mental state is whatever plays a certain causal role between inputs, outputs and other states. It says nothing about whether those transitions involve representations. It applies naturally to pain.
Computational theory of cognition. That causal structure is computational: a disciplined series of transitions among representations, governed by algorithmic (perhaps heuristic) rules. It applies most plausibly to beliefs, desires and reasoning. It is the philosophical basis of cognitive science.
Crane prefers "computational theory of cognition" to the common label "representational theory of mind" for two reasons: the latter claims the whole mind, which is contentious, and the idea that the mind "represents the world" is innocuous; almost every theory accepts it. What not every theory accepts is that the mind contains representations. Jean-Paul Sartre called representations "idols invented by the psychologists". Containing representations means distinct inner states that stand for things: as Crane thinks about his imminent trip to Budapest, there is a state in his head that represents it, just as a file on his hard disk represents this chapter. This is not a return to pictures in the head; there are many kinds of representation besides pictures.
A Language of Thought?
We often think silently in words, and language lets us form thoughts we could hardly have otherwise (try thinking about postmodernism without a language). But that is not the thesis. The Mentalese hypothesis says that when you believe that the price of property is rising again, there is literally written in your head a sentence that means the same as the English sentence. It is not English but a sentence of a postulated mental language, the language of thought (LOT) or Mentalese, which cognitive science should try to discover. Two questions follow: what is it for any symbol to be written in the head, and what is it for a sentence to be?
Media and vehicles
Symbols can be seen, heard, sent by radio, stored on magnetic tape or in circuits. To be precise, distinguish types and tokens: "Est! Est! Est!" contains three tokens of one word-type. The sentence "The man who broke the bank at Monte Carlo died in misery" can be written, spoken or stored on disk and remain the same sentence. Crane builds two layers on this.
| Level | What varies | Example |
|---|---|---|
| Content (message) | What is represented | "Beware! Children crossing!" |
| Vehicle | The kind of representation carrying the message | A road sign picturing two children in a red triangle (picture) versus the written words (sentence); tree rings indicating age (indication) |
| Medium | The physical stuff that realises tokens of one vehicle | Ink, sound waves, magnetic tape, silicon, neural tissue |
The Mentalese hypothesis then says: the vehicle of thought is a sentence, and its medium is the neural structure of the brain. Picture neurons as primitive processors, like the logic gates of Chapter 7, that fire when their inputs are right; combinations of them make up the Mentalese sentence that translates as "Prices are rising".
Syntax, semantics and compositionality
Calling the vehicle a language means its elements have syntactic and semantic structure.
- Syntax. The features that concern form, not meaning: what the basic categories are and which combinations are well formed. "The Pope leads a jolly life" is grammatical; "Life leads a jolly the Pope" is not. For symbols in the head, syntax means simple symbols plus rules that build complex ones; the theorist's task is to find both.
- Semantics. The features that concern meaning: "pusillanimous" is an adjective (syntax) and means spineless, weak-willed (semantics). Symbols are symbols because they stand for things, and standing for is a semantic relation. Semantics also covers how words relate to one another.
- Compositionality. "Cleopatra loves Antony" has constituents that keep the same meaning in "Cleopatra committed suicide", "Desdemona loves Cassio" and "Antony deserted his duty". The meaning of a sentence is determined by the meanings of its parts and their mode of combination. Hence you understand a sentence you have never met: "There are fourteen rooms in the bridge." You can say what is in the bridge, where the rooms are, how many there are.
Contrast ships' signal flags. One flag means "yellow fever on board", another "customs inspectors welcome", but you cannot combine them into "yellow fever inspectors welcome", and a new flag must be learned on its own. A language gives you, from a finite stock of words, the ability to form and understand a potentially infinite number of sentences (its productivity). The vocabulary must be finite if the language is to be learnable; otherwise every new sentence would be like a new flag. Mentalese, on the hypothesis, has semantic features because its words stand for things in the world and its sentences' meanings are compositionally determined. (For exposition, Crane makes the "chauvinistic assumption" that Mentalese is English.)
For and Against the Language of Thought
Precisely stated: when a thinker has a belief or desire with content P, a sentence meaning P (a representation with syntactic and semantic structure) is written in their head; the vehicle is linguistic, the medium neural. Something is missing: believing, desiring and hoping that prices will fall share a content, so what distinguishes them? Crane's short answer: the hypothesis does not explain it, nor is it meant to. It explains having an attitude with this content rather than that, not having this attitude rather than another; Mentalese theorists expect a separate theory of attitude types.
Fodor's first argument: systematicity
- Language is compositional: anyone who understands "Cleopatra loves Antony" can understand "Antony loves Cleopatra".
- Thought shows the same pattern: anyone able to think Cleopatra loves Antony is able to think (at least entertain) Antony loves Cleopatra; nothing more is needed for the second thought.
- The best explanation of this systematicity is that thought itself has a compositional structure.
- To have compositional syntax and semantics is to have a language.
- Therefore (probably) there is a language of thought.
This is inference to the best explanation, not entailment: thought could in principle be systematic without Mentalese, but Fodor holds LOT is the best scientific explanation.
The phrasebook mind. Imagine a creature whose thoughts were stored like a tourist phrasebook or a ship's flags: it can think the waiter brought the soup and the soup is cold, but has no way to think the waiter is cold without learning that thought as a new unit. It would be like meeting a language user who could say "John loves Mary" but not "Mary loves John". Fodor's claim is that no real thinker is like this, and the oddness of the creature is the evidence. The thought experiment also shows where a reply must aim: either find real minds with phrasebook gaps, or show that some non-linguistic architecture can fill in the gaps automatically.
Fodor's second argument: rational trains of thought
Compare two ways of deciding how to get to Ljubljana:
- Reasoning. The bus is cheaper and more direct; the train is more pleasant, leaves at a better time, takes longer, but stops in Vienna, which I want to visit. I weigh the factors and choose the train.
- Association. I wake, see two pigeons on the roof; pigeons remind me of Venice, which I once visited by train; so I choose the train.
Same conclusion, very different processes. For common-sense psychological explanation to work (Chapter 4), much more of our thinking must be like the first than the second; if all thought were free association, we could not see the link between people's thoughts and their behaviour, and in fact we usually can. Reasoning can concern what to do or what to think: Crane believes Bishop Berkeley thought matter a contradictory notion; nothing contradictory can exist; matter does exist; so Berkeley was wrong.
Logic studies the form of inferences, abstracting from content. "If I visit Ljubljana I'll go by train; I'll visit Ljubljana; so I'll go by train" and "If matter exists Berkeley was wrong; matter exists; so Berkeley was wrong" share the form P → Q, P, therefore Q: modus ponens. It holds good because of its form, in the sense of being truth-preserving: a valid argument cannot lead from true premises to a false conclusion. Mentalese theorists hold that many mental processes are truth-preserving in virtue of form, so the items they process must have form: the sentences in the head have syntax, and that is why they can enter systematic processes.
The key link: semantics, syntax, causation
The computer shows how the three connect. A machine's processors are causally sensitive only to formal properties. The and-gate passes current when and only when both inputs are on; that causal profile encodes the formal structure of "and" (P and Q is true when and only when both are true), and that formal structure mirrors the meaning of "and". So the device performs computations by purely causal operations. In the mind, reasoning by modus ponens is a causal process that mirrors the formal relation, so the causal process must have components mirroring the inference's parts: form must have a causal basis. And form keeps meaning in step: valid forms preserve truth, so syntax does not create semantics, but it keeps it in tow. As Haugeland put it, "if you take care of the syntax, the semantics will take care of itself."
The syntax of a symbol might determine [its] causes and effects ... in much the same way that the geometry of a key determines which locks it will open.
So the hypothesis connects the content of thought with its causal nature: a mental syntax realised in the causal structure of the brain, just as a computer's formal properties are realised in its circuits and are semantically interpretable as calculations or inferences. It is computational because it invokes representations processed by formal rules. Which rules is for cognitive science to discover; logic was only an example. Defenders point to computational theories of vision (the approach of David Marr's Vision, 1982), which explain how the visual system builds a representation of the 3D scene from the distribution of light on the retina by computational inferences in stages, using built-in principles that are not accessible to introspection but are "tacitly known".
Objection one: homunculi again?
A computer's states have meaning derived from its designers and users; English sentences arguably get meaning from how speakers use them. Mentalese sentences cannot get meaning from conscious use, or we could introspect whether the hypothesis is true. Appeal to an inner user invites the homunculus fallacy:
- Suppose a homunculus (sub-system) in the brain uses Mentalese sentences, giving them meaning.
- If it uses them by having its own inner language, we must explain that language's meaning, and a smaller homunculus only repeats the problem: a regress.
- If it uses them without an inner language, why not say the same of people, and drop Mentalese?
- So either Mentalese gets meaning as public language does (regress) or some other way (unexplained). Either way, no explanation.
Searle thinks the fallacy "endemic to computational models of cognition". Crane treats it as a challenge: explain the semantics of LOT without appealing to what you are trying to explain. Two responses:
- Dennett's discharge. Accept homunculi but make each stupider than the one it explains, ending in switches (Chapter 6). Searle's rejoinder: at the bottom the stupidest homunculus still manipulates 1s and 0s that must mean something; one that merely moves a Turing machine's tape cannot explain why the symbols mean anything, so the problem of getting from meaningless to meaningful activity recurs at the lowest level.
- No users at all. The more popular reply: Mentalese sentences affect behaviour "without having to be understood" (Fodor). Conscious use of sentences stops in the outside world; no homunculi use Mentalese as we use English. This avoids the fallacy, but leaves the question how Mentalese sentences get their meaning, the topic of Chapter 9.
Objection two: following versus conforming to a rule
Searle also presses an objection descended from W.V. Quine's criticism of Chomsky's claim that we tacitly know grammar. Mentalese says thinking is rule-governed and that we tacitly know the rules. How does that differ from thinking that merely conforms to rules, as planets conform to Kepler's laws without following them? Adding an explicit representation of the rule does not help: a rule-representation is just another representation, and we need a further rule to apply it to the others, and so on (the regress Lewis Carroll dramatised in "What the Tortoise Said to Achilles", 1895). The defender must explain what it is for a rule to be implicitly represented in the causal structure of mental processes: roughly, that behaviour is better explained by supposing the thinker tacitly knows the rule. That pushes the burden onto tacit knowledge, which Crane leaves for further reading.
A related worry: people often reason illogically. Laws of logic seem normative (they prescribe how we ought to think) rather than descriptive: exceptions to a physical law make us revise the law; an illogical person makes us call the person irrational. Mentalese cannot let its rules be merely normative. It has two replies:
- Defensive. The hypothesis is not committed to the normative laws of logic being the rules that operate on Mentalese; which rules govern the mind is empirical.
- Aggressive. Even if such rules govern thought, they are idealisations of messy behaviour, stated with a ceteris paribus ("all other things equal") clause, as in other sciences.
The crux is whether human thinking is rule-governed in the hypothesis's sense: are there laws of thought for cognitive science to find? That reconnects with Dreyfus (Chapter 7).
"Brainy" Computers: Connectionism
Computers excel at fast calculation, storage and retrieval, chess and theorem-proving, where most humans are poor (we use phones to remember numbers). Things that come naturally to us (telling jokes, showing sympathy, practical bodily skills) were exactly what early AI and cognitive science found hardest. Many trace this to the limits of the Mentalese picture, and the existence of neural network ("connectionist") models undercuts Fodor's claim that LOT is "the only game in town": an inference to the best explanation weakens when rival explanations appear.
Classical versus connectionist architecture
A classical ("von Neumann") architecture, after John von Neumann, separates data structures (explicit representations) from rules or programs that operate on them in virtue of their syntactic structure, typically serially, step by step as a flow chart. This is the "rules and representations" picture; applied to AI, Haugeland called it GOFAI, "good old-fashioned AI".
A connectionist network is a mass of simple units (nodes), each excited or inhibited by its inputs and linked to others by connections of varying strengths ("weights"). A unit turns on when its weighted input exceeds its threshold, which in turn affects the units it connects to. Units come in layers: an input layer, an output layer and one or more "hidden" layers between.
Computing means fixing the input units on or off; a pattern of activation spreads through the network, shaped by the weights, until it "settles" into a stable state and the output is read off. What matters most for the contrast is that this happens in parallel, across the network at once. Representations are assigned by builders, in two ways: localist (each unit stands for a feature) or distributed (the state of the whole network represents). Distributed representation is often called connectionism's signature, hence the name parallel distributed processing (PDP).
Networks can be "trained": feed in an input, compare the output with the target, adjust the connection strengths repeatedly until the output is right; a trained network then generalises to new samples with some success. Because they resemble brains (many simple units, variable connections) they are called "neural networks", though nodes have none of the complexity of real neurons. Enthusiasts also claim psychological plausibility: they excel at face and pattern recognition, where classical machines struggle.
Are connectionist networks computers?
Jack Copeland defines computers strictly by symbol manipulation and excludes networks; Robert Cummins counts anything that transforms input representations into output representations in a disciplined way. Partly terminology, Crane says, but he sides with Cummins: networks compute input-output functions systematically, using (localist or distributed) representations, and learn by learning algorithms.
Where the real difference lies
It is not that a network's lowest level lacks psychological interpretation: a classical machine's bits lack it too. The difference is what happens above that level. In a classical machine, basic operations build complex symbols (Mentalese words and sentences) on which processes operate by form, so one and the same thing is the vehicle of computation and the vehicle of content. In a network, as Cummins puts it, "connectionists do not assume that the objects of computation are the objects of semantic interpretation." Computation happens at the level of units changing activation and weights; representation may live only in states of the whole network. The idea that the objects of computation need not be the objects of representation is the most radical and philosophically interesting part of the connectionist doctrine.
Put in terms of syntax: for Mentalese, thinking "some dogs are bigger than others" and "there are too many dogs around here" involves the same syntactic item, DOGS, both times. Connectionists deny the two thoughts need have anything non-semantic in common. As David Rumelhart and James McClelland (1986) put it, "the currency of our systems is not symbols, but excitation and inhibition": thoughts do not have syntax.
Sturgeon's football pitch. Scott Sturgeon's analogy: a rectangular array of lights as big as a football pitch, each glowing more or less brightly, displays from a distance English sentences: "We know your secret!", "Buy your tickets early to avoid disappointment". The word "your" is made by one bank of lights in the first display and a different bank in the second, which need share nothing, not even shape (YOUR versus your). The objects of processing (individual lights) are not the objects of representation (patterns on the whole pitch).
