“The limits of my language mean the limits of my world.”

— Ludwig Wittgenstein, Tractatus Logico-Philosophicus, 5.6

Large language models such as GPT, Claude, and Gemini invite questions that go beyond engineering. What does it mean to use language? Where does meaning come from? How much can the structure of language tell us about the world?

These questions run through the work of Ludwig Wittgenstein. His philosophy is often divided into two phases: an early account of language as a logical picture of reality, and a later account that places meaning in use.

I see a useful analogy between these two perspectives and two aspects of LLMs: pretraining and conversation. This is a philosophical lens, rather than a claim that the models implement Wittgenstein’s theories. It helps us examine both the structures they learn and what happens when people use them.

Two perspectives on language

In the Tractatus Logico-Philosophicus, Wittgenstein explores how propositions can represent reality. A proposition can picture a possible state of affairs because it shares a logical form with what it represents.

In Philosophical Investigations, his attention shifts to the many ways people use language. Asking a question, giving an order, telling a story, and making a promise are different activities. Their meaning depends on the practices in which they occur.

He calls these activities language games. Language has no single function that explains all of them. To understand an expression, we must look at its use within a shared practice.

The shift is from the structure that makes representation possible to the activities that give language a role in our lives.

Pretraining: learning linguistic structure

During pretraining, an LLM learns from large collections of text. For an autoregressive model, the immediate task is to predict the next token from the tokens that come before it. Repeated across many examples, this task produces a model of patterns and relationships in language.

The comparison with early Wittgenstein is suggestive:

• The training corpus supplies descriptions of the world.
• Token sequences express claims, questions, and possible situations.
• The model learns relationships that help it produce coherent language.

In this sense, pretraining builds a statistical picture of linguistic experience. It gives the model a structure through which it can represent and generate text.

The analogy has limits. A corpus contains fiction, errors, and contradictions as well as facts. Statistical relationships are not the same as the logical form described in the Tractatus. A model’s ability to produce a plausible sentence does not establish that the sentence is true.

Still, the parallel raises a useful question: how far can learned linguistic structure take us toward representing the world?

Conversation: taking part in language games

When we use an LLM in conversation, structure alone is not enough to describe what happens. The same words can call for different responses depending on the task, tone, and preceding dialogue.

An explanation for a beginner differs from one for an expert. A request for a joke differs from a request for a factual answer. Each exchange has expectations about what counts as an appropriate response.

This recalls later Wittgenstein:

• Meaning depends on the context of use.
• Conversation follows rules that are often left unstated.
• Language helps people perform tasks, as well as describe things.

A prompt is a move within an activity. The response gains significance through its role in that activity. People can use it to clarify an idea, revise a plan, or continue a story.

This does not mean that ordinary conversation changes the model’s weights. The context changes what it generates. Nor does participation remove the need for truth: in a factual exchange, accuracy is part of what makes a response useful.

A philosophical artifact

Wittgenstein’s work asks us to examine how language relates to thought and life. His early work emphasizes representation; his later work emphasizes use.

LLMs bring both questions into view. They learn linguistic structures, then generate responses within activities directed by people. Whether this amounts to human-like understanding remains a separate question. Their behavior gives us something concrete to examine without settling that debate first.

They are not philosophical theorists. But they can be viewed as philosophical artifacts: systems that make old questions about language visible in a new form.

Learning to “speak”

Children learn language through participation in life, not only through definitions. An LLM’s training is different from that experience, but the comparison directs attention to an important issue: successful language use requires more than a dictionary of symbols.

As a metaphor, we might say that a pretrained GPT model resembles a student of early Wittgenstein, learning structures through which language can picture the world. A conversational system such as ChatGPT brings later Wittgenstein into focus, because its responses acquire a role through use and interaction.

The two perspectives help explain why fluent text generation is only part of the story. We also need to examine what people do with the text, which expectations govern the exchange, and whether the response serves its purpose.

From philosophy to practice

In Philosophical Investigations, Wittgenstein urges us: “Don’t think, but look.” His point is to attend to actual language use instead of forcing it into a single abstract account.

The same approach is useful for AI. Alongside asking whether a model understands, we can observe how it participates in activities where meaning matters. What does its response help someone do? When does it fit the situation? When does it fail?

Artificial intelligence has not solved Wittgenstein’s questions. It has given us a new place to explore them.

---

Adapted from my article “Wittgenstein and Large Language Models: A Philosophical Lens on AI’s Rise — From Logical Pictures to Language Games”, originally published on LinkedIn on May 10, 2025. This blog edition retains the original publication date.

References

• Ludwig Wittgenstein. Tractatus Logico-Philosophicus. Proposition 5.6.
• Ludwig Wittgenstein. Philosophical Investigations. Sections 7, 43, and 66.