Wittgenstein’s Blue Book and the Philosophy of Large Language Models: Do Machines Understand What They Say?

By Matthew Parish

Saturday 17 October 2026

The extraordinary development of artificial intelligence in the twenty-first century has revived a series of philosophical questions that many people had imagined were settled, or at least safely confined to the more obscure corridors of university philosophy departments. Can a machine think? What does it mean to understand a sentence? Is language merely a system of symbols, or does its meaningful use require consciousness, intention and an understanding of the world? When a large language model responds to a question with apparent intelligence, is it genuinely engaged in an intellectual activity, or is it merely imitating the external manifestations of human thought? These questions have acquired an unexpected commercial urgency, because some of the most valuable technological enterprises in the world now depend upon persuading their customers that machines are capable of doing things that, until recently, we regarded as exclusively human.

It is therefore an especially curious historical coincidence that one of the most illuminating guides to these questions was written almost a century ago, before the invention of the modern electronic computer. Ludwig Wittgenstein’s Blue Book, dictated to his students at Cambridge during the academic year 1933–1934 and subsequently published in 1958, is amongst the most important transitional works in twentieth-century philosophy. It marks Wittgenstein’s movement away from the austere logical architecture of his earlier Tractatus Logico-Philosophicus towards the more subtle, practical and frequently disconcerting investigations into language that would characterise his later philosophy. Although comparatively short, the Blue Book contains the intellectual foundations of an approach to language that may be more relevant to artificial intelligence than almost anything written by contemporary computer scientists.

Wittgenstein begins by asking what it means for a word to have meaning. This apparently elementary question immediately exposes a philosophical difficulty. We are inclined to suppose that behind every meaningful word there must exist something corresponding to it: an object, a mental image, a concept, or some mysterious psychological process through which the word acquires its significance. Yet this assumption, Wittgenstein argues, arises principally from a misunderstanding of the way language actually functions. We have been seduced by the grammatical structure of our own sentences into imagining that because a word is a noun, there must be some corresponding thing to which it refers. The result is that we begin inventing metaphysical entities to explain something that requires no such explanation.

Consider the word ‘red’. We might imagine that understanding this word requires the possession of a mental image of redness, which we consult whenever we encounter an object and wish to determine whether the word applies to it. Wittgenstein points out that ordinary language does not function in this manner. If somebody asks us to fetch a red flower, we do not normally summon an internal representation of redness and compare it with the flowers before us. We simply select a red flower. Our understanding is demonstrated in the activity of using the expression appropriately, rather than through the presence of some additional mental object accompanying the words.

This observation leads Wittgenstein towards one of the most influential propositions in modern philosophy: that the significance of a linguistic sign is to be understood through its use. Language is not a collection of inert symbols animated by an invisible metaphysical substance called meaning. Rather, words acquire their significance through their participation in the activities, conventions and practices of those who employ them. To understand language is not necessarily to possess a particular internal experience, but to know how to operate within the various circumstances in which language has a function.

Here we encounter the first profound implication for large language models. Such models operate by processing linguistic tokens through extraordinarily elaborate mathematical structures whose parameters have been adjusted through training on vast quantities of information. They identify relationships between expressions and produce responses that may demonstrate impressive grammatical sophistication, contextual sensitivity and apparent reasoning. Nevertheless, their operations differ considerably from the familiar psychological processes through which human beings experience language. The question is whether that difference disqualifies them from understanding anything at all.

A conventional philosophical response is to insist that it must. Whatever the sophistication of their linguistic output, machines supposedly lack the inner consciousness, subjective awareness or intentionality that gives human language its meaning. Their words are therefore said to be empty imitations of genuine communication. The difficulty with this argument, from a Wittgensteinian perspective, is that it may simply reintroduce the very metaphysical assumption that the Blue Book seeks to dismantle. We imagine that meaningful language must be accompanied by something concealed within the speaker, and then conclude that a machine cannot understand because we have no evidence that it possesses this mysterious accompaniment.