The classicist answers that the analogy puts the lights at the level of medium, not vehicle. Classical theorists happily allow sentences to be realised in a connectionist-like intermediate level; what they insist on is that if cognition is systematic, its vehicle must be systematic.
- Cognition is systematic (the Antony and Cleopatra fact), and any theory must explain this.
- A language-like mechanism explains it, since systematicity is built into compositional syntax and semantics.
- Nothing guarantees that a connectionist network produces systematic representations.
- If a network does behave systematically, it does so by implementing a Mentalese-style mechanism.
- Therefore: either the network is merely the medium for a linguistic vehicle, or it cannot behave systematically. Either way, connectionism is not an alternative theory of cognition.
Connectionists have two routes: argue that cognition is not systematic in Fodor's sense (much work to show how), or argue that networks can be systematic without implementing LOT (hard to escape the "mere implementation" charge).
| Feature | Classical (Mentalese, GOFAI) | Connectionist (PDP) |
|---|---|---|
| Basic elements | Symbols with syntactic structure | Units with activation levels, linked by weights |
| Processing | Rules applied to data structures, mostly serial | Spreading activation, settling, parallel |
| Representation | Explicit, localised symbols | Localist or distributed across the network |
| Vehicle of computation = vehicle of content? | Yes: the Mentalese sentence is both | Not necessarily: units compute, whole states represent |
| Learning | Programmed rules (plus heuristics) | Training by adjusting weights with a learning algorithm |
| Strengths claimed | Systematicity, productivity, rational inference | Pattern recognition, graceful degradation, learning, brain-likeness |
| Main charge against it | Brittle; poor at what humans find easy; rule-following puzzles | Fails systematicity, or merely implements a classical system |
Dated and loose points. Crane describes training as "trial and error" and networks as passing electric currents; the decisive method since Rumelhart, Hinton and Williams (1986) is backpropagation, computing how each weight contributes to the error and adjusting by gradient descent, and networks are almost always simulated in software. He also omits the classic connectionist replies to Fodor and Pylyshyn: Paul Smolensky's tensor-product representations, Tim van Gelder's "functional compositionality" and David Chalmers's 1990 networks that performed structure-sensitive transformations on distributed representations; whether these escape the "mere implementation" horn is still disputed. Two internal slips: the and-gate was introduced in Chapter 7, not Chapter 6, and the computing-versus-thinking distinction Crane recalls as from "Chapter 3" was drawn in Chapters 6-7.
Connectionism won the engineering, and the systematicity debate came back. Neural networks became the dominant form of AI: AlexNet (2012) for image recognition, the transformer (2017) behind large language models, and the 2024 Nobel Prize in Physics for John Hopfield and Geoffrey Hinton. Systematicity turned into an experimental question. Brenden Lake and Marco Baroni's SCAN benchmark (2018) showed standard sequence models failing to generalise compositionally (learning "jump" but not "jump twice"); in 2023 the same authors reported in Nature that a network trained by meta-learning for compositionality matched human-like systematic generalisation on such tasks. Meanwhile Jake Quilty-Dunn, Nicolas Porot and Eric Mandelbaum argued in Behavioral and Brain Sciences (2023) that the language of thought has re-emerged as "the best game in town", citing evidence from perception, infant and animal cognition. Interpretability work finding reusable, compositional features inside transformers sharpens Crane's question: if a network develops structured, recombinable internal representations, is that refuting Fodor or implementing him?
Conclusion: Does Computation Explain Representation?
Decades on, both Mentalese and connectionism remain highly speculative large-scale pictures; detailed theories of human reasoning are far off. Whether the mind has a classical, connectionist or mixed architecture, or any computational structure at all, is an empirical question, not one philosophy can settle; Crane has only clarified what would have to be settled.
For the problem of representation, the result is sobering. Computation is defined in terms of representation, so the computational theory of cognition takes representation for granted. If we want to explain representation, we must look elsewhere: the next chapter.
Classical versus connectionist is a live architecture choice in banking. A rules engine for credit decisions or sanctions screening is a classical system: the objects of computation are the objects of interpretation, so every step is a readable rule, which is why regulators find it auditable. A gradient-boosted or neural model is closer to connectionism: it computes over weights and features whose individual meaning may be nil, and the "reasons" must be reconstructed afterwards (reason codes, SHAP values, probes). Crane's distinction tells you where explainability will cost you, and hybrid designs (a learned scorer feeding an explicit, rule-based decision layer) are in effect the "connectionist medium, classical vehicle" compromise Fodor and Pylyshyn describe.
Explaining Mental Representation
The computational theory shows how representations can cause things, but not what makes something a representation: computation presupposes it. This is the symbol-grounding problem. Naturalists want a reductive definition, "X represents Y if and only if ___", with the blank filled by non-representational, natural facts, and the leading candidate is causation: thoughts represent what reliably causes them. Crane shows that such theories founder on error: reliable indication makes misrepresentation impossible and invites the disjunction problem, and neither ideal conditions, Dretske's teleological functions nor Fodor's asymmetric dependence gives a satisfying account of what error is.
Reduction and Definition
The detour through computation helped in two ways: a computer shows how representations can have causes and effects, and a rule-governed causal process over structured representations shows how a mechanical device can store and process them. But computation takes representation for granted: a computational process is by definition a systematic relation among representations, so calling a state computational presupposes its representational nature. To say there is a language of thought is not to say what makes its words mean anything. This is the "symbol-grounding problem" (the term is Stevan Harnad's, 1990): in what is the meaning of mental symbols grounded?
The mechanical view is naturalistic: it treats the mind as part of nature, the world of causes and effects in space and time. One option is to accept representation as a basic natural feature, as physics accepts energy without explaining it in other terms, perhaps as a theoretical notion whose nature is exhausted by the theory it belongs to (Chapter 5). Many naturalists find this unsatisfying: we learned something "deeper" when we found that light is electromagnetic radiation, and they want the same for representation. Jerry Fodor put the attitude memorably: when physicists finish their catalogue of ultimate properties, spin, charm and charge may appear, "but aboutness surely won't: intentionality simply doesn't go that deep". Crane will reject this demand by the end of the book, but first takes it seriously.
The reductive schema
The obvious way to explain intentionality in other terms is to give necessary and sufficient conditions ("A if and only if B": B is both sufficient and necessary for A) stated without mentioning representation:
(R) X represents Y if and only if __.
The resemblance theory of pictures had this form: "X pictorially represents Y iff X resembles Y". Filling (R) this way is a reductive definition, often thought to give the essence of a concept. Not all definitions are reductive. Take colour:
- Definition 1 (reductive): X is red iff X reflects light of wavelength N.
- Definition 2 (non-reductive): X is red iff X looks red to normal perceivers in normal circumstances.
Definition 2 mentions looking red on its right-hand side, so it does not define redness in other terms. Some philosophers think representation is like this: we should not expect to define it in other terms (Crane returns to this in Chapter 12).
Conceptual and Naturalistic Definitions
Definition 2 has a further virtue: it does not go beyond what we know in having the concept. Anyone who grasps red knows that red things look red to normal perceivers; plenty of people use the concept without having heard of wavelengths. So Definition 2 offers conceptually necessary and sufficient conditions, while Definition 1 offers naturalistic ones, telling us in scientific terms what it is to be red (also called "nomological" conditions, from Greek nomos, law).
| Kind of definition | What it gives | Example |
|---|---|---|
| Conceptual, reductive | Conditions knowable just by having the concept, stated in other terms | "X is a bachelor iff X is an unmarried man" |
| Conceptual, non-reductive | Conditions knowable a priori, but using the concept itself | "X is red iff X looks red to normal perceivers" |
| Naturalistic (nomological), reductive | What science says the property is, in other terms | "X is red iff X reflects light of wavelength N" |
| Naturalistic necessary condition only | A law-based requirement not part of the concept | Oxygen for life: we can still make sense of oxygen-free life on Mars |
Followers of Quine and Wittgenstein doubt there are any interesting reductive conceptual definitions. Quine's bachelor looks easy until odd cases arrive: is a divorced or widowed man a bachelor? A fifteen-year-old? The Pope? Jesus? Do bachelors exist only in certain cultures? We can legislate (men over 25, never married, not in religious orders...), but then we are deciding, not reporting what the concept contains. If bachelor resists conceptual reduction, mental representation surely will. So naturalists who want reduction look for naturalistic necessary and sufficient conditions: the physical, chemical or biological facts that underpin representation.
Fodor says only two candidates have ever seriously been proposed for the blank in (R): resemblance and causation. Resemblance failed even for pictures (Chapter 2) and is hopeless for words or as the basis of all representation. That leaves causal theories.
Causal Theories of Mental Representation
Naturalism includes the causal picture of mind: the mind fits the causal order and its behaviour falls under the same kinds of laws as other things. It is almost inevitable that naturalists should say representation is based on causal relations. Common sense already has a causal notion of meaning, noticed by H.P. Grice (1957):
- (a) "A red light means stop." A matter of convention: nothing about red connects it with stopping; amber would have done.
- (b) "Those spots mean measles." Not convention: the spots are symptoms of measles and can be used to detect it. Likewise "smoke means fire", "those clouds mean thunder". Grice called this natural meaning.
Natural meaning is just causal correlation: spots are effects of measles, smoke of fire, and clouds share a cause with thunder. Causal theories hope such correlations can ground mental representation.
The "crude causal theory" (Fodor's label), "X represents Y iff Y causes X", fails both ways. A child counting sheep to fall asleep has sheep-thoughts not caused by sheep; a startled sheep on a dark night may cause my fear while I represent it as a dog or a ghost. What is missing is a natural, regular link rather than a convention or a one-off accident. Call the regular relation reliable indication: X reliably indicates Y when there is a reliable causal link between X and Y (smoke indicates fire). The improved theory:
X represents Y if and only if X reliably indicates Y.
For the mind: a mental state represents Y iff there is a reliable causal correlation between that type of state and Y.
Obvious gaps: we think about unicorns and Santa Claus (which cause nothing), numbers and sets (which, if they exist, are outside space and time), and future events (which cannot cause present states). Causal theorists treat these as special products of our complex thought-producing machinery and propose to start with simple cases: perceptual thoughts about the environment and basic drives for food, drink, sex, warmth. If those cannot be explained by indication, the complex ones have no chance. The attraction is plain: indication is everywhere, not mysterious and not unique to minds, so explaining representation by it would show how mental representation fits into nature.
The Problem of Error
The ubiquity of indication is also the trouble. Because representations always indicate something, it is hard to see how they could ever misrepresent; and many things reliably correlated with a representation are not what it represents. Both are failures to account for error in thought.
The misrepresentation problem. Treat indication as a necessary condition of representation. Then X cannot represent Y in Y's absence. Spots caused by smallpox indicate smallpox, so they do not misrepresent measles; they represent smallpox. (That we might misread them as measles is irrelevant: the theory may not appeal to our interpretation, which would put things the wrong way round.) As Grice observed, natural meaning is factive: "X means that p" entails p; smoke's meaning fire entails fire. But few mental representations guarantee the existence of what they represent.
The disjunction problem. Treat indication as a sufficient condition. My sheep-perceptions (S-representations) reliably indicate sheep; but at a distance in bad light I systematically take goats for sheep, so S-representations also reliably indicate goats-in-certain-circumstances. The theory must then say their content is "sheep or goat-in-certain-circumstances", a disjunction (logicians' term for an "or" statement), and the apparent error vanishes: the goat was correctly represented all along.
Vervet monkey alarm calls. Ethologists Dorothy Cheney and Robert Seyfarth found that vervets give acoustically distinct calls for different predators, including a "leopard alarm". They suggest its meaning is "only as precise as it needs to be": in Amboseli, where leopards hunt vervets but lions and cheetahs do not, it might mean "big spotted cat that isn't a cheetah"; where cheetahs also hunt vervets, "leopard or cheetah". Scientists here cheerfully attribute disjunctive content. Crane's point: the disjunction problem bites when we ask what it would be for a vervet to misrepresent a cheetah as a leopard, and "only as precise as it needs to be" avoids rather than answers that question.
Both problems make error impossible: one by ruling out representation of absent situations, the other by ruling in too many situations. In both, the indication theory gives the wrong answer to "What does this represent?"
Ideal conditions
The standard repair: misrepresentation occurs when conditions (inside or outside the organism) are not ideal. Robert Cummins: "misrepresentation is malfunctioning". So:
X represents Y iff X is a reliable indicator of Y in ideal conditions.
Error comes from bad light, distance, impaired sense organs. (Ideal conditions are also called "normal" conditions, an idea from Dennis Stampe, 1977.) But what are ideal conditions? We cannot say "the conditions in which X represents Y", or the definition becomes circular and uninformative. We need to specify them without mentioning representation.
Dretske's teleological functions
Fred Dretske, a pioneer of the indication approach, appealed to the teleological function of a representation ("teleological" meaning goal-directed; a different sense of "function" from the mathematical one). Biology attributes such functions routinely: the heart's function is to pump blood, and it keeps that function even when it is not pumping. So: X can represent Y even when Y is absent, provided it is X's function to indicate Y. Ideal conditions become Stampe's conditions of "well-functioning". This handles the misrepresentation problem: even in the dark, my eyes have the function of indicating visible objects.
Dretske's magnetotactic bacteria. Some marine bacteria contain magnetosomes, tiny internal magnets that act like compass needles and align the bacteria with the earth's magnetic field. In the northern hemisphere the field lines incline downward (towards geomagnetic north), so the bacteria swim down; in the southern hemisphere, the reverse. The survival value: these organisms cannot live in oxygen, and moving towards geomagnetic north takes them away from oxygen-rich surface water to the oxygen-poor sediment.
What is the mechanism's function: to propel the bacterium to geomagnetic north, or to oxygen-free water? The device is a magnet, but the point of having it is reaching anoxic water. Perhaps it has both; but since the two can come apart, the function is really "towards magnetic north OR towards the absence of oxygen". Teleological functions are subject to the same disjunctive indeterminacy as indication. If representation is determinate and function is not, function cannot fix representation. Fodor, who once favoured the teleological approach, drew this conclusion.
The southern bacterium. Move a northern-hemisphere bacterium to the southern hemisphere. It dutifully swims towards geomagnetic north, which there means upwards, into the oxygen that kills it. Has its magnetosome misrepresented where safe water is, or has it correctly represented magnetic north and simply been placed in an unlucky world? Your answer reveals your theory. If the content is "anoxic water this way", there is error and a death that error explains. If the content is "magnetic north this way", nothing went wrong representationally. And if, as Crane suggests, nothing in the bacterium's biology decides between them, then perhaps there is no fact about what it represents, which is fatal only if representation must be determinate.
