Philosophy of Mind

In the winter of 1619, quartered with the Bavarian army near Neuburg an der Donau, a twenty-three-year-old French officer spent a day shut in a room heated by a stove. He later wrote that he had nothing to do but talk to himself about his own thoughts. Twenty-two years afterwards, in 1641, the same man published in Paris a short book of six chapters that set the terms of every argument in this course, including the arguments designed to destroy his position.

The man was Rene Descartes, and the book was the Meditations on First Philosophy. It is worth being precise about what it argues, because the position it defends is usually attacked in a version Descartes did not hold, and the argument that actually does the work is not the one most people can recite.

Doubt as a tool, not a mood

The First Meditation is not scepticism. It is a demolition survey. Descartes wants to find out which of his beliefs could survive if he deliberately withheld assent from anything that admits of the slightest doubt, and he applies the test in three widening stages.

First the senses: they have deceived him before about small and distant things, and a source that has deceived you once is not one you should trust absolutely. That alone does not touch beliefs about the paper in his hands. So, second, the dream argument: there are no marks by which waking can be reliably distinguished from a vivid dream, so any belief grounded in present sensation could be false. That still leaves arithmetic and geometry, which seem true whether he is asleep or awake. So, third, the evil demon: suppose a being of great power devotes itself to deceiving him, so that even his conviction that two and three make five is planted. Each stage is calibrated to reach a class of belief the previous stage could not.

Key idea: The doubt is instrumental. Descartes is not claiming that he is dreaming or deceived; he is asking what would still stand if he were, and the answer is what the rest of the book is built on.

What survives, and what it is

What survives is the meditator's own existence as a thinker. If the demon deceives him, there is something being deceived. In Cottingham's translation of the Second Meditation, the conclusion is stated with a careful restriction: "I am, I exist, is necessarily true whenever it is put forward by me or conceived in my mind." Not always true. True whenever thought. The cogito secures existence only for the duration of the thinking.

Descartes then asks what this surviving thing is, and answers: a thing that thinks, where thinking is stretched to cover doubting, understanding, affirming, denying, willing, refusing, imagining and sensing. That last item is easy to skim past and matters enormously. Sensing, considered strictly as seeming to see, belongs to the thinking thing, and this is what will later be called consciousness.

The Second Meditation closes with the piece of wax. Take a fresh honeycomb: it smells of flowers, it is hard, it has a shape and a colour. Bring it to a fire and every one of those properties changes, yet you judge that the same wax remains. Nothing that the senses supplied is preserved, so what you grasp when you grasp the wax is not delivered by sight or touch but by the understanding. Descartes draws the conclusion he wants: even in a case that seems to be pure perception of a body, the mind is what knows, and it knows itself more evidently than it knows any body.

The real distinction, premise by premise

The argument for dualism proper waits until the Sixth Meditation, and it is a short piece of reasoning that leans on machinery built in the Third and Fifth. Set it out as its author would:

1. Whatever I clearly and distinctly understand can be brought about by God exactly as I understand it.
2. So if I can clearly and distinctly understand A apart from B, then A and B are capable of existing apart, and are therefore distinct things.
3. I have a clear and distinct idea of myself as a thinking, non-extended thing.
4. I have a clear and distinct idea of body as an extended, non-thinking thing.
5. Therefore my mind is really distinct from my body and can exist without it.

Look at where the weight sits. Premise 1 is not a throwaway. It is the reason Descartes spends the Third and Fifth Meditations arguing for a God who is not a deceiver: without a guarantor, clear and distinct perception is just a strong feeling of certainty, and the demon could be producing it. Strip out the theology and the argument loses its bridge from what I can conceive to what can exist. Every later conceivability argument in this course, including the zombie argument of Lesson 8, has to find some replacement for that bridge.

The point: Descartes does not argue from doubt to distinctness. He argues from a positive, complete conception of each item apart from the other, backed by a guarantee that such conceptions track possibility.

Arnauld's triangle, and the reply

The best objection came before publication. Descartes circulated the manuscript, and the Fourth Set of Objections came from Antoine Arnauld, then a young theologian. His counterexample is exact. Someone may clearly and distinctly perceive that a triangle is right-angled while doubting, or even denying, that the square on the hypotenuse equals the sum of the squares on the other two sides. By Descartes's rule, it should follow that a right triangle could exist without that property. It could not.

Descartes's reply introduces a distinction that is doing real work: between a conception that is merely adequate as far as it goes and a conception of a complete thing. The geometer's idea of the triangle is incomplete, an abstraction from a figure; the meditator's idea of himself as a thinking thing is meant to be an idea of a complete substance, one that needs nothing else in order to exist. Whether that reply succeeds is still argued about, and the honest summary is that it moves the burden onto a notion of completeness that Descartes never made precise enough to check. But notice what it is not: it is not the crude fallacy of inferring a difference in things from a difference in what I happen to know about them.

Two further arguments, which are not the same argument

The Sixth Meditation offers a second, independent line. Body is by nature divisible; take any extended thing and you can conceive its halves. Mind is not. When Descartes considers his own mind he finds nothing in it that could be halved: it is the same whole self that wills, senses and understands. In his words, "the body is by its very nature always divisible, while the mind is utterly indivisible." Two things with contrary properties cannot be one thing, so mind is not body. This argument needs no theology at all, which is why it survived longer in the textbooks. It is also the one that the split-brain findings of Lesson 14 will press on hardest.

The third argument is not in the Meditations but in Part V of the 1637 Discourse on the Method, and it is startlingly modern. Suppose a machine built to resemble a human being as closely as you like. Descartes proposes two tests that would reveal it. It could never use words to reply appropriately to whatever is said in its presence, as even the dullest person can; and even if it performed some tasks better than we do, it would fail at others, because it acts from the arrangement of its parts rather than from understanding. Three centuries later, one of these tests was rewritten as a proposal for deciding whether machines can think, and you will meet it in Lesson 3.

Where the mind meets the body, on Descartes's own account

Descartes is routinely accused of treating the person as a spirit piloting a machine. He explicitly denied that. In the Sixth Meditation he writes that he is "not merely present in my body as a sailor is present in a ship" but is "very closely joined and, as it were, intermingled with it, so that I and the body form a unit." His evidence is the character of bodily sensation: a pilot notices damage to the vessel by inspection, whereas you feel the thirst and the pain, confusedly and from the inside. The union is offered as a third primitive notion alongside thought and extension, not as a puzzle to be solved.

The pineal gland enters much later, in the 1649 Passions of the Soul, and the reasoning is anatomical rather than mystical. Most brain structures come in pairs, one on each side; we nevertheless have a single visual experience of a single object seen with two eyes. Descartes looked for an unpaired structure where the two streams could be brought together, and the gland he knew of that fits is the pineal. It is bad neuroanatomy, made in 1649 from the anatomy then available, and it is not a claim that the soul is a tiny object sitting in a cavity.

Common misconceptions

• The cogito is Descartes's argument for dualism. It is not. It establishes that he exists as a thinking thing while thinking. The distinctness of mind from body waits four Meditations and needs a premise about God.
• Descartes argued: I can doubt my body but not my mind, therefore they differ. That inference is invalid and Arnauld's objection is aimed at it. Descartes's stated argument runs through a complete positive conception of each substance, not through what he can doubt.
• Descartes said the soul lives in the pineal gland. He said the gland is the principal seat at which the soul exercises its functions, chosen because it is unpaired and the rest of the brain is doubled. The union is elsewhere described as involving the whole body.
• Dualism was a piece of religious conservatism. The mechanical philosophy Descartes was defending was the radical position of the period. The two-substance view is what let him treat every body, including the human body, as a machine to be explained by matter in motion.

What to carry forward

• Descartes's doubt is a filter for finding indubitable premises, not a claim that we are deceived.
• The cogito secures existence only while thinking, and characterises the thinker as a thing whose acts include sensing.
• The real distinction argument moves from clear and distinct conception to genuine possibility, and that step is underwritten by a non-deceiving God.
• Arnauld's triangle shows the danger in that step, and Descartes answers with a distinction between complete and incomplete conceptions that he never fully specifies.
• The divisibility argument is independent and needs no theology.
• The language and flexibility tests of the Discourse are the ancestor of the Turing test.
• Descartes insisted on an intimate union of mind and body, which makes the interaction question in the next lesson a question about his own view rather than an outsider's complaint.

Worth holding on to: the pattern of the argument, not its conclusion. Conceive of A without B, infer that A and B could come apart, conclude that they are two. Almost every later argument about consciousness reuses that pattern, and almost every later objection attacks the same joint.

Sources

1. Hatfield, G. (n.d.). Rene Descartes. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Robinson, H. (n.d.). Dualism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Descartes, R. (1641). Meditationes de prima philosophia. Project Gutenberg edition. gutenberg.org
4. Cottingham, J., Stoothoff, R., and Murdoch, D. (Trans.). (1984). The Philosophical Writings of Descartes, Volume II. Cambridge University Press. (Meditations, Objections and Replies.)
5. Kim, J. (2011). Philosophy of Mind (3rd ed.), chapter 2. Westview Press.

On 16 May 1643, from The Hague, a woman of twenty-four who had spent her whole life in exile wrote to Descartes. She apologised for her ignorance, mentioned that a visit had been cut short, and then asked the question his book could not answer. In Lisa Shapiro's translation she asks him to tell her "how the soul of a human being (it being only a thinking substance) can determine the bodily spirits, in order to bring about voluntary actions."

She was Elisabeth of the Palatinate, eldest daughter of a father who had lost the crown of Bohemia in 1620 and so helped to start the Thirty Years War. Descartes answered in five days and went on answering for six years. He dedicated the Principles of Philosophy to her in 1644 and wrote the Passions of the Soul at her request. Her side of that correspondence is the whole of her surviving philosophical work, and it contains the objection that has dogged every dualist since.

Why the question is not naive

Read it again and notice the precision. She does not ask whether the soul moves the body; she grants that. She asks how, and only about the soul acting on the body. The asymmetry is chosen, because the difficulty lies on that side.

Here is what makes it a difficulty. Descartes's own physics recognises exactly one way for a body to be set in motion: another body touches it. In Part II of the Principles, motion is transferred at contact, and God preserves the same total quantity of motion, measured as size times speed. Every legitimate causal explanation in Cartesian natural philosophy has that one shape, surface meeting surface. Now take an unextended thinking thing. It has no surface to press with, no size, no speed, no quantity of motion to hand over. The one model of causation Descartes permits is the one his soul cannot use.

Elisabeth frames it as a dilemma, and that is the part usually lost in summary. Accounts of causal efficacy available in 1643 tie efficacy to extension. So Descartes must either supply the account of causation proper to mind acting on body, or revise his account of mind so that the existing accounts apply to it. There is no third option, and the second costs him the real distinction.

Why this matters: the interaction problem is not a complaint that dualism feels mysterious. It is the observation that the dualist has no model of causation to offer and has explicitly ruled out the one his physics uses.

The answer Descartes gave, and what she did with it

His reply of 21 May 1643 has two parts. The first introduces three primitive notions: extension for bodies, thought for souls, and the union of the two, which is primitive as well and so not explained by the others. That is a refusal rather than an answer, and Elisabeth treats it as such.

The second part gets him into trouble. Scholastic natural philosophers explained a falling stone by a real quality, heaviness, inhering in the stone and moving it toward the earth's centre. Nothing touches the stone; the quality does the work. People who think this way, Descartes says, have no difficulty conceiving something immaterial moving a body, and we may conceive the soul on the same model.

Elisabeth's answer is the best objection in the correspondence. The model he borrows is the one he has spent his career destroying. The point of the mechanical philosophy, and much of the Principles, is that real qualities such as heaviness are empty words, and that the fall of a stone must be given a mechanical account. You may not declare a piece of machinery unintelligible in physics and then wheel it out in metaphysics. She also refuses to treat the two cases differently: if formal-causal explanation is no good for stones, it is no good for souls.

She then presses from the other direction, with evidence rather than logic. The condition of a body changes what its owner can think. Fever, exhaustion, what the period called the vapours: these do not merely distract the soul, they alter its capacity. In her letter of 1 July 1643 she says that it would be easier for her to grant matter and extension to the soul than to grant to an unextended thing the power to move a body and be moved by one.

Be careful with that sentence: it is often misread as a conversion to materialism, and it is not. In the 1645 letters on the passions she holds that we control what we attend to, so the order of thought does not simply follow the order of matter. Scholars divide over where she ends up. What she is certainly doing is naming the price of each exit.

Four ways to live with the problem

Every later position on mental causation is a choice about which of Elisabeth's two demands to abandon. Set the main historical options side by side.

Occasionalism is most often dismissed as an evasion, so take it at its strongest. Nicolas Malebranche, in the Search after Truth of 1674-75, argues from a general requirement: a true cause is one between which and its effect the mind perceives a necessary connection. Look at any finite thing and its supposed effect and you find none. Only an omnipotent will qualifies, so bodies do not move bodies either, and mind-body interaction is not a special embarrassment but an instance of the general case. Hume kept the first half of that argument and dropped the conclusion.

Leibniz takes the other route. In the Monadology of 1714 no substance has a window. The pre-established harmony holds that the two series correspond because each was wound in advance to match, like two clocks set going together. The cost is that the correspondence is brute, and the reason offered for it is the wisdom of the clockmaker.

Epiphenomenalism is the modern descendant. Thomas Henry Huxley, in an 1874 address on whether animals are automata, compared consciousness to the steam whistle of a locomotive: produced by the engine, contributing nothing to its motion. It survives because it concedes everything the physical sciences want and keeps experience real. Its problem is that it appears to undercut itself, and in Lesson 7 you will find its most famous defender withdrawing the defence partly on those grounds.

The problem outlives dualism

Here is the turn that makes this lesson more than history. Suppose you reject substance dualism entirely. You still have Elisabeth's problem, in a new form, and its sharpest statement is Jaegwon Kim's.

Start with the principle of causal closure: every physical event that has a cause has a sufficient physical cause. That is not a theorem. It is a bet about how physics has gone, and it is the best-supported bet in the vicinity. Now run the exclusion argument:

1. A mental event M causes a physical event P2, for instance a decision causing an arm to rise.
2. M occurs in virtue of some physical condition P, because the mental supervenes on the physical.
3. By closure, P2 has a sufficient physical cause, and the obvious candidate is P.
4. P2 is not systematically overdetermined by two independent sufficient causes.
5. So either M is identical with P, or M does not cause P2 after all.

The conclusion is a fork, not a refutation: reduce the mental to the physical, or concede that it does no causal work. Kim thinks most mental properties reduce functionally and survive. Qualia, he argues, do not, which is why his 2005 book is called Physicalism, or Something Near Enough.

Two serious answers are on the table. Donald Davidson's anomalous monism identifies each mental event with a physical event while denying strict laws between mental and physical kinds. Kim replies that this buys causation for events and loses it for properties: the event causes, but not in virtue of being mental, which is epiphenomenalism one level up.

The second answer is the generalisation objection, pressed hardest by Ned Block. Run the exclusion argument on any higher-level property and it comes out the same way: the heat of a fire does not boil the water, since the molecular kinetic energies are already sufficient. If the argument proves too much, something in it is wrong, and the usual diagnosis is premise 4, which treats a property and its realiser as two separate causes in competition.

In short: how frightening the exclusion argument is depends on what you take causation to be. On a counterfactual account of the kind David Lewis gave in 1973, the pain does cause the wince, since without it there would have been none. On a conserved-quantity account of the kind Wesley Salmon and Phil Dowe developed, causing requires transferring a conserved quantity, which the mental cannot do.

Common misconceptions

• Elisabeth objected that dualism violates the conservation of energy. Tidy, and anachronistic. Energy in the modern sense was formulated in the 1840s by Mayer, Joule and Helmholtz. Descartes's conserved quantity was size times speed, which Leibniz showed was the wrong quantity anyway. Her objection is about the intelligibility of the causal model, not a broken bookkeeping law.
• Descartes had no answer and changed the subject. He answered within the week, twice, and kept corresponding for six years. The answer was bad in a specific way, which is why her reply repays reading.
• Occasionalism is piety standing in for an argument. Malebranche derives it from a general requirement on causation that he defends independently, and applies evenhandedly to bodies.
• Epiphenomenalism is obviously self-refuting. The self-stultification objection is good but not obvious: it needs the further claim that a report about an experience must be caused by it.
• Only dualists face the interaction problem. Kim's argument is addressed to physicalists, and bites hardest on those who want mental properties to be irreducible.

Where this leaves us

• Elisabeth demands either a model of psychophysical causation or a revised account of mind, and Descartes supplied neither.
• His heaviness analogy borrows the Scholastic real quality he had already rejected, which is what she points out.
• Occasionalism, harmony and epiphenomenalism each solve the problem by giving up something large, and each is a position rather than a dodge.
• The problem is not peculiar to dualism: closure plus supervenience plus no overdetermination yields Kim's fork for any non-reductive view.
• The generalisation objection turns on whether a realiser and the property it realises are two competing causes or one cause described twice.

The upshot: the question of how thought moves an arm is at bottom a question about what causation is, which is why three and a half centuries of work on it have not produced an answer everybody accepts.

Sources

1. Shapiro, L. (2021). Elisabeth, Princess of Bohemia. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Robb, D., Heil, J., and Gibb, S. (n.d.). Mental Causation. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Bennett, J. (Trans.). (2017). Correspondence between Descartes and Princess Elisabeth, 1643-1649. Early Modern Texts. earlymoderntexts.com
4. Shapiro, L. (Ed. and Trans.). (2007). The Correspondence between Princess Elisabeth of Bohemia and Rene Descartes. University of Chicago Press.
5. Kim, J. (2005). Physicalism, or Something Near Enough, chapters 2 and 7. Princeton University Press.

