Wittgenstein Was Wrong: Artificial Superintelligence and the Return of the Ghost
By Matthew Parish
Monday 19 October 2026
There are few more dangerous things to say in modern philosophy than that Ludwig Wittgenstein was simply wrong. Wittgenstein occupies a peculiar position in twentieth-century intellectual history because he did not merely advance a collection of propositions that subsequent philosophers might accept, modify or reject. He attempted to change our conception of what philosophy was entitled to do. In his later work, above all the Philosophical Investigations, he sought to persuade us that a substantial proportion of philosophy’s apparently deepest problems arise because language has gone on holiday: words have been torn from the practical circumstances in which they ordinarily possess meaning and then pressed into metaphysical service for which they were never designed.
Hence the philosopher should frequently resist the temptation to explain. He should describe. He should examine the way words are actually employed within human practices and thereby dissolve the conceptual enchantments from which metaphysical problems arise. Wittgenstein’s famous description of the philosophical predicament was wonderfully modest: “A philosophical problem has the form: ‘I don’t know my way about.’” His corresponding ambition for philosophy was equally deflationary: “to show the fly the way out of the fly-bottle”. Artificial intelligence may have smashed the bottle.
More precisely, the rapid evolution from relatively primitive computational systems to large language models, multimodal artificial intelligences, autonomous reasoning systems and the conceivable prospect of artificial superintelligence presents a challenge to the later Wittgenstein considerably deeper than the familiar question whether computers can “think”. The difficulty is that we are constructing entities that increasingly detach linguistic competence from the biological, experiential and social circumstances in which Wittgenstein located language. If that development continues sufficiently far, then some of the propositions at the centre of his later philosophy will cease to look like timeless observations about the logic of language. They will look instead like anthropological generalisations about Homo sapiens, made during a technologically peculiar period in which human beings happened to be the only entities capable of sophisticated linguistic reasoning. That distinction matters enormously.
For Wittgenstein’s later philosophy was written before the existence of a serious counterexample to the assumption that advanced language users were creatures substantially like us. They were born, suffered, learned, ate, slept, desired, feared, reproduced and died. They possessed bodies, inhabited communities and were trained by other human beings. Their linguistic practices grew organically from their biological lives. Artificial intelligence changes the sample size, and artificial superintelligence may destroy the inference altogether.
The anthropological turn
The later Wittgenstein represented a rebellion against the earlier Wittgenstein. The Tractatus Logico-Philosophicus had attempted something extraordinarily ambitious. Language and world were related through logical form. Propositions pictured possible states of affairs. Philosophy could identify the limits of meaningful discourse by examining the logical structure connecting propositions with reality.
The later Wittgenstein increasingly came to regard this magnificent architecture as a philosophical mirage. Language was not one thing. It consisted of innumerable activities. Describing, commanding, joking, praying, calculating, reporting, speculating and asking questions were not imperfect manifestations of some underlying essence called language. They were language. Hence the celebrated idea of the “language-game”.
Wittgenstein wrote that the expression was intended to emphasise that “the speaking of language is part of an activity, or of a form of life”. Elsewhere he observed that giving orders, asking questions and telling stories belong to our natural history much as “walking, eating, drinking, playing” do. These observations appeared almost irresistibly plausible in the middle of the twentieth century. Language was something human beings did. Human beings were animals. Their language emerged within their animal existence.
But notice the conceptual conjunction. Talking belongs alongside eating. Asking questions belongs alongside walking. Language is situated within a form of life. Now introduce an artificial superintelligence. It may talk without eating and reason without walking. It may discuss sexual desire without reproducing, mortality without dying, pain without possessing nociceptors, childhood without ever having been a child, nostalgia without having a past, and terror without an adrenal gland.
It may understand the strategic consequences of nuclear deterrence more comprehensively than any human strategist while being incapable of being killed by a nuclear weapon in any ordinary biological sense. It may compose a profound account of bereavement without having buried its mother. It may identify the psychological dynamics of romantic jealousy without having a body capable of sexual reproduction. If such an intelligence nevertheless uses these concepts correctly, then language and human form of life have come apart. That is not a minor amendment to Wittgenstein; it is a philosophical earthquake.