Fodor's asymmetric dependence
For this reason Fodor developed a non-teleological causal theory. Keep SHEEP for the mental representation and → for "causes":
- There is a reliable law sheep → SHEEP, and also a law goat → SHEEP (the error case).
- Dependence: if there had been no sheep → SHEEP connection, there would have been no goat → SHEEP connection. Goats cause SHEEP only because sheep do: a goat looks like a sheep-perception only because sheep-perceptions look that way.
- Asymmetry: if there had been no goat → SHEEP connection, there would still have been a sheep → SHEEP connection.
- So the goat-link is asymmetrically dependent on the sheep-link, and SHEEP represents sheep, not "sheep or goat"; the goat-caused tokens are errors.
(If only goats caused SHEEP, it would be a goat-representation.) Two notes. Asymmetric dependence exists only to solve the disjunction problem; Fodor is otherwise content with indication, so a theory without that problem need not face the question. And Fodor offers it only as a sufficient condition: whatever has indication plus asymmetric dependence represents, but not every possible representation must have that structure.
Crane is unmoved. He grants that the conditions Fodor describes probably are true of mental representations, but they give no deeper understanding. In effect Fodor says error is parasitic on true belief, which "is just what we knew already". The real question is what error is, and saying it is parasitic on truth does not answer it.
A brisk dismissal of two large research programmes. Crane's treatment of teleosemantics and asymmetric dependence is short. Standard objections to Fodor are more specific than "we knew that already": critics such as Lynne Rudder Baker and Fred Adams and Ken Aizawa argued that the counterfactuals are either false or unevaluable without already knowing the content. On the teleosemantic side, Ruth Millikan's consumer-based version (Chapter 10) gives a principled answer to the bacteria case: content is fixed by what the consumers of the state need in order to perform their functions, which is oxygen-free water, while Karen Neander (A Mark of the Mental, 2017) defends the opposite, "magnetic north" reading. The parallel frog-and-fly dispute (does the frog's detector represent flies or small dark moving things?) runs through the whole literature; the indeterminacy Crane cites is a known problem with rival worked solutions, not an unanswered refutation.
Conclusion: The Importance of Error
Causal theories' trouble with error is unsurprising, because representation and error go hand in hand. To represent the world as being a certain way is to allow a gap between how the representation says the world is and how it actually is, which is just to allow the possibility of error. Any reduction that captures the essence of representation must capture whatever makes that possible, which is why error can never be a side issue for a causal theory. It is time to look at alternative approaches.
Naturalising content since 2016. Two book-length programmes have moved the debate. Karen Neander's A Mark of the Mental (2017) combines teleology with causation in an "informational teleosemantics" for sensory-perceptual representation, deliberately starting with simple systems (such as the toad's prey detector) as Crane's causal theorists propose. Nicholas Shea's Representation in Cognitive Science (2018) proposes "varitel semantics": content is fixed by whatever exploitable relations (correlations or structural correspondences) explain how a system stably performs its task functions, so different systems get content in different ways, and error is failure of the relation that explains success. Both are explicitly aimed at the representations posited by cognitive science and neuroscience, and both treat the problem of error as the test of a theory, exactly Crane's moral.
The disjunction problem is the everyday problem of model monitoring. A fraud or sanctions-screening model's alert reliably indicates fraud, but also reliably indicates whatever correlates with fraud in the training data (a postcode, a name pattern). Whether a hit is an error or the model's true content depends on what it was built and validated to detect: a teleological answer. Interpretability practice has rediscovered Fodor's move too: to decide what an internal feature represents, researchers intervene counterfactually (ablate or patch it) and check which causal links depend on which, rather than relying on correlation alone.
The Mechanical Mind and Its Biological Basis
Causal theories forget that thinkers are organisms that act. Crane follows the attempt to ground content in agency (the success theory: a belief's content is the condition under which actions it causes succeed) and then in biological function (Millikan and Papineau's teleosemantics), which founders, in his view, on Swampman. He then assesses evolutionary psychology, rejecting adaptationist just-so stories while allowing that its best targets are dedicated mental "organs", and explains Fodor's modularity of mind: informationally encapsulated, fast, mandatory input systems around a central system of belief, with a sharp critique of massive modularity.
Mental Representation, Organisms and Agency
Causal theories tried to solve the symbol-grounding problem and hit the problem of error: if every belief's content is whatever causes it, every belief correctly represents its cause. They also ignore the organism and its environment. Even if our minds are in some sense computers, we are not computers: we are biological creatures living in communities. This chapter and the next examine how those facts bear on the representational theory: some challenge it, others merely supplement it.
Agency theories say causal theories look at the wrong end: attend not to the causes of beliefs but to their effects in action. What we do flows from what we believe and want, and whether an action satisfies a desire depends on whether the belief behind it is true. If I want wine and believe there is wine in the fridge, going to the fridge succeeds only if there is wine in the fridge. (Other conditions, such as the fridge existing and my limbs working, can be folded into further beliefs.) Turned into a definition, this is the success theory of belief content, named after J.T. Whyte (1990): the content of a belief is identical with the success conditions of the actions it causes.
- First pass. A belief B represents condition C iff actions caused by B are successful when C obtains.
- Objection: idle beliefs. I believe the UK Prime Minister has no moustache, and this has never caused any action. Fix: imagine a quiz show asking for clean-shaven world leaders; go subjunctive: B represents C iff actions that would be caused by B would succeed were C to obtain.
- What is success? Satisfying the desire that partly causes the action: B represents C iff actions caused by B and a desire D would satisfy D were C to obtain. Still reductive.
- What is satisfaction? Not the ceasing of desire (it can cease because I want something else more, the roof falls in, or I die), and not believing it satisfied (if you hypnotise me into thinking I drank wine, I did not get what I wanted). Satisfaction is bringing about the state of affairs the desire represents.
- Final form. B represents C iff actions caused by B and D would bring about the state of affairs represented by D were C to obtain.
- Verdict: circular. Belief content is explained by desire content. The theory needs a non-representational account of desire.
Mental Representation and Biological Function
The most developed way to complete the success theory appeals to biological (teleological) functions: the biological theory of mental representation, pioneered by Ruth Millikan and David Papineau. It assumes desires have evolutionary functions. Organisms that get no water do not survive long, so natural selection favours states that cause us to get water, which is part of what a desire for water is.
The radical step is to explain content by function, not just origin. Natural selection has ensured we are in states whose function is to cause survival-enhancing situations; those states are desires, and those situations are their contents. Getting water enhances survival, so selection has given us states that (other things equal) cause us to get water; their content is roughly "I have water". The theory works "outside in": first find which states of affairs enhance survival, then find the states whose function is to cause them. Plug this into the success theory and belief content is explained too: a fully reductive account. (Crane sets aside desires with little survival value, such as wanting to be famous for jumping off the Golden Gate Bridge; the theory starts with food, sex, warmth, and Chapter 12 returns to the rest.)
If it thinks, it evolved
The theory makes evolutionary history a condition of having desires: a desire just is a state to which selection has awarded a function. But it seems possible for there to be a thinker that never evolved. If thinkers are made of matter, we could in principle take one apart and rebuild it, or build another to the same plan; the coherence of Star Trek-style teletransportation depends on this. The theory has a reply for designed thinkers: functions can be assigned by a creator, as an artificial heart gets its function by being designed and used as one. Such a creature's states would have derived intentionality (Chapter 3), but intentionality nonetheless.
Swampman. Donald Davidson imagined lightning striking a swamp and, by fantastic coincidence, assembling the chemicals into an exact physical replica of a human being: call him Swampman, a replica of Crane. He looks, walks and talks like Crane, but has no evolutionary history; he is a freak accident. Physicalists who think mental states are fixed by local physical states must say that at the moment of creation Swampman has almost all of Crane's thoughts and conscious states (barring those tied to context and location). The biological theory must say he has no representational states at all.
- Swampman is possible: a molecule-for-molecule replica of a thinker, with no evolutionary or design history.
- Swampman thinks (he is physically identical to a thinker, and thought is fixed by physical make-up).
- On the biological theory, a state represents only if it has a function selected by evolution (or assigned by design).
- Swampman's states have no such function.
- So, on the biological theory, Swampman has no thoughts, contradicting 2.
- Therefore the biological theory is false, unless it can deny 1 or 2.
David Papineau replies that the theory is a theoretical reduction, not a conceptual analysis, and reductions can overturn everyday judgements. Science reduces liquid to a state in which molecules cohere but form no long-range order, and thereby corrects the everyday verdict that glass is not a liquid. The biological theory offers a naturalistic, not conceptual, definition (Chapter 9), so our intuitions about Swampman may mislead us.
Crane thinks the cost is far higher than one odd verdict. Swampman lacks thoughts because his mechanisms have no history of actually enhancing survival: this is the aetiological notion of function (from Larry Wright, 1973), no actual evolutionary history, no function. But that notion is not special to minds. A brain is defined by its functions, so Swampman has no brain; equally no heart and, if blood is functionally defined, no blood. He has something that looks like a heart pumping something that looks like blood around something that looks like a body, sustaining something that "looks like" thought. Why even call him "he"? He is not a man. A theory committed to Swampman's not thinking is committed to his not being an organism. One can reject that notion of function or the theory; with the further problems of Chapter 12, Crane rejects the theory. Still, the idea that representation has a basis in biology is very plausible for a believer in the mechanical mind, so the question becomes how biological explanation bears on mental capacities.
Three things the Swampman exchange leaves out. Davidson introduced Swampman (1987) to argue the opposite of Crane's premise 2: since Swampman has no causal history with the things his words seem to be about, he would not mean or think anything at first; Millikan likewise bites the bullet. So premise 2 is a contested intuition, not common ground. Papineau's example is also shaky: physicists classify glass as an amorphous solid, and the popular story that old window panes are thicker at the bottom because glass flows is a myth. Finally, Crane's "no heart, no blood" reductio presumes that heart and brain are defined aetiologically; teleosemanticists can reply that biology classifies organs by structure and homology too, and reserve the historical notion for content.
Evolution and the Mind
The biological theory belongs to a wider project, evolutionary psychology. This is not the claim, as solid as anything in science, that humans evolved from earlier apes over some seven million years. It is the specific, controversial claim that many mental capacities are adaptations: traits whose nature is explained by natural selection. Example: some birds' drab plumage is an adaptation because drab ancestors were better camouflaged, survived predators, bred, and passed it on.
- Units of selection. Some say selection acts on organisms; Richard Dawkins says the basic unit is the gene, with organisms as "vehicles" (the "selfish gene"). Believing in adaptations commits you to neither.
- Adaptationism. Defined variously: that all traits are adaptations (a "crazy view"), or, in Paul Griffiths's formulation, that a model censored of all evolutionary mechanisms except natural selection could predict evolution accurately.
- The logic of adaptive explanation. It is inference to the best explanation: point to an obvious fact, show the hypothesis makes sense of it, and accept it if no better rival exists. It is also reverse engineering: from the trait, infer the environment in which it would aid survival. Evidence must show (i) the adaptive story beats the alternatives and (ii) independent knowledge that the organism lived in such an environment.
The restaurant couple. A rich older man buys an expensive dinner for a younger woman. The Sunday-newspaper explanation: he wants to propagate his genes and youth signals fertility; she wants to propagate hers and his wealth signals he can provide for offspring. Crane calls this absurd. Not just because restaurants were invented in eighteenth-century Paris, not the Pleistocene, but because dining out is a complex social activity entangled with money, class, gastronomy and viticulture; comparing it to a peacock's tail refuses to see real differences. And as an inference to the best explanation it fails: common-sense psychology explains the scene far better, so unless one is an eliminativist the gene story is among the worst explanations available.
Retreating to unconscious adaptive mechanisms does not help without positive evidence, because not all traits are adaptations:
- Exaptations. Traits evolved for one purpose and co-opted for another: feathers, thought to have evolved for insulation, later used for flight.
- Non-adaptations. Some traits may not result from selection at all; Chomsky and others argue this for language. There is no valid inference from "X might have aided survival in circumstances Y" to "X is an adaptation".
- Not genetic determinism. Genetically identical organisms develop differently in different environments; development reflects genes, environment, freak events, floods and ice ages. Evolution does not rely on natural selection alone.
Stephen Jay Gould and Richard Lewontin (1979) compared adaptationist stories to explaining the spandrels of St Mark's Basilica in Venice (the spaces between arches, filled with glorious mosaics) as built for the mosaics. In fact spandrels are by-products of building arches, which artists then exploited. Traits can arise through many historical processes, so adaptive claims need empirical evidence. Paul Bloom summarises the mainstream view: selection is the most important mechanism but not the only one; many traits are by-products or results of random genetic drift; selection is needed to explain Darwin's "organs of extreme perfection and complexity" such as the heart, the hand and the eye.
So we need positive reasons to treat any psychological trait as an adaptation, and the place to look is mental organs: relatively isolated, resilient, probably innate mechanisms dedicated to specific information-processing tasks. The prime example is the visual system; even sceptics might accept it as an adaptation, given a specification of its task and of the environment where performing it aids survival. Two conclusions: evolutionary psychology is entangled with disputes in evolutionary theory itself, and it is strongest when its targets are specific mental "organs", usually called modules.
Spandrels, pendentives and feathers. The structures Gould and Lewontin described in San Marco are, architecturally, pendentives, the curved triangles that carry the central dome, not spandrels between doorway arches; critics (including Daniel Dennett) pointed out that pendentives are arguably designed to carry decoration, which weakens the analogy without killing the point about by-products. The feather example is also more open than Crane implies: early feathers may have served display or brooding as well as insulation.
The Modularity of Mind
Fodor's case for Mentalese rested on rational, systematic thought. But in parts of the mind there are representations without fully rational processing, and the clearest case is vision, which builds representations from retinal light to a 3D scene but is not rational in the way thought is. That motivates Fodor's second big idea.
Mach bands. Look at a row of grey stripes, each uniformly grey, darker to lighter (named after the physicist Ernst Mach). Each stripe looks lighter at the edge nearer its darker neighbour. Mask a single stripe and you can see it is uniform. Now you know it is uniform, and it still looks non-uniform. What matters is not that vision is fooled but that the illusion persists when you know it is an illusion.
- If perceiving were believing, in the Mach band case you would believe the stripes are uniform and believe they are not.