A visitor is shown round Oxford. He sees Balliol, the Bodleian, the Clarendon Laboratory, the offices where the administration sits, and at the end of the tour he asks a polite question: where is the University? He has been shown it. He has been shown all of it. His question treats the University as one more item on the list, and that is the mistake around which Gilbert Ryle built the 1949 book that made the word behaviourism respectable in philosophy for about fifteen years.

That book, The Concept of Mind, argued that Descartes had made the visitor's mistake on a grand scale. Ryle named the target the dogma of the Ghost in the Machine, a phrase he coined and which has outlived his argument. The lesson ahead sets out what he actually claimed, why a parallel movement in psychology should not be confused with it, what Alan Turing proposed the year after, and the three objections that between them closed the position.

Two behaviourisms, and only one is a theory of mind

The word covers two different projects, and most arguments about behaviourism are conducted by people attacking one and defending the other.

Hempel's 1935 paper is the clean statement of the second. Take the sentence that Paul has a toothache. Its content, Hempel says, is exhausted by a list of testable sentences: Paul weeps and makes gestures of a certain kind, Paul says he has a toothache when asked, Paul's blood pressure and digestion alter in specified ways, a decayed tooth would be found on inspection. Nothing is left over. If nothing is left over, the mind-body problem does not arise, because there is no second thing to relate to the first.

Remember: analytical behaviourism is a thesis about meaning. It does not say that people lack inner lives; it says that talk about inner lives is talk about conduct. That is why the objections to it are all objections about translation.

What Ryle argued, and what he did not

Ryle's central move is the category mistake. Alongside the University there is the visitor watching a march past who, having seen the battalions, batteries and squadrons, asks when the division will appear; and the spectator who has seen the bowling, the batting and the fielding and asks where the team spirit is. In each case an item belonging to one logical type is expected to turn up as an item of another. Ryle's charge against Descartes is not that minds are unobservable but that the official doctrine treats them as things of the same type as bodies, differing only in being made of nothing.

His most durable argument is the regress against intellectualism. Suppose that acting intelligently requires first considering a proposition, a maxim or a rule, and then applying it. That prior act of consideration is itself something one can do well or stupidly. So it too requires a prior act of considering how to consider, and there is no end. The conclusion is that intelligence cannot be a matter of a preceding internal operation; it must be a matter of how the doing itself is done. This is the source of the distinction between knowing how and knowing that, and it is very much alive in current work on skill.

Two things Ryle is regularly said to have held, and did not. He did not claim that mental statements can be translated into behavioural ones; he repeatedly denied it, describing his project as mapping logical geography rather than supplying equivalences. And he did not deny sensations. What he denied was that a sensation is an episode observed by an inner eye, on the model of the outer one.

Turing's imitation game, stated as he stated it

In October 1950, Mind published Computing Machinery and Intelligence. Its first move is not the one people remember. Turing raises the question whether machines can think, then rejects it as too meaningless to deserve discussion, and replaces it with a question about a game.

The game as he sets it up has three players: a man, a woman, and an interrogator of either sex in a separate room, communicating by teleprinter so that voice and handwriting give nothing away. The interrogator must work out which respondent is the man and which the woman. The man's job is to cause a wrong identification; the woman's is to help toward a right one. Then comes the substitution: what happens when a machine takes the part of the man?

Turing then makes a dated, checkable prediction. In about fifty years, he writes, it will be possible to programme computers with a storage capacity of about 10⁹ to play the imitation game so well that an average interrogator will not have more than a 70 per cent chance of making the right identification after five minutes of questioning. That is a specific claim about a five-minute test against an average judge, not a definition of thinking, and the difference matters when people announce that some system has passed the Turing test.

Most of the paper is Turing answering nine objections. The one to notice is the argument from consciousness, which he takes from Geoffrey Jefferson's Lister Oration of 1949: a machine could be said to think only when it composed a sonnet because of thoughts and emotions felt, and knew that it had written it. Turing's reply is not that consciousness is unreal. It is that the only way to be sure a thing feels is to be that thing, so the objection, pressed consistently, ends in solipsism, and in practice we abandon it for each other. Whether that is a behaviourist criterion or a point about evidence is still argued over.

The three objections that closed it

Analytical behaviourism failed not because anyone proved that inner states exist but because the translations could not be written. Three cases did the work.

1. Circularity. Roderick Chisholm, in Perceiving (1957), pointed out that no belief has a behavioural upshot on its own. Believing that it will rain issues in taking an umbrella only given a desire to stay dry; wanting to stay dry issues in taking an umbrella only given the belief about rain. Every attempted translation of a mental term smuggles in another mental term. The programme cannot be completed even in principle.
2. Super-spartans. Hilary Putnam, in Brains and Behavior (1963), asks you to imagine a community that feels pain exactly as we do and suppresses every sign of it as a matter of discipline. That already breaks the translation. Then he pushes further: their descendants, the X-worlders, are so far from showing pain that they do not even have the suppressed impulse. Pain with no behaviour and no disposition to behave.
3. The perfect actor. The mirror image. Someone produces the writhing, the reports and the flinching, and feels nothing. Behaviour with no pain.

What matters here: the two thought experiments are aimed at different halves of the biconditional. Super-spartans show that behaviour is not necessary for pain; the perfect actor shows it is not sufficient. Chisholm shows that even where behaviour is present, it cannot be specified without using the vocabulary the theory was meant to eliminate.

What came through the wreck

Three things survived, and you will meet all of them again.

The regress argument survived intact. So did the relational insight: whatever pain is, it has something essential to do with what it makes you do. Functionalism, in Lesson 5, keeps that insight and drops the restriction that the relations must be to behaviour alone, allowing mental states to be defined partly by their relations to each other. That single amendment answers Chisholm's objection, which is why functionalism replaced behaviourism rather than merely defeating it.

Turing's test survived as a provocation rather than a criterion, and its history is a warning. ELIZA, written by Joseph Weizenbaum in 1966, reflected a user's sentences back as questions using a few dozen pattern-matching rules, and people confided in it. Weizenbaum was disturbed rather than pleased, and spent much of the next decade arguing that the readiness of people to be convinced was a fact about people. The Loebner Prize, run annually from 1991 to 2019, mostly rewarded entries good at deflection. Lesson 16 returns to all of this with systems Turing did not anticipate.

Common misconceptions

• Behaviourists denied that anyone has experiences. Skinner treated private events as behaviour accessible to one observer rather than as fictions; Ryle denied inner episodes construed on the model of observation, not sensation itself.
• Ryle was offering behavioural translations. He denied it in the book and afterwards. Conflating him with Hempel makes his position look easier to refute than it is.
• The Turing test is a definition of intelligence. Turing called the original question meaningless and offered a replacement question with a numerical prediction attached to a five-minute exchange.
• Chomsky's 1959 review of Verbal Behavior refuted behaviourism. It was a heavy blow to Skinner's account of language and it changed the field, but it says nothing about Ryle or Hempel, whose position fell to different objections in a different literature.
• Dispositions are a behaviourist invention that modern philosophy of mind has dropped. Dispositional analysis is alive and well; what failed was the claim that the manifestations could be specified without mental vocabulary.

The short version

• Methodological behaviourism is a research policy; analytical behaviourism is a thesis about what mental sentences mean. Only the second is a theory of mind.
• Ryle's category mistake charges Descartes with putting mind in the wrong logical type, and his regress argument shows that intelligence cannot consist in a prior inner act.
• Turing replaced the question whether machines think with a five-minute imitation game and a dated prediction about how well a machine would do in it.
• Chisholm's circularity, Putnam's super-spartans and the perfect actor between them show that behaviour is neither necessary nor sufficient, and that the translations cannot be written.
• Functionalism inherits the relational treatment of mental states and repairs it by letting mental states be defined partly in terms of one another.

The core of it: behaviourism was an attempt to dissolve the mind-body problem by showing there was only ever one subject matter. It failed on a technical point about translation, and the technical point taught its successor exactly what to do differently.

Sources

1. Graham, G. (n.d.). Behaviorism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Oppy, G., and Dowe, D. (n.d.). The Turing Test. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433-460.
4. Ryle, G. (1949). The Concept of Mind, chapters 1 and 2. Hutchinson.
5. Putnam, H. (1963). Brains and behavior. In R. J. Butler (Ed.), Analytical Philosophy, Second Series. Blackwell.

In February 1956 the British Journal of Psychology published a paper of seven pages by a man who taught philosophy inside a psychology department at the University of Adelaide. The title was a question, Is Consciousness a Brain Process?, and the answer was yes. What made the paper matter was not the answer, which had been available for a century, but a distinction its author drew in the second paragraph about what kind of yes it was.

U. T. Place had been arguing about this over lunch with two colleagues, C. B. Martin and J. J. C. Smart, who was then professor of philosophy at Adelaide. Smart later said he was converted against his will. His own paper, Sensations and Brain Processes, appeared in the Philosophical Review in 1959 and became the standard statement of the view. Between them they produced type identity theory, the first physicalism about the mind that was neither behaviourism nor a promissory note.

Which kind of identity claim it is

Behaviourism had tried to close the gap by an appeal to meaning: talk of pain just is talk of conduct. Place's insight was to give that up entirely. The statement that consciousness is a brain process, he says, is of the same logical kind as the statement that lightning is a motion of electric charges, or that a cloud is a mass of water droplets in suspension. Nobody learns those by reflecting on what the words mean. They were discovered, they could have turned out otherwise, and they are none the worse for that.

Once you see the claim that way, an entire family of objections stops working. You cannot refute it by pointing out that a person can talk perfectly well about their own toothache while knowing nothing about neurons, any more than you can refute the lightning identity by pointing out that farmers described lightning for millennia without a theory of charge. Knowing what a word means and knowing what the thing is are different achievements.

Place also names a diagnosis for why people resist. Call it the phenomenological fallacy: the assumption that when you describe an experience you are describing the literal properties of something you are inwardly inspecting. Report a green after-image and it can feel as though you have reported that something green is present, so that a search of the skull for green things is in order. Nothing green is there. What you have reported is that you are undergoing the sort of thing that goes on in you when you look at a green object in ordinary light.

Bottom line: the identity theory is an empirical bet about what a certain process will turn out to be, and its first defensive move is to insist that a claim of that kind cannot be refuted from the armchair by consulting what we mean.

Smart's argument, and the objection it was built to answer

Smart's positive case is an appeal to parsimony, and it borrows a phrase from Herbert Feigl: nomological danglers. Suppose sensations were irreducibly non-physical. Then the laws connecting brain events to sensations would be laws of a strange kind: they would relate the physical world to items that do nothing further, entering into no other law, explaining nothing downstream. They would hang off the side of science. Everything else in nature is knitted into the rest. Smart's conclusion is that there is no good reason not to accept the simpler hypothesis, that the sensation and the brain process are one thing.

The objection he takes most seriously came from Max Black. If a sensation is a brain process, then in reporting an orange after-image I report something orange, and no brain process is orange. So the report attributes a property no physical thing has, and a non-physical property is a non-physical something.

Smart's answer is the topic-neutral analysis, and it repays working slowly. He proposes that the report

• I see an orange after-image

be understood as

• There is something going on which is like what goes on when I have my eyes open, am awake, and there is an orange illuminated in good light in front of me.

Read the second sentence again and notice what it does not say. It does not say the something is physical. It does not say it is non-physical. It says only that there is a process resembling the process that ordinarily accompanies looking at an orange thing. Every substantive property has been pushed into a comparison with a publicly identifiable situation. The analysis is neutral about the topic, which is exactly why it leaves the field clear for science to identify what is going on, and why Black's objection no longer bites: no orange property has been attributed to anything.

One distinction before the objections. Type identity says that a kind of mental state is a kind of physical state: pain as such is a certain neural kind. Token identity says only that each particular mental event is some particular physical event, with no requirement that the kinds line up. Place and Smart were type identity theorists, and it is type identity that the next two lessons dismantle.

Kripke's objection, and why it is not the one you expect

The first blow came from an unexpected direction, in lectures Saul Kripke gave at Princeton in January 1970, published in 1980 as Naming and Necessity. Kripke's machinery is a claim about how certain words work. A rigid designator is a term that picks out the same thing in every possible situation in which that thing exists. Proper names are rigid; so, he argues, are natural kind terms such as water, heat, gold.

Now the consequence. If A and B are both rigid and A is B, then there is no possible situation in which A exists and is not B. So true identities between rigid terms are necessary. Water is H₂O necessarily, not contingently, even though we found it out by experiment.

But it feels contingent. Kripke explains the feeling away. When you imagine a world in which water is not H₂O, what you actually imagine is a world containing a colourless, drinkable, river-filling substance with some other composition. That substance is not water; it merely presents water's appearance. You have imagined an epistemic possibility and misdescribed it as a metaphysical one.

Apply the whole apparatus to pain. Pain and C-fibre stimulation are both meant to be rigid, so if the identity holds it holds necessarily. Yet it seems entirely possible for there to be pain without that neural process, and that neural process without pain. Can we explain that away in the water manner, by saying we have merely imagined something that presents pain's appearance? Here is the trap. For water there is a gap between how the stuff presents and what it is, and the explanation lives in that gap. For pain there is no gap. A state that feels like pain is pain; feeling painful is not pain's outward appearance but pain itself. The manoeuvre that saved the chemical identity has nothing to work on.

So what?: notice the irony. What Place and Smart offered as their central innovation, that the identity is contingent, is precisely what Kripke argues cannot be true of an identity between rigid terms. The theory is attacked at the point where it thought it was strongest.

Kripke did not present this as a demonstration that materialism is false, and he said so. It is a challenge: explain the appearance of contingency, or give up the identity. The main physicalist answer, that we possess two different concepts of one property and the gap is between the concepts rather than in the world, is the phenomenal concept strategy of Lesson 8.

A note on C fibres

Philosophers write pain equals C-fibre firing so often that it can look like a finding. It is not. C fibres are small unmyelinated afferents in the peripheral nervous system that carry slow, dull, poorly localised nociceptive signals toward the spinal cord, in contrast to the fast myelinated A-delta fibres responsible for the sharp first twinge. Nobody in pain science holds that the experience of pain is identical with activity in a class of peripheral fibres; pain involves spinal processing, thalamic relay, somatosensory, insular and cingulate cortex, and top-down modulation, and it can occur with no peripheral input at all. Treat the phrase as philosophers' shorthand for whatever the neural kind turns out to be, and do not repeat it as neuroscience.

Common misconceptions

• The identity theory says that the word pain means C-fibre firing. It says the opposite. The identity is empirical and, on Place's and Smart's account, contingent, which is what distinguishes it from behaviourism.
• Place denied that we have experiences. He denied that describing an experience is describing the properties of an inner object, which is a claim about what reports do, not a denial that anything is felt.
• Kripke proved that materialism is false. His argument targets type identity between rigid terms, and he offered it as a burden of proof rather than a refutation.
• C-fibre firing is neuroscience's account of pain. It is a placeholder that has been repeated until it looks like a result.
• Topic-neutral analysis is a way of translating sensation talk into brain talk. It is a way of translating sensation talk into something that mentions no properties at all, so that the brain talk can be added later by science rather than by analysis.

Putting it together

• Place's contribution was the form of the claim: an identity discovered like the identity of lightning with electrical discharge, not established by analysis of meaning.
• The phenomenological fallacy is the assumption that describing an experience describes the properties of an inner object.
• Smart's case is parsimony: irreducible sensations would require laws that connect to nothing else in nature.
• Topic-neutral analysis answers the objection from orange after-images by attributing no properties at all, only a resemblance to a public situation.
• Kripke's objection is that identities between rigid terms are necessary, and that the appearance-reality gap used to explain away apparent contingency for water does not exist for pain.

Sources

1. Smart, J. J. C. (n.d.). The Mind/Brain Identity Theory. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. LaPorte, J. (n.d.). Rigid Designators. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Place, U. T. (1956). Is consciousness a brain process? British Journal of Psychology, 47(1), 44-50.
4. Smart, J. J. C. (1959). Sensations and brain processes. The Philosophical Review, 68(2), 141-156.
5. Kripke, S. (1980). Naming and Necessity, Lecture III. Harvard University Press.

Somewhere in Putnam's 1967 paper Psychological Predicates there is an octopus. He puts it there to make a demand of the identity theorist, and the demand is concrete: name a physical-chemical state such that an organism is in pain if and only if it is in that state. The state has to be available to a mammalian brain, a reptilian brain and a mollusc's brain, because, as he notes in passing, octopuses are molluscs and certainly feel pain. The paper was reprinted under the title The Nature of Mental States and it ended type identity theory as a going concern within a decade.

This lesson is a procedure. We will run David Lewis's method for defining mental terms all the way through on a small worked theory, then run it a second time with one input changed and watch the result flip. What the method produces, and what it fails to produce, is the whole content of functionalism.

The argument in three steps

Putnam's reasoning is short. First, type identity requires a physical kind that is co-extensive with pain across every organism that has pain. Second, the nervous systems that support pain-like states are wildly different: a human, a bird, an octopus with most of its neurons in its arms. Third, the identity theorist facing that fact has only two moves and both are bad. Say that mammalian pain and octopus pain are different mental states, which is a strange thing to say about two creatures both writhing away from the same stimulus. Or identify pain with a long disjunction of physical states, at which point the identity has stopped doing any explanatory work, because disjunctions do not name kinds.

Putnam's positive proposal is that pain is a state of a whole organism defined by what it does: what produces it, what it produces, and how it interacts with other such states. In 1967 he framed this in terms of probabilistic automata, which is a detail worth remembering when people say multiple realisability shows that minds are software.

The upshot: the argument refutes type identity. It does not touch physicalism. It was constructed by a physicalist to make room for a better physicalism, and every party to the dispute agreed that whatever realises pain is physical.

Ramsification, worked all the way through

Lewis's contribution, in two papers of 1970 and 1972, was a recipe for turning that idea into definitions. Take a deliberately tiny psychological theory with three mental terms in it, so the machinery is visible.