Wittgenstein would almost certainly have found this reasoning suspicious. Indeed, the Blue Book expressly raises the question of whether a machine could think, considering it alongside the more peculiar question of whether a machine could suffer toothache. His purpose is not to demonstrate that machines possess consciousness, nor to establish that they are constitutionally incapable of acquiring it. Rather, he asks us to examine what we mean when we describe something as capable or incapable of thinking. Are we reporting an empirical discovery about machinery, or are we expressing a convention about the circumstances in which we are prepared to use particular words?

This distinction is of extraordinary importance. The assertion that machines cannot think may sound like a scientific proposition, but it can conceal a grammatical prejudice. If by thinking we mean an activity necessarily performed by a biological human brain, then naturally a machine cannot think, but the conclusion follows from the definition rather than from any discovery about artificial intelligence. If, on the other hand, thinking is identified through the performance of activities such as reasoning, calculating, drawing inferences, correcting mistakes and explaining conclusions, then it becomes considerably more difficult to exclude machines merely because they are made of silicon rather than organic tissue.

Yet the opposite conclusion would be equally premature. Wittgenstein does not invite us to abandon every distinction between human understanding and mechanical symbol manipulation. His argument is more subtle, and consequently more demanding. The mistake consists in supposing that there must be one universal property called understanding, the possession of which can be determined by inspecting the internal constitution of whatever entity we are considering. Understanding, like so many other philosophical concepts, may have several different uses depending upon the activity in question.

This brings us to another central theme of the Blue Book: Wittgenstein’s criticism of what he calls the craving for generality. Philosophers, he observes, repeatedly assume that everything described by the same word must possess some common essential characteristic. Yet many concepts do not function in this manner. Games, for example, may involve competition, amusement, rules, physical exercise, chance or intellectual skill, without every game necessarily possessing all these characteristics. What unites them is a network of overlapping similarities rather than a single universal essence.

The same observation may apply to intelligence. A mathematician solving an equation, a child learning a language, a chess player anticipating an opponent’s next move, and a diplomat negotiating a ceasefire may all be described as exercising intelligence. Yet the activities involved are strikingly different. It is therefore not immediately obvious that artificial intelligence must be judged according to one universal criterion, any more than it follows that every human act of intelligence must involve the same psychological processes.

A large language model may exhibit genuine competence in translation while remaining unreliable in mathematical reasoning. It may explain a philosophical argument with considerable sophistication while inventing a judicial precedent that does not exist. It may construct a persuasive commercial proposal while possessing no independent commercial interests of its own. These phenomena are not necessarily contradictions. They suggest that the various activities we collect under the heading of intelligence should be examined individually, according to their different purposes and standards of correctness, rather than being forced into an artificial metaphysical unity.

There is, however, an important difficulty with interpreting Wittgenstein as a philosopher who anticipated and justified modern artificial intelligence. His account of meaning as use does not entail that every system capable of producing plausible sentences genuinely understands them. Statistical regularities in language are not identical to the social practices through which language acquires its significance. This distinction is fundamental, because the activities Wittgenstein examines are embedded in forms of human behaviour involving expectations, correction, purposes, consequences and shared standards of judgment.

One of the Blue Book’s most instructive examples involves a written instruction to fetch six apples from a grocer. The significance of the words does not arise from a private mental image of six apples, but from a sequence of practical operations through which the instruction is fulfilled. There is a shopkeeper, a system of identifying goods, a practice of counting, and an established relationship between the written instruction and the activity it directs. The language is intelligible because it participates in an organised human practice.

Imagine now asking a large language model to fetch six apples. It may explain the instruction perfectly, translate it into several languages, calculate the cost of the purchase, or formulate a plan for obtaining the fruit. If equipped with appropriate tools and authorised access to a delivery service, it might even arrange for the apples to arrive. Yet these different achievements should not be confused. Explaining an instruction, planning its execution and carrying it out are distinguishable activities, and competence in one does not automatically establish competence in the others.