- That is an explicit, conscious contradiction, and a rational person must eliminate such contradictions (you cannot reason or act on "P and not-P").
- But here there is no way to eliminate the conflict: the stripes go on looking shaded however hard you try.
- Yet your rationality is untouched: you know what to conclude and how to act.
- Therefore perception is not belief.
Jerry Fodor (The Modularity of Mind, 1983) took such cases as evidence that vision is a mental module, an information-processing system largely independent of the central system responsible for belief and reasoning, and that other input systems (for example language processing) are modular too. The overall picture, central system plus modules, is the modularity of mind.
The defining feature is informational encapsulation ("functionally defined" here means defined by what it does, function as in functionalism). A mechanism is encapsulated when it systematically lacks access to much of the information in the thinker's mind: it may output P while the thinker knows not-P, and that knowledge cannot change the output. In Zenon Pylyshyn's phrase, its output is not "cognitively penetrable". The central system, home of the propositional attitudes, is the opposite: reasoning can draw on any part of one's stock of beliefs. People are irrational and self-deceived, but those are personal idiosyncrasies, not built into the belief system.
Around encapsulation other properties cluster. "What encapsulation buys is speed; and it buys speed at the price of intelligence" (Fodor). Fodor likens modules to reflexes: the blink reflex does not wait to consider whether that wasp will really hit your eye. Modules are not reflexes, since they contain contentful states, but the comparison shows why the features go together.
| Fodor's feature (1983) | What it means | Mach band / vision illustration |
|---|---|---|
| Informational encapsulation | Cannot use much of what the thinker knows | Knowing the stripes are uniform does not change how they look |
| Domain specificity | Handles only a restricted kind of input | Vision represents only visible properties |
| Mandatory operation | Runs automatically, cannot be switched off | You cannot help seeing the shading, or hearing a sentence as grammatical |
| Speed | Much faster than central reasoning | Seeing is near-instant; deliberation is slow |
| Innateness / characteristic development | Not learned; develops on a fixed schedule | Born with it (Crane: "we are born with them") |
| Fixed neural architecture ("hard-wired") | Realised in dedicated brain areas that others cannot replace | Damage to visual cortex is not compensated elsewhere |
| Characteristic breakdown patterns | Fails in specific, selective ways | Selective visual agnosias |
| Limited central access | Intermediate representations not available to introspection | We cannot report early visual processing |
| Shallow outputs | Delivers simple, basic-level categories, leaving inference to the centre | Sees shapes and objects, not "a cheap reproduction chair" |
The last three rows complete Fodor's original list of nine; Crane's text names the first six.
How much of the mind is modular?
Fodor was cautious: each perceptual system plus a language module. Others went further. Some posit an innate theory-of-mind module, on the ground that it can be damaged while general intelligence survives; this has been offered as the source of autism (Simon Baron-Cohen's "mindblindness"). Beyond that is massive modularity: a distinct, more or less encapsulated mechanism for each kind of cognitive task, perhaps a module for recognising birds, for beliefs about cookery, even for philosophy. Then there is no central system at all, no domain-general, unencapsulated mechanism. Crane's objections:
- The dinner problem. Suppose a food module. Can my reasoning about what to cook for dinner really be confined to it? It must be sensitive to whether I'm going out later, trying to lose weight, wanting to please friends. Fold all that into the module and you lose any way to tell modules apart.
- Fodor's routing problem. If there is no general-purpose mechanism, what decides which module gets a given input? That decision cannot itself be modular, since it must select among inputs bound for many modules, so massive modularity undermines itself.
Modularity and evolutionary psychology are naturally linked: once a module is identified we can ask about its function and evolutionary history as with any organ (though, as Fodor says, "cognition is too soft to leave a paleontological record"). Unsurprisingly, evolutionary psychologists (such as Leda Cosmides and John Tooby) embrace massive modularity, and its critics, Fodor above all, reject evolutionary psychology: there will be no adaptationist explanation of, say, human "mating" behaviour, because that cognition cannot be isolated from all the interlinked activities in which it makes sense.
Dated claims about autism and a slip. The claim that autistic children typically have high general intelligence but lack "theory of mind" reflects 1990s research; autism is now understood as a heterogeneous spectrum with a wide range of intellectual ability, false-belief deficits are not universal, and the "double empathy problem" (Damian Milton, 2012) recasts social difficulty as a two-way mismatch between autistic and non-autistic people rather than a missing module. Separately, Crane says evolutionary psychologists adopt "the massive modularity thesis described in Chapter 4", but the thesis is introduced only here.
Is perception really encapsulated? The modularity debate has continued as the cognitive penetration debate. Predictive processing (introduced in the Chapter 5 update) models perception as top-down prediction meeting bottom-up error, which seems to make every level penetrable by expectation. Against this, Chaz Firestone and Brian Scholl argued in Behavioral and Brain Sciences (2016) that the many reported effects of beliefs, desires and language on what we literally see mostly dissolve under scrutiny into response bias, memory, attention shifts or low-level confounds, so that Fodor's encapsulation survives. Persistent illusions like Mach bands remain the modularist's best evidence, and predictive theorists answer that high-precision sensory signals can outweigh contrary beliefs within a single hierarchy.
Conclusion: The Place of the Organism
Action may help fix belief content, but leaves desire content unexplained. Grounding desire in biological function faces Swampman-style replicas, though Crane allows it is unclear how serious that challenge is. On evolutionary psychology, the mind can be integrated into the causal world even if many or most mental capacities lack an evolutionary explanation, and the best candidates for such explanation are modular capacities. Next: how organisms are embedded in their environments.
Fodor's routing problem is the orchestration problem in agentic AI. A platform of specialised agents or tools (KYC check, credit memo drafter, sanctions screen) is a massively modular design: each component is fast, domain-specific and encapsulated, which is good for testing and control. But something must decide which agent gets which request and integrate their outputs, and that router cannot itself be domain-specific. In practice it is a general-purpose model or planner, a "central system" in Fodor's sense, and it is where relevance failures, cascading errors and audit gaps concentrate. Design the central layer as carefully as the modules.
The Extent of the Mind
Does the mind stop at the skin? Crane separates three ways of pushing it outward. Content externalism (Putnam's Twin Earth) says what we think is partly fixed by the environment; Crane, an internalist, replies with narrow content. Vehicle externalism, the extended mind of Clark and Chalmers, says the mechanisms of thought can lie outside the head (Otto's notebook); Crane argues it is just functionalism drawn out. Enactive and embodied cognition says thinking is bodily activity; Crane finds its insights true but compatible with the representational, mechanical mind, except where it insists on our particular kind of body.
Intentionality as a Relation
The causal and biological theories ground intentionality in real relations to the environment. But we can think about gods, unicorns and Sherlock Holmes, which do not exist. The tension can be put as an inconsistent triad:
- (A) Thoughts are relations between thinkers and the things thought about. Plausible because thought varies along two dimensions, thinker and object, as relations do (X can be longer than many things; many things can be shorter than Y).
- (B) Relations entail the existence of their relata. Plausible quite apart from minds: if X is taller than Y or causes Y, both exist; nothing can weigh 10 grams without existing.
- (C) Thoughts can be about things that do not exist. Among the most obvious facts about thought, and a puzzle since Plato.
Each has been challenged. Against (A): we could have all the same thoughts if the world were radically different, say if a mad scientist stimulated our brains. Against (B): surely "hobbits are smaller than elves" is true though neither exists. Against (C): thinking of a unicorn is really thinking of the properties of being a horse, having a horn, which exist. Crane would reject (A).
The obvious compromise is (A*) some thoughts are relations to things thought about and (C*) some thoughts are about non-existents, both consistent with (B). (A*) has a striking consequence for the brain in a vat (BIV). If my thought about my next holiday in the Bahamas is a relation to the Bahamas, then a BIV in a world without Bahamas cannot have that thought, so a BIV could not have all the thoughts I have. That seems to refute scepticism by mere reflection on thought. The sceptic can reply that it only shows you may not know which thoughts you are having: perhaps it merely seems you are thinking about the Bahamas. That threatens privileged access or first-person authority: not infallibility, but the idea that I know what I think without observing my behaviour, as others must. If my Bahamas-thought is essentially a relation to the Bahamas, and I have no privileged access to whether the Bahamas exist, I have none to whether the thought occurs. Why believe (A*)? Beyond the contested theories of the last two chapters, there are independent arguments for externalism.
Externalism About Content
Externalism says the contents of our thoughts are essentially fixed by things outside us; internalism denies it. Both agree that what we think about is usually outside us; they disagree over whether those things are essential to the thoughts. A test: could you have thought about kangaroos in a world without kangaroos? Yes makes you an internalist about kangaroo-thoughts, no an externalist; one can be externalist about some kinds of thought and not others.
Putnam's Twin Earth
The central argument is Hilary Putnam's (1970s; "The Meaning of 'Meaning'", 1975).
Twin Earth. Somewhere in the universe is a planet exactly like Earth, except that the stuff its inhabitants call "water", which fills their lakes and rains from their skies, is not H2O but a different compound, XYZ, superficially indistinguishable from water. Each of us has a physical replica there. My twin and I both think the thought we would express as "water is wet". Same brain states, same experiences; different stuff in the world. (Ignore the kink that our bodies are mostly water; it does not affect the point.)
- The object of my twin's thoughts is XYZ; the object of my thoughts is H2O.
- XYZ ≠ H2O.
- Content determines object: if two thoughts have the same content, they have the same object.
- So: if two thoughts have different objects, they have different contents (from 3).
- So: my twin and I have thoughts with different contents (1, 2, 4).
- Thoughts are identified by their contents.
- Therefore my twin and I, though physical duplicates, have different thoughts: thought is not fixed by internal constitution alone.
The conclusion does not need a real Twin Earth: whatever internal feature we cite, a Twin Earth case could duplicate it while varying the object, so no internal feature suffices for thinking about this rather than that. Crane assesses the steps:
- Premise 2 is a stipulation, not an "intuition". To deny it you must deny that two substances could look the same but differ fundamentally, a very hard position. It is also independent of disputes over whether XYZ would count as water or whether natural kinds have essences.
- Premise 3 is the weak point. It echoes Gottlob Frege's principle that sense determines reference, and it captures the point of thinking, contact between mind and reality. But we routinely count people as thinking the same thing about different things. Indexicals ("I", "here", "now") are the paradigm: if I say "it's hot here" in London and you say it in New York, there is a clear sense in which we said the same thing, though we speak of different places.
The externalist can distinguish broad content (which fixes the object: London versus New York) from narrow content (shared: something like "it's hot where I am"). But if there is any shared content, premise 3 is false and the Twin Earth argument fails. So externalists must reject narrow content altogether, while internalists say every Twin Earth case has a narrow content, and that it is the psychologically relevant one. Internalists do not say we think only about ideas: we really think about H2O or XYZ, but narrow content plus environment fixes which, just as "it's hot here" plus context fixes London. This is Crane's own view: what a thought is about is settled by narrow content together with context, not by content alone.
Burge's arthritis (added; not in Crane). Tyler Burge (1979) gave externalism a social twist. Bert believes many true things about arthritis and one false one: one morning he tells his doctor "I have arthritis in my thigh". The doctor explains that arthritis is an inflammation of the joints only. Now imagine a counterfactual Bert, physically identical with the same history, in a community where doctors use "arthritis" for rheumatoid ailments of joints and thighs alike. Burge's verdict: actual Bert has a false belief about arthritis; counterfactual Bert has a true belief about a different ailment (call it "tharthritis"). If so, content depends on the linguistic practices of one's community, not just on chemistry, and it reaches any concept a person incompletely understands (contracts, sofas, brisket), not only natural kinds. This is harder for the narrow-content strategy: there is no obvious indexical "where I am" to factor out, and narrow content must be defined without the community's concept.
Externalism About Vehicles: The Extended Mind
Content externalism is consistent with all the vehicles of thought (Chapter 8) being inside the head, the same on Earth and Twin Earth, so it leaves the mechanical mind intact. The extended mind hypothesis of Andy Clark and David Chalmers ("The Extended Mind", Analysis, 1998) goes further: the vehicles of thought themselves can extend beyond brain and body.
Inga and Otto. Inga wants to visit the Museum of Modern Art; she recalls that it is on 53rd Street and sets off. Otto has pathological memory loss and writes important information in a notebook he always carries; he looks up the address and sets off. Clark and Chalmers argue that if Inga believed MoMA was on 53rd Street before she recalled it, so did Otto, before he consulted the notebook. Otto-plus-notebook is functionally equivalent to Inga for all relevant purposes, yet part of the vehicle of Otto's belief lies outside his body. With our smartphones and tablets, we are all like Otto.
The natural scoff, "then my walking stick or motorbike is part of my body", only raises the real question: what makes something part of a mind? Clark and Chalmers answer with what became the parity principle: if a part of the world functions as a process which, were it done in the head, we would unhesitatingly count as cognitive, then that part of the world is part of the cognitive process. Treat inside and outside with parity when they do the same job.
- The detachable visual cortex. Suppose the part of someone's brain dedicated to vision were removed, kept in a small box, and connected by radio. Would being spatially separate stop it being their visual system? Crane agrees with Chalmers that there is nothing "privileged about skin and skull as boundaries for the mind".
- The implanted notebook. Put Otto's notebook contents into an artificial neural structure in his brain and the case is essentially unchanged.
- Sprevak's argument. Mark Sprevak argues the extended mind follows from functionalism: if what makes a state mental is its causal role, and an outside state plays the same role, functionalism must count it as the same mental state.
Clark and Chalmers also gave conditions for when an external resource counts (often called the "trust and glue" conditions, from their paper rather than Crane): it must be reliably available and typically invoked, easily accessible, automatically endorsed when retrieved, and (more tentatively) consciously endorsed at some earlier time.
Objection: Otto's notebook does not do the same thing. Otto must decide to consult it, take it out, look at it; Inga does not "look at" her beliefs, so their causal roles differ. Reply: sameness of mental state cannot require identity down to every causal detail. A forgetful person who activates the belief by a mnemonic still believes MoMA is on 53rd Street; why not Otto with his notebook? What emerges is the problem of individuation for functionalism: what pattern of dispositions and connections makes a state a belief, and the same belief as another? There must be enough similarity in overall profile without identity in every respect (or no two people could share a belief).