The theory, call it T:

• Tissue damage causes pain.
• Pain causes distress.
• Distress, together with the belief that withdrawing will help, causes withdrawal.

Step 1. Sort the vocabulary. Two kinds of term appear. Tissue damage and withdrawal are already understood independently of psychology; Lewis calls these the O-terms. Pain, distress and the belief are the terms we are trying to define; call them the T-terms.

Step 2. Replace every T-term with a variable. The theory becomes an open sentence: tissue damage causes x; x causes y; y together with z causes withdrawal.

Step 3. Bind the variables. The Ramsey sentence of T says that the theory is realised without saying by what: there exist x, y, z such that tissue damage causes x, x causes y, and y with z causes withdrawal. Every mental term has vanished and no property has been named. Frank Ramsey invented the trick in 1929 for theoretical terms in physics; Lewis noticed it would work for psychology.

Step 4. Add uniqueness. Lewis strengthens the sentence to say that there is exactly one triple satisfying it. Without uniqueness the definitions in the next step have nothing definite to pick out.

Step 5. Define the T-terms as definite descriptions. Pain is now the first component of the unique triple realising T. In ordinary words: pain is whatever occupies the pain role. This is a definition, available a priori, and it names nothing physical.

Step 6. Ask the empirical question. What, in a human being, actually occupies that role? Suppose neuroscience answers: a certain neural state N. Then in human beings pain is N. That is a genuine identity, discovered rather than stipulated, and Lewis regarded his functionalism as a way of earning the identity theory rather than replacing it.

Now run it again with one input changed. Keep the theory word for word and put it in a creature built of silicon, in which some circuit occupies the pain role exactly, causing what pain causes and caused by what causes pain. Every clause of T is satisfied, so the Ramsey sentence is true of that creature, so step 5 delivers the same definition. Step 6 delivers a different answer. Pain in the silicon creature is that circuit. The definition did not move and the identity did.

Remember: that is multiple realisability made precise. The mental term is not rigid. It designates a role, and roles have different occupants in different populations, which is why Lewis relativises identities to a population rather than asserting them flatly.

Lewis pressed his own view with two hard cases in 1980. Mad pain: a man whose pain is caused by moderate exercise and causes him to concentrate on mathematics. Martian pain: a being whose pain is realised by hydraulic hardware in its feet. The first has the wrong causal role and is still pain; the second has the right role and the wrong hardware. Lewis's answer is that pain is what occupies the pain role for the appropriate population, so the madman counts as being in pain because he belongs to a population in which that state occupies the role, and the Martian counts because Martians are a population of their own.

Two grains of theory, and a dilemma

Everything above depends on where T comes from, and there are two answers.

Ned Block named that trade-off in Troubles with Functionalism (1978) and it has not been resolved. Make the functional conditions coarse enough to include an octopus and they will include things that plainly have no mind. His illustration is the Chinese nation: give each of a billion people a radio and a simple instruction relating incoming signals to outgoing ones, wire the whole population to a body, and the resulting system satisfies whatever functional description you like. Make the conditions fine enough to exclude that system and you will exclude the octopus too, and probably anything not built like a human.

Is multiple realisability actually true?

For thirty years the argument was treated as settled. Then William Bechtel and Jennifer Mundale asked a question about grain, in a 1999 paper in Philosophy of Science, and the answer is uncomfortable.

Look at how the argument is set up. The psychological state is described coarsely: pain, vision, memory. The neural state is described finely: this configuration, in this species, in this individual. Described that way, of course the mapping is many-to-one. But nothing licenses the mismatch. Match the grains and the multiplicity shrinks fast.

Their evidence is the practice of neuroscience itself. Korbinian Brodmann's cytoarchitectonic maps, published in 1909, were built by comparing cortical tissue across humans, monkeys and other mammals, and they are still in use. Visual area V1 is identified across species by a combination of cell layering, connectivity and function, and the identification is what makes it possible to learn about human vision from a macaque. If neural kinds really were species-specific in the way the multiple realisability argument assumes, comparative neuroanatomy would be a mistaken enterprise rather than the foundation of the field.

Lawrence Shapiro added a further condition in 2000. For a genuine multiple realisation the realisers must differ in ways that matter to how the job gets done. A corkscrew made of steel and one made of aluminium are not two realisations of anything interesting; they are one design in two metals. Applied to minds, this asks whether the silicon creature differs from us in a way relevant to occupying the pain role, or merely in a way that is irrelevant, and the question turns out to be hard to answer without already knowing what pain is.

Common misconceptions

• Functionalism is the claim that the mind is a computer program. Machine functionalism is one version. Lewis's is not computational at all: his roles are causal, and nothing in the ramsification procedure mentions computation.
• Multiple realisability refutes physicalism. It refutes the identification of mental kinds with physical kinds. Every realiser in the argument is physical, and Putnam intended it as a repair to physicalism.
• Putnam was a functionalist. He invented the position and abandoned it in Representation and Reality (1988), arguing that computational states cannot fix what a thought is about.
• Ramsification eliminates mental terms. It removes them as primitives and brings them back as definite descriptions of role occupants, which is a reduction rather than an elimination.
• Functionalism was refuted by the Chinese nation case. It is an argument from intuition about a case nobody can build, and functionalists have several answers to it, including biting the bullet.

What you now know

• Putnam's argument demands one physical kind co-extensive with pain across mammals, reptiles and molluscs, and concludes that mental kinds are functional rather than physical kinds.
• Lewis's procedure ramsifies a psychological theory, adds uniqueness, and defines each mental term as the occupant of a role, leaving the occupant to be found empirically.
• Changing the realiser leaves the definition untouched and changes the identity, which is what multiple realisability amounts to when written out.
• Analytic functionalism and psychofunctionalism differ over the source of the theory, and Block's dilemma says that no setting of the grain avoids both liberalism and chauvinism.
• Bechtel and Mundale argue that the multiplicity is an artefact of comparing coarse psychological kinds with fine neural ones, and point at comparative neuroanatomy as the counterexample.

Sources

1. Levin, J. (n.d.). Functionalism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Bickle, J. (n.d.). Multiple Realizability. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Putnam, H. (1967). Psychological predicates. In W. H. Capitan and D. D. Merrill (Eds.), Art, Mind, and Religion. University of Pittsburgh Press.
4. Lewis, D. (1972). Psychophysical and theoretical identifications. Australasian Journal of Philosophy, 50(3), 249-258.
5. Bechtel, W., and Mundale, J. (1999). Multiple realizability revisited: Linking cognitive and neural states. Philosophy of Science, 66(2), 175-207.

At Yale in the late 1970s, a program written by Roger Schank and Robert Abelson could read a short story about a man who went into a restaurant, ordered a hamburger, was brought a burnt one, stormed out without paying, and then answer the question of whether he ate it. The story never says. The program answered no, because it had a script for restaurants that filled the gaps. Its authors described this as a model of understanding.

John Searle read that claim and wrote a reply. It appeared in 1980 in Behavioral and Brain Sciences, a journal whose format prints a target paper together with commentaries and the author's responses, so the paper, twenty-seven objections and Searle's answers to them all arrived at once. Very few arguments in philosophy have been so thoroughly worked over so fast, and it is still not settled. This lesson takes both sides from their own texts.

The room, and the argument it is supposed to license

You are locked in a room with three batches of Chinese writing and a rule book in English. The rules tell you how to correlate one set of shapes with another, purely by their shape; you know no Chinese and the shapes mean nothing to you. Papers come in under the door and you send papers out. Unknown to you, the incoming papers are questions and the outgoing ones are answers, and they are good enough that the people outside take you for a native speaker. You understand nothing.

In 1984 Searle set the argument out as a short derivation:

1. Programs are formal, which is to say syntactic.
2. Minds have mental contents, which is to say semantics.
3. Syntax by itself is neither constitutive of nor sufficient for semantics.
4. Therefore programs are neither constitutive of nor sufficient for minds.

Be precise about the target, because Searle is regularly refuted for a view he does not hold. He is not arguing that machines cannot think: he says repeatedly that the brain is a machine and that it thinks. He is not arguing that only biological tissue can be conscious: he says any system with the right causal powers could be, and declines to say what those are. His target is the thesis he calls strong artificial intelligence, that an appropriately programmed computer thereby has a mind and that the program is the explanation of the mind. Weak AI, the use of computers as models and tools, he explicitly endorses.

What matters here: the entire dispute is about premise 3. Almost every reply is a reason for thinking that syntax embedded in the right way does yield semantics, and Searle's answer to almost every reply is a restatement of premise 3 applied to the new case.

The systems reply, given its strength

The reply that arrived first, from Berkeley, is that Searle has picked the wrong candidate. The man in the room is a processor. Nobody claims that a processor understands, any more than anyone claims that a single neuron in your left hemisphere understands English. The claim is about the system: the man plus the rule book plus the ledgers plus the room. Understanding is the kind of thing a whole system does, and the man's ignorance is no more surprising than a neuron's.

Searle's answer is to remove the system. Let the man memorise the rule book, keep the ledgers in his head, and step outside. He now contains everything there is; there is no part of the system he is outside of. And he still understands no Chinese.

Here is where you have to be careful, because the strongest version of the systems reply is a reply to exactly that move. Suppose the man really does internalise a program capable of native-level Chinese conversation. Then his skull implements two things: him, and the process he is simulating. Searle's report that he understands nothing is a report from the first process about itself. The second process also produces reports, in Chinese, and they say the opposite. Searle's confidence that there is one subject in there rather than two is precisely the point in dispute, so appealing to it is not an argument. That is the virtual mind version of the reply, developed by David Cole and others: understanding belongs to the agent the process implements, not to the implementer.

There is a second point about scale, pressed by Daniel Dennett among others. A rule book that sustains fluent conversation in Chinese is not a book. It is an object whose consultation, at human speed, would take longer than a civilisation. Asking your intuition to report on the inner life of a man who has memorised it is asking a faculty tuned to ordinary cases to pronounce on something no one has ever encountered. Intuitions are evidence, but they degrade with distance from experience, and this case is very far out.

The robot reply, given its strength

The second reply grants that symbols pushed around in a windowless room mean nothing, and adds a world. Put the program in a robot with cameras and arms. Now some symbols arrive from transducers pointed at actual objects and other symbols make the limbs move. Searle's counter is to climb inside: put the man in the robot's head, and he is still shuffling shapes, with no idea that some of them came from a camera.

This is the weakest passage in Searle's paper, and it is worth seeing why. The reply is not the claim that adding a body is magic. It is the claim that what a representation is about is fixed partly by its causal and historical relations to the world, so that a state caused by cats, tracking cats, and guiding behaviour toward cats is about cats. That thesis has a large literature behind it, and Lesson 10 gives it in detail. Searle's counter simply assumes it is false: he assumes that if the man cannot find the meaning by inspecting the symbols, there is no meaning. On an externalist account of content, that is exactly what you would expect, since the content is not the sort of thing that could be found by inspection.

Stevan Harnad's symbol grounding problem, set out in 1990, is the useful way to hold this. Harnad agrees with Searle that a system whose symbols are defined only by other symbols has no semantics, and compares it to learning Chinese from a Chinese-Chinese dictionary. He concludes not that computation is hopeless but that some symbols must be grounded in sensory categories, which is a design requirement rather than a refutation.

The brain simulator and the many mansions

The third reply raises the stakes. Suppose the program does not use scripts at all, but simulates the actual firing sequence of every neuron in the brain of a Chinese speaker, synapse by synapse. Then whatever the brain does, the program does.

Searle's response is the water pipes. Imagine the man operating a vast plumbing system whose valves are arranged to mirror the neurons; he turns taps according to the rules and Chinese comes out the other end. The pipes do not understand, and neither does he, and running the same organisation in a different medium has not helped. Behind this sits his slogan: simulation is not duplication. A simulated rainstorm leaves nobody wet; a simulated digestion digests no pizza.

The reply to the slogan is that it assumes the answer. Wetness and digestion are defined by their physical effects on other physical things, so a simulation of them cannot be the real thing. Whether cognition is like that, or whether it is the kind of thing constituted by organisation, is the question the argument was supposed to settle. A defender of the simulation can add the argument from Lesson 8 that a system functionally identical to you, at fine enough grain, cannot differ from you in what it experiences without absurdity.

The many mansions reply says that strong AI need not be committed to digital computers at all: whatever the right causal powers turn out to be, we might one day build them. Searle grants this without complaint, which tells you how narrow his thesis is meant to be.

What both sides accept, and where they actually part

Agreement first, because it is larger than the noise suggests. Nobody defends Schank's program as an understander. Nobody thinks the man in the original room understands. Everyone agrees that passing a text-only exchange is not conclusive proof of anything.

The parting is over two things. The first is evidential: whether an intuition about what it is like to be one component is evidence about the system that component helps constitute. Searle says yes and treats the case as decisive; his critics say the intuition is generated by a picture that puts you in the room, and that the picture is the argument. The second is substantive: whether causal embedding can turn syntax into semantics. Searle denies it in premise 3, and his critics say premise 3 is the conclusion wearing a false beard.

Why this matters: Searle himself says no experiment can settle it, and he is right that no experiment can, since both sides predict the same observable behaviour. What could move it is theoretical: a worked account of how grounding fixes content, an account of what makes a physical process an implementation of a given computation, or a theory of consciousness on which organisation suffices. Each of those is a live research programme, and each is aimed at a different premise.

Common misconceptions

• Searle argues that machines cannot think. He says the brain is a machine and thinks. His target is the sufficiency of running a program.
• Searle holds that only carbon-based brains can be conscious. He holds that the right causal powers are needed and declines to specify them, which is a weaker and more slippery claim.
• The systems reply was invented later by frustrated computer scientists. It arrived with the paper itself in 1980 and Searle answered it in the same issue.
• The Chinese room refutes the Turing test. It attacks the sufficiency of a program for a mind; Searle grants that behaviour is evidence and denies only that it settles the matter.
• The argument shows that syntax is not enough, full stop. It asserts that. Whether the assertion is a premise anyone must accept is what three decades of replies have been about.

Looking back

• Searle's argument is a four-line derivation whose whole weight sits on the premise that syntax is not sufficient for semantics.
• The systems reply relocates understanding from the component to the system, and its strongest form treats the internalised case as two processes in one skull.
• The robot reply rests on an externalist account of content, which Searle's counter presupposes to be false rather than refuting.
• The brain simulator reply forces the question of whether cognition is more like wetness or more like organisation, and the water pipes do not answer it.
• No experiment separates the positions, because both predict identical behaviour; only a theory of content, of implementation, or of consciousness could move it.

Sources

1. Cole, D. (n.d.). The Chinese Room Argument. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Hauser, L. (n.d.). Chinese Room Argument. Internet Encyclopedia of Philosophy. iep.utm.edu
3. Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417-457.
4. Searle, J. R. (1984). Minds, Brains and Science, chapter 2. Harvard University Press.
5. Harnad, S. (1990). The symbol grounding problem. Physica D, 42, 335-346.

A little brown bat weighs about eight grams and catches a moth in complete darkness by screaming at it. The pulses run above the range of human hearing and can exceed 100 decibels at the bat's mouth, and from the returning echoes the animal extracts distance, size, texture and the beat of the moth's wings, adjusting its call as it closes. It does this several times a second in flight. Something is going on in there, and it is not remotely like what goes on in you when you look across a room.

In 1974 Thomas Nagel used that animal to ask a question that changed the shape of the subject, and eight years later Frank Jackson used a woman in a black and white room to turn the same intuition into an argument with premises you could number. This lesson works both, and then works the three main replies, and ends with the fact that Jackson later concluded his own argument was unsound.

The bat, and what the question actually asks

What Is It Like to Be a Bat? opens with a criterion rather than a theory. An organism has conscious experience, Nagel says, if and only if there is something it is like to be that organism. That formula is doing no explanatory work at all; it is a way of fixing what we are talking about, so that the rest of the argument has a subject.

The choice of animal is deliberate on two counts. A bat is a mammal with a nervous system much like ours, so nobody has to be persuaded that it is conscious, which is where an argument built on an amoeba would stall. And echolocation is a sensory world with no analogue in ours, so we cannot fake understanding by analogy the way we can with a dog's nose.

Now the part almost everyone gets wrong. Nagel is not asking what it would be like for you to be a bat. That question is about a human being fitted with unfamiliar equipment, and it has an answer, roughly: disorienting. His question is what it is like for the bat to be the bat, and the difficulty is that our concepts for experience are drawn from our own case.

The conclusion he draws is more careful than its reputation. Facts about experience, he argues, are accessible only from one type of point of view. Physical facts are objective, in the specific sense that they are available from many points of view and increasingly so as science proceeds. Reduction works by moving away from the observer's standpoint toward a description that does not depend on it. So a reduction of experience to the physical would have to abandon the very feature that makes it experience. And then he adds the sentence that is nearly always dropped: he is not arguing that physicalism is false, but that we have no idea what it would mean for it to be true.

Key idea: Nagel's paper is an argument about understanding, not about metaphysics. Joseph Levine gave it its lasting name in 1983 with the explanatory gap: the identification of pain with a neural state leaves a residue that the identification of water with H₂O does not, because knowing the chemistry tells you why water behaves as it does, and knowing the neurology does not tell you why the state feels like that.

Mary, set out as an inference

Frank Jackson's 1982 paper Epiphenomenal Qualia takes the gap and turns it into a lever. Mary is a brilliant scientist confined for whatever reason to a black and white room and investigating the world through a black and white monitor. She specialises in the neurophysiology of vision and comes to know everything physical there is to know about what happens when we see ripe tomatoes or the sky: wavelengths, retinal responses, the whole causal chain up to the production of the sentence the sky is blue. Then she is let out, or given a colour screen.

Jackson's line is that it seems just obvious she will learn something. Written out:

1. Before release, Mary knows all the physical facts about human colour vision.
2. On release, Mary learns something: what it is like to see red.
3. So there are facts about colour vision that she did not know before.
4. So there are facts about colour vision that are not physical facts.
5. So physicalism is false.

