Large language model (LLM) hallucination is almost universally framed as a pathology to be corrected. This paper proposes an alternative theoretical framework, arguing that hallucination is the textual equivalent of a computational glitch—a stochastic deviation that simultaneously reveals the underlying architecture of the system and produces semantically generative output. Drawing on three theoretical traditions—glitch art theory Menkman, 2010, Betancourt, 2016, poststructuralist authorship theory Barthes, 1977, and the philosophy of mind Searle, 1980—we develop what we term the Stochastic Glitch / Zero-Author (SGZA) framework. This framework holds that: (1) hallucination and creativity are mechanically identical processes, distinguished only by evaluative frame; (2) the LLM constitutes the most literal realization of the Barthesian scriptor —a zero-author that produces meaning-bearing text without intention; and (3) the interpretive labor released by this zero-authorship falls entirely to the reader, in a stronger sense than Barthes originally envisaged. Crucially, we distinguish this from claims that LLMs “are creative”: following Searle’s Chinese Room argument, we maintain that glitch-as-beauty resides in the accident, not in the agent. The SGZA framework offers a productive alternative to both the “bullshit” frame Hicks et al., 2024 and uncritical celebrations of machine creativity, with implications for AI aesthetics, human–AI collaboration, and the epistemology of generative systems.
Xavier Vinaixa Roselló (Thu,) studied this question.