This is where the limits of contemporary language models become philosophically interesting. Human linguistic competence develops within a world of practical engagement. We learn what promises mean partly because we are expected to keep them, what danger means partly because we can be harmed, and what an apology means partly because we inhabit relationships in which responsibility and reconciliation matter. Language is embedded in action, and action has consequences that extend beyond the production of further sentences.

A large language model can reproduce the linguistic conventions associated with these activities, sometimes with remarkable sensitivity. It can write a persuasive apology, compose a declaration of love, explain the obligations arising under a contract, or express sympathy to somebody who has suffered a bereavement. But does it thereby apologise, love, undertake contractual obligations or sympathise? The answer cannot be obtained merely by analysing the grammatical quality of its responses. We must examine the place those responses occupy within the relevant human practices.

If a company uses an automated system to issue a binding contractual acceptance, the absence of human consciousness within the software does not necessarily prevent legal consequences from arising. The system may operate within an institutional arrangement through which its outputs are attributed to the company. Conversely, when an ordinary conversational model declares that it personally promises to perform some future act, we have good reason to distinguish the fluent linguistic expression from the assumption of personal responsibility that ordinarily accompanies such a promise. The words may be identical, but their practical significance is different.

This is a particularly useful correction to the increasingly theatrical manner in which artificial intelligence is marketed. Contemporary systems are encouraged to communicate in the first person, to express apparent preferences, and to adopt the conversational manners of sympathetic human companions. Such practices make the systems more convenient to use, but they also encourage philosophical confusion. The word ‘I’ ordinarily identifies the person speaking, but its appearance in machine-generated dialogue does not itself demonstrate the existence of an independent personality possessing desires, experiences or interests.

Nor does the absence of proof of such experiences automatically establish their impossibility. The question of machine consciousness remains a serious philosophical and scientific problem, and Wittgenstein offers no convenient formula for resolving it. What he does offer is a warning against mistaking the familiar grammar of personal expressions for evidence of an invisible mental life. Equally, he would caution against imagining that the lack of access to such an inner life is sufficient to settle every practical question about the capacities of a machine.

A further and frequently neglected aspect of the Blue Book concerns Wittgenstein’s distinction between reasons and causes. When somebody explains why they performed an action, they may be giving a reason that justifies the action, or describing the causal circumstances that produced it. These are different kinds of explanation, although ordinary language frequently conceals the distinction. A person might explain a mathematical conclusion by demonstrating the calculation that supports it, without thereby describing every psychological event that occurred while solving the problem.

This distinction has immense implications for artificial intelligence. When a language model explains why it reached a conclusion, the explanation may constitute a perfectly intelligible justification of that conclusion. Yet it does not necessarily provide an accurate causal account of the computational processes that generated the response. An explanation can be logically persuasive without being a faithful reconstruction of the internal mechanisms responsible for its production. Conversely, an exhaustive technical description of neural activations may tell us comparatively little about whether the conclusion was justified.

The distinction matters particularly in law, medicine and public administration, where artificial intelligence is increasingly employed to organise information and assist human judgment. A lawyer evaluating a legal proposition does not merely want to know which sequence of computational operations produced it. The lawyer wants to know whether the proposition follows from the relevant legislation, judicial authorities and established principles of interpretation. The justificatory relationship between a conclusion and its supporting reasons cannot be replaced by a description of the machinery that generated the conclusion.

Nevertheless, the machinery matters for a different purpose. If we wish to understand why a system repeatedly makes a particular error, a causal investigation into its training, architecture, available information and operating conditions may be indispensable. Wittgenstein’s contribution is to remind us that these are two different investigations. The question of why a model generated a particular sentence is not necessarily the same question as whether that sentence is correct, meaningful or justified.

There is a further irony in the relationship between Wittgenstein and modern machine learning. Much of contemporary artificial intelligence depends upon the extraction of generalisable patterns from enormous quantities of data. Wittgenstein, by contrast, repeatedly warns against the philosophical temptation to replace the examination of particular cases with sweeping generalisations. Yet these positions are not necessarily incompatible. The discovery of statistical regularities is a legitimate scientific enterprise; the difficulty arises when such discoveries are mistaken for complete philosophical explanations of meaning, intention or understanding.

The mathematician may describe a model as calculating probabilities over possible sequences of tokens, and that description may be entirely appropriate at a particular level of analysis. But to conclude that the system therefore understands nothing is to make a further philosophical claim that does not follow automatically from the mathematics. Equally, to conclude that the system understands everything because it performs impressively on linguistic benchmarks is to confuse a collection of technical achievements with an account of the nature of understanding.

This tension has already produced substantial disagreement amongst contemporary researchers. Emily Bender and Alexander Koller, in their influential 2020 paper on meaning and natural language understanding, argued that learning linguistic form alone cannot establish the connection between language and communicative meaning. Other philosophers and computational linguists have challenged the assumption that systems trained on text must necessarily be incapable of reference, pointing out that linguistic expressions themselves possess histories connecting them to objects, institutions and events in the world. These disagreements are important because they reveal that the philosophical significance of language-model training cannot be settled simply by repeating that the systems predict tokens.

Wittgenstein’s contribution is to change the character of the investigation. Instead of searching for an abstract substance called meaning concealed within the artificial neural network, we should examine the different circumstances in which its language is employed. When a model translates a Ukrainian medical instruction into English, what standards determine whether the translation is correct? When it summarises an arbitration award, what makes the summary faithful to the decision? When it explains a scientific theory, what distinguishes genuine explanatory competence from the production of plausible but erroneous technical vocabulary?

These questions admit of practical examination. We can test the translation against the intentions and conventions of the original language, compare the legal summary with the text of the award, and evaluate the scientific explanation against established knowledge and experimental evidence. We can also investigate whether the system responds appropriately to correction, preserves relevant distinctions, acknowledges uncertainty and identifies the limits of its own information. Such inquiries are more illuminating than an interminable metaphysical dispute about whether silicon possesses a soul.

The Ukrainian experience provides a particularly striking illustration of the distinction between linguistic competence and practical understanding. In a city such as Lviv, the expression announcing an air raid alert is not merely a sequence of words. It participates in an elaborate network of public warnings, institutional responsibilities, military assessments and individual decisions about personal safety. The meaning of the announcement is inseparable from the practical circumstances in which it is issued and the consequences that may follow from ignoring it.

An artificial intelligence system may translate the announcement, explain the relevant terminology or assist in making emergency information accessible to foreigners. These may be valuable and meaningful linguistic functions. Yet if it begins independently assuring people that an alert can safely be ignored, it has moved from linguistic assistance towards a form of practical judgment with potentially grave consequences. The distinction is not that only a conscious human being could ever perform such a function. Automated warning systems already participate in important institutional practices. Rather, the question is whether the system is reliably connected to the necessary evidence, authority, procedures and responsibilities.

Here the practical force of Wittgenstein’s philosophy becomes apparent. Understanding is not a magical property that mysteriously appears when a language model reaches a certain number of parameters. Nor is it necessarily an exclusively biological phenomenon. It is a concept whose application depends upon the activities being described, the conventions governing those activities, and the standards by which successful performance is distinguished from failure.

The increasing integration of artificial intelligence with external tools, scientific instruments, databases and automated systems makes these distinctions progressively more important. A language model confined to conversation occupies a different practical position from a system capable of conducting experiments, operating machinery, implementing software changes or making decisions within an institutional framework. The extension of its capabilities may expand the range of activities in which we can meaningfully describe it as reasoning, investigating or understanding. But no single technical development automatically establishes the full collection of characteristics ordinarily associated with human thought.

We should also resist the temptation to treat human beings as possessing perfect understanding merely because we are conscious. Human language is full of conventional expressions, unconscious habits, mistakes, misunderstandings and apparently intelligent remarks that conceal extraordinary ignorance. Much of ordinary conversation involves the skilful reproduction of familiar linguistic patterns. Indeed, Wittgenstein’s descriptions of the everyday use of language frequently make human linguistic activity appear considerably less mysterious than philosophers have traditionally supposed.