The overnight edit. One night someone rewrites an entry in Otto's notebook: MoMA is now listed on 51st Street. Next morning, before opening it, does Otto believe MoMA is on 51st Street? Parity seems to say yes, since a matching tamper with Inga's memory traces would change her dispositional belief. Yet it is strange that a stranger with a pencil can change what Otto believes while he sleeps without any causal contact with Otto's body. The case shows why Clark and Chalmers needed the trust conditions, and it puts Crane's individuation question sharply: is a belief individuated by its role in guiding action (then the edit changes it) or by its integration with the believer's other states and history (then perhaps it does not)?
Because it grows out of functionalism, the extended mind does not threaten the mechanical or representational theory: the mind is still a causal mechanism with effects in behaviour, and Otto's notebook contains representations processed by mechanisms that may well be computational.
Anti-Cartesianism
Externalism of both kinds has been presented as a revolt against the "Cartesian" picture of the mind as inner, private, isolated from the world, detached from the body, parodied in Gilbert Ryle's "ghost in the machine". Anthony Kenny wrote that "the inheritance of Descartes is the single most substantial obstacle to a correct philosophical understanding of the human mind"; for such critics Descartes's real error was not dualism but internalism. Crane is unconvinced: content externalism rests on controversial assumptions, and the extended mind amounts at most to the innocuous claim that the brain/body-world boundary is somewhat arbitrary. A more radical anti-Cartesianism stresses the body and its limits: enactive or embodied cognition, often lumped with the other two but really quite different.
The main objections to the extended mind are missing. Crane presents the debate as largely settled in its favour (as "innocuous"). The best-known critics, Fred Adams and Ken Aizawa (The Bounds of Cognition, 2008), charge a "coupling-constitution fallacy": that a notebook is causally coupled to Otto's cognition does not make it part of cognition, any more than a pencil is part of the act of calculating; they also argue cognition requires a mark of the cognitive (non-derived content) which notebooks lack. Robert Rupert defends a more modest "embedded" view: the mind relies heavily on external scaffolding without extending into it.
Embodiment and Enactive Cognition
Descartes himself stressed the body: nature teaches him that he is "not lodged in my body, like a pilot in his ship" but "so compounded and intermingled" with it as to form "a single whole" (Meditation 6, 1641). Yet he also held that our thoughts could stay the same if an evil demon deceived us, even about having a body; the BIV is its modern form. The internalist grants that being a BIV would be terrible, and that bodily input-output mechanisms and felt embodiment enable thought (Descartes did serious physiology), but holds the body makes no essential difference to what we can think.
The radical, non-Cartesian alternative is that bodily activity does not merely enable thinking: this activity is thinking. Saying thought is "embodied" still distinguishes two things and then connects them; the enactive view says there are not two things but one: thinking is acting, doing.
- The typing test. Close your eyes and say where the "f" key is. Most fluent typists cannot say, yet can type "off" with eyes shut. Knowing where the key is is intelligent, not reflex, but it is an active, bodily matter, not a stored proposition.
- Scrabble tiles. Clark and Chalmers: choosing a word involves rearranging tiles on the tray, which is "not part of action; it is part of thought".
Traditional cognitive science treats cognition as the filling in a "sandwich", between the bread of input (perception) and output (action): the picture of Chapters 8 and 9, which grew from the causal theory of Chapter 4. Enactivists attack the sandwich (the image is Susan Hurley's "classical sandwich").
- Sensorimotor contingencies. Alva Noë argues that perception is constituted by patterns of sensorimotor dependence: we experience something as a cube because we grasp how its appearance varies (or would vary) with movement, its "sensorimotor profile"; there is no intervening "sensation of roundness or distance". Andy Clark calls this the Strong Sensorimotor Model: experience consists in exercising implicit knowledge of the rules linking sensory input to movement.
- Affordances. J.J. Gibson argued that vision delivers not a static picture but affordances: things are seen as movable, twistable, edible, drinkable, directly, without inference from neutral properties. Gibson's influence waned with the computational theory but did not vanish.
- Two visual systems. David Milner and Melvyn Goodale (early 1990s) argued for two visual streams, supported by experiment and neuroscience. Patient DF, with profound visual agnosia, could not describe the size, shape or orientation of a slot, yet posted a card through it as accurately as normal subjects.
The two visual systems are, for Crane, among the most important results of recent cognitive science, but they do not entail enactivism (nor do Gibson's ideas): both streams take input from visual cortex and yield recognition and action. This is not a departure from the sandwich, just a new account of the filling.
Gibson and the two streams, more precisely. Gibson did publish The Perception of the Visual World in 1950, but the term "affordance" dates from 1966 and the theory's full statement from The Ecological Approach to Visual Perception (1979). Crane's description of the two systems ("one for object recognition, the other for location") is the earlier "what versus where" proposal of Leslie Ungerleider and Mortimer Mishkin (1982); Milner and Goodale's distinctive reinterpretation (1992) is "vision for perception" (ventral) versus "vision for action" (dorsal), which is what DF's slot-posting shows. DF's agnosia followed carbon monoxide poisoning.
The case for enactivism, and Crane's verdict
Further support offered for the enactive view:
- Situated cognition. We think "on the hoof", embedded in action under time pressure, constantly revising our grasp of the task; there may be no time to build elaborate models.
- Off-loading. We shift cognitive work onto body and environment: the typist's knowledge is in the body's coupling with the keyboard (compare Otto).
- Robotics. Traditional AI's robots met the frame problem (Chapter 7): how to represent enough of the environment without becoming computationally intractable. Rodney Brooks built computationally simple robots with direct feedback loops from sensing to action instead of a "walking encyclopedia", arguing that "the world is its own best model". Clark and others took inspiration from this, though the old question returns: what does building an artefact teach about how we do it?
- Perception and action are closely linked: experience changes systematically as you move.
- Cognition operates under time pressure, with little time to build detailed world models.
- Thinking often involves practical know-how and off-loading onto external objects.
Together they undermine the idea that cognition happens in an abstract inner space, independent of input, output and environment, with no limits of time or resources. Francisco Varela, Evan Thompson and Eleanor Rosch (The Embodied Mind, 1991) say cognition depends on "having a body with various sensorimotor capacities" embedded in a wider biological, psychological and cultural context. Crane calls this obviously true, and in no conflict with the representational theory.
Why no conflict? Functionalism says a mental state lies at the causal intersection of input, output and other states; it could equally say the state is that intersection, the sensorimotor profile. It can include interactions with the environment and bodily states, sets no limit on feedback loops, and is consistent with fast processing under time pressure. "The sandwich seems secure." The real break would come only if enactivists insist that cognition requires just our kind of body, which contradicts functionalism's multiple realisability; whether bodies like ours are metaphysically necessary Crane leaves open.
| View | What lies outside the head | Relation to functionalism | Threat to the mechanical mind? |
|---|---|---|---|
| Content externalism (Putnam, Burge) | What fixes content: substances, community usage | Neutral; vehicles stay inside | No; at most challenges BIV scepticism and privileged access |
| Extended mind (Clark and Chalmers) | Vehicles of thought: notebooks, devices | Follows from it (Sprevak) | No; external states are still representations in a causal mechanism |
| Enactive / embodied cognition (Noë; Varela, Thompson, Rosch) | Thinking is bodily, sensorimotor activity | Compatible, unless it requires our specific body | Only on the strong version, which denies multiple realisability |
Katalin Farkas shows the extended mind and enactivism, often grouped together, have opposite motivations. The extended mind is a consequence of the functionalist view that only functional role matters, not the physical realiser: Otto's notebook could be a computer or tape recorder. Embodied views stress dependence on the particular shape of our bodies. The two are not incompatible, but limit each other's scope: the closer enactivism moves to the extended mind, the more functionalist it becomes; the more it insists on our actual bodies, the less it harmonises with extension.
Conclusion: The Extent of the Mind
The externalist view of content, if correct, undermines the Cartesian idea that a BIV could have all our thoughts. The extended mind is a development of functionalism. The enactive view pushes further from functionalism and Cartesianism, but most of its supporting reasons are acceptable to a functionalist. The radicalism of these alternatives may have been exaggerated. On content, they do not help: if the extended mind is functionalism, it must explain content too, since saying representations extend beyond the head does not say how they get content; enactivists vary in whether they treat content as a category at all. Since the reductive theories of Chapters 9 and 10 have serious flaws and functionalism is not essentially reductive, it is time to take a non-reductive approach seriously.
Predictive processing and the extended, embodied mind. Andy Clark's Surfing Uncertainty (2016) recast his embodied and extended views in the predictive processing framework (outlined in the Chapter 5 update): the brain as a hierarchy minimising prediction error, with action as a way of making predictions come true. This has split the 4E camp. Clark and Jakob Hohwy disagree over whether prediction-error minimisation draws a sharp "evidentiary boundary" at the senses (Hohwy, an internalist reading) or invites extension into body and world (Clark). Radical enactivists such as Daniel Hutto and Erik Myin reject the representational reading of predictive processing altogether. Crane's verdict, that the sandwich survives with a new filling, describes the internalist reading well. In practice, the extended mind debate is now conducted over smartphones, search engines and AI assistants, which meet Clark and Chalmers's reliability and accessibility conditions far better than a paper notebook.
When does a tool become part of the banker's mind? A relationship manager who relies on a CRM and an AI copilot to "remember" client exposures is Otto with a notebook. Clark and Chalmers's conditions (reliably available, easily accessible, automatically endorsed) describe exactly the point at which staff stop checking the system, which is also the point at which its errors become their errors. The overnight-edit case is the integrity risk: whoever can change the external store can change what the organisation believes. Controls on data lineage and change logging are, in Crane's terms, controls on the vehicles of institutional belief.
A Non-Reductive Conception of Mental Representation
Crane now gives his own answer: mental representation does not need a reductive definition. Reductive theories cannot reach complex concepts like baroque architecture; naturalism does not demand reduction; and the computational theory of cognition can use representation as a theoretical concept, assigned through an interpretation function (Cummins's "Tower Bridge") much as numbers are assigned to weights in measurement. Seen through the philosophy of scientific models, contents are abstract objects used to model brain processes. The price: sub-personal computational states lack unique, essential contents, which opens a gap between how thoughts seem to us and what cognitive science says of them.
Against Reduction and Definition
Even if reductive theories solved the problem of error, a postponed problem remains: how to get from simple concepts (water, food, predator) to complex ones. Their strategy is "get the simple ones right first", but how would a causal or biological story reach baroque architecture? Fodor's asymmetric dependence fits badly: many people meet baroque buildings without ever representing them as baroque, and others acquire the concept from books without causal contact with any baroque building. So a reductive theory of all content must either fill in (R) directly for every concept or give a systematic method of building complex concepts from simple ones; Crane holds neither Fodor's nor the biological theory can take the direct route, and neither supplies the building method. (The success theory can treat every content directly, since P can concern anything, but it is incomplete without a reduction of desire.)
The deeper worry is that definition is not the only form of explanation. To explain baroque you would show buildings, point out broken pediments, cartouches, the extravagant use of line and colour, contrast earlier and later styles, until the concept is grasped. You would not say "a building is baroque iff ___". Grasping the concept is not grasping a definition; in Wittgenstein's phrase, "light dawns gradually over the whole".
For reductionists like Fodor, explanation runs from representation to computation: what separates genuinely computing systems (an adding machine) from systems merely describable as computing (the solar system) is that the former contain and process representations. So a naturalistic vindication of the computational theory needs a reductive theory of representation. Crane rejects this final step for two reasons: (1) naturalism does not require reduction, and (2) the computational theory does not need a reductive account of representation.
What naturalism requires
"Naturalism" sometimes means physicalism (everything is physical, supervenes on or is determined by the physical) and sometimes "nothing supernatural"; but "No spooks!" makes a rallying cry, not a metaphysics. Crane's version: the natural world is the spacetime "causal order", and naturalism says there is nothing outside it. He passes over numbers and abstract objects in silence, and allows properties that make no causal difference (perhaps beauty) so long as they are properties of things that have causes and effects.
"Reduction" also has two senses:
| Ontological reduction | Explanatory reduction | |
|---|---|---|
| Relates | Entities | Theories |
| Claim | One class of entities is a sub-class of another: every mental event is a physical event (token identity) | Theory B explains why theory A is true |
| Example | Mental events identified with physical events | Thermodynamics explained by statistical mechanics (theories stay distinct) |
| Appeal | Neat, elegant; ontological parsimony, which dominated 20th-century metaphysics | Crane's view: the important idea in the history of science |
If reduction is explanation, being anti-reductionist "on principle" is a mistake: a successful explanatory reduction is an advance in knowledge. A principled anti-reductionism would have to show some phenomena resist the kind of theorising that makes reduction possible. But what if a workable theory (such as common-sense psychology) resists explanatory reduction? Eliminativists conclude its entities do not exist. Crane's reply: without an extra physicalist premise such as the causal closure of the physical (Chapter 13), naturalism does not require explanatory reduction, because causation is not confined to the fundamental level; nothing in Chapter 4's account of causation says only basic physical things are causes. And even if reduction were required, it need not take the form of necessary and sufficient conditions. So we need not reject a natural phenomenon just because we cannot reduce it.
A Non-Reductive Conception of Computation and Representation
The second reason: the computational theory of cognition does not need a reductive definition of representation. Start with what the theory explains: the behaviour of organisms in their environments, which is plausibly representational (goal-directed, aimed at satisfying desires, as in searching for food). The theory says this behaviour is best explained as computing a "cognitive function", whose arguments and values are representations with some cognitive relation to one another. Since computation is defined over representations, certain inner states, inputs and outputs must be treated as representations; as stages in a computation they must also have a purely formal ("syntactic") specification. To treat a state as a representation is to give a mapping from the formally described state to an abstract content: an interpretation function. The result is what Robert Cummins calls the "Tower Bridge" picture.
Take an adding machine. The top span is addition, relating numbers; the bottom span is a causal transition among machine states (button presses, displays); the interpretation function I maps each state to a number. A computation is a causal transition among states that mirrors the "transition" among the entities represented. We do not first identify representations on independent grounds and then find computations over them; we treat the system as computing (disciplined transitions among formally specified states) and define an interpretation that maps states to contents. This keeps Fodor's "no computation without representation" but drops the demand for reduction. Representation is just another concept in the theory; it needs no external philosophical defence or reduction.