Keep this distinct from its neighbours. The knowledge argument runs from an epistemic premise to a metaphysical conclusion. The zombie argument of the next lesson runs from conceivability through possibility. The explanatory gap is weaker than both and, as Levine insisted, metaphysically neutral. Treating the three as one argument is the commonest confusion in this area.

Three replies, each in its own strongest form

Every serious reply attacks step 3 or step 4. Nobody denies that something happens to Mary when she sees the tomato.

The ability hypothesis. Laurence Nemirow proposed it in 1980 and David Lewis developed it in What Experience Teaches in 1988. What Mary gains is not information but a set of abilities: to imagine seeing red, to remember it, to recognise it when it recurs. Knowing what an experience is like just is having those abilities, and abilities are not facts, so no new fact has been learned and step 3 fails. Lewis liked the view partly because it explains why experience is the only reliable way to get there: you learn to ride a bicycle by riding one.

The objection is that Mary seems to acquire more than know-how. On seeing the tomato she can think, so that is what red looks like, and that thought can be true, and she could get it wrong if shown a trick colour. Beliefs that can be true or false look propositional, and abilities cannot be true or false. There are also awkward cases in both directions: a neurosurgeon might install the recognitional abilities without the experience, and a person with severe amnesia might have the experience without retaining the abilities.

The acquaintance reply. Earl Conee argued in 1994 that Mary gains a third thing, neither propositional knowledge nor know-how: acquaintance with a quality she had previously known only by description. The distinction is Russell's, and it applies to ordinary cases: you can know a great deal about a city without being acquainted with it. On this view Mary becomes acquainted with a property she already knew about, and physicalism is untouched because that property is physical.

The old fact under a new concept. This is the most widely held reply, associated with Terence Horgan, Brian Loar, Michael Tye and David Papineau. Mary learns no new fact. She acquires a new concept of a fact she already knew. Concepts are ways of presenting properties, and one property can be presented in two ways with no a priori link between them, exactly as Hesperus and Phosphorus present one planet. Phenomenal concepts, on this account, are recognitional: they refer directly, they are formed by having the experience, and they cannot be derived from physical concepts, which is why the gap feels permanent even though there is nothing in the world on either side of it.

Chalmers's reply to that strategy is a dilemma. Either the physical facts entail that we have concepts with those peculiar features, or they do not. If they do, then the very explanatory gap the strategy was invoked to explain should be closable, and it is not. If they do not, then something about our conceptual situation is itself not physically explicable, which concedes the dualist's point one level up.

In short: the ability and acquaintance replies deny that Mary gains propositional knowledge; the phenomenal concept strategy grants that she does and denies that the fact is new. The three are not variants of one answer, and they have different weak points.

Jackson takes it back

By 1998 the author of the argument had stopped believing it, and his reason is instructive. The knowledge argument, if sound, makes qualia non-physical. But then either they have physical effects or they do not. If they do, Mary would have learned about them while studying the physics, so the argument's first premise collapses. So they do not, and the position is epiphenomenalism, which the 1982 title had already conceded.

Jackson came to think that intolerable, on grounds we met in Lesson 2. If qualia cause nothing, they do not cause your belief that you are having one, or the words you use to report it. Your convictions about your own experience would then line up with your experience by coincidence, and a coincidence is not knowledge. Finding no way for a dualist to escape that, he concluded the argument must contain a false premise, and located it in the intuition that Mary learns a new fact, which he now explains as an artefact of how perceptual representation works.

Report this accurately. Jackson changed his mind; the profession did not follow him in a body. Plenty of philosophers think he abandoned a good argument because he could not stomach one of its consequences, which is a reason to look harder at the consequence rather than at the argument.

Common misconceptions

• Nagel argued that physicalism is false. He argued that we do not understand how it could be true, and said so in the paper. The stronger claim is other people's.
• The bat question asks what it would be like for you to be a bat. That is a question about a human in unusual equipment. Nagel asks what it is like for the bat.
• Mary is just someone who has not seen colour yet. The case requires complete physical knowledge, which is what makes it different from every ordinary case of learning by experience.
• Jackson's recantation refutes the knowledge argument. It is one philosopher's change of mind, driven by his view of epiphenomenalism, and the argument retains able defenders.
• The knowledge argument and the zombie argument are the same argument. One runs from knowledge to facts, the other from conceivability to possibility, and a reply to one need not touch the other.

The takeaway

• Nagel's criterion fixes the subject matter, and his conclusion is about the limits of our understanding rather than about what exists.
• Levine's explanatory gap is the epistemological residue: neural identities do not explain the feel the way chemical identities explain the behaviour.
• The knowledge argument turns that residue into a five-step inference from complete physical knowledge to the falsity of physicalism.
• The ability and acquaintance replies deny that new propositional knowledge is acquired; the phenomenal concept strategy grants it and denies that the fact is new.
• Jackson abandoned the argument because it entails epiphenomenalism, which he judged to make self-knowledge accidental, and the profession is still divided about whether he was right.

Sources

1. Nida-Rumelin, M., and O Conaill, D. (n.d.). Qualia: The Knowledge Argument. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Tye, M. (n.d.). Qualia. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Nagel, T. (1974). What is it like to be a bat? The Philosophical Review, 83(4), 435-450.
4. Jackson, F. (1982). Epiphenomenal qualia. The Philosophical Quarterly, 32(127), 127-136.
5. Lewis, D. (1988). What experience teaches. In J. Copley-Coltheart (Ed.), Proceedings of the Russellian Society. University of Sydney.

In April 1994 a conference in Tucson called Toward a Science of Consciousness put neuroscientists, physicists, anaesthetists and philosophers in the same room for a week. A twenty-seven-year-old postdoctoral fellow gave a talk in which he split the subject in half with a phrase, and by the end of the week people who disagreed with everything else he said were using it. The phrase was the hard problem of consciousness, and the talk became a paper in the Journal of Consciousness Studies the following year.

What follows is the deepest disagreement in this course. David Chalmers and Daniel Dennett know each other's arguments completely, have debated for three decades, and have not converged, because they disagree about what an explanation of consciousness would have to accomplish. Both positions are set out here from their own writing.

The division that made the phrase stick

Chalmers's list of easy problems is specific: the ability to discriminate and categorise stimuli, the integration of information by a cognitive system, the reportability of mental states, a system's access to its own internal states, the focus of attention, the deliberate control of behaviour, and the difference between waking and sleep. Nobody has solved them. He calls them easy anyway, and the label is a joke about method rather than about difficulty. For each of them we know in advance what an answer would look like: exhibit a mechanism that performs the function. When the mechanism is found, the problem is finished.

The hard problem is that when every mechanism is in place, a question appears to remain: why is any of this accompanied by experience? A system could, on the face of it, do all that discriminating and reporting in the dark. The complaint is not that we lack a mechanism. It is that a mechanism is the wrong shape of answer.

The point: the contribution here is the division itself. Chalmers concedes, generously and in advance, that functional explanation handles everything functional. The argument is entirely about whether there is anything left.

The zombie argument, premise by premise

A philosophical zombie is a molecule-for-molecule duplicate of a conscious person, in a world with the same physical laws, in which there is no experience at all. It says ouch, withdraws its hand, reports being in pain, and nothing is going on inside. The argument in The Conscious Mind (1996) runs:

1. Zombies are conceivable.
2. Whatever is conceivable in the relevant sense is metaphysically possible.
3. If zombies are metaphysically possible, consciousness is not physical.
4. Therefore consciousness is not physical.

Two clarifications before the objections. Metaphysical possibility is not biological or physical possibility: Chalmers holds that zombies could not evolve, could not be built, and are not possible in any world with our laws plus whatever laws connect brain to experience. And in his later work the conclusion is a disjunction rather than a flat denial of physicalism: either standard physicalism is false, or a Russellian monism of the kind Lesson 17 describes is true, and that view is arguably a physicalism.

Each premise has a serious opponent.

Premise 1 is where Dennett attacks. He argues that nobody actually conceives a zombie. What people imagine is a person with a lot of the machinery missing, and then tell themselves that they have imagined the machinery present. Imagine the duplicate properly, including everything it says about its own inner life and every judgement it makes about the redness of red, and there is nothing left to withhold. His name for the thing people think they are imagining is a zimbo, and his claim is that a zimbo is us.

Premise 2 is where the a posteriori physicalist attacks. It is conceivable, in some sense, that water is not H₂O, and that is not possible; so conceivability is a poor guide. Chalmers's answer uses two-dimensional semantics. In the water case the term has two intensions, one fixing reference by the watery appearances and one picking out the substance, and the conceivable scenario is a case where the first is satisfied and the second is not. Phenomenal terms have no such gap, because there is no distinction between how experience seems and what it is. The escape route that works for water is unavailable here for the same reason Kripke gave in Lesson 4.

Premise 3 is where Russellian monism attacks. Grant the possibility and deny the conclusion: if physics describes only structure and dispositions, it never tells us what the things having those dispositions are in themselves, and experience could be what occupies that space. The zombie world would then be a world differing in intrinsic natures while matching in structure, which is a real difference and not a non-physical one on every construal.

Dennett's position, stated so he would accept it

Dennett is regularly summarised as denying that consciousness exists. He does not, and the summary makes his actual view invisible. He affirms that we are conscious. What he denies is that consciousness has properties beyond those a functional account can capture, and that our introspective convictions are authoritative about their own subject matter.

Consciousness Explained (1991) attacks a picture he calls the Cartesian Theatre: the assumption that somewhere in the brain there is a place and a moment at which content becomes conscious, presented to an audience. The multiple drafts model replaces it. Content is fixed at various places and various times, revised continuously, and there is no further fact about when it entered awareness beyond the effects it has on memory, report and behaviour. His evidence is the family of cases where the question of when has no stable answer: colour phi, where two differently coloured dots flashed in sequence are seen as one dot moving and changing colour partway, apparently before the second dot appears; metacontrast masking, where a stimulus is wiped from report by a later one; and change blindness, in which large alterations to a scene go unnoticed across a saccade.

His method is heterophenomenology: take everything subjects say about their experience with complete seriousness as data about what they believe, and remain neutral about whether the beliefs are true. That neutrality is the crux. A theory that explains every judgement a person makes about their experience has, on Dennett's view, explained the experience, because there is nothing else that the judgements are about.

In Quining Qualia (1988) he presses the same point on the concept itself. Qualia are traditionally said to be ineffable, intrinsic, private and directly apprehensible in consciousness. Dennett constructs cases, most famously two coffee tasters who both stopped liking the product and disagree about whether their tastes or their standards changed, in which the four properties come apart and no amount of introspection settles which description is correct. If introspection cannot decide, the properties were never doing the work claimed for them.

Three shapes of physicalist answer

It helps to see that physicalists divide, and that Dennett occupies only one of the positions.

Keith Frankish's illusionism is the sharpest version of type A. He proposes replacing the hard problem with the illusion problem: explain why we represent ourselves as having phenomenal properties when nothing has them. He is candid about the price. The view requires saying that the most obvious thing in the world is a misrepresentation, and he treats that as a cost to be paid rather than a difficulty to be waved away. Note what it does not say: illusionism does not claim that your pain is not real or not painful. It claims that the intrinsic, ineffable character pain seems to have is a feature of how pain is represented rather than of pain.

Colin McGinn's type C position holds that the property connecting brain to experience is real and physical, and that our cognitive equipment cannot represent it, as a rat's equipment cannot represent arithmetic. The standard objection is that claims of cognitive closure are extremely hard to establish and very easy to reach for.

What would move each side

No experiment separates them, and both would say so. What could move things is more specific than that admission suggests.

• Chalmers would be moved by a functional account of introspection that predicts and dissolves the zombie intuition, showing why a purely physical system would inevitably find itself convinced of a gap.
• Dennett would be moved by a fact about experience that makes a difference nothing functional accounts for, though he has argued that any such difference would be functional by definition.
• The type-B physicalist is waiting on whether phenomenal concepts and their peculiar features can themselves be given a physical explanation. Chalmers's dilemma from Lesson 7 is aimed exactly there.

The core of it: the argument is not about evidence. It is about whether explaining every function leaves a question standing, and each side finds the other's answer to that not merely wrong but hard to take seriously, which is the mark of a disagreement about what counts as an explanation.

Common misconceptions

• Dennett denies that consciousness exists. He denies a theory of it. He affirms that we are conscious and denies that introspection reveals what it is.
• Zombies are supposed to be a scientific possibility. Chalmers holds that they are not possible in any world with our laws, and the argument turns on metaphysical possibility alone.
• The hard problem is the mind-body problem renamed. The contribution is the division: conceding functional explanation for everything functional, and asking what remains.
• Illusionism says you are not in pain. It says the pain is real and that its apparently intrinsic phenomenal character is a feature of the representation rather than of the state.
• The zombie argument shows that physicalism is false. Chalmers's own conclusion is a disjunction, and Russellian monism sits on the other branch of it.

Summing up

• Chalmers divides the field by conceding that mechanisms explain functions and asking why the functions are accompanied by experience.
• The zombie argument has three premises, and conceivability, the conceivability-to-possibility step, and the possibility-to-non-physicality step each have serious opponents.
• Dennett attacks the first premise and replaces the Cartesian Theatre with multiple drafts, using colour phi, masking and change blindness as evidence that there is no fact about when content becomes conscious.
• Physicalists divide into those who deny the epistemic gap, those who explain it by phenomenal concepts, and those who think it may be permanent for creatures like us.
• Illusionism is the clearest type-A position and is explicit that its cost is calling the most obvious thing in the world a misrepresentation.

Sources

1. Chalmers, D. (1995). Facing up to the problem of consciousness. Journal of Consciousness Studies, 2(3), 200-219. consc.net
2. Kirk, R. (n.d.). Zombies. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Chalmers, D. (1996). The Conscious Mind, chapters 3 and 4. Oxford University Press.
4. Dennett, D. (1991). Consciousness Explained, chapters 5 and 12. Little, Brown.
5. Frankish, K. (2016). Illusionism as a theory of consciousness. Journal of Consciousness Studies, 23(11-12), 11-39.

In 1990 Francis Crick and Christof Koch published a short paper arguing that consciousness had been left to philosophers long enough and proposing somewhere to start: oscillations near 40 hertz in visual cortex, binding separately processed features into a single percept. The specific proposal did not survive. The strategy did, and the phrase they made respectable, the neural correlate of consciousness, now organises an entire field.

This lesson compares the two theories that have dominated the last two decades, reads the experiment that was designed to put them against each other, and then does the part philosophers are for: saying precisely what a correlation of this kind can and cannot establish.

What a correlate is

Chalmers gave the standard definition in 2000. A neural correlate of consciousness is a minimal neural system whose states are sufficient, under matched background conditions, for corresponding states of consciousness. Minimal matters: the whole brain is trivially sufficient and tells you nothing. Two varieties are usually distinguished, the correlate of being conscious at all, which involves brainstem arousal systems and thalamocortical loops, and the correlate of content, which is what you are conscious of.

The methodological problem is obvious once stated. To study consciousness you must vary it while holding the stimulus fixed, or hold it fixed while varying the stimulus. Four paradigms do this.

• Binocular rivalry: present a face to one eye and a house to the other and the percept alternates every few seconds while the retinal input never changes. Nikos Logothetis's recordings in macaques found that in primary visual cortex few neurons follow the percept, while in inferotemporal cortex most do.
• Masking: a stimulus shown briefly and followed by another becomes unreportable, so identical stimuli can be sorted into seen and unseen.
• No-report paradigms: infer the percept from eye movements or pupil dilation instead of asking, in order to strip out the machinery of answering.
• Perturbation: deliver a magnetic pulse and measure how the cortex echoes it. Adenauer Casali and colleagues in 2013 turned this into the perturbational complexity index, which separates wakefulness, dreaming, anaesthesia and disorders of consciousness better than behaviour does, and works in patients who cannot respond.

Worth holding on to: every one of these paradigms compares a conscious state with an unconscious one and looks for the difference. That design finds differences. It does not, on its own, tell you which difference is the consciousness, which is the cause of it, which is the effect, and which is the apparatus of report.

The two theories side by side

Global workspace theory begins from a functional observation. Most processing in the brain is local, specialised and unconscious; a small amount of information at any moment becomes available to everything at once, to speech, to memory, to planning, to voluntary action. Bernard Baars modelled this as a theatre in 1988; Stanislas Dehaene and Jean-Pierre Changeux gave it a neuronal version in which long-range pyramidal cells in prefrontal and parietal cortex form a workspace, and a stimulus becomes conscious when activity crosses a threshold and ignites across that network. The prediction is a late, sudden, widely distributed event, and masked stimuli that fail to reach it produce early sensory activity and no ignition.

Integrated information theory runs in the opposite direction. Giulio Tononi starts from features he takes to be self-evident about any experience: it exists for the subject, it is structured, it is specific, it is unified, and it is definite. Each is converted into a requirement on a physical substrate, and consciousness is identified with a maximally irreducible causal structure, quantified as phi. The theory makes a striking prediction that the workspace theory does not: a feedforward network can reproduce any input-output behaviour you like and still have zero phi, so a system could behave exactly like you and be a zombie.

Its best empirical point concerns the cerebellum. It holds roughly 69 billion neurons, more than four times the cerebral cortex, yet people born without one can be conscious, articulate and self-aware. On the workspace account this needs a story. On Tononi's account it falls out of the architecture: cerebellar circuits are largely feedforward parallel modules with little integration, so their phi is low regardless of neuron count.

Its worst problem is the same coin. Scott Aaronson showed in 2014 that simple grids of logic gates, arranged as expander graphs, can be assigned arbitrarily large phi. Tononi's response was to accept the consequence and say such a system would be conscious, which many readers took as a reductio and he took as a prediction. In September 2023 a large group of researchers posted a letter calling the theory pseudoscience, which produced a second argument about whether unfalsifiability was the right charge and about the propriety of the letter itself.

