Language Game

In the AI Story

The concept begins with observation. Wittgenstein asks the reader to look at ordinary speech — a builder calling for slabs, a child learning to name colors, strangers exchanging greetings — and notice how radically different these activities are. They share the surface feature of using words; they share almost nothing else. What counts as a successful move differs in each. What counts as understanding differs. What counts as the same word being used for the same purpose differs.

This diversity defeats the search for a single essence of language. There is no one thing language does. There are many things language does, and they resemble each other the way members of a family resemble each other — through overlapping similarities rather than shared essence. The name for this pattern is family resemblance, and it is the positive concept that replaces the Tractarian appeal to shared logical form.

Applied to the AI language interface, the concept reveals something no previous framework captured. The pre-2025 computational paradigm supported exactly one language game: formal instruction. Command and execution. Two moves, fully determined by specification. The large language model participates in many games — describing, speculating, suggesting, revising, questioning — shifting between them as the dialogue moves. When Edo Segal describes the experience of working with Claude as feeling met, what he is describing, in Wittgenstein's terms, is having his language game recognized and played along with.

Whether the machine really plays the game or merely produces appropriate-seeming moves is the question the beetle in the box argument dissolves. For the game to proceed, the moves must be accepted by the other players as moves. Whether some inner accompaniment validates the moves from the machine's side drops out of the game. The question of meaning is a question about the moves, not about hidden inner states.

Origin

The term Sprachspiel appears in the 1930s manuscripts and is developed across the Philosophical Investigations. Wittgenstein was partly inspired by watching children play — observing how language and activity grew together in a single texture of learning.

Key Ideas

Language and activity interwoven. A language game is a form of life in miniature; words and actions are inseparable within it.

Meaning is role in the game. The meaning of a word is its function within the specific activity in which it is employed.

Grammar as rules of the game. The grammar of a language game is not syntax but the set of conditions that determine what counts as a move.

Many games, not one language. Ordering, describing, joking, consoling are structurally different; they cannot be reduced to a single underlying activity.

AI as partial participant. The machine can make moves the other player recognizes as appropriate; whether this constitutes playing the game is the question the beetle dissolves.

Debates & Critiques

Whether AI participation in language games is genuine or merely simulated depends on where one locates the criteria for participation. Wittgenstein's own framework tends to locate them in the public acceptance of moves rather than in hidden inner states — a framework that cuts against both naive enthusiasm and reflexive skepticism about what large language models have learned.