By 'success' we think of current AI/LLMs capacity of producing text that is regarded as coherent, informative, even convincing, by human readers [see for instance Spitale et al. and Salvi et al.]

Wittgenstein's position is

For a large class of cases of the employment of the word “meaning” - though not for all - this word can be explained in this way: the meaning of a word is its use in the language.

in his "Philosophical Investigations"

Notice that the question is not about "Does the machine understand? Can the machine think? Does it have a mind/conciousness?" or anything of the sort - the man himself says "But surely a machine cannot think!" -, but only about language, as in Given the machine produces text based on statistical analysis, and that the texts seem to us to be 'meaningful', is 'meaning' really just use?

  • 4
    In Seven Pillars for the Future of Artificial Intelligence by Cambrai, Wang, Ho': The main goal of most tech companies is not designing the building blocks of intelligence but simply creating products that existing and potential customers deem intelligent. In this context, instead of labeling it as ‘artificial’ intelligence, it may be more apt to characterize such research as ‘pareidoliac’ intelligence. Commented Apr 16 at 19:25
  • 2
  • 1
    Too short to be a proper answer, but LLMs are actually really bad at corresponding language to non-linguistic objects. Absent either very careful training or a secondary technology for handling it, a generative LLM will answer, "Done, here you go" or similar to requests for some non-linguistic object (like ordering take-out and sending a receipt, or drawing a picture), while neither completing the task nor exhibiting any awareness that it hasn't completed the task.
    – Tim C
    Commented Apr 16 at 20:50
  • 1
    @user66697 no no, i do mean 'support'
    – ac15
    Commented May 3 at 14:52
  • 1
    i started a question about my confusion about this apologies to anyone who finds that annoying
    – andrós
    Commented May 3 at 15:55

9 Answers 9


Yes, indeed: According to the post-Tractatus Wittgenstein, words are "meaning families"; the specific "meaning" of a word is determined by (or perhaps is) its use in context. Speaker and listener must share part of that context lest they could not communicate (see the "private language" discussion of e.g. a sensation peculiar to a single person in the Investigations).

In this sense, surely the words and sentences produced by LLMs have meaning. The LLMs have been trained with a lot of texts, in which the words appear in various contexts; the models have "absorbed" these contexts and internalized them. When they construct texts, they use the words based on these stored contexts. In a way, they are an intermediary between a collective of original human speakers and a listener. They use the words the same way the original authors use them; that's why the words in the produced texts have their intelligible meaning, even though the model is not conscious and "doesn't know what it's talking about"; in fact, the word "know" is categorically inapplicable because there is no facility for knowing in the machine, beyond word contexts.

And yet it produces meaning. Wittgenstein, indeed.

  • 1
    +1 "Speaker and listener must share part of that context lest they could not communicate" AI like ChatGPT includes algorithms and statistical analysis to determine context based on the question asked. Commented Apr 16 at 16:33
  • 4
    "no facility for knowing in the machine, beyond word contexts." - so, you can train a LLM on transcripts of a board game like reversi. And then you can play reversi with that LLM. What more, we can read that LLM's mind and determine what it thinks the board looks like; we can go further, and perform mind surgery and change the state of the board in the LLM's mind, and it have it continue to play as if the board looked different. We can even write an impossible board state into its mind, and watch as it plays consistent with it. How is this not some evidence of "facility for knowing"?
    – Yakk
    Commented Apr 16 at 19:20
  • 1
    @Peter-ReinstateMonica, a relatively small chess model can play entirely novel games with an error rate of 0.2%, which would be in the same realm as humans. Humans also make errors. Commented Apr 17 at 2:55
  • 2
    This is a complete misunderstanding of what LLMs do and a complete misunderstanding of what language is, and a complete misunderstanding of what Wittgenstein was talking about when he talked about 'meaning of a word is its use in the language'. Wittgenstein was particularly talking bout the referents of words. Because LLMs just produce strings which look grammatical, their strings have no referents. They never intend any particular word to refer to any particular thing, and the individual words cannot be understood to 'stand for' anything in particular on the LLM's part. Commented Apr 17 at 11:29
  • 2
    @Araucaria With respect to the one concrete criticism: Yes, it is true that LLMs do not "intend" anything. They can't. What happens here is something linguistically very interesting: We can now observe a separation of intent and domain knowledge on one hand, which only people or other sentient beings have, from usage. The LLMs have slurped up all the usage (largely: context) and it turns out that they can (without doubt) produce "meaningful" (intelligible, informative) texts just with that context knowledge. The human intent and understanding -- meaning! -- is preserved in the context! Commented Apr 17 at 13:14

