Does AI really communicate or just distribute information?

The Shannon information-theoretic framing of communication treats it as transmission of content from sender to receiver across a channel. The framing has been useful for engineering communication systems but has gradually been treated as a definition of what communication is. On this conflation, anything that moves content from one place to another communicates, and AI's prolific generation of content is communication.

Sociolinguistics, speech act theory, and pragmatic theory all point to a different definition. Communication is social action: a relational act between persons in which the act of saying something does work in the relationship — establishes shared understanding, performs a commitment, takes a position, registers acknowledgment. The content is part of the act but the act is not exhausted by the content. To communicate is to do something with someone, not just to convey something to someone.

This distinction is not metaphysical hairsplitting. It tracks what receivers experience as the difference between "being communicated with" and "being broadcast to." A friend's text registers as communication; a marketing email does not, even when both contain similar content, because one is a relational act and the other is content-distribution dressed up to look like a relational act. The receiver is competent at the distinction even when they cannot articulate it.

AI distributes information. It does not communicate, in the speech-act sense. The output is generated, not enacted. The "you" the output addresses is a placeholder for the prompt-source, not a person the AI is in relation with. The act of saying does no work in any relationship because there is no relationship for the act to do work in. The content lands as information; whatever relational uptake the receiver performs is performed entirely on their side, projected onto a system that did not perform a relational act.

The implication: the information-distribution framing of AI output systematically overstates what AI is doing. The right description is closer to publishing a generated artifact than to communicating with a person. The conversational interface obscures this — it presents content-distribution as if it were communication — but the structural difference is real and matters for how AI output should be received. Does AI writing lack the internal appeal to attention that humans use? is the post-level companion of this claim.

The strongest counterargument: communication-as-relational-act is too narrow — broadcast media communicates, even though the relation is asymmetric. The reply is that broadcast media has a speaker who took a position and bears responsibility for what was said; the relational structure is asymmetric but present. AI generation has no speaker who bears responsibility, so even the asymmetric communication structure is absent.