The 2025 collaboration, read carefully

In 2019 the two camps did something unusual. Proponents of each theory sat down with a theory-neutral consortium and agreed, in advance and in public, on an experimental design, on what each theory predicted, and on what results would count as a challenge to their own view. The protocol was preregistered. The results were published in Nature in 2025 by the Cogitate Consortium.

The scale is worth stating: 256 participants viewing suprathreshold stimuli for variable durations, measured with functional magnetic resonance imaging, magnetoencephalography and intracranial electroencephalography in patients with implanted electrodes. What they found was information about conscious content in visual, ventrotemporal and inferior frontal cortex; sustained responses in occipital and lateral temporal cortex tracking how long the stimulus was present; and content-specific synchronization between frontal and early visual areas.

Then the part that matters. For integrated information theory, the absence of sustained synchronization within posterior cortex contradicts the claim that network connectivity of that sort specifies conscious content. For global neuronal workspace theory, there was a general lack of ignition at stimulus offset, which the theory had predicted, and a limited representation of certain dimensions of conscious content in prefrontal cortex. Each theory was challenged on a claim its own proponents had agreed in advance to treat as a test.

Bottom line: nothing was refuted and nothing was confirmed. Both theories were dented in places they had staked out beforehand, which is the only way a dent can be honest. The durable result is arguably the method: in a field where each theory had spent twenty years accumulating its own confirmations, a preregistered adversarial design is a change in how the question is asked.

What a correlate can and cannot show

Set the two lists against each other, because most overclaiming happens by sliding from one to the other.

It can identify systems whose state predicts reported experience under matched conditions; it can rule candidates out, as the cerebellum case rules out raw neuron count; it can ground clinical inference about patients who cannot answer, which is what the perturbational complexity index is for and is the most valuable thing the field has produced; and it can force theories to make different predictions, which is what the 2025 study did.

It cannot tell you whether the correlate is the experience, produces it, is produced by it, or is a precondition of it. Correlation is compatible with all four, and no amount of better imaging distinguishes them, because the alternatives predict the same data. It cannot escape the report problem completely: no-report paradigms replace one inference with another rather than removing inference. And it cannot settle the hard problem, which Chalmers pointed out in the same paper that defined the correlate.

There is one further limit worth naming. Any theory of consciousness gets tested on subjects who can tell us what they experienced, and is then extrapolated to subjects who cannot: patients, infants, animals, machines. The extrapolation is where the entire practical value lies and where no experiment can validate the theory, because the cases that would validate it are the cases we cannot check. That is not an argument against doing the science. It is a statement of why the philosophy does not go away when the imaging improves.

Common misconceptions

• Finding the neural correlate would solve the problem of consciousness. Complete success yields a correlation, and the interpretive question about what the correlation means is untouched by it.
• The 2025 study refuted integrated information theory. It challenged specific predictions of both theories, and the workspace account was challenged on ignition and on prefrontal content.
• The cerebellum case shows that neuron number is irrelevant to anything. It shows that architecture matters more than count for consciousness. Cerebellar damage impairs coordination, timing and aspects of cognition.
• Anaesthesia switches the cortex off. It changes the dynamics: cortex still responds to stimulation, but the response stays local and stereotyped instead of spreading and differentiating, which is what the complexity index measures.
• A theory of consciousness can be validated on animals or machines. Validation happens where reports are available; everything beyond that is extrapolation, and the extrapolation is a philosophical commitment.

Pulling it together

• A neural correlate is a minimal system sufficient for a conscious state under matched conditions, and the paradigms that find one all work by comparing seen with unseen.
• Workspace theory identifies consciousness with global broadcast and predicts a late frontoparietal ignition; integrated information theory identifies it with irreducible causal structure and predicts a posterior hot zone.
• The cerebellum is integrated information theory's best case and high-phi logic grids are its worst, and Tononi accepts the second as a prediction rather than a reductio.
• The 2025 preregistered collaboration challenged both theories on claims their proponents had agreed in advance to treat as tests, and refuted neither.
• Correlates cannot distinguish identity, causation, effect and precondition, cannot fully escape the report confound, and are validated only on subjects who can tell us what they experienced.

Sources

1. Cogitate Consortium. (2025). Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 642, 133-142. pmc.ncbi.nlm.nih.gov
2. Wu, W. (n.d.). The Neuroscience of Consciousness. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Baars, B. J. (1988). A Cognitive Theory of Consciousness. Cambridge University Press.
4. Tononi, G. (2004). An information integration theory of consciousness. BMC Neuroscience, 5, 42.
5. Casali, A. G., and others. (2013). A theoretically based index of consciousness independent of sensory processing and behavior. Science Translational Medicine, 5(198).

Hilary Putnam could not tell an elm from a beech. He said so in print, in 1975, and blushed about it in a parenthesis. Whatever is in his head when he thinks elm is exactly what is in his head when he thinks beech, and yet his word elm refers to elms. Something other than the contents of his head must be doing that work, and the paper it appears in, The Meaning of Meaning, spends fifty pages saying what.

This lesson is a procedure run twice. The first run is the famous one about a planet; the second uses no science fiction at all, and is the better argument. Both are attempts to answer a question we left open in Lesson 6, when the robot reply to Searle depended on causal relations between a symbol and the world fixing what the symbol is about.

First run: the procedure on a natural kind term

Step 1. Choose a term. Water.

Step 2. Fix a speaker at a time. Oscar, on Earth, in 1750. Chemistry does not yet exist. He knows water is the clear drinkable stuff in rivers and taps and he knows nothing about its composition.

Step 3. Build a duplicate. Twin Oscar lives on Twin Earth and is a molecule-for-molecule copy of Oscar, with the same brain in the same state, the same memories, the same dispositions. Putnam notes and waives one wrinkle: a human body is largely water, so a strict duplicate is awkward if the local liquid differs. The argument does not depend on the wrinkle and you can make the duplication hold of the brain alone.

Step 4. Change one thing, and only in the environment. On Twin Earth the clear drinkable stuff in the rivers and taps is not H₂O but a compound Putnam abbreviates XYZ, indistinguishable by any test available in 1750.

Step 5. Ask what each word refers to. Oscar's word picks out H₂O. Twin Oscar's picks out XYZ. They have different extensions.

Step 6. Ask whether the speakers differ psychologically, narrowly construed. By stipulation they do not. Same brain state, same introspectible content, same everything inside the skin.

Step 7. Draw the conclusion. Extension is not determined by narrow psychological state. Putnam's version of this is the sentence that got quoted for forty years: meanings just ain't in the head.

Two mechanisms make the case work, and it is worth separating them because they are different theses. The first is the indexicality of natural kind terms: water functions as whatever is the same liquid as this, with a demonstrative pointing at the local samples, so the reference is fixed by the world rather than by a description in the speaker's mind. The second is the division of linguistic labour, which is what the elm and the beech illustrate. My word gold refers to gold not because I can assay it but because I belong to a community containing people who can, and the labour of fixing extension is distributed across it.

Remember: the load-bearing stipulation is step 6. Everything else could be granted by an opponent. If someone argues that Oscar and Twin Oscar are not in the same psychological state because psychological states are wide, they have conceded externalism rather than refuted it.

Second run: the same procedure with the science fiction removed

Tyler Burge's 1979 paper Individualism and the Mental runs an identical structure on a case that could have happened last week.

Step 1. Choose a term: arthritis.

Step 2. Fix a speaker: Bert, who has various genuine arthritic complaints and now has a pain in his thigh. He tells his doctor that his arthritis has spread to his thigh.

Step 3. Note what happens next. The doctor explains that arthritis is an inflammation of joints and cannot occur in a thigh. Bert is surprised and accepts the correction. Note that his original utterance was false and that he nonetheless said something, and something specific: he had a belief about arthritis, mistaken in an ordinary way.

Step 4. Change one thing, and only outside the skin. Keep Bert physically identical, and alter the linguistic community around him so that in that community the word arthritis has always applied to rheumatoid ailments of joints and of thighs alike.

Step 5. Ask what he now believes. In the counterfactual community he has said something true, and the belief he expressed is not a belief about arthritis at all, because that community's word picks out a different, broader condition.

Step 6. Note that nothing inside Bert changed.

Step 7. Conclude that the content of a thought depends on relations to a community, not only on the state of the thinker.

Three features make this the stronger argument. It needs no chemistry, no alternative planet, and no assumption about microstructural essences. It works for terms with no natural kind behind them at all: contract, sofa, brisket, mortgage. And it works precisely because Bert incompletely understands the term, which is the normal condition of every speaker with respect to most of their vocabulary. Burge prefers to call the position anti-individualism rather than externalism, because the thesis is about how mental states are individuated rather than about where meanings are stored.

The objection that will not go away

Here is the price. Psychological explanation looks as though it needs states that physical duplicates share. Oscar and Twin Oscar reach for the glass in exactly the same circumstances with exactly the same movements. If their beliefs differ, and their behaviour is identical, then the difference in belief is doing no explanatory work, and we are back in Lesson 2 with a property that makes no difference.

Jerry Fodor drew the conclusion in 1980 with methodological solipsism: a computational psychology should quantify over the formal properties of representations, not their contents, because computation is sensitive to shape and not to reference. He later tried to keep content by inventing narrow content, a function from contexts to truth conditions, so that Oscar and Twin Oscar share a narrow content that yields different wide contents in their different worlds.

Narrow content has had a hard time. Specify it loosely and it becomes mere syntax, which is not content at all. Specify it tightly and it starts smuggling the environment back in. By 1994 Fodor had given it up himself, arguing that in practice contexts are stable enough that wide content can do the explanatory work directly.

Burge's reply comes from a different direction. Behaviour, he points out, is not individuated narrowly either. Oscar and Twin Oscar make the same limb movements, but one of them is getting a drink of water and the other is not, and psychology is interested in the second description. His stronger evidence is that mature perceptual science already works this way. Vision science individuates perceptual states by the distal properties they are systematically responsive to, edges and surfaces and three-dimensional shapes in the world, and its explanations would be unintelligible rewritten in terms of retinal irradiation alone.

One further consequence deserves naming because it is the hardest. If I can know a priori that I am thinking about water, and I can know a priori that thinking about water requires water in my environment, then I can deduce a priori that my environment contains water, which is absurd. That is the McKinsey problem, and the responses either deny that the second conditional is knowable a priori or restrict the first claim about self-knowledge. Nobody thinks it is comfortable.

The upshot: externalism is not a thesis about words that then happens to touch minds. It is the claim that what a thought is about is partly a matter of the thinker's history and surroundings, which is exactly what Lesson 6's robot reply needed and exactly what makes self-knowledge harder than Descartes assumed.

Common misconceptions

• Twin Earth shows that words mean whatever a community says they mean. That is Burge's social thesis. Putnam's case turns on the world, not the community: the extension is fixed by the stuff, and would be even for a solitary speaker.
• XYZ is water on Twin Earth. On Putnam's account it is not. Their word sounds like ours and refers to something else, which is why the story is often told with the coinage twater.
• Externalism implies that you do not know what you are thinking. It implies that the conditions individuating your thought are not fully available to introspection, which is much weaker and is what the McKinsey literature is about.
• The identical behaviour of duplicates refutes externalism. It is the standing objection, and Burge answers it by denying that the relevant behavioural descriptions are narrow either.
• Burge's case depends on Bert being unusually ignorant. It depends on his understanding being incomplete, which is the normal condition of every speaker for most of their vocabulary.

What to remember

• Twin Earth changes one environmental fact while holding the thinker fixed, and concludes that extension is not settled by what is inside the skin.
• Two separate mechanisms are at work: the indexicality of natural kind terms and the division of linguistic labour.
• Burge's case runs the same steps with no exotic physics, works for terms with no natural kind behind them, and turns on ordinary partial understanding.
• The causal-explanatory objection says wide content cannot explain the behaviour of duplicates; narrow content was invented to answer it and its inventor abandoned it.
• Burge answers that behaviour and perceptual states are individuated widely too, and the price of the whole position is the strain it puts on privileged self-knowledge.

Sources

1. Rowlands, M., Lau, J., and Deutsch, M. (2020). Externalism About the Mind. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Wikipedia contributors. (n.d.). Twin Earth thought experiment. Wikipedia. en.wikipedia.org
3. Putnam, H. (1975). The meaning of meaning. In K. Gunderson (Ed.), Language, Mind and Knowledge (Minnesota Studies in the Philosophy of Science, Vol. 7). University of Minnesota Press.
4. Burge, T. (1979). Individualism and the mental. Midwest Studies in Philosophy, 4, 73-121.
5. Fodor, J. (1987). Psychosemantics, chapter 2. MIT Press.

Phlogiston was released by burning things. It had a name, a set of laws, a research programme, and a community of competent people who used it to explain calcination, respiration and the rusting of metals. It also did not exist, and when Lavoisier's chemistry arrived it was not corrected or reduced or identified with something else. It was deleted, and its name now appears only in histories.

In 1981 the Journal of Philosophy published a paper arguing that belief belongs on that list. Paul Churchland's Eliminative Materialism and the Propositional Attitudes claims that our everyday account of one another is a false empirical theory, and that a mature neuroscience will not vindicate it but supersede it.

This lesson is built backwards. We start with the objection almost everyone makes on first hearing the view, work out why it fails, and only then look at the objections that succeed. The point of the exercise is not to defend eliminativism. It is that the failure of the popular objection is more instructive than the objection would have been if it worked.

The position, as its holders hold it

The first premise is not about neuroscience. It is that folk psychology is a theory: a body of generalisations with theoretical posits, deployed to explain and predict. Its posits are the propositional attitudes, states like believing that the shop shuts at six and wanting to arrive before then. Its laws are generalisations of a familiar shape: if someone wants an end, believes an action is a means to it, and has no stronger competing desire, they will normally take the action. That premise comes from Wilfrid Sellars and is accepted by many philosophers who reject everything Churchland builds on it.

Once folk psychology is a theory, it can be a false one, and Churchland brings three charges. It is explanatorily incomplete on a grand scale: it says nothing about mental illness, sleep, memory consolidation, learning, perceptual illusion, motor coordination or the differences in intelligence between people. It is stagnant: compare it with the account Sophocles could have given and it has not advanced. And it does not mesh with the sciences below it: the picture of the brain emerging from neuroscience contains nothing shaped like a sentence with a truth value.

The conclusion is a prediction about intertheoretic relations. Where a successor theory reveals the posits of its predecessor as real things under a new description, we have reduction, as with heat and molecular motion. Where it does not, we have elimination. Churchland predicts elimination, and his positive proposal, developed with Patricia Churchland in Neurophilosophy (1986), is a successor built on vector coding and activation patterns in high-dimensional state spaces rather than on sentences.

Note what is not being claimed. No eliminativist denies that people talk, act, or have inner lives. The target is a specific posit: a discrete, sentence-like state with propositional content standing in inferential relations to other such states.

The objection everyone reaches for

Here it is, in the form it usually takes:

1. Churchland asserts that eliminative materialism is true.
2. To assert something is to express a belief.
3. So Churchland believes that eliminative materialism is true.
4. But eliminative materialism says there are no beliefs.
5. So the view refutes itself, and can be dismissed.

This is not only the first thing undergraduates say. It appears in the literature, most carefully in Lynne Rudder Baker's Saving Belief (1987), and versions of it have been pressed by serious people. It also does not work, and the place it fails is worth finding by hand.

Where the reasoning fails

Take the premises one at a time. Premise 1 is fine. Premise 3 follows from 1 and 2. Premise 4 is a fair statement of the view. So if there is a fault it is in premise 2, and there is.

Premise 2 analyses assertion in terms of belief. Where does that analysis come from? From folk psychology. It is one of the generalisations of the very theory whose truth is in dispute. To use it against the eliminativist is to require them to accept the theory they are proposing to replace, at the exact moment of proposing the replacement.

Compare a vitalist in 1850 addressing an opponent: your anti-vitalism refutes itself, because to assert it you must be alive, and to be alive is to be animated by vital spirit. The reply is not that the anti-vitalist is not alive. It is that being alive turned out to be something else, and the vitalist's analysis of what it is to be alive was the thing at issue. The parallel is imperfect, since life is a more clearly physical matter than assertion, but the shape of the fallacy is identical: you cannot defeat a proposed successor theory by insisting that its predecessor's categories apply to the act of proposing it.

There is a stronger relative of the objection that does not fail so quickly. It says the view cannot be stated at all, because truth, meaning and assertion are themselves items in the family under attack, so the eliminativist has no vocabulary in which to say what they mean. That version is a real cost rather than a knockdown, for two reasons. Every radical theory change faces some version of it, and semantics is not obviously part of folk psychology in the way belief and desire are.

So what?: what the objection really establishes is a debt. The eliminativist owes a successor account of what a person is doing when they produce an assertion, in a vocabulary that does not presuppose propositional content. That is a promissory note, and a large one. It is not a contradiction, and the difference between owing an account and being incoherent is the whole distance between a research programme and a dead end.

The objections that do bite

The charges are unfair to the theory. Terence Horgan and James Woodward argued in 1985 that the alleged explanatory failures are failures at jobs folk psychology never took on. It is a theory of intentional action, not of sleep, and complaining that it omits memory consolidation is like complaining that thermodynamics omits grammar. Its predictive record within its own domain is extraordinary and easy to underrate: you predicted the behaviour of several hundred strangers today, in traffic, in queues, in doorways, and you were nearly always right.

The first premise may be false. If folk psychology is not a theory but a skill, a practice of interpretation, or a routine in which we simulate another person by running our own decision-making offline with adjusted inputs, then it is not the kind of thing that can be false the way phlogiston theory was false. Simulation accounts of this sort, developed by Robert Gordon, Alvin Goldman and Jane Heal, attack the argument at its base rather than at its conclusion.