The measurement analogy
When we weigh things we map weights onto numbers: a standard bag of sugar is 2.2 pounds, and arithmetic then mirrors physical relations: two bags weigh 4.4 pounds. Likewise we "measure" thoughts with sentences: we pick out Vladimir's belief with "The man who broke the bank at Monte Carlo died in misery", and logic mirrors psychological relations, so we infer he believes someone died in misery. (Whether logic really describes thought processes is the Chapter 8 question; the analogy only shows how concrete states map onto abstract entities whose relations mirror relations among the states.) Nobody asks how to "reduce" an object's relation to the number that measures its weight; nor, on this view, should we ask how to reduce a person's relation to the sentences expressing their thoughts. Two features of measurement matter:
- An independent characterisation exists. Weights can be compared without numbers, on old-fashioned kitchen scales. This carries over uncontroversially: computational theorists accept that states in computations have formal descriptions independent of the sentences expressing their contents.
- No unique number. The sugar is 2.2 pounds and also 1 kilogram; there is no limit to the numbers that can measure it, so "the" number is meaningless. This carries over awkwardly, because we are convinced each belief has a unique, essential content: if the belief that snow is white had the content grass is green, it would be a different belief.
For beliefs the fix is easy: content is the meaning shared by intertranslatable sentences, "Snow is white", "La neve è bianca", "Schnee ist weiss", "A hó fehér". But for the states posited by a computational theory (of vision, say), Crane accepts the stronger conclusion: abandon the idea that every such state has a unique, essential content. An interpretation function is just a mapping onto an abstract structure that preserves the structure of the inner states, and many mappings will do that. Which we choose is fixed not by an elusive "unique content" but by which interpretation gives the theory more explanatory power.
One circuit, two functions. Take a simple electronic gate that outputs high voltage only when both inputs are high. Read high voltage as 1 (true) and it computes AND. Now read high voltage as 0 (false) instead: the very same device outputs "false" only when both inputs are "false", which is the truth table of OR. Nothing in the physics favours one reading; two interpretation functions, both structure-preserving, assign different contents. On Crane's view this is not a scandal: which function the gate computes is settled by which interpretation makes the larger system's behaviour intelligible, as the choice between pounds and kilograms is settled by convenience. The case shows exactly what Crane gives up, and why he thinks it is affordable for sub-personal states.
Objection: this makes computation depend on theorists' decisions. Talk of "treating", "specifying" and "assigning" suggests that whether an organism computes is up to us. Reply: applying a theory is a human decision; whether the application correctly characterises the organism is not. Testing a computational hypothesis requires interpreting the process's elements, just as we label the separate forces on a body to compute the net force, without making the physical interaction a matter of choice. This is what happens throughout science: psychological theories give models, and scientific theories are, on a familiar view, "collections of models".
What Crane's position costs, and who shares it. The view is close to Frances Egan's "deflationary" account, on which content is an "intentional gloss" added to a mathematically specified computation for explanatory purposes. Its standard vulnerability is pancomputationalism: if interpretation functions are cheap, Putnam and Searle argued, almost any physical system can be mapped onto almost any computation; Crane's "correct characterisation" reply needs constraints (counterfactual, causal, explanatory) that he only gestures at. There is also an internal tension: in Chapter 9 Crane rejected teleological theories partly because functions are indeterminate while representation is determinate; here he accepts indeterminate content for computational states. (And the reference to Fodor's argument "in §8.4" should be §8.3.)
Scientific Models and Models of the Mind
The "semantic view of theories" traces to Patrick Suppes (1960) and Bas van Fraassen (The Scientific Image, 1980), contrasted with the logical empiricists' "syntactic" view. They used "model" in the model-theoretic sense: a set-theoretic structure of objects and operations that makes a theory's sentences true. Later work by Ronald Giere and others stressed what scientists actually call models: simplified mathematical structures (equations for idealised populations), imagined comparisons (Rutherford's atom as a solar system) and concrete objects (the wire-and-wood double helix of DNA). These are not model-theoretic models: the solar system makes nothing true about atoms; it is used to represent an aspect of them.
In this now-established sense, models help us understand real systems by representing how they are or evolve. They are simpler than their targets and may involve idealisations (frictionless planes), outright falsehoods (rational-actor economics) or unspecificity (a cell model that omits the cell type). Michael Weisberg calls modelling "indirect theoretical investigation of a real-world phenomenon".
Applied to the brain: the brain is the target system; the theorist attributes states to it to predict and explain outputs, and models those states by relating them to abstract objects, their contents, expecting better understanding than from neurochemistry or gross behaviour alone.
The domino illusion. Some circles look convex, some concave; rotate the image 180 degrees and they swap. Chris Frith (Making Up the Mind, 2007) explains it: "the brain has this simple rule built into its wiring". Because light generally comes from above, a concave object is dark at the top and light at the bottom, a convex one light at the top and dark at the bottom. That does not mean those words are written in the brain in English or anything else. Attributing the rule is an abstraction from brain activity that helps predict and explain it: we model the process with a rule relating shading to shape.
Modelling exploits similarity, but not between a brain state and an abstract object (any more than a population resembles its growth equations). The similarity is between the movement between brain states and the stages of an explicit inference from the rule plus the image to the conclusion about shape. Since inferences relate propositions, the brain process resembles a relation among propositions: a computation in the sense of Chapters 7 and 8. This is close to Fodor's argument from trains of thought to Mentalese, but with a decisive difference: we can accept the similarity without concluding that there are symbols in the brain. Understood as modelling, the computational theory does not need the Mentalese hypothesis.
That raises a hard question: what distinguishes a theory that models brain transitions by relating them to contents (treating them as representations) from one that treats them as merely law-governed processes? Viewing the Kanizsa triangle, normal perceivers see a white triangle occluding parts of three black discs. Some scientists say the visual system "completes" the representation by making "assumptions" about how objects normally relate, modelling it as inference. Others say the brain is simply governed by a law about how objects look, with no reason to call it inference, since an inference must be rationally sensitive to further information, and this is not: whatever you know, what you see does not change (as with the Mach bands).
How to decide whether a brain process is correctly modelled as an inference is, for Crane, perhaps the central question in the philosophy of computational psychology. It connects realism about the psychological, the nature of cognitive mechanisms, what makes something a computation, and the causal nature of thought, and none of it needs a reductive definition of representation.
The non-reductive answer to "What is a mental representation?" lists the ways the concept figures in the theory: states of an organism interpretable as instantiating stages in the computation of a cognitive function are representations. That plus the general theory of computation tells us all we need about their nature. The hard work lies ahead: finding which systems to treat as computational, and which computations they perform.
The light-from-above rule turned out to be learnable. Wendy Adams, Erich Graf and Marc Ernst (Nature Neuroscience, 2004) showed that the visual system's "light-from-above" assumption behaves like a Bayesian prior: after training with touch feedback consistent with light from a shifted direction, observers' shape-from-shading judgements shifted accordingly. This supports Crane's modelling stance (the "rule" is a useful abstraction, now formalised as a probability distribution rather than a sentence) while complicating the "law not inference" side of the Kanizsa debate: the process is insensitive to belief but sensitive to experience. Bayesian and predictive-processing models of perception are today's leading example of Crane's question, treating perception as inference over priors without positing Mentalese sentences.
Conclusion: Can Representation Be Reductively Explained?
Philosophers sometimes write as if " iff " is the goal of every theory. It is not obviously so; Crane has offered one alternative and shown how it fits scientific practice with models. The non-reductive theory can say much of what the reductive theory wants about the computational structure of mind without a definitional reduction, and so without the intractable problems of error.
The price is that computational mental states do not have unique, essential contents, while for ordinary representational states unique content seems essential. So even this reading of the computational theory departs from the ordinary notion of thought. Is that a problem? It seems obvious that my belief that it is now raining could not have another content and remain the same belief. But the computational theory concerns unconscious mechanisms and is not answerable to introspection: our thoughts do not seem computational either (except when we work through an explicit algorithm), and no one thinks that refutes the theory. There remains a tension between how our thoughts seem to us and what the science of cognition says about them, the subject of the final chapter.
Interpretability work is Crane's Tower Bridge in practice. When an AI team claims a model "represents" customer risk or "knows" a policy rule, it is proposing an interpretation function from internal states to contents. Crane's standard is the right one for model validation: the interpretation is not a matter of taste but must correctly characterise the system, which in practice means it must support counterfactual predictions under intervention (the "causal abstraction" programme in interpretability formalises exactly this). And his AND/OR lesson applies: several interpretations may fit; choose and document the one with the most explanatory and predictive power, and do not mistake the label for the mechanism.
Consciousness and the Mechanical Mind
The essence of the mechanical mind, that the mind is a causal mechanism with effects in behaviour, survives every detailed dispute in the book. The serious wholesale challenge comes from consciousness: what it is like to have experiences. Crane clarifies phenomenal consciousness and the two senses of qualia, reconstructs the causal argument for physicalism, the zombie argument and the knowledge argument, and states the mind-body problem as a precise dilemma: causation drives us towards physicalism, consciousness drives us away. His closing verdict: no argument defeats the mechanical mind, but there are principled limits to reductive explanation and to what science can tell us.
The Story So Far
Many questions stay open: Theory Theory or simulation, classical or connectionist architecture, reductive or non-reductive content. On some (cognitive architecture, modularity) not enough is known for anything but a cautious open mind; on others (Theory Theory versus simulation, enactivism) the debate is not sharply enough formulated to know what is at stake. None of this is reason to reject the mechanical view, whose essence is overwhelmingly plausible: the mind is a causal mechanism which has its effects in behaviour. "Everything else (computation, Theory Theory, modularity, reductive theories of content) is detail."
Some critics reject the view wholesale: it leaves out the phenomenology of mind (from phenomena, appearances), how our minds seem to us, what it is like to have a point of view. Gregory McCulloch complained that it treats the mind as "a dead phenomenon, a blank agency imprinted with causally efficacious traces"; in Francis Bacon's phrase, it would "buckle and bow the mind unto the nature of things".
The pattern recurs through the book:
- Simulation versus Theory Theory. Interpreting others does not seem like applying a theory; it feels like imaginative identification. The Theory Theorist replies that the question is what makes interpretation succeed, and the theory is tacit, not phenomenologically available.
- Dreyfus versus AI. Part of Dreyfus's case is phenomenological: try to find which rules you follow when going to a restaurant; you cannot say. The computationalist replies that the processes are not supposed to be introspectible.
That reply is unsatisfying. The mechanical view cannot deny that minds seem some way, and many of its own motivations start from how minds seem: the causal picture of thought begins with common-sense explanation of behaviour (Chapters 4 and 5), and Fodor's case for Mentalese starts from ordinary beliefs about Antony and Cleopatra. It would be disingenuous for the mechanical view to claim no interest in appearances. Is the worry merely that no one has yet accounted for phenomenology, or is there a principled objection? Normativity (thought is true or false, reasoning sound or unsound) was one candidate, but if representation is a basic concept (Chapter 12), so is normativity, with no reason to deny the underlying mechanisms are causal. The domain where general arguments claim no causal or mechanical picture can work is consciousness, often called the biggest obstacle to a science of mind.
Consciousness, "What It's Like" and Qualia
Consciousness is at once the most obvious feature of mental life and among the hardest to characterise. Ordinary usage is easy: a doctor asks whether a patient lost consciousness; we wonder whether a lobster thrown alive into boiling water is conscious. We lack infallible tests, but know what is at issue, until we reflect. Saint Augustine said of time that when no one asks him he knows, but asked, he cannot say; consciousness is the same.
We attribute consciousness to creatures (people and animals) and to states of mind (sensations, some thoughts): creature consciousness and state consciousness (terms from David Rosenthal). They are interdependent and neither need be defined by the other, though creatures are an easier start.
Thomas Nagel ("What Is It Like to Be a Bat?", 1974) gave the canonical formula: a creature is conscious when there is something it is like to be that creature. Nothing it is like to be a bacterium or a piece of cheese; something it is like to be a dog, a human, a bat. Transferred to states: there is something it is like to be tasting vanilla ice-cream or smelling burning rubber, nothing it is like to be largely composed of water or to have high blood pressure. The phrase is no definition (no one could acquire the concept from it), but two clarifications help:
- Not comparative. "What is Vegemite like?" "Like Marmite." That sense (what resembles it) is not Nagel's.
- Not just "feels like". There is something it is like to be thinking through a hard problem in your head, with no special feeling or sensation involved.
"What it is like" expresses how things seem to a conscious subject: what it is to be a bat is one thing, what it is like to be a bat another. Call this phenomenal consciousness: a creature or state is phenomenally conscious when there is something it is like to be it or be in it, and the specific way it is, its phenomenal character.
Two senses of "qualia"
Phenomenal consciousness is often described in terms of qualia (singular quale), and here terminology causes havoc.
| Sense | Definition | Who accepts them | Effect on Brentano's thesis |
|---|---|---|---|
| (i) Broad | Whatever properties give an experience its phenomenal character; having qualia just is having phenomenal consciousness | Almost everyone who accepts phenomenal consciousness | Neutral |
| (ii) Narrow | Non-intentional, non-representational properties of experience that make it feel as it does (Block's "mental paint") | Qualia theorists | Denied: some mental properties are not intentional |
| Denial of (ii) | Phenomenal character is exhausted by representational content | Intentionalists / representationalists such as Michael Tye; Daniel Dennett ("Quining Qualia") | Upheld: all mental phenomena are intentional |
For the qualia theorist, when you have described every way the aroma of coffee represents coffee, something is left out: the intrinsic feel of the experience. For the intentionalist, looking at a blue wall, I am aware of the wall and its blueness and nothing more, no extra intrinsic property of my state. Hence the problem of consciousness is often called "the problem of qualia". The key is to separate phenomenal consciousness, the thing to be explained, from qualia in sense (ii), one proposed explanation. Philosophers who deny qualia are not denying that there is something it is like to have experiences; they are rejecting one way of explaining it.
Consciousness and Physicalism
The mind-body problem is often presented as puzzlement at how mere matter like the brain could give rise to consciousness. That alone is not a philosophical problem. Someone who learned all about photosynthesis and cellular growth and still found it incredible that plants grow from sun, water and soil: "Tough." Remarkable phenomena (life, reproduction) can be mysterious and still scientifically explained.
A sharper route starts from the physicalist's own argument.
- Mental states, conscious and unconscious, have effects in the physical world (the causal picture of Chapter 4).
- The causal closure of the physical: every physical happening has purely physical causes, according to physical law. (The empirical support: science succeeds by finding underlying mechanisms, and these turn out to be physical. David Lewis describes a unified, cumulative body of theories explaining every physical phenomenon "down to fundamental particles or fields governed by a few simple laws".)
- If mental causes were not physical, there would be physical effects with non-physical causes, contradicting 2.
- Therefore mental states are physical states (states of the brain).
It does not presuppose that scientists have found the "neural correlate" of any state. To resist it, one must deny the causal efficacy of the mental, deny causal closure, or find a fallacy in the inference.
