Does AI text generation unfold through temporal reflection?

Human writing is temporal in a specific sense. A writer reflects in time, and the sentence that follows emerges from the time spent thinking about the sentence before it. The order of one thought after another is a temporal order: the later thought is later because something happened in the interval — consideration, revision, reaction. Time is constitutive of what the next thought becomes.

LLM generation also produces one token after another, but the ordering principle is different. The next token is selected by probability conditional on the prior sequence. Nothing happens in the interval between tokens except the computation of the next distribution. There is no reflection, no revision, no duration in which the claim is tested against what has come before. The order is sequential — strictly — but it is not temporal in the reflective sense. It is computed ordering, not lived ordering.

This matters for how AI-generated text relates to discourse. Human discourse is temporal because it is made of moves that respond to prior moves, anticipate future moves, and take time to make. AI text has the surface form of such a move but lacks the temporal structure that would give it its meaning. The text appears, in a sense, all at once — even though it was produced sequentially — because the production time is not the time of anyone's thinking.

This is adjacent to but distinct from Does LLM generation explore competing claims while producing text?. Smoothness describes the absence of turbulent counter-exploration. Atemporality describes the absence of duration-in-reflection. Both properties follow from the same generative process but bear on different dimensions of what makes discourse discursive.

Source: Epistemic Inflation