This observation should introduce an element of humility into debates about artificial intelligence. If the ability to produce language according to learned conventions is regarded as disqualifying evidence against machine understanding, we may find ourselves applying an unexpectedly severe standard to human beings. On the other hand, the fact that human thought frequently employs habits and conventions does not mean that every system reproducing those habits occupies the same practical or moral position as a person.

There is therefore no necessary contradiction between acknowledging that large language models perform intellectually valuable activities and denying that they possess human personalities. We can accept that their linguistic capabilities are remarkable without attributing consciousness to them, and recognise that their outputs may have genuine significance without assuming that every statement they produce expresses a personal intention. The philosophical challenge is to preserve the distinctions that our increasingly anthropomorphic technological vocabulary threatens to obliterate.

One suspects that Wittgenstein would have found the contemporary debate about artificial intelligence both fascinating and exasperating. Silicon Valley has turned the attribution of human characteristics to machines into a major commercial enterprise, while some of its critics have responded by insisting upon equally sweeping metaphysical prohibitions. One faction appears determined to manufacture artificial persons as quickly as possible, while another insists that no amount of computational achievement could ever amount to genuine thinking. Both may be suffering from precisely the craving for generality against which Wittgenstein warned.

The Blue Book offers an escape from this sterile confrontation. Instead of demanding a definitive answer to the question of whether artificial intelligence understands language, we should investigate the different things we mean by understanding and the circumstances in which those descriptions are justified. This is not an evasion of the philosophical problem. It is a recognition that the problem may have been badly formulated in the first place.

There is something peculiarly appropriate about the fact that a work composed in Cambridge during the 1930s should illuminate the most commercially consequential intellectual development of the twenty-first century. Wittgenstein inhabited a world in which language was overwhelmingly regarded as an activity of human beings, and machines were not yet capable of engaging in apparently intelligent conversation. Nevertheless, he recognised that the greatest obstacles to understanding language were not necessarily technical. They arose from the misleading pictures that language itself encouraged us to construct.

Large language models have made those misleading pictures vastly more seductive. They can speak in the first person, deploy sophisticated philosophical terminology, offer apparently sensitive moral judgments and discuss the nature of their own supposed existence. Yet the more convincingly they employ the grammar of human experience, the more carefully we must examine what their words are actually doing. Fluency is not an adequate substitute for justification, and the appearance of personal intention is not evidence of responsibility.

Perhaps the most important lesson of the Blue Book is therefore neither that machines think nor that they cannot think. It is that philosophy must cease searching for an occult ingredient that transforms mere signs into meaningful language. The significance of language emerges from the complex practices in which signs are employed, and the philosophical task is to understand those practices without confusing them with imaginary metaphysical mechanisms.

This conclusion is simultaneously reassuring and unsettling. It is reassuring because artificial intelligence need not possess a mysterious inner consciousness in order to perform useful and perhaps genuinely intellectual activities. It is unsettling because the distinctions between human and artificial intelligence may prove more complicated, more contextual and less absolute than either enthusiasts or sceptics would prefer. Understanding is not a precious metaphysical fluid confined to the human skull, but neither is it something that can be established simply by manufacturing increasingly plausible sentences.

We may therefore discover that the most important philosophical question about artificial intelligence is not whether machines can think like us, but whether we have understood what we mean when we say that we ourselves are thinking. Wittgenstein’s answer would require us to examine our words, our activities and the conventions through which they acquire significance. In doing so, we might discover that artificial intelligence has not merely presented humanity with a new technological problem. It has exposed the inadequacy of some of our oldest philosophical assumptions about the nature of the human mind.

The Blue Book was written before computers could converse, before artificial neural networks existed, and before the technological world became preoccupied with the possibility of artificial general intelligence. Yet it may have anticipated the central intellectual predicament of our age more clearly than many contemporary manifestos on artificial intelligence. The greatest philosophical danger is not necessarily that machines will begin to think, nor even that they will persuade us that they can think. It is that we shall continue to use the words thought, meaning and understanding without having adequately considered what we are saying.

And that, rather than any deficiency in the machines themselves, may turn out to be the most distinctively human failure of all.