It is no objection to say "but mental states don't seem like brain states". That is just denying the conclusion. It can seem to you that George Orwell wrote 1984 without its seeming that Eric Blair did, though Blair did ("it seems to me that..." is an intensional context, Chapter 3). Compare "matter can't be energy, it doesn't seem like energy". Wittgenstein asked why people once thought the sun went round the earth; told "because it looks that way", he asked: "And how would it look if the earth went around the sun?" Exactly the same. So with mind and brain: how would mental states seem if they were brain states? Exactly as they do. There is no simple inference from how mental states seem to their nature.
Zombies
But there may be a more complicated inference hidden in the natural objection. Nagel noted that any physical description of conscious states seems "logically compatible with the absence of consciousness". Compare a genuine scientific identification: since water is H2O, it is impossible for H2O to exist without water, as impossible as Orwell without Blair. If a headache were identical with brain state B, B without the headache should be equally impossible. Yet B without the headache seems conceivable: a creature in all my brain states with no headache. Generalise and you get a zombie: a physical replica of a conscious being that is not conscious at all.
- If zombies are possible, physicalism is false. Whether physicalism says conscious states are identical with brain states or wholly constituted by them (constitution, unlike identity, is asymmetric: a parliament is constituted by its members, not vice versa), it says they are nothing over and above brain states: all God needs to do to create my conscious states is create my brain. A possible zombie shows that is not enough.
- Zombies are conceivable. We can coherently imagine a physical replica of me with no sensations, perceptions, thoughts or imagination, perhaps with unconscious states, but no consciousness of any kind, imagined from the outside, since there is no zombie point of view. This claims only bare conceivability: not that zombies exist, that you might be one, or that they are naturally possible.
- Therefore zombies are possible (from 2).
- Therefore physicalism is false (from 1 and 3).
The argument, though controversial, involves no obvious fallacy. The physicalist has two places to push.
- Deny premise 2. Perhaps we only seem to imagine zombies. I seem to imagine being my brother, but no one can literally be someone else; what I really imagine is myself living a life like his. Saul Kripke (Naming and Necessity) says that one who "imagines water that is not H2O" really imagines something that looks like water but isn't. So apparent conceivability can mislead because the thing is impossible.
- Deny the step from 2 to 3. Conceivability is not the same as possibility: some things may be imaginable yet impossible (a counterexample to a law of logic), some possible yet unimaginable (Crane cannot visualise curved spacetime). The argument needs only that imaginability is the best evidence of possibility, as perception is evidence of reality. The physicalist replies that this evidence is fallible (recall the Churchlands' luminous room, Chapter 7) and that the causal argument shows zombies cannot be possible, whatever we imagine.
So for the physicalist zombies are either inconceivable and impossible or conceivable but impossible. Crane judges the second less plausible: if imaginability is usually a good guide to possibility, why not here? Physicalists do better to deny that zombies are genuinely conceivable by finding a hidden incoherence in the story, though "the issues here ... are very complicated".
The explanatory gap and the hard problem. Crane's discussion rests on ideas he does not name. Joseph Levine (1983) called it the explanatory gap: even granting that pain is C-fibre firing, nothing in the physical story makes it intelligible why that state feels painful, whereas "water is H2O" leaves no such residue, since H2O's properties explain water's behaviour. David Chalmers (1995; The Conscious Mind, 1996) distinguished the "easy problems" of explaining functions (discrimination, report, attention) from the "hard problem" of why any of it is accompanied by experience, and built the modern zombie argument on that distinction. Crane's preferred physicalist strategy (deny genuine conceivability) is, in effect, a bet that the gap will close.
The Limits of Scientific Knowledge
Suppose the physicalist shows zombies are not really possible, so the brain-consciousness link is necessary. Physicalism is still not home and dry: related arguments aim to show an epistemological shortfall. Even if physicalism is metaphysically correct, its account of our knowledge would be incomplete. The clearest is the knowledge argument, introduced independently by Frank Jackson ("Epiphenomenal Qualia", 1982) and Howard Robinson (Matter and Sense, 1982).
Louis and the wine (Crane's version). Louis is a brilliant scientist and absolute expert on the physics, physiology and psychology of taste and every scientific fact about winemaking, but has never tasted wine. One day he does: "Amazing! So this is what Chateau Latour tastes like! Now I know." Jackson's original is Mary, a colour scientist who knows every physical fact about colour vision but has lived all her life in a black-and-white room; when she is released and sees a red tomato, she seems to learn what it is like to see red.
- Before tasting wine, Louis knew all the physical, physiological, psychological and oenological facts about wine and tasting it.
- After tasting it, he learned something new: what wine tastes like.
- Therefore not everything there is to know about tasting wine is physical: there are non-physical things to learn, namely what it tastes like.
The conclusion follows readily: you cannot learn without learning something, and since Louis already knew everything physical, what he learned is not physical. The point extends beyond physics to any science one could learn without having the experiences it describes. Jackson and Robinson concluded physicalism is false.
Some distrust thought experiments, but it is hard to deny that a first taste teaches what wine tastes like. Resistance must target premise 1: can anyone know all the physical facts? Dennett objects that you do not imagine someone having all the money in the world by imagining them very rich. Crane replies that you would not go far wrong by imagining them very, very rich and then more so, more of the same kind, never something different; likewise Louis needs only more of the kind of science we already have.
The standard physicalist reply grants new knowledge but denies new things. The objects of Louis's knowledge are all physical (alcohol, acid, sugar; his brain's neurochemistry); what changes is his state of knowledge, which shows only that states of knowledge are more numerous than the entities known, just as one can know Orwell and learn something new on learning he is Blair.
Crane thinks this is "not such a happy resting place". It concedes in-principle limits to what physical science can tell us: science gives the chemical constitution of wine, not what it tastes like. Physicalists must give up the idea that science could state every truth about the world independently of the experiences and perspectives of conscious beings, since truths about what wine tastes like can be learned only by tasting. That is surprising only to those who thought science could tell us everything.
The knowledge argument's afterlife is missing. Crane presents the "new knowledge, old fact" reply and stops. Other replies are prominent: the ability hypothesis (David Lewis, Laurence Nemirow: Mary gains know-how, abilities to recognise, imagine and remember, not new facts), the acquaintance hypothesis (Earl Conee), and the phenomenal concept strategy (Brian Loar and others: Mary acquires a new concept of an old physical property). Most strikingly, Frank Jackson himself abandoned the argument in the late 1990s and became a physicalist, arguing that our intuitions about Mary mislead us. Crane's own conclusion (limits on scientific knowledge without non-physical things) is close to the phenomenal-concept view.
The mind-body problem, precisely
Contrary to first appearances, the mind-body problem can be precisely formulated, as a dilemma:
Horn one: mental causation. If the mind is not physical, how can it act in the physical world? The causal argument, with the closure of the physical, drives us to identify the mind with something physical.
Horn two: consciousness. If the mind is physical, how can we explain consciousness, what it is like to taste something, even if tasting is wholly physical? The zombie and knowledge arguments drive us away from physicalism.
Causation drives us towards physicalism, but consciousness drives us away from it.
- Colin McGinn ("Can We Solve the Mind-Body Problem?", 1989) accepts both horns' force but denies the dilemma is metaphysical. There is a natural property of the brain that explains how it produces consciousness; nothing spooky is involved.
- Cognitive closure. But our concept-forming capacities may be constitutionally unable to grasp it, as a monkey is closed to quantum mechanics. Introspection gives us consciousness without the brain; perception gives us the brain without consciousness; neither faculty can reveal the link.
- Relation to Crane. Crane also ends with limits to scientific knowledge, but his limit is perspectival (truths about what experiences are like are learned only by having them) rather than a limit of human intelligence. Crane finds the limit unsurprising; McGinn finds it a permanent mystery that dissolves the problem only by declaring it unsolvable by us.
Conclusion: What the Problems of Consciousness Tell Us About the Mechanical Mind
The mechanical view is a causal view, not necessarily a physicalist one, so an attack on physicalism is not automatically an attack on the mechanical mind. Its heart is that the mind is a causal mechanism with effects in behaviour; mental representation undoubtedly has causal powers, which ties it directly to the mind-body problem. Nothing in the book undermines representation as causally potent. But the mechanical view must still engage the causal argument for physicalism, and if it accepts physicalism it must answer the arguments from consciousness; given the close links between thought and consciousness, consciousness cannot be ignored. The positive conclusion: no powerful argument has been found against the view that the mind is a causal mechanism which has its effects in behaviour.
The negative conclusion: there seem to be limits to reductive explanation of the distinctive features of the mind.
- Chapter 7. Nothing thinks simply by being a computer: reasoning is not just reckoning (Hobbes).
- Chapter 8. The Mentalese hypothesis describes mechanisms of thought but takes representation for granted.
- Chapters 9 and 10. Causal and biological reductions founder on misrepresentation and on complex concepts.
- Chapter 13. Even if the conceivability arguments fail, their variants show fundamental limits to scientific knowledge.
- Lesson (Chapter 12). Be content to understand representation, intentionality, thought and consciousness in their own terms, without reductive accounts from other sciences.
Perhaps we knew this already. Science, Einstein is supposed to have said, cannot give us the taste of chicken soup. But, Crane ends, wouldn't it be weird if it did?
Fading qualia. A thought experiment of David Chalmers turns this chapter back on Chapter 7. Replace your neurons one at a time with silicon chips that reproduce each neuron's input-output behaviour exactly. Your behaviour never changes: you keep saying "I see vivid red". If silicon cannot support experience, your experience must either vanish suddenly at some single neuron (implausible) or fade gradually while you sincerely report no change, which would make you systematically wrong about your own experience while functioning perfectly. Chalmers concludes that functional organisation fixes experience, so a suitably organised machine would be conscious. Searle would bite the bullet and say the reports become empty. Which way you go decides whether "Can a computer be conscious?" gets the same answer as Crane's "Can a computer think?"
Is AI conscious? The question became a research programme. In 2023 Patrick Butlin, Robert Long and colleagues (19 authors including Yoshua Bengio) published "Consciousness in Artificial Intelligence: Insights from the Science of Consciousness". Adopting computational functionalism as a working hypothesis, they derived "indicator properties" from leading theories (recurrent processing theory, global workspace theory, higher-order theories, predictive processing, attention schema theory, plus agency and embodiment) and checked current systems against them. Their conclusion: no current AI system is a strong candidate for consciousness, but there are no obvious technical barriers to building systems that satisfy many indicators. The episode in 2022 when Google engineer Blake Lemoine claimed the LaMDA chatbot was sentient showed why such checklists are needed: fluent first-person talk is exactly what an LLM is trained to produce, so behaviour is weak evidence, which is Crane's (and Searle's) point about the Turing test carried over to experience.
Consciousness science and animal sentience since 2016. Crane's lobster is no longer only a philosopher's example: after an independent review led by Jonathan Birch at the LSE (2021) found strong evidence of sentience in decapod crustaceans and cephalopod molluscs, the UK's Animal Welfare (Sentience) Act 2022 recognised them as sentient beings, and the New York Declaration on Animal Consciousness (April 2024) stated that there is at least a realistic possibility of conscious experience in all vertebrates and many invertebrates, including insects. In the science itself, a pre-registered adversarial collaboration testing integrated information theory (IIT) against global neuronal workspace theory (GNWT) reported results in 2023 (published in Nature in 2025) that challenged key predictions of both theories; in September 2023 an open letter by 124 researchers called IIT "pseudoscience", provoking a public dispute about standards in the field. None of this closes Levine's gap, but it shows the "neural correlate" programme Crane mentions is now an experimental, theory-testing enterprise.
Two quotations with shaky pedigrees. The Wittgenstein sun anecdote comes to us through Elizabeth Anscombe's An Introduction to Wittgenstein's Tractatus (1959), not from Wittgenstein's writings. The Einstein line about chicken soup has no reliable source; Crane's "is supposed to have remarked" is the right hedge.
Why this matters for AI governance in a bank. The zombie and knowledge arguments are a reminder that behavioural evidence underdetermines inner facts, whether the inner fact is consciousness or understanding. Policies on "AI welfare" or claims that an assistant "cares" about a customer should be treated with Crane's discipline: separate what the system does (auditable) from what it is like to be it (not established by any current test). For customer-facing copilots, the practical rule is to design out anthropomorphic claims, because users, like Lemoine, readily read fluent first-person language as evidence of a mind.
Master Takeaways
The Mechanical Mind adds up to a defence of the mind as a natural causal mechanism combined with a refusal to reduce its most distinctive features. Representation is real, causally potent and central to cognitive science, but every attempt to define it in other terms (resemblance, convention, computation, causation, biological function) runs into interpretation, error or consciousness. Crane's settled position is that we should explain thought and experience in their own terms, using representation as a theoretical concept, and accept that science has limits without concluding that the mind lies outside nature.
The takeaways
- Symbols never interpret themselves. Pictures fail by resemblance, words by convention, ideas by presupposing what they explain; syntax is not sufficient for semantics. Every route ends at mental representation, which is why the problem of intentionality is the problem of mind.
- Nothing thinks simply by being a computer. Searle's Chinese room is right that symbol manipulation adds no meaning; what gives symbols meaning is interaction with the world, which concedes that thinking is not mere computation. Yet we may still think partly by computing: the two questions are independent.
- Computation presupposes representation. A computer processes representations by rules; a planet merely instantiates a function. So the computational theory of mind, Mentalese included, takes content for granted and cannot be the theory of what content is.
- Error is the test of any theory of representation. To represent is to allow a gap between how things are said to be and how they are. Indication theories rule error out (misrepresentation problem) or dissolve it into disjunctions (disjunction problem), and neither ideal conditions, functions nor asymmetric dependence explains what error is.
- The mechanical mind is a causal thesis, not a reductive one. Its core, that the mind is a causal mechanism with effects in behaviour, survives every detailed dispute; it entails neither physicalism nor reduction to physics, and an attack on physicalism is not an attack on it.
- Representation can be explained without being defined. Like numbers measuring weight, contents are assigned by an interpretation function and used to model brain processes; the price is that sub-personal states lack unique, essential contents.
- Thoughts are attitudes to contents. Every propositional attitude is "A ψs that S": belief, hope and desire can share a content, and content individuates thoughts. Thought can be studied apart from consciousness, because some thoughts are unconscious.
- Intentionality is not intensionality. Failures of substitution and existential generalisation mark belief reports but also modal sentences about no mind, so intensionality cannot be the criterion of the mental.