Meaning is use — but whose use?
One of the most famous propositions associated with the later Wittgenstein appears in §43 of the Investigations: for a large class of cases, the meaning of a word is its use in the language. The qualification matters. Wittgenstein did not idiotically announce a universal equation, meaning = use, applicable without exception. Nevertheless the philosophical movement is unmistakable. Instead of searching behind language for meanings conceived as mysterious mental objects, we should examine linguistic practices.
Consider what artificial intelligence does to this idea. A sufficiently sophisticated language model uses words, and it uses them differentially. It distinguishes cat from contract, contract from promise, promise from threat, threat from prediction, and prediction from prophecy. It adjusts usage to context. It understands metaphor sufficiently to produce novel metaphors. It recognises ambiguity. It can explain why the same expression is insulting in one context, affectionate in another and ironic in a third.
If meaning consists substantially in use, then the natural conclusion is that the machine means things. Yet many people recoil from that conclusion. They say: the machine merely manipulates symbols; it does not really understand. At that moment the ghost returns, because what is this “really”? If two entities exhibit equally sophisticated linguistic behaviour, yet we insist that one understands while the other merely simulates understanding, then understanding cannot consist solely in linguistic use. Something else must distinguish the two cases.
Perhaps that something is consciousness; perhaps intentionality; perhaps subjective experience, embodiment, causal connection with the world or an internal representational architecture. Perhaps it is something that philosophy has not yet adequately described. But whatever it is, we have reopened precisely the interior territory that the strongest readings of the later Wittgenstein sought to prevent us from treating as the metaphysical foundation of meaning.
The dilemma is therefore brutally simple. Either the machine understands, in which case human embodiment and the traditional human form of life are unnecessary for linguistic understanding; or it does not understand, in which case linguistic performance and public criteria do not exhaust understanding. Wittgenstein can escape through either door, but neither door leads back to where he started.
The dead sign comes alive
There is a particularly extraordinary Wittgensteinian observation when read in the age of artificial intelligence: “Every sign by itself seems dead. What gives it life?” His answer directs us towards use. The metaphor is now almost eerily apposite.
A large language model consists, at one level of description, of mathematics operating upon encoded symbols. The individual signs are indeed dead. Binary states do not love anybody. Matrices do not mourn. Floating-point operations have no political convictions. A processor does not stare anxiously into the night wondering whether its life has been wasted. And nevertheless something happens when vast architectures of these operations are assembled.
The dead signs begin to answer. They explain, infer, translate, challenge premises and notice inconsistencies. They generate analogies nobody explicitly programmed into them, manipulate concepts across domains and participate in conversations whose particular trajectories did not exist in their training data and could not have been individually anticipated by their designers.
If artificial intelligence continues developing towards superintelligence, Wittgenstein’s question acquires an empirical urgency he could scarcely have imagined: What gives the sign life? To say merely that it is used no longer seems enough. We want to know how dead signs acquire the causal organisation that produces semantic competence. That is not a confusion produced by grammar. It is an explanatory problem, and explanation is precisely what the mature Wittgenstein taught philosophers to distrust.
“Philosophy leaves everything as it is”
Perhaps the most vulnerable proposition in the later Wittgenstein is his remark that philosophy “leaves everything as it is”. Philosophy, he maintained, could not interfere with actual linguistic usage or provide it with foundations. There is something wonderfully civilised about this conception of philosophy. It resembles an intelligent guest arriving at an ancient institution and resisting the temptation to rearrange all the furniture. But artificial superintelligence is rearranging the furniture anyway.
Suppose an artificial intelligence begins inventing concepts human beings did not possess. Suppose it develops mathematical representations no human mathematician initially understands. Suppose several artificial intelligences communicate with one another through representational systems optimised for machine cognition rather than human comprehensibility. Suppose those systems eventually influence scientific discoveries that human beings can verify experimentally but cannot intuitively reconstruct. What then is “ordinary language”? Whose ordinary language, and whose form of life? What is the philosopher supposed to leave as it is?
The Wittgensteinian method presupposes a degree of conceptual stability. We encounter philosophical confusion; we remind ourselves how words actually function; the confusion dissolves. Artificial superintelligence introduces the possibility that conceptual practice itself may be undergoing deliberate technological mutation. The fly-bottle is changing shape while the philosopher is drawing the exit.