No. Wittgenstein would probably be the first to argue that the bare existence of a functioning Large Language Model does not by itself have any philosophical importance. The construction of an LLM is a mathematical or an engineering problem. Curb your enthusiasm, also, by considering that the recent proliferation of the development of LLMs have been largely the result of overcoming of barriers to purely practical limitaitons on the infrastructure of hardware, scale, and profitability; the theoretical infrastructure has been in place for a very long time. It would be silly to think that the large scale investment of capital into these technologies signals any kind of philosophical evidence of anything (does a functioning quantum computer prove the many worlds interpretation?).

Further, the LLM has no obvious bearing on the question of "meaning as use" because the LLM is not "using a word" in Wittgenstein's sense of 'use'. "Using a word" is a complex set of activities and social institutions which go well beyond the mechanical process of slotting in a word in the "right" spot to generate a coherent syntactical string that is responsive to an artificial context elaborated from the outside by determinate input parameters.

  • 2
    "It would be silly to think that [an artificial entity that gives exam-grade answers in any of the major languages to a vast spectrum of spoken questions across academic fields as a result of] the large scale investment of capital into these technologies signals any kind of philosophical evidence"!? Silliness is in the eye of the beholder, apparently. Commented Apr 16 at 8:57
  • 7
    Re: " the theoretical infrastructure has been in place for a very long time" not at all. The idea to scale such models up dramatically has existed for a very long time. Nobody seriously suspected that it would produce results anywhere near as good as it now does. No theoretical infrastructure ever suggested it would.
    – user73763
    Commented Apr 16 at 17:48
  • 4
    @Fattie We all mostly regurgitate what we have read and heard; "truly new" ideas are few and far in between. And calling mastership of all major languages a party trick is a text book example for Ray Kurzweil's observation that all skills mastered by automates are immediately regarded inferior and unimportant and certainly not a sign of intelligence. I can only state, hopefully undisputed, that passing college exams in 20 languages and a handful of subjects plus being a chess and Go master would have been a sign of the highest intelligence until about mid-20th century. Commented Apr 16 at 21:42
  • 2
    @Peter-ReinstateMonica - cheers, it's pointless to debate such a massive topic in comments. The go and chess enterprises (which are astonishing and amazing) have no connection at all to the "it really usually often sounds just like a normal language" output of chat.gpt. As I said, that's an awesome "trick" (or whatever one wishes to call it). You must realize though that when you ask chat.gpt a (say) programming question, you literally get one of (say) my answers from SO (if it's an incredibly tough question, rarely discussed) or just any old answer from SO (if a common question)...
    – Fattie
    Commented Apr 17 at 0:38
  • 1
    @Fattie, It's well-known that GPTs can play chess games that have never been played before. GPTs have also been observed to infer the concept of an Othello or chess board purely from the text of moves. Not infer the state of the board. Infer the concept. Search engine? Architecturally, and behaviourally, it has more in common with a human brain (connections, weights, signals) than it does with a search engine (indexes, keywords). Commented Apr 17 at 2:39

From SEP Wittgenstein 3.3 Meaning as Use:


Ascertainment of the use (of a word, of a proposition), however, is not given to any sort of constructive theory building, as in the Tractatus. Rather, when investigating meaning, the philosopher must “look and see” the variety of uses to which the word is put. An analogy with tools sheds light on the nature of words. When we think of tools in a toolbox, we do not fail to see their variety; but the “functions of words are as diverse as the functions of these objects” (PI 11).

