Does AI generate genuine utterances or just text patterns?

If language is event, and subjecthood is produced within the event, then what AI generates is not a defective version of speech but a categorically different kind of output. Event-residue is text that carries the marks of communicative events — register, turn-structure, hedging, politeness markers, argument form — without having been produced by an event. The marks are inherited from the training distribution, where they were produced by actual communicative events between actual subjects. They are now reproduced without the event, the way a fossil preserves the form of a living thing without the life.

The human user supplies what the AI does not: orientation toward the text as communication. The user reads the output as a turn in an exchange, attributes communicative intent, infers beliefs and commitments, and responds accordingly. This is not illusion in the dismissive sense — it is genuine interpretive labor. The user is doing the work that would, in a real exchange, be distributed across two participants. In human-human communication, both parties orient toward mutual understanding. In human-AI interaction, the human orients unilaterally, and the AI generates text that happens to be interpretable by someone doing that work.

The result is a pseudo-event: something that has the structure of a communicative exchange from the user's side but is not an exchange from the system's side. The distinction matters because pseudo-events cannot generate the normative consequences real events generate. A real communicative exchange creates mutual commitments (you said X, and I can hold you to it); produces updated common ground (we now share an understanding); establishes accountability (if your claim was wrong, the falsity is attributable to you). The pseudo-event does none of these because the AI side does not hold commitments, does not share ground, and is not accountable in the sense the word requires.