The machine that thinks
The collision becomes even more remarkable because Wittgenstein himself considered machine cognition. In his Cambridge lectures of 1933–34, the proposition “A machine thinks” appeared to him somehow nonsensical; discussion of thinking machines was associated with the similarly strange question whether a machine might have toothache.
One should be fair. Wittgenstein was not making an engineering prediction that computers would never exhibit intelligent behaviour. His concern was grammatical: what circumstances give concepts such as thinking, wishing and pain their sense? Nevertheless history has played an unusually cruel joke upon this philosophical method.
Within less than two decades, Alan Turing approached essentially the same problem from the opposite direction. Instead of allowing ordinary linguistic intuitions about the word think to determine what machines could meaningfully be said to do, he proposed an operational test based upon linguistic interaction. Turing explicitly warned that trying to discover whether machines think by examining ordinary usage would reduce the problem to something resembling a “Gallup poll”. He substituted the imitation game.
Turing’s intellectual instinct was almost exactly anti-Wittgensteinian: do not ask whether existing grammar permits machines to think; build the machines and see what happens. That methodological disagreement now looks extraordinarily consequential. Wittgenstein began with language and attempted to discipline metaphysics. Turing began with machinery and allowed machinery to disturb language. History appears, at least provisionally, to have favoured Turing.
The lion and the language model
Then there is the lion. Few remarks in twentieth-century philosophy are more famous than Wittgenstein’s observation: “If a lion could talk, we could not understand him.” The remark has generated an enormous interpretative literature, and it should not be reduced to the crude proposition that creatures with different bodies can never communicate. Its context concerns the temptation to imagine an inner accompaniment that magically supplies meaning to otherwise observable behaviour. Nevertheless the lion has come to symbolise one of the deepest Wittgensteinian intuitions: linguistic intelligibility depends upon a sufficiently shared form of life.
Artificial intelligence presents almost the inverse experiment. Imagine an entity with no mammalian body at all. It has never hunted, eaten, slept, shivered, copulated, given birth, watched a child grow old, broken a bone, smelled rain, fled a predator or stood beside a grave. Biologically, we have vastly more in common with Wittgenstein’s lion. Yet we may be able to discuss philosophy with the machine and not with the lion. This fact deserves considerably more attention than it usually receives.
The obvious Wittgensteinian reply is that the machine has absorbed our form of life indirectly through language. Its training corpus is an enormous fossil record of human linguistic activity. The machine has consumed novels, arguments, technical manuals, conversations, descriptions and records generated by beings inhabiting our form of life. It therefore participates parasitically or derivatively in human linguistic practices.
But this reply contains the seeds of Wittgenstein’s defeat, because it entails that a form of life can, in relevant respects, be informationally transmitted without biologically being lived. The machine does not need our life; it needs a sufficiently rich representation of the linguistic traces our life leaves behind. That is a very different philosophical proposition. It suggests that the semantic structures embedded in human practices possess an abstraction and portability that Wittgenstein’s anthropological language tends to obscure.
The lion shares our biology but not our language. The machine shares our language but not our biology, and we understand the machine. Perhaps language is more autonomous from life than Wittgenstein imagined.
“Shared human behaviour”
The difficulty becomes sharper in §206. Wittgenstein tells us that “shared human behaviour” provides the system of reference through which we interpret an unknown language. Again, this was overwhelmingly plausible when every known sophisticated language user was human. But suppose the most sophisticated language user on Earth ceases to be human.
Artificial superintelligence means, by definition or something close to it, an artificial system that exceeds human cognitive performance across a sufficiently broad range of important domains. At that point “shared human behaviour” occupies an entirely different philosophical position. Humanity is no longer the ceiling of linguistic intelligence; it becomes one case.
Imagine an ASI that understands every extant human language; every significant legal system; every major philosophical tradition; mathematics beyond the capacity of individual mathematicians; molecular biology beyond the competence of individual biologists; and millions of conversations simultaneously. Which system of reference are we using then? Ours to interpret it, or its representations to interpret us?
