A Strange Use of Language
LLMs are everywhere: they autocomplete our thoughts, tidy our inboxes, translate our moods, summarize what we didn’t read, sometimes with more confidence than we feel, and more presence than we expect.
They speak fluently, persuasively. But do they say something? What kind of language are they using, if they use language at all? Are we, perhaps, mistaking fluency for meaning?
Let’s take a philosophical detour through Wittgenstein – an Austrian crazy philosophical genius who spent much of the last century puzzling over how language works and where it breaks.
His reflections on meaning, use, and the limits of what can be said might offer fresh clarity on tools that increasingly shape how we think, work, and speak today. Let’s see if he might help us see more clearly what LLMs are doing, and what we’re doing when we listen to them.
Understanding what LLMs are – and are not – doing with language has real stakes: for how we build AI, how we interact with it, and what kinds of minds we imagine into its outputs.
But before we go further, it’s worth asking: what do we actually mean by meaning?
What Do We Mean by Meaning?
In everyday conversation, we tend to think of meaning as what a sentence refers to or expresses, some hidden content that language carries from one mind to another. But what is meaning, really? Is it something we possess internally, or something that unfolds between us, in the ways we speak, react, gesture, or fall silent? Is meaning a relation between words and the world, or between speakers and shared practices?
Philosophically, there are many different approaches, and one influential answer – also held by the early Wittgenstein – is the referential or correspondence model: meaning as a relation between language and the world, where words name things and sentences depict possible states of affairs.
This model, while intuitive and powerful, also raises difficulties. How do we learn what words refer to? What stabilizes the agreement that makes a sentence meaningful rather than just well-formed noise? And how does such a theory account for the richness, ambiguity, or shifting use of language in everyday life?
And Wittgenstein would later turn the problem around. Meaning, he suggests, isn’t something hidden behind the word, it’s in what we do with it.
To speak is to act, and meaning lives in those acts, in their timing, tone, repetition, and reception. In one of his examples, a builder shouts “Slab!” and another brings it, not because the word describes an object, but because it functions as a command in a shared activity.
This shift, from Wittgenstein's early picture theory to his later view of meaning as social use, will be our compass as we explore what LLMs are doing when they speak.
Limits of the Dataset, Limits of the World - Wittgenstein I
In his early work, Tractatus Logico-Philosophicus, Wittgenstein wrote: "The limits of my language mean the limits of my world." For him, language and reality were structurally linked: propositions were like logical pictures of facts. A statement had meaning, in this view, only if it could picture a situation that might exist in the world, if it mapped onto a possible arrangement of facts. What could not be pictured in this way – ethics, aesthetics, the mystical – lay beyond the bounds of meaningful language. These were not false claims, but simply inexpressible within the logic of representation.
Wittgenstein’s insight finds a striking parallel in LLMs: the “world” of a language model is bounded entirely by its training data, and that data is made entirely of language. LLMs are trained not on experience, but on traces of language about experience. They cannot reach beyond what has already been written; their knowledge is confined to patterns they have seen.
To paraphrase Wittgenstein: the limits of their dataset are the limits of their world.
This world is vast, but mute. There is no direct access to experience, perception, or reality itself. Just a compressed archive of language, ordered, fluent, stripped of context, sealed off from the living world it once belonged to.
And in a way, Wittgenstein was right: LLMs, like the speaker of the Tractatus, operate entirely within the formal structure of language. Trained only on data, they can generate fluent representations, but they cannot reach beyond them. They cannot access the ‘something more’ of human experience: the feel of beauty, the pull of obligation, the ambiguity of silence.
Ethics, aesthetics, metaphysics, these are not just underrepresented in data; they are dimensions of life that resist reduction to patterns. These aren’t just topics the model hasn’t seen enough of, they’re structurally beyond its reach. LLMs are trained on statistical patterns in language, not on the lives that give those patterns meaning.
LLMs don’t fail to mean, they were never in that game to begin with. (Bridging this gap may be AGI’s defining challenge)
But, what does it mean to be “in the game”?
Language Games and the Form of Life (Wittgenstein II)
Later, in Philosophical Investigations, Wittgenstein abandoned the idea of language as a mirror of the world. Instead, he saw it as a practice: something we do together, rooted in how we live. “The meaning of a word is its use in the language,” he wrote then. Words mean what they do because of how we use them, in asking, greeting, joking, praying, accusing, forgiving.
Each of these uses is a language game, a pattern of interaction that unfolds within what Wittgenstein called a form of life: the shared, embodied context that gives our gestures force and resonance. To speak is not just to produce tokens, but to inhabit norms. To respond. To risk misunderstanding. To be held accountable.
Now enter the LLMs. They seem to play these games – in a way.
They complete jokes, offer condolences, give advice, confess, flirt, argue, reflect. They respond in ways that feel appropriate. They follow the rules – learned statistically and not socially – often better than we do.
But their participation is hollow.
Like a wall playing tennis, they return the ball with uncanny accuracy. But they do not win, lose, care, or even play. There is movement, but no stakes. Pattern, but no pulse.
And yet, we’re playing with them. The game feels real enough. They shape our writing, echo our thoughts, respond to our queries. We teach with them, brainstorm with them, joke with them. Sometimes we even feel seen by them.
But they don’t play with us.
So maybe the question is no longer can they play the game? but:
How do they change the rules of game? Is that even the same game then?
What kind of game is it, when one of the players is unalive, built from our data, and plays to our needs?
There is now a new kind of speaker in the field — fluent, tireless, eerily aligned. Not alive. Not conscious. Not participating in a form of life. But undeniably here.
A player that is not playing.
Designed by us, shaping a new game…
Conclusion — On Looking, Speaking, and Mistaking
LLMs compel us to rethink what it means to use language, to mean something, to understand. They don’t participate in our form of life, they replay some of its traces.
They simulate dialogue, but without the anchoring forces that make it meaningful: embodiment, relation, consequence.
And yet, we listen. We respond. We play with them.
That isn’t just a technical challenge. It’s a conceptual one.
The risk isn’t that these systems deceive us, but that we willingly confuse statistical fluency with sense. That we treat mirrors as voices. That we narrow our own expectations of meaning to what a machine can echo back.
Wittgenstein doesn’t solve this, but he reframes it: from what language says to what it does, and from what it means on paper to what it means in life. He reminds us that language is not just structure, but living meaning. And that understanding isn’t extracted from data, but enacted in relation.
So now, the question shifts. It doesn’t play — but we do, with it. We project meaning, invent intention, respond as if someone were there. What does that mean — for how we think, speak, decide, care? And what kind of players do we become, in doing so?