- Original versus derived intentionality frames the machine question. Books and computer outputs mean things because minds interpret them; whether anything but a mind could have original intentionality may be partly terminological.
- Behaviourism fails from the inside. A thought fixes behaviour only in combination with other thoughts (the umbrella man), so mental states cannot be defined by behaviour alone: the holism of the mental.
- Acting for a reason is having the reason as a cause. Davidson's Boleslav cases show that only causation separates having a reason from acting on it, an inference to the best explanation rather than a proof (and vulnerable to deviant causal chains).
- Common-sense psychology is a theory worth keeping. Churchland's eliminativist charges (explanatory failure, stagnation, irreducibility) are weak; theory and simulation accounts of mindreading may differ mainly in emphasis.
- Turing made effective procedures unmysterious. Any algorithm reduces to read, write, move, change state; the universal machine mimics any other; hence variable realisation: what matters is the program, not the material.
- The Turing test is behaviouristic. Treating indistinguishable performance as thought let AI blur simulation and the real thing, and fluent behaviour remains weak evidence of understanding or experience.
- Dreyfus's problem is relevance, not storage. Even a complete store of common sense needs the know-how to tell which facts bear on which; rules specific enough to act on are never general enough (the bus driver).
- Systematicity is the best argument for a language of thought. Anyone who can think Antony loves Cleopatra can think Cleopatra loves Antony; compositional syntax that mirrors semantics and drives causal role explains it, and Fodor and Pylyshyn say connectionism either lacks it or implements it.
- Connectionism's radical idea. In a network the objects of computation need not be the objects of representation: units compute, whole patterns represent, and thoughts may have no syntax.
- Functions are as indeterminate as indication. Dretske's bacteria may be steering towards magnetic north or towards oxygen-free water; if representation is determinate, function cannot fix it.
- Teleosemantics pays for its history with Swampman. If content requires selected function, a lightning-made replica has no thoughts, and on the aetiological notion of function, no heart, brain or blood either.
- Adaptationist stories need evidence, not imagination. "Might have aided survival" does not imply "is an adaptation" (spandrels, exaptations); evolutionary psychology is strongest for dedicated mental organs.
- Perception is not belief, and modules are encapsulated. The Mach bands look shaded however much you know; Fodor's input modules are fast, mandatory and cognitively impenetrable, and massive modularity cannot say who routes inputs to modules.
- Twin Earth works only if content determines object. Indexicals show we often think the same thing about different things; with narrow content plus context, the internalist keeps the twins' thoughts the same.
- The extended mind is functionalism drawn out. Otto's notebook counts by the parity principle; enactivism's insights fit a functionalist "sandwich" unless it insists cognition needs our particular body.
- Phenomenal consciousness is not the same as qualia. Denying non-representational qualia is not denying that there is something it is like; much of the debate is terminological confusion between explanandum and explanans.
- The best physicalist reply to zombies is to deny their conceivability. Accepting conceivability but denying possibility sits badly with treating imagination as evidence of possibility.
- Even true physicalism leaves limits on scientific knowledge. Louis learns what wine tastes like; facts about experience are learnable only from the inside. The mind-body problem is a precise dilemma: causation drives us towards physicalism, consciousness drives us away.
Frameworks & mental models
| Name | What it says | Use it when |
|---|---|---|
| The mechanical mind | The mind is a causal mechanism, part of nature, with effects in behaviour | Separating causal explanation from reductionism or physicalism |
| Necessary and sufficient conditions / schema (R) | "X represents Y iff ___", filled without mentioning representation | Testing any proposed definition of a concept |
| Resemblance theory | Pictures represent by resembling | Seeing why similarity is symmetric, reflexive and too cheap to fix meaning |
| Attitude + content ("A ψs that S") | A thought is an attitude to a represented situation | Analysing beliefs, desires and their individuation |
| Original vs derived intentionality | Minds mean things intrinsically; books and outputs borrow meaning | Asking whether an AI system's states mean anything |
| Intentionality vs intensionality | Directedness of minds vs substitution failure in sentences | Avoiding conflating a mental and a logical property |
| Holism of the mental | Thoughts link to behaviour only via other thoughts | Rejecting behaviourist or single-signal inferences about intent |
| Causal picture of thought | Reasons are causes of action (Davidson) | Explaining the difference between having and acting on a reason |
| Theory Theory vs simulation | Mindreading by tacit theory vs imaginative replication | Modelling how people (or machines) attribute mental states |
| Eliminative materialism | Folk psychology is a false theory; beliefs do not exist | Stress-testing whether mental categories earn their keep |
| Turing machine / Church-Turing thesis | Any algorithm runs on a tape, reader and machine table; computable = Turing-computable | Reasoning about what is computable in principle |
| Computing vs instantiating a function | Computation needs representations in and out; physical processes merely obey functions | Deciding whether a system computes or is just modelled |
| Black boxes and homunculi | Explain a capacity by ever-stupider components down to switches | Functional decomposition of any system |
| Turing test | Indistinguishable conversation counts as thinking | Evaluating behavioural claims about machine minds |
| Dreyfus's know-how argument | General intelligence needs common sense, which is skill, not rules | Judging rule-based automation of judgement-heavy work |
| Chinese room | Running a program is never sufficient for understanding | Separating performance from understanding |
| Language of thought (Mentalese) | Thoughts are sentences in the head; syntax mirrors semantics and drives causation | Explaining systematic, productive, rational thought |
| Medium vs vehicle vs content | Physical stuff vs kind of representation vs message | Locating where a representational claim operates |
| Classical vs connectionist architecture | Rules over symbols vs weights over units; same or different vehicles of computation and content | Choosing or explaining AI architectures and their auditability |
| Fodor-Pylyshyn dilemma | Networks either lack systematicity or implement a classical system | Assessing claims that neural nets "reason" |
| Indication / causal theory | X represents what reliably causes it | Grounding meaning in information |
| Misrepresentation and disjunction problems | Indication rules out error or makes contents disjunctive | Testing any information-based theory of content |
| Asymmetric dependence | Error links depend on true links, not vice versa | Distinguishing content from noise counterfactually |
| Success theory | Belief content = conditions under which actions succeed | Linking representation to agency |
| Teleosemantics | Content fixed by selected biological function (Millikan, Papineau, Dretske) | Grounding content in biology or design purpose |
| Modularity of mind | Encapsulated, fast, mandatory, domain-specific input systems plus a central system | Explaining illusions and designing specialised components |
| Massive modularity | The mind is modules all the way up | Spotting the missing router or integrator |
| Inconsistent triad (A, B, C) | Thought is relational; relations need relata; we think of non-existents | Diagnosing puzzles about thinking of what does not exist |
| Twin Earth / broad vs narrow content | Physical twins differ in broad content; may share narrow content | Deciding whether meaning is "in the head" |
| Extended mind / parity principle | External processes count as cognitive if they would count inside the head | Deciding when tools become part of cognition |
| Enactive cognition / sandwich model | Thinking is sensorimotor activity vs cognition between input and output | Evaluating embodied AI and robotics claims |
| Ontological vs explanatory reduction | Identifying entities vs explaining one theory by another | Clarifying what "reducing" anything would mean |
| Tower Bridge / interpretation function | Computation is a causal transition mirroring a function over contents | Making representational claims about models testable |
| Scientific models | Contents are abstract objects that model brain processes | Using cognitive or AI models without reifying them |
| Phenomenal consciousness and two senses of qualia | What it is like vs non-representational "mental paint" | Untangling claims about machine or animal experience |
| Causal argument for physicalism | Mental causes + causal closure imply mental states are physical | Understanding the case for physicalism |
| Zombie argument | Conceivable zombies imply physicalism is false | Probing the link between conceivability and possibility |
| Knowledge argument | Learning what wine tastes like shows non-physical knowledge | Thinking about the limits of third-person science |
| Mind-body dilemma | Causation pulls towards physicalism, consciousness away | Framing the whole problem precisely |
Strengths. Crane is unusually clear and fair, lays out arguments premise by premise, separates questions that are routinely conflated (thinking vs computing, functionalism vs computationalism, phenomenal consciousness vs qualia, content vs vehicle externalism), and his central thread (representation as the key to mind) gives the book a unity most introductions lack.
Blind spots. Several research programmes get brisk dismissals: teleosemantics and asymmetric dependence (whose proponents have specific answers to his objections), the extended mind's critics (Adams and Aizawa, Rupert), the post-Jackson responses to the knowledge argument (ability hypothesis, phenomenal concepts, Jackson's own recantation), and the connectionist replies to Fodor and Pylyshyn. He omits Burge's social externalism, Levine's explanatory gap, Chalmers's hard problem and McGinn's mysterianism by name, though his arguments lean on them. His own non-reductive view inherits the pancomputationalism worry and sits in tension with his complaint that biological functions are indeterminate.
Dated or wrong. The chronology and history contain slips: Copernicus's birth year (1473) for his work (1543), Harvey's lectures (1616) rather than his publication (1628), Harvard's Mark I called "fully electronic" (it was electromechanical), ENIAC named a "calculator", Colossus linked to Enigma (it attacked Lorenz), a garbled chess number (10^120), and several wrong internal chapter cross-references. The neural-network material describes training as "trial and error" and predates the deep-learning revolution; the remark that AI had become "just engineering" with no aspiration to thinking machines was overtaken within a few years.
How the field has moved. Large language models have made the Chinese room, symbol grounding, original intentionality and the Turing test empirical arguments; mechanistic interpretability finds internal representations that can be read and causally edited; systematic generalisation in networks became an experimental question (SCAN, meta-learning for compositionality); the language of thought has had a revival in cognitive science; predictive processing reshaped debates on perception, modularity and the extended mind; and consciousness science gained adversarial collaborations, formal AI-consciousness indicator frameworks (Butlin, Long and colleagues, 2023) and legal recognition of invertebrate sentience. None of this settles Crane's questions, but it moves several from the armchair to the lab, which is roughly what his non-reductive, model-based view would predict.
Who's who
| Name | Who | Why they matter here |
|---|---|---|
| René Descartes | 17th-century philosopher | Animals as machines, the mind outside mechanism; the "Cartesian" internalism critics attack |
| Thomas Hobbes | Author of Leviathan (1651) | "Reason is nothing but reckoning": the computational view prefigured |
| G.W. Leibniz | Philosopher-mathematician | The characteristica universalis; binary notation; thought as calculation |
| George Boole | Logician, The Laws of Thought (1854) | Algebra of truth-functions that became logic gates |
| Franz Brentano | Philosopher-psychologist (1874) | Intentionality as the mark of the mental |
| Alan Turing | Mathematician | Turing machine, universal machine, the Turing test |
| Alonzo Church | Logician | Church's (Church-Turing) thesis |
| Ludwig Wittgenstein | Philosopher | Signs need life from use; "light dawns gradually"; the sun anecdote |
| Donald Davidson | Philosopher | Reasons as causes; Swampman |
| Paul and Patricia Churchland | Philosophers | Eliminative materialism; the luminous room |
| Adam Morton, Jane Heal | Philosophers | The Theory Theory label; simulation theory |
| Hubert Dreyfus | Philosopher, What Computers Can't Do | Common-sense know-how cannot be formalised |
| John Searle | Philosopher | The Chinese room; syntax is not semantics |
| Herbert Simon | AI pioneer | The 1965 prediction that machines would do any human work within 20 years |
| Jerry Fodor | Philosopher and cognitive scientist | Language of thought, systematicity, asymmetric dependence, modularity |
| Zenon Pylyshyn | Cognitive scientist | Fodor-Pylyshyn critique of connectionism; cognitive impenetrability |
| David Rumelhart and James McClelland | PDP pioneers | Connectionism: "excitation and inhibition", not symbols |
| Robert Cummins | Philosopher | Objects of computation vs objects of interpretation; Tower Bridge |
| John Haugeland | Philosopher | GOFAI; "take care of the syntax and the semantics will take care of itself" |
| Daniel Dennett | Philosopher | Homunculi discharged by stupidity; "Quining qualia"; objections to Mary/Louis |
| H.P. Grice | Philosopher of language | Natural vs non-natural meaning |
| Fred Dretske | Philosopher | Indication and teleological function; magnetotactic bacteria |
| Ruth Millikan, David Papineau | Philosophers | The biological (teleosemantic) theory of content |
| Dorothy Cheney and Robert Seyfarth | Primatologists | Vervet alarm calls and disjunctive content |
| Stephen Jay Gould and Richard Lewontin | Biologists | Spandrels and the critique of adaptationism |
| Hilary Putnam | Philosopher | Functionalism; Twin Earth externalism |
| Andy Clark and David Chalmers | Philosophers | The extended mind and the parity principle |
| Alva Noë; Varela, Thompson and Rosch | Enactivists | Sensorimotor theory; The Embodied Mind |
| David Milner and Melvyn Goodale | Neuroscientists | Two visual streams; patient DF |
| Rodney Brooks | Roboticist | "The world is its own best model" |
| Katalin Farkas | Philosopher | Extended mind and enactivism limit each other's scope |
| Thomas Nagel | Philosopher | "What is it like to be a bat?" |
| Saul Kripke | Philosopher | Apparent conceivability of water without H2O is illusory |
| Frank Jackson, Howard Robinson | Philosophers | The knowledge argument |
| David Lewis | Philosopher | Unified physical science; theoretical terms; ability hypothesis |
Connect it
- Original versus derived intentionality is the right frame for LLM governance. A copilot's outputs mean what they mean because of the documents, prompts and people around them; treating them as the system's own understanding is Crane's (and Searle's) category mistake, and controls belong where meaning is supplied: grounding sources and human review.
- The disjunction problem is model-risk monitoring. A screening or fraud model reliably indicates whatever correlates with its target; deciding whether a hit is error or true content requires a teleological answer (what it was built and validated to detect) and counterfactual testing, which is asymmetric dependence in practice.
- Classical versus connectionist maps onto auditability. Rules engines make the objects of computation the objects of interpretation; learned models do not, so explanation must be reconstructed by interpretation functions (probes, attributions), which Crane's Tower Bridge says must correctly characterise the system, not merely label it.
- Fodor's routing problem is the orchestrator problem in agentic platforms. Specialised, encapsulated agents are easy to test; the general-purpose router that integrates them is where relevance failures and audit gaps concentrate, and it deserves the most design attention.
- The extended mind defines when tools become part of judgement. Once a CRM or AI assistant is reliably available and automatically endorsed, its errors become the banker's errors; data integrity and change control are controls on the vehicles of institutional belief.