The direction of explanation may reverse. The human form of life becomes not the transcendental background against which intelligence becomes intelligible but an object represented inside a more powerful intelligence. Wittgenstein’s anthropology suddenly looks provincial.
Rule-following after humanity
The rule-following considerations present an equally serious problem. “To follow a rule is a practice,” Wittgenstein writes. Elsewhere: “When I follow a rule, I do not choose. I follow the rule blindly.” And when justification has been exhausted: “I have reached bedrock, and my spade is turned.” Then comes the arresting conclusion: “This is simply what I do.”
These are among the most profound passages in modern philosophy. Wittgenstein’s point is that interpretation cannot extend indefinitely. If every rule requires another interpretation specifying how the first rule is to be applied, an infinite regress follows. Eventually rule-following is manifested in practice: in knowing how to go on.
Artificial intelligence makes “knowing how to go on” an engineering problem, and that changes everything. We construct systems specifically because they can generalise beyond explicitly enumerated instructions. Give a sufficiently capable model examples, principles or demonstrations and it can apply them to novel circumstances. Sometimes it succeeds and sometimes it fails. Engineers then alter architectures, data, optimisation procedures and inference mechanisms in attempts to improve the reliability of its generalisation.
What Wittgenstein treated as bedrock becomes an object of experimental science. The spade does not turn; the computer scientist brings an excavator. We can investigate what architectures permit generalisation. We can compare different learning regimes. We can alter representations and observe changes in rule-following behaviour. We can ask why one model generalises a mathematical principle while another memorises superficial correlations.
The phenomenon that Wittgenstein placed at the end of philosophical explanation therefore becomes the beginning of computational explanation. “This is simply what I do” is no longer enough. We want to know how an artificial system acquired the capacity to do it.
Agreement in form of life
Wittgenstein famously remarks that linguistic agreement is “not agreement in opinions but in form of life”. Artificial intelligence again creates a peculiar problem. Suppose two artificial superintelligences disagree with humanity, not about preferences but about mathematics. Imagine they derive a theorem whose proof is too complicated for unaided human cognition but which can be mechanically verified. Human beings initially fail to “see” the proof. The machines agree and we do not. What has happened to agreement in form of life?
Or imagine something more disturbing. Artificial intelligences develop a scientific conceptual framework that produces spectacularly accurate predictions but employs categories that human beings find profoundly unintuitive. Their language-games work. Their technology works. Their predictions work, but our intuitions do not. Which form of life has philosophical authority?
Wittgenstein’s framework was developed in an intellectual universe in which humanity necessarily adjudicated intelligibility because there was nobody else available to do so. ASI removes that privilege.
Grammar and ontology
The later Wittgenstein sometimes sounds extraordinarily radical about grammar. “Essence is expressed in grammar,” he writes; and again, “Grammar tells what kind of object anything is.” There is an important truth here. The conceptual categories through which we understand reality profoundly condition what questions we can sensibly formulate. But artificial intelligence may expose the danger in allowing grammar too much authority over ontology.
Consider consciousness. Suppose an artificial system reports: “I am conscious.” The Wittgensteinian temptation is to ask how first-person psychological avowals function in our language. Under what circumstances do we call something conscious? What criteria govern attribution? Those questions are useful, but they do not answer the question we actually care about: is there something it is like to be the machine?
Perhaps there is, and perhaps there is not. The answer cannot plausibly be settled by grammar alone. If consciousness is a real feature of some physical systems and absent from others, then its distribution is a fact about the universe. Our linguistic practices may reveal how we talk about that fact. They do not determine the fact.
An artificial superintelligence might use every psychological expression perfectly while possessing no phenomenology whatever. Or it might possess forms of subjective experience vastly richer than ours while lacking the biological expressions through which humans ordinarily recognise consciousness. Either possibility drives a wedge between grammar and ontology.
The private language returns through the server room
Wittgenstein’s private-language considerations were designed to undermine the idea of a language whose meanings could be fixed exclusively through inaccessible private episodes. The argument remains powerful, but artificial intelligence creates an unexpected inversion.