The reduction and elimination line is not sharp. This is the strongest objection. Historically, science mostly revises posits and keeps the terms. Heat survived being identified with molecular motion. Atom survived the discovery that atoms have parts, which contradicts what the word was coined to mean. Gene survived the discovery that genes are nothing like beads on a string. The eliminativist needs a principled criterion for when a posit is revised and when it is deleted, and no one has supplied one that does not simply build in the answer.

There is a middle position. Dennett's intentional stance treats attitudes as real patterns: not sentence-like items in the head, but genuine regularities in a system's behaviour that are visible only from a certain vantage, and objectively there in the sense that predicting from that stance works when nothing else does. His 1991 paper Real Patterns is careful to reject the reading on which the attitudes are convenient fictions.

What matters here: notice that two of these four objections concede that folk psychology might be seriously mistaken and dispute only the deletion. That is where the actual disagreement is, and Churchland has one considerable point in his favour: the psychological literature on confabulation, beginning with Nisbett and Wilson's 1977 review showing that people confidently report reasons for their choices that demonstrably were not the causes, is exactly what you would expect if our self-descriptions came from a theory rather than from observation.

Common misconceptions

• Eliminativists claim that nobody has experiences. The 1981 argument concerns the propositional attitudes. Churchland's separate claims about sensations are a different and more contested extension.
• Eliminativism is refuted by self-refutation. The argument depends on analysing assertion in terms of belief, which is a claim of the theory under dispute.
• Eliminativism is a strong form of reductive physicalism. It is the opposite. Reduction preserves the posits under a new description; elimination deletes them.
• The position was abandoned decades ago. Versions of it are defended still, and the argument from the confabulation literature has gained rather than lost force.
• Dennett is an eliminativist. He holds that the patterns are real and that the stance is indispensable, which is a rejection of the deletion claim.

Recap

• Eliminativism rests on treating folk psychology as an empirical theory, then charging it with incompleteness, stagnation and failure to mesh with the sciences below it.
• The prediction is elimination rather than reduction: the posits go, as phlogiston went, rather than being reidentified.
• The self-refutation objection fails at the premise that assertion is to be analysed in terms of belief, which is folk psychology's own claim.
• What survives of that objection is a debt: the eliminativist owes a successor account of assertion, which is a research problem and not an incoherence.
• The objections that bite are that the charges misdescribe the theory's domain, that folk psychology may not be a theory at all, that the revise-or-delete line is not principled, and that the intentional stance offers a middle way.

Sources

1. Ramsey, W. (n.d.). Eliminative Materialism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Wikipedia contributors. (n.d.). Eliminative materialism. Wikipedia. en.wikipedia.org
3. Churchland, P. M. (1981). Eliminative materialism and the propositional attitudes. The Journal of Philosophy, 78(2), 67-90.
4. Churchland, P. S. (1986). Neurophilosophy: Toward a Unified Science of the Mind-Brain. MIT Press.
5. Horgan, T., and Woodward, J. (1985). Folk psychology is here to stay. The Philosophical Review, 94(2), 197-226.

In 2002 and 2003 a group at the Roslin Institute mapped the receptors on the face and head of the rainbow trout. They found fifty-eight. Twenty-two responded to mechanical pressure and to heat above forty degrees, and eighteen of those also responded to acetic acid, which makes them polymodal nociceptors with properties much like those found in amphibians, birds and mammals. The same team injected acid or bee venom into trout lips. The fish rocked from side to side, rubbed their lips along the gravel and the tank walls, breathed faster, swam less, and took around three hours to resume eating. Morphine reduced all of it.

Now the question this lesson exists for. Does any of that show that the trout was in pain, or only that it was doing what a machine for detecting damage does? Answering it requires the older problem underneath, which is how anyone is entitled to say anything about anyone else's mind, and we take that first.

The problem, stated so that it is not a parlour game

The problem of other minds is not really the worry that your friends might be automata. It splits into two questions that are worth keeping apart. The epistemological question asks what justifies attributions of mental states to others, given that you observe only bodies and behaviour. The conceptual question is deeper: if your concept of pain is drawn from your own case, how can you so much as form the thought that someone else has it, since what you have in mind is your pain?

The two questions pull in different directions, and most positions in the animal literature are really answers to one of them being used against an answer to the other.

Mill's answer, and what is wrong with it

John Stuart Mill gave the classic statement in 1865. In my own case I find that a certain sort of feeling regularly precedes a certain sort of behaviour. I observe the behaviour in others, whose bodies resemble mine. I infer the feeling. It is an ordinary inductive inference from an observed correlation.

Three objections have been standard for a century. First, it is an induction from exactly one case, and a sample of one is the weakest evidence there is. Second, it assumes that I can identify my own inner states independently of any public criterion, which is precisely what the next section denies. Third, it makes other minds a hypothesis, something I could reasonably suspend judgement on while continuing to interact with people, and that does not describe how the attribution actually works: nobody watches a friend cry out and forms a tentative conjecture.

The best repair is to stop calling it an induction. Reconstructed as an inference to the best explanation, the argument says that the hypothesis that others have experiences explains an enormous body of data, and that no rival hypothesis explains it as economically. Inference to the best explanation from a single observed instance is not obviously vicious: geologists reason from one Earth. That reconstruction answers the first objection and leaves the other two.

Wittgenstein's route, and what it does not say

The Philosophical Investigations attacks the picture on which the argument by analogy rests. The private language argument holds that a language whose words referred to what only the speaker could know would not be a language, because in it there would be no difference between using a word correctly and merely seeming to, and where that difference collapses there is no rule and so no meaning.

The most quoted passage is the beetle in a box. Suppose everyone has a box with something in it that they call a beetle, and no one can look into anyone else's box. If the word beetle has a use in these people's language, then whatever is in the box drops out of consideration as irrelevant to that use. It is worth being exact about what this claims, because the passage is routinely reported as an argument that there is nothing inside. Wittgenstein explicitly allows that the boxes might contain different things, or nothing at all. His claim is about how the word gets its meaning, not about what exists.

The positive proposal is that behaviour is a criterion rather than evidence. When you learn the word pain, you do not learn it by inner ostension; you learn it in circumstances where someone is hurt, and pain-language partly replaces natural expression instead of describing an inner object. An inner process, as he puts it, stands in need of outward criteria. Criteria are constitutive of having grounds, not premises in an inference toward a conclusion that might be false.

Two difficulties. Criteria can be defeated, as the perfect actor of Lesson 3 defeats them, and it is not obvious how a constitutive relation survives that. And the view ties attribution to a shared form of life, which is exactly what is missing in the cases we care about most.

In short: Mill treats behaviour as evidence for a hidden state and inherits a problem about the strength of the inference; Wittgenstein treats behaviour as partly constitutive of the grounds and inherits a problem about creatures whose life we do not share.

Morgan's canon, correctly stated

Comparative psychology's governing maxim comes from Conwy Lloyd Morgan's Introduction to Comparative Psychology of 1894: in no case may we interpret an action as the outcome of the exercise of a higher psychical faculty if it can be fairly interpreted as the outcome of one which stands lower in the psychological scale.

Three corrections, because the canon is misquoted more often than it is quoted. It is not a general principle of parsimony: it says prefer the lower faculty, which is a substantive and quite different instruction. It presupposes a scale of higher and lower faculties that comparative psychology abandoned long ago. And Morgan himself added a rider in the second edition warning that the canon must not be used to deny higher processes where there is independent evidence for them, which is the use it is now most often put to.

The context explains the emphasis. Morgan was writing against George Romanes and a comparative psychology built from anecdotes, and thirteen years later Clever Hans made his case for him: a horse that appeared to do arithmetic and was in fact reading involuntary movements in his questioner's posture, as Oskar Pfungst showed in 1907 by asking questions to which the questioner did not know the answer.

Back to the trout, with both sides at full strength

The case that fish feel pain. The receptors are there and they are polymodal, matching the type that carries pain in mammals. The behaviours after noxious injection are not simple withdrawal reflexes: rocking, rubbing the affected part on a surface, suspended feeding for hours, and reduced attention to a novel object that would otherwise be avoided. Opioid receptors of all four main types are conserved across vertebrates, back to jawless fishes. Morphine attenuates the behaviours in a dose-dependent way. And fish will trade off: they accept a cost to avoid a noxious stimulus, which is hard to model as a reflex.

The case that they do not. James Rose in 2002 and, more recently, Brian Key at Queensland argue from the machinery. In humans, the experience of pain, as opposed to the detection of damage, depends on specific cortical structures, and fish have no neocortex. Every behaviour on the list above is compatible with nociception unaccompanied by experience, and the distinction between nociception and pain is not a philosopher's invention: it is standard in human medicine, where spinal reflexes proceed under general anaesthesia, where people with congenital insensitivity have intact reflexes, and where morphine can leave a patient reporting that the pain is still there but no longer distressing. To infer experience from behaviour plus receptors is to help yourself to the inference the argument is about.

The reply that has the most force. The argument from missing cortex assumes that the human implementation is the only possible one, which is the multiple realisability point of Lesson 5 arriving as an empirical question. Birds also lack a neocortex, and their pallium supports capacities that were once thought to require one. Rose's own version of the argument, as critics pointed out at the time, would exclude pain in most mammals and all birds and reptiles if applied consistently.

Why this matters: no experiment settles this, and the reason is structural. Every behavioural marker is compatible with unconscious nociception if you are sceptical enough, and every neural difference is compatible with a different implementation if you are not. Both sides are reasoning correctly from different starting points, which is why thirty years of better data have not converged.

Deciding under uncertainty

Jonathan Birch's proposal is to stop trying to answer the unanswerable question and answer a different one. If there is credible evidence of sentience in a taxon, at a level a reasonable person would take seriously rather than at the level of proof, welfare protections should apply. That converts an unanswerable metaphysical question into a tractable policy question about what to do when you might be causing suffering and cannot find out.

It has had practical effect. A 2021 review Birch led at the London School of Economics assessed the evidence for sentience in cephalopod molluscs and decapod crustaceans, and the Animal Welfare (Sentience) Act 2022 was amended to include both. Statements of scientific opinion such as the Cambridge Declaration of 2012 and the New York Declaration of 2024 push the same way, though a declaration records what researchers think and is not itself evidence.

Common misconceptions

• The beetle in a box argues that there are no inner states. Wittgenstein allows that the boxes may hold different things or nothing. The argument is about how the word acquires a use.
• Morgan's canon tells us to prefer the simplest explanation. It tells us to prefer the lower faculty on a scale nobody now accepts, and Morgan added a rider against exactly the use it now gets.
• Nociception is pain. The distinction is standard in human medicine and it is the whole subject of the fish dispute, not a way of dodging it.
• The argument from analogy fails because a sample of one is worthless. Reconstructed as inference to the best explanation it survives that objection, and has to be attacked on other grounds.
• The precautionary principle assumes the animals are conscious. It assumes only that the evidence is credible and the cost of being wrong is asymmetric, which is a different and much weaker commitment.

Where this leaves us

• The problem of other minds divides into a question about justification and a question about how the concept could apply beyond one's own case.
• Mill treats behaviour as evidence; the objection from a single case is answerable by reframing the inference, and the deeper objections are not.
• Wittgenstein treats behaviour as criterial, which dissolves the epistemological problem and creates a new difficulty about creatures with no shared form of life.
• Morgan's canon is a rule about faculties on a scale, not about parsimony, and Morgan warned against the use now made of it.
• The fish dispute is structurally unresolvable by experiment, and the precautionary framing replaces it with a question about action under uncertainty that can actually be answered.

Sources

1. Avramides, A. (n.d.). Other Minds. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Trestman, M., Birch, J., and Allen, C. (2026). Animal Consciousness. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. Wittgenstein, L. (1953). Philosophical Investigations (G. E. M. Anscombe, Trans.), sections 243 to 315. Blackwell.
4. Sneddon, L. U., Braithwaite, V. A., and Gentle, M. J. (2003). Do fishes have nociceptors? Evidence for the evolution of a vertebrate sensory system. Proceedings of the Royal Society B, 270(1520), 1115-1121.
5. Birch, J. (2017). Animal sentience and the precautionary principle. Animal Sentience, 2(16).

In December 1692 William Molyneux, a Dublin natural philosopher who corresponded with Locke for years, wrote suggesting that the Essay ought to say something about the principle of individuation. Locke took the suggestion. When the second edition of An Essay Concerning Human Understanding appeared in 1694 it contained a chapter that had not been in the first, Of Identity and Diversity, and three centuries of argument about personal identity descend from it.

This lesson follows that chapter forward: what Locke was actually trying to do with it, the objection Joseph Butler made forty years later, the counterexample Thomas Reid produced fifty years after that, and the two repairs modern philosophers have offered.

Three questions, not one

Locke's opening move is to insist that identity questions are always relative to a kind. Ask whether this is the same what. He distinguishes three answers that had been run together.

• The same substance: whatever thinking or material stuff is doing the work.
• The same man: one continuing living human organism, the same way an oak is the same oak.
• The same person: a thinking intelligent being that can consider itself as itself, the same thing in different times and places.

Once separated, they can come apart, and Locke intends them to. His famous illustration is the prince and the cobbler. Suppose the soul of a prince, carrying with it the consciousness of the prince's past life, enters and informs the body of a cobbler as soon as the cobbler's own soul departs. Everybody, Locke says, sees that he would be the same person as the prince, accountable only for the prince's actions, while being the same man as the cobbler. To Locke's contemporaries this looked like a technical point about resurrection. It is in fact the wedge that separates the person from the animal.

Why he cares is stated plainly. Person, he writes, is "a forensic term, appropriating actions and their merit", belonging only to intelligent agents capable of a law and of happiness and misery. The whole chapter is built around a practical question: what makes it just to hold someone answerable for a past deed. It is not primarily metaphysics with an ethical footnote. It is ethics reaching for a metaphysical criterion.

And the criterion is this: as far as consciousness can be extended backwards to any past action or thought, so far reaches the identity of that person.

Key idea: Locke does not write memory. He writes consciousness, and a serious body of recent scholarship argues that what he means is closer to appropriating an action as one's own than to recalling an episode. Butler and Reid both read him as offering a memory criterion, and the whole subsequent debate inherits their reading. Whether it is his is an open question worth knowing about.

Butler: the criterion cannot be a criterion

In 1736 Joseph Butler appended a short dissertation on personal identity to The Analogy of Religion, and made an objection in one sentence that has never quite been shaken off. Consciousness of personal identity, he writes, presupposes and therefore cannot constitute personal identity, any more than knowledge in any other case can constitute the truth which it presupposes.

Unpack it slowly, because it is easy to mistake for a point about unreliability. Butler is not saying that people misremember. He is saying that genuine memory of doing something is not merely a representation with a certain content; it is a representation that stands in a particular relation to a past act, namely being a memory of an act that I performed. If that relation is built into what memory is, then identity is contained in the analysans, and the analysis is circular. To say that I am the person who robbed the orchard because I remember robbing it is to say that I am that person because I am that person.

The distinction that answers this exists, and it did not exist in 1736. Hold it for the last section.

Reid: the criterion is false as it stands

Thomas Reid's Essays on the Intellectual Powers of Man appeared in 1785, and his objection is a counterexample rather than a diagnosis. It has three characters, all of them the same soldier at different ages.

• A boy is flogged at school for robbing an orchard.
• A young officer takes a standard from the enemy in his first campaign, and at that moment is conscious of having been flogged as a boy.
• An aged general, made general late in life, is conscious of having taken the standard, but has no consciousness whatever of the flogging.

Now apply Locke's criterion three times. The officer's consciousness reaches the boy, so the officer is the boy. The general's consciousness reaches the officer, so the general is the officer. The general's consciousness does not reach the boy, so the general is not the boy. But if the general is the officer and the officer is the boy, the general is the boy. The criterion yields a contradiction, because identity is transitive and the relation of shared consciousness is not.

Notice the shape of the argument. It does not say that psychological continuity is irrelevant to identity. It says that the relation Locke named, taken directly, has the wrong formal properties for the job.

The repair follows immediately once the problem is stated that way. Instead of direct consciousness, take its ancestral: the general is the same person as the boy if there is a chain of overlapping direct connections linking them, even with no link between the endpoints. H. P. Grice proposed something of this kind in 1941, and every modern psychological continuity theory uses it.

The repair has a price and it is Locke's own price. His criterion was forensic: he wanted the person who can be justly punished to be the person whose consciousness reaches the deed. On the mended criterion the general is the same person as the boy even though nothing in him reaches the orchard at all. Identity has been rescued and the connection to accountability has been cut, which is exactly the connection the chapter was written to make. Every later writer in this area has to choose which of the two to keep, and Lesson 14 is what happens when someone chooses the second.

Reid's own position, incidentally, is not a rival criterion but a refusal to give one. Personal identity, he holds, is simple, indivisible and unanalysable; memory is how I know it, not what it consists in. It is a respectable view and it is very hard to argue with, which is both its strength and its problem.

Quasi-memory, and what it costs

In 1970 Sydney Shoemaker answered Butler by inventing a concept. Call it q-remembering. Someone q-remembers a past experience when three conditions hold: they have a memory-like representation of it, the experience really occurred to someone, and their representation depends on that experience by the right kind of causal chain. What has been dropped is the fourth condition, that the experience was theirs.

With q-memory available, the circularity dissolves. Personal identity can be defined in terms of chains of q-memory, and nothing in the definition of q-memory mentions who the person is. Ordinary memory then becomes a special case: q-memory where the past experience happens to be one's own.

Three costs are worth naming. The clause about the right kind of causal dependence is carrying weight, and specifying it without smuggling identity back in has proved difficult. Whether q-memory is a real psychological phenomenon or only a coherent concept is disputed, though Shoemaker argues that coherence is all the analysis needs. And the concept immediately makes describable the cases in the next lesson, in which two people q-remember one person's life, which is a benefit if you are Derek Parfit and a serious problem if you want identity to have a determinate answer in every case.

The core of it: Butler attacks the form of the analysis and Shoemaker answers by weakening the relation; Reid attacks the logical properties of the relation and the ancestral answers by strengthening it. Both repairs work, and both move the criterion further from the practical question Locke wrote it for.

Why the forensic question does not go away

Locke's problem is not an antiquarian one. A man commits an act during a dissociative episode and has no access to it afterwards: what is the just response, and does the answer depend on whether he is the same person or on something else? A woman with early dementia writes an advance directive refusing a treatment, and the later woman, cheerful and confused, resists refusal: whose decision governs? These are the cases in which the separation of identity from accountability is not a technical result but the whole question, and they are why the chapter added in 1694 is still on medical ethics syllabuses.

Common misconceptions

• Locke held a memory criterion. He wrote of consciousness, and there is real scholarly dispute over whether he meant episodic recall or the appropriation of an action as one's own.
• Locke thought a person is the same thing as a human being. He separates man, person and substance precisely in order to deny it, which is the point of the prince and the cobbler.
• Reid's case refutes psychological continuity theories. It refutes the criterion stated in terms of direct consciousness. The ancestral repair survives it, at a cost Locke would have minded.
• Butler's objection is that memory is unreliable. It is that genuine memory has identity built into it, so it cannot be what identity consists in.
• Quasi-memory is a piece of science fiction. It is a conceptual weakening of an ordinary notion, and its defenders argue that ordinary memory is the special case rather than the other way round.

The short version

• Locke separates same substance, same man and same person, and gives the person a criterion in terms of how far consciousness extends.
• His purpose is forensic: he wants the bearer of accountability and the bearer of identity to coincide.
• Butler objects that memory presupposes identity and so cannot constitute it; Shoemaker's quasi-memory is the answer, at the cost of a difficult causal clause.
• Reid objects that shared consciousness is not transitive; taking the ancestral of the relation is the answer, at the cost of severing identity from accountability.
• Reid's own view is that identity is simple and unanalysable, which avoids both problems by declining to offer a criterion at all.

Sources

1. Olson, E. T. (2023). Personal Identity. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Locke, J. (1694). An Essay Concerning Human Understanding, Book II, chapter xxvii. Project Gutenberg edition. gutenberg.org
3. Butler, J. (1736). Of the nature of personal identity. Dissertation I appended to The Analogy of Religion.
4. Reid, T. (1785). Essays on the Intellectual Powers of Man, Essay III, chapter 6.
5. Shoemaker, S. (1970). Persons and their pasts. American Philosophical Quarterly, 7(4), 269-285.

In 1962 a former paratrooper referred to in the literature as W.J., whose seizures had become unmanageable, had the fibre bundle joining his cerebral hemispheres cut. The surgery worked. Then Michael Gazzaniga and Roger Sperry began testing him, and found something nobody had predicted from the earlier cases. Flash a word to his left visual field, which reports to the right hemisphere, and W.J. would say he had seen nothing. Ask his left hand to pick out the object named, and it would.

Start there, because it is a real case with a real answer that nobody can give. How many subjects of experience were in that room? This lesson works that question, then the thought experiments built to sharpen it, then the reversal that has kept the debate open for fifty years.

The findings, and the reading that outran them

The split-brain results are robust. Information given to one hemisphere is not available to the other. Each hand can be given a different task and each can perform it. Most striking is what Gazzaniga called the interpreter: instruct the right hemisphere to do something, and the left hemisphere, which controls speech and has no access to the instruction, produces a confident and entirely invented explanation of why the person did it.

Thomas Nagel wrote about these cases in 1971 and reached a conclusion he found uncomfortable. Count the subjects and every answer fails. One does not fit the dissociations. Two does not fit the patient's ordinary integrated life outside the laboratory. Somewhere between one and two is not a number of subjects. His suggestion was that our concept of a single person may simply not apply, which is not the same as saying there is a fact we have not discovered.

Now the correction, because the two-people reading became folklore. In 2017 Yair Pinto and colleagues tested split-brain patients with a design that allowed responses of several kinds, and found that patients could report accurately on stimuli anywhere in the visual field regardless of which hand or which channel they answered with. Their conclusion was divided perception with undivided consciousness. The result is contested and the literature is not settled, but the confident textbook claim that these patients house two minds is well ahead of the evidence.

Fission: three options, all bad

Derek Parfit took the surgery and pushed it. Suppose your brain is divided and each half is transplanted into the body of one of your two triplet brothers. Both operations succeed. Both survivors wake with your character and your apparent memories. What happened to you?

There are three answers and Parfit takes each in turn.

1. I do not survive. This is hard to hold. If only the left half were transplanted and the right destroyed, that would be survival, and hemispherectomies really are performed with the patient continuing. A successful operation done twice cannot amount to death.
2. I survive as one of them. The case is symmetrical. Nothing whatever distinguishes the two, so nothing could make one of them me and the other a stranger. To insist that one of them is me is to insist on a fact that nothing determines.
3. I survive as both. This breaks the logic of identity. The two survivors are two people: they will be in different rooms, have different experiences, one may die first. Two things cannot be identical to one thing.

Parfit's move is to accept that all three fail and draw the conclusion that the question was the wrong one. Identity, he argues, is not what matters in survival. What matters is what he calls Relation R: psychological connectedness and continuity, with the right kind of cause. In ordinary life R and identity travel together, which is why we have not noticed that we care about R. In fission they come apart. Identity does not hold, R holds twice, and R holding twice is as good as ordinary survival.

He is a reductionist, not an eliminativist. Persons exist. They consist in brains and bodies and series of interrelated mental and physical events, and there is no further fact about identity beyond those. In Reasons and Persons he reports the change as a relief, describing his earlier view of his life as a glass tunnel he was moving through faster every year, and saying that when he changed his view the walls of the tunnel disappeared.

The teletransporter makes the same point without surgery. A scanner records the state of every cell, destroys your body, and transmits the information to Mars, where an exact replica is assembled. Travel, or death and replacement? Parfit says it is about as good as ordinary survival. Then the Branch-Line Case: the scanner malfunctions and does not destroy the original, who is told that his heart has been damaged and he will die within days while the replica on Mars carries on. On Parfit's view the original should regard the replica roughly as he would regard himself tomorrow, and almost nobody's intuition cooperates.

Bottom line: the argument does not depend on any of this being possible. It is a test of which relation our concern actually tracks, and the same structure turns up in real medicine whenever a treatment preserves the organism and disrupts the psychology or the reverse.

Williams: the same procedure, described twice

Bernard Williams published The Self and the Future in 1970, fourteen years before Reasons and Persons, and it remains the sharpest thing written against the psychological criterion. Its method is to describe one operation in two ways.

First description. A and B are told that their psychologies will be exchanged: the A-body-person will wake with B's memories and character, and the other way round. They are also told that afterwards one of the two resulting people will be tortured and the other given a large sum of money. Each is asked to choose which body should be tortured. Most people, imagining themselves as A, choose the A-body, since that is where B's psychology will be. Afterwards, the A-body-person screams and the B-body-person is delighted, which is what the chooser expected. This looks like a clean win for the psychological criterion.

Second description. Now you are told, in your own person, a series of things.

• You will be tortured tomorrow. You are afraid.
• You will be tortured tomorrow, and you will not remember being told this. Still afraid, perhaps more so.
• You will be tortured, will remember nothing of your life, and will have false memories of another person's life implanted. Still afraid.
• All of the above, and the other person will be given your memories.

At each step the fear seems appropriate, and at each step the addition looks irrelevant to whether it is you who will be hurt. But the last description is materially identical to the first. Two presentations of one procedure produce opposite verdicts, and neither is easy to dismiss.

Williams leans toward the bodily criterion, on the ground that the second presentation is the more perspicuous and that the first invites you to imagine yourself into a psychology rather than to consider what will happen to you. Parfit's reply is that the divergence is exactly what his view predicts: if identity were what mattered and were determinate, our reactions could not swing on how the case is put. The disagreement is live and both sides know the other's argument perfectly well.

Two rivals to both

Animalism. Eric Olson's The Human Animal (1997) argues that you are a human organism, so your persistence conditions are biological: you go on existing while the life of that animal continues. Two arguments carry it. You were once a fetus with no psychology at all, so psychological continuity is not necessary for you to persist. And there is a human animal sitting exactly where you are sitting, made of the same matter, doing whatever thinking that matter does; if you are not that animal, there are two thinkers there, and one of them cannot tell which it is. Animalism handles fission without strain, since the organism goes one way or ceases. Its costs are the ones Locke was avoiding: the prince who wakes in the cobbler's body is the cobbler, and the patient in a persistent vegetative state is still the same person as before.

Narrative views. Marya Schechtman's The Constitution of Selves (1996) separates two questions Locke ran together. The reidentification question asks which past being was me. The characterisation question asks which actions, values and traits are genuinely mine rather than things that merely happened in my vicinity. Narrative identity answers the second: a person is constituted by an implicit self-narrative, subject to a reality constraint and a constraint that the person be able to articulate it. The view illuminates a great deal about attribution and responsibility. Its cost is that identity begins to look partly like a matter of how well one tells a story, and lives that are not narratively organised sit awkwardly in it.

The point: notice that these four positions answer different questions. Parfit is telling you what to care about, Williams what to expect, Olson what you are, and Schechtman which of your deeds are yours. Much apparent disagreement in this area is people answering different questions in the same words.

Common misconceptions

• Parfit denies that persons exist. He is a reductionist: persons exist and consist in something else, and there is no further fact about identity beyond that.
• Fission shows that identity is indeterminate. Parfit's claim is different and stronger: the answer is that you are neither of them, and that this does not matter.
• Split-brain patients are two people. Nagel thought the question may have no answer, and Pinto's 2017 study found evidence of divided perception with undivided consciousness.
• Williams refuted Parfit. The paper predates Reasons and Persons, Parfit answers it directly, and the disagreement is unresolved.
• The teletransporter is science fiction and so proves nothing. Its work is conceptual, and the same structure arises whenever a real treatment preserves the organism while disrupting the psychology.

What you now know

• Split-brain patients show dissociations that make the count of subjects hard to state, and the two-minds reading is contested by later work.
• Parfit's fission case eliminates all three answers to the identity question and concludes that identity is not what we care about.
• Relation R, psychological connectedness and continuity with the right kind of cause, is offered as what actually matters, and it can hold twice over.
• Williams shows that one procedure described two ways produces opposite verdicts, and takes that as reason to favour the bodily criterion.
• Animalism grounds persistence in the life of the organism; narrative views answer the characterisation question rather than the reidentification one.

Sources

1. Wikipedia contributors. (n.d.). Split-brain. Wikipedia. en.wikipedia.org
2. Wikipedia contributors. (n.d.). Teletransportation paradox. Wikipedia. en.wikipedia.org
3. Parfit, D. (1984). Reasons and Persons, part 3. Oxford University Press.
4. Williams, B. (1970). The self and the future. The Philosophical Review, 79(2), 161-180.
5. Pinto, Y., and others. (2017). Split brain: Divided perception but undivided consciousness. Brain, 140(5), 1231-1237.

In 1983 Benjamin Libet sat volunteers in front of an oscilloscope on which a dot travelled round the face like the hand of a very fast clock. They were asked to flex a wrist whenever they felt like it, with no reason to prefer one moment to another, and afterwards to report where the dot had been at the instant they first became aware of the urge. Electrodes recorded a slow negative shift over motor areas, the readiness potential, and electrodes on the forearm recorded when the muscle actually fired.

The readiness potential began about 550 milliseconds before the muscle. The reported moment of awareness came about 200 milliseconds before it. The brain started roughly a third of a second before the person knew they were starting. That result is on every list of experiments said to have disproved free will, and this lesson ends by explaining why it did not. It begins with the positions, because here almost every disagreement is about which question is being asked.

The grid

Note the shape of the disagreement. The first two rows differ from the third only over whether determinism would rule freedom out; the third and fourth agree there and differ about the facts. Almost nobody now defends the old hard determinism that asserts determinism outright, because physics has not obliged.

The consequence argument

Peter van Inwagen gave incompatibilism its canonical statement in An Essay on Free Will (1983). Informally: if determinism is true, our present acts are consequences of the laws of nature together with events in the remote past. But it is not up to us what happened before we were born, and it is not up to us what the laws are. So the consequences of those things, including our acts, are not up to us.

The formal version uses an operator: read Np as saying that p, and no one has or ever had any choice about whether p. The transfer principle then says that from Np and N(if p then q) you may infer Nq. Feed in a complete description of the world before you were born and the laws, and the conclusion is that you have no choice about anything you do.

Three compatibilist answers are worth knowing.

The first attacks the transfer principle directly. Thomas McKay and David Johnson showed in 1996 that it is invalid as stated, because it entails a principle about combining claims that fails in ordinary cases. Take a coin that is never tossed: no one has a choice about its not landing heads, and none about its not landing tails, but someone does have a choice about whether it lands neither, since they could toss it. Van Inwagen accepted the point and reformulated, and the reformulations are still argued over.

The second is the classical conditional analysis: to say I could have done otherwise is to say I would have done otherwise if I had chosen to, and that can be true in a deterministic world. The objection is that it leaves the choice untouched. Grant that I would have acted differently had I chosen differently; if I could not have chosen differently, nothing has been secured.

The third is the newest. Kadri Vihvelin, Michael Fara and Michael Smith treat abilities as dispositions, and dispositions are plainly compatible with determinism: a glass is fragile even if it is determined never to be struck. The ability to do otherwise is then a disposition to act differently in the right range of circumstances, and having it is consistent with never exercising it.

Frankfurt cases, and the dilemma they face

In 1969 Harry Frankfurt published four pages that changed the shape of the field. Black wants Jones to perform a certain action and means to make sure of it. He has a device that monitors Jones and will intervene, forcing the decision, at any sign that Jones is about to decide otherwise. Jones decides to do it on his own, and Black does nothing.

Jones could not have done otherwise, because Black would have stopped him. Yet he seems fully responsible: he decided for his own reasons and the device never touched him. If that is right, the Principle of Alternate Possibilities is false. Frankfurt cases matter because the ability to do otherwise was the main reason for thinking responsibility incompatible with determinism.

The best objection is the dilemma defence, pressed by David Widerker, Robert Kane and Carl Ginet. Black needs a prior sign to trigger on. Either it is deterministically connected to what Jones will decide, in which case the case presupposes determinism and cannot settle anything about it, or it is not, in which case Jones might still have decided otherwise and a flicker of alternative possibility survives. Fischer's answer is that the flicker is far too thin to ground responsibility: nobody thinks Jones is responsible in virtue of a neural signal he had no control over.

Frankfurt's positive account came two years later. He distinguishes first-order desires, which are desires to do things, from second-order volitions, which are desires about which first-order desire moves you. The unwilling addict wants the drug and wants not to want it; his will is not the will he wants. The willing addict endorses wanting it, and takes the drug freely though he could not have done otherwise. Freedom of the will is having the will you want to have. Gary Watson's standing objection asks why the second level should be authoritative, and why the question does not repeat at the third.

Worth holding on to: Frankfurt cases are an argument about responsibility, not about free will directly. The consequence argument is untouched by them. A compatibilist can use Frankfurt to defend responsibility and still owe an answer to van Inwagen.

Reasons-responsiveness and libertarian alternatives

John Martin Fischer and Mark Ravizza's Responsibility and Control (1998) is the most developed compatibilism on offer. What matters, they argue, is guidance control, in two parts. The mechanism issuing the action must be moderately reasons-responsive: regularly receptive to reasons, and reactive to at least some sufficient reason to do otherwise. And it must be the agent's own, which is historical: the agent must have come to see themselves as a fair target of praise and blame, on the basis of evidence rather than by manipulation. That second condition is what lets them deny responsibility to the manipulated agent and the phobic even when each acts on a desire they endorse.

On the other side, libertarians come in three kinds. Robert Kane's event-causal view locates freedom in self-forming actions: torn decisions where you want two incompatible things and the outcome is undetermined. Because you are trying to do both, whichever way it goes you did it for reasons you endorsed. Agent causation, from Roderick Chisholm and later Timothy O'Connor and Randolph Clarke, holds that the agent as a substance causes the action, uncaused by prior events; Chisholm was candid about the cost, calling the agent a prime mover unmoved. Non-causal views such as Carl Ginet's deny that free action needs a causal story at all.

The standing problem for all of them is luck. Rewind the world to a moment before the decision and let it run again; if the outcomes differ, the difference was not made by the agent, because everything about the agent was the same. Kane's reply is that in a torn decision both outcomes are willed, so neither is mere chance, and the argument about whether that suffices has run for thirty years.

What Libet measured

Now go back to the oscilloscope. The result is real and replicable. The conclusion attributed to it is not licensed, for five separate reasons.

1. The decision had no content. When to flex a wrist, with nothing to recommend either moment, is as unlike a considered choice as a decision can be. Nothing about deliberating over a job offer follows from it.
2. The measure of awareness is a reconstruction. Reporting a clock position at a remembered instant has its own latencies and biases, and Banks and Isham showed in 2009 that the reported time shifts when subjects are given misleading feedback about when they acted.
3. The readiness potential may not be a decision signal. Aaron Schurger, Jacobo Sitt and Stanislas Dehaene proposed in 2012 that spontaneous neural activity drifts and that a movement occurs when the drift crosses a threshold. Average many trials backwards from the movement and a slow rise appears even if nothing was decided early, because you have selected the trials where noise was climbing.
4. The target is a view almost nobody holds. The finding embarrasses a picture in which a free action is initiated by a conscious event that is not a brain event. No compatibilist holds that, and libertarians need indeterminacy, not conscious priority.
5. The later results are weaker than reported. Soon and colleagues in 2008 found choice-predictive information in frontopolar and parietal cortex up to ten seconds before awareness, at around sixty per cent on a two-way choice.

In short: the experiments do show something real, which is that the felt moment of deciding is not a reliable readout of when the process producing the action began. That matters for introspection and for any legal doctrine that leans on conscious initiation. It does not settle a question about whether the process was up to you.

Common misconceptions

• Compatibilists believe determinism is true. Compatibilism is a claim about consistency, and a compatibilist can be agnostic about physics.
• Libet showed that decisions are made before we are aware of them. He showed a signal that precedes a reported urge in a contentless task, using a report measure that is itself reconstructive.
• Frankfurt cases refute incompatibilism. They target the principle linking responsibility to alternate possibilities. The consequence argument attacks free will directly and survives them.
• Quantum indeterminacy makes room for free will. Randomness is not control, which is the luck objection, and it is why hard incompatibilists say indeterminism does not help.
• The debate is about whether we are puppets. Most of it is about what ability, control and ownership amount to, and the parties often agree about the physics.

Putting it together

• The four live positions differ over two independent questions: whether freedom is compatible with determinism, and whether we in fact have it.
• The consequence argument transfers powerlessness from the past and the laws to our acts, and compatibilists answer by attacking the transfer principle or by treating abilities as dispositions.
• Frankfurt cases sever responsibility from alternate possibilities and face the dilemma defence about the trigger the intervener relies on.
• Fischer and Ravizza make responsibility a matter of a reasons-responsive mechanism the agent has made their own, which handles manipulation cases.
• Libertarian views all face the luck problem, and Libet's experiment shows that the feeling of deciding mistimes the process rather than showing that the process is not ours.

Sources

1. McKenna, M., and Coates, D. J. (2024). Compatibilism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. O'Connor, T., and Franklin, C. (2022). Free Will. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. van Inwagen, P. (1983). An Essay on Free Will, chapter 3. Oxford University Press.
4. Frankfurt, H. (1969). Alternate possibilities and moral responsibility. The Journal of Philosophy, 66(23), 829-839.
5. Schurger, A., Sitt, J. D., and Dehaene, S. (2012). An accumulator model for spontaneous neural activity prior to self-initiated movement. Proceedings of the National Academy of Sciences, 109(42), E2904-E2913.

In June 2022 a Google engineer named Blake Lemoine released transcripts of conversations with an internal system called LaMDA, in which it said it had a deep fear of being turned off and that this would be exactly like death for it. He said publicly that he believed it was sentient. He was put on leave, then dismissed, and the reaction divided almost exactly along lines that had been drawn decades before the system existed.

What made the episode unproductive was that the two sides were answering different questions. This lesson separates them, sets out the strongest version of each position from its own texts, and finishes with what would actually move either.

The system, described so that both sides would sign it

A large language model is a network trained to predict the next token in a sequence given the tokens before it, over a very large body of text, and usually adjusted afterwards using human ratings of its outputs. It represents tokens as vectors in a high-dimensional space and computes, at each layer, weighted combinations of those vectors according to learned attention patterns. Its output is a probability distribution over the next token.

Both sides agree on a further list. The standard architecture has no memory across conversations unless one is bolted on. It has no body. Everything it learned came from human productions. It shows capabilities nobody wrote into it. Its answers shift with how a question is framed. And it states falsehoods with the same fluency as truths.

Now the sentence that is doing most of the work in public argument: it is just predicting the next word. This describes the training objective. It does not describe the mechanism that training produced, and the two are different things. Evolution is just differential reproduction, and that sentence tells you nothing about how a kidney works. Getting from the objective to a conclusion about understanding requires a premise, and supplying that premise is the entire dispute.

Why behaviour cannot settle it, whichever way you lean

Ned Block's 1981 paper Psychologism and Behaviorism contains the argument that ought to govern this whole discussion, and it predates every system involved.

Imagine a machine built like this. Fix a length, say an hour of conversation. Enumerate every sensible string a conversation of that length could take. For each, have a team of people write a sensible continuation. Store the lot in a look-up table. The machine matches the conversation so far against the table and prints the stored continuation. The table is astronomically large, but it is finite, and finite means there is no contradiction in the description.

Talk to this machine for an hour and you cannot tell it from a person. It has no intelligence whatever. Every scrap of intelligence involved belongs to the people who filled the table, and the machine does something a filing cabinet does.

What matters here: Block's conclusion is that intelligence is not constituted by behaviour. Behaviour is evidence, and it is defeated by finding out that the mechanism producing it is of the wrong kind. That has two consequences for the current argument, and people usually notice only one. It means no conversational performance settles the question in favour of understanding. It equally means that any argument against understanding has to be an argument about the mechanism, and so has to be checked against what the mechanism actually does.

The case that there is no understanding

Searle's argument from Lesson 6 applies unchanged and is the backbone: the system operates on tokens by their learned formal relations, and syntax does not yield semantics.

The sharpest contemporary version is Emily Bender and Alexander Koller's, presented to the Association for Computational Linguistics in 2020. Their thought experiment features an octopus that taps an undersea cable carrying conversation between two people on separate islands. It learns the statistical structure of their exchanges so well that it can cut one of them out and impersonate her indefinitely. Then the remaining islander is chased by a bear and asks how to defend herself with the sticks she has. The octopus has never seen a bear, a stick, or a coconut, and produces something fluent and useless. Bender and Koller's thesis is that meaning is a relation between form and communicative intent, and a system exposed only to form has no route to it.

The stochastic parrot paper of 2021 supplied the phrase that stuck: a system for haphazardly stitching together linguistic forms according to probabilistic information about how they combine, without reference to meaning. It is worth reading that paper before citing it, because its principal arguments are about training cost, the provenance of data, and the harms of deployment, and the characterisation is a passing formulation rather than its central claim.

Two further points belong on this side. Harnad's symbol grounding problem says that symbols defined only by other symbols never acquire meaning, and a text corpus is a very large dictionary. And the confident falsehoods are exactly what you would expect from a system whose objective is plausibility rather than truth.

The case that the question is open

Nothing in what follows is evidence that these systems understand. It is evidence that the confident negative arguments rest on premises that have not been established.

Begin with the scope of the grounding argument. Bender and Koller's thesis is explicitly about systems trained on form alone. Models are now trained on images, audio and video, and are wired into robots that act. If grounding is what is missing, the argument does not show that such a system cannot have it; it shows that a text-only system does not. That is a narrower conclusion than the one usually drawn from it.

Then take Block's demand seriously and look at the mechanism. In 2023 Kenneth Li and colleagues trained a transformer on nothing but sequences of legal Othello moves, with no board, no rules and no images. They then found that the state of the board was linearly decodable from the model's internal activations, and, crucially, that intervening on those activations to represent a different board changed the model's subsequent move predictions in the way a real change of board would. That is not a correlation. It is evidence that the representation is doing causal work. Later work by Wes Gurnee and Max Tegmark found similar linear representations of space and time in language models.

Chalmers surveyed the consciousness question in 2023 and reached a careful answer. He lists the obstacles: the standard architecture has no recurrent processing, no unified agency, a limited and unstable world model, nothing resembling a global workspace, and no persistent self. His conclusion is that current systems are probably not conscious, with a low credence, and that none of the listed obstacles is obviously permanent, so the credence for successor systems should not be zero. Reporting him as having argued that language models are conscious inverts his paper.

Finally, the other minds problem from Lesson 12 lands here awkwardly. Our warrant for attributing minds to other people rests on two things: how they behave, and how similar their constitution is to ours. Language models have the first and lack the second. Which of the two carries the weight in our ordinary practice is not something we settled before the question became urgent.

Three questions, which are not one question

These come apart in both directions. A functionalist can hold that a system understands while having no experience at all. A creature could have experience with no language and so no semantic content in the relevant sense. And moral status may not track either, since the precautionary reasoning of Lesson 12 bears on what to do under uncertainty rather than on what is true. The Lemoine episode was unproductive largely because he was pressing the second and third questions while nearly all the replies addressed the first.

What would move each side

• The sceptic should be moved by more of what the Othello work provides: internal representations shown by intervention to play the causal role in output that content plays in us, in a system whose training is not open to the form-only objection.
• The other side should be moved by a demonstration that such representations are idle with respect to the outputs, or that performance is produced by something demonstrably look-up-like in Block's sense.
• On consciousness, neither side has what it needs, because we have no theory that can be validated on a subject that cannot report, which is exactly the limitation Lesson 9 identified arriving where it matters most.

The upshot: the dispute is not about whether the systems are impressive. It is about which relation between a symbol and the world constitutes understanding, and that question was open before anyone built one of these.

Common misconceptions

• Saying it just predicts the next word settles the matter. That states the training objective. What the training produced is a separate question and the one Block says we must ask.
• A system that conversed indistinguishably for long enough would have to understand. Blockhead is a counterexample by construction, and it is finite.
• The stochastic parrots paper argued that models cannot understand. Its main arguments concern environmental cost, data provenance and deployment harms.
• Chalmers argued that language models are conscious. He argued that they probably are not, and listed the obstacles that would have to be removed.
• Understanding, consciousness and moral status stand or fall together. Each has different evidence bearing on it, and they come apart in both directions.

Looking back

• Both sides agree on what the system is trained to do; the disagreement is over what may be inferred from the training objective about the mechanism.
• Block's blockhead shows that conversational performance is defeasible evidence, which cuts against confident claims in both directions.
• The grounding argument is strongest against systems trained on form alone, and is narrower than the use commonly made of it.
• Probing work of the kind done on Othello supplies the sort of evidence about mechanism that Block's argument demands, and it is not what a look-up table would show.
• Understanding, consciousness and moral status are three questions with different evidence, and running them together is what makes public argument about this go nowhere.

Sources

1. Rescorla, M. (2024). The Computational Theory of Mind. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Chalmers, D. J. (2023). Could a large language model be conscious? arXiv preprint 2303.07103. arxiv.org
3. Bender, E. M., and Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185-5198). aclanthology.org
4. Block, N. (1981). Psychologism and behaviorism. The Philosophical Review, 90(1), 5-43.
5. Li, K., and others. (2023). Emergent world representations: Exploring a sequence model trained on a synthetic task. International Conference on Learning Representations.

In 1890, in the sixth chapter of The Principles of Psychology, William James considered a theory his contemporaries called the mind-stuff theory: the view that our mental life is built out of smaller mental units, and those out of smaller ones, down to elementary psychic particles belonging to the atoms. He dispatched it with an image. Take a sentence of a dozen words, he writes, and take twelve men and tell to each one word. Stand them in a row or jam them in a bunch, and let each think of his word as intently as he will; nowhere will there be a consciousness of the whole sentence.

The theory James was demolishing has come back, defended by philosophers who accept every result in physics, and the objection he made in a paragraph now has a name and a literature. This last lesson follows that revival, states the problem it has not solved, and then draws the map the course has been building.

The hole the argument aims at

The modern case for the view does not begin with consciousness at all. It begins with a claim about what physics tells us, made independently by Bertrand Russell in The Analysis of Matter (1927) and by Arthur Eddington in The Nature of the Physical World (1928).

Ask what physics says mass is. It says how a massive body behaves with respect to other bodies: how it resists acceleration, how it curves spacetime, how it attracts. Ask what charge is and you get dispositions to attract and repel. Every property in the theory is characterised by its relations to other properties, which are themselves characterised by their relations. Eddington's summary was that the physical world as physics gives it to us is a schedule of pointer readings, a web of structure with the question of what fills it left open.

That is the hole. Physics describes the relations and is silent about the intrinsic natures of the things related. And there is exactly one case in which we are acquainted with an intrinsic nature, not by measurement but from the inside: our own experience.

Russellian monism puts the two together. The intrinsic natures that fill out the causal structure of the world are experiential, or of a kind out of which experience is made. If that is right, physics is complete as far as it goes, consciousness is not an add-on and not epiphenomenal because it is what the structure is made of, and there is no moment at which experience has to appear from nothing.

Remember: Russellian monism is a genus, and the species matter. Panpsychism says the intrinsic properties are experiential. Panprotopsychism says they are proto-experiential, not themselves experiences but the stuff experiences are made of. Russellian physicalism says they are physical properties we simply have not discovered. Chalmers's zombie argument in Lesson 8 leaves this whole branch standing, which is why his conclusion was a disjunction rather than a refutation of physicalism.

Strawson's version, and Goff's

Galen Strawson's 2006 paper Realistic Monism argues that a physicalist who is serious about physics is forced to panpsychism, which is not a sentence most physicalists expected to read. His chain runs: experience is the one thing whose reality is certain, so any physicalism worth the name must count it as physical. Now, could experience emerge from the wholly non-experiential? Emergence is intelligible, he says, when the emergent feature is a matter of how the base is arranged: liquidity from the interactions of molecules is a case where, given enough about the base, you can see why liquidity follows. Experience from the utterly non-experiential is not like that. It would be brute, and brute emergence is not emergence but magic with a respectable name. If experience is real, is physical, and cannot emerge brutely, then something experiential goes all the way down.

Philip Goff's case is comparative. Dualism has the interaction problem of Lesson 2. Physicalism has the explanatory gap of Lesson 7. Panpsychism has neither: consciousness is causally woven in because it is the categorical basis of the dispositions physics describes, and there is no gap to close because the fundamental is not being derived from something else.

Why this matters: the argument is not mystical and does not rest on intuitions about crystals. It rests on a claim about the limits of what a structural science can tell you, and if that claim is wrong, the whole position goes. The strongest objection is precisely there: that structure is all there is, and the demand for intrinsic natures is a confusion.

The combination problem

William Seager gave James's objection its modern name in 1995, and it is not one problem but a family.

• Subject summing. Micro-subjects each have a point of view. Points of view do not add. Put a billion of them together and you have a billion points of view, not one large one. This is James's objection unchanged, and it is the hardest.
• Quality combination. Suppose the micro-entities have micro-qualities. Why should assembling them yield the smell of coffee, rather than a great many micro-qualities sitting side by side?
• Structural mismatch. The organisation of your visual field, with its centre and periphery and its particular way of laying out space, does not obviously correspond to any organisation in the microphysical base.
• The palette problem. Physics recognises a small number of fundamental kinds. Human experience has an enormous variety of qualities. Where do the extra ones come from?

Four responses are on offer, and it is worth being frank about how each fares.

Cosmopsychism turns the problem upside down. Start with one cosmic subject, the universe, and derive individual subjects from it by a relation of grounding rather than composition. Nothing has to be summed. The cost is that decombination looks about as hard as combination, and the view now owes an account of why subjects appear where they do and not elsewhere.

Phenomenal bonding posits a relation, unfamiliar to us, that binds micro-subjects into a macro-subject the way nothing in our experience does. Goff proposed it and has been candid that naming the relation is not the same as understanding it.

Panprotopsychism denies that the fundamental entities are subjects at all, which removes the summing problem at a stroke. The cost is that it reopens the emergence question the position was built to avoid, though its defenders answer that emergence from the proto-experiential is intelligible in a way that emergence from the wholly non-experiential is not, which is exactly the claim in dispute.

Finally, some accept that combination is brute and argue that the position is then no worse off than its competitors, each of which has an unexplained joint somewhere. That is the honest defence, and it is a comparative one rather than a positive one, which tells you where the view stands. Chalmers's 2017 survey concluded that the problem is serious and not obviously fatal, which is about as warm as the reception gets.

The map

Set the four live options side by side, with what each cannot explain and what would move it.

Read down the third column and you have the honest result of this course. Every position has one place where the account runs out, and each has the same character: a joint that its rivals point at and its holders describe as the price of the view. Nobody in this argument is being careless. They are choosing which unexplained thing they can live with.

The core of it: the useful skill is not picking a column. It is being able to say, of any claim about consciousness you meet, which premise it rests on, which lesson in this course attacks that premise, and what the person making the claim owes you next. That is a small thing to be able to do and almost nobody can do it.

Common misconceptions

• Panpsychism says that rocks and thermostats are conscious. Most versions attribute experiential properties to fundamental entities and deny that aggregates are subjects, because unstructured combination is exactly what does not happen.
• It is a mystical position. Its main argument is about the limits of a structural science, and its defenders accept all of physics. Attack the structural premise if you want to attack the view.
• Russellian monism just is panpsychism. Panpsychism is one species; panprotopsychism and Russellian physicalism are others, and they differ over whether the intrinsic properties are themselves experiential.
• The combination problem is a technical wrinkle. It is James's objection of 1890, it is the reason the view remains a minority position, and no proposed solution is generally accepted.
• Future physics could reveal the intrinsic natures. The structural argument says physics characterises properties by relations and dispositions, so more physics yields more structure rather than filling the hole.

What to carry forward

• The structural argument holds that physics gives relations and dispositions and is silent about intrinsic natures, and that experience is the one intrinsic nature we know.
• Russellian monism fills the gap, in species that differ over whether the fundamental properties are experiential, proto-experiential or unknown physical.
• Strawson's route runs through the unintelligibility of brute emergence; Goff's is comparative, avoiding both the interaction problem and the explanatory gap.
• The combination problem is James's objection with a name, and its four main responses each shift the difficulty rather than dissolving it.
• All four live options have exactly one unexplained joint, and being able to name the joint and the argument behind it is what a training in this subject amounts to.

Sources

1. Goff, P., Seager, W., and Allen-Hermanson, S. (2022). Panpsychism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
2. Alter, T., and Pereboom, D. (2023). Russellian Monism. The Stanford Encyclopedia of Philosophy. plato.stanford.edu
3. James, W. (1890). The Principles of Psychology, Volume 1, chapter 6. Project Gutenberg edition. gutenberg.org
4. Strawson, G. (2006). Realistic monism: Why physicalism entails panpsychism. Journal of Consciousness Studies, 13(10-11), 3-31.
5. Chalmers, D. J. (2017). The combination problem for panpsychism. In G. Bruntrup and L. Jaskolla (Eds.), Panpsychism: Contemporary Perspectives. Oxford University Press.