Work

Beyond Hallucination: Cultural Entropy in Multi-Turn Conversation with Large Language Models

Type: article
Abstract: Cultural entropy — the progressive reduction of the space of symbolic differences that can achieve visibility in a media environment — was proposed in a previous paper as a structural property of algorithmic media at the scale of populations. This paper argues that the same dynamic operates within the single human–AI exchange. Four independent strands of technical literature document the degradation of generative trajectories: exposure bias and error accumulation, hallucination snowballing within a response, entrenchment across the turns of a conversation, and the non-uniform processing of accumulated context. The paper reads this record as the phenomenology of an operationally closed code — consistency with context — that governs continuation independently of the validity of what is continued, as the population- scale code governed selection independently of symbolic value. It specifies what at this scale is literal — a Shannon entropy already used in production — and what remains heuristic — the cultural magnitude, which supervenes on the substrate without reducing to it — defending the two-level statute through McAllister's critique of effective complexity and Eco's information aesthetics. It then argues that closure retrodicts what the engineering literature leaves as coincidence, that every effective remedy operates from outside the exchange, and extends the mechanism, as a declared hypothesis, to interpretive corruption: the entrenchment of framings rather than facts, for which detection is conceptually undefined. Read against Monod, the configuration is conservation without selection; the human participant's return passage is where the missing term re-enters. Five open questions close the argument.
Language: English
Open Access status: green
OpenAlex ID: W7214512781
Cites: 14
Cited by:
Related to: 10