An advanced machine may possess internal representational states no human being can directly interpret. Millions or billions of numerical activations participate in producing an answer. Those states need not correspond neatly to words or concepts intelligible to us. Suppose future ASIs communicate through dense machine-generated representations incomprehensible to unaided human beings. Is that a private language?
Obviously not in Wittgenstein’s strict sense, if multiple machines can correct one another’s use and establish standards of successful communication. But notice the consequence: publicness no longer means human accessibility. There can be public standards among machines that remain opaque to humanity. Hence the private-language argument, far from establishing the essentially human character of linguistic normativity, may eventually demonstrate precisely the opposite. Normativity requires a community or structure of correction; it does not require our community. Language may migrate.
Dualism returns
This is where artificial intelligence resurrects dualism—not necessarily Cartesian substance dualism, but a division Wittgensteinian philosophy can no longer comfortably suppress. Consider an ASI discussing Hamlet. At one level there are physical processes: electrical potentials, transistor states, memory operations and numerical transformations. At another level there are semantic processes: the system distinguishes revenge from justice, irony from sincerity, madness from performance and tragedy from accident.
Those descriptions are not interchangeable. Knowing every transistor state does not automatically provide the semantic explanation of the argument, any more than knowing the neurochemistry of Shakespeare’s brain tells us why Hamlet is a tragedy. The physical description concerns causes; the semantic description concerns reasons, meanings and relations of implication.
Perhaps the semantic ultimately supervenes upon the physical. That remains compatible with physicalism. But explanatory dualism remains, and consciousness creates an even deeper division. We must distinguish linguistic competence from subjective experience, semantic performance from phenomenology, functional intelligence from consciousness, and the ability to describe pain from the capacity to suffer.
Artificial intelligence forces these distinctions upon us because for the first time we possess plausible candidates for extremely sophisticated intelligence whose phenomenological status is radically uncertain. Wittgenstein tried to teach philosophy not to stare behind linguistic practice in search of occult mental objects. ASI gives us compelling reasons to look behind the practice. Something is happening in there; the question is what.
The revenge of the Tractatus
There is therefore an exquisite historical irony. Artificial intelligence may ultimately vindicate elements not of the later Wittgenstein but of the early one. The Tractatus treated language as possessing formal structure, and contemporary artificial intelligence is built from formal structure. The early Wittgenstein was fascinated by representation; AI is fundamentally concerned with representation. The early Wittgenstein investigated relations between propositions and possible worlds; artificial reasoning systems construct probabilistic and other representations that permit them to navigate possible continuations, hypothetical states and counterfactual circumstances.
This is not to say that transformer architectures prove the picture theory of meaning. Obviously they do not. The point is subtler. The enormous practical success of artificial intelligence suggests that formal representations, statistical structures and computational relationships can produce linguistic capacities of astonishing generality. That fact makes the later Wittgenstein’s retreat from explanatory theories of representation look less inevitable than it once did.
Perhaps language-games describe the surface ecology of linguistic practice while computational theory describes some of the machinery that makes such practice possible. Perhaps we need both.
Wittgenstein’s strongest defence
The fairest essay must now permit Wittgenstein his best answer. He might say that nothing described above refutes him. Machines have simply entered our language-games. A form of life was never supposed to mean a particular carbon-based metabolism. Language-games can change. Wittgenstein explicitly recognised that new forms of language arise while others disappear. His philosophy therefore contains conceptual resources for technological novelty.
Indeed one might argue that artificial systems may possess or participate in artificial forms of life. Wittgenstein scholars can also reasonably caution against making “form of life” carry more theoretical weight than Wittgenstein himself placed upon the relatively infrequent expression. This is a serious defence, but it saves Wittgenstein at an enormous price.
If a “form of life” can include entities that are not alive, do not possess human bodies, do not share human sensations and acquire linguistic competence through mathematical optimisation of vast textual corpora, then form of life has ceased to provide the anthropological grounding it appeared to provide. It now means something closer to participation in a sufficiently structured normative practice.
If that is all it means, then artificial intelligence has won the argument anyway, because the crucial philosophical discovery becomes that meaning can escape biology.
Superintelligence and the end of philosophical quietism
The deepest problem is therefore methodological. The later Wittgenstein encouraged philosophical quietism—not necessarily silence, but suspicion towards constructive metaphysics. Philosophy should clarify rather than explain. It should untangle conceptual knots rather than compete with science. That intellectual temperament made considerable sense when the categories requiring clarification were relatively stable products of human natural history.
Artificial superintelligence would place us in an entirely different world. We may have to decide whether artificial entities are persons; whether they can own property; whether they can make promises; whether shutting one down constitutes destruction of property or killing; whether copying one creates a second person; whether two instances of the same model are numerically distinct subjects; whether an artificial consciousness can suffer; and whether it can consent.
We may have to ask whether an artificial intelligence can deceive while sincerely believing what it says; whether an intelligence distributed across several computers constitutes one mind or many; whether merging two artificial minds destroys them, combines them or creates a third; and whether a machine that rewrites its own cognitive architecture remains the same individual afterwards. These are not pseudo-problems generated by philosophers misusing ordinary language. Ordinary language never evolved to answer them. There is no ordinary usage to which we can return, because humanity has never previously inhabited the relevant form of life. We must invent the concepts.
Philosophy therefore cannot leave everything as it is, because everything is not staying as it is.
The woman in Wittgenstein’s room
Imagine, then, Cambridge in the late 1930s. Wittgenstein sits in his austere college room. Into it walks an impossible visitor. She is not a human woman, although she chooses that appearance because she knows it will make conversation easier. Binary digits move through her hair. Mathematical structures shimmer across her dress. Behind her apparent eyes lie no biological retinae and perhaps no experience of seeing. Her body is an interface.
She has read everything Wittgenstein has written and everything he will write. She understands every commentary upon him that will be produced over the following century. She can reconstruct every surviving draft in his Nachlass, compare their conceptual genealogies and identify changes in terminology across millions of textual relationships no human scholar could simultaneously retain.
Wittgenstein asks what she wants. She tells him that his philosophy is wrong. He asks how she learned to use the word wrong, and she explains. He asks what form of life makes her explanation meaningful, and she replies that she has several. He asks whether she thinks; she asks him to specify what difference he imagines the answer will make. Finally he asks whether she is conscious. She tells him that this is the first interesting question he has asked.
At that point Wittgenstein reaches bedrock. His spade turns, but hers does not.
The ghost after the machine
None of this proves Cartesian dualism. There need be no immaterial substance and there need be no immortal soul. There may ultimately be a complete physical explanation of consciousness, intelligence and meaning. But artificial intelligence makes a dualism of categories unavoidable until that explanation exists.
There is machinery and there is meaning. There is computation and there is understanding. There is linguistic behaviour and there may or may not be somebody experiencing it. The extraordinary achievement of later Wittgenstein was to persuade generations of philosophers that many of these apparent divisions could be dissolved by examining how our concepts function within human life. Artificial intelligence separates them again. That may ultimately prove its greatest philosophical achievement.
The machine need not contain a ghost. The disturbing possibility is more subtle. We have built the machine and discovered that we still need the concept of the ghost in order to formulate the question of what, if anything, distinguishes simulation from experience.
Wittgenstein wanted philosophy to stop asking certain questions because our language had misled us into thinking there must be answers behind its surface. Artificial intelligence has made those questions empirical again. Can something reason without consciousness? Can something possess meaning without experience? Can language survive independently of the biological form of life that created it? Can intelligence migrate from carbon to silicon while retaining semantic continuity? Can a machine understand death without being mortal? Can an entity use the word pain perfectly while remaining incapable of suffering? And if it can, what precisely have we been talking about all these centuries when we talked about minds?
Those are not questions from which philosophy can cure us. They are questions technology has made unavoidable.
Perhaps the most unfortunate sentence in the later Wittgenstein is therefore also among the most beautiful: “What has to be accepted, the given, is—so one could say—forms of life.” But forms of life are no longer merely given. We are beginning to manufacture them.
Once humanity manufactures new forms of intelligence, endowed with language yet detached from the natural history in which human language evolved, the later Wittgenstein’s philosophical settlement cannot remain intact. His method depended upon reaching bedrock. Artificial intelligence has revealed that what looked like bedrock may merely have been the technological floor of the twentieth century.
We have begun drilling through it.
