Is all human language use fundamentally communicative?

The acquisition story for human language is that children learn it through social interaction. Caregivers address children, children attempt to address back, the loop produces the gradual accumulation of linguistic competence. Acquisition theorists across schools (Tomasello on social pragmatics, Chomsky on universal grammar, social-interactionists on scaffolded learning) agree on this much: language is acquired in communicative contexts, even when the underlying mechanism debated.

What is less commonly remarked is that the communicative mode persists across all subsequent uses. There is no point at which humans transition from learning-language-by-communicating to using-language-without-communicating. Even the apparently solitary uses — private writing, internal monologue, talking to a dog — preserve the communicative orientation. Private writing addresses a future reader, often the writer themselves. Internal monologue addresses an internal interlocutor, often a hypothetical questioner or interlocutor. Talking to a dog addresses the dog as a hearer. The communicative scaffold is so deep in the operation that it persists when the addressee is hypothetical, internal, or absent.

This bears on what counts as "use of language." A common cognitive-science framing treats language as an information-encoding system that humans deploy for various purposes including but not limited to communication. The acquisition and persistence story above suggests this framing is upside down. Language is the medium through which the addressing-an-other operation happens; humans do not have access to language outside this operation, because they did not learn it outside this operation. The "uses" of language are all communicative uses; non-communicative use is a category that does not occur.

This produces a sharp asymmetry with LLMs. LLMs were trained on the surface output of human communicative acts (text), not on the communicative acts themselves. The training corpus is what communicative acts produce, not what produces them. So LLMs learned the surface form of language without learning the communicative scaffold that constitutes it for humans. They have access to language in a mode humans never have access to it: as strings without addressing. Are language models and human speakers doing the same thing? is the meta-discourse claim; this is the acquisition-theoretic claim that grounds it.

The diagnostic implication is that comparisons between LLM and human language use that ignore the acquisition difference are systematically misframed. Asking whether LLMs "understand" or "communicate" treats these as comparable verbs across the two systems. They are not — for humans, both verbs name properties of an addressing operation that LLMs do not perform. The verbs do not transpose.

The strongest counterargument: LLMs trained on dialogue corpora and fine-tuned with conversational data have learned something communicative-adjacent. The reply is that they have learned the surface form of communicative acts, not the operation; the asymmetry is in what was trained on, which cannot be repaired by more training of the same kind.

Source: Communication vs Language