Words can have different meanings depending on context. Using the tool analogy, a screwdriver has one meaning in the context of driving screws and a different meaning when prying open a can of paint.

For fun, I asked ChatGPT:

Is a screwdriver equivalent to a prybar?

The answer:

While both screwdrivers and pry bars can be used for leverage, they serve different primary purposes. Screwdrivers are primarily used for driving screws into or out of materials, while pry bars are specifically designed for prying or lifting objects apart. While a screwdriver can sometimes be used as a makeshift pry bar in certain situations, it's not its intended purpose, and using it as such can damage the screwdriver and possibly the object you're working on. It's generally safer and more effective to use tools for their intended purposes.

Based on this answer, the LLM used by ChatGPT both supports and incorporates Wittgenstein's position.

  • I feel there is a version or an anti-version of "Searl's Room" in there somewhere. Can a glorified highly compressed text dictionary really be said to "incorporate Wittgenstein's position" just because it is based on a server room and CPUs/GPUs rather than a library room and librarian who has never seen a prybar or screwdriver but has excellent mastery of the english language? 🤔 Commented Apr 16 at 19:48
  • 2
    @DavidTonhofer Searl (and his antipode Turing) are always present in LLM discussions. But there is a difference here to the room with a dictionary: An LLM is more than a dictionary, it stores context, that is word relations. This is where Wittgenstein's Untersuchung comes into play: The use -- which arguably is mainly context -- is meaning. In this sense, the LLMs are more than just dumb dictionaries because they actually store meaning. Commented Apr 17 at 7:58
  • 1
    @Peter-ReinstateMonica I thought it was interesting that the answer to my question to ChatGPT provided context without me explicitly asking for it. A simple (correct) answer without context is "No. Screwdrivers screw and Prybars pry". Commented Apr 17 at 16:39
  • 1
    @uhClem The issue is exactly in the definition of "meaning". I don't claim I understand Wittgenstein perfectly well; but I think he has a similar idea as Turing. Instead of looking at "hidden, real" things (Turing with intelligence, Wittgenstein with meaning, intent) we look at the performance, the use. Barks like a dog, smells like a dog, is a dog. And I am always astonished by people saying "but the model has just stored adjacency": How else do you think any of us learned to speak? Yes, there was also adjacency to emotion and physical objects (a wider context), but a context it was.- Commented Apr 19 at 14:53

No. As @transitionsynthesis says, LLMs do not come in contact with Wittgenstein and so can't support or refute him. Meaning (and Wittgenstein's "use") requires intention, which LLMs do not have. Strictly speaking they are only pattern-matching systems and don't even have anything to do with language -- language only happens to be the medium in which they match patterns. If their responses to our prompts seem meaningful to us that's a coincidence -- a hallucination on our part -- and the meaning we find is our use of the words, which involves the intention of our reception ("How can I apply this string of terms to my question about good ways to melt an egg?"). You might take that as support of Wittgenstein's proposition, but it has nothing to do with the LLMs -- it's the same use we make of all utterances we receive.

So, frame challenge: I think OP's consideration of LLMs in the question is an error. The last clause, "Is 'meaning' really just use?" is a good question. But since the speaker (or anyway, word-emitter) in the situation is an intentionless automaton the only "use" involved is that of the receiver. That being the case, the fact that this producer is an LLM is irrelevant -- the source might be ChatGPT or words drawn from a hat or a Magic 8-Ball.

So the success of LLMs at producing word-strings to which interpretations can be applied neither supports nor refutes the idea that "meaning is use". With respect to that proposition we're left in the same position we've always been in when considering "meaning" from only the recipient's point of view.

  • 1
    I think you should remove the disclaimer at the beginning, and link to transitionsynthesis's answer wirh the words "As @transitionsynthesis says", because this can function as a standalone answer to the question.
    – wizzwizz4
    Commented Apr 17 at 21:56
  • You're right @wizzwizz4, except that I didn't consider (or intend) this to directly answer OP's question; I only meant to address a faulty assumption. But since you suggest it I'll try to make the answer a little more complete.
    – uhClem
    Commented Apr 18 at 17:17
  • The hallucination is not in our AIs but in ourselves. I hope we don't take it beyond language to, driving or surgery or warfare or something.
    – Scott Rowe
    Commented Apr 19 at 13:23
  • You can say that again, @ScottRowe. But then... is there much evidence that we haven't always?
    – uhClem
    Commented Apr 19 at 13:44


Outside from the philosophy perspective which has been touched on by other answers, from a machine learning perspective, the meaning of words in a transformer is only partially defined by their use.

The attention mechanism does modify the meaning of a word based on its relative position to other words, but it does not define it. Each word starts with a default definition (embedding), which the attention mechanism then modifies. Thus words in a transformer do have a default meaning. The context, or use, only modifies the default meaning.

  • Where are these default definitions provided? You might just be providing a distinction (the embedding is not part of the LLM training), when you can use LLMs to build embeddings; they might just not be the most efficient way to do it.
    – Yakk
    Commented Apr 16 at 19:23
  • @Yakk Ultimately embeddings come from a learned one-to-one mapping between words and embeddings. Sometimes the mapping is learned at the same time as the rest of the network. Sometimes it uses a pretrained model like Word2Vec link. Importantly though the "mapping" between words and embeddings only ever takes in the word as input, never the context. Commented Apr 17 at 15:21

When I read LLM, I am using language, and often (if not always) it is meaningful to me. Though this may be his position, I wouldn't say this supports Wittgenstein's claim, as it seems too general to be refuted or confirmed my our use of LLM, not unless we take reading LLM as paradigmatic of all language.

In the Philosophical Investigations (156-178) and in the Brown Book (78-87) Wittgenstein gives a "deconstruction" of the "reading"

You could look there to find out if he thinks it is a language game.


Reading the above discussion I tend to a qualifies "yes" as an answer, especially I agree with Jo Wehler when he writes "Insofar is ChatGPT a good example how the meaning of a word is fixed by its use in different contexts.".

I used the argument in my paper https://doi.org/10.32388/9FH6AD . However, one should note that "language game" is usually an interactive enterprise with an active self, while LMM do simply passive observations of language. In a sense, they are like scientists that do not perform experiments. Hence, they need much more learning material than human beings to learn a language. But, if they are strong enough to provide a level that reliably passes the Turing test, then it can be said that (even passive) language game is enough to fix meaning.


The quote is from "Ludwig Wittgenstein: Philosophical Investigations, Section 43".

  • I simply take Wittgenstein's statement as exactly what is says: It gives an operational definition of the term "meaning of a word". The definition applies for a large class of cases. I do not search for further philosophical profoundness.
  • And similarly, AI-tools like ChatGPT learn how to use the words from the context of the textual examples from its training base. The tool has no possibility to learn the meaning from acting in the world.

Insofar is ChatGPT a good example how the meaning of a word is fixed by its use in different contexts.

  • 1
    And similarly, AI-tools like ChatGPT learn how to use the words from the context of the textual examples from its training base ... that should be "determine how to use the words", shurely? (Although they may "learn" during the Reinforcement Learning phase rather the pre-training phase) Commented Apr 16 at 19:53

Very much so, I would say. But of course Wittgenstein wrote in a time before the current success of AI and I think he would express himself a little differently in the present day world.

The problem is that approximate use is not approximate meaning. Modern day AI seems to get the use right - at certain angles and if you squint your eyes a bit, but there are still differences.

Some of these differences are related to what you can call 'subjective meaning': the (feeling of) "understanding". This line of argument is of course what Wittgenstein himself argues against with his "beetle in a box" argument: you cannot share what is going on in your head and "that whereof you cannot speak, you shall be silent", and "the limits of my language is the borders of my world" (which again brings us back to "meaning is use"). I agree with Wittgenstein, but think that a philosopher like Daniel Dennett has developed this line of thinking into a much more elaborate philosophy with his "